diff options
42 files changed, 3888 insertions, 2214 deletions
diff --git a/media/libaom/README_MCP b/media/libaom/README_MCP index ad04810a30..803417195b 100644 --- a/media/libaom/README_MCP +++ b/media/libaom/README_MCP @@ -9,4 +9,4 @@ The upstream aom git repository is: https://aomedia.googlesource.com/aom -The version used was tagged 2.0.2. +The version used was tagged 3.4.0. diff --git a/media/libaom/aom_version.h b/media/libaom/aom_version.h index 92de763be8..c03b348d01 100644 --- a/media/libaom/aom_version.h +++ b/media/libaom/aom_version.h @@ -1,7 +1,7 @@ -#define VERSION_MAJOR 2 -#define VERSION_MINOR 0 -#define VERSION_PATCH 2 +#define VERSION_MAJOR 3 +#define VERSION_MINOR 4 +#define VERSION_PATCH 0 #define VERSION_EXTRA "" #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH)) -#define VERSION_STRING_NOSP "v2.0.2" -#define VERSION_STRING " v2.0.2" +#define VERSION_STRING_NOSP "v3.4.0" +#define VERSION_STRING " v3.4.0" diff --git a/media/libaom/config/aom_version.h b/media/libaom/config/aom_version.h index 92de763be8..c03b348d01 100644 --- a/media/libaom/config/aom_version.h +++ b/media/libaom/config/aom_version.h @@ -1,7 +1,7 @@ -#define VERSION_MAJOR 2 -#define VERSION_MINOR 0 -#define VERSION_PATCH 2 +#define VERSION_MAJOR 3 +#define VERSION_MINOR 4 +#define VERSION_PATCH 0 #define VERSION_EXTRA "" #define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH)) -#define VERSION_STRING_NOSP "v2.0.2" -#define VERSION_STRING " v2.0.2" +#define VERSION_STRING_NOSP "v3.4.0" +#define VERSION_STRING " v3.4.0" diff --git a/media/libaom/config/generic/config/aom_config.asm b/media/libaom/config/generic/config/aom_config.asm index af2521fa4c..53cdceae13 100644 --- a/media/libaom/config/generic/config/aom_config.asm +++ b/media/libaom/config/generic/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 0 HAVE_SSSE3 equ 0 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/generic/config/aom_config.h b/media/libaom/config/generic/config/aom_config.h index 6e300df375..fa9f0d16de 100644 --- a/media/libaom/config/generic/config/aom_config.h +++ b/media/libaom/config/generic/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/generic/config/aom_dsp_rtcd.h b/media/libaom/config/generic/config/aom_dsp_rtcd.h index f769d1180d..ff786a39e8 100644 --- a/media/libaom/config/generic/config/aom_dsp_rtcd.h +++ b/media/libaom/config/generic/config/aom_dsp_rtcd.h @@ -31,13 +31,16 @@ void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0 void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); #define aom_blend_a64_vmask aom_blend_a64_vmask_c +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); #define aom_convolve8_horiz aom_convolve8_horiz_c void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); #define aom_convolve8_vert aom_convolve8_vert_c -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); #define aom_convolve_copy aom_convolve_copy_c void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -55,9 +58,6 @@ void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c @@ -115,9 +115,6 @@ void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c @@ -175,9 +172,6 @@ void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c @@ -235,9 +229,6 @@ void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c @@ -295,9 +286,6 @@ void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_c -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_c @@ -358,7 +346,7 @@ void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); #define aom_highbd_convolve8_vert aom_highbd_convolve8_vert_c -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); #define aom_highbd_convolve_copy aom_highbd_convolve_copy_c void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -376,9 +364,6 @@ void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c @@ -436,9 +421,6 @@ void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, con void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c @@ -496,9 +478,6 @@ void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const ui void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c @@ -556,9 +535,6 @@ void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c @@ -616,9 +592,6 @@ void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c @@ -724,9 +697,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -784,9 +754,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -844,9 +811,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -904,9 +868,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -964,9 +925,6 @@ void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c @@ -1018,48 +976,72 @@ void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_c +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_14_quad aom_lpf_horizontal_14_quad_c + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_c void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_c +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_c + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_6 aom_lpf_horizontal_6_c void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_c +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_6_quad aom_lpf_horizontal_6_quad_c + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_c void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_c +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_8_quad aom_lpf_horizontal_8_quad_c + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_14 aom_lpf_vertical_14_c void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_c +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_c + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_4 aom_lpf_vertical_4_c void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_c +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_c + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_6 aom_lpf_vertical_6_c void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_c +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_c + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_8 aom_lpf_vertical_8_c void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_c +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_c + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c @@ -1075,9 +1057,6 @@ void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c @@ -1120,6 +1099,9 @@ void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_scaled_2d aom_scaled_2d_c + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c @@ -1135,9 +1117,6 @@ void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c @@ -1195,9 +1174,6 @@ void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c @@ -1255,9 +1231,6 @@ void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c @@ -1315,9 +1288,6 @@ void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_c -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_32x16 aom_v_predictor_32x16_c diff --git a/media/libaom/config/generic/config/av1_rtcd.h b/media/libaom/config/generic/config/av1_rtcd.h index dd8536b690..36510c8a56 100644 --- a/media/libaom/config/generic/config/av1_rtcd.h +++ b/media/libaom/config/generic/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -91,27 +91,6 @@ void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE m int64_t av1_calc_frame_error_c(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); #define av1_calc_frame_error av1_calc_frame_error_c -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -#define av1_convolve_2d_copy_sr av1_convolve_2d_copy_sr_c - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); #define av1_convolve_2d_scale av1_convolve_2d_scale_c @@ -121,22 +100,22 @@ void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); #define av1_convolve_horiz_rs av1_convolve_horiz_rs_c -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); #define av1_convolve_x_sr av1_convolve_x_sr_c -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); #define av1_convolve_y_sr av1_convolve_y_sr_c void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); #define av1_dist_wtd_convolve_2d av1_dist_wtd_convolve_2d_c -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); #define av1_dist_wtd_convolve_2d_copy av1_dist_wtd_convolve_2d_copy_c -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); #define av1_dist_wtd_convolve_x av1_dist_wtd_convolve_x_c -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); #define av1_dist_wtd_convolve_y av1_dist_wtd_convolve_y_c void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); @@ -166,9 +145,6 @@ void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -#define av1_highbd_convolve_2d_copy_sr av1_highbd_convolve_2d_copy_sr_c - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_convolve_2d_scale av1_highbd_convolve_2d_scale_c @@ -184,22 +160,22 @@ void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); #define av1_highbd_convolve_horiz_rs av1_highbd_convolve_horiz_rs_c -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_convolve_x_sr av1_highbd_convolve_x_sr_c -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); #define av1_highbd_convolve_y_sr av1_highbd_convolve_y_sr_c void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_2d av1_highbd_dist_wtd_convolve_2d_c -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_2d_copy av1_highbd_dist_wtd_convolve_2d_copy_c -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_x av1_highbd_dist_wtd_convolve_x_c -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_y av1_highbd_dist_wtd_convolve_y_c void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -214,9 +190,30 @@ void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); #define av1_highbd_inv_txfm_add av1_highbd_inv_txfm_add_c +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); #define av1_highbd_inv_txfm_add_16x4 av1_highbd_inv_txfm_add_16x4_c +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); #define av1_highbd_inv_txfm_add_4x16 av1_highbd_inv_txfm_add_4x16_c @@ -226,6 +223,21 @@ void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); #define av1_highbd_inv_txfm_add_4x8 av1_highbd_inv_txfm_add_4x8_c +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); #define av1_highbd_inv_txfm_add_8x4 av1_highbd_inv_txfm_add_8x4_c @@ -304,12 +316,15 @@ void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); #define av1_inv_txfm_add av1_inv_txfm_add_c +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +#define av1_resize_and_extend_frame av1_resize_and_extend_frame_c + void av1_round_shift_array_c(int32_t *arr, int size, int bit); #define av1_round_shift_array av1_round_shift_array_c int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); #define av1_selfguided_restoration av1_selfguided_restoration_c void av1_upsample_intra_edge_c(uint8_t *p, int sz); @@ -330,12 +345,36 @@ void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); #define cdef_copy_rect8_8bit_to_16bit cdef_copy_rect8_8bit_to_16bit_c -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -#define cdef_filter_block cdef_filter_block_c +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_16_0 cdef_filter_16_0_c + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_16_1 cdef_filter_16_1_c + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_16_2 cdef_filter_16_2_c + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_16_3 cdef_filter_16_3_c + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_8_0 cdef_filter_8_0_c + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_8_1 cdef_filter_8_1_c + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_8_2 cdef_filter_8_2_c + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +#define cdef_filter_8_3 cdef_filter_8_3_c int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); #define cdef_find_dir cdef_find_dir_c +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +#define cdef_find_dir_dual cdef_find_dir_dual_c + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); #define cfl_get_luma_subsampling_420_hbd cfl_get_luma_subsampling_420_hbd_c diff --git a/media/libaom/config/linux/arm/config/aom_config.asm b/media/libaom/config/linux/arm/config/aom_config.asm index a5d8cc9b99..fa3f6286c5 100644 --- a/media/libaom/config/linux/arm/config/aom_config.asm +++ b/media/libaom/config/linux/arm/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ .equ CONFIG_AV1_DECODER, 1 .equ CONFIG_AV1_ENCODER, 0 .equ CONFIG_AV1_HIGHBITDEPTH, 1 +.equ CONFIG_AV1_TEMPORAL_DENOISING, 0 .equ CONFIG_BIG_ENDIAN, 0 +.equ CONFIG_BITRATE_ACCURACY, 0 +.equ CONFIG_BITRATE_ACCURACY_BL, 0 .equ CONFIG_BITSTREAM_DEBUG, 0 .equ CONFIG_COEFFICIENT_RANGE_CHECKING, 0 .equ CONFIG_COLLECT_COMPONENT_TIMING, 0 @@ -28,33 +31,40 @@ .equ CONFIG_DEBUG, 0 .equ CONFIG_DENOISE, 1 .equ CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8, 1 -.equ CONFIG_DIST_8X8, 0 .equ CONFIG_ENTROPY_STATS, 0 +.equ CONFIG_EXCLUDE_SIMD_MISMATCH, 0 +.equ CONFIG_FPMT_TEST, 0 +.equ CONFIG_FRAME_PARALLEL_ENCODE, 0 +.equ CONFIG_FRAME_PARALLEL_ENCODE_2, 0 .equ CONFIG_GCC, 1 .equ CONFIG_GCOV, 0 .equ CONFIG_GPROF, 0 -.equ CONFIG_HTB_TRELLIS, 0 .equ CONFIG_INSPECTION, 0 .equ CONFIG_INTERNAL_STATS, 0 .equ CONFIG_INTER_STATS_ONLY, 0 .equ CONFIG_LIBYUV, 0 -.equ CONFIG_LPF_MASK, 0 .equ CONFIG_MAX_DECODE_PROFILE, 2 .equ CONFIG_MISMATCH_DEBUG, 0 .equ CONFIG_MULTITHREAD, 1 .equ CONFIG_NN_V2, 0 .equ CONFIG_NORMAL_TILE_MODE, 0 +.equ CONFIG_OPTICAL_FLOW_API, 0 .equ CONFIG_OS_SUPPORT, 1 +.equ CONFIG_PARTITION_SEARCH_ORDER, 0 .equ CONFIG_PIC, 1 +.equ CONFIG_RATECTRL_LOG, 0 +.equ CONFIG_RD_COMMAND, 0 .equ CONFIG_RD_DEBUG, 0 .equ CONFIG_REALTIME_ONLY, 0 +.equ CONFIG_RT_ML_PARTITIONING, 0 .equ CONFIG_RUNTIME_CPU_DETECT, 1 .equ CONFIG_SHARED, 0 -.equ CONFIG_SHARP_SETTINGS, 0 .equ CONFIG_SIZE_LIMIT, 0 .equ CONFIG_SPATIAL_RESAMPLING, 1 .equ CONFIG_SPEED_STATS, 0 -.equ CONFIG_SUPERRES_IN_RECODE, 1 +.equ CONFIG_TFLITE, 0 +.equ CONFIG_THREE_PASS, 0 +.equ CONFIG_TUNE_BUTTERAUGLI, 0 .equ CONFIG_TUNE_VMAF, 0 .equ CONFIG_WEBM_IO, 0 .equ DECODE_HEIGHT_LIMIT, 0 @@ -77,4 +87,5 @@ .equ HAVE_SSSE3, 0 .equ HAVE_VSX, 0 .equ HAVE_WXWIDGETS, 0 +.equ STATIC_LINK_JXL, 0 .section .note.GNU-stack,"",%progbits
\ No newline at end of file diff --git a/media/libaom/config/linux/arm/config/aom_config.h b/media/libaom/config/linux/arm/config/aom_config.h index 5fc909a91c..4a3dc92ba6 100644 --- a/media/libaom/config/linux/arm/config/aom_config.h +++ b/media/libaom/config/linux/arm/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 1 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h index c6e420acba..9498c97b63 100644 --- a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h +++ b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h @@ -33,14 +33,18 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_neon(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); #define aom_convolve8_horiz aom_convolve8_horiz_c void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); #define aom_convolve8_vert aom_convolve8_vert_c -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -#define aom_convolve_copy aom_convolve_copy_c +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -58,9 +62,6 @@ void aom_dc_128_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c @@ -122,9 +123,6 @@ void aom_dc_left_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c @@ -186,9 +184,6 @@ void aom_dc_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c @@ -250,9 +245,6 @@ void aom_dc_top_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c @@ -314,9 +306,6 @@ void aom_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_c -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_c @@ -380,7 +369,7 @@ void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); #define aom_highbd_convolve8_vert aom_highbd_convolve8_vert_c -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); #define aom_highbd_convolve_copy aom_highbd_convolve_copy_c void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -398,9 +387,6 @@ void aom_highbd_dc_128_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c @@ -458,9 +444,6 @@ void aom_highbd_dc_left_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, con void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c @@ -519,9 +502,6 @@ void aom_highbd_dc_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const ui void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c @@ -583,9 +563,6 @@ void aom_highbd_dc_top_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c @@ -643,9 +620,6 @@ void aom_highbd_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c @@ -689,352 +663,448 @@ void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint1 #define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c void aom_highbd_lpf_horizontal_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_horizontal_14 aom_highbd_lpf_horizontal_14_c +void aom_highbd_lpf_horizontal_14_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_horizontal_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd); -#define aom_highbd_lpf_horizontal_14_dual aom_highbd_lpf_horizontal_14_dual_c +void aom_highbd_lpf_horizontal_14_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,int bd); void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_c +void aom_highbd_lpf_horizontal_4_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_horizontal_4_dual aom_highbd_lpf_horizontal_4_dual_c +void aom_highbd_lpf_horizontal_4_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_horizontal_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_horizontal_6 aom_highbd_lpf_horizontal_6_c +void aom_highbd_lpf_horizontal_6_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_horizontal_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_horizontal_6_dual aom_highbd_lpf_horizontal_6_dual_c +void aom_highbd_lpf_horizontal_6_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_c +void aom_highbd_lpf_horizontal_8_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_horizontal_8_dual aom_highbd_lpf_horizontal_8_dual_c +void aom_highbd_lpf_horizontal_8_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_vertical_14_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_vertical_14 aom_highbd_lpf_vertical_14_c +void aom_highbd_lpf_vertical_14_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_14)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_vertical_14_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_vertical_14_dual aom_highbd_lpf_vertical_14_dual_c +void aom_highbd_lpf_vertical_14_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_14_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_c +void aom_highbd_lpf_vertical_4_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_4)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_vertical_4_dual aom_highbd_lpf_vertical_4_dual_c +void aom_highbd_lpf_vertical_4_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_vertical_6_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_vertical_6 aom_highbd_lpf_vertical_6_c +void aom_highbd_lpf_vertical_6_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_6)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_vertical_6_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_vertical_6_dual aom_highbd_lpf_vertical_6_dual_c +void aom_highbd_lpf_vertical_6_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_6_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); -#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_c +void aom_highbd_lpf_vertical_8_neon(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_8)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd); void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); -#define aom_highbd_lpf_vertical_8_dual aom_highbd_lpf_vertical_8_dual_c +void aom_highbd_lpf_vertical_8_dual_neon(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); +RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd); void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c +void aom_highbd_paeth_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c +void aom_highbd_paeth_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_16x4 aom_highbd_paeth_predictor_16x4_c +void aom_highbd_paeth_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_16x64 aom_highbd_paeth_predictor_16x64_c +void aom_highbd_paeth_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c - -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c +void aom_highbd_paeth_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c +void aom_highbd_paeth_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c +void aom_highbd_paeth_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_32x64 aom_highbd_paeth_predictor_32x64_c +void aom_highbd_paeth_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_32x8 aom_highbd_paeth_predictor_32x8_c +void aom_highbd_paeth_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_4x16 aom_highbd_paeth_predictor_4x16_c +void aom_highbd_paeth_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c +void aom_highbd_paeth_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c +void aom_highbd_paeth_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_64x16 aom_highbd_paeth_predictor_64x16_c +void aom_highbd_paeth_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_64x32 aom_highbd_paeth_predictor_64x32_c +void aom_highbd_paeth_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_64x64 aom_highbd_paeth_predictor_64x64_c +void aom_highbd_paeth_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c +void aom_highbd_paeth_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_8x32 aom_highbd_paeth_predictor_8x32_c +void aom_highbd_paeth_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c +void aom_highbd_paeth_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c +void aom_highbd_paeth_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_paeth_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c +void aom_highbd_smooth_h_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c +void aom_highbd_smooth_h_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_16x4 aom_highbd_smooth_h_predictor_16x4_c +void aom_highbd_smooth_h_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_16x64 aom_highbd_smooth_h_predictor_16x64_c +void aom_highbd_smooth_h_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c - -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c +void aom_highbd_smooth_h_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c +void aom_highbd_smooth_h_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c +void aom_highbd_smooth_h_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_32x64 aom_highbd_smooth_h_predictor_32x64_c +void aom_highbd_smooth_h_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_32x8 aom_highbd_smooth_h_predictor_32x8_c +void aom_highbd_smooth_h_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_4x16 aom_highbd_smooth_h_predictor_4x16_c +void aom_highbd_smooth_h_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c +void aom_highbd_smooth_h_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c +void aom_highbd_smooth_h_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_64x16 aom_highbd_smooth_h_predictor_64x16_c +void aom_highbd_smooth_h_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_64x32 aom_highbd_smooth_h_predictor_64x32_c +void aom_highbd_smooth_h_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_64x64 aom_highbd_smooth_h_predictor_64x64_c +void aom_highbd_smooth_h_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c +void aom_highbd_smooth_h_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_8x32 aom_highbd_smooth_h_predictor_8x32_c +void aom_highbd_smooth_h_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c +void aom_highbd_smooth_h_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c +void aom_highbd_smooth_h_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_h_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c +void aom_highbd_smooth_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c +void aom_highbd_smooth_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_16x4 aom_highbd_smooth_predictor_16x4_c +void aom_highbd_smooth_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_16x64 aom_highbd_smooth_predictor_16x64_c +void aom_highbd_smooth_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c - -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c +void aom_highbd_smooth_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c +void aom_highbd_smooth_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c +void aom_highbd_smooth_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_32x64 aom_highbd_smooth_predictor_32x64_c +void aom_highbd_smooth_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_32x8 aom_highbd_smooth_predictor_32x8_c +void aom_highbd_smooth_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_4x16 aom_highbd_smooth_predictor_4x16_c +void aom_highbd_smooth_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c +void aom_highbd_smooth_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c +void aom_highbd_smooth_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_64x16 aom_highbd_smooth_predictor_64x16_c +void aom_highbd_smooth_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_64x32 aom_highbd_smooth_predictor_64x32_c +void aom_highbd_smooth_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_64x64 aom_highbd_smooth_predictor_64x64_c +void aom_highbd_smooth_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c +void aom_highbd_smooth_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_8x32 aom_highbd_smooth_predictor_8x32_c +void aom_highbd_smooth_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c +void aom_highbd_smooth_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c +void aom_highbd_smooth_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c +void aom_highbd_smooth_v_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c +void aom_highbd_smooth_v_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_16x4 aom_highbd_smooth_v_predictor_16x4_c +void aom_highbd_smooth_v_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_16x64 aom_highbd_smooth_v_predictor_16x64_c +void aom_highbd_smooth_v_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c - -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c +void aom_highbd_smooth_v_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c +void aom_highbd_smooth_v_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c +void aom_highbd_smooth_v_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_32x64 aom_highbd_smooth_v_predictor_32x64_c +void aom_highbd_smooth_v_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_32x8 aom_highbd_smooth_v_predictor_32x8_c +void aom_highbd_smooth_v_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_4x16 aom_highbd_smooth_v_predictor_4x16_c +void aom_highbd_smooth_v_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c +void aom_highbd_smooth_v_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c +void aom_highbd_smooth_v_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_64x16 aom_highbd_smooth_v_predictor_64x16_c +void aom_highbd_smooth_v_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_64x32 aom_highbd_smooth_v_predictor_64x32_c +void aom_highbd_smooth_v_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_64x64 aom_highbd_smooth_v_predictor_64x64_c +void aom_highbd_smooth_v_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c +void aom_highbd_smooth_v_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_8x32 aom_highbd_smooth_v_predictor_8x32_c +void aom_highbd_smooth_v_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c +void aom_highbd_smooth_v_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c +void aom_highbd_smooth_v_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_smooth_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_c +void aom_highbd_v_predictor_16x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_c +void aom_highbd_v_predictor_16x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_16x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_16x4 aom_highbd_v_predictor_16x4_c +void aom_highbd_v_predictor_16x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_16x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_16x64 aom_highbd_v_predictor_16x64_c +void aom_highbd_v_predictor_16x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_16x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c - -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c +void aom_highbd_v_predictor_16x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c +void aom_highbd_v_predictor_32x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_c +void aom_highbd_v_predictor_32x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_32x64 aom_highbd_v_predictor_32x64_c +void aom_highbd_v_predictor_32x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_32x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_32x8 aom_highbd_v_predictor_32x8_c +void aom_highbd_v_predictor_32x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_32x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_4x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_4x16 aom_highbd_v_predictor_4x16_c +void aom_highbd_v_predictor_4x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_4x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_c +void aom_highbd_v_predictor_4x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_4x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_c +void aom_highbd_v_predictor_4x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_4x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_64x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_64x16 aom_highbd_v_predictor_64x16_c +void aom_highbd_v_predictor_64x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_64x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_64x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_64x32 aom_highbd_v_predictor_64x32_c +void aom_highbd_v_predictor_64x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_64x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_64x64_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_64x64 aom_highbd_v_predictor_64x64_c +void aom_highbd_v_predictor_64x64_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_64x64)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_c +void aom_highbd_v_predictor_8x16_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_8x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_8x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_8x32 aom_highbd_v_predictor_8x32_c +void aom_highbd_v_predictor_8x32_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_8x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_c +void aom_highbd_v_predictor_8x4_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_8x4)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_c +void aom_highbd_v_predictor_8x8_neon(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); +RTCD_EXTERN void (*aom_highbd_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params); void aom_lowbd_blend_a64_d16_mask_neon(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params); @@ -1045,296 +1115,404 @@ void aom_lpf_horizontal_14_neon(uint8_t *s, int pitch, const uint8_t *blimit, co RTCD_EXTERN void (*aom_lpf_horizontal_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_c +void aom_lpf_horizontal_14_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_c +void aom_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_c +void aom_lpf_horizontal_6_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_c +void aom_lpf_horizontal_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_c +void aom_lpf_vertical_14_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_c +void aom_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_c +void aom_lpf_vertical_6_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); -#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_c +void aom_lpf_vertical_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); + +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c +void aom_paeth_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c +void aom_paeth_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_16x4 aom_paeth_predictor_16x4_c +void aom_paeth_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_16x64 aom_paeth_predictor_16x64_c +void aom_paeth_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c - -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c +void aom_paeth_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c +void aom_paeth_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c +void aom_paeth_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_32x64 aom_paeth_predictor_32x64_c +void aom_paeth_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_32x8 aom_paeth_predictor_32x8_c +void aom_paeth_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_4x16 aom_paeth_predictor_4x16_c +void aom_paeth_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c +void aom_paeth_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c +void aom_paeth_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_64x16 aom_paeth_predictor_64x16_c +void aom_paeth_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_64x32 aom_paeth_predictor_64x32_c +void aom_paeth_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_64x64 aom_paeth_predictor_64x64_c +void aom_paeth_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c +void aom_paeth_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_8x32 aom_paeth_predictor_8x32_c +void aom_paeth_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c +void aom_paeth_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c +void aom_paeth_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); + +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c +void aom_smooth_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c +void aom_smooth_h_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_16x4 aom_smooth_h_predictor_16x4_c +void aom_smooth_h_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_16x64 aom_smooth_h_predictor_16x64_c +void aom_smooth_h_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c - -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c +void aom_smooth_h_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c +void aom_smooth_h_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c +void aom_smooth_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_32x64 aom_smooth_h_predictor_32x64_c +void aom_smooth_h_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_32x8 aom_smooth_h_predictor_32x8_c +void aom_smooth_h_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_4x16 aom_smooth_h_predictor_4x16_c +void aom_smooth_h_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c +void aom_smooth_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c +void aom_smooth_h_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_64x16 aom_smooth_h_predictor_64x16_c +void aom_smooth_h_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_64x32 aom_smooth_h_predictor_64x32_c +void aom_smooth_h_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_64x64 aom_smooth_h_predictor_64x64_c +void aom_smooth_h_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c +void aom_smooth_h_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_8x32 aom_smooth_h_predictor_8x32_c +void aom_smooth_h_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c +void aom_smooth_h_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c +void aom_smooth_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_16x16 aom_smooth_predictor_16x16_c +void aom_smooth_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c +void aom_smooth_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_16x4 aom_smooth_predictor_16x4_c +void aom_smooth_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_16x64 aom_smooth_predictor_16x64_c +void aom_smooth_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c - -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c +void aom_smooth_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c +void aom_smooth_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_32x32 aom_smooth_predictor_32x32_c +void aom_smooth_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_32x64 aom_smooth_predictor_32x64_c +void aom_smooth_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_32x8 aom_smooth_predictor_32x8_c +void aom_smooth_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_4x16 aom_smooth_predictor_4x16_c +void aom_smooth_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_4x4 aom_smooth_predictor_4x4_c +void aom_smooth_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c +void aom_smooth_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_64x16 aom_smooth_predictor_64x16_c +void aom_smooth_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_64x32 aom_smooth_predictor_64x32_c +void aom_smooth_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_64x64 aom_smooth_predictor_64x64_c +void aom_smooth_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c +void aom_smooth_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_8x32 aom_smooth_predictor_8x32_c +void aom_smooth_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_8x4 aom_smooth_predictor_8x4_c +void aom_smooth_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_8x8 aom_smooth_predictor_8x8_c +void aom_smooth_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c +void aom_smooth_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c +void aom_smooth_v_predictor_16x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_16x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_16x4 aom_smooth_v_predictor_16x4_c +void aom_smooth_v_predictor_16x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_16x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_16x64 aom_smooth_v_predictor_16x64_c +void aom_smooth_v_predictor_16x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_16x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c - -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c +void aom_smooth_v_predictor_16x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c +void aom_smooth_v_predictor_32x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c +void aom_smooth_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_32x64 aom_smooth_v_predictor_32x64_c +void aom_smooth_v_predictor_32x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_32x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_32x8 aom_smooth_v_predictor_32x8_c +void aom_smooth_v_predictor_32x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_32x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_4x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_4x16 aom_smooth_v_predictor_4x16_c +void aom_smooth_v_predictor_4x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_4x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c +void aom_smooth_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c +void aom_smooth_v_predictor_4x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_64x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_64x16 aom_smooth_v_predictor_64x16_c +void aom_smooth_v_predictor_64x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_64x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_64x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_64x32 aom_smooth_v_predictor_64x32_c +void aom_smooth_v_predictor_64x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_64x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_64x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_64x64 aom_smooth_v_predictor_64x64_c +void aom_smooth_v_predictor_64x64_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_64x64)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c +void aom_smooth_v_predictor_8x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_8x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_8x32 aom_smooth_v_predictor_8x32_c +void aom_smooth_v_predictor_8x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_8x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c +void aom_smooth_v_predictor_8x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c +void aom_smooth_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +RTCD_EXTERN void (*aom_smooth_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1352,9 +1530,6 @@ void aom_v_predictor_16x64_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_c -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_32x16 aom_v_predictor_32x16_c @@ -1416,6 +1591,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_NEON) aom_blend_a64_hmask = aom_blend_a64_hmask_neon; aom_blend_a64_vmask = aom_blend_a64_vmask_c; if (flags & HAS_NEON) aom_blend_a64_vmask = aom_blend_a64_vmask_neon; + aom_convolve_copy = aom_convolve_copy_c; + if (flags & HAS_NEON) aom_convolve_copy = aom_convolve_copy_neon; aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c; if (flags & HAS_NEON) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_neon; aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_c; @@ -1466,24 +1643,432 @@ static void setup_rtcd_internal(void) if (flags & HAS_NEON) aom_highbd_dc_predictor_64x64 = aom_highbd_dc_predictor_64x64_neon; aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_c; if (flags & HAS_NEON) aom_highbd_dc_predictor_8x8 = aom_highbd_dc_predictor_8x8_neon; + aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_14 = aom_highbd_lpf_horizontal_14_neon; + aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_14_dual = aom_highbd_lpf_horizontal_14_dual_neon; + aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_4 = aom_highbd_lpf_horizontal_4_neon; + aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_neon; + aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_6 = aom_highbd_lpf_horizontal_6_neon; + aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_6_dual = aom_highbd_lpf_horizontal_6_dual_neon; + aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_8 = aom_highbd_lpf_horizontal_8_neon; + aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_neon; + aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_14 = aom_highbd_lpf_vertical_14_neon; + aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_14_dual = aom_highbd_lpf_vertical_14_dual_neon; + aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_4 = aom_highbd_lpf_vertical_4_neon; + aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_neon; + aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_6 = aom_highbd_lpf_vertical_6_neon; + aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_6_dual = aom_highbd_lpf_vertical_6_dual_neon; + aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_8 = aom_highbd_lpf_vertical_8_neon; + aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_c; + if (flags & HAS_NEON) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_neon; + aom_highbd_paeth_predictor_16x16 = aom_highbd_paeth_predictor_16x16_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x16 = aom_highbd_paeth_predictor_16x16_neon; + aom_highbd_paeth_predictor_16x32 = aom_highbd_paeth_predictor_16x32_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x32 = aom_highbd_paeth_predictor_16x32_neon; + aom_highbd_paeth_predictor_16x4 = aom_highbd_paeth_predictor_16x4_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x4 = aom_highbd_paeth_predictor_16x4_neon; + aom_highbd_paeth_predictor_16x64 = aom_highbd_paeth_predictor_16x64_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x64 = aom_highbd_paeth_predictor_16x64_neon; + aom_highbd_paeth_predictor_16x8 = aom_highbd_paeth_predictor_16x8_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_16x8 = aom_highbd_paeth_predictor_16x8_neon; + aom_highbd_paeth_predictor_32x16 = aom_highbd_paeth_predictor_32x16_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x16 = aom_highbd_paeth_predictor_32x16_neon; + aom_highbd_paeth_predictor_32x32 = aom_highbd_paeth_predictor_32x32_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x32 = aom_highbd_paeth_predictor_32x32_neon; + aom_highbd_paeth_predictor_32x64 = aom_highbd_paeth_predictor_32x64_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x64 = aom_highbd_paeth_predictor_32x64_neon; + aom_highbd_paeth_predictor_32x8 = aom_highbd_paeth_predictor_32x8_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_32x8 = aom_highbd_paeth_predictor_32x8_neon; + aom_highbd_paeth_predictor_4x16 = aom_highbd_paeth_predictor_4x16_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x16 = aom_highbd_paeth_predictor_4x16_neon; + aom_highbd_paeth_predictor_4x4 = aom_highbd_paeth_predictor_4x4_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x4 = aom_highbd_paeth_predictor_4x4_neon; + aom_highbd_paeth_predictor_4x8 = aom_highbd_paeth_predictor_4x8_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_4x8 = aom_highbd_paeth_predictor_4x8_neon; + aom_highbd_paeth_predictor_64x16 = aom_highbd_paeth_predictor_64x16_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x16 = aom_highbd_paeth_predictor_64x16_neon; + aom_highbd_paeth_predictor_64x32 = aom_highbd_paeth_predictor_64x32_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x32 = aom_highbd_paeth_predictor_64x32_neon; + aom_highbd_paeth_predictor_64x64 = aom_highbd_paeth_predictor_64x64_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_64x64 = aom_highbd_paeth_predictor_64x64_neon; + aom_highbd_paeth_predictor_8x16 = aom_highbd_paeth_predictor_8x16_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x16 = aom_highbd_paeth_predictor_8x16_neon; + aom_highbd_paeth_predictor_8x32 = aom_highbd_paeth_predictor_8x32_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x32 = aom_highbd_paeth_predictor_8x32_neon; + aom_highbd_paeth_predictor_8x4 = aom_highbd_paeth_predictor_8x4_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x4 = aom_highbd_paeth_predictor_8x4_neon; + aom_highbd_paeth_predictor_8x8 = aom_highbd_paeth_predictor_8x8_c; + if (flags & HAS_NEON) aom_highbd_paeth_predictor_8x8 = aom_highbd_paeth_predictor_8x8_neon; + aom_highbd_smooth_h_predictor_16x16 = aom_highbd_smooth_h_predictor_16x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x16 = aom_highbd_smooth_h_predictor_16x16_neon; + aom_highbd_smooth_h_predictor_16x32 = aom_highbd_smooth_h_predictor_16x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x32 = aom_highbd_smooth_h_predictor_16x32_neon; + aom_highbd_smooth_h_predictor_16x4 = aom_highbd_smooth_h_predictor_16x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x4 = aom_highbd_smooth_h_predictor_16x4_neon; + aom_highbd_smooth_h_predictor_16x64 = aom_highbd_smooth_h_predictor_16x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x64 = aom_highbd_smooth_h_predictor_16x64_neon; + aom_highbd_smooth_h_predictor_16x8 = aom_highbd_smooth_h_predictor_16x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_16x8 = aom_highbd_smooth_h_predictor_16x8_neon; + aom_highbd_smooth_h_predictor_32x16 = aom_highbd_smooth_h_predictor_32x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x16 = aom_highbd_smooth_h_predictor_32x16_neon; + aom_highbd_smooth_h_predictor_32x32 = aom_highbd_smooth_h_predictor_32x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x32 = aom_highbd_smooth_h_predictor_32x32_neon; + aom_highbd_smooth_h_predictor_32x64 = aom_highbd_smooth_h_predictor_32x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x64 = aom_highbd_smooth_h_predictor_32x64_neon; + aom_highbd_smooth_h_predictor_32x8 = aom_highbd_smooth_h_predictor_32x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_32x8 = aom_highbd_smooth_h_predictor_32x8_neon; + aom_highbd_smooth_h_predictor_4x16 = aom_highbd_smooth_h_predictor_4x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x16 = aom_highbd_smooth_h_predictor_4x16_neon; + aom_highbd_smooth_h_predictor_4x4 = aom_highbd_smooth_h_predictor_4x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x4 = aom_highbd_smooth_h_predictor_4x4_neon; + aom_highbd_smooth_h_predictor_4x8 = aom_highbd_smooth_h_predictor_4x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_4x8 = aom_highbd_smooth_h_predictor_4x8_neon; + aom_highbd_smooth_h_predictor_64x16 = aom_highbd_smooth_h_predictor_64x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x16 = aom_highbd_smooth_h_predictor_64x16_neon; + aom_highbd_smooth_h_predictor_64x32 = aom_highbd_smooth_h_predictor_64x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x32 = aom_highbd_smooth_h_predictor_64x32_neon; + aom_highbd_smooth_h_predictor_64x64 = aom_highbd_smooth_h_predictor_64x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_64x64 = aom_highbd_smooth_h_predictor_64x64_neon; + aom_highbd_smooth_h_predictor_8x16 = aom_highbd_smooth_h_predictor_8x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x16 = aom_highbd_smooth_h_predictor_8x16_neon; + aom_highbd_smooth_h_predictor_8x32 = aom_highbd_smooth_h_predictor_8x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x32 = aom_highbd_smooth_h_predictor_8x32_neon; + aom_highbd_smooth_h_predictor_8x4 = aom_highbd_smooth_h_predictor_8x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x4 = aom_highbd_smooth_h_predictor_8x4_neon; + aom_highbd_smooth_h_predictor_8x8 = aom_highbd_smooth_h_predictor_8x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_h_predictor_8x8 = aom_highbd_smooth_h_predictor_8x8_neon; + aom_highbd_smooth_predictor_16x16 = aom_highbd_smooth_predictor_16x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x16 = aom_highbd_smooth_predictor_16x16_neon; + aom_highbd_smooth_predictor_16x32 = aom_highbd_smooth_predictor_16x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x32 = aom_highbd_smooth_predictor_16x32_neon; + aom_highbd_smooth_predictor_16x4 = aom_highbd_smooth_predictor_16x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x4 = aom_highbd_smooth_predictor_16x4_neon; + aom_highbd_smooth_predictor_16x64 = aom_highbd_smooth_predictor_16x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x64 = aom_highbd_smooth_predictor_16x64_neon; + aom_highbd_smooth_predictor_16x8 = aom_highbd_smooth_predictor_16x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_16x8 = aom_highbd_smooth_predictor_16x8_neon; + aom_highbd_smooth_predictor_32x16 = aom_highbd_smooth_predictor_32x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x16 = aom_highbd_smooth_predictor_32x16_neon; + aom_highbd_smooth_predictor_32x32 = aom_highbd_smooth_predictor_32x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x32 = aom_highbd_smooth_predictor_32x32_neon; + aom_highbd_smooth_predictor_32x64 = aom_highbd_smooth_predictor_32x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x64 = aom_highbd_smooth_predictor_32x64_neon; + aom_highbd_smooth_predictor_32x8 = aom_highbd_smooth_predictor_32x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_32x8 = aom_highbd_smooth_predictor_32x8_neon; + aom_highbd_smooth_predictor_4x16 = aom_highbd_smooth_predictor_4x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x16 = aom_highbd_smooth_predictor_4x16_neon; + aom_highbd_smooth_predictor_4x4 = aom_highbd_smooth_predictor_4x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x4 = aom_highbd_smooth_predictor_4x4_neon; + aom_highbd_smooth_predictor_4x8 = aom_highbd_smooth_predictor_4x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_4x8 = aom_highbd_smooth_predictor_4x8_neon; + aom_highbd_smooth_predictor_64x16 = aom_highbd_smooth_predictor_64x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x16 = aom_highbd_smooth_predictor_64x16_neon; + aom_highbd_smooth_predictor_64x32 = aom_highbd_smooth_predictor_64x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x32 = aom_highbd_smooth_predictor_64x32_neon; + aom_highbd_smooth_predictor_64x64 = aom_highbd_smooth_predictor_64x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_64x64 = aom_highbd_smooth_predictor_64x64_neon; + aom_highbd_smooth_predictor_8x16 = aom_highbd_smooth_predictor_8x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x16 = aom_highbd_smooth_predictor_8x16_neon; + aom_highbd_smooth_predictor_8x32 = aom_highbd_smooth_predictor_8x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x32 = aom_highbd_smooth_predictor_8x32_neon; + aom_highbd_smooth_predictor_8x4 = aom_highbd_smooth_predictor_8x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x4 = aom_highbd_smooth_predictor_8x4_neon; + aom_highbd_smooth_predictor_8x8 = aom_highbd_smooth_predictor_8x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_predictor_8x8 = aom_highbd_smooth_predictor_8x8_neon; + aom_highbd_smooth_v_predictor_16x16 = aom_highbd_smooth_v_predictor_16x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x16 = aom_highbd_smooth_v_predictor_16x16_neon; + aom_highbd_smooth_v_predictor_16x32 = aom_highbd_smooth_v_predictor_16x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x32 = aom_highbd_smooth_v_predictor_16x32_neon; + aom_highbd_smooth_v_predictor_16x4 = aom_highbd_smooth_v_predictor_16x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x4 = aom_highbd_smooth_v_predictor_16x4_neon; + aom_highbd_smooth_v_predictor_16x64 = aom_highbd_smooth_v_predictor_16x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x64 = aom_highbd_smooth_v_predictor_16x64_neon; + aom_highbd_smooth_v_predictor_16x8 = aom_highbd_smooth_v_predictor_16x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_16x8 = aom_highbd_smooth_v_predictor_16x8_neon; + aom_highbd_smooth_v_predictor_32x16 = aom_highbd_smooth_v_predictor_32x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x16 = aom_highbd_smooth_v_predictor_32x16_neon; + aom_highbd_smooth_v_predictor_32x32 = aom_highbd_smooth_v_predictor_32x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x32 = aom_highbd_smooth_v_predictor_32x32_neon; + aom_highbd_smooth_v_predictor_32x64 = aom_highbd_smooth_v_predictor_32x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x64 = aom_highbd_smooth_v_predictor_32x64_neon; + aom_highbd_smooth_v_predictor_32x8 = aom_highbd_smooth_v_predictor_32x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_32x8 = aom_highbd_smooth_v_predictor_32x8_neon; + aom_highbd_smooth_v_predictor_4x16 = aom_highbd_smooth_v_predictor_4x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x16 = aom_highbd_smooth_v_predictor_4x16_neon; + aom_highbd_smooth_v_predictor_4x4 = aom_highbd_smooth_v_predictor_4x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x4 = aom_highbd_smooth_v_predictor_4x4_neon; + aom_highbd_smooth_v_predictor_4x8 = aom_highbd_smooth_v_predictor_4x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_4x8 = aom_highbd_smooth_v_predictor_4x8_neon; + aom_highbd_smooth_v_predictor_64x16 = aom_highbd_smooth_v_predictor_64x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x16 = aom_highbd_smooth_v_predictor_64x16_neon; + aom_highbd_smooth_v_predictor_64x32 = aom_highbd_smooth_v_predictor_64x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x32 = aom_highbd_smooth_v_predictor_64x32_neon; + aom_highbd_smooth_v_predictor_64x64 = aom_highbd_smooth_v_predictor_64x64_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_64x64 = aom_highbd_smooth_v_predictor_64x64_neon; + aom_highbd_smooth_v_predictor_8x16 = aom_highbd_smooth_v_predictor_8x16_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x16 = aom_highbd_smooth_v_predictor_8x16_neon; + aom_highbd_smooth_v_predictor_8x32 = aom_highbd_smooth_v_predictor_8x32_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x32 = aom_highbd_smooth_v_predictor_8x32_neon; + aom_highbd_smooth_v_predictor_8x4 = aom_highbd_smooth_v_predictor_8x4_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x4 = aom_highbd_smooth_v_predictor_8x4_neon; + aom_highbd_smooth_v_predictor_8x8 = aom_highbd_smooth_v_predictor_8x8_c; + if (flags & HAS_NEON) aom_highbd_smooth_v_predictor_8x8 = aom_highbd_smooth_v_predictor_8x8_neon; + aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_16x16 = aom_highbd_v_predictor_16x16_neon; + aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_16x32 = aom_highbd_v_predictor_16x32_neon; + aom_highbd_v_predictor_16x4 = aom_highbd_v_predictor_16x4_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_16x4 = aom_highbd_v_predictor_16x4_neon; + aom_highbd_v_predictor_16x64 = aom_highbd_v_predictor_16x64_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_16x64 = aom_highbd_v_predictor_16x64_neon; + aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_16x8 = aom_highbd_v_predictor_16x8_neon; + aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_32x16 = aom_highbd_v_predictor_32x16_neon; + aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_32x32 = aom_highbd_v_predictor_32x32_neon; + aom_highbd_v_predictor_32x64 = aom_highbd_v_predictor_32x64_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_32x64 = aom_highbd_v_predictor_32x64_neon; + aom_highbd_v_predictor_32x8 = aom_highbd_v_predictor_32x8_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_32x8 = aom_highbd_v_predictor_32x8_neon; + aom_highbd_v_predictor_4x16 = aom_highbd_v_predictor_4x16_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_4x16 = aom_highbd_v_predictor_4x16_neon; + aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_4x4 = aom_highbd_v_predictor_4x4_neon; + aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_4x8 = aom_highbd_v_predictor_4x8_neon; + aom_highbd_v_predictor_64x16 = aom_highbd_v_predictor_64x16_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_64x16 = aom_highbd_v_predictor_64x16_neon; + aom_highbd_v_predictor_64x32 = aom_highbd_v_predictor_64x32_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_64x32 = aom_highbd_v_predictor_64x32_neon; + aom_highbd_v_predictor_64x64 = aom_highbd_v_predictor_64x64_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_64x64 = aom_highbd_v_predictor_64x64_neon; + aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_8x16 = aom_highbd_v_predictor_8x16_neon; + aom_highbd_v_predictor_8x32 = aom_highbd_v_predictor_8x32_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_8x32 = aom_highbd_v_predictor_8x32_neon; + aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_8x4 = aom_highbd_v_predictor_8x4_neon; + aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_c; + if (flags & HAS_NEON) aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_neon; aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c; if (flags & HAS_NEON) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_neon; aom_lpf_horizontal_14 = aom_lpf_horizontal_14_c; if (flags & HAS_NEON) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_neon; + aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c; + if (flags & HAS_NEON) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_neon; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c; + if (flags & HAS_NEON) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_neon; aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c; if (flags & HAS_NEON) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_neon; + aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c; + if (flags & HAS_NEON) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_neon; + aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c; + if (flags & HAS_NEON) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_neon; aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c; if (flags & HAS_NEON) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_neon; + aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c; + if (flags & HAS_NEON) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_neon; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c; + if (flags & HAS_NEON) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_neon; aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c; if (flags & HAS_NEON) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_neon; + aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c; + if (flags & HAS_NEON) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_neon; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c; + if (flags & HAS_NEON) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_neon; aom_lpf_vertical_14 = aom_lpf_vertical_14_c; if (flags & HAS_NEON) aom_lpf_vertical_14 = aom_lpf_vertical_14_neon; + aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c; + if (flags & HAS_NEON) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_neon; + aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c; + if (flags & HAS_NEON) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_neon; aom_lpf_vertical_4 = aom_lpf_vertical_4_c; if (flags & HAS_NEON) aom_lpf_vertical_4 = aom_lpf_vertical_4_neon; + aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c; + if (flags & HAS_NEON) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_neon; + aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c; + if (flags & HAS_NEON) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_neon; aom_lpf_vertical_6 = aom_lpf_vertical_6_c; if (flags & HAS_NEON) aom_lpf_vertical_6 = aom_lpf_vertical_6_neon; + aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c; + if (flags & HAS_NEON) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_neon; + aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c; + if (flags & HAS_NEON) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_neon; aom_lpf_vertical_8 = aom_lpf_vertical_8_c; if (flags & HAS_NEON) aom_lpf_vertical_8 = aom_lpf_vertical_8_neon; + aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c; + if (flags & HAS_NEON) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_neon; + aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c; + if (flags & HAS_NEON) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_neon; + aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; + if (flags & HAS_NEON) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_neon; + aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c; + if (flags & HAS_NEON) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_neon; + aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_c; + if (flags & HAS_NEON) aom_paeth_predictor_16x4 = aom_paeth_predictor_16x4_neon; + aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_c; + if (flags & HAS_NEON) aom_paeth_predictor_16x64 = aom_paeth_predictor_16x64_neon; + aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c; + if (flags & HAS_NEON) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_neon; + aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c; + if (flags & HAS_NEON) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_neon; + aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c; + if (flags & HAS_NEON) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_neon; + aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_c; + if (flags & HAS_NEON) aom_paeth_predictor_32x64 = aom_paeth_predictor_32x64_neon; + aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_c; + if (flags & HAS_NEON) aom_paeth_predictor_32x8 = aom_paeth_predictor_32x8_neon; + aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_c; + if (flags & HAS_NEON) aom_paeth_predictor_4x16 = aom_paeth_predictor_4x16_neon; + aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c; + if (flags & HAS_NEON) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_neon; + aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c; + if (flags & HAS_NEON) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_neon; + aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_c; + if (flags & HAS_NEON) aom_paeth_predictor_64x16 = aom_paeth_predictor_64x16_neon; + aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_c; + if (flags & HAS_NEON) aom_paeth_predictor_64x32 = aom_paeth_predictor_64x32_neon; + aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_c; + if (flags & HAS_NEON) aom_paeth_predictor_64x64 = aom_paeth_predictor_64x64_neon; + aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c; + if (flags & HAS_NEON) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_neon; + aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_c; + if (flags & HAS_NEON) aom_paeth_predictor_8x32 = aom_paeth_predictor_8x32_neon; + aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c; + if (flags & HAS_NEON) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_neon; + aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; + if (flags & HAS_NEON) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_neon; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_NEON) aom_scaled_2d = aom_scaled_2d_neon; + aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_neon; + aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_neon; + aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_16x4 = aom_smooth_h_predictor_16x4_neon; + aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_16x64 = aom_smooth_h_predictor_16x64_neon; + aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_16x8 = aom_smooth_h_predictor_16x8_neon; + aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_32x16 = aom_smooth_h_predictor_32x16_neon; + aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_32x32 = aom_smooth_h_predictor_32x32_neon; + aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_32x64 = aom_smooth_h_predictor_32x64_neon; + aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_32x8 = aom_smooth_h_predictor_32x8_neon; + aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_4x16 = aom_smooth_h_predictor_4x16_neon; + aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_4x4 = aom_smooth_h_predictor_4x4_neon; + aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_4x8 = aom_smooth_h_predictor_4x8_neon; + aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_64x16 = aom_smooth_h_predictor_64x16_neon; + aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_64x32 = aom_smooth_h_predictor_64x32_neon; + aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_64x64 = aom_smooth_h_predictor_64x64_neon; + aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_8x16 = aom_smooth_h_predictor_8x16_neon; + aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_8x32 = aom_smooth_h_predictor_8x32_neon; + aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_8x4 = aom_smooth_h_predictor_8x4_neon; + aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_c; + if (flags & HAS_NEON) aom_smooth_h_predictor_8x8 = aom_smooth_h_predictor_8x8_neon; + aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c; + if (flags & HAS_NEON) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_neon; + aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c; + if (flags & HAS_NEON) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_neon; + aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_c; + if (flags & HAS_NEON) aom_smooth_predictor_16x4 = aom_smooth_predictor_16x4_neon; + aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_c; + if (flags & HAS_NEON) aom_smooth_predictor_16x64 = aom_smooth_predictor_16x64_neon; + aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c; + if (flags & HAS_NEON) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_neon; + aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c; + if (flags & HAS_NEON) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_neon; + aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c; + if (flags & HAS_NEON) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_neon; + aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_c; + if (flags & HAS_NEON) aom_smooth_predictor_32x64 = aom_smooth_predictor_32x64_neon; + aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_c; + if (flags & HAS_NEON) aom_smooth_predictor_32x8 = aom_smooth_predictor_32x8_neon; + aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_c; + if (flags & HAS_NEON) aom_smooth_predictor_4x16 = aom_smooth_predictor_4x16_neon; + aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c; + if (flags & HAS_NEON) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_neon; + aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c; + if (flags & HAS_NEON) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_neon; + aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_c; + if (flags & HAS_NEON) aom_smooth_predictor_64x16 = aom_smooth_predictor_64x16_neon; + aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_c; + if (flags & HAS_NEON) aom_smooth_predictor_64x32 = aom_smooth_predictor_64x32_neon; + aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_c; + if (flags & HAS_NEON) aom_smooth_predictor_64x64 = aom_smooth_predictor_64x64_neon; + aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c; + if (flags & HAS_NEON) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_neon; + aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_c; + if (flags & HAS_NEON) aom_smooth_predictor_8x32 = aom_smooth_predictor_8x32_neon; + aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c; + if (flags & HAS_NEON) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_neon; + aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c; + if (flags & HAS_NEON) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_neon; + aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_16x16 = aom_smooth_v_predictor_16x16_neon; + aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_16x32 = aom_smooth_v_predictor_16x32_neon; + aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_16x4 = aom_smooth_v_predictor_16x4_neon; + aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_16x64 = aom_smooth_v_predictor_16x64_neon; + aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_16x8 = aom_smooth_v_predictor_16x8_neon; + aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_32x16 = aom_smooth_v_predictor_32x16_neon; + aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_32x32 = aom_smooth_v_predictor_32x32_neon; + aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_32x64 = aom_smooth_v_predictor_32x64_neon; + aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_32x8 = aom_smooth_v_predictor_32x8_neon; + aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_4x16 = aom_smooth_v_predictor_4x16_neon; + aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_4x4 = aom_smooth_v_predictor_4x4_neon; + aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_4x8 = aom_smooth_v_predictor_4x8_neon; + aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_64x16 = aom_smooth_v_predictor_64x16_neon; + aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_64x32 = aom_smooth_v_predictor_64x32_neon; + aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_64x64 = aom_smooth_v_predictor_64x64_neon; + aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_8x16 = aom_smooth_v_predictor_8x16_neon; + aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_8x32 = aom_smooth_v_predictor_8x32_neon; + aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_8x4 = aom_smooth_v_predictor_8x4_neon; + aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_c; + if (flags & HAS_NEON) aom_smooth_v_predictor_8x8 = aom_smooth_v_predictor_8x8_neon; aom_v_predictor_16x16 = aom_v_predictor_16x16_c; if (flags & HAS_NEON) aom_v_predictor_16x16 = aom_v_predictor_16x16_neon; aom_v_predictor_32x32 = aom_v_predictor_32x32_c; diff --git a/media/libaom/config/linux/arm/config/av1_rtcd.h b/media/libaom/config/linux/arm/config/av1_rtcd.h index 7a346efd26..747a74ae5a 100644 --- a/media/libaom/config/linux/arm/config/av1_rtcd.h +++ b/media/libaom/config/linux/arm/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -93,28 +93,6 @@ void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE m int64_t av1_calc_frame_error_c(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); #define av1_calc_frame_error av1_calc_frame_error_c -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); #define av1_convolve_2d_scale av1_convolve_2d_scale_c @@ -125,38 +103,41 @@ RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8 void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); #define av1_convolve_horiz_rs av1_convolve_horiz_rs_c -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); -#define av1_dr_prediction_z1 av1_dr_prediction_z1_c +void av1_dr_prediction_z1_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); -#define av1_dr_prediction_z2 av1_dr_prediction_z2_c +void av1_dr_prediction_z2_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); -#define av1_dr_prediction_z3 av1_dr_prediction_z3_c +void av1_dr_prediction_z3_neon(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength); #define av1_filter_intra_edge av1_filter_intra_edge_c @@ -165,7 +146,8 @@ void av1_filter_intra_edge_high_c(uint16_t *p, int sz, int strength); #define av1_filter_intra_edge_high av1_filter_intra_edge_high_c void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode); -#define av1_filter_intra_predictor av1_filter_intra_predictor_c +void av1_filter_intra_predictor_neon(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode); +RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode); void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8 av1_highbd_convolve8_c @@ -176,9 +158,6 @@ void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -#define av1_highbd_convolve_2d_copy_sr av1_highbd_convolve_2d_copy_sr_c - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_convolve_2d_scale av1_highbd_convolve_2d_scale_c @@ -194,22 +173,22 @@ void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); #define av1_highbd_convolve_horiz_rs av1_highbd_convolve_horiz_rs_c -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_convolve_x_sr av1_highbd_convolve_x_sr_c -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); #define av1_highbd_convolve_y_sr av1_highbd_convolve_y_sr_c void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_2d av1_highbd_dist_wtd_convolve_2d_c -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_2d_copy av1_highbd_dist_wtd_convolve_2d_copy_c -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_x av1_highbd_dist_wtd_convolve_x_c -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); #define av1_highbd_dist_wtd_convolve_y av1_highbd_dist_wtd_convolve_y_c void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -222,25 +201,76 @@ void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c void av1_highbd_inv_txfm_add_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add av1_highbd_inv_txfm_add_c +void av1_highbd_inv_txfm_add_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_16x4 av1_highbd_inv_txfm_add_16x4_c +void av1_highbd_inv_txfm_add_16x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_32x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_32x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_32x64_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x64)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_4x16 av1_highbd_inv_txfm_add_4x16_c +void av1_highbd_inv_txfm_add_4x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_4x4 av1_highbd_inv_txfm_add_4x4_c +void av1_highbd_inv_txfm_add_4x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_4x8 av1_highbd_inv_txfm_add_4x8_c +void av1_highbd_inv_txfm_add_4x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x32_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_64x32)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x64_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_64x64)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_8x16_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_8x4 av1_highbd_inv_txfm_add_8x4_c +void av1_highbd_inv_txfm_add_8x4_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); -#define av1_highbd_inv_txfm_add_8x8 av1_highbd_inv_txfm_add_8x8_c +void av1_highbd_inv_txfm_add_8x8_neon(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c @@ -258,76 +288,98 @@ void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stri #define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c +void av1_inv_txfm2d_add_16x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_16x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c +void av1_inv_txfm2d_add_16x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_16x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c +void av1_inv_txfm2d_add_16x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_16x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c +void av1_inv_txfm2d_add_16x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_16x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c +void av1_inv_txfm2d_add_32x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_32x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c +void av1_inv_txfm2d_add_32x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c +void av1_inv_txfm2d_add_32x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_32x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c +void av1_inv_txfm2d_add_32x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_32x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c +void av1_inv_txfm2d_add_4x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_4x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_4x4 av1_inv_txfm2d_add_4x4_c +void av1_inv_txfm2d_add_4x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c +void av1_inv_txfm2d_add_4x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_4x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c +void av1_inv_txfm2d_add_64x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_64x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c +void av1_inv_txfm2d_add_64x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_64x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c +void av1_inv_txfm2d_add_64x64_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_64x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c +void av1_inv_txfm2d_add_8x16_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_8x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c +void av1_inv_txfm2d_add_8x32_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_8x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c +void av1_inv_txfm2d_add_8x4_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_8x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); -#define av1_inv_txfm2d_add_8x8 av1_inv_txfm2d_add_8x8_c +void av1_inv_txfm2d_add_8x8_neon(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); +RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd); void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); void av1_inv_txfm_add_neon(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_neon(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_neon(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); #define av1_upsample_intra_edge av1_upsample_intra_edge_c @@ -351,14 +403,46 @@ void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t * void cdef_copy_rect8_8bit_to_16bit_neon(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_neon(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_neon(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_neon(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_neon(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_neon(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_neon(TX_SIZE tx_size); RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size); @@ -411,8 +495,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_NEON) av1_apply_selfguided_restoration = av1_apply_selfguided_restoration_neon; av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c; if (flags & HAS_NEON) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_neon; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c; - if (flags & HAS_NEON) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_neon; av1_convolve_2d_sr = av1_convolve_2d_sr_c; if (flags & HAS_NEON) av1_convolve_2d_sr = av1_convolve_2d_sr_neon; av1_convolve_x_sr = av1_convolve_x_sr_c; @@ -427,8 +509,84 @@ static void setup_rtcd_internal(void) if (flags & HAS_NEON) av1_dist_wtd_convolve_x = av1_dist_wtd_convolve_x_neon; av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_c; if (flags & HAS_NEON) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_neon; + av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_NEON) av1_dr_prediction_z1 = av1_dr_prediction_z1_neon; + av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_NEON) av1_dr_prediction_z2 = av1_dr_prediction_z2_neon; + av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_NEON) av1_dr_prediction_z3 = av1_dr_prediction_z3_neon; + av1_filter_intra_predictor = av1_filter_intra_predictor_c; + if (flags & HAS_NEON) av1_filter_intra_predictor = av1_filter_intra_predictor_neon; + av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_neon; + av1_highbd_inv_txfm_add_16x32 = av1_highbd_inv_txfm_add_16x32_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x32 = av1_highbd_inv_txfm_add_16x32_neon; + av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x4 = av1_highbd_inv_txfm_add_16x4_neon; + av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_neon; + av1_highbd_inv_txfm_add_32x16 = av1_highbd_inv_txfm_add_32x16_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x16 = av1_highbd_inv_txfm_add_32x16_neon; + av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_neon; + av1_highbd_inv_txfm_add_32x64 = av1_highbd_inv_txfm_add_32x64_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_32x64 = av1_highbd_inv_txfm_add_32x64_neon; + av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x16 = av1_highbd_inv_txfm_add_4x16_neon; + av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_neon; + av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_4x8 = av1_highbd_inv_txfm_add_4x8_neon; + av1_highbd_inv_txfm_add_64x32 = av1_highbd_inv_txfm_add_64x32_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_64x32 = av1_highbd_inv_txfm_add_64x32_neon; + av1_highbd_inv_txfm_add_64x64 = av1_highbd_inv_txfm_add_64x64_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_64x64 = av1_highbd_inv_txfm_add_64x64_neon; + av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_neon; + av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_neon; + av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; + if (flags & HAS_NEON) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_neon; + av1_inv_txfm2d_add_16x32 = av1_inv_txfm2d_add_16x32_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_16x32 = av1_inv_txfm2d_add_16x32_neon; + av1_inv_txfm2d_add_16x4 = av1_inv_txfm2d_add_16x4_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_16x4 = av1_inv_txfm2d_add_16x4_neon; + av1_inv_txfm2d_add_16x64 = av1_inv_txfm2d_add_16x64_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_16x64 = av1_inv_txfm2d_add_16x64_neon; + av1_inv_txfm2d_add_16x8 = av1_inv_txfm2d_add_16x8_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_16x8 = av1_inv_txfm2d_add_16x8_neon; + av1_inv_txfm2d_add_32x16 = av1_inv_txfm2d_add_32x16_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_32x16 = av1_inv_txfm2d_add_32x16_neon; + av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_neon; + av1_inv_txfm2d_add_32x64 = av1_inv_txfm2d_add_32x64_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_32x64 = av1_inv_txfm2d_add_32x64_neon; + av1_inv_txfm2d_add_32x8 = av1_inv_txfm2d_add_32x8_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_32x8 = av1_inv_txfm2d_add_32x8_neon; + av1_inv_txfm2d_add_4x16 = av1_inv_txfm2d_add_4x16_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_4x16 = av1_inv_txfm2d_add_4x16_neon; + av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_neon; + av1_inv_txfm2d_add_4x8 = av1_inv_txfm2d_add_4x8_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_4x8 = av1_inv_txfm2d_add_4x8_neon; + av1_inv_txfm2d_add_64x16 = av1_inv_txfm2d_add_64x16_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_64x16 = av1_inv_txfm2d_add_64x16_neon; + av1_inv_txfm2d_add_64x32 = av1_inv_txfm2d_add_64x32_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_64x32 = av1_inv_txfm2d_add_64x32_neon; + av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_neon; + av1_inv_txfm2d_add_8x16 = av1_inv_txfm2d_add_8x16_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_8x16 = av1_inv_txfm2d_add_8x16_neon; + av1_inv_txfm2d_add_8x32 = av1_inv_txfm2d_add_8x32_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_8x32 = av1_inv_txfm2d_add_8x32_neon; + av1_inv_txfm2d_add_8x4 = av1_inv_txfm2d_add_8x4_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_8x4 = av1_inv_txfm2d_add_8x4_neon; + av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c; + if (flags & HAS_NEON) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_neon; av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_NEON) av1_inv_txfm_add = av1_inv_txfm_add_neon; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_NEON) av1_resize_and_extend_frame = av1_resize_and_extend_frame_neon; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_NEON) av1_round_shift_array = av1_round_shift_array_neon; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -441,10 +599,26 @@ static void setup_rtcd_internal(void) if (flags & HAS_NEON) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_neon; cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c; if (flags & HAS_NEON) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_neon; - cdef_filter_block = cdef_filter_block_c; - if (flags & HAS_NEON) cdef_filter_block = cdef_filter_block_neon; + cdef_filter_16_0 = cdef_filter_16_0_c; + if (flags & HAS_NEON) cdef_filter_16_0 = cdef_filter_16_0_neon; + cdef_filter_16_1 = cdef_filter_16_1_c; + if (flags & HAS_NEON) cdef_filter_16_1 = cdef_filter_16_1_neon; + cdef_filter_16_2 = cdef_filter_16_2_c; + if (flags & HAS_NEON) cdef_filter_16_2 = cdef_filter_16_2_neon; + cdef_filter_16_3 = cdef_filter_16_3_c; + if (flags & HAS_NEON) cdef_filter_16_3 = cdef_filter_16_3_neon; + cdef_filter_8_0 = cdef_filter_8_0_c; + if (flags & HAS_NEON) cdef_filter_8_0 = cdef_filter_8_0_neon; + cdef_filter_8_1 = cdef_filter_8_1_c; + if (flags & HAS_NEON) cdef_filter_8_1 = cdef_filter_8_1_neon; + cdef_filter_8_2 = cdef_filter_8_2_c; + if (flags & HAS_NEON) cdef_filter_8_2 = cdef_filter_8_2_neon; + cdef_filter_8_3 = cdef_filter_8_3_c; + if (flags & HAS_NEON) cdef_filter_8_3 = cdef_filter_8_3_neon; cdef_find_dir = cdef_find_dir_c; if (flags & HAS_NEON) cdef_find_dir = cdef_find_dir_neon; + cdef_find_dir_dual = cdef_find_dir_dual_c; + if (flags & HAS_NEON) cdef_find_dir_dual = cdef_find_dir_dual_neon; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_NEON) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_neon; cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c; diff --git a/media/libaom/config/linux/ia32/config/aom_config.asm b/media/libaom/config/linux/ia32/config/aom_config.asm index 591f86e6c9..028a043478 100644 --- a/media/libaom/config/linux/ia32/config/aom_config.asm +++ b/media/libaom/config/linux/ia32/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 1 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/linux/ia32/config/aom_config.h b/media/libaom/config/linux/ia32/config/aom_config.h index d2e651f19b..7ce2646afb 100644 --- a/media/libaom/config/linux/ia32/config/aom_config.h +++ b/media/libaom/config/linux/ia32/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 1 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h index 9018213bd6..c8041dd7ca 100644 --- a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h +++ b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1740,6 +1723,7 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; aom_convolve_copy = aom_convolve_copy_c; if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c; if (flags & HAS_SSE2) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_sse2; aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c; @@ -2138,34 +2122,53 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2; aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_sse2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_sse2; aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_sse2; + aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_sse2; aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_sse2; aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_sse2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_sse2; aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_sse2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_lpf_vertical_14 = aom_lpf_vertical_14_c; if (flags & HAS_SSE2) aom_lpf_vertical_14 = aom_lpf_vertical_14_sse2; aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_sse2; + aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2; aom_lpf_vertical_4 = aom_lpf_vertical_4_c; if (flags & HAS_SSE2) aom_lpf_vertical_4 = aom_lpf_vertical_4_sse2; aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_sse2; + aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_sse2; aom_lpf_vertical_6 = aom_lpf_vertical_6_c; if (flags & HAS_SSE2) aom_lpf_vertical_6 = aom_lpf_vertical_6_sse2; aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_sse2; + aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_sse2; aom_lpf_vertical_8 = aom_lpf_vertical_8_c; if (flags & HAS_SSE2) aom_lpf_vertical_8 = aom_lpf_vertical_8_sse2; aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_sse2; + aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_sse2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -2214,6 +2217,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/linux/ia32/config/av1_rtcd.h b/media/libaom/config/linux/ia32/config/av1_rtcd.h index 7ecd05f380..14c816a8a9 100644 --- a/media/libaom/config/linux/ia32/config/av1_rtcd.h +++ b/media/libaom/config/linux/ia32/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -201,11 +181,6 @@ void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -225,35 +200,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -272,10 +247,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -288,6 +284,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -297,13 +308,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -375,22 +388,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -424,12 +441,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -438,6 +504,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -508,9 +581,6 @@ static void setup_rtcd_internal(void) av1_calc_frame_error = av1_calc_frame_error_c; if (flags & HAS_SSE2) av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_c; @@ -538,10 +608,13 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -549,9 +622,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -598,8 +668,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -610,6 +683,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -635,16 +710,56 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_c; - if (flags & HAS_SSE2) cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_c; + if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_c; + if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_c; + if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_c; + if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_c; + if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_c; + if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_c; + if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_c; + if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_c; if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_c; + if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/linux/x64/config/aom_config.asm b/media/libaom/config/linux/x64/config/aom_config.asm index 9ce9b2090e..36581168c1 100644 --- a/media/libaom/config/linux/x64/config/aom_config.asm +++ b/media/libaom/config/linux/x64/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/linux/x64/config/aom_config.h b/media/libaom/config/linux/x64/config/aom_config.h index 12997ea4dd..52a6e0c58c 100644 --- a/media/libaom/config/linux/x64/config/aom_config.h +++ b/media/libaom/config/linux/x64/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h index 36c170a05d..8c332069c3 100644 --- a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h +++ b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -#define aom_convolve_copy aom_convolve_copy_sse2 +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2 -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2 -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2 -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2 -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2 -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2 @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2 -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2 @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2 -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2 @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2 -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2 @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2 -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2 @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2 -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2 @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2 -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2 @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2 +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2 @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2 +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2 + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2 @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2 +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2 @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2 +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2 @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2 +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_sse2 + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2 @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2 +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2 + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2 @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2 +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2 + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2 @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2 +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2 + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2 -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1736,6 +1719,8 @@ static void setup_rtcd_internal(void) aom_convolve8_vert = aom_convolve8_vert_sse2; if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3; if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; + aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2; if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2; aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2; @@ -1816,6 +1801,12 @@ static void setup_rtcd_internal(void) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c; if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1; if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -1864,6 +1855,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/linux/x64/config/av1_rtcd.h b/media/libaom/config/linux/x64/config/av1_rtcd.h index 8f20181b42..a8a2afa03e 100644 --- a/media/libaom/config/linux/x64/config/av1_rtcd.h +++ b/media/libaom/config/linux/x64/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -204,11 +184,6 @@ void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8 void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2 -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -228,35 +203,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -275,10 +250,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -291,6 +287,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -300,13 +311,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -378,22 +391,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -427,12 +444,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -441,6 +507,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -510,8 +583,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2; av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_sse2; @@ -532,10 +603,13 @@ static void setup_rtcd_internal(void) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -543,8 +617,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -591,8 +663,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -603,6 +678,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -625,14 +702,46 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/mac/x64/config/aom_config.asm b/media/libaom/config/mac/x64/config/aom_config.asm index 9ce9b2090e..36581168c1 100644 --- a/media/libaom/config/mac/x64/config/aom_config.asm +++ b/media/libaom/config/mac/x64/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/mac/x64/config/aom_config.h b/media/libaom/config/mac/x64/config/aom_config.h index 12997ea4dd..52a6e0c58c 100644 --- a/media/libaom/config/mac/x64/config/aom_config.h +++ b/media/libaom/config/mac/x64/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h index 36c170a05d..8c332069c3 100644 --- a/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h +++ b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -#define aom_convolve_copy aom_convolve_copy_sse2 +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2 -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2 -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2 -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2 -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2 -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2 @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2 -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2 @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2 -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2 @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2 -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2 @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2 -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2 @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2 -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2 @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2 -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2 @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2 +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2 @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2 +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2 + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2 @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2 +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2 @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2 +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2 @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2 +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_sse2 + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2 @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2 +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2 + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2 @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2 +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2 + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2 @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2 +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2 + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2 -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1736,6 +1719,8 @@ static void setup_rtcd_internal(void) aom_convolve8_vert = aom_convolve8_vert_sse2; if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3; if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; + aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2; if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2; aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2; @@ -1816,6 +1801,12 @@ static void setup_rtcd_internal(void) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c; if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1; if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -1864,6 +1855,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/mac/x64/config/av1_rtcd.h b/media/libaom/config/mac/x64/config/av1_rtcd.h index 8f20181b42..a8a2afa03e 100644 --- a/media/libaom/config/mac/x64/config/av1_rtcd.h +++ b/media/libaom/config/mac/x64/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -204,11 +184,6 @@ void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8 void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2 -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -228,35 +203,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -275,10 +250,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -291,6 +287,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -300,13 +311,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -378,22 +391,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -427,12 +444,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -441,6 +507,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -510,8 +583,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2; av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_sse2; @@ -532,10 +603,13 @@ static void setup_rtcd_internal(void) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -543,8 +617,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -591,8 +663,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -603,6 +678,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -625,14 +702,46 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/win/ia32/config/aom_config.asm b/media/libaom/config/win/ia32/config/aom_config.asm index baf05cb910..e704f3f332 100644 --- a/media/libaom/config/win/ia32/config/aom_config.asm +++ b/media/libaom/config/win/ia32/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/win/ia32/config/aom_config.h b/media/libaom/config/win/ia32/config/aom_config.h index fc1271c743..8a73a20b11 100644 --- a/media/libaom/config/win/ia32/config/aom_config.h +++ b/media/libaom/config/win/ia32/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h index 9018213bd6..c8041dd7ca 100644 --- a/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h +++ b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1740,6 +1723,7 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; aom_convolve_copy = aom_convolve_copy_c; if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c; if (flags & HAS_SSE2) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_sse2; aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c; @@ -2138,34 +2122,53 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2; aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_sse2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_sse2; aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_sse2; + aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_sse2; aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_sse2; aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_sse2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_sse2; aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_sse2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_lpf_vertical_14 = aom_lpf_vertical_14_c; if (flags & HAS_SSE2) aom_lpf_vertical_14 = aom_lpf_vertical_14_sse2; aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_sse2; + aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2; aom_lpf_vertical_4 = aom_lpf_vertical_4_c; if (flags & HAS_SSE2) aom_lpf_vertical_4 = aom_lpf_vertical_4_sse2; aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_sse2; + aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_sse2; aom_lpf_vertical_6 = aom_lpf_vertical_6_c; if (flags & HAS_SSE2) aom_lpf_vertical_6 = aom_lpf_vertical_6_sse2; aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_sse2; + aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_sse2; aom_lpf_vertical_8 = aom_lpf_vertical_8_c; if (flags & HAS_SSE2) aom_lpf_vertical_8 = aom_lpf_vertical_8_sse2; aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_sse2; + aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_sse2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -2214,6 +2217,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/win/ia32/config/av1_rtcd.h b/media/libaom/config/win/ia32/config/av1_rtcd.h index 7ecd05f380..14c816a8a9 100644 --- a/media/libaom/config/win/ia32/config/av1_rtcd.h +++ b/media/libaom/config/win/ia32/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -201,11 +181,6 @@ void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -225,35 +200,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -272,10 +247,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -288,6 +284,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -297,13 +308,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -375,22 +388,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -424,12 +441,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -438,6 +504,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -508,9 +581,6 @@ static void setup_rtcd_internal(void) av1_calc_frame_error = av1_calc_frame_error_c; if (flags & HAS_SSE2) av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_c; @@ -538,10 +608,13 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -549,9 +622,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -598,8 +668,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -610,6 +683,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -635,16 +710,56 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_c; - if (flags & HAS_SSE2) cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_c; + if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_c; + if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_c; + if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_c; + if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_c; + if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_c; + if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_c; + if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_c; + if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_c; if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_c; + if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/win/mingw32/config/aom_config.asm b/media/libaom/config/win/mingw32/config/aom_config.asm index baf05cb910..e704f3f332 100644 --- a/media/libaom/config/win/mingw32/config/aom_config.asm +++ b/media/libaom/config/win/mingw32/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/win/mingw32/config/aom_config.h b/media/libaom/config/win/mingw32/config/aom_config.h index fc1271c743..8a73a20b11 100644 --- a/media/libaom/config/win/mingw32/config/aom_config.h +++ b/media/libaom/config/win/mingw32/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/win/mingw32/config/aom_dsp_rtcd.h b/media/libaom/config/win/mingw32/config/aom_dsp_rtcd.h index 9018213bd6..c8041dd7ca 100644 --- a/media/libaom/config/win/mingw32/config/aom_dsp_rtcd.h +++ b/media/libaom/config/win/mingw32/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_128_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_left_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_dc_top_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_128_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_left_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_dc_top_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_h_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); RTCD_EXTERN void (*aom_highbd_v_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_horizontal_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_horizontal_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_14)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_14_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_4)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_4_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_4_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_6)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_6_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); RTCD_EXTERN void (*aom_lpf_vertical_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); RTCD_EXTERN void (*aom_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_vertical_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1740,6 +1723,7 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; aom_convolve_copy = aom_convolve_copy_c; if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_c; if (flags & HAS_SSE2) aom_dc_128_predictor_16x16 = aom_dc_128_predictor_16x16_sse2; aom_dc_128_predictor_16x32 = aom_dc_128_predictor_16x32_c; @@ -2138,34 +2122,53 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2; aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_sse2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; aom_lpf_horizontal_4 = aom_lpf_horizontal_4_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4 = aom_lpf_horizontal_4_sse2; aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_4_dual = aom_lpf_horizontal_4_dual_sse2; + aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_4_quad = aom_lpf_horizontal_4_quad_sse2; aom_lpf_horizontal_6 = aom_lpf_horizontal_6_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6 = aom_lpf_horizontal_6_sse2; aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_6_dual = aom_lpf_horizontal_6_dual_sse2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; aom_lpf_horizontal_8 = aom_lpf_horizontal_8_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8 = aom_lpf_horizontal_8_sse2; aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_c; if (flags & HAS_SSE2) aom_lpf_horizontal_8_dual = aom_lpf_horizontal_8_dual_sse2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_lpf_vertical_14 = aom_lpf_vertical_14_c; if (flags & HAS_SSE2) aom_lpf_vertical_14 = aom_lpf_vertical_14_sse2; aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_14_dual = aom_lpf_vertical_14_dual_sse2; + aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_14_quad = aom_lpf_vertical_14_quad_sse2; aom_lpf_vertical_4 = aom_lpf_vertical_4_c; if (flags & HAS_SSE2) aom_lpf_vertical_4 = aom_lpf_vertical_4_sse2; aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_4_dual = aom_lpf_vertical_4_dual_sse2; + aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_4_quad = aom_lpf_vertical_4_quad_sse2; aom_lpf_vertical_6 = aom_lpf_vertical_6_c; if (flags & HAS_SSE2) aom_lpf_vertical_6 = aom_lpf_vertical_6_sse2; aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_6_dual = aom_lpf_vertical_6_dual_sse2; + aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_6_quad = aom_lpf_vertical_6_quad_sse2; aom_lpf_vertical_8 = aom_lpf_vertical_8_c; if (flags & HAS_SSE2) aom_lpf_vertical_8 = aom_lpf_vertical_8_sse2; aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_c; if (flags & HAS_SSE2) aom_lpf_vertical_8_dual = aom_lpf_vertical_8_dual_sse2; + aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_c; + if (flags & HAS_SSE2) aom_lpf_vertical_8_quad = aom_lpf_vertical_8_quad_sse2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -2214,6 +2217,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/win/mingw32/config/av1_rtcd.h b/media/libaom/config/win/mingw32/config/av1_rtcd.h index 7ecd05f380..14c816a8a9 100644 --- a/media/libaom/config/win/mingw32/config/av1_rtcd.h +++ b/media/libaom/config/win/mingw32/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -201,11 +181,6 @@ void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -225,35 +200,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -272,10 +247,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -288,6 +284,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -297,13 +308,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -375,22 +388,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -424,12 +441,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -438,6 +504,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -508,9 +581,6 @@ static void setup_rtcd_internal(void) av1_calc_frame_error = av1_calc_frame_error_c; if (flags & HAS_SSE2) av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_c; @@ -538,10 +608,13 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -549,9 +622,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_c; - if (flags & HAS_SSE2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -598,8 +668,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -610,6 +683,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -635,16 +710,56 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_c; - if (flags & HAS_SSE2) cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_c; + if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_c; + if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_c; + if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_c; + if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_c; + if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_c; + if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_c; + if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_c; + if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_c; if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_c; + if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/win/mingw64/config/aom_config.asm b/media/libaom/config/win/mingw64/config/aom_config.asm index 9ce9b2090e..36581168c1 100644 --- a/media/libaom/config/win/mingw64/config/aom_config.asm +++ b/media/libaom/config/win/mingw64/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/win/mingw64/config/aom_config.h b/media/libaom/config/win/mingw64/config/aom_config.h index 12997ea4dd..52a6e0c58c 100644 --- a/media/libaom/config/win/mingw64/config/aom_config.h +++ b/media/libaom/config/win/mingw64/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/win/mingw64/config/aom_dsp_rtcd.h b/media/libaom/config/win/mingw64/config/aom_dsp_rtcd.h index 36c170a05d..8c332069c3 100644 --- a/media/libaom/config/win/mingw64/config/aom_dsp_rtcd.h +++ b/media/libaom/config/win/mingw64/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -#define aom_convolve_copy aom_convolve_copy_sse2 +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2 -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2 -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2 -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2 -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2 -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2 @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2 -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2 @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2 -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2 @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2 -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2 @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2 -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2 @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2 -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2 @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2 -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2 @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2 +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2 @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2 +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2 + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2 @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2 +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2 @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2 +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2 @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2 +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_sse2 + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2 @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2 +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2 + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2 @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2 +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2 + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2 @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2 +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2 + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2 -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1736,6 +1719,8 @@ static void setup_rtcd_internal(void) aom_convolve8_vert = aom_convolve8_vert_sse2; if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3; if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; + aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2; if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2; aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2; @@ -1816,6 +1801,12 @@ static void setup_rtcd_internal(void) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c; if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1; if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -1864,6 +1855,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/win/mingw64/config/av1_rtcd.h b/media/libaom/config/win/mingw64/config/av1_rtcd.h index 8f20181b42..a8a2afa03e 100644 --- a/media/libaom/config/win/mingw64/config/av1_rtcd.h +++ b/media/libaom/config/win/mingw64/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -204,11 +184,6 @@ void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8 void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2 -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -228,35 +203,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -275,10 +250,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -291,6 +287,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -300,13 +311,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -378,22 +391,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -427,12 +444,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -441,6 +507,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -510,8 +583,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2; av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_sse2; @@ -532,10 +603,13 @@ static void setup_rtcd_internal(void) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -543,8 +617,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -591,8 +663,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -603,6 +678,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -625,14 +702,46 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/config/win/x64/config/aom_config.asm b/media/libaom/config/win/x64/config/aom_config.asm index 9ce9b2090e..36581168c1 100644 --- a/media/libaom/config/win/x64/config/aom_config.asm +++ b/media/libaom/config/win/x64/config/aom_config.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2021, Alliance for Open Media. All rights reserved +; Copyright (c) 2022, Alliance for Open Media. All rights reserved ; ; This source code is subject to the terms of the BSD 2 Clause License and ; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -19,7 +19,10 @@ CONFIG_ANALYZER equ 0 CONFIG_AV1_DECODER equ 1 CONFIG_AV1_ENCODER equ 0 CONFIG_AV1_HIGHBITDEPTH equ 1 +CONFIG_AV1_TEMPORAL_DENOISING equ 0 CONFIG_BIG_ENDIAN equ 0 +CONFIG_BITRATE_ACCURACY equ 0 +CONFIG_BITRATE_ACCURACY_BL equ 0 CONFIG_BITSTREAM_DEBUG equ 0 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0 CONFIG_COLLECT_COMPONENT_TIMING equ 0 @@ -28,33 +31,40 @@ CONFIG_COLLECT_RD_STATS equ 0 CONFIG_DEBUG equ 0 CONFIG_DENOISE equ 1 CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 equ 1 -CONFIG_DIST_8X8 equ 0 CONFIG_ENTROPY_STATS equ 0 +CONFIG_EXCLUDE_SIMD_MISMATCH equ 0 +CONFIG_FPMT_TEST equ 0 +CONFIG_FRAME_PARALLEL_ENCODE equ 0 +CONFIG_FRAME_PARALLEL_ENCODE_2 equ 0 CONFIG_GCC equ 1 CONFIG_GCOV equ 0 CONFIG_GPROF equ 0 -CONFIG_HTB_TRELLIS equ 0 CONFIG_INSPECTION equ 0 CONFIG_INTERNAL_STATS equ 0 CONFIG_INTER_STATS_ONLY equ 0 CONFIG_LIBYUV equ 0 -CONFIG_LPF_MASK equ 0 CONFIG_MAX_DECODE_PROFILE equ 2 CONFIG_MISMATCH_DEBUG equ 0 CONFIG_MULTITHREAD equ 1 CONFIG_NN_V2 equ 0 CONFIG_NORMAL_TILE_MODE equ 0 +CONFIG_OPTICAL_FLOW_API equ 0 CONFIG_OS_SUPPORT equ 1 +CONFIG_PARTITION_SEARCH_ORDER equ 0 CONFIG_PIC equ 0 +CONFIG_RATECTRL_LOG equ 0 +CONFIG_RD_COMMAND equ 0 CONFIG_RD_DEBUG equ 0 CONFIG_REALTIME_ONLY equ 0 +CONFIG_RT_ML_PARTITIONING equ 0 CONFIG_RUNTIME_CPU_DETECT equ 1 CONFIG_SHARED equ 0 -CONFIG_SHARP_SETTINGS equ 0 CONFIG_SIZE_LIMIT equ 0 CONFIG_SPATIAL_RESAMPLING equ 1 CONFIG_SPEED_STATS equ 0 -CONFIG_SUPERRES_IN_RECODE equ 1 +CONFIG_TFLITE equ 0 +CONFIG_THREE_PASS equ 0 +CONFIG_TUNE_BUTTERAUGLI equ 0 CONFIG_TUNE_VMAF equ 0 CONFIG_WEBM_IO equ 0 DECODE_HEIGHT_LIMIT equ 0 @@ -77,3 +87,4 @@ HAVE_SSE4_2 equ 1 HAVE_SSSE3 equ 1 HAVE_VSX equ 0 HAVE_WXWIDGETS equ 0 +STATIC_LINK_JXL equ 0 diff --git a/media/libaom/config/win/x64/config/aom_config.h b/media/libaom/config/win/x64/config/aom_config.h index 12997ea4dd..52a6e0c58c 100644 --- a/media/libaom/config/win/x64/config/aom_config.h +++ b/media/libaom/config/win/x64/config/aom_config.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, Alliance for Open Media. All rights reserved + * Copyright (c) 2022, Alliance for Open Media. All rights reserved * * This source code is subject to the terms of the BSD 2 Clause License and * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License @@ -21,7 +21,10 @@ #define CONFIG_AV1_DECODER 1 #define CONFIG_AV1_ENCODER 0 #define CONFIG_AV1_HIGHBITDEPTH 1 +#define CONFIG_AV1_TEMPORAL_DENOISING 0 #define CONFIG_BIG_ENDIAN 0 +#define CONFIG_BITRATE_ACCURACY 0 +#define CONFIG_BITRATE_ACCURACY_BL 0 #define CONFIG_BITSTREAM_DEBUG 0 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0 #define CONFIG_COLLECT_COMPONENT_TIMING 0 @@ -30,33 +33,40 @@ #define CONFIG_DEBUG 0 #define CONFIG_DENOISE 1 #define CONFIG_DISABLE_FULL_PIXEL_SPLIT_8X8 1 -#define CONFIG_DIST_8X8 0 #define CONFIG_ENTROPY_STATS 0 +#define CONFIG_EXCLUDE_SIMD_MISMATCH 0 +#define CONFIG_FPMT_TEST 0 +#define CONFIG_FRAME_PARALLEL_ENCODE 0 +#define CONFIG_FRAME_PARALLEL_ENCODE_2 0 #define CONFIG_GCC 1 #define CONFIG_GCOV 0 #define CONFIG_GPROF 0 -#define CONFIG_HTB_TRELLIS 0 #define CONFIG_INSPECTION 0 #define CONFIG_INTERNAL_STATS 0 #define CONFIG_INTER_STATS_ONLY 0 #define CONFIG_LIBYUV 0 -#define CONFIG_LPF_MASK 0 #define CONFIG_MAX_DECODE_PROFILE 2 #define CONFIG_MISMATCH_DEBUG 0 #define CONFIG_MULTITHREAD 1 #define CONFIG_NN_V2 0 #define CONFIG_NORMAL_TILE_MODE 0 +#define CONFIG_OPTICAL_FLOW_API 0 #define CONFIG_OS_SUPPORT 1 +#define CONFIG_PARTITION_SEARCH_ORDER 0 #define CONFIG_PIC 0 +#define CONFIG_RATECTRL_LOG 0 +#define CONFIG_RD_COMMAND 0 #define CONFIG_RD_DEBUG 0 #define CONFIG_REALTIME_ONLY 0 +#define CONFIG_RT_ML_PARTITIONING 0 #define CONFIG_RUNTIME_CPU_DETECT 1 #define CONFIG_SHARED 0 -#define CONFIG_SHARP_SETTINGS 0 #define CONFIG_SIZE_LIMIT 0 #define CONFIG_SPATIAL_RESAMPLING 1 #define CONFIG_SPEED_STATS 0 -#define CONFIG_SUPERRES_IN_RECODE 1 +#define CONFIG_TFLITE 0 +#define CONFIG_THREE_PASS 0 +#define CONFIG_TUNE_BUTTERAUGLI 0 #define CONFIG_TUNE_VMAF 0 #define CONFIG_WEBM_IO 0 #define DECODE_HEIGHT_LIMIT 0 @@ -80,4 +90,5 @@ #define HAVE_VSX 0 #define HAVE_WXWIDGETS 0 #define INLINE inline +#define STATIC_LINK_JXL 0 #endif /* AOM_CONFIG_H_ */ diff --git a/media/libaom/config/win/x64/config/aom_dsp_rtcd.h b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h index 36c170a05d..8c332069c3 100644 --- a/media/libaom/config/win/x64/config/aom_dsp_rtcd.h +++ b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h @@ -35,6 +35,9 @@ void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h); +void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +#define aom_convolve8 aom_convolve8_c + void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); @@ -47,9 +50,10 @@ void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h); -void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h); -#define aom_convolve_copy aom_convolve_copy_sse2 +void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -71,9 +75,6 @@ void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2 -void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c - void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -156,9 +157,6 @@ void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_ void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2 -void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c - void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -241,9 +239,6 @@ void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2 -void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c - void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -326,9 +321,6 @@ void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2 -void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c - void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -411,9 +403,6 @@ void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2 -void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_h_predictor_2x2 aom_h_predictor_2x2_c - void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2 @@ -498,10 +487,10 @@ void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, ui void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bd); -void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); -RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, const int16_t *filter_y, int filter_y_stride, int w, int h, int bd); +void aom_highbd_convolve_copy_c(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); +RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, int w, int h); void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); @@ -521,9 +510,6 @@ void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2 -void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c - void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2 @@ -591,9 +577,6 @@ void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2 -void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c - void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2 @@ -661,9 +644,6 @@ void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uin void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2 -void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c - void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2 @@ -731,9 +711,6 @@ void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2 -void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c - void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2 @@ -801,9 +778,6 @@ void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2 -void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c - void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2 @@ -938,9 +912,6 @@ void aom_highbd_paeth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, const void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c -void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c - void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c @@ -998,9 +969,6 @@ void aom_highbd_smooth_h_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c -void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c - void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c @@ -1058,9 +1026,6 @@ void aom_highbd_smooth_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, cons void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c -void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c - void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c @@ -1118,9 +1083,6 @@ void aom_highbd_smooth_v_predictor_16x64_c(uint16_t *dst, ptrdiff_t y_stride, co void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c -void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c - void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c @@ -1181,9 +1143,6 @@ void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2 -void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); -#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c - void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd); #define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2 @@ -1246,6 +1205,11 @@ void aom_lpf_horizontal_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_14_dual aom_lpf_horizontal_14_dual_sse2 +void aom_lpf_horizontal_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_14_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_14_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2 @@ -1254,6 +1218,10 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2 +void aom_lpf_horizontal_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_horizontal_4_quad aom_lpf_horizontal_4_quad_sse2 + void aom_lpf_horizontal_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_6 aom_lpf_horizontal_6_sse2 @@ -1262,6 +1230,11 @@ void aom_lpf_horizontal_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_6_dual aom_lpf_horizontal_6_dual_sse2 +void aom_lpf_horizontal_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_6_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_6_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2 @@ -1270,6 +1243,11 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2 +void aom_lpf_horizontal_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_horizontal_8_quad_avx2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +RTCD_EXTERN void (*aom_lpf_horizontal_8_quad)(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); + void aom_lpf_vertical_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_14_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_14 aom_lpf_vertical_14_sse2 @@ -1278,6 +1256,10 @@ void aom_lpf_vertical_14_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, c void aom_lpf_vertical_14_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_14_dual aom_lpf_vertical_14_dual_sse2 +void aom_lpf_vertical_14_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_14_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_14_quad aom_lpf_vertical_14_quad_sse2 + void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2 @@ -1286,6 +1268,10 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2 +void aom_lpf_vertical_4_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_4_quad aom_lpf_vertical_4_quad_sse2 + void aom_lpf_vertical_6_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_6_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_6 aom_lpf_vertical_6_sse2 @@ -1294,6 +1280,10 @@ void aom_lpf_vertical_6_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_6_dual aom_lpf_vertical_6_dual_sse2 +void aom_lpf_vertical_6_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_6_quad aom_lpf_vertical_6_quad_sse2 + void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh); #define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2 @@ -1302,6 +1292,10 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1); #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2 +void aom_lpf_vertical_8_quad_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0); +#define aom_lpf_vertical_8_quad aom_lpf_vertical_8_quad_sse2 + void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1326,9 +1320,6 @@ void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c - void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1391,6 +1382,10 @@ void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t * void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); +void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); +RTCD_EXTERN void (*aom_scaled_2d)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h); + void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1411,9 +1406,6 @@ void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_h_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c - void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_h_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_h_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1490,9 +1482,6 @@ void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c - void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1569,9 +1558,6 @@ void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8 void aom_smooth_v_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c - void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_smooth_v_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); RTCD_EXTERN void (*aom_smooth_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1648,9 +1634,6 @@ void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); #define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2 -void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); -#define aom_v_predictor_2x2 aom_v_predictor_2x2_c - void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left); @@ -1736,6 +1719,8 @@ static void setup_rtcd_internal(void) aom_convolve8_vert = aom_convolve8_vert_sse2; if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3; if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2; + aom_convolve_copy = aom_convolve_copy_sse2; + if (flags & HAS_AVX2) aom_convolve_copy = aom_convolve_copy_avx2; aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2; if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2; aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2; @@ -1816,6 +1801,12 @@ static void setup_rtcd_internal(void) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c; if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1; if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2; + aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_14_quad = aom_lpf_horizontal_14_quad_avx2; + aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_6_quad = aom_lpf_horizontal_6_quad_avx2; + aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_sse2; + if (flags & HAS_AVX2) aom_lpf_horizontal_8_quad = aom_lpf_horizontal_8_quad_avx2; aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c; if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3; if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2; @@ -1864,6 +1855,8 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3; aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c; if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3; + aom_scaled_2d = aom_scaled_2d_c; + if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3; aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_c; if (flags & HAS_SSSE3) aom_smooth_h_predictor_16x16 = aom_smooth_h_predictor_16x16_ssse3; aom_smooth_h_predictor_16x32 = aom_smooth_h_predictor_16x32_c; diff --git a/media/libaom/config/win/x64/config/av1_rtcd.h b/media/libaom/config/win/x64/config/av1_rtcd.h index 8f20181b42..a8a2afa03e 100644 --- a/media/libaom/config/win/x64/config/av1_rtcd.h +++ b/media/libaom/config/win/x64/config/av1_rtcd.h @@ -13,6 +13,7 @@ */ #include "aom/aom_integer.h" +#include "aom_dsp/odintrin.h" #include "aom_dsp/txfm_common.h" #include "av1/common/common.h" #include "av1/common/enums.h" @@ -20,7 +21,6 @@ #include "av1/common/filter.h" #include "av1/common/convolve.h" #include "av1/common/av1_txfm.h" -#include "av1/common/odintrin.h" #include "av1/common/restoration.h" struct macroblockd; @@ -101,29 +101,6 @@ int64_t av1_calc_frame_error_sse2(const uint8_t *const ref, int stride, const ui int64_t av1_calc_frame_error_avx2(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); RTCD_EXTERN int64_t (*av1_calc_frame_error)(const uint8_t *const ref, int stride, const uint8_t *const dst, int p_width, int p_height, int p_stride); -void av1_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation); -#define av1_cnn_activate av1_cnn_activate_c - -void av1_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add); -#define av1_cnn_add av1_cnn_add_c - -void av1_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std); -#define av1_cnn_batchnorm av1_cnn_batchnorm_c - -void av1_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step); -#define av1_cnn_convolve av1_cnn_convolve_c - -void av1_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride); -#define av1_cnn_deconvolve av1_cnn_deconvolve_c - -void av1_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct); -#define av1_cnn_predict av1_cnn_predict_c - -void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); - void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params); @@ -137,15 +114,15 @@ void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, i void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn); -void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); +RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn); void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); @@ -153,30 +130,33 @@ void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); -RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +void av1_dist_wtd_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); +RTCD_EXTERN void (*av1_dist_wtd_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params); void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); +void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z1)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy); void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); +void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z2)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); +void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); RTCD_EXTERN void (*av1_dr_prediction_z3)(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy); @@ -204,11 +184,6 @@ void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8 void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps); #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2 -void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); - void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd); @@ -228,35 +203,35 @@ void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_ void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd); -void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); +RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, int bd); void av1_highbd_dist_wtd_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dist_wtd_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); -RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +void av1_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); +RTCD_EXTERN void (*av1_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd); void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd); @@ -275,10 +250,31 @@ void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest, int void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x32 av1_highbd_inv_txfm_add_16x32_c + void av1_highbd_inv_txfm_add_16x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_16x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_16x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x64 av1_highbd_inv_txfm_add_16x64_c + +void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c + +void av1_highbd_inv_txfm_add_32x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x16 av1_highbd_inv_txfm_add_32x16_c + +void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c + +void av1_highbd_inv_txfm_add_32x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x64 av1_highbd_inv_txfm_add_32x64_c + +void av1_highbd_inv_txfm_add_32x8_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_32x8 av1_highbd_inv_txfm_add_32x8_c + void av1_highbd_inv_txfm_add_4x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_4x16_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x16)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -291,6 +287,21 @@ void av1_highbd_inv_txfm_add_4x8_c(const tran_low_t *input, uint8_t *dest, int s void av1_highbd_inv_txfm_add_4x8_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +void av1_highbd_inv_txfm_add_64x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x16 av1_highbd_inv_txfm_add_64x16_c + +void av1_highbd_inv_txfm_add_64x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x32 av1_highbd_inv_txfm_add_64x32_c + +void av1_highbd_inv_txfm_add_64x64_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_64x64 av1_highbd_inv_txfm_add_64x64_c + +void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c + +void av1_highbd_inv_txfm_add_8x32_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); +#define av1_highbd_inv_txfm_add_8x32 av1_highbd_inv_txfm_add_8x32_c + void av1_highbd_inv_txfm_add_8x4_c(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_inv_txfm_add_8x4_sse4_1(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x4)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); @@ -300,13 +311,15 @@ void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest, RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *input, uint8_t *dest, int stride, const TxfmParam *txfm_param); void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); -#define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c +void av1_highbd_iwht4x4_16_add_sse4_1(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); +RTCD_EXTERN void (*av1_highbd_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd); #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); +void av1_highbd_warp_affine_avx2(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta); void av1_highbd_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bd); @@ -378,22 +391,26 @@ void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param); +void av1_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +void av1_resize_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); +RTCD_EXTERN void (*av1_resize_and_extend_frame)(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes); + void av1_round_shift_array_c(int32_t *arr, int size, int bit); void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit); RTCD_EXTERN void (*av1_round_shift_array)(int32_t *arr, int size, int bit); int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height, - int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, - int sgr_params_idx, int bit_depth, int highbd); + int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride, + int sgr_params_idx, int bit_depth, int highbd); void av1_upsample_intra_edge_c(uint8_t *p, int sz); void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz); @@ -427,12 +444,61 @@ void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h); -void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); -RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int coeff_shift); +void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); + +void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); +RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height); int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift); int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); @@ -441,6 +507,13 @@ int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coef int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift); RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift); +void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); +RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2); + cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size); cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size); @@ -510,8 +583,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2; av1_calc_frame_error = av1_calc_frame_error_sse2; if (flags & HAS_AVX2) av1_calc_frame_error = av1_calc_frame_error_avx2; - av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2; av1_convolve_2d_scale = av1_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1; av1_convolve_2d_sr = av1_convolve_2d_sr_sse2; @@ -532,10 +603,13 @@ static void setup_rtcd_internal(void) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_sse2; if (flags & HAS_AVX2) av1_dist_wtd_convolve_y = av1_dist_wtd_convolve_y_avx2; av1_dr_prediction_z1 = av1_dr_prediction_z1_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z1 = av1_dr_prediction_z1_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z1 = av1_dr_prediction_z1_avx2; av1_dr_prediction_z2 = av1_dr_prediction_z2_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z2 = av1_dr_prediction_z2_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z2 = av1_dr_prediction_z2_avx2; av1_dr_prediction_z3 = av1_dr_prediction_z3_c; + if (flags & HAS_SSE4_1) av1_dr_prediction_z3 = av1_dr_prediction_z3_sse4_1; if (flags & HAS_AVX2) av1_dr_prediction_z3 = av1_dr_prediction_z3_avx2; av1_filter_intra_edge = av1_filter_intra_edge_c; if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1; @@ -543,8 +617,6 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1; av1_filter_intra_predictor = av1_filter_intra_predictor_c; if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1; - av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2; - if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2; av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c; if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1; av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c; @@ -591,8 +663,11 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x4 = av1_highbd_inv_txfm_add_8x4_sse4_1; av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c; if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1; + av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_c; + if (flags & HAS_SSE4_1) av1_highbd_iwht4x4_16_add = av1_highbd_iwht4x4_16_add_sse4_1; av1_highbd_warp_affine = av1_highbd_warp_affine_c; if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1; + if (flags & HAS_AVX2) av1_highbd_warp_affine = av1_highbd_warp_affine_avx2; av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c; if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3; if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2; @@ -603,6 +678,8 @@ static void setup_rtcd_internal(void) av1_inv_txfm_add = av1_inv_txfm_add_c; if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3; if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2; + av1_resize_and_extend_frame = av1_resize_and_extend_frame_c; + if (flags & HAS_SSSE3) av1_resize_and_extend_frame = av1_resize_and_extend_frame_ssse3; av1_round_shift_array = av1_round_shift_array_c; if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1; av1_selfguided_restoration = av1_selfguided_restoration_c; @@ -625,14 +702,46 @@ static void setup_rtcd_internal(void) if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3; if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1; if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2; - cdef_filter_block = cdef_filter_block_sse2; - if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3; - if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1; - if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2; + cdef_filter_16_0 = cdef_filter_16_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2; + cdef_filter_16_1 = cdef_filter_16_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2; + cdef_filter_16_2 = cdef_filter_16_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2; + cdef_filter_16_3 = cdef_filter_16_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2; + cdef_filter_8_0 = cdef_filter_8_0_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2; + cdef_filter_8_1 = cdef_filter_8_1_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2; + cdef_filter_8_2 = cdef_filter_8_2_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2; + cdef_filter_8_3 = cdef_filter_8_3_sse2; + if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3; + if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1; + if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2; cdef_find_dir = cdef_find_dir_sse2; if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3; if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1; if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2; + cdef_find_dir_dual = cdef_find_dir_dual_sse2; + if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3; + if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1; + if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2; cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c; if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3; if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2; diff --git a/media/libaom/generate_sources_mozbuild.py b/media/libaom/generate_sources_mozbuild.py index 2f89c99488..b8ed07b1f7 100644 --- a/media/libaom/generate_sources_mozbuild.py +++ b/media/libaom/generate_sources_mozbuild.py @@ -9,7 +9,7 @@ import os import re import subprocess -AOM_DIR = '/media/libaom/src' +AOM_DIR = './src' def write_aom_config(system, arch, variables, cache_variables): # read template cmake file diff --git a/media/libaom/generate_sources_mozbuild.sh b/media/libaom/generate_sources_mozbuild.sh index d712d145d7..85a051dcdb 100755 --- a/media/libaom/generate_sources_mozbuild.sh +++ b/media/libaom/generate_sources_mozbuild.sh @@ -15,7 +15,7 @@ export LC_ALL=C BASE_DIR=$(pwd) -LIBAOM_SRC_DIR="/media/libaom/src" +LIBAOM_SRC_DIR="src" LIBAOM_CONFIG_DIR="config" # Print license header. diff --git a/media/libaom/sources.mozbuild b/media/libaom/sources.mozbuild index e6ddf7b79d..7490a3888d 100644 --- a/media/libaom/sources.mozbuild +++ b/media/libaom/sources.mozbuild @@ -2,543 +2,483 @@ files = { 'ARM_EXPORTS': [ - '/media/libaom/src/aom/aom.h', - '/media/libaom/src/aom/aom_codec.h', - '/media/libaom/src/aom/aom_decoder.h', - '/media/libaom/src/aom/aom_encoder.h', - '/media/libaom/src/aom/aom_frame_buffer.h', - '/media/libaom/src/aom/aom_image.h', - '/media/libaom/src/aom/aom_integer.h', - '/media/libaom/src/aom/aomcx.h', - '/media/libaom/src/aom/aomdx.h', - '/media/libaom/src/aom_mem/aom_mem.h', - '/media/libaom/src/aom_ports/aom_timer.h', - '/media/libaom/src/aom_ports/arm.h', - '/media/libaom/src/aom_ports/bitops.h', - '/media/libaom/src/aom_ports/mem.h', - '/media/libaom/src/aom_ports/sanitizer.h', - '/media/libaom/src/aom_ports/system_state.h', - '/media/libaom/src/aom_scale/aom_scale.h', - '/media/libaom/src/aom_scale/yv12config.h', ], 'ARM_SOURCES': [ - '/media/libaom/src/aom/src/aom_codec.c', - '/media/libaom/src/aom/src/aom_decoder.c', - '/media/libaom/src/aom/src/aom_encoder.c', - '/media/libaom/src/aom/src/aom_image.c', - '/media/libaom/src/aom/src/aom_integer.c', - '/media/libaom/src/aom_dsp/aom_convolve.c', - '/media/libaom/src/aom_dsp/aom_dsp_rtcd.c', - '/media/libaom/src/aom_dsp/arm/blend_a64_mask_neon.c', - '/media/libaom/src/aom_dsp/arm/fwd_txfm_neon.c', - '/media/libaom/src/aom_dsp/arm/intrapred_neon.c', - '/media/libaom/src/aom_dsp/arm/loopfilter_neon.c', - '/media/libaom/src/aom_dsp/arm/subtract_neon.c', - '/media/libaom/src/aom_dsp/avg.c', - '/media/libaom/src/aom_dsp/binary_codes_reader.c', - '/media/libaom/src/aom_dsp/bitreader.c', - '/media/libaom/src/aom_dsp/bitreader_buffer.c', - '/media/libaom/src/aom_dsp/bitwriter_buffer.c', - '/media/libaom/src/aom_dsp/blend_a64_hmask.c', - '/media/libaom/src/aom_dsp/blend_a64_mask.c', - '/media/libaom/src/aom_dsp/blend_a64_vmask.c', - '/media/libaom/src/aom_dsp/entcode.c', - '/media/libaom/src/aom_dsp/entdec.c', - '/media/libaom/src/aom_dsp/fft.c', - '/media/libaom/src/aom_dsp/grain_synthesis.c', - '/media/libaom/src/aom_dsp/intrapred.c', - '/media/libaom/src/aom_dsp/loopfilter.c', - '/media/libaom/src/aom_dsp/subtract.c', - '/media/libaom/src/aom_mem/aom_mem.c', - '/media/libaom/src/aom_ports/arm_cpudetect.c', - '/media/libaom/src/aom_scale/aom_scale_rtcd.c', - '/media/libaom/src/aom_scale/generic/aom_scale.c', - '/media/libaom/src/aom_scale/generic/gen_scalers.c', - '/media/libaom/src/aom_scale/generic/yv12config.c', - '/media/libaom/src/aom_scale/generic/yv12extend.c', - '/media/libaom/src/aom_util/aom_thread.c', - '/media/libaom/src/aom_util/debug_util.c', - '/media/libaom/src/av1/av1_dx_iface.c', - '/media/libaom/src/av1/common/alloccommon.c', - '/media/libaom/src/av1/common/arm/av1_inv_txfm_neon.c', - '/media/libaom/src/av1/common/arm/av1_txfm_neon.c', - '/media/libaom/src/av1/common/arm/blend_a64_hmask_neon.c', - '/media/libaom/src/av1/common/arm/blend_a64_vmask_neon.c', - '/media/libaom/src/av1/common/arm/cfl_neon.c', - '/media/libaom/src/av1/common/arm/convolve_neon.c', - '/media/libaom/src/av1/common/arm/jnt_convolve_neon.c', - '/media/libaom/src/av1/common/arm/reconinter_neon.c', - '/media/libaom/src/av1/common/arm/selfguided_neon.c', - '/media/libaom/src/av1/common/arm/warp_plane_neon.c', - '/media/libaom/src/av1/common/arm/wiener_convolve_neon.c', - '/media/libaom/src/av1/common/av1_inv_txfm1d.c', - '/media/libaom/src/av1/common/av1_inv_txfm2d.c', - '/media/libaom/src/av1/common/av1_loopfilter.c', - '/media/libaom/src/av1/common/av1_rtcd.c', - '/media/libaom/src/av1/common/av1_txfm.c', - '/media/libaom/src/av1/common/blockd.c', - '/media/libaom/src/av1/common/cdef.c', - '/media/libaom/src/av1/common/cdef_block.c', - '/media/libaom/src/av1/common/cdef_block_neon.c', - '/media/libaom/src/av1/common/cfl.c', - '/media/libaom/src/av1/common/convolve.c', - '/media/libaom/src/av1/common/debugmodes.c', - '/media/libaom/src/av1/common/entropy.c', - '/media/libaom/src/av1/common/entropymode.c', - '/media/libaom/src/av1/common/entropymv.c', - '/media/libaom/src/av1/common/frame_buffers.c', - '/media/libaom/src/av1/common/idct.c', - '/media/libaom/src/av1/common/mvref_common.c', - '/media/libaom/src/av1/common/obu_util.c', - '/media/libaom/src/av1/common/odintrin.c', - '/media/libaom/src/av1/common/pred_common.c', - '/media/libaom/src/av1/common/quant_common.c', - '/media/libaom/src/av1/common/reconinter.c', - '/media/libaom/src/av1/common/reconintra.c', - '/media/libaom/src/av1/common/resize.c', - '/media/libaom/src/av1/common/restoration.c', - '/media/libaom/src/av1/common/scale.c', - '/media/libaom/src/av1/common/scan.c', - '/media/libaom/src/av1/common/seg_common.c', - '/media/libaom/src/av1/common/thread_common.c', - '/media/libaom/src/av1/common/tile_common.c', - '/media/libaom/src/av1/common/timing.c', - '/media/libaom/src/av1/common/txb_common.c', - '/media/libaom/src/av1/common/warped_motion.c', - '/media/libaom/src/av1/decoder/decodeframe.c', - '/media/libaom/src/av1/decoder/decodemv.c', - '/media/libaom/src/av1/decoder/decoder.c', - '/media/libaom/src/av1/decoder/decodetxb.c', - '/media/libaom/src/av1/decoder/detokenize.c', - '/media/libaom/src/av1/decoder/obu.c', - '/media/libaom/src/av1/encoder/arm/neon/av1_error_neon.c', - '/media/libaom/src/av1/encoder/arm/neon/quantize_neon.c', + './src/aom/src/aom_codec.c', + './src/aom/src/aom_decoder.c', + './src/aom/src/aom_encoder.c', + './src/aom/src/aom_image.c', + './src/aom/src/aom_integer.c', + './src/aom_dsp/aom_convolve.c', + './src/aom_dsp/aom_dsp_rtcd.c', + './src/aom_dsp/arm/aom_convolve_copy_neon.c', + './src/aom_dsp/arm/blend_a64_mask_neon.c', + './src/aom_dsp/arm/fwd_txfm_neon.c', + './src/aom_dsp/arm/highbd_intrapred_neon.c', + './src/aom_dsp/arm/highbd_loopfilter_neon.c', + './src/aom_dsp/arm/intrapred_neon.c', + './src/aom_dsp/arm/loopfilter_neon.c', + './src/aom_dsp/arm/subtract_neon.c', + './src/aom_dsp/binary_codes_reader.c', + './src/aom_dsp/bitreader.c', + './src/aom_dsp/bitreader_buffer.c', + './src/aom_dsp/bitwriter_buffer.c', + './src/aom_dsp/blend_a64_hmask.c', + './src/aom_dsp/blend_a64_mask.c', + './src/aom_dsp/blend_a64_vmask.c', + './src/aom_dsp/entcode.c', + './src/aom_dsp/entdec.c', + './src/aom_dsp/fft.c', + './src/aom_dsp/intrapred.c', + './src/aom_dsp/loopfilter.c', + './src/aom_dsp/odintrin.c', + './src/aom_dsp/subtract.c', + './src/aom_mem/aom_mem.c', + './src/aom_ports/arm_cpudetect.c', + './src/aom_scale/aom_scale_rtcd.c', + './src/aom_scale/generic/aom_scale.c', + './src/aom_scale/generic/gen_scalers.c', + './src/aom_scale/generic/yv12config.c', + './src/aom_scale/generic/yv12extend.c', + './src/aom_util/aom_thread.c', + './src/aom_util/debug_util.c', + './src/av1/arg_defs.c', + './src/av1/av1_dx_iface.c', + './src/av1/common/alloccommon.c', + './src/av1/common/arm/av1_inv_txfm_neon.c', + './src/av1/common/arm/av1_txfm_neon.c', + './src/av1/common/arm/blend_a64_hmask_neon.c', + './src/av1/common/arm/blend_a64_vmask_neon.c', + './src/av1/common/arm/cdef_block_neon.c', + './src/av1/common/arm/cfl_neon.c', + './src/av1/common/arm/convolve_neon.c', + './src/av1/common/arm/highbd_inv_txfm_neon.c', + './src/av1/common/arm/jnt_convolve_neon.c', + './src/av1/common/arm/reconinter_neon.c', + './src/av1/common/arm/reconintra_neon.c', + './src/av1/common/arm/resize_neon.c', + './src/av1/common/arm/selfguided_neon.c', + './src/av1/common/arm/warp_plane_neon.c', + './src/av1/common/arm/wiener_convolve_neon.c', + './src/av1/common/av1_inv_txfm1d.c', + './src/av1/common/av1_inv_txfm2d.c', + './src/av1/common/av1_loopfilter.c', + './src/av1/common/av1_rtcd.c', + './src/av1/common/av1_txfm.c', + './src/av1/common/blockd.c', + './src/av1/common/cdef.c', + './src/av1/common/cdef_block.c', + './src/av1/common/cfl.c', + './src/av1/common/convolve.c', + './src/av1/common/debugmodes.c', + './src/av1/common/entropy.c', + './src/av1/common/entropymode.c', + './src/av1/common/entropymv.c', + './src/av1/common/frame_buffers.c', + './src/av1/common/idct.c', + './src/av1/common/mvref_common.c', + './src/av1/common/obu_util.c', + './src/av1/common/pred_common.c', + './src/av1/common/quant_common.c', + './src/av1/common/reconinter.c', + './src/av1/common/reconintra.c', + './src/av1/common/resize.c', + './src/av1/common/restoration.c', + './src/av1/common/scale.c', + './src/av1/common/scan.c', + './src/av1/common/seg_common.c', + './src/av1/common/thread_common.c', + './src/av1/common/tile_common.c', + './src/av1/common/timing.c', + './src/av1/common/txb_common.c', + './src/av1/common/warped_motion.c', + './src/av1/decoder/decodeframe.c', + './src/av1/decoder/decodemv.c', + './src/av1/decoder/decoder.c', + './src/av1/decoder/decodetxb.c', + './src/av1/decoder/detokenize.c', + './src/av1/decoder/grain_synthesis.c', + './src/av1/decoder/obu.c', + './src/common/args_helper.c', ], 'GENERIC_EXPORTS': [ - '/media/libaom/src/aom/aom.h', - '/media/libaom/src/aom/aom_codec.h', - '/media/libaom/src/aom/aom_decoder.h', - '/media/libaom/src/aom/aom_encoder.h', - '/media/libaom/src/aom/aom_frame_buffer.h', - '/media/libaom/src/aom/aom_image.h', - '/media/libaom/src/aom/aom_integer.h', - '/media/libaom/src/aom/aomcx.h', - '/media/libaom/src/aom/aomdx.h', - '/media/libaom/src/aom_mem/aom_mem.h', - '/media/libaom/src/aom_ports/aom_timer.h', - '/media/libaom/src/aom_ports/bitops.h', - '/media/libaom/src/aom_ports/mem.h', - '/media/libaom/src/aom_ports/sanitizer.h', - '/media/libaom/src/aom_ports/system_state.h', - '/media/libaom/src/aom_scale/aom_scale.h', - '/media/libaom/src/aom_scale/yv12config.h', ], 'GENERIC_SOURCES': [ - '/media/libaom/src/aom/src/aom_codec.c', - '/media/libaom/src/aom/src/aom_decoder.c', - '/media/libaom/src/aom/src/aom_encoder.c', - '/media/libaom/src/aom/src/aom_image.c', - '/media/libaom/src/aom/src/aom_integer.c', - '/media/libaom/src/aom_dsp/aom_convolve.c', - '/media/libaom/src/aom_dsp/aom_dsp_rtcd.c', - '/media/libaom/src/aom_dsp/avg.c', - '/media/libaom/src/aom_dsp/binary_codes_reader.c', - '/media/libaom/src/aom_dsp/bitreader.c', - '/media/libaom/src/aom_dsp/bitreader_buffer.c', - '/media/libaom/src/aom_dsp/bitwriter_buffer.c', - '/media/libaom/src/aom_dsp/blend_a64_hmask.c', - '/media/libaom/src/aom_dsp/blend_a64_mask.c', - '/media/libaom/src/aom_dsp/blend_a64_vmask.c', - '/media/libaom/src/aom_dsp/entcode.c', - '/media/libaom/src/aom_dsp/entdec.c', - '/media/libaom/src/aom_dsp/fft.c', - '/media/libaom/src/aom_dsp/grain_synthesis.c', - '/media/libaom/src/aom_dsp/intrapred.c', - '/media/libaom/src/aom_dsp/loopfilter.c', - '/media/libaom/src/aom_dsp/subtract.c', - '/media/libaom/src/aom_mem/aom_mem.c', - '/media/libaom/src/aom_scale/aom_scale_rtcd.c', - '/media/libaom/src/aom_scale/generic/aom_scale.c', - '/media/libaom/src/aom_scale/generic/gen_scalers.c', - '/media/libaom/src/aom_scale/generic/yv12config.c', - '/media/libaom/src/aom_scale/generic/yv12extend.c', - '/media/libaom/src/aom_util/aom_thread.c', - '/media/libaom/src/aom_util/debug_util.c', - '/media/libaom/src/av1/av1_dx_iface.c', - '/media/libaom/src/av1/common/alloccommon.c', - '/media/libaom/src/av1/common/av1_inv_txfm1d.c', - '/media/libaom/src/av1/common/av1_inv_txfm2d.c', - '/media/libaom/src/av1/common/av1_loopfilter.c', - '/media/libaom/src/av1/common/av1_rtcd.c', - '/media/libaom/src/av1/common/av1_txfm.c', - '/media/libaom/src/av1/common/blockd.c', - '/media/libaom/src/av1/common/cdef.c', - '/media/libaom/src/av1/common/cdef_block.c', - '/media/libaom/src/av1/common/cfl.c', - '/media/libaom/src/av1/common/convolve.c', - '/media/libaom/src/av1/common/debugmodes.c', - '/media/libaom/src/av1/common/entropy.c', - '/media/libaom/src/av1/common/entropymode.c', - '/media/libaom/src/av1/common/entropymv.c', - '/media/libaom/src/av1/common/frame_buffers.c', - '/media/libaom/src/av1/common/idct.c', - '/media/libaom/src/av1/common/mvref_common.c', - '/media/libaom/src/av1/common/obu_util.c', - '/media/libaom/src/av1/common/odintrin.c', - '/media/libaom/src/av1/common/pred_common.c', - '/media/libaom/src/av1/common/quant_common.c', - '/media/libaom/src/av1/common/reconinter.c', - '/media/libaom/src/av1/common/reconintra.c', - '/media/libaom/src/av1/common/resize.c', - '/media/libaom/src/av1/common/restoration.c', - '/media/libaom/src/av1/common/scale.c', - '/media/libaom/src/av1/common/scan.c', - '/media/libaom/src/av1/common/seg_common.c', - '/media/libaom/src/av1/common/thread_common.c', - '/media/libaom/src/av1/common/tile_common.c', - '/media/libaom/src/av1/common/timing.c', - '/media/libaom/src/av1/common/txb_common.c', - '/media/libaom/src/av1/common/warped_motion.c', - '/media/libaom/src/av1/decoder/decodeframe.c', - '/media/libaom/src/av1/decoder/decodemv.c', - '/media/libaom/src/av1/decoder/decoder.c', - '/media/libaom/src/av1/decoder/decodetxb.c', - '/media/libaom/src/av1/decoder/detokenize.c', - '/media/libaom/src/av1/decoder/obu.c', + './src/aom/src/aom_codec.c', + './src/aom/src/aom_decoder.c', + './src/aom/src/aom_encoder.c', + './src/aom/src/aom_image.c', + './src/aom/src/aom_integer.c', + './src/aom_dsp/aom_convolve.c', + './src/aom_dsp/aom_dsp_rtcd.c', + './src/aom_dsp/binary_codes_reader.c', + './src/aom_dsp/bitreader.c', + './src/aom_dsp/bitreader_buffer.c', + './src/aom_dsp/bitwriter_buffer.c', + './src/aom_dsp/blend_a64_hmask.c', + './src/aom_dsp/blend_a64_mask.c', + './src/aom_dsp/blend_a64_vmask.c', + './src/aom_dsp/entcode.c', + './src/aom_dsp/entdec.c', + './src/aom_dsp/fft.c', + './src/aom_dsp/intrapred.c', + './src/aom_dsp/loopfilter.c', + './src/aom_dsp/odintrin.c', + './src/aom_dsp/subtract.c', + './src/aom_mem/aom_mem.c', + './src/aom_scale/aom_scale_rtcd.c', + './src/aom_scale/generic/aom_scale.c', + './src/aom_scale/generic/gen_scalers.c', + './src/aom_scale/generic/yv12config.c', + './src/aom_scale/generic/yv12extend.c', + './src/aom_util/aom_thread.c', + './src/aom_util/debug_util.c', + './src/av1/arg_defs.c', + './src/av1/av1_dx_iface.c', + './src/av1/common/alloccommon.c', + './src/av1/common/av1_inv_txfm1d.c', + './src/av1/common/av1_inv_txfm2d.c', + './src/av1/common/av1_loopfilter.c', + './src/av1/common/av1_rtcd.c', + './src/av1/common/av1_txfm.c', + './src/av1/common/blockd.c', + './src/av1/common/cdef.c', + './src/av1/common/cdef_block.c', + './src/av1/common/cfl.c', + './src/av1/common/convolve.c', + './src/av1/common/debugmodes.c', + './src/av1/common/entropy.c', + './src/av1/common/entropymode.c', + './src/av1/common/entropymv.c', + './src/av1/common/frame_buffers.c', + './src/av1/common/idct.c', + './src/av1/common/mvref_common.c', + './src/av1/common/obu_util.c', + './src/av1/common/pred_common.c', + './src/av1/common/quant_common.c', + './src/av1/common/reconinter.c', + './src/av1/common/reconintra.c', + './src/av1/common/resize.c', + './src/av1/common/restoration.c', + './src/av1/common/scale.c', + './src/av1/common/scan.c', + './src/av1/common/seg_common.c', + './src/av1/common/thread_common.c', + './src/av1/common/tile_common.c', + './src/av1/common/timing.c', + './src/av1/common/txb_common.c', + './src/av1/common/warped_motion.c', + './src/av1/decoder/decodeframe.c', + './src/av1/decoder/decodemv.c', + './src/av1/decoder/decoder.c', + './src/av1/decoder/decodetxb.c', + './src/av1/decoder/detokenize.c', + './src/av1/decoder/grain_synthesis.c', + './src/av1/decoder/obu.c', + './src/common/args_helper.c', ], 'IA32_EXPORTS': [ - '/media/libaom/src/aom/aom.h', - '/media/libaom/src/aom/aom_codec.h', - '/media/libaom/src/aom/aom_decoder.h', - '/media/libaom/src/aom/aom_encoder.h', - '/media/libaom/src/aom/aom_frame_buffer.h', - '/media/libaom/src/aom/aom_image.h', - '/media/libaom/src/aom/aom_integer.h', - '/media/libaom/src/aom/aomcx.h', - '/media/libaom/src/aom/aomdx.h', - '/media/libaom/src/aom_mem/aom_mem.h', - '/media/libaom/src/aom_ports/aom_timer.h', - '/media/libaom/src/aom_ports/bitops.h', - '/media/libaom/src/aom_ports/mem.h', - '/media/libaom/src/aom_ports/sanitizer.h', - '/media/libaom/src/aom_ports/system_state.h', - '/media/libaom/src/aom_scale/aom_scale.h', - '/media/libaom/src/aom_scale/yv12config.h', ], 'IA32_SOURCES': [ - '/media/libaom/src/aom/src/aom_codec.c', - '/media/libaom/src/aom/src/aom_decoder.c', - '/media/libaom/src/aom/src/aom_encoder.c', - '/media/libaom/src/aom/src/aom_image.c', - '/media/libaom/src/aom/src/aom_integer.c', - '/media/libaom/src/aom_dsp/aom_convolve.c', - '/media/libaom/src/aom_dsp/aom_dsp_rtcd.c', - '/media/libaom/src/aom_dsp/avg.c', - '/media/libaom/src/aom_dsp/binary_codes_reader.c', - '/media/libaom/src/aom_dsp/bitreader.c', - '/media/libaom/src/aom_dsp/bitreader_buffer.c', - '/media/libaom/src/aom_dsp/bitwriter_buffer.c', - '/media/libaom/src/aom_dsp/blend_a64_hmask.c', - '/media/libaom/src/aom_dsp/blend_a64_mask.c', - '/media/libaom/src/aom_dsp/blend_a64_vmask.c', - '/media/libaom/src/aom_dsp/entcode.c', - '/media/libaom/src/aom_dsp/entdec.c', - '/media/libaom/src/aom_dsp/fft.c', - '/media/libaom/src/aom_dsp/grain_synthesis.c', - '/media/libaom/src/aom_dsp/intrapred.c', - '/media/libaom/src/aom_dsp/loopfilter.c', - '/media/libaom/src/aom_dsp/subtract.c', - '/media/libaom/src/aom_dsp/x86/aom_asm_stubs.c', - '/media/libaom/src/aom_dsp/x86/aom_convolve_copy_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_ssse3.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm', - '/media/libaom/src/aom_dsp/x86/avg_intrin_avx2.c', - '/media/libaom/src/aom_dsp/x86/avg_intrin_sse2.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_hmask_sse4.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_mask_avx2.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_mask_sse4.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_vmask_sse4.c', - '/media/libaom/src/aom_dsp/x86/fft_avx2.c', - '/media/libaom/src/aom_dsp/x86/fft_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_avx2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_ssse3.c', - '/media/libaom/src/aom_dsp/x86/highbd_intrapred_asm_sse2.asm', - '/media/libaom/src/aom_dsp/x86/highbd_intrapred_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_loopfilter_avx2.c', - '/media/libaom/src/aom_dsp/x86/highbd_loopfilter_sse2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_asm_sse2.asm', - '/media/libaom/src/aom_dsp/x86/intrapred_avx2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_sse2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_ssse3.c', - '/media/libaom/src/aom_dsp/x86/inv_wht_sse2.asm', - '/media/libaom/src/aom_dsp/x86/loopfilter_sse2.c', - '/media/libaom/src/aom_mem/aom_mem.c', - '/media/libaom/src/aom_ports/emms.asm', - '/media/libaom/src/aom_ports/x86_abi_support.asm', - '/media/libaom/src/aom_scale/aom_scale_rtcd.c', - '/media/libaom/src/aom_scale/generic/aom_scale.c', - '/media/libaom/src/aom_scale/generic/gen_scalers.c', - '/media/libaom/src/aom_scale/generic/yv12config.c', - '/media/libaom/src/aom_scale/generic/yv12extend.c', - '/media/libaom/src/aom_util/aom_thread.c', - '/media/libaom/src/aom_util/debug_util.c', - '/media/libaom/src/av1/av1_dx_iface.c', - '/media/libaom/src/av1/common/alloccommon.c', - '/media/libaom/src/av1/common/av1_inv_txfm1d.c', - '/media/libaom/src/av1/common/av1_inv_txfm2d.c', - '/media/libaom/src/av1/common/av1_loopfilter.c', - '/media/libaom/src/av1/common/av1_rtcd.c', - '/media/libaom/src/av1/common/av1_txfm.c', - '/media/libaom/src/av1/common/blockd.c', - '/media/libaom/src/av1/common/cdef.c', - '/media/libaom/src/av1/common/cdef_block.c', - '/media/libaom/src/av1/common/cdef_block_avx2.c', - '/media/libaom/src/av1/common/cdef_block_sse2.c', - '/media/libaom/src/av1/common/cdef_block_sse4.c', - '/media/libaom/src/av1/common/cdef_block_ssse3.c', - '/media/libaom/src/av1/common/cfl.c', - '/media/libaom/src/av1/common/convolve.c', - '/media/libaom/src/av1/common/debugmodes.c', - '/media/libaom/src/av1/common/entropy.c', - '/media/libaom/src/av1/common/entropymode.c', - '/media/libaom/src/av1/common/entropymv.c', - '/media/libaom/src/av1/common/frame_buffers.c', - '/media/libaom/src/av1/common/idct.c', - '/media/libaom/src/av1/common/mvref_common.c', - '/media/libaom/src/av1/common/obu_util.c', - '/media/libaom/src/av1/common/odintrin.c', - '/media/libaom/src/av1/common/pred_common.c', - '/media/libaom/src/av1/common/quant_common.c', - '/media/libaom/src/av1/common/reconinter.c', - '/media/libaom/src/av1/common/reconintra.c', - '/media/libaom/src/av1/common/resize.c', - '/media/libaom/src/av1/common/restoration.c', - '/media/libaom/src/av1/common/scale.c', - '/media/libaom/src/av1/common/scan.c', - '/media/libaom/src/av1/common/seg_common.c', - '/media/libaom/src/av1/common/thread_common.c', - '/media/libaom/src/av1/common/tile_common.c', - '/media/libaom/src/av1/common/timing.c', - '/media/libaom/src/av1/common/txb_common.c', - '/media/libaom/src/av1/common/warped_motion.c', - '/media/libaom/src/av1/common/x86/av1_convolve_horiz_rs_sse4.c', - '/media/libaom/src/av1/common/x86/av1_convolve_scale_sse4.c', - '/media/libaom/src/av1/common/x86/av1_inv_txfm_avx2.c', - '/media/libaom/src/av1/common/x86/av1_inv_txfm_ssse3.c', - '/media/libaom/src/av1/common/x86/av1_txfm_sse4.c', - '/media/libaom/src/av1/common/x86/cfl_avx2.c', - '/media/libaom/src/av1/common/x86/cfl_sse2.c', - '/media/libaom/src/av1/common/x86/cfl_ssse3.c', - '/media/libaom/src/av1/common/x86/convolve_2d_avx2.c', - '/media/libaom/src/av1/common/x86/convolve_2d_sse2.c', - '/media/libaom/src/av1/common/x86/convolve_avx2.c', - '/media/libaom/src/av1/common/x86/convolve_sse2.c', - '/media/libaom/src/av1/common/x86/filterintra_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_sse2.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_ssse3.c', - '/media/libaom/src/av1/common/x86/highbd_inv_txfm_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_inv_txfm_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_jnt_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_jnt_convolve_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_warp_plane_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_wiener_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_wiener_convolve_ssse3.c', - '/media/libaom/src/av1/common/x86/intra_edge_sse4.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_sse2.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_ssse3.c', - '/media/libaom/src/av1/common/x86/reconinter_avx2.c', - '/media/libaom/src/av1/common/x86/reconinter_sse4.c', - '/media/libaom/src/av1/common/x86/reconinter_ssse3.c', - '/media/libaom/src/av1/common/x86/selfguided_avx2.c', - '/media/libaom/src/av1/common/x86/selfguided_sse4.c', - '/media/libaom/src/av1/common/x86/warp_plane_avx2.c', - '/media/libaom/src/av1/common/x86/warp_plane_sse2.c', - '/media/libaom/src/av1/common/x86/warp_plane_sse4.c', - '/media/libaom/src/av1/common/x86/wiener_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/wiener_convolve_sse2.c', - '/media/libaom/src/av1/decoder/decodeframe.c', - '/media/libaom/src/av1/decoder/decodemv.c', - '/media/libaom/src/av1/decoder/decoder.c', - '/media/libaom/src/av1/decoder/decodetxb.c', - '/media/libaom/src/av1/decoder/detokenize.c', - '/media/libaom/src/av1/decoder/obu.c', + './src/aom/src/aom_codec.c', + './src/aom/src/aom_decoder.c', + './src/aom/src/aom_encoder.c', + './src/aom/src/aom_image.c', + './src/aom/src/aom_integer.c', + './src/aom_dsp/aom_convolve.c', + './src/aom_dsp/aom_dsp_rtcd.c', + './src/aom_dsp/binary_codes_reader.c', + './src/aom_dsp/bitreader.c', + './src/aom_dsp/bitreader_buffer.c', + './src/aom_dsp/bitwriter_buffer.c', + './src/aom_dsp/blend_a64_hmask.c', + './src/aom_dsp/blend_a64_mask.c', + './src/aom_dsp/blend_a64_vmask.c', + './src/aom_dsp/entcode.c', + './src/aom_dsp/entdec.c', + './src/aom_dsp/fft.c', + './src/aom_dsp/intrapred.c', + './src/aom_dsp/loopfilter.c', + './src/aom_dsp/odintrin.c', + './src/aom_dsp/subtract.c', + './src/aom_dsp/x86/aom_asm_stubs.c', + './src/aom_dsp/x86/aom_convolve_copy_avx2.c', + './src/aom_dsp/x86/aom_convolve_copy_sse2.c', + './src/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm', + './src/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c', + './src/aom_dsp/x86/aom_subpixel_8t_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_8t_ssse3.asm', + './src/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm', + './src/aom_dsp/x86/blend_a64_hmask_sse4.c', + './src/aom_dsp/x86/blend_a64_mask_avx2.c', + './src/aom_dsp/x86/blend_a64_mask_sse4.c', + './src/aom_dsp/x86/blend_a64_vmask_sse4.c', + './src/aom_dsp/x86/fft_avx2.c', + './src/aom_dsp/x86/fft_sse2.c', + './src/aom_dsp/x86/highbd_convolve_avx2.c', + './src/aom_dsp/x86/highbd_convolve_sse2.c', + './src/aom_dsp/x86/highbd_convolve_ssse3.c', + './src/aom_dsp/x86/highbd_intrapred_asm_sse2.asm', + './src/aom_dsp/x86/highbd_intrapred_sse2.c', + './src/aom_dsp/x86/highbd_loopfilter_avx2.c', + './src/aom_dsp/x86/highbd_loopfilter_sse2.c', + './src/aom_dsp/x86/intrapred_asm_sse2.asm', + './src/aom_dsp/x86/intrapred_avx2.c', + './src/aom_dsp/x86/intrapred_sse2.c', + './src/aom_dsp/x86/intrapred_sse4.c', + './src/aom_dsp/x86/intrapred_ssse3.c', + './src/aom_dsp/x86/inv_wht_sse2.asm', + './src/aom_dsp/x86/loopfilter_avx2.c', + './src/aom_dsp/x86/loopfilter_sse2.c', + './src/aom_mem/aom_mem.c', + './src/aom_scale/aom_scale_rtcd.c', + './src/aom_scale/generic/aom_scale.c', + './src/aom_scale/generic/gen_scalers.c', + './src/aom_scale/generic/yv12config.c', + './src/aom_scale/generic/yv12extend.c', + './src/aom_util/aom_thread.c', + './src/aom_util/debug_util.c', + './src/av1/arg_defs.c', + './src/av1/av1_dx_iface.c', + './src/av1/common/alloccommon.c', + './src/av1/common/av1_inv_txfm1d.c', + './src/av1/common/av1_inv_txfm2d.c', + './src/av1/common/av1_loopfilter.c', + './src/av1/common/av1_rtcd.c', + './src/av1/common/av1_txfm.c', + './src/av1/common/blockd.c', + './src/av1/common/cdef.c', + './src/av1/common/cdef_block.c', + './src/av1/common/cfl.c', + './src/av1/common/convolve.c', + './src/av1/common/debugmodes.c', + './src/av1/common/entropy.c', + './src/av1/common/entropymode.c', + './src/av1/common/entropymv.c', + './src/av1/common/frame_buffers.c', + './src/av1/common/idct.c', + './src/av1/common/mvref_common.c', + './src/av1/common/obu_util.c', + './src/av1/common/pred_common.c', + './src/av1/common/quant_common.c', + './src/av1/common/reconinter.c', + './src/av1/common/reconintra.c', + './src/av1/common/resize.c', + './src/av1/common/restoration.c', + './src/av1/common/scale.c', + './src/av1/common/scan.c', + './src/av1/common/seg_common.c', + './src/av1/common/thread_common.c', + './src/av1/common/tile_common.c', + './src/av1/common/timing.c', + './src/av1/common/txb_common.c', + './src/av1/common/warped_motion.c', + './src/av1/common/x86/av1_convolve_horiz_rs_sse4.c', + './src/av1/common/x86/av1_convolve_scale_sse4.c', + './src/av1/common/x86/av1_inv_txfm_avx2.c', + './src/av1/common/x86/av1_inv_txfm_ssse3.c', + './src/av1/common/x86/av1_txfm_sse4.c', + './src/av1/common/x86/cdef_block_avx2.c', + './src/av1/common/x86/cdef_block_sse2.c', + './src/av1/common/x86/cdef_block_sse4.c', + './src/av1/common/x86/cdef_block_ssse3.c', + './src/av1/common/x86/cfl_avx2.c', + './src/av1/common/x86/cfl_sse2.c', + './src/av1/common/x86/cfl_ssse3.c', + './src/av1/common/x86/convolve_2d_avx2.c', + './src/av1/common/x86/convolve_2d_sse2.c', + './src/av1/common/x86/convolve_avx2.c', + './src/av1/common/x86/convolve_sse2.c', + './src/av1/common/x86/filterintra_sse4.c', + './src/av1/common/x86/highbd_convolve_2d_avx2.c', + './src/av1/common/x86/highbd_convolve_2d_sse4.c', + './src/av1/common/x86/highbd_convolve_2d_ssse3.c', + './src/av1/common/x86/highbd_inv_txfm_avx2.c', + './src/av1/common/x86/highbd_inv_txfm_sse4.c', + './src/av1/common/x86/highbd_jnt_convolve_avx2.c', + './src/av1/common/x86/highbd_jnt_convolve_sse4.c', + './src/av1/common/x86/highbd_warp_affine_avx2.c', + './src/av1/common/x86/highbd_warp_plane_sse4.c', + './src/av1/common/x86/highbd_wiener_convolve_avx2.c', + './src/av1/common/x86/highbd_wiener_convolve_ssse3.c', + './src/av1/common/x86/intra_edge_sse4.c', + './src/av1/common/x86/jnt_convolve_avx2.c', + './src/av1/common/x86/jnt_convolve_sse2.c', + './src/av1/common/x86/jnt_convolve_ssse3.c', + './src/av1/common/x86/reconinter_avx2.c', + './src/av1/common/x86/reconinter_sse4.c', + './src/av1/common/x86/reconinter_ssse3.c', + './src/av1/common/x86/resize_ssse3.c', + './src/av1/common/x86/selfguided_avx2.c', + './src/av1/common/x86/selfguided_sse4.c', + './src/av1/common/x86/warp_plane_avx2.c', + './src/av1/common/x86/warp_plane_sse2.c', + './src/av1/common/x86/warp_plane_sse4.c', + './src/av1/common/x86/wiener_convolve_avx2.c', + './src/av1/common/x86/wiener_convolve_sse2.c', + './src/av1/decoder/decodeframe.c', + './src/av1/decoder/decodemv.c', + './src/av1/decoder/decoder.c', + './src/av1/decoder/decodetxb.c', + './src/av1/decoder/detokenize.c', + './src/av1/decoder/grain_synthesis.c', + './src/av1/decoder/obu.c', + './src/common/args_helper.c', ], 'X64_EXPORTS': [ - '/media/libaom/src/aom/aom.h', - '/media/libaom/src/aom/aom_codec.h', - '/media/libaom/src/aom/aom_decoder.h', - '/media/libaom/src/aom/aom_encoder.h', - '/media/libaom/src/aom/aom_frame_buffer.h', - '/media/libaom/src/aom/aom_image.h', - '/media/libaom/src/aom/aom_integer.h', - '/media/libaom/src/aom/aomcx.h', - '/media/libaom/src/aom/aomdx.h', - '/media/libaom/src/aom_mem/aom_mem.h', - '/media/libaom/src/aom_ports/aom_timer.h', - '/media/libaom/src/aom_ports/bitops.h', - '/media/libaom/src/aom_ports/mem.h', - '/media/libaom/src/aom_ports/sanitizer.h', - '/media/libaom/src/aom_ports/system_state.h', - '/media/libaom/src/aom_scale/aom_scale.h', - '/media/libaom/src/aom_scale/yv12config.h', ], 'X64_SOURCES': [ - '/media/libaom/src/aom/src/aom_codec.c', - '/media/libaom/src/aom/src/aom_decoder.c', - '/media/libaom/src/aom/src/aom_encoder.c', - '/media/libaom/src/aom/src/aom_image.c', - '/media/libaom/src/aom/src/aom_integer.c', - '/media/libaom/src/aom_dsp/aom_convolve.c', - '/media/libaom/src/aom_dsp/aom_dsp_rtcd.c', - '/media/libaom/src/aom_dsp/avg.c', - '/media/libaom/src/aom_dsp/binary_codes_reader.c', - '/media/libaom/src/aom_dsp/bitreader.c', - '/media/libaom/src/aom_dsp/bitreader_buffer.c', - '/media/libaom/src/aom_dsp/bitwriter_buffer.c', - '/media/libaom/src/aom_dsp/blend_a64_hmask.c', - '/media/libaom/src/aom_dsp/blend_a64_mask.c', - '/media/libaom/src/aom_dsp/blend_a64_vmask.c', - '/media/libaom/src/aom_dsp/entcode.c', - '/media/libaom/src/aom_dsp/entdec.c', - '/media/libaom/src/aom_dsp/fft.c', - '/media/libaom/src/aom_dsp/grain_synthesis.c', - '/media/libaom/src/aom_dsp/intrapred.c', - '/media/libaom/src/aom_dsp/loopfilter.c', - '/media/libaom/src/aom_dsp/subtract.c', - '/media/libaom/src/aom_dsp/x86/aom_asm_stubs.c', - '/media/libaom/src/aom_dsp/x86/aom_convolve_copy_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_8t_ssse3.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm', - '/media/libaom/src/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm', - '/media/libaom/src/aom_dsp/x86/avg_intrin_avx2.c', - '/media/libaom/src/aom_dsp/x86/avg_intrin_sse2.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_hmask_sse4.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_mask_avx2.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_mask_sse4.c', - '/media/libaom/src/aom_dsp/x86/blend_a64_vmask_sse4.c', - '/media/libaom/src/aom_dsp/x86/fft_avx2.c', - '/media/libaom/src/aom_dsp/x86/fft_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_avx2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_convolve_ssse3.c', - '/media/libaom/src/aom_dsp/x86/highbd_intrapred_asm_sse2.asm', - '/media/libaom/src/aom_dsp/x86/highbd_intrapred_sse2.c', - '/media/libaom/src/aom_dsp/x86/highbd_loopfilter_avx2.c', - '/media/libaom/src/aom_dsp/x86/highbd_loopfilter_sse2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_asm_sse2.asm', - '/media/libaom/src/aom_dsp/x86/intrapred_avx2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_sse2.c', - '/media/libaom/src/aom_dsp/x86/intrapred_ssse3.c', - '/media/libaom/src/aom_dsp/x86/inv_wht_sse2.asm', - '/media/libaom/src/aom_dsp/x86/loopfilter_sse2.c', - '/media/libaom/src/aom_mem/aom_mem.c', - '/media/libaom/src/aom_ports/emms.asm', - '/media/libaom/src/aom_scale/aom_scale_rtcd.c', - '/media/libaom/src/aom_scale/generic/aom_scale.c', - '/media/libaom/src/aom_scale/generic/gen_scalers.c', - '/media/libaom/src/aom_scale/generic/yv12config.c', - '/media/libaom/src/aom_scale/generic/yv12extend.c', - '/media/libaom/src/aom_util/aom_thread.c', - '/media/libaom/src/aom_util/debug_util.c', - '/media/libaom/src/av1/av1_dx_iface.c', - '/media/libaom/src/av1/common/alloccommon.c', - '/media/libaom/src/av1/common/av1_inv_txfm1d.c', - '/media/libaom/src/av1/common/av1_inv_txfm2d.c', - '/media/libaom/src/av1/common/av1_loopfilter.c', - '/media/libaom/src/av1/common/av1_rtcd.c', - '/media/libaom/src/av1/common/av1_txfm.c', - '/media/libaom/src/av1/common/blockd.c', - '/media/libaom/src/av1/common/cdef.c', - '/media/libaom/src/av1/common/cdef_block.c', - '/media/libaom/src/av1/common/cdef_block_avx2.c', - '/media/libaom/src/av1/common/cdef_block_sse2.c', - '/media/libaom/src/av1/common/cdef_block_sse4.c', - '/media/libaom/src/av1/common/cdef_block_ssse3.c', - '/media/libaom/src/av1/common/cfl.c', - '/media/libaom/src/av1/common/convolve.c', - '/media/libaom/src/av1/common/debugmodes.c', - '/media/libaom/src/av1/common/entropy.c', - '/media/libaom/src/av1/common/entropymode.c', - '/media/libaom/src/av1/common/entropymv.c', - '/media/libaom/src/av1/common/frame_buffers.c', - '/media/libaom/src/av1/common/idct.c', - '/media/libaom/src/av1/common/mvref_common.c', - '/media/libaom/src/av1/common/obu_util.c', - '/media/libaom/src/av1/common/odintrin.c', - '/media/libaom/src/av1/common/pred_common.c', - '/media/libaom/src/av1/common/quant_common.c', - '/media/libaom/src/av1/common/reconinter.c', - '/media/libaom/src/av1/common/reconintra.c', - '/media/libaom/src/av1/common/resize.c', - '/media/libaom/src/av1/common/restoration.c', - '/media/libaom/src/av1/common/scale.c', - '/media/libaom/src/av1/common/scan.c', - '/media/libaom/src/av1/common/seg_common.c', - '/media/libaom/src/av1/common/thread_common.c', - '/media/libaom/src/av1/common/tile_common.c', - '/media/libaom/src/av1/common/timing.c', - '/media/libaom/src/av1/common/txb_common.c', - '/media/libaom/src/av1/common/warped_motion.c', - '/media/libaom/src/av1/common/x86/av1_convolve_horiz_rs_sse4.c', - '/media/libaom/src/av1/common/x86/av1_convolve_scale_sse4.c', - '/media/libaom/src/av1/common/x86/av1_inv_txfm_avx2.c', - '/media/libaom/src/av1/common/x86/av1_inv_txfm_ssse3.c', - '/media/libaom/src/av1/common/x86/av1_txfm_sse4.c', - '/media/libaom/src/av1/common/x86/cfl_avx2.c', - '/media/libaom/src/av1/common/x86/cfl_sse2.c', - '/media/libaom/src/av1/common/x86/cfl_ssse3.c', - '/media/libaom/src/av1/common/x86/convolve_2d_avx2.c', - '/media/libaom/src/av1/common/x86/convolve_2d_sse2.c', - '/media/libaom/src/av1/common/x86/convolve_avx2.c', - '/media/libaom/src/av1/common/x86/convolve_sse2.c', - '/media/libaom/src/av1/common/x86/filterintra_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_sse2.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_convolve_2d_ssse3.c', - '/media/libaom/src/av1/common/x86/highbd_inv_txfm_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_inv_txfm_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_jnt_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_jnt_convolve_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_warp_plane_sse4.c', - '/media/libaom/src/av1/common/x86/highbd_wiener_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/highbd_wiener_convolve_ssse3.c', - '/media/libaom/src/av1/common/x86/intra_edge_sse4.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_sse2.c', - '/media/libaom/src/av1/common/x86/jnt_convolve_ssse3.c', - '/media/libaom/src/av1/common/x86/reconinter_avx2.c', - '/media/libaom/src/av1/common/x86/reconinter_sse4.c', - '/media/libaom/src/av1/common/x86/reconinter_ssse3.c', - '/media/libaom/src/av1/common/x86/selfguided_avx2.c', - '/media/libaom/src/av1/common/x86/selfguided_sse4.c', - '/media/libaom/src/av1/common/x86/warp_plane_avx2.c', - '/media/libaom/src/av1/common/x86/warp_plane_sse2.c', - '/media/libaom/src/av1/common/x86/warp_plane_sse4.c', - '/media/libaom/src/av1/common/x86/wiener_convolve_avx2.c', - '/media/libaom/src/av1/common/x86/wiener_convolve_sse2.c', - '/media/libaom/src/av1/decoder/decodeframe.c', - '/media/libaom/src/av1/decoder/decodemv.c', - '/media/libaom/src/av1/decoder/decoder.c', - '/media/libaom/src/av1/decoder/decodetxb.c', - '/media/libaom/src/av1/decoder/detokenize.c', - '/media/libaom/src/av1/decoder/obu.c', + './src/aom/src/aom_codec.c', + './src/aom/src/aom_decoder.c', + './src/aom/src/aom_encoder.c', + './src/aom/src/aom_image.c', + './src/aom/src/aom_integer.c', + './src/aom_dsp/aom_convolve.c', + './src/aom_dsp/aom_dsp_rtcd.c', + './src/aom_dsp/binary_codes_reader.c', + './src/aom_dsp/bitreader.c', + './src/aom_dsp/bitreader_buffer.c', + './src/aom_dsp/bitwriter_buffer.c', + './src/aom_dsp/blend_a64_hmask.c', + './src/aom_dsp/blend_a64_mask.c', + './src/aom_dsp/blend_a64_vmask.c', + './src/aom_dsp/entcode.c', + './src/aom_dsp/entdec.c', + './src/aom_dsp/fft.c', + './src/aom_dsp/intrapred.c', + './src/aom_dsp/loopfilter.c', + './src/aom_dsp/odintrin.c', + './src/aom_dsp/subtract.c', + './src/aom_dsp/x86/aom_asm_stubs.c', + './src/aom_dsp/x86/aom_convolve_copy_avx2.c', + './src/aom_dsp/x86/aom_convolve_copy_sse2.c', + './src/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm', + './src/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c', + './src/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c', + './src/aom_dsp/x86/aom_subpixel_8t_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_8t_ssse3.asm', + './src/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm', + './src/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm', + './src/aom_dsp/x86/blend_a64_hmask_sse4.c', + './src/aom_dsp/x86/blend_a64_mask_avx2.c', + './src/aom_dsp/x86/blend_a64_mask_sse4.c', + './src/aom_dsp/x86/blend_a64_vmask_sse4.c', + './src/aom_dsp/x86/fft_avx2.c', + './src/aom_dsp/x86/fft_sse2.c', + './src/aom_dsp/x86/highbd_convolve_avx2.c', + './src/aom_dsp/x86/highbd_convolve_sse2.c', + './src/aom_dsp/x86/highbd_convolve_ssse3.c', + './src/aom_dsp/x86/highbd_intrapred_asm_sse2.asm', + './src/aom_dsp/x86/highbd_intrapred_sse2.c', + './src/aom_dsp/x86/highbd_loopfilter_avx2.c', + './src/aom_dsp/x86/highbd_loopfilter_sse2.c', + './src/aom_dsp/x86/intrapred_asm_sse2.asm', + './src/aom_dsp/x86/intrapred_avx2.c', + './src/aom_dsp/x86/intrapred_sse2.c', + './src/aom_dsp/x86/intrapred_sse4.c', + './src/aom_dsp/x86/intrapred_ssse3.c', + './src/aom_dsp/x86/inv_wht_sse2.asm', + './src/aom_dsp/x86/loopfilter_avx2.c', + './src/aom_dsp/x86/loopfilter_sse2.c', + './src/aom_mem/aom_mem.c', + './src/aom_scale/aom_scale_rtcd.c', + './src/aom_scale/generic/aom_scale.c', + './src/aom_scale/generic/gen_scalers.c', + './src/aom_scale/generic/yv12config.c', + './src/aom_scale/generic/yv12extend.c', + './src/aom_util/aom_thread.c', + './src/aom_util/debug_util.c', + './src/av1/arg_defs.c', + './src/av1/av1_dx_iface.c', + './src/av1/common/alloccommon.c', + './src/av1/common/av1_inv_txfm1d.c', + './src/av1/common/av1_inv_txfm2d.c', + './src/av1/common/av1_loopfilter.c', + './src/av1/common/av1_rtcd.c', + './src/av1/common/av1_txfm.c', + './src/av1/common/blockd.c', + './src/av1/common/cdef.c', + './src/av1/common/cdef_block.c', + './src/av1/common/cfl.c', + './src/av1/common/convolve.c', + './src/av1/common/debugmodes.c', + './src/av1/common/entropy.c', + './src/av1/common/entropymode.c', + './src/av1/common/entropymv.c', + './src/av1/common/frame_buffers.c', + './src/av1/common/idct.c', + './src/av1/common/mvref_common.c', + './src/av1/common/obu_util.c', + './src/av1/common/pred_common.c', + './src/av1/common/quant_common.c', + './src/av1/common/reconinter.c', + './src/av1/common/reconintra.c', + './src/av1/common/resize.c', + './src/av1/common/restoration.c', + './src/av1/common/scale.c', + './src/av1/common/scan.c', + './src/av1/common/seg_common.c', + './src/av1/common/thread_common.c', + './src/av1/common/tile_common.c', + './src/av1/common/timing.c', + './src/av1/common/txb_common.c', + './src/av1/common/warped_motion.c', + './src/av1/common/x86/av1_convolve_horiz_rs_sse4.c', + './src/av1/common/x86/av1_convolve_scale_sse4.c', + './src/av1/common/x86/av1_inv_txfm_avx2.c', + './src/av1/common/x86/av1_inv_txfm_ssse3.c', + './src/av1/common/x86/av1_txfm_sse4.c', + './src/av1/common/x86/cdef_block_avx2.c', + './src/av1/common/x86/cdef_block_sse2.c', + './src/av1/common/x86/cdef_block_sse4.c', + './src/av1/common/x86/cdef_block_ssse3.c', + './src/av1/common/x86/cfl_avx2.c', + './src/av1/common/x86/cfl_sse2.c', + './src/av1/common/x86/cfl_ssse3.c', + './src/av1/common/x86/convolve_2d_avx2.c', + './src/av1/common/x86/convolve_2d_sse2.c', + './src/av1/common/x86/convolve_avx2.c', + './src/av1/common/x86/convolve_sse2.c', + './src/av1/common/x86/filterintra_sse4.c', + './src/av1/common/x86/highbd_convolve_2d_avx2.c', + './src/av1/common/x86/highbd_convolve_2d_sse4.c', + './src/av1/common/x86/highbd_convolve_2d_ssse3.c', + './src/av1/common/x86/highbd_inv_txfm_avx2.c', + './src/av1/common/x86/highbd_inv_txfm_sse4.c', + './src/av1/common/x86/highbd_jnt_convolve_avx2.c', + './src/av1/common/x86/highbd_jnt_convolve_sse4.c', + './src/av1/common/x86/highbd_warp_affine_avx2.c', + './src/av1/common/x86/highbd_warp_plane_sse4.c', + './src/av1/common/x86/highbd_wiener_convolve_avx2.c', + './src/av1/common/x86/highbd_wiener_convolve_ssse3.c', + './src/av1/common/x86/intra_edge_sse4.c', + './src/av1/common/x86/jnt_convolve_avx2.c', + './src/av1/common/x86/jnt_convolve_sse2.c', + './src/av1/common/x86/jnt_convolve_ssse3.c', + './src/av1/common/x86/reconinter_avx2.c', + './src/av1/common/x86/reconinter_sse4.c', + './src/av1/common/x86/reconinter_ssse3.c', + './src/av1/common/x86/resize_ssse3.c', + './src/av1/common/x86/selfguided_avx2.c', + './src/av1/common/x86/selfguided_sse4.c', + './src/av1/common/x86/warp_plane_avx2.c', + './src/av1/common/x86/warp_plane_sse2.c', + './src/av1/common/x86/warp_plane_sse4.c', + './src/av1/common/x86/wiener_convolve_avx2.c', + './src/av1/common/x86/wiener_convolve_sse2.c', + './src/av1/decoder/decodeframe.c', + './src/av1/decoder/decodemv.c', + './src/av1/decoder/decoder.c', + './src/av1/decoder/decodetxb.c', + './src/av1/decoder/detokenize.c', + './src/av1/decoder/grain_synthesis.c', + './src/av1/decoder/obu.c', + './src/common/args_helper.c', ], } |