Bug 1314147 - Generate build description for libaom. r=froydnj,kinetik
authorRalph Giles <giles@mozilla.com>
Fri, 14 Apr 2017 10:36:43 -0700
changeset 356837 69beada73635a06d1577c73876e708b067a3ab35
parent 356836 e9cfbd2b4fe2654522d8d325228a529571f4a55f
child 356838 c57f94d10914aa00968ac3d61f85fbf224b639c0
push id31775
push userihsiao@mozilla.com
push dateMon, 08 May 2017 03:10:38 +0000
treeherdermozilla-central@22aaf8bad4df [default view] [failures only]
perfherder[talos] [build metrics] [platform microbench] (compared to previous push)
reviewersfroydnj, kinetik
bugs1314147
milestone55.0a1
first release with
nightly linux32
nightly linux64
nightly mac
nightly win32
nightly win64
last release without
nightly linux32
nightly linux64
nightly mac
nightly win32
nightly win64
Bug 1314147 - Generate build description for libaom. r=froydnj,kinetik Result of running generate_sources_mozbuild.sh. MozReview-Commit-ID: DFmYUs5LRlG
media/libaom/README_MOZILLA
media/libaom/config/aom_version.h
media/libaom/config/generic/aom_config.asm
media/libaom/config/generic/aom_config.c
media/libaom/config/generic/aom_config.h
media/libaom/config/generic/aom_dsp_rtcd.h
media/libaom/config/generic/aom_scale_rtcd.h
media/libaom/config/generic/av1_rtcd.h
media/libaom/config/linux/arm/aom_config.asm
media/libaom/config/linux/arm/aom_config.c
media/libaom/config/linux/arm/aom_config.h
media/libaom/config/linux/arm/aom_dsp_rtcd.h
media/libaom/config/linux/arm/aom_scale_rtcd.h
media/libaom/config/linux/arm/av1_rtcd.h
media/libaom/config/linux/ia32/aom_config.asm
media/libaom/config/linux/ia32/aom_config.c
media/libaom/config/linux/ia32/aom_config.h
media/libaom/config/linux/ia32/aom_dsp_rtcd.h
media/libaom/config/linux/ia32/aom_scale_rtcd.h
media/libaom/config/linux/ia32/av1_rtcd.h
media/libaom/config/linux/x64/aom_config.asm
media/libaom/config/linux/x64/aom_config.c
media/libaom/config/linux/x64/aom_config.h
media/libaom/config/linux/x64/aom_dsp_rtcd.h
media/libaom/config/linux/x64/aom_scale_rtcd.h
media/libaom/config/linux/x64/av1_rtcd.h
media/libaom/config/mac/x64/aom_config.asm
media/libaom/config/mac/x64/aom_config.c
media/libaom/config/mac/x64/aom_config.h
media/libaom/config/mac/x64/aom_dsp_rtcd.h
media/libaom/config/mac/x64/aom_scale_rtcd.h
media/libaom/config/mac/x64/av1_rtcd.h
media/libaom/config/win/ia32/aom_config.asm
media/libaom/config/win/ia32/aom_config.c
media/libaom/config/win/ia32/aom_config.h
media/libaom/config/win/ia32/aom_dsp_rtcd.h
media/libaom/config/win/ia32/aom_scale_rtcd.h
media/libaom/config/win/ia32/av1_rtcd.h
media/libaom/config/win/x64/aom_config.asm
media/libaom/config/win/x64/aom_config.c
media/libaom/config/win/x64/aom_config.h
media/libaom/config/win/x64/aom_dsp_rtcd.h
media/libaom/config/win/x64/aom_scale_rtcd.h
media/libaom/config/win/x64/av1_rtcd.h
media/libaom/sources.mozbuild
--- a/media/libaom/README_MOZILLA
+++ b/media/libaom/README_MOZILLA
@@ -12,8 +12,10 @@ To update the library source and build c
 
 To update to a specific upstream git tag or commit, use
 
   ./mach vendor aom -r <commit>
 
 The upstream aom git repository is:
 
     https://aomedia.googlesource.com/aom
+
+The git commit ID used was 4d668d7feb1f8abd809d1bca0418570a7f142a36 (Wed May 03 17:10:13 2017 +0000).
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/aom_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  0
+#define VERSION_MINOR  0
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "Release"
+#define VERSION_STRING      " Release"
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/aom_config.asm
@@ -0,0 +1,155 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  0
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  0
+.equ HAVE_NEON ,  0
+.equ HAVE_NEON_ASM ,  0
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_AOM_PORTS ,  1
+.equ HAVE_FEXCEPT ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_WXWIDGETS ,  0
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  1
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  0
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_AV1_ENCODER ,  1
+.equ CONFIG_AV1_DECODER ,  1
+.equ CONFIG_AV1 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  0
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  0
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_ACCOUNTING ,  0
+.equ CONFIG_INSPECTION ,  0
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_LOWBITDEPTH ,  1
+.equ CONFIG_HIGHBITDEPTH ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  1
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_CDEF ,  1
+.equ CONFIG_VAR_TX ,  0
+.equ CONFIG_RECT_TX ,  1
+.equ CONFIG_REF_MV ,  1
+.equ CONFIG_TPL_MV ,  0
+.equ CONFIG_DUAL_FILTER ,  1
+.equ CONFIG_CONVOLVE_ROUND ,  0
+.equ CONFIG_COMPOUND_ROUND ,  0
+.equ CONFIG_EXT_TX ,  0
+.equ CONFIG_TX64X64 ,  0
+.equ CONFIG_SUB8X8_MC ,  0
+.equ CONFIG_EXT_INTRA ,  1
+.equ CONFIG_INTRA_INTERP ,  0
+.equ CONFIG_FILTER_INTRA ,  0
+.equ CONFIG_INTRABC ,  0
+.equ CONFIG_EXT_INTER ,  0
+.equ CONFIG_INTERINTRA ,  0
+.equ CONFIG_WEDGE ,  0
+.equ CONFIG_COMPOUND_SEGMENT ,  0
+.equ CONFIG_EXT_REFS ,  1
+.equ CONFIG_GLOBAL_MOTION ,  1
+.equ CONFIG_NEW_QUANT ,  0
+.equ CONFIG_SUPERTX ,  0
+.equ CONFIG_ANS ,  0
+.equ CONFIG_EC_MULTISYMBOL ,  1
+.equ CONFIG_NEW_TOKENSET ,  1
+.equ CONFIG_LOOP_RESTORATION ,  0
+.equ CONFIG_EXT_PARTITION ,  0
+.equ CONFIG_EXT_PARTITION_TYPES ,  0
+.equ CONFIG_UNPOISON_PARTITION_CTX ,  0
+.equ CONFIG_EXT_TILE ,  0
+.equ CONFIG_MOTION_VAR ,  1
+.equ CONFIG_NCOBMC ,  0
+.equ CONFIG_WARPED_MOTION ,  1
+.equ CONFIG_Q_ADAPT_PROBS ,  0
+.equ CONFIG_SUBFRAME_PROB_UPDATE ,  0
+.equ CONFIG_BITSTREAM_DEBUG ,  0
+.equ CONFIG_ALT_INTRA ,  1
+.equ CONFIG_PALETTE ,  1
+.equ CONFIG_PALETTE_DELTA_ENCODING ,  0
+.equ CONFIG_DAALA_EC ,  1
+.equ CONFIG_RAWBITS ,  0
+.equ CONFIG_EC_SMALLMUL ,  0
+.equ CONFIG_PVQ ,  0
+.equ CONFIG_CFL ,  0
+.equ CONFIG_XIPHRC ,  0
+.equ CONFIG_CB4X4 ,  1
+.equ CONFIG_CHROMA_2X2 ,  0
+.equ CONFIG_CHROMA_SUB8X8 ,  0
+.equ CONFIG_FRAME_SIZE ,  0
+.equ CONFIG_DELTA_Q ,  1
+.equ CONFIG_EXT_DELTA_Q ,  1
+.equ CONFIG_ADAPT_SCAN ,  0
+.equ CONFIG_FILTER_7BIT ,  1
+.equ CONFIG_PARALLEL_DEBLOCKING ,  0
+.equ CONFIG_PARALLEL_DEBLOCKING_15TAP ,  0
+.equ CONFIG_LOOPFILTERING_ACROSS_TILES ,  0
+.equ CONFIG_TILE_GROUPS ,  1
+.equ CONFIG_EC_ADAPT ,  1
+.equ CONFIG_TEMPMV_SIGNALING ,  0
+.equ CONFIG_RD_DEBUG ,  0
+.equ CONFIG_REFERENCE_BUFFER ,  1
+.equ CONFIG_COEF_INTERLEAVE ,  0
+.equ CONFIG_ENTROPY_STATS ,  0
+.equ CONFIG_MASKED_TX ,  0
+.equ CONFIG_DEPENDENT_HORZTILES ,  0
+.equ CONFIG_DAALA_DIST ,  0
+.equ CONFIG_TRIPRED ,  0
+.equ CONFIG_PALETTE_THROUGHPUT ,  1
+.equ CONFIG_REF_ADAPT ,  0
+.equ CONFIG_LV_MAP ,  0
+.equ CONFIG_TXK_SEL ,  0
+.equ CONFIG_MV_COMPRESS ,  1
+.equ CONFIG_FRAME_SUPERRES ,  0
+.equ CONFIG_NEW_MULTISYMBOL ,  0
+.equ CONFIG_COMPOUND_SINGLEREF ,  0
+.equ CONFIG_AOM_QM ,  0
+.equ CONFIG_LOWDELAY_COMPOUND ,  0
+.equ CONFIG_ANALYZER ,  0
+	.section	.note.GNU-stack,"",%progbits
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/aom_config.c
@@ -0,0 +1,11 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+#include "aom/aom_codec.h"
+static const char* const cfg = "--target=generic-gnu --enable-external-build --disable-examples --disable-docs --disable-unit-tests --size-limit=8192x4608 --enable-pic";
+const char *aom_codec_build_config(void) {return cfg;}
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/aom_config.h
@@ -0,0 +1,167 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef AOM_CONFIG_H
+#define AOM_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_AOM_PORTS 1
+#define HAVE_FEXCEPT 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_WXWIDGETS 0
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 1
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 0
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_HIGHBITDEPTH 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_CDEF 1
+#define CONFIG_VAR_TX 0
+#define CONFIG_RECT_TX 1
+#define CONFIG_REF_MV 1
+#define CONFIG_TPL_MV 0
+#define CONFIG_DUAL_FILTER 1
+#define CONFIG_CONVOLVE_ROUND 0
+#define CONFIG_COMPOUND_ROUND 0
+#define CONFIG_EXT_TX 0
+#define CONFIG_TX64X64 0
+#define CONFIG_SUB8X8_MC 0
+#define CONFIG_EXT_INTRA 1
+#define CONFIG_INTRA_INTERP 0
+#define CONFIG_FILTER_INTRA 0
+#define CONFIG_INTRABC 0
+#define CONFIG_EXT_INTER 0
+#define CONFIG_INTERINTRA 0
+#define CONFIG_WEDGE 0
+#define CONFIG_COMPOUND_SEGMENT 0
+#define CONFIG_EXT_REFS 1
+#define CONFIG_GLOBAL_MOTION 1
+#define CONFIG_NEW_QUANT 0
+#define CONFIG_SUPERTX 0
+#define CONFIG_ANS 0
+#define CONFIG_EC_MULTISYMBOL 1
+#define CONFIG_NEW_TOKENSET 1
+#define CONFIG_LOOP_RESTORATION 0
+#define CONFIG_EXT_PARTITION 0
+#define CONFIG_EXT_PARTITION_TYPES 0
+#define CONFIG_UNPOISON_PARTITION_CTX 0
+#define CONFIG_EXT_TILE 0
+#define CONFIG_MOTION_VAR 1
+#define CONFIG_NCOBMC 0
+#define CONFIG_WARPED_MOTION 1
+#define CONFIG_Q_ADAPT_PROBS 0
+#define CONFIG_SUBFRAME_PROB_UPDATE 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_ALT_INTRA 1
+#define CONFIG_PALETTE 1
+#define CONFIG_PALETTE_DELTA_ENCODING 0
+#define CONFIG_DAALA_EC 1
+#define CONFIG_RAWBITS 0
+#define CONFIG_EC_SMALLMUL 0
+#define CONFIG_PVQ 0
+#define CONFIG_CFL 0
+#define CONFIG_XIPHRC 0
+#define CONFIG_CB4X4 1
+#define CONFIG_CHROMA_2X2 0
+#define CONFIG_CHROMA_SUB8X8 0
+#define CONFIG_FRAME_SIZE 0
+#define CONFIG_DELTA_Q 1
+#define CONFIG_EXT_DELTA_Q 1
+#define CONFIG_ADAPT_SCAN 0
+#define CONFIG_FILTER_7BIT 1
+#define CONFIG_PARALLEL_DEBLOCKING 0
+#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+#define CONFIG_TILE_GROUPS 1
+#define CONFIG_EC_ADAPT 1
+#define CONFIG_TEMPMV_SIGNALING 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REFERENCE_BUFFER 1
+#define CONFIG_COEF_INTERLEAVE 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_MASKED_TX 0
+#define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DAALA_DIST 0
+#define CONFIG_TRIPRED 0
+#define CONFIG_PALETTE_THROUGHPUT 1
+#define CONFIG_REF_ADAPT 0
+#define CONFIG_LV_MAP 0
+#define CONFIG_TXK_SEL 0
+#define CONFIG_MV_COMPRESS 1
+#define CONFIG_FRAME_SUPERRES 0
+#define CONFIG_NEW_MULTISYMBOL 0
+#define CONFIG_COMPOUND_SINGLEREF 0
+#define CONFIG_AOM_QM 0
+#define CONFIG_LOWDELAY_COMPOUND 0
+#define CONFIG_ANALYZER 0
+#define DECODE_WIDTH_LIMIT 8192
+#define DECODE_HEIGHT_LIMIT 4608
+#endif /* AOM_CONFIG_H */
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/aom_dsp_rtcd.h
@@ -0,0 +1,901 @@
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_c
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_c
+
+void aom_blend_a64_hmask_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 h, int w);
+#define aom_blend_a64_hmask aom_blend_a64_hmask_c
+
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+#define aom_blend_a64_mask aom_blend_a64_mask_c
+
+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 h, int w);
+#define aom_blend_a64_vmask aom_blend_a64_vmask_c
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_pred aom_comp_avg_pred_c
+
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_c
+
+void aom_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);
+#define aom_convolve8 aom_convolve8_c
+
+void aom_convolve8_avg_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_avg aom_convolve8_avg_c
+
+void aom_convolve8_avg_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_avg_horiz aom_convolve8_avg_horiz_c
+
+void aom_convolve8_avg_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_avg_vert aom_convolve8_avg_vert_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_avg_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_convolve_avg aom_convolve_avg_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 x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve_copy aom_convolve_copy_c
+
+void aom_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
+
+void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
+
+void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
+
+void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
+
+void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
+
+void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
+
+void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
+
+void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
+
+void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x4 aom_d135_predictor_4x4_c
+
+void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
+
+void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x16 aom_d153_predictor_16x16_c
+
+void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
+
+void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_32x32 aom_d153_predictor_32x32_c
+
+void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_4x4 aom_d153_predictor_4x4_c
+
+void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x8 aom_d153_predictor_8x8_c
+
+void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
+
+void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
+
+void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
+
+void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
+
+void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
+
+void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
+
+void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
+
+void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
+
+void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
+
+void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
+
+void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
+
+void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
+
+void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
+
+void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_4x4 aom_d63e_predictor_4x4_c
+
+void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
+
+void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x32 aom_dc_128_predictor_32x32_c
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_c
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_c
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x32 aom_dc_left_predictor_32x32_c
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_c
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_c
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x32 aom_dc_predictor_32x32_c
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_c
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_c
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x32 aom_dc_top_predictor_32x32_c
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_c
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_c
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct16x16 aom_fdct16x16_c
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct16x16_1 aom_fdct16x16_1_c
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32 aom_fdct32x32_c
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32_1 aom_fdct32x32_1_c
+
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32_rd aom_fdct32x32_rd_c
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_c
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4_1 aom_fdct4x4_1_c
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct8x8 aom_fdct8x8_c
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct8x8_1 aom_fdct8x8_1_c
+
+void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_get16x16var aom_get16x16var_c
+
+unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+#define aom_get4x4sse_cs aom_get4x4sse_cs_c
+
+void aom_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_get8x8var aom_get8x8var_c
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_c
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x32 aom_h_predictor_32x32_c
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_c
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_c
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define aom_hadamard_16x16 aom_hadamard_16x16_c
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define aom_hadamard_8x8 aom_hadamard_8x8_c
+
+void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_10_add aom_idct16x16_10_add_c
+
+void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_1_add aom_idct16x16_1_add_c
+
+void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_256_add aom_idct16x16_256_add_c
+
+void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_38_add aom_idct16x16_38_add_c
+
+void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct32x32_1024_add aom_idct32x32_1024_add_c
+
+void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct32x32_135_add aom_idct32x32_135_add_c
+
+void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct32x32_1_add aom_idct32x32_1_add_c
+
+void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct32x32_34_add aom_idct32x32_34_add_c
+
+void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct4x4_16_add aom_idct4x4_16_add_c
+
+void aom_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct4x4_1_add aom_idct4x4_1_add_c
+
+void aom_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct8x8_12_add aom_idct8x8_12_add_c
+
+void aom_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct8x8_1_add aom_idct8x8_1_add_c
+
+void aom_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct8x8_64_add aom_idct8x8_64_add_c
+
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width);
+#define aom_int_pro_col aom_int_pro_col_c
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+#define aom_int_pro_row aom_int_pro_row_c
+
+void aom_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_16_add aom_iwht4x4_16_add_c
+
+void aom_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_1_add aom_iwht4x4_1_add_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_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_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_edge_16 aom_lpf_horizontal_edge_16_c
+
+void aom_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_edge_8 aom_lpf_horizontal_edge_8_c
+
+void aom_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_16 aom_lpf_vertical_16_c
+
+void aom_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_16_dual aom_lpf_vertical_16_dual_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_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_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_c
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x16 aom_mse16x16_c
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_c
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_c
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_c
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x16 aom_obmc_sad16x16_c
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x32 aom_obmc_sad16x32_c
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x8 aom_obmc_sad16x8_c
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x16 aom_obmc_sad32x16_c
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x32 aom_obmc_sad32x32_c
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x64 aom_obmc_sad32x64_c
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x4 aom_obmc_sad4x4_c
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x8 aom_obmc_sad4x8_c
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x32 aom_obmc_sad64x32_c
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x64 aom_obmc_sad64x64_c
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x16 aom_obmc_sad8x16_c
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x4 aom_obmc_sad8x4_c
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x8 aom_obmc_sad8x8_c
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x16 aom_obmc_variance16x16_c
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x32 aom_obmc_variance16x32_c
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x8 aom_obmc_variance16x8_c
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x16 aom_obmc_variance32x16_c
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x32 aom_obmc_variance32x32_c
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x64 aom_obmc_variance32x64_c
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x4 aom_obmc_variance4x4_c
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x8 aom_obmc_variance4x8_c
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x32 aom_obmc_variance64x32_c
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x64 aom_obmc_variance64x64_c
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x16 aom_obmc_variance8x16_c
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x4 aom_obmc_variance8x4_c
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x8 aom_obmc_variance8x8_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
+
+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_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_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_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_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b aom_quantize_b_c
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32 aom_quantize_b_32x32_c
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64 aom_quantize_b_64x64_c
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_c
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_c
+
+void aom_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x3 aom_sad16x16x3_c
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x4d aom_sad16x16x4d_c
+
+void aom_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x8 aom_sad16x16x8_c
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_c
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_c
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x32x4d aom_sad16x32x4d_c
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_c
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_c
+
+void aom_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x3 aom_sad16x8x3_c
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x4d aom_sad16x8x4d_c
+
+void aom_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x8 aom_sad16x8x8_c
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x16 aom_sad32x16_c
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x16_avg aom_sad32x16_avg_c
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x16x4d aom_sad32x16x4d_c
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x32 aom_sad32x32_c
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x32_avg aom_sad32x32_avg_c
+
+void aom_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x3 aom_sad32x32x3_c
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x4d aom_sad32x32x4d_c
+
+void aom_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x8 aom_sad32x32x8_c
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x64 aom_sad32x64_c
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x64_avg aom_sad32x64_avg_c
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x64x4d aom_sad32x64x4d_c
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_c
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_c
+
+void aom_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x3 aom_sad4x4x3_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x4d aom_sad4x4x4d_c
+
+void aom_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x8 aom_sad4x4x8_c
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_c
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x4d aom_sad4x8x4d_c
+
+void aom_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x8 aom_sad4x8x8_c
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x32 aom_sad64x32_c
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x32_avg aom_sad64x32_avg_c
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad64x32x4d aom_sad64x32x4d_c
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x64 aom_sad64x64_c
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x64_avg aom_sad64x64_avg_c
+
+void aom_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x3 aom_sad64x64x3_c
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x4d aom_sad64x64x4d_c
+
+void aom_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x8 aom_sad64x64x8_c
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_c
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_c
+
+void aom_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x3 aom_sad8x16x3_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x4d aom_sad8x16x4d_c
+
+void aom_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x8 aom_sad8x16x8_c
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_c
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x4d aom_sad8x4x4d_c
+
+void aom_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x8 aom_sad8x4x8_c
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_c
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_c
+
+void aom_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x3 aom_sad8x8x3_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x4d aom_sad8x8x4d_c
+
+void aom_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x8 aom_sad8x8x8_c
+
+int aom_satd_c(const int16_t *coeff, int length);
+#define aom_satd aom_satd_c
+
+void aom_scaled_2d_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_scaled_2d aom_scaled_2d_c
+
+void aom_scaled_avg_2d_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_scaled_avg_2d aom_scaled_avg_2d_c
+
+void aom_scaled_avg_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_scaled_avg_horiz aom_scaled_avg_horiz_c
+
+void aom_scaled_avg_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_scaled_avg_vert aom_scaled_avg_vert_c
+
+void aom_scaled_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_scaled_horiz aom_scaled_horiz_c
+
+void aom_scaled_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_scaled_vert aom_scaled_vert_c
+
+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_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_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_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_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
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x16 aom_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x32 aom_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x8 aom_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x16 aom_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x32 aom_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x64 aom_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x4 aom_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x8 aom_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x32 aom_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x64 aom_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x16 aom_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x4 aom_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x8 aom_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x16 aom_sub_pixel_variance16x16_c
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x32 aom_sub_pixel_variance16x32_c
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x8 aom_sub_pixel_variance16x8_c
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x16 aom_sub_pixel_variance32x16_c
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x32 aom_sub_pixel_variance32x32_c
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x64 aom_sub_pixel_variance32x64_c
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x4 aom_sub_pixel_variance4x4_c
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x8 aom_sub_pixel_variance4x8_c
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x32 aom_sub_pixel_variance64x32_c
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x64 aom_sub_pixel_variance64x64_c
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x16 aom_sub_pixel_variance8x16_c
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x4 aom_sub_pixel_variance8x4_c
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x8 aom_sub_pixel_variance8x8_c
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define aom_subtract_block aom_subtract_block_c
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+#define aom_sum_squares_2d_i16 aom_sum_squares_2d_i16_c
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_c
+
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_upsampled_pred aom_upsampled_pred_c
+
+void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x16 aom_v_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x32 aom_v_predictor_32x32_c
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x4 aom_v_predictor_4x4_c
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x8 aom_v_predictor_8x8_c
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x16 aom_variance16x16_c
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x32 aom_variance16x32_c
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x8 aom_variance16x8_c
+
+unsigned int aom_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x2 aom_variance2x2_c
+
+unsigned int aom_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x4 aom_variance2x4_c
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x16 aom_variance32x16_c
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x32 aom_variance32x32_c
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x64 aom_variance32x64_c
+
+unsigned int aom_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x2 aom_variance4x2_c
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_c
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_c
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x32 aom_variance64x32_c
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance64x64 aom_variance64x64_c
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x16 aom_variance8x16_c
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_c
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x8 aom_variance8x8_c
+
+uint32_t aom_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_h aom_variance_halfpixvar16x16_h_c
+
+uint32_t aom_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_hv aom_variance_halfpixvar16x16_hv_c
+
+uint32_t aom_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_v aom_variance_halfpixvar16x16_v_c
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+#define aom_vector_var aom_vector_var_c
+
+void aom_dsp_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/aom_scale_rtcd.h
@@ -0,0 +1,75 @@
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_scale_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/generic/av1_rtcd.h
@@ -0,0 +1,223 @@
+#ifndef AOM_RTCD_H_
+#define AOM_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct aom_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+typedef uint16_t od_dering_in;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_clpf_block_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+#define aom_clpf_block aom_clpf_block_c
+
+void aom_clpf_block_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+#define aom_clpf_block_hbd aom_clpf_block_hbd_c
+
+void aom_clpf_hblock_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+#define aom_clpf_hblock aom_clpf_hblock_c
+
+void aom_clpf_hblock_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+#define aom_clpf_hblock_hbd aom_clpf_hblock_hbd_c
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define av1_block_error av1_block_error_c
+
+int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define av1_block_error_fp av1_block_error_fp_c
+
+void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+#define av1_convolve_horiz av1_convolve_horiz_c
+
+void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+#define av1_convolve_vert av1_convolve_vert_c
+
+int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_diamond_search_sad av1_diamond_search_sad_c
+
+void av1_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_fdct8x8_quant av1_fdct8x8_quant_c
+
+void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x16 av1_fht16x16_c
+
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x32 av1_fht16x32_c
+
+void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x4 av1_fht16x4_c
+
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x8 av1_fht16x8_c
+
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x16 av1_fht32x16_c
+
+void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x32 av1_fht32x32_c
+
+void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x8 av1_fht32x8_c
+
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x16 av1_fht4x16_c
+
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x4 av1_fht4x4_c
+
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x8 av1_fht4x8_c
+
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x16 av1_fht8x16_c
+
+void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x32 av1_fht8x32_c
+
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x4 av1_fht8x4_c
+
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x8 av1_fht8x8_c
+
+int av1_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_full_range_search av1_full_range_search_c
+
+int av1_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define av1_full_search_sad av1_full_search_sad_c
+
+void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
+#define av1_fwd_idtx av1_fwd_idtx_c
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define av1_fwht4x4 av1_fwht4x4_c
+
+void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define av1_iht16x16_256_add av1_iht16x16_256_add_c
+
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x32_512_add av1_iht16x32_512_add_c
+
+void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x4_64_add av1_iht16x4_64_add_c
+
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x8_128_add av1_iht16x8_128_add_c
+
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht32x16_512_add av1_iht32x16_512_add_c
+
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define av1_iht32x32_1024_add av1_iht32x32_1024_add_c
+
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht32x8_256_add av1_iht32x8_256_add_c
+
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x16_64_add av1_iht4x16_64_add_c
+
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x4_16_add av1_iht4x4_16_add_c
+
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x8_32_add av1_iht4x8_32_add_c
+
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x16_128_add av1_iht8x16_128_add_c
+
+void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x32_256_add av1_iht8x32_256_add_c
+
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x4_32_add av1_iht8x4_32_add_c
+
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x8_64_add av1_iht8x8_64_add_c
+
+void av1_lowbd_convolve_init_c(void);
+#define av1_lowbd_convolve_init av1_lowbd_convolve_init_c
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp av1_quantize_fp_c
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
+
+void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define av1_temporal_filter_apply av1_temporal_filter_apply_c
+
+void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+#define av1_warp_affine av1_warp_affine_c
+
+void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+#define copy_4x4_16bit_to_16bit copy_4x4_16bit_to_16bit_c
+
+void copy_4x4_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+#define copy_4x4_16bit_to_8bit copy_4x4_16bit_to_8bit_c
+
+void copy_8x8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+#define copy_8x8_16bit_to_16bit copy_8x8_16bit_to_16bit_c
+
+void copy_8x8_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+#define copy_8x8_16bit_to_8bit copy_8x8_16bit_to_8bit_c
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+#define copy_rect8_16bit_to_16bit copy_rect8_16bit_to_16bit_c
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+#define copy_rect8_8bit_to_16bit copy_rect8_8bit_to_16bit_c
+
+int od_dir_find8_c(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+#define od_dir_find8 od_dir_find8_c
+
+void od_filter_dering_direction_4x4_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+#define od_filter_dering_direction_4x4 od_filter_dering_direction_4x4_c
+
+void od_filter_dering_direction_8x8_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+#define od_filter_dering_direction_8x8 od_filter_dering_direction_8x8_c
+
+void aom_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/aom_config.asm
@@ -0,0 +1,155 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  1
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  1
+.equ HAVE_NEON ,  1
+.equ HAVE_NEON_ASM ,  1
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_AOM_PORTS ,  1
+.equ HAVE_FEXCEPT ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_WXWIDGETS ,  0
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  1
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  1
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_AV1_ENCODER ,  1
+.equ CONFIG_AV1_DECODER ,  1
+.equ CONFIG_AV1 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  1
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  0
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_ACCOUNTING ,  0
+.equ CONFIG_INSPECTION ,  0
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_LOWBITDEPTH ,  1
+.equ CONFIG_HIGHBITDEPTH ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  1
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_CDEF ,  1
+.equ CONFIG_VAR_TX ,  0
+.equ CONFIG_RECT_TX ,  1
+.equ CONFIG_REF_MV ,  1
+.equ CONFIG_TPL_MV ,  0
+.equ CONFIG_DUAL_FILTER ,  1
+.equ CONFIG_CONVOLVE_ROUND ,  0
+.equ CONFIG_COMPOUND_ROUND ,  0
+.equ CONFIG_EXT_TX ,  0
+.equ CONFIG_TX64X64 ,  0
+.equ CONFIG_SUB8X8_MC ,  0
+.equ CONFIG_EXT_INTRA ,  1
+.equ CONFIG_INTRA_INTERP ,  0
+.equ CONFIG_FILTER_INTRA ,  0
+.equ CONFIG_INTRABC ,  0
+.equ CONFIG_EXT_INTER ,  0
+.equ CONFIG_INTERINTRA ,  0
+.equ CONFIG_WEDGE ,  0
+.equ CONFIG_COMPOUND_SEGMENT ,  0
+.equ CONFIG_EXT_REFS ,  1
+.equ CONFIG_GLOBAL_MOTION ,  1
+.equ CONFIG_NEW_QUANT ,  0
+.equ CONFIG_SUPERTX ,  0
+.equ CONFIG_ANS ,  0
+.equ CONFIG_EC_MULTISYMBOL ,  1
+.equ CONFIG_NEW_TOKENSET ,  1
+.equ CONFIG_LOOP_RESTORATION ,  0
+.equ CONFIG_EXT_PARTITION ,  0
+.equ CONFIG_EXT_PARTITION_TYPES ,  0
+.equ CONFIG_UNPOISON_PARTITION_CTX ,  0
+.equ CONFIG_EXT_TILE ,  0
+.equ CONFIG_MOTION_VAR ,  1
+.equ CONFIG_NCOBMC ,  0
+.equ CONFIG_WARPED_MOTION ,  1
+.equ CONFIG_Q_ADAPT_PROBS ,  0
+.equ CONFIG_SUBFRAME_PROB_UPDATE ,  0
+.equ CONFIG_BITSTREAM_DEBUG ,  0
+.equ CONFIG_ALT_INTRA ,  1
+.equ CONFIG_PALETTE ,  1
+.equ CONFIG_PALETTE_DELTA_ENCODING ,  0
+.equ CONFIG_DAALA_EC ,  1
+.equ CONFIG_RAWBITS ,  0
+.equ CONFIG_EC_SMALLMUL ,  0
+.equ CONFIG_PVQ ,  0
+.equ CONFIG_CFL ,  0
+.equ CONFIG_XIPHRC ,  0
+.equ CONFIG_CB4X4 ,  1
+.equ CONFIG_CHROMA_2X2 ,  0
+.equ CONFIG_CHROMA_SUB8X8 ,  0
+.equ CONFIG_FRAME_SIZE ,  0
+.equ CONFIG_DELTA_Q ,  1
+.equ CONFIG_EXT_DELTA_Q ,  1
+.equ CONFIG_ADAPT_SCAN ,  0
+.equ CONFIG_FILTER_7BIT ,  1
+.equ CONFIG_PARALLEL_DEBLOCKING ,  0
+.equ CONFIG_PARALLEL_DEBLOCKING_15TAP ,  0
+.equ CONFIG_LOOPFILTERING_ACROSS_TILES ,  0
+.equ CONFIG_TILE_GROUPS ,  1
+.equ CONFIG_EC_ADAPT ,  1
+.equ CONFIG_TEMPMV_SIGNALING ,  0
+.equ CONFIG_RD_DEBUG ,  0
+.equ CONFIG_REFERENCE_BUFFER ,  1
+.equ CONFIG_COEF_INTERLEAVE ,  0
+.equ CONFIG_ENTROPY_STATS ,  0
+.equ CONFIG_MASKED_TX ,  0
+.equ CONFIG_DEPENDENT_HORZTILES ,  0
+.equ CONFIG_DAALA_DIST ,  0
+.equ CONFIG_TRIPRED ,  0
+.equ CONFIG_PALETTE_THROUGHPUT ,  1
+.equ CONFIG_REF_ADAPT ,  0
+.equ CONFIG_LV_MAP ,  0
+.equ CONFIG_TXK_SEL ,  0
+.equ CONFIG_MV_COMPRESS ,  1
+.equ CONFIG_FRAME_SUPERRES ,  0
+.equ CONFIG_NEW_MULTISYMBOL ,  0
+.equ CONFIG_COMPOUND_SINGLEREF ,  0
+.equ CONFIG_AOM_QM ,  0
+.equ CONFIG_LOWDELAY_COMPOUND ,  0
+.equ CONFIG_ANALYZER ,  0
+	.section	.note.GNU-stack,"",%progbits
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/aom_config.c
@@ -0,0 +1,11 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+#include "aom/aom_codec.h"
+static const char* const cfg = "--target=armv7-linux-gcc --enable-external-build --disable-examples --disable-docs --disable-unit-tests --size-limit=8192x4608 --enable-pic --enable-runtime-cpu-detect --enable-realtime-only";
+const char *aom_codec_build_config(void) {return cfg;}
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/aom_config.h
@@ -0,0 +1,167 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef AOM_CONFIG_H
+#define AOM_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 1
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 1
+#define HAVE_NEON 1
+#define HAVE_NEON_ASM 1
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_AOM_PORTS 1
+#define HAVE_FEXCEPT 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_WXWIDGETS 0
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 1
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 0
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_HIGHBITDEPTH 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_CDEF 1
+#define CONFIG_VAR_TX 0
+#define CONFIG_RECT_TX 1
+#define CONFIG_REF_MV 1
+#define CONFIG_TPL_MV 0
+#define CONFIG_DUAL_FILTER 1
+#define CONFIG_CONVOLVE_ROUND 0
+#define CONFIG_COMPOUND_ROUND 0
+#define CONFIG_EXT_TX 0
+#define CONFIG_TX64X64 0
+#define CONFIG_SUB8X8_MC 0
+#define CONFIG_EXT_INTRA 1
+#define CONFIG_INTRA_INTERP 0
+#define CONFIG_FILTER_INTRA 0
+#define CONFIG_INTRABC 0
+#define CONFIG_EXT_INTER 0
+#define CONFIG_INTERINTRA 0
+#define CONFIG_WEDGE 0
+#define CONFIG_COMPOUND_SEGMENT 0
+#define CONFIG_EXT_REFS 1
+#define CONFIG_GLOBAL_MOTION 1
+#define CONFIG_NEW_QUANT 0
+#define CONFIG_SUPERTX 0
+#define CONFIG_ANS 0
+#define CONFIG_EC_MULTISYMBOL 1
+#define CONFIG_NEW_TOKENSET 1
+#define CONFIG_LOOP_RESTORATION 0
+#define CONFIG_EXT_PARTITION 0
+#define CONFIG_EXT_PARTITION_TYPES 0
+#define CONFIG_UNPOISON_PARTITION_CTX 0
+#define CONFIG_EXT_TILE 0
+#define CONFIG_MOTION_VAR 1
+#define CONFIG_NCOBMC 0
+#define CONFIG_WARPED_MOTION 1
+#define CONFIG_Q_ADAPT_PROBS 0
+#define CONFIG_SUBFRAME_PROB_UPDATE 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_ALT_INTRA 1
+#define CONFIG_PALETTE 1
+#define CONFIG_PALETTE_DELTA_ENCODING 0
+#define CONFIG_DAALA_EC 1
+#define CONFIG_RAWBITS 0
+#define CONFIG_EC_SMALLMUL 0
+#define CONFIG_PVQ 0
+#define CONFIG_CFL 0
+#define CONFIG_XIPHRC 0
+#define CONFIG_CB4X4 1
+#define CONFIG_CHROMA_2X2 0
+#define CONFIG_CHROMA_SUB8X8 0
+#define CONFIG_FRAME_SIZE 0
+#define CONFIG_DELTA_Q 1
+#define CONFIG_EXT_DELTA_Q 1
+#define CONFIG_ADAPT_SCAN 0
+#define CONFIG_FILTER_7BIT 1
+#define CONFIG_PARALLEL_DEBLOCKING 0
+#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+#define CONFIG_TILE_GROUPS 1
+#define CONFIG_EC_ADAPT 1
+#define CONFIG_TEMPMV_SIGNALING 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REFERENCE_BUFFER 1
+#define CONFIG_COEF_INTERLEAVE 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_MASKED_TX 0
+#define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DAALA_DIST 0
+#define CONFIG_TRIPRED 0
+#define CONFIG_PALETTE_THROUGHPUT 1
+#define CONFIG_REF_ADAPT 0
+#define CONFIG_LV_MAP 0
+#define CONFIG_TXK_SEL 0
+#define CONFIG_MV_COMPRESS 1
+#define CONFIG_FRAME_SUPERRES 0
+#define CONFIG_NEW_MULTISYMBOL 0
+#define CONFIG_COMPOUND_SINGLEREF 0
+#define CONFIG_AOM_QM 0
+#define CONFIG_LOWDELAY_COMPOUND 0
+#define CONFIG_ANALYZER 0
+#define DECODE_WIDTH_LIMIT 8192
+#define DECODE_HEIGHT_LIMIT 4608
+#endif /* AOM_CONFIG_H */
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/aom_dsp_rtcd.h
@@ -0,0 +1,1200 @@
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_neon(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_8x8)(const uint8_t *, int p);
+
+void aom_blend_a64_hmask_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 h, int w);
+#define aom_blend_a64_hmask aom_blend_a64_hmask_c
+
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+#define aom_blend_a64_mask aom_blend_a64_mask_c
+
+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 h, int w);
+#define aom_blend_a64_vmask aom_blend_a64_vmask_c
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_pred aom_comp_avg_pred_c
+
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_c
+
+void aom_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);
+void aom_convolve8_neon(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)(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_avg_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_avg_neon(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_avg)(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_avg_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_avg_horiz_neon(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_avg_horiz)(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_avg_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);
+void aom_convolve8_avg_vert_neon(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_avg_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_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_neon(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_horiz)(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_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);
+void aom_convolve8_vert_neon(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_avg_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_convolve_avg_neon(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_convolve_avg)(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 x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve_copy_neon(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_convolve_copy)(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_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
+
+void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
+
+void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
+
+void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
+
+void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
+
+void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
+
+void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
+
+void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
+
+void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d135_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
+
+void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x16 aom_d153_predictor_16x16_c
+
+void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
+
+void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_32x32 aom_d153_predictor_32x32_c
+
+void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_4x4 aom_d153_predictor_4x4_c
+
+void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x8 aom_d153_predictor_8x8_c
+
+void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
+
+void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
+
+void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
+
+void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
+
+void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
+
+void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
+
+void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
+
+void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
+
+void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
+
+void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
+
+void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
+
+void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
+
+void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
+
+void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_4x4 aom_d63e_predictor_4x4_c
+
+void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
+
+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);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct16x16 aom_fdct16x16_c
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct16x16_1 aom_fdct16x16_1_c
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32 aom_fdct32x32_c
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32_1 aom_fdct32x32_1_c
+
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct32x32_rd aom_fdct32x32_rd_c
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_c
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4_1 aom_fdct4x4_1_c
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_neon(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct8x8)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct8x8_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*aom_get16x16var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+unsigned int aom_get4x4sse_cs_neon(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_get4x4sse_cs)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+
+void aom_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get8x8var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*aom_get8x8var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_c
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_16x16_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, int src_stride, int16_t *coeff);
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_8x8_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, int src_stride, int16_t *coeff);
+
+void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_38_add aom_idct16x16_38_add_c
+
+void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width);
+int16_t aom_int_pro_col_neon(const uint8_t *ref, int width);
+RTCD_EXTERN int16_t (*aom_int_pro_col)(const uint8_t *ref, int width);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+void aom_int_pro_row_neon(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+
+void aom_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_16_add aom_iwht4x4_16_add_c
+
+void aom_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_1_add aom_iwht4x4_1_add_c
+
+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);
+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_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);
+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_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_edge_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_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_edge_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+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);
+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_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);
+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_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_media(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_c
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_c
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_c
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x16 aom_obmc_sad16x16_c
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x32 aom_obmc_sad16x32_c
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x8 aom_obmc_sad16x8_c
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x16 aom_obmc_sad32x16_c
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x32 aom_obmc_sad32x32_c
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x64 aom_obmc_sad32x64_c
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x4 aom_obmc_sad4x4_c
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x8 aom_obmc_sad4x8_c
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x32 aom_obmc_sad64x32_c
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad64x64 aom_obmc_sad64x64_c
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x16 aom_obmc_sad8x16_c
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x4 aom_obmc_sad8x4_c
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x8 aom_obmc_sad8x8_c
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x16 aom_obmc_variance16x16_c
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x32 aom_obmc_variance16x32_c
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x8 aom_obmc_variance16x8_c
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x16 aom_obmc_variance32x16_c
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x32 aom_obmc_variance32x32_c
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x64 aom_obmc_variance32x64_c
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x4 aom_obmc_variance4x4_c
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x8 aom_obmc_variance4x8_c
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x32 aom_obmc_variance64x32_c
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance64x64 aom_obmc_variance64x64_c
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x16 aom_obmc_variance8x16_c
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x4 aom_obmc_variance8x4_c
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x8 aom_obmc_variance8x8_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
+
+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_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_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_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_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b aom_quantize_b_c
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32 aom_quantize_b_32x32_c
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64 aom_quantize_b_64x64_c
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_media(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_c
+
+void aom_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x3 aom_sad16x16x3_c
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x8 aom_sad16x16x8_c
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_c
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_c
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x32x4d aom_sad16x32x4d_c
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_c
+
+void aom_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x3 aom_sad16x8x3_c
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x4d aom_sad16x8x4d_c
+
+void aom_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x8 aom_sad16x8x8_c
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x16 aom_sad32x16_c
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x16_avg aom_sad32x16_avg_c
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x16x4d aom_sad32x16x4d_c
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x32_avg aom_sad32x32_avg_c
+
+void aom_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x3 aom_sad32x32x3_c
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x8 aom_sad32x32x8_c
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x64 aom_sad32x64_c
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x64_avg aom_sad32x64_avg_c
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x64x4d aom_sad32x64x4d_c
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_c
+
+void aom_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x3 aom_sad4x4x3_c
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x4d aom_sad4x4x4d_c
+
+void aom_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x8 aom_sad4x4x8_c
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_c
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_c
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x4d aom_sad4x8x4d_c
+
+void aom_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x8 aom_sad4x8x8_c
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad64x32 aom_sad64x32_c
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x32_avg aom_sad64x32_avg_c
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad64x32x4d aom_sad64x32x4d_c
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad64x64_avg aom_sad64x64_avg_c
+
+void aom_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x3 aom_sad64x64x3_c
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x8 aom_sad64x64x8_c
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_c
+
+void aom_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x3 aom_sad8x16x3_c
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x4d aom_sad8x16x4d_c
+
+void aom_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x8 aom_sad8x16x8_c
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_c
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_c
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x4d aom_sad8x4x4d_c
+
+void aom_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x8 aom_sad8x4x8_c
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_c
+
+void aom_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x3 aom_sad8x8x3_c
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x4d aom_sad8x8x4d_c
+
+void aom_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x8 aom_sad8x8x8_c
+
+int aom_satd_c(const int16_t *coeff, int length);
+int aom_satd_neon(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_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_scaled_2d aom_scaled_2d_c
+
+void aom_scaled_avg_2d_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_scaled_avg_2d aom_scaled_avg_2d_c
+
+void aom_scaled_avg_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_scaled_avg_horiz aom_scaled_avg_horiz_c
+
+void aom_scaled_avg_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_scaled_avg_vert aom_scaled_avg_vert_c
+
+void aom_scaled_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_scaled_horiz aom_scaled_horiz_c
+
+void aom_scaled_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_scaled_vert aom_scaled_vert_c
+
+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_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_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_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_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
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x16 aom_sub_pixel_avg_variance16x16_c
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x32 aom_sub_pixel_avg_variance16x32_c
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x8 aom_sub_pixel_avg_variance16x8_c
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x16 aom_sub_pixel_avg_variance32x16_c
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x32 aom_sub_pixel_avg_variance32x32_c
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x64 aom_sub_pixel_avg_variance32x64_c
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x4 aom_sub_pixel_avg_variance4x4_c
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x8 aom_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x32 aom_sub_pixel_avg_variance64x32_c
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance64x64 aom_sub_pixel_avg_variance64x64_c
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x16 aom_sub_pixel_avg_variance8x16_c
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x4 aom_sub_pixel_avg_variance8x4_c
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x8 aom_sub_pixel_avg_variance8x8_c
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x32 aom_sub_pixel_variance16x32_c
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x8 aom_sub_pixel_variance16x8_c
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x16 aom_sub_pixel_variance32x16_c
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x64 aom_sub_pixel_variance32x64_c
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x4 aom_sub_pixel_variance4x4_c
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x8 aom_sub_pixel_variance4x8_c
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance64x32 aom_sub_pixel_variance64x32_c
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x16 aom_sub_pixel_variance8x16_c
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x4 aom_sub_pixel_variance8x4_c
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+#define aom_sum_squares_2d_i16 aom_sum_squares_2d_i16_c
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+#define aom_sum_squares_i16 aom_sum_squares_i16_c
+
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_upsampled_pred aom_upsampled_pred_c
+
+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);
+RTCD_EXTERN void (*aom_v_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x32 aom_variance16x32_c
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x2 aom_variance2x2_c
+
+unsigned int aom_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x4 aom_variance2x4_c
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x16 aom_variance32x16_c
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x2 aom_variance4x2_c
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x4 aom_variance4x4_c
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x8 aom_variance4x8_c
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x4 aom_variance8x4_c
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+uint32_t aom_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_h_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_h aom_variance_halfpixvar16x16_h_media
+
+uint32_t aom_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_hv_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_hv aom_variance_halfpixvar16x16_hv_media
+
+uint32_t aom_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_v_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define aom_variance_halfpixvar16x16_v aom_variance_halfpixvar16x16_v_media
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_neon(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+void aom_dsp_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+    aom_avg_4x4 = aom_avg_4x4_c;
+    if (flags & HAS_NEON) aom_avg_4x4 = aom_avg_4x4_neon;
+    aom_avg_8x8 = aom_avg_8x8_c;
+    if (flags & HAS_NEON) aom_avg_8x8 = aom_avg_8x8_neon;
+    aom_convolve8 = aom_convolve8_c;
+    if (flags & HAS_NEON) aom_convolve8 = aom_convolve8_neon;
+    aom_convolve8_avg = aom_convolve8_avg_c;
+    if (flags & HAS_NEON) aom_convolve8_avg = aom_convolve8_avg_neon;
+    aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_c;
+    if (flags & HAS_NEON) aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_neon;
+    aom_convolve8_avg_vert = aom_convolve8_avg_vert_c;
+    if (flags & HAS_NEON) aom_convolve8_avg_vert = aom_convolve8_avg_vert_neon;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
+    if (flags & HAS_NEON) aom_convolve8_horiz = aom_convolve8_horiz_neon;
+    aom_convolve8_vert = aom_convolve8_vert_c;
+    if (flags & HAS_NEON) aom_convolve8_vert = aom_convolve8_vert_neon;
+    aom_convolve_avg = aom_convolve_avg_c;
+    if (flags & HAS_NEON) aom_convolve_avg = aom_convolve_avg_neon;
+    aom_convolve_copy = aom_convolve_copy_c;
+    if (flags & HAS_NEON) aom_convolve_copy = aom_convolve_copy_neon;
+    aom_d135_predictor_4x4 = aom_d135_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_d135_predictor_4x4 = aom_d135_predictor_4x4_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;
+    if (flags & HAS_NEON) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_neon;
+    aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_neon;
+    aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_neon;
+    aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_neon;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_neon;
+    aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_neon;
+    aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_neon;
+    aom_dc_predictor_16x16 = aom_dc_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_predictor_16x16 = aom_dc_predictor_16x16_neon;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_neon;
+    aom_dc_predictor_4x4 = aom_dc_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_predictor_4x4 = aom_dc_predictor_4x4_neon;
+    aom_dc_predictor_8x8 = aom_dc_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_predictor_8x8 = aom_dc_predictor_8x8_neon;
+    aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_neon;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_neon;
+    aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_neon;
+    aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_neon;
+    aom_fdct8x8 = aom_fdct8x8_c;
+    if (flags & HAS_NEON) aom_fdct8x8 = aom_fdct8x8_neon;
+    aom_fdct8x8_1 = aom_fdct8x8_1_c;
+    if (flags & HAS_NEON) aom_fdct8x8_1 = aom_fdct8x8_1_neon;
+    aom_get16x16var = aom_get16x16var_c;
+    if (flags & HAS_NEON) aom_get16x16var = aom_get16x16var_neon;
+    aom_get4x4sse_cs = aom_get4x4sse_cs_c;
+    if (flags & HAS_NEON) aom_get4x4sse_cs = aom_get4x4sse_cs_neon;
+    aom_get8x8var = aom_get8x8var_c;
+    if (flags & HAS_NEON) aom_get8x8var = aom_get8x8var_neon;
+    aom_h_predictor_16x16 = aom_h_predictor_16x16_c;
+    if (flags & HAS_NEON) aom_h_predictor_16x16 = aom_h_predictor_16x16_neon;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_c;
+    if (flags & HAS_NEON) aom_h_predictor_32x32 = aom_h_predictor_32x32_neon;
+    aom_h_predictor_4x4 = aom_h_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_h_predictor_4x4 = aom_h_predictor_4x4_neon;
+    aom_h_predictor_8x8 = aom_h_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_h_predictor_8x8 = aom_h_predictor_8x8_neon;
+    aom_hadamard_16x16 = aom_hadamard_16x16_c;
+    if (flags & HAS_NEON) aom_hadamard_16x16 = aom_hadamard_16x16_neon;
+    aom_hadamard_8x8 = aom_hadamard_8x8_c;
+    if (flags & HAS_NEON) aom_hadamard_8x8 = aom_hadamard_8x8_neon;
+    aom_idct16x16_10_add = aom_idct16x16_10_add_c;
+    if (flags & HAS_NEON) aom_idct16x16_10_add = aom_idct16x16_10_add_neon;
+    aom_idct16x16_1_add = aom_idct16x16_1_add_c;
+    if (flags & HAS_NEON) aom_idct16x16_1_add = aom_idct16x16_1_add_neon;
+    aom_idct16x16_256_add = aom_idct16x16_256_add_c;
+    if (flags & HAS_NEON) aom_idct16x16_256_add = aom_idct16x16_256_add_neon;
+    aom_idct32x32_1024_add = aom_idct32x32_1024_add_c;
+    if (flags & HAS_NEON) aom_idct32x32_1024_add = aom_idct32x32_1024_add_neon;
+    aom_idct32x32_135_add = aom_idct32x32_135_add_c;
+    if (flags & HAS_NEON) aom_idct32x32_135_add = aom_idct32x32_1024_add_neon;
+    aom_idct32x32_1_add = aom_idct32x32_1_add_c;
+    if (flags & HAS_NEON) aom_idct32x32_1_add = aom_idct32x32_1_add_neon;
+    aom_idct32x32_34_add = aom_idct32x32_34_add_c;
+    if (flags & HAS_NEON) aom_idct32x32_34_add = aom_idct32x32_1024_add_neon;
+    aom_idct4x4_16_add = aom_idct4x4_16_add_c;
+    if (flags & HAS_NEON) aom_idct4x4_16_add = aom_idct4x4_16_add_neon;
+    aom_idct4x4_1_add = aom_idct4x4_1_add_c;
+    if (flags & HAS_NEON) aom_idct4x4_1_add = aom_idct4x4_1_add_neon;
+    aom_idct8x8_12_add = aom_idct8x8_12_add_c;
+    if (flags & HAS_NEON) aom_idct8x8_12_add = aom_idct8x8_12_add_neon;
+    aom_idct8x8_1_add = aom_idct8x8_1_add_c;
+    if (flags & HAS_NEON) aom_idct8x8_1_add = aom_idct8x8_1_add_neon;
+    aom_idct8x8_64_add = aom_idct8x8_64_add_c;
+    if (flags & HAS_NEON) aom_idct8x8_64_add = aom_idct8x8_64_add_neon;
+    aom_int_pro_col = aom_int_pro_col_c;
+    if (flags & HAS_NEON) aom_int_pro_col = aom_int_pro_col_neon;
+    aom_int_pro_row = aom_int_pro_row_c;
+    if (flags & HAS_NEON) aom_int_pro_row = aom_int_pro_row_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_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_edge_16 = aom_lpf_horizontal_edge_16_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_neon;
+    aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_c;
+    if (flags & HAS_NEON) aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_neon;
+    aom_lpf_vertical_16 = aom_lpf_vertical_16_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_16 = aom_lpf_vertical_16_neon;
+    aom_lpf_vertical_16_dual = aom_lpf_vertical_16_dual_c;
+    if (flags & HAS_NEON) aom_lpf_vertical_16_dual = aom_lpf_vertical_16_dual_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_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_minmax_8x8 = aom_minmax_8x8_c;
+    if (flags & HAS_NEON) aom_minmax_8x8 = aom_minmax_8x8_neon;
+    aom_mse16x16 = aom_mse16x16_media;
+    if (flags & HAS_NEON) aom_mse16x16 = aom_mse16x16_neon;
+    aom_sad16x16 = aom_sad16x16_media;
+    if (flags & HAS_NEON) aom_sad16x16 = aom_sad16x16_neon;
+    aom_sad16x16x4d = aom_sad16x16x4d_c;
+    if (flags & HAS_NEON) aom_sad16x16x4d = aom_sad16x16x4d_neon;
+    aom_sad16x8 = aom_sad16x8_c;
+    if (flags & HAS_NEON) aom_sad16x8 = aom_sad16x8_neon;
+    aom_sad32x32 = aom_sad32x32_c;
+    if (flags & HAS_NEON) aom_sad32x32 = aom_sad32x32_neon;
+    aom_sad32x32x4d = aom_sad32x32x4d_c;
+    if (flags & HAS_NEON) aom_sad32x32x4d = aom_sad32x32x4d_neon;
+    aom_sad4x4 = aom_sad4x4_c;
+    if (flags & HAS_NEON) aom_sad4x4 = aom_sad4x4_neon;
+    aom_sad64x64 = aom_sad64x64_c;
+    if (flags & HAS_NEON) aom_sad64x64 = aom_sad64x64_neon;
+    aom_sad64x64x4d = aom_sad64x64x4d_c;
+    if (flags & HAS_NEON) aom_sad64x64x4d = aom_sad64x64x4d_neon;
+    aom_sad8x16 = aom_sad8x16_c;
+    if (flags & HAS_NEON) aom_sad8x16 = aom_sad8x16_neon;
+    aom_sad8x8 = aom_sad8x8_c;
+    if (flags & HAS_NEON) aom_sad8x8 = aom_sad8x8_neon;
+    aom_satd = aom_satd_c;
+    if (flags & HAS_NEON) aom_satd = aom_satd_neon;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_media;
+    if (flags & HAS_NEON) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_neon;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_neon;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
+    if (flags & HAS_NEON) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_neon;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_media;
+    if (flags & HAS_NEON) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_neon;
+    aom_subtract_block = aom_subtract_block_c;
+    if (flags & HAS_NEON) aom_subtract_block = aom_subtract_block_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;
+    if (flags & HAS_NEON) aom_v_predictor_32x32 = aom_v_predictor_32x32_neon;
+    aom_v_predictor_4x4 = aom_v_predictor_4x4_c;
+    if (flags & HAS_NEON) aom_v_predictor_4x4 = aom_v_predictor_4x4_neon;
+    aom_v_predictor_8x8 = aom_v_predictor_8x8_c;
+    if (flags & HAS_NEON) aom_v_predictor_8x8 = aom_v_predictor_8x8_neon;
+    aom_variance16x16 = aom_variance16x16_media;
+    if (flags & HAS_NEON) aom_variance16x16 = aom_variance16x16_neon;
+    aom_variance16x8 = aom_variance16x8_c;
+    if (flags & HAS_NEON) aom_variance16x8 = aom_variance16x8_neon;
+    aom_variance32x32 = aom_variance32x32_c;
+    if (flags & HAS_NEON) aom_variance32x32 = aom_variance32x32_neon;
+    aom_variance32x64 = aom_variance32x64_c;
+    if (flags & HAS_NEON) aom_variance32x64 = aom_variance32x64_neon;
+    aom_variance64x32 = aom_variance64x32_c;
+    if (flags & HAS_NEON) aom_variance64x32 = aom_variance64x32_neon;
+    aom_variance64x64 = aom_variance64x64_c;
+    if (flags & HAS_NEON) aom_variance64x64 = aom_variance64x64_neon;
+    aom_variance8x16 = aom_variance8x16_c;
+    if (flags & HAS_NEON) aom_variance8x16 = aom_variance8x16_neon;
+    aom_variance8x8 = aom_variance8x8_media;
+    if (flags & HAS_NEON) aom_variance8x8 = aom_variance8x8_neon;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_NEON) aom_vector_var = aom_vector_var_neon;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/aom_scale_rtcd.h
@@ -0,0 +1,80 @@
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_scale_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/arm/av1_rtcd.h
@@ -0,0 +1,282 @@
+#ifndef AOM_RTCD_H_
+#define AOM_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct aom_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+typedef uint16_t od_dering_in;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_clpf_block_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_neon(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_block)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_block_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_hbd_neon(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_block_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_hblock_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_neon(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_hblock)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_hblock_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_hbd_neon(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define av1_block_error av1_block_error_c
+
+int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t av1_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+RTCD_EXTERN int64_t (*av1_block_error_fp)(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+
+void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+#define av1_convolve_horiz av1_convolve_horiz_c
+
+void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+#define av1_convolve_vert av1_convolve_vert_c
+
+int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_diamond_search_sad av1_diamond_search_sad_c
+
+void av1_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_fdct8x8_quant_neon(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_fdct8x8_quant)(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x16 av1_fht16x16_c
+
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x32 av1_fht16x32_c
+
+void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x4 av1_fht16x4_c
+
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x8 av1_fht16x8_c
+
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x16 av1_fht32x16_c
+
+void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x32 av1_fht32x32_c
+
+void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x8 av1_fht32x8_c
+
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x16 av1_fht4x16_c
+
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x4 av1_fht4x4_c
+
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x8 av1_fht4x8_c
+
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x16 av1_fht8x16_c
+
+void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x32 av1_fht8x32_c
+
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x4 av1_fht8x4_c
+
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x8 av1_fht8x8_c
+
+int av1_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_full_range_search av1_full_range_search_c
+
+int av1_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define av1_full_search_sad av1_full_search_sad_c
+
+void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
+#define av1_fwd_idtx av1_fwd_idtx_c
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define av1_fwht4x4 av1_fwht4x4_c
+
+void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define av1_iht16x16_256_add av1_iht16x16_256_add_c
+
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x32_512_add av1_iht16x32_512_add_c
+
+void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x4_64_add av1_iht16x4_64_add_c
+
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x8_128_add av1_iht16x8_128_add_c
+
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht32x16_512_add av1_iht32x16_512_add_c
+
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define av1_iht32x32_1024_add av1_iht32x32_1024_add_c
+
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht32x8_256_add av1_iht32x8_256_add_c
+
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x16_64_add av1_iht4x16_64_add_c
+
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x8_32_add av1_iht4x8_32_add_c
+
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x16_128_add av1_iht8x16_128_add_c
+
+void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x32_256_add av1_iht8x32_256_add_c
+
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x4_32_add av1_iht8x4_32_add_c
+
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_lowbd_convolve_init_c(void);
+#define av1_lowbd_convolve_init av1_lowbd_convolve_init_c
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
+
+void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define av1_temporal_filter_apply av1_temporal_filter_apply_c
+
+void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+#define av1_warp_affine av1_warp_affine_c
+
+void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_16bit_neon(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_4x4_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_4x4_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_8bit_neon(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_4x4_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_8x8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_16bit_neon(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_8x8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_8x8_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_8bit_neon(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_8x8_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_neon(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_neon(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+
+int od_dir_find8_c(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+int od_dir_find8_neon(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*od_dir_find8)(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+
+void od_filter_dering_direction_4x4_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_4x4_neon(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+RTCD_EXTERN void (*od_filter_dering_direction_4x4)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+
+void od_filter_dering_direction_8x8_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_8x8_neon(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+RTCD_EXTERN void (*od_filter_dering_direction_8x8)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+
+void aom_rtcd(void);
+
+#include "aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+    aom_clpf_block = aom_clpf_block_c;
+    if (flags & HAS_NEON) aom_clpf_block = aom_clpf_block_neon;
+    aom_clpf_block_hbd = aom_clpf_block_hbd_c;
+    if (flags & HAS_NEON) aom_clpf_block_hbd = aom_clpf_block_hbd_neon;
+    aom_clpf_hblock = aom_clpf_hblock_c;
+    if (flags & HAS_NEON) aom_clpf_hblock = aom_clpf_hblock_neon;
+    aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_c;
+    if (flags & HAS_NEON) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_neon;
+    av1_block_error_fp = av1_block_error_fp_c;
+    if (flags & HAS_NEON) av1_block_error_fp = av1_block_error_fp_neon;
+    av1_fdct8x8_quant = av1_fdct8x8_quant_c;
+    if (flags & HAS_NEON) av1_fdct8x8_quant = av1_fdct8x8_quant_neon;
+    av1_iht4x4_16_add = av1_iht4x4_16_add_c;
+    if (flags & HAS_NEON) av1_iht4x4_16_add = av1_iht4x4_16_add_neon;
+    av1_iht8x8_64_add = av1_iht8x8_64_add_c;
+    if (flags & HAS_NEON) av1_iht8x8_64_add = av1_iht8x8_64_add_neon;
+    av1_quantize_fp = av1_quantize_fp_c;
+    if (flags & HAS_NEON) av1_quantize_fp = av1_quantize_fp_neon;
+    copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_c;
+    if (flags & HAS_NEON) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_neon;
+    copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_c;
+    if (flags & HAS_NEON) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_neon;
+    copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_c;
+    if (flags & HAS_NEON) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_neon;
+    copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_c;
+    if (flags & HAS_NEON) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_neon;
+    copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_NEON) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_neon;
+    copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_NEON) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_neon;
+    od_dir_find8 = od_dir_find8_c;
+    if (flags & HAS_NEON) od_dir_find8 = od_dir_find8_neon;
+    od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_c;
+    if (flags & HAS_NEON) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_neon;
+    od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_c;
+    if (flags & HAS_NEON) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_neon;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/aom_config.asm
@@ -0,0 +1,151 @@
+%define ARCH_ARM 0
+%define ARCH_MIPS 0
+%define ARCH_X86 1
+%define ARCH_X86_64 0
+%define HAVE_EDSP 0
+%define HAVE_MEDIA 0
+%define HAVE_NEON 0
+%define HAVE_NEON_ASM 0
+%define HAVE_MIPS32 0
+%define HAVE_DSPR2 0
+%define HAVE_MSA 0
+%define HAVE_MIPS64 0
+%define HAVE_MMX 1
+%define HAVE_SSE 1
+%define HAVE_SSE2 1
+%define HAVE_SSE3 1
+%define HAVE_SSSE3 1
+%define HAVE_SSE4_1 1
+%define HAVE_AVX 1
+%define HAVE_AVX2 1
+%define HAVE_AOM_PORTS 1
+%define HAVE_FEXCEPT 1
+%define HAVE_PTHREAD_H 1
+%define HAVE_WXWIDGETS 0
+%define CONFIG_DEPENDENCY_TRACKING 1
+%define CONFIG_EXTERNAL_BUILD 1
+%define CONFIG_INSTALL_DOCS 1
+%define CONFIG_INSTALL_BINS 1
+%define CONFIG_INSTALL_LIBS 1
+%define CONFIG_INSTALL_SRCS 0
+%define CONFIG_DEBUG 0
+%define CONFIG_GPROF 0
+%define CONFIG_GCOV 0
+%define CONFIG_RVCT 0
+%define CONFIG_GCC 1
+%define CONFIG_MSVS 0
+%define CONFIG_PIC 1
+%define CONFIG_BIG_ENDIAN 0
+%define CONFIG_CODEC_SRCS 0
+%define CONFIG_DEBUG_LIBS 0
+%define CONFIG_RUNTIME_CPU_DETECT 1
+%define CONFIG_POSTPROC 1
+%define CONFIG_MULTITHREAD 1
+%define CONFIG_INTERNAL_STATS 0
+%define CONFIG_AV1_ENCODER 1
+%define CONFIG_AV1_DECODER 1
+%define CONFIG_AV1 1
+%define CONFIG_ENCODERS 1
+%define CONFIG_DECODERS 1
+%define CONFIG_STATIC_MSVCRT 0
+%define CONFIG_SPATIAL_RESAMPLING 1
+%define CONFIG_REALTIME_ONLY 0
+%define CONFIG_ONTHEFLY_BITPACKING 0
+%define CONFIG_ERROR_CONCEALMENT 0
+%define CONFIG_SHARED 0
+%define CONFIG_STATIC 1
+%define CONFIG_SMALL 0
+%define CONFIG_POSTPROC_VISUALIZER 0
+%define CONFIG_OS_SUPPORT 1
+%define CONFIG_UNIT_TESTS 0
+%define CONFIG_WEBM_IO 1
+%define CONFIG_LIBYUV 1
+%define CONFIG_ACCOUNTING 0
+%define CONFIG_INSPECTION 0
+%define CONFIG_DECODE_PERF_TESTS 0
+%define CONFIG_ENCODE_PERF_TESTS 0
+%define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+%define CONFIG_LOWBITDEPTH 1
+%define CONFIG_HIGHBITDEPTH 0
+%define CONFIG_EXPERIMENTAL 0
+%define CONFIG_SIZE_LIMIT 1
+%define CONFIG_FP_MB_STATS 0
+%define CONFIG_CDEF 1
+%define CONFIG_VAR_TX 0
+%define CONFIG_RECT_TX 1
+%define CONFIG_REF_MV 1
+%define CONFIG_TPL_MV 0
+%define CONFIG_DUAL_FILTER 1
+%define CONFIG_CONVOLVE_ROUND 0
+%define CONFIG_COMPOUND_ROUND 0
+%define CONFIG_EXT_TX 0
+%define CONFIG_TX64X64 0
+%define CONFIG_SUB8X8_MC 0
+%define CONFIG_EXT_INTRA 1
+%define CONFIG_INTRA_INTERP 0
+%define CONFIG_FILTER_INTRA 0
+%define CONFIG_INTRABC 0
+%define CONFIG_EXT_INTER 0
+%define CONFIG_INTERINTRA 0
+%define CONFIG_WEDGE 0
+%define CONFIG_COMPOUND_SEGMENT 0
+%define CONFIG_EXT_REFS 1
+%define CONFIG_GLOBAL_MOTION 1
+%define CONFIG_NEW_QUANT 0
+%define CONFIG_SUPERTX 0
+%define CONFIG_ANS 0
+%define CONFIG_EC_MULTISYMBOL 1
+%define CONFIG_NEW_TOKENSET 1
+%define CONFIG_LOOP_RESTORATION 0
+%define CONFIG_EXT_PARTITION 0
+%define CONFIG_EXT_PARTITION_TYPES 0
+%define CONFIG_UNPOISON_PARTITION_CTX 0
+%define CONFIG_EXT_TILE 0
+%define CONFIG_MOTION_VAR 1
+%define CONFIG_NCOBMC 0
+%define CONFIG_WARPED_MOTION 1
+%define CONFIG_Q_ADAPT_PROBS 0
+%define CONFIG_SUBFRAME_PROB_UPDATE 0
+%define CONFIG_BITSTREAM_DEBUG 0
+%define CONFIG_ALT_INTRA 1
+%define CONFIG_PALETTE 1
+%define CONFIG_PALETTE_DELTA_ENCODING 0
+%define CONFIG_DAALA_EC 1
+%define CONFIG_RAWBITS 0
+%define CONFIG_EC_SMALLMUL 0
+%define CONFIG_PVQ 0
+%define CONFIG_CFL 0
+%define CONFIG_XIPHRC 0
+%define CONFIG_CB4X4 1
+%define CONFIG_CHROMA_2X2 0
+%define CONFIG_CHROMA_SUB8X8 0
+%define CONFIG_FRAME_SIZE 0
+%define CONFIG_DELTA_Q 1
+%define CONFIG_EXT_DELTA_Q 1
+%define CONFIG_ADAPT_SCAN 0
+%define CONFIG_FILTER_7BIT 1
+%define CONFIG_PARALLEL_DEBLOCKING 0
+%define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+%define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+%define CONFIG_TILE_GROUPS 1
+%define CONFIG_EC_ADAPT 1
+%define CONFIG_TEMPMV_SIGNALING 0
+%define CONFIG_RD_DEBUG 0
+%define CONFIG_REFERENCE_BUFFER 1
+%define CONFIG_COEF_INTERLEAVE 0
+%define CONFIG_ENTROPY_STATS 0
+%define CONFIG_MASKED_TX 0
+%define CONFIG_DEPENDENT_HORZTILES 0
+%define CONFIG_DAALA_DIST 0
+%define CONFIG_TRIPRED 0
+%define CONFIG_PALETTE_THROUGHPUT 1
+%define CONFIG_REF_ADAPT 0
+%define CONFIG_LV_MAP 0
+%define CONFIG_TXK_SEL 0
+%define CONFIG_MV_COMPRESS 1
+%define CONFIG_FRAME_SUPERRES 0
+%define CONFIG_NEW_MULTISYMBOL 0
+%define CONFIG_COMPOUND_SINGLEREF 0
+%define CONFIG_AOM_QM 0
+%define CONFIG_LOWDELAY_COMPOUND 0
+%define CONFIG_ANALYZER 0
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/aom_config.c
@@ -0,0 +1,11 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+#include "aom/aom_codec.h"
+static const char* const cfg = "--target=x86-linux-gcc --enable-external-build --disable-examples --disable-docs --disable-unit-tests --size-limit=8192x4608 --enable-pic --enable-postproc --as=yasm";
+const char *aom_codec_build_config(void) {return cfg;}
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/aom_config.h
@@ -0,0 +1,167 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef AOM_CONFIG_H
+#define AOM_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 1
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_AOM_PORTS 1
+#define HAVE_FEXCEPT 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_WXWIDGETS 0
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 1
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_POSTPROC 1
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 0
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_HIGHBITDEPTH 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_CDEF 1
+#define CONFIG_VAR_TX 0
+#define CONFIG_RECT_TX 1
+#define CONFIG_REF_MV 1
+#define CONFIG_TPL_MV 0
+#define CONFIG_DUAL_FILTER 1
+#define CONFIG_CONVOLVE_ROUND 0
+#define CONFIG_COMPOUND_ROUND 0
+#define CONFIG_EXT_TX 0
+#define CONFIG_TX64X64 0
+#define CONFIG_SUB8X8_MC 0
+#define CONFIG_EXT_INTRA 1
+#define CONFIG_INTRA_INTERP 0
+#define CONFIG_FILTER_INTRA 0
+#define CONFIG_INTRABC 0
+#define CONFIG_EXT_INTER 0
+#define CONFIG_INTERINTRA 0
+#define CONFIG_WEDGE 0
+#define CONFIG_COMPOUND_SEGMENT 0
+#define CONFIG_EXT_REFS 1
+#define CONFIG_GLOBAL_MOTION 1
+#define CONFIG_NEW_QUANT 0
+#define CONFIG_SUPERTX 0
+#define CONFIG_ANS 0
+#define CONFIG_EC_MULTISYMBOL 1
+#define CONFIG_NEW_TOKENSET 1
+#define CONFIG_LOOP_RESTORATION 0
+#define CONFIG_EXT_PARTITION 0
+#define CONFIG_EXT_PARTITION_TYPES 0
+#define CONFIG_UNPOISON_PARTITION_CTX 0
+#define CONFIG_EXT_TILE 0
+#define CONFIG_MOTION_VAR 1
+#define CONFIG_NCOBMC 0
+#define CONFIG_WARPED_MOTION 1
+#define CONFIG_Q_ADAPT_PROBS 0
+#define CONFIG_SUBFRAME_PROB_UPDATE 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_ALT_INTRA 1
+#define CONFIG_PALETTE 1
+#define CONFIG_PALETTE_DELTA_ENCODING 0
+#define CONFIG_DAALA_EC 1
+#define CONFIG_RAWBITS 0
+#define CONFIG_EC_SMALLMUL 0
+#define CONFIG_PVQ 0
+#define CONFIG_CFL 0
+#define CONFIG_XIPHRC 0
+#define CONFIG_CB4X4 1
+#define CONFIG_CHROMA_2X2 0
+#define CONFIG_CHROMA_SUB8X8 0
+#define CONFIG_FRAME_SIZE 0
+#define CONFIG_DELTA_Q 1
+#define CONFIG_EXT_DELTA_Q 1
+#define CONFIG_ADAPT_SCAN 0
+#define CONFIG_FILTER_7BIT 1
+#define CONFIG_PARALLEL_DEBLOCKING 0
+#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+#define CONFIG_TILE_GROUPS 1
+#define CONFIG_EC_ADAPT 1
+#define CONFIG_TEMPMV_SIGNALING 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REFERENCE_BUFFER 1
+#define CONFIG_COEF_INTERLEAVE 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_MASKED_TX 0
+#define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DAALA_DIST 0
+#define CONFIG_TRIPRED 0
+#define CONFIG_PALETTE_THROUGHPUT 1
+#define CONFIG_REF_ADAPT 0
+#define CONFIG_LV_MAP 0
+#define CONFIG_TXK_SEL 0
+#define CONFIG_MV_COMPRESS 1
+#define CONFIG_FRAME_SUPERRES 0
+#define CONFIG_NEW_MULTISYMBOL 0
+#define CONFIG_COMPOUND_SINGLEREF 0
+#define CONFIG_AOM_QM 0
+#define CONFIG_LOWDELAY_COMPOUND 0
+#define CONFIG_ANALYZER 0
+#define DECODE_WIDTH_LIMIT 8192
+#define DECODE_HEIGHT_LIMIT 4608
+#endif /* AOM_CONFIG_H */
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/aom_dsp_rtcd.h
@@ -0,0 +1,1692 @@
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_4x4)(const uint8_t *, int p);
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+RTCD_EXTERN unsigned int (*aom_avg_8x8)(const uint8_t *, int p);
+
+void aom_blend_a64_hmask_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 h, int w);
+void aom_blend_a64_hmask_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 h, int w);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(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 h, int w);
+
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+RTCD_EXTERN void (*aom_blend_a64_mask)(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, uint32_t mask_stride, int h, int w, int suby, int subx);
+
+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 h, int w);
+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 h, int w);
+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 h, int w);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_pred aom_comp_avg_pred_c
+
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_comp_avg_upsampled_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+void aom_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);
+void aom_convolve8_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_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);
+void aom_convolve8_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)(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_avg_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_avg_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_avg_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);
+RTCD_EXTERN void (*aom_convolve8_avg)(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_avg_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_avg_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_avg_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);
+RTCD_EXTERN void (*aom_convolve8_avg_horiz)(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_avg_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);
+void aom_convolve8_avg_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);
+void aom_convolve8_avg_vert_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);
+RTCD_EXTERN void (*aom_convolve8_avg_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_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);
+void aom_convolve8_horiz_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_horiz)(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_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);
+void aom_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);
+void aom_convolve8_vert_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);
+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_avg_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_convolve_avg_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);
+RTCD_EXTERN void (*aom_convolve_avg)(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 x_step_q4, const int16_t *filter_y, int y_step_q4, 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 x_step_q4, const int16_t *filter_y, int y_step_q4, 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 x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
+
+void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
+
+void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
+
+void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
+
+void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
+
+void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
+
+void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
+
+void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
+
+void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x4 aom_d135_predictor_4x4_c
+
+void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
+
+void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
+
+void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
+
+void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
+
+void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
+
+void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
+
+void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
+
+void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
+
+void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
+
+void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
+
+void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
+
+void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
+
+void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
+
+void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
+
+void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
+
+void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d63e_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d63e_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
+
+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);
+RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct16x16)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_1_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct16x16_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_1_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_rd_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct4x4_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct8x8)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct8x8_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get16x16var_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*aom_get16x16var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+#define aom_get4x4sse_cs aom_get4x4sse_cs_c
+
+void aom_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*aom_get8x8var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+RTCD_EXTERN unsigned int (*aom_get_mb_ss)(const int16_t *);
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, int src_stride, int16_t *coeff);
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, int src_stride, int16_t *coeff);
+
+void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_38_add aom_idct16x16_38_add_c
+
+void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct4x4_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_64_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width);
+int16_t aom_int_pro_col_sse2(const uint8_t *ref, int width);
+RTCD_EXTERN int16_t (*aom_int_pro_col)(const uint8_t *ref, int width);
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+RTCD_EXTERN void (*aom_int_pro_row)(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+
+void aom_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_iwht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_1_add aom_iwht4x4_1_add_c
+
+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);
+
+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);
+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_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);
+
+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);
+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_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_edge_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_8_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_edge_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_vertical_16_dual)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+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);
+
+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);
+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_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);
+
+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);
+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_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+RTCD_EXTERN void (*aom_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse8x8)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+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_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_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_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_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_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_32x32 aom_quantize_b_32x32_c
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64 aom_quantize_b_64x64_c
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x8x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x3 aom_sad32x32x3_c
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x8 aom_sad32x32x8_c
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x4x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad4x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x8 aom_sad4x8x8_c
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x3 aom_sad64x64x3_c
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x8 aom_sad64x64x8_c
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x16x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x4)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x8 aom_sad8x4x8_c
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad8x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x8x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+int aom_satd_c(const int16_t *coeff, int length);
+int aom_satd_sse2(const int16_t *coeff, int length);
+RTCD_EXTERN int (*aom_satd)(const int16_t *coeff, int length);
+
+void aom_scaled_2d_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_scaled_2d_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);
+RTCD_EXTERN void (*aom_scaled_2d)(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_scaled_avg_2d_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_scaled_avg_2d aom_scaled_avg_2d_c
+
+void aom_scaled_avg_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_scaled_avg_horiz aom_scaled_avg_horiz_c
+
+void aom_scaled_avg_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_scaled_avg_vert aom_scaled_avg_vert_c
+
+void aom_scaled_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_scaled_horiz aom_scaled_horiz_c
+
+void aom_scaled_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_scaled_vert aom_scaled_vert_c
+
+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_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_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_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_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
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void aom_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+RTCD_EXTERN void (*aom_subtract_block)(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, int height);
+RTCD_EXTERN uint64_t (*aom_sum_squares_2d_i16)(const int16_t *src, int stride, int width, int height);
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t N);
+RTCD_EXTERN uint64_t (*aom_sum_squares_i16)(const int16_t *src, uint32_t N);
+
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_upsampled_pred_sse2(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+RTCD_EXTERN void (*aom_upsampled_pred)(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+
+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_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_16x16)(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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x2 aom_variance2x2_c
+
+unsigned int aom_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance2x4 aom_variance2x4_c
+
+unsigned int aom_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x2 aom_variance4x2_c
+
+unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+uint32_t aom_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_h_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_variance_halfpixvar16x16_h)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+
+uint32_t aom_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_hv_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_variance_halfpixvar16x16_hv)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+
+uint32_t aom_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t aom_variance_halfpixvar16x16_v_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_variance_halfpixvar16x16_v)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
+int aom_vector_var_sse2(const int16_t *ref, const int16_t *src, int bwl);
+RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
+
+void aom_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_avg_4x4 = aom_avg_4x4_c;
+    if (flags & HAS_SSE2) aom_avg_4x4 = aom_avg_4x4_sse2;
+    aom_avg_8x8 = aom_avg_8x8_c;
+    if (flags & HAS_SSE2) aom_avg_8x8 = aom_avg_8x8_sse2;
+    aom_blend_a64_hmask = aom_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
+    aom_blend_a64_mask = aom_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    aom_blend_a64_vmask = aom_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
+    aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_comp_avg_upsampled_pred = aom_comp_avg_upsampled_pred_sse2;
+    aom_convolve8 = aom_convolve8_c;
+    if (flags & HAS_SSE2) aom_convolve8 = aom_convolve8_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8 = aom_convolve8_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8 = aom_convolve8_avx2;
+    aom_convolve8_avg = aom_convolve8_avg_c;
+    if (flags & HAS_SSE2) aom_convolve8_avg = aom_convolve8_avg_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_avg = aom_convolve8_avg_ssse3;
+    aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_c;
+    if (flags & HAS_SSE2) aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_ssse3;
+    aom_convolve8_avg_vert = aom_convolve8_avg_vert_c;
+    if (flags & HAS_SSE2) aom_convolve8_avg_vert = aom_convolve8_avg_vert_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_avg_vert = aom_convolve8_avg_vert_ssse3;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
+    if (flags & HAS_SSE2) aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) aom_convolve8_horiz = aom_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) aom_convolve8_horiz = aom_convolve8_horiz_avx2;
+    aom_convolve8_vert = aom_convolve8_vert_c;
+    if (flags & HAS_SSE2) 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_avg = aom_convolve_avg_c;
+    if (flags & HAS_SSE2) aom_convolve_avg = aom_convolve_avg_sse2;
+    aom_convolve_copy = aom_convolve_copy_c;
+    if (flags & HAS_SSE2) aom_convolve_copy = aom_convolve_copy_sse2;
+    aom_d153_predictor_16x16 = aom_d153_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_d153_predictor_16x16 = aom_d153_predictor_16x16_ssse3;
+    aom_d153_predictor_32x32 = aom_d153_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_d153_predictor_32x32 = aom_d153_predictor_32x32_ssse3;
+    aom_d153_predictor_4x4 = aom_d153_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_d153_predictor_4x4 = aom_d153_predictor_4x4_ssse3;
+    aom_d153_predictor_8x8 = aom_d153_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_d153_predictor_8x8 = aom_d153_predictor_8x8_ssse3;
+    aom_d63e_predictor_4x4 = aom_d63e_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_d63e_predictor_4x4 = aom_d63e_predictor_4x4_ssse3;
+    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_32x32 = aom_dc_128_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_4x4 = aom_dc_128_predictor_4x4_sse2;
+    aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_128_predictor_8x8 = aom_dc_128_predictor_8x8_sse2;
+    aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_16x16 = aom_dc_left_predictor_16x16_sse2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_4x4 = aom_dc_left_predictor_4x4_sse2;
+    aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_left_predictor_8x8 = aom_dc_left_predictor_8x8_sse2;
+    aom_dc_predictor_16x16 = aom_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_16x16 = aom_dc_predictor_16x16_sse2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    aom_dc_predictor_4x4 = aom_dc_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_4x4 = aom_dc_predictor_4x4_sse2;
+    aom_dc_predictor_8x8 = aom_dc_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_predictor_8x8 = aom_dc_predictor_8x8_sse2;
+    aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_16x16 = aom_dc_top_predictor_16x16_sse2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_4x4 = aom_dc_top_predictor_4x4_sse2;
+    aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_dc_top_predictor_8x8 = aom_dc_top_predictor_8x8_sse2;
+    aom_fdct16x16 = aom_fdct16x16_c;
+    if (flags & HAS_SSE2) aom_fdct16x16 = aom_fdct16x16_sse2;
+    aom_fdct16x16_1 = aom_fdct16x16_1_c;
+    if (flags & HAS_SSE2) aom_fdct16x16_1 = aom_fdct16x16_1_sse2;
+    if (flags & HAS_AVX2) aom_fdct16x16_1 = aom_fdct16x16_1_avx2;
+    aom_fdct32x32 = aom_fdct32x32_c;
+    if (flags & HAS_SSE2) aom_fdct32x32 = aom_fdct32x32_sse2;
+    if (flags & HAS_AVX2) aom_fdct32x32 = aom_fdct32x32_avx2;
+    aom_fdct32x32_1 = aom_fdct32x32_1_c;
+    if (flags & HAS_SSE2) aom_fdct32x32_1 = aom_fdct32x32_1_sse2;
+    if (flags & HAS_AVX2) aom_fdct32x32_1 = aom_fdct32x32_1_avx2;
+    aom_fdct32x32_rd = aom_fdct32x32_rd_c;
+    if (flags & HAS_SSE2) aom_fdct32x32_rd = aom_fdct32x32_rd_sse2;
+    if (flags & HAS_AVX2) aom_fdct32x32_rd = aom_fdct32x32_rd_avx2;
+    aom_fdct4x4 = aom_fdct4x4_c;
+    if (flags & HAS_SSE2) aom_fdct4x4 = aom_fdct4x4_sse2;
+    aom_fdct4x4_1 = aom_fdct4x4_1_c;
+    if (flags & HAS_SSE2) aom_fdct4x4_1 = aom_fdct4x4_1_sse2;
+    aom_fdct8x8 = aom_fdct8x8_c;
+    if (flags & HAS_SSE2) aom_fdct8x8 = aom_fdct8x8_sse2;
+    aom_fdct8x8_1 = aom_fdct8x8_1_c;
+    if (flags & HAS_SSE2) aom_fdct8x8_1 = aom_fdct8x8_1_sse2;
+    aom_get16x16var = aom_get16x16var_c;
+    if (flags & HAS_SSE2) aom_get16x16var = aom_get16x16var_sse2;
+    if (flags & HAS_AVX2) aom_get16x16var = aom_get16x16var_avx2;
+    aom_get8x8var = aom_get8x8var_c;
+    if (flags & HAS_SSE2) aom_get8x8var = aom_get8x8var_sse2;
+    aom_get_mb_ss = aom_get_mb_ss_c;
+    if (flags & HAS_SSE2) aom_get_mb_ss = aom_get_mb_ss_sse2;
+    aom_h_predictor_16x16 = aom_h_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_h_predictor_16x16 = aom_h_predictor_16x16_sse2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    aom_h_predictor_4x4 = aom_h_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_h_predictor_4x4 = aom_h_predictor_4x4_sse2;
+    aom_h_predictor_8x8 = aom_h_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_h_predictor_8x8 = aom_h_predictor_8x8_sse2;
+    aom_hadamard_16x16 = aom_hadamard_16x16_c;
+    if (flags & HAS_SSE2) aom_hadamard_16x16 = aom_hadamard_16x16_sse2;
+    aom_hadamard_8x8 = aom_hadamard_8x8_c;
+    if (flags & HAS_SSE2) aom_hadamard_8x8 = aom_hadamard_8x8_sse2;
+    aom_idct16x16_10_add = aom_idct16x16_10_add_c;
+    if (flags & HAS_SSE2) aom_idct16x16_10_add = aom_idct16x16_10_add_sse2;
+    aom_idct16x16_1_add = aom_idct16x16_1_add_c;
+    if (flags & HAS_SSE2) aom_idct16x16_1_add = aom_idct16x16_1_add_sse2;
+    aom_idct16x16_256_add = aom_idct16x16_256_add_c;
+    if (flags & HAS_SSE2) aom_idct16x16_256_add = aom_idct16x16_256_add_sse2;
+    aom_idct32x32_1024_add = aom_idct32x32_1024_add_c;
+    if (flags & HAS_SSE2) aom_idct32x32_1024_add = aom_idct32x32_1024_add_sse2;
+    if (flags & HAS_SSSE3) aom_idct32x32_1024_add = aom_idct32x32_1024_add_ssse3;
+    aom_idct32x32_135_add = aom_idct32x32_135_add_c;
+    if (flags & HAS_SSE2) aom_idct32x32_135_add = aom_idct32x32_1024_add_sse2;
+    if (flags & HAS_SSSE3) aom_idct32x32_135_add = aom_idct32x32_135_add_ssse3;
+    aom_idct32x32_1_add = aom_idct32x32_1_add_c;
+    if (flags & HAS_SSE2) aom_idct32x32_1_add = aom_idct32x32_1_add_sse2;
+    aom_idct32x32_34_add = aom_idct32x32_34_add_c;
+    if (flags & HAS_SSE2) aom_idct32x32_34_add = aom_idct32x32_34_add_sse2;
+    if (flags & HAS_SSSE3) aom_idct32x32_34_add = aom_idct32x32_34_add_ssse3;
+    aom_idct4x4_16_add = aom_idct4x4_16_add_c;
+    if (flags & HAS_SSE2) aom_idct4x4_16_add = aom_idct4x4_16_add_sse2;
+    aom_idct4x4_1_add = aom_idct4x4_1_add_c;
+    if (flags & HAS_SSE2) aom_idct4x4_1_add = aom_idct4x4_1_add_sse2;
+    aom_idct8x8_12_add = aom_idct8x8_12_add_c;
+    if (flags & HAS_SSE2) aom_idct8x8_12_add = aom_idct8x8_12_add_sse2;
+    if (flags & HAS_SSSE3) aom_idct8x8_12_add = aom_idct8x8_12_add_ssse3;
+    aom_idct8x8_1_add = aom_idct8x8_1_add_c;
+    if (flags & HAS_SSE2) aom_idct8x8_1_add = aom_idct8x8_1_add_sse2;
+    aom_idct8x8_64_add = aom_idct8x8_64_add_c;
+    if (flags & HAS_SSE2) aom_idct8x8_64_add = aom_idct8x8_64_add_sse2;
+    if (flags & HAS_SSSE3) aom_idct8x8_64_add = aom_idct8x8_64_add_ssse3;
+    aom_int_pro_col = aom_int_pro_col_c;
+    if (flags & HAS_SSE2) aom_int_pro_col = aom_int_pro_col_sse2;
+    aom_int_pro_row = aom_int_pro_row_c;
+    if (flags & HAS_SSE2) aom_int_pro_row = aom_int_pro_row_sse2;
+    aom_iwht4x4_16_add = aom_iwht4x4_16_add_c;
+    if (flags & HAS_SSE2) aom_iwht4x4_16_add = aom_iwht4x4_16_add_sse2;
+    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_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_edge_16 = aom_lpf_horizontal_edge_16_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_avx2;
+    aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_c;
+    if (flags & HAS_SSE2) aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_sse2;
+    if (flags & HAS_AVX2) aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_avx2;
+    aom_lpf_vertical_16 = aom_lpf_vertical_16_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_16 = aom_lpf_vertical_16_sse2;
+    aom_lpf_vertical_16_dual = aom_lpf_vertical_16_dual_c;
+    if (flags & HAS_SSE2) aom_lpf_vertical_16_dual = aom_lpf_vertical_16_dual_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_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_minmax_8x8 = aom_minmax_8x8_c;
+    if (flags & HAS_SSE2) aom_minmax_8x8 = aom_minmax_8x8_sse2;
+    aom_mse16x16 = aom_mse16x16_c;
+    if (flags & HAS_SSE2) aom_mse16x16 = aom_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
+    aom_mse16x8 = aom_mse16x8_c;
+    if (flags & HAS_SSE2) aom_mse16x8 = aom_mse16x8_sse2;
+    aom_mse8x16 = aom_mse8x16_c;
+    if (flags & HAS_SSE2) aom_mse8x16 = aom_mse8x16_sse2;
+    aom_mse8x8 = aom_mse8x8_c;
+    if (flags & HAS_SSE2) aom_mse8x8 = aom_mse8x8_sse2;
+    aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x16 = aom_obmc_sad16x16_sse4_1;
+    aom_obmc_sad16x32 = aom_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x32 = aom_obmc_sad16x32_sse4_1;
+    aom_obmc_sad16x8 = aom_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad16x8 = aom_obmc_sad16x8_sse4_1;
+    aom_obmc_sad32x16 = aom_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x16 = aom_obmc_sad32x16_sse4_1;
+    aom_obmc_sad32x32 = aom_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x32 = aom_obmc_sad32x32_sse4_1;
+    aom_obmc_sad32x64 = aom_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad32x64 = aom_obmc_sad32x64_sse4_1;
+    aom_obmc_sad4x4 = aom_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x4 = aom_obmc_sad4x4_sse4_1;
+    aom_obmc_sad4x8 = aom_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad4x8 = aom_obmc_sad4x8_sse4_1;
+    aom_obmc_sad64x32 = aom_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x32 = aom_obmc_sad64x32_sse4_1;
+    aom_obmc_sad64x64 = aom_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad64x64 = aom_obmc_sad64x64_sse4_1;
+    aom_obmc_sad8x16 = aom_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x16 = aom_obmc_sad8x16_sse4_1;
+    aom_obmc_sad8x4 = aom_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x4 = aom_obmc_sad8x4_sse4_1;
+    aom_obmc_sad8x8 = aom_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_sad8x8 = aom_obmc_sad8x8_sse4_1;
+    aom_obmc_variance16x16 = aom_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x16 = aom_obmc_variance16x16_sse4_1;
+    aom_obmc_variance16x32 = aom_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x32 = aom_obmc_variance16x32_sse4_1;
+    aom_obmc_variance16x8 = aom_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance16x8 = aom_obmc_variance16x8_sse4_1;
+    aom_obmc_variance32x16 = aom_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x16 = aom_obmc_variance32x16_sse4_1;
+    aom_obmc_variance32x32 = aom_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x32 = aom_obmc_variance32x32_sse4_1;
+    aom_obmc_variance32x64 = aom_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance32x64 = aom_obmc_variance32x64_sse4_1;
+    aom_obmc_variance4x4 = aom_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x4 = aom_obmc_variance4x4_sse4_1;
+    aom_obmc_variance4x8 = aom_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance4x8 = aom_obmc_variance4x8_sse4_1;
+    aom_obmc_variance64x32 = aom_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x32 = aom_obmc_variance64x32_sse4_1;
+    aom_obmc_variance64x64 = aom_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance64x64 = aom_obmc_variance64x64_sse4_1;
+    aom_obmc_variance8x16 = aom_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x16 = aom_obmc_variance8x16_sse4_1;
+    aom_obmc_variance8x4 = aom_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
+    aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    aom_quantize_b = aom_quantize_b_c;
+    if (flags & HAS_SSE2) aom_quantize_b = aom_quantize_b_sse2;
+    aom_sad16x16 = aom_sad16x16_c;
+    if (flags & HAS_SSE2) aom_sad16x16 = aom_sad16x16_sse2;
+    aom_sad16x16_avg = aom_sad16x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x16_avg = aom_sad16x16_avg_sse2;
+    aom_sad16x16x3 = aom_sad16x16x3_c;
+    if (flags & HAS_SSE3) aom_sad16x16x3 = aom_sad16x16x3_sse3;
+    if (flags & HAS_SSSE3) aom_sad16x16x3 = aom_sad16x16x3_ssse3;
+    aom_sad16x16x4d = aom_sad16x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x16x4d = aom_sad16x16x4d_sse2;
+    aom_sad16x16x8 = aom_sad16x16x8_c;
+    if (flags & HAS_SSE4_1) aom_sad16x16x8 = aom_sad16x16x8_sse4_1;
+    aom_sad16x32 = aom_sad16x32_c;
+    if (flags & HAS_SSE2) aom_sad16x32 = aom_sad16x32_sse2;
+    aom_sad16x32_avg = aom_sad16x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x32_avg = aom_sad16x32_avg_sse2;
+    aom_sad16x32x4d = aom_sad16x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x32x4d = aom_sad16x32x4d_sse2;
+    aom_sad16x8 = aom_sad16x8_c;
+    if (flags & HAS_SSE2) aom_sad16x8 = aom_sad16x8_sse2;
+    aom_sad16x8_avg = aom_sad16x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad16x8_avg = aom_sad16x8_avg_sse2;
+    aom_sad16x8x3 = aom_sad16x8x3_c;
+    if (flags & HAS_SSE3) aom_sad16x8x3 = aom_sad16x8x3_sse3;
+    if (flags & HAS_SSSE3) aom_sad16x8x3 = aom_sad16x8x3_ssse3;
+    aom_sad16x8x4d = aom_sad16x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad16x8x4d = aom_sad16x8x4d_sse2;
+    aom_sad16x8x8 = aom_sad16x8x8_c;
+    if (flags & HAS_SSE4_1) aom_sad16x8x8 = aom_sad16x8x8_sse4_1;
+    aom_sad32x16 = aom_sad32x16_c;
+    if (flags & HAS_SSE2) aom_sad32x16 = aom_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16 = aom_sad32x16_avx2;
+    aom_sad32x16_avg = aom_sad32x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x16_avg = aom_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x16_avg = aom_sad32x16_avg_avx2;
+    aom_sad32x16x4d = aom_sad32x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x16x4d = aom_sad32x16x4d_sse2;
+    aom_sad32x32 = aom_sad32x32_c;
+    if (flags & HAS_SSE2) aom_sad32x32 = aom_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32 = aom_sad32x32_avx2;
+    aom_sad32x32_avg = aom_sad32x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x32_avg = aom_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32_avg = aom_sad32x32_avg_avx2;
+    aom_sad32x32x4d = aom_sad32x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x32x4d = aom_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x32x4d = aom_sad32x32x4d_avx2;
+    aom_sad32x64 = aom_sad32x64_c;
+    if (flags & HAS_SSE2) aom_sad32x64 = aom_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64 = aom_sad32x64_avx2;
+    aom_sad32x64_avg = aom_sad32x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad32x64_avg = aom_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64_avg = aom_sad32x64_avg_avx2;
+    aom_sad32x64x4d = aom_sad32x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad32x64x4d = aom_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad32x64x4d = aom_sad32x64x4d_avx2;
+    aom_sad4x4 = aom_sad4x4_c;
+    if (flags & HAS_SSE2) aom_sad4x4 = aom_sad4x4_sse2;
+    aom_sad4x4_avg = aom_sad4x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x4_avg = aom_sad4x4_avg_sse2;
+    aom_sad4x4x3 = aom_sad4x4x3_c;
+    if (flags & HAS_SSE3) aom_sad4x4x3 = aom_sad4x4x3_sse3;
+    aom_sad4x4x4d = aom_sad4x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x4x4d = aom_sad4x4x4d_sse2;
+    aom_sad4x4x8 = aom_sad4x4x8_c;
+    if (flags & HAS_SSE4_1) aom_sad4x4x8 = aom_sad4x4x8_sse4_1;
+    aom_sad4x8 = aom_sad4x8_c;
+    if (flags & HAS_SSE2) aom_sad4x8 = aom_sad4x8_sse2;
+    aom_sad4x8_avg = aom_sad4x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad4x8_avg = aom_sad4x8_avg_sse2;
+    aom_sad4x8x4d = aom_sad4x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad4x8x4d = aom_sad4x8x4d_sse2;
+    aom_sad64x32 = aom_sad64x32_c;
+    if (flags & HAS_SSE2) aom_sad64x32 = aom_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32 = aom_sad64x32_avx2;
+    aom_sad64x32_avg = aom_sad64x32_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x32_avg = aom_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32_avg = aom_sad64x32_avg_avx2;
+    aom_sad64x32x4d = aom_sad64x32x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x32x4d = aom_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x32x4d = aom_sad64x32x4d_avx2;
+    aom_sad64x64 = aom_sad64x64_c;
+    if (flags & HAS_SSE2) aom_sad64x64 = aom_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64 = aom_sad64x64_avx2;
+    aom_sad64x64_avg = aom_sad64x64_avg_c;
+    if (flags & HAS_SSE2) aom_sad64x64_avg = aom_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64_avg = aom_sad64x64_avg_avx2;
+    aom_sad64x64x4d = aom_sad64x64x4d_c;
+    if (flags & HAS_SSE2) aom_sad64x64x4d = aom_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_sad64x64x4d = aom_sad64x64x4d_avx2;
+    aom_sad8x16 = aom_sad8x16_c;
+    if (flags & HAS_SSE2) aom_sad8x16 = aom_sad8x16_sse2;
+    aom_sad8x16_avg = aom_sad8x16_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x16_avg = aom_sad8x16_avg_sse2;
+    aom_sad8x16x3 = aom_sad8x16x3_c;
+    if (flags & HAS_SSE3) aom_sad8x16x3 = aom_sad8x16x3_sse3;
+    aom_sad8x16x4d = aom_sad8x16x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x16x4d = aom_sad8x16x4d_sse2;
+    aom_sad8x16x8 = aom_sad8x16x8_c;
+    if (flags & HAS_SSE4_1) aom_sad8x16x8 = aom_sad8x16x8_sse4_1;
+    aom_sad8x4 = aom_sad8x4_c;
+    if (flags & HAS_SSE2) aom_sad8x4 = aom_sad8x4_sse2;
+    aom_sad8x4_avg = aom_sad8x4_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x4_avg = aom_sad8x4_avg_sse2;
+    aom_sad8x4x4d = aom_sad8x4x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x4x4d = aom_sad8x4x4d_sse2;
+    aom_sad8x8 = aom_sad8x8_c;
+    if (flags & HAS_SSE2) aom_sad8x8 = aom_sad8x8_sse2;
+    aom_sad8x8_avg = aom_sad8x8_avg_c;
+    if (flags & HAS_SSE2) aom_sad8x8_avg = aom_sad8x8_avg_sse2;
+    aom_sad8x8x3 = aom_sad8x8x3_c;
+    if (flags & HAS_SSE3) aom_sad8x8x3 = aom_sad8x8x3_sse3;
+    aom_sad8x8x4d = aom_sad8x8x4d_c;
+    if (flags & HAS_SSE2) aom_sad8x8x4d = aom_sad8x8x4d_sse2;
+    aom_sad8x8x8 = aom_sad8x8x8_c;
+    if (flags & HAS_SSE4_1) aom_sad8x8x8 = aom_sad8x8x8_sse4_1;
+    aom_satd = aom_satd_c;
+    if (flags & HAS_SSE2) aom_satd = aom_satd_sse2;
+    aom_scaled_2d = aom_scaled_2d_c;
+    if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_subtract_block = aom_subtract_block_c;
+    if (flags & HAS_SSE2) aom_subtract_block = aom_subtract_block_sse2;
+    aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_2d_i16 = aom_sum_squares_2d_i16_sse2;
+    aom_sum_squares_i16 = aom_sum_squares_i16_c;
+    if (flags & HAS_SSE2) aom_sum_squares_i16 = aom_sum_squares_i16_sse2;
+    aom_upsampled_pred = aom_upsampled_pred_c;
+    if (flags & HAS_SSE2) aom_upsampled_pred = aom_upsampled_pred_sse2;
+    aom_v_predictor_16x16 = aom_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) aom_v_predictor_16x16 = aom_v_predictor_16x16_sse2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    aom_v_predictor_4x4 = aom_v_predictor_4x4_c;
+    if (flags & HAS_SSE2) aom_v_predictor_4x4 = aom_v_predictor_4x4_sse2;
+    aom_v_predictor_8x8 = aom_v_predictor_8x8_c;
+    if (flags & HAS_SSE2) aom_v_predictor_8x8 = aom_v_predictor_8x8_sse2;
+    aom_variance16x16 = aom_variance16x16_c;
+    if (flags & HAS_SSE2) aom_variance16x16 = aom_variance16x16_sse2;
+    if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
+    aom_variance16x32 = aom_variance16x32_c;
+    if (flags & HAS_SSE2) aom_variance16x32 = aom_variance16x32_sse2;
+    aom_variance16x8 = aom_variance16x8_c;
+    if (flags & HAS_SSE2) aom_variance16x8 = aom_variance16x8_sse2;
+    aom_variance32x16 = aom_variance32x16_c;
+    if (flags & HAS_SSE2) aom_variance32x16 = aom_variance32x16_sse2;
+    if (flags & HAS_AVX2) aom_variance32x16 = aom_variance32x16_avx2;
+    aom_variance32x32 = aom_variance32x32_c;
+    if (flags & HAS_SSE2) aom_variance32x32 = aom_variance32x32_sse2;
+    if (flags & HAS_AVX2) aom_variance32x32 = aom_variance32x32_avx2;
+    aom_variance32x64 = aom_variance32x64_c;
+    if (flags & HAS_SSE2) aom_variance32x64 = aom_variance32x64_sse2;
+    aom_variance4x4 = aom_variance4x4_c;
+    if (flags & HAS_SSE2) aom_variance4x4 = aom_variance4x4_sse2;
+    aom_variance4x8 = aom_variance4x8_c;
+    if (flags & HAS_SSE2) aom_variance4x8 = aom_variance4x8_sse2;
+    aom_variance64x32 = aom_variance64x32_c;
+    if (flags & HAS_SSE2) aom_variance64x32 = aom_variance64x32_sse2;
+    if (flags & HAS_AVX2) aom_variance64x32 = aom_variance64x32_avx2;
+    aom_variance64x64 = aom_variance64x64_c;
+    if (flags & HAS_SSE2) aom_variance64x64 = aom_variance64x64_sse2;
+    if (flags & HAS_AVX2) aom_variance64x64 = aom_variance64x64_avx2;
+    aom_variance8x16 = aom_variance8x16_c;
+    if (flags & HAS_SSE2) aom_variance8x16 = aom_variance8x16_sse2;
+    aom_variance8x4 = aom_variance8x4_c;
+    if (flags & HAS_SSE2) aom_variance8x4 = aom_variance8x4_sse2;
+    aom_variance8x8 = aom_variance8x8_c;
+    if (flags & HAS_SSE2) aom_variance8x8 = aom_variance8x8_sse2;
+    aom_variance_halfpixvar16x16_h = aom_variance_halfpixvar16x16_h_c;
+    if (flags & HAS_SSE2) aom_variance_halfpixvar16x16_h = aom_variance_halfpixvar16x16_h_sse2;
+    aom_variance_halfpixvar16x16_hv = aom_variance_halfpixvar16x16_hv_c;
+    if (flags & HAS_SSE2) aom_variance_halfpixvar16x16_hv = aom_variance_halfpixvar16x16_hv_sse2;
+    aom_variance_halfpixvar16x16_v = aom_variance_halfpixvar16x16_v_c;
+    if (flags & HAS_SSE2) aom_variance_halfpixvar16x16_v = aom_variance_halfpixvar16x16_v_sse2;
+    aom_vector_var = aom_vector_var_c;
+    if (flags & HAS_SSE2) aom_vector_var = aom_vector_var_sse2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/aom_scale_rtcd.h
@@ -0,0 +1,78 @@
+#ifndef AOM_SCALE_RTCD_H_
+#define AOM_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders aom_extend_frame_borders_c
+
+void aom_extend_frame_borders_y_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_borders_y aom_extend_frame_borders_y_c
+
+void aom_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define aom_extend_frame_inner_borders aom_extend_frame_inner_borders_c
+
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_2_1_scale aom_horizontal_line_2_1_scale_c
+
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_3_scale aom_horizontal_line_5_3_scale_c
+
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define aom_horizontal_line_5_4_scale aom_horizontal_line_5_4_scale_c
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale aom_vertical_band_2_1_scale_c
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_2_1_scale_i aom_vertical_band_2_1_scale_i_c
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_3_scale aom_vertical_band_5_3_scale_c
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width);
+#define aom_vertical_band_5_4_scale aom_vertical_band_5_4_scale_c
+
+void aom_yv12_copy_frame_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_frame aom_yv12_copy_frame_c
+
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_u aom_yv12_copy_u_c
+
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc);
+#define aom_yv12_copy_v aom_yv12_copy_v_c
+
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define aom_yv12_copy_y aom_yv12_copy_y_c
+
+void aom_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define aom_yv12_extend_frame_borders aom_yv12_extend_frame_borders_c
+
+void aom_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/ia32/av1_rtcd.h
@@ -0,0 +1,416 @@
+#ifndef AOM_RTCD_H_
+#define AOM_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct aom_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+typedef uint16_t od_dering_in;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void aom_clpf_block_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_sse2(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_ssse3(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_sse4_1(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_block)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_block_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_hbd_sse2(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_hbd_ssse3(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_block_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_block_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_hblock_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_sse2(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_ssse3(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_sse4_1(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_hblock)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+void aom_clpf_hblock_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_hbd_sse2(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_hbd_ssse3(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void aom_clpf_hblock_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+RTCD_EXTERN void (*aom_clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t av1_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+RTCD_EXTERN int64_t (*av1_block_error_fp)(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+
+void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+void av1_convolve_horiz_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_horiz)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+
+void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+void av1_convolve_vert_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_vert)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
+
+int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_diamond_search_sad av1_diamond_search_sad_c
+
+void av1_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_fdct8x8_quant_sse2(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_fdct8x8_quant_ssse3(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_fdct8x8_quant)(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x16_avx2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht16x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x32_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht16x32)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht16x4 av1_fht16x4_c
+
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht16x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht32x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x32_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x32_avx2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht32x32)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht32x8 av1_fht32x8_c
+
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht4x16 av1_fht4x16_c
+
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht4x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht8x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define av1_fht8x32 av1_fht8x32_c
+
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht8x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*av1_fht8x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+
+int av1_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define av1_full_range_search av1_full_range_search_c
+
+int av1_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+int av1_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+int av1_full_search_sadx8(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+RTCD_EXTERN int (*av1_full_search_sad)(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+
+void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
+#define av1_fwd_idtx av1_fwd_idtx_c
+
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define av1_fwht4x4 av1_fwht4x4_c
+
+void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+RTCD_EXTERN void (*av1_iht16x16_256_add)(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht16x32_512_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht16x32_512_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht16x4_64_add av1_iht16x4_64_add_c
+
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht16x8_128_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht16x8_128_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht32x16_512_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht32x16_512_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define av1_iht32x32_1024_add av1_iht32x32_1024_add_c
+
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht32x8_256_add av1_iht32x8_256_add_c
+
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht4x16_64_add av1_iht4x16_64_add_c
+
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x8_32_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht4x8_32_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x16_128_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht8x16_128_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define av1_iht8x32_256_add av1_iht8x32_256_add_c
+
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x4_32_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht8x4_32_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+RTCD_EXTERN void (*av1_iht8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+void av1_lowbd_convolve_init_c(void);
+void av1_lowbd_convolve_init_ssse3(void);
+RTCD_EXTERN void (*av1_lowbd_convolve_init)(void);
+
+void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
+#define av1_quantize_b av1_quantize_b_c
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void av1_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*av1_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
+
+void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void av1_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+RTCD_EXTERN void (*av1_temporal_filter_apply)(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+
+void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse2(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_4x4_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_4x4_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_8bit_sse2(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_8bit_ssse3(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_4x4_16bit_to_8bit_sse4_1(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_4x4_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_8x8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_8x8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_8x8_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_8bit_sse2(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_8bit_ssse3(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+void copy_8x8_16bit_to_8bit_sse4_1(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+RTCD_EXTERN void (*copy_8x8_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+
+int od_dir_find8_c(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+int od_dir_find8_sse2(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+int od_dir_find8_ssse3(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+int od_dir_find8_sse4_1(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*od_dir_find8)(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
+
+void od_filter_dering_direction_4x4_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_4x4_sse2(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_4x4_ssse3(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_4x4_sse4_1(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+RTCD_EXTERN void (*od_filter_dering_direction_4x4)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+
+void od_filter_dering_direction_8x8_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_8x8_sse2(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_8x8_ssse3(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+void od_filter_dering_direction_8x8_sse4_1(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+RTCD_EXTERN void (*od_filter_dering_direction_8x8)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
+
+void aom_rtcd(void);
+
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+    aom_clpf_block = aom_clpf_block_c;
+    if (flags & HAS_SSE2) aom_clpf_block = aom_clpf_block_sse2;
+    if (flags & HAS_SSSE3) aom_clpf_block = aom_clpf_block_ssse3;
+    if (flags & HAS_SSE4_1) aom_clpf_block = aom_clpf_block_sse4_1;
+    aom_clpf_block_hbd = aom_clpf_block_hbd_c;
+    if (flags & HAS_SSE2) aom_clpf_block_hbd = aom_clpf_block_hbd_sse2;
+    if (flags & HAS_SSSE3) aom_clpf_block_hbd = aom_clpf_block_hbd_ssse3;
+    if (flags & HAS_SSE4_1) aom_clpf_block_hbd = aom_clpf_block_hbd_sse4_1;
+    aom_clpf_hblock = aom_clpf_hblock_c;
+    if (flags & HAS_SSE2) aom_clpf_hblock = aom_clpf_hblock_sse2;
+    if (flags & HAS_SSSE3) aom_clpf_hblock = aom_clpf_hblock_ssse3;
+    if (flags & HAS_SSE4_1) aom_clpf_hblock = aom_clpf_hblock_sse4_1;
+    aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_c;
+    if (flags & HAS_SSE2) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse2;
+    if (flags & HAS_SSSE3) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_ssse3;
+    if (flags & HAS_SSE4_1) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse4_1;
+    av1_block_error = av1_block_error_c;
+    if (flags & HAS_SSE2) av1_block_error = av1_block_error_sse2;
+    if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_block_error_fp = av1_block_error_fp_c;
+    if (flags & HAS_SSE2) av1_block_error_fp = av1_block_error_fp_sse2;
+    av1_convolve_horiz = av1_convolve_horiz_c;
+    if (flags & HAS_SSSE3) av1_convolve_horiz = av1_convolve_horiz_ssse3;
+    av1_convolve_vert = av1_convolve_vert_c;
+    if (flags & HAS_SSSE3) av1_convolve_vert = av1_convolve_vert_ssse3;
+    av1_fdct8x8_quant = av1_fdct8x8_quant_c;
+    if (flags & HAS_SSE2) av1_fdct8x8_quant = av1_fdct8x8_quant_sse2;
+    if (flags & HAS_SSSE3) av1_fdct8x8_quant = av1_fdct8x8_quant_ssse3;
+    av1_fht16x16 = av1_fht16x16_c;
+    if (flags & HAS_SSE2) av1_fht16x16 = av1_fht16x16_sse2;
+    if (flags & HAS_AVX2) av1_fht16x16 = av1_fht16x16_avx2;
+    av1_fht16x32 = av1_fht16x32_c;
+    if (flags & HAS_SSE2) av1_fht16x32 = av1_fht16x32_sse2;
+    av1_fht16x8 = av1_fht16x8_c;
+    if (flags & HAS_SSE2) av1_fht16x8 = av1_fht16x8_sse2;
+    av1_fht32x16 = av1_fht32x16_c;
+    if (flags & HAS_SSE2) av1_fht32x16 = av1_fht32x16_sse2;
+    av1_fht32x32 = av1_fht32x32_c;
+    if (flags & HAS_SSE2) av1_fht32x32 = av1_fht32x32_sse2;
+    if (flags & HAS_AVX2) av1_fht32x32 = av1_fht32x32_avx2;
+    av1_fht4x4 = av1_fht4x4_c;
+    if (flags & HAS_SSE2) av1_fht4x4 = av1_fht4x4_sse2;
+    av1_fht4x8 = av1_fht4x8_c;
+    if (flags & HAS_SSE2) av1_fht4x8 = av1_fht4x8_sse2;
+    av1_fht8x16 = av1_fht8x16_c;
+    if (flags & HAS_SSE2) av1_fht8x16 = av1_fht8x16_sse2;
+    av1_fht8x4 = av1_fht8x4_c;
+    if (flags & HAS_SSE2) av1_fht8x4 = av1_fht8x4_sse2;
+    av1_fht8x8 = av1_fht8x8_c;
+    if (flags & HAS_SSE2) av1_fht8x8 = av1_fht8x8_sse2;
+    av1_full_search_sad = av1_full_search_sad_c;
+    if (flags & HAS_SSE3) av1_full_search_sad = av1_full_search_sadx3;
+    if (flags & HAS_SSE4_1) av1_full_search_sad = av1_full_search_sadx8;
+    av1_iht16x16_256_add = av1_iht16x16_256_add_c;
+    if (flags & HAS_SSE2) av1_iht16x16_256_add = av1_iht16x16_256_add_sse2;
+    if (flags & HAS_AVX2) av1_iht16x16_256_add = av1_iht16x16_256_add_avx2;
+    av1_iht16x32_512_add = av1_iht16x32_512_add_c;
+    if (flags & HAS_SSE2) av1_iht16x32_512_add = av1_iht16x32_512_add_sse2;
+    av1_iht16x8_128_add = av1_iht16x8_128_add_c;
+    if (flags & HAS_SSE2) av1_iht16x8_128_add = av1_iht16x8_128_add_sse2;
+    av1_iht32x16_512_add = av1_iht32x16_512_add_c;
+    if (flags & HAS_SSE2) av1_iht32x16_512_add = av1_iht32x16_512_add_sse2;
+    av1_iht4x4_16_add = av1_iht4x4_16_add_c;
+    if (flags & HAS_SSE2) av1_iht4x4_16_add = av1_iht4x4_16_add_sse2;
+    av1_iht4x8_32_add = av1_iht4x8_32_add_c;
+    if (flags & HAS_SSE2) av1_iht4x8_32_add = av1_iht4x8_32_add_sse2;
+    av1_iht8x16_128_add = av1_iht8x16_128_add_c;
+    if (flags & HAS_SSE2) av1_iht8x16_128_add = av1_iht8x16_128_add_sse2;
+    av1_iht8x4_32_add = av1_iht8x4_32_add_c;
+    if (flags & HAS_SSE2) av1_iht8x4_32_add = av1_iht8x4_32_add_sse2;
+    av1_iht8x8_64_add = av1_iht8x8_64_add_c;
+    if (flags & HAS_SSE2) av1_iht8x8_64_add = av1_iht8x8_64_add_sse2;
+    av1_lowbd_convolve_init = av1_lowbd_convolve_init_c;
+    if (flags & HAS_SSSE3) av1_lowbd_convolve_init = av1_lowbd_convolve_init_ssse3;
+    av1_quantize_fp = av1_quantize_fp_c;
+    if (flags & HAS_SSE2) av1_quantize_fp = av1_quantize_fp_sse2;
+    av1_temporal_filter_apply = av1_temporal_filter_apply_c;
+    if (flags & HAS_SSE2) av1_temporal_filter_apply = av1_temporal_filter_apply_sse2;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE2) av1_warp_affine = av1_warp_affine_sse2;
+    copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse4_1;
+    copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_c;
+    if (flags & HAS_SSE2) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_sse2;
+    if (flags & HAS_SSSE3) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_sse4_1;
+    copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_sse4_1;
+    copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_c;
+    if (flags & HAS_SSE2) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_sse2;
+    if (flags & HAS_SSSE3) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_sse4_1;
+    copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
+    copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
+    od_dir_find8 = od_dir_find8_c;
+    if (flags & HAS_SSE2) od_dir_find8 = od_dir_find8_sse2;
+    if (flags & HAS_SSSE3) od_dir_find8 = od_dir_find8_ssse3;
+    if (flags & HAS_SSE4_1) od_dir_find8 = od_dir_find8_sse4_1;
+    od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_c;
+    if (flags & HAS_SSE2) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_sse2;
+    if (flags & HAS_SSSE3) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_ssse3;
+    if (flags & HAS_SSE4_1) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_sse4_1;
+    od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_c;
+    if (flags & HAS_SSE2) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_sse2;
+    if (flags & HAS_SSSE3) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_ssse3;
+    if (flags & HAS_SSE4_1) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_sse4_1;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/x64/aom_config.asm
@@ -0,0 +1,151 @@
+%define ARCH_ARM 0
+%define ARCH_MIPS 0
+%define ARCH_X86 0
+%define ARCH_X86_64 1
+%define HAVE_EDSP 0
+%define HAVE_MEDIA 0
+%define HAVE_NEON 0
+%define HAVE_NEON_ASM 0
+%define HAVE_MIPS32 0
+%define HAVE_DSPR2 0
+%define HAVE_MSA 0
+%define HAVE_MIPS64 0
+%define HAVE_MMX 1
+%define HAVE_SSE 1
+%define HAVE_SSE2 1
+%define HAVE_SSE3 1
+%define HAVE_SSSE3 1
+%define HAVE_SSE4_1 1
+%define HAVE_AVX 1
+%define HAVE_AVX2 1
+%define HAVE_AOM_PORTS 1
+%define HAVE_FEXCEPT 1
+%define HAVE_PTHREAD_H 1
+%define HAVE_WXWIDGETS 0
+%define CONFIG_DEPENDENCY_TRACKING 1
+%define CONFIG_EXTERNAL_BUILD 1
+%define CONFIG_INSTALL_DOCS 1
+%define CONFIG_INSTALL_BINS 1
+%define CONFIG_INSTALL_LIBS 1
+%define CONFIG_INSTALL_SRCS 0
+%define CONFIG_DEBUG 0
+%define CONFIG_GPROF 0
+%define CONFIG_GCOV 0
+%define CONFIG_RVCT 0
+%define CONFIG_GCC 1
+%define CONFIG_MSVS 0
+%define CONFIG_PIC 1
+%define CONFIG_BIG_ENDIAN 0
+%define CONFIG_CODEC_SRCS 0
+%define CONFIG_DEBUG_LIBS 0
+%define CONFIG_RUNTIME_CPU_DETECT 1
+%define CONFIG_POSTPROC 1
+%define CONFIG_MULTITHREAD 1
+%define CONFIG_INTERNAL_STATS 0
+%define CONFIG_AV1_ENCODER 1
+%define CONFIG_AV1_DECODER 1
+%define CONFIG_AV1 1
+%define CONFIG_ENCODERS 1
+%define CONFIG_DECODERS 1
+%define CONFIG_STATIC_MSVCRT 0
+%define CONFIG_SPATIAL_RESAMPLING 1
+%define CONFIG_REALTIME_ONLY 0
+%define CONFIG_ONTHEFLY_BITPACKING 0
+%define CONFIG_ERROR_CONCEALMENT 0
+%define CONFIG_SHARED 0
+%define CONFIG_STATIC 1
+%define CONFIG_SMALL 0
+%define CONFIG_POSTPROC_VISUALIZER 0
+%define CONFIG_OS_SUPPORT 1
+%define CONFIG_UNIT_TESTS 0
+%define CONFIG_WEBM_IO 1
+%define CONFIG_LIBYUV 1
+%define CONFIG_ACCOUNTING 0
+%define CONFIG_INSPECTION 0
+%define CONFIG_DECODE_PERF_TESTS 0
+%define CONFIG_ENCODE_PERF_TESTS 0
+%define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+%define CONFIG_LOWBITDEPTH 1
+%define CONFIG_HIGHBITDEPTH 0
+%define CONFIG_EXPERIMENTAL 0
+%define CONFIG_SIZE_LIMIT 1
+%define CONFIG_FP_MB_STATS 0
+%define CONFIG_CDEF 1
+%define CONFIG_VAR_TX 0
+%define CONFIG_RECT_TX 1
+%define CONFIG_REF_MV 1
+%define CONFIG_TPL_MV 0
+%define CONFIG_DUAL_FILTER 1
+%define CONFIG_CONVOLVE_ROUND 0
+%define CONFIG_COMPOUND_ROUND 0
+%define CONFIG_EXT_TX 0
+%define CONFIG_TX64X64 0
+%define CONFIG_SUB8X8_MC 0
+%define CONFIG_EXT_INTRA 1
+%define CONFIG_INTRA_INTERP 0
+%define CONFIG_FILTER_INTRA 0
+%define CONFIG_INTRABC 0
+%define CONFIG_EXT_INTER 0
+%define CONFIG_INTERINTRA 0
+%define CONFIG_WEDGE 0
+%define CONFIG_COMPOUND_SEGMENT 0
+%define CONFIG_EXT_REFS 1
+%define CONFIG_GLOBAL_MOTION 1
+%define CONFIG_NEW_QUANT 0
+%define CONFIG_SUPERTX 0
+%define CONFIG_ANS 0
+%define CONFIG_EC_MULTISYMBOL 1
+%define CONFIG_NEW_TOKENSET 1
+%define CONFIG_LOOP_RESTORATION 0
+%define CONFIG_EXT_PARTITION 0
+%define CONFIG_EXT_PARTITION_TYPES 0
+%define CONFIG_UNPOISON_PARTITION_CTX 0
+%define CONFIG_EXT_TILE 0
+%define CONFIG_MOTION_VAR 1
+%define CONFIG_NCOBMC 0
+%define CONFIG_WARPED_MOTION 1
+%define CONFIG_Q_ADAPT_PROBS 0
+%define CONFIG_SUBFRAME_PROB_UPDATE 0
+%define CONFIG_BITSTREAM_DEBUG 0
+%define CONFIG_ALT_INTRA 1
+%define CONFIG_PALETTE 1
+%define CONFIG_PALETTE_DELTA_ENCODING 0
+%define CONFIG_DAALA_EC 1
+%define CONFIG_RAWBITS 0
+%define CONFIG_EC_SMALLMUL 0
+%define CONFIG_PVQ 0
+%define CONFIG_CFL 0
+%define CONFIG_XIPHRC 0
+%define CONFIG_CB4X4 1
+%define CONFIG_CHROMA_2X2 0
+%define CONFIG_CHROMA_SUB8X8 0
+%define CONFIG_FRAME_SIZE 0
+%define CONFIG_DELTA_Q 1
+%define CONFIG_EXT_DELTA_Q 1
+%define CONFIG_ADAPT_SCAN 0
+%define CONFIG_FILTER_7BIT 1
+%define CONFIG_PARALLEL_DEBLOCKING 0
+%define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+%define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+%define CONFIG_TILE_GROUPS 1
+%define CONFIG_EC_ADAPT 1
+%define CONFIG_TEMPMV_SIGNALING 0
+%define CONFIG_RD_DEBUG 0
+%define CONFIG_REFERENCE_BUFFER 1
+%define CONFIG_COEF_INTERLEAVE 0
+%define CONFIG_ENTROPY_STATS 0
+%define CONFIG_MASKED_TX 0
+%define CONFIG_DEPENDENT_HORZTILES 0
+%define CONFIG_DAALA_DIST 0
+%define CONFIG_TRIPRED 0
+%define CONFIG_PALETTE_THROUGHPUT 1
+%define CONFIG_REF_ADAPT 0
+%define CONFIG_LV_MAP 0
+%define CONFIG_TXK_SEL 0
+%define CONFIG_MV_COMPRESS 1
+%define CONFIG_FRAME_SUPERRES 0
+%define CONFIG_NEW_MULTISYMBOL 0
+%define CONFIG_COMPOUND_SINGLEREF 0
+%define CONFIG_AOM_QM 0
+%define CONFIG_LOWDELAY_COMPOUND 0
+%define CONFIG_ANALYZER 0
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/x64/aom_config.c
@@ -0,0 +1,11 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+#include "aom/aom_codec.h"
+static const char* const cfg = "--target=x86_64-linux-gcc --enable-external-build --disable-examples --disable-docs --disable-unit-tests --size-limit=8192x4608 --enable-pic --enable-postproc --as=yasm";
+const char *aom_codec_build_config(void) {return cfg;}
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/x64/aom_config.h
@@ -0,0 +1,167 @@
+/* Copyright (c) 2016, 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 */
+/* was not distributed with this source code in the LICENSE file, you can */
+/* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
+/* Media Patent License 1.0 was not distributed with this source code in the */
+/* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef AOM_CONFIG_H
+#define AOM_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 1
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 1
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_AOM_PORTS 1
+#define HAVE_FEXCEPT 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_WXWIDGETS 0
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 1
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_POSTPROC 1
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_AV1_ENCODER 1
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 0
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_HIGHBITDEPTH 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_CDEF 1
+#define CONFIG_VAR_TX 0
+#define CONFIG_RECT_TX 1
+#define CONFIG_REF_MV 1
+#define CONFIG_TPL_MV 0
+#define CONFIG_DUAL_FILTER 1
+#define CONFIG_CONVOLVE_ROUND 0
+#define CONFIG_COMPOUND_ROUND 0
+#define CONFIG_EXT_TX 0
+#define CONFIG_TX64X64 0
+#define CONFIG_SUB8X8_MC 0
+#define CONFIG_EXT_INTRA 1
+#define CONFIG_INTRA_INTERP 0
+#define CONFIG_FILTER_INTRA 0
+#define CONFIG_INTRABC 0
+#define CONFIG_EXT_INTER 0
+#define CONFIG_INTERINTRA 0
+#define CONFIG_WEDGE 0
+#define CONFIG_COMPOUND_SEGMENT 0
+#define CONFIG_EXT_REFS 1
+#define CONFIG_GLOBAL_MOTION 1
+#define CONFIG_NEW_QUANT 0
+#define CONFIG_SUPERTX 0
+#define CONFIG_ANS 0
+#define CONFIG_EC_MULTISYMBOL 1
+#define CONFIG_NEW_TOKENSET 1
+#define CONFIG_LOOP_RESTORATION 0
+#define CONFIG_EXT_PARTITION 0
+#define CONFIG_EXT_PARTITION_TYPES 0
+#define CONFIG_UNPOISON_PARTITION_CTX 0
+#define CONFIG_EXT_TILE 0
+#define CONFIG_MOTION_VAR 1
+#define CONFIG_NCOBMC 0
+#define CONFIG_WARPED_MOTION 1
+#define CONFIG_Q_ADAPT_PROBS 0
+#define CONFIG_SUBFRAME_PROB_UPDATE 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_ALT_INTRA 1
+#define CONFIG_PALETTE 1
+#define CONFIG_PALETTE_DELTA_ENCODING 0
+#define CONFIG_DAALA_EC 1
+#define CONFIG_RAWBITS 0
+#define CONFIG_EC_SMALLMUL 0
+#define CONFIG_PVQ 0
+#define CONFIG_CFL 0
+#define CONFIG_XIPHRC 0
+#define CONFIG_CB4X4 1
+#define CONFIG_CHROMA_2X2 0
+#define CONFIG_CHROMA_SUB8X8 0
+#define CONFIG_FRAME_SIZE 0
+#define CONFIG_DELTA_Q 1
+#define CONFIG_EXT_DELTA_Q 1
+#define CONFIG_ADAPT_SCAN 0
+#define CONFIG_FILTER_7BIT 1
+#define CONFIG_PARALLEL_DEBLOCKING 0
+#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+#define CONFIG_TILE_GROUPS 1
+#define CONFIG_EC_ADAPT 1
+#define CONFIG_TEMPMV_SIGNALING 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REFERENCE_BUFFER 1
+#define CONFIG_COEF_INTERLEAVE 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_MASKED_TX 0
+#define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DAALA_DIST 0
+#define CONFIG_TRIPRED 0
+#define CONFIG_PALETTE_THROUGHPUT 1
+#define CONFIG_REF_ADAPT 0
+#define CONFIG_LV_MAP 0
+#define CONFIG_TXK_SEL 0
+#define CONFIG_MV_COMPRESS 1
+#define CONFIG_FRAME_SUPERRES 0
+#define CONFIG_NEW_MULTISYMBOL 0
+#define CONFIG_COMPOUND_SINGLEREF 0
+#define CONFIG_AOM_QM 0
+#define CONFIG_LOWDELAY_COMPOUND 0
+#define CONFIG_ANALYZER 0
+#define DECODE_WIDTH_LIMIT 8192
+#define DECODE_HEIGHT_LIMIT 4608
+#endif /* AOM_CONFIG_H */
new file mode 100644
--- /dev/null
+++ b/media/libaom/config/linux/x64/aom_dsp_rtcd.h
@@ -0,0 +1,1434 @@
+#ifndef AOM_DSP_RTCD_H_
+#define AOM_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/enums.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int aom_avg_4x4_c(const uint8_t *, int p);
+unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
+#define aom_avg_4x4 aom_avg_4x4_sse2
+
+unsigned int aom_avg_8x8_c(const uint8_t *, int p);
+unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
+#define aom_avg_8x8 aom_avg_8x8_sse2
+
+void aom_blend_a64_hmask_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 h, int w);
+void aom_blend_a64_hmask_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 h, int w);
+RTCD_EXTERN void (*aom_blend_a64_hmask)(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 h, int w);
+
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+void aom_blend_a64_mask_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, uint32_t mask_stride, int h, int w, int suby, int subx);
+RTCD_EXTERN void (*aom_blend_a64_mask)(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, uint32_t mask_stride, int h, int w, int suby, int subx);
+
+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 h, int w);
+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 h, int w);
+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 h, int w);
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_pred aom_comp_avg_pred_c
+
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
+
+void aom_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);
+void aom_convolve8_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_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);
+void aom_convolve8_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)(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_avg_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_avg_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_avg_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);
+RTCD_EXTERN void (*aom_convolve8_avg)(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_avg_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_avg_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_avg_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);
+RTCD_EXTERN void (*aom_convolve8_avg_horiz)(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_avg_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);
+void aom_convolve8_avg_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);
+void aom_convolve8_avg_vert_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);
+RTCD_EXTERN void (*aom_convolve8_avg_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_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);
+void aom_convolve8_horiz_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_horiz)(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_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);
+void aom_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);
+void aom_convolve8_vert_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);
+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_avg_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_convolve_avg_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);
+#define aom_convolve_avg aom_convolve_avg_sse2
+
+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 x_step_q4, const int16_t *filter_y, int y_step_q4, 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 x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve_copy aom_convolve_copy_sse2
+
+void aom_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
+
+void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
+
+void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
+
+void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
+
+void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
+
+void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
+
+void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
+
+void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
+
+void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x4 aom_d135_predictor_4x4_c
+
+void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
+
+void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
+
+void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
+
+void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
+
+void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
+
+void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
+
+void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
+
+void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
+
+void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
+
+void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
+
+void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
+
+void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
+
+void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
+
+void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
+
+void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
+
+void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_d63e_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_d63e_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
+
+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);
+#define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x32 aom_dc_128_predictor_32x32_sse2
+
+void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_sse2
+
+void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_sse2
+
+void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x32 aom_dc_left_predictor_32x32_sse2
+
+void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_sse2
+
+void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_sse2
+
+void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x16 aom_dc_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x32 aom_dc_predictor_32x32_sse2
+
+void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x4 aom_dc_predictor_4x4_sse2
+
+void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x8 aom_dc_predictor_8x8_sse2
+
+void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x32 aom_dc_top_predictor_32x32_sse2
+
+void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_sse2
+
+void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_sse2
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct16x16 aom_fdct16x16_sse2
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct16x16_1_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct16x16_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_1_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32_1)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct32x32_rd_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4 aom_fdct4x4_sse2
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct4x4_1 aom_fdct4x4_1_sse2
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_ssse3(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*aom_fdct8x8)(const int16_t *input, tran_low_t *output, int stride);
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_fdct8x8_1 aom_fdct8x8_1_sse2
+
+void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get16x16var_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*aom_get16x16var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
+#define aom_get4x4sse_cs aom_get4x4sse_cs_c
+
+void aom_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void aom_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_get8x8var aom_get8x8var_sse2
+
+unsigned int aom_get_mb_ss_c(const int16_t *);
+unsigned int aom_get_mb_ss_sse2(const int16_t *);
+#define aom_get_mb_ss aom_get_mb_ss_sse2
+
+void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x16 aom_h_predictor_16x16_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_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x32 aom_h_predictor_32x32_sse2
+
+void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x4 aom_h_predictor_4x4_sse2
+
+void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x8 aom_h_predictor_8x8_sse2
+
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define aom_hadamard_16x16 aom_hadamard_16x16_sse2
+
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void aom_hadamard_8x8_ssse3(const int16_t *src_diff, int src_stride, int16_t *coeff);
+RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, int src_stride, int16_t *coeff);
+
+void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_10_add aom_idct16x16_10_add_sse2
+
+void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_1_add aom_idct16x16_1_add_sse2
+
+void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_256_add aom_idct16x16_256_add_sse2
+
+void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct16x16_38_add aom_idct16x16_38_add_c
+
+void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct32x32_1_add aom_idct32x32_1_add_sse2
+
+void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct4x4_16_add aom_idct4x4_16_add_sse2
+
+void aom_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct4x4_1_add aom_idct4x4_1_add_sse2
+
+void aom_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_12_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+void aom_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_idct8x8_1_add aom_idct8x8_1_add_sse2
+
+void aom_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct8x8_64_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
+
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width);
+int16_t aom_int_pro_col_sse2(const uint8_t *ref, int width);
+#define aom_int_pro_col aom_int_pro_col_sse2
+
+void aom_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, int ref_stride, int height);
+#define aom_int_pro_row aom_int_pro_row_sse2
+
+void aom_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_16_add aom_iwht4x4_16_add_sse2
+
+void aom_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define aom_iwht4x4_1_add aom_iwht4x4_1_add_c
+
+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
+
+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);
+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_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
+
+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);
+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_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_edge_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_horizontal_edge_8_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*aom_lpf_horizontal_edge_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void aom_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_16 aom_lpf_vertical_16_sse2
+
+void aom_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void aom_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_vertical_16_dual aom_lpf_vertical_16_dual_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
+
+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);
+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_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
+
+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);
+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_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_minmax_8x8 aom_minmax_8x8_sse2
+
+unsigned int aom_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int aom_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse16x8 aom_mse16x8_sse2
+
+unsigned int aom_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x16 aom_mse8x16_sse2
+
+unsigned int aom_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_mse8x8 aom_mse8x8_sse2
+
+unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+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_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_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_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_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_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*aom_quantize_b_32x32)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_quantize_b_64x64 aom_quantize_b_64x64_c
+
+unsigned int aom_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x16 aom_sad16x16_sse2
+
+unsigned int aom_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x16_avg aom_sad16x16_avg_sse2
+
+void aom_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x16x4d aom_sad16x16x4d_sse2
+
+void aom_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x16x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x32 aom_sad16x32_sse2
+
+unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x32_avg aom_sad16x32_avg_sse2
+
+void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x32x4d aom_sad16x32x4d_sse2
+
+unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x8 aom_sad16x8_sse2
+
+unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x8_avg aom_sad16x8_avg_sse2
+
+void aom_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x8x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x8x4d aom_sad16x8x4d_sse2
+
+void aom_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad16x8x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad16x8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x16x4d aom_sad32x16x4d_sse2
+
+unsigned int aom_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x3 aom_sad32x32x3_c
+
+void aom_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad32x32x8 aom_sad32x32x8_c
+
+unsigned int aom_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x4 aom_sad4x4_sse2
+
+unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x4_avg aom_sad4x4_avg_sse2
+
+void aom_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x4x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x4x4d aom_sad4x4x4d_sse2
+
+void aom_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad4x4x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad4x4x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x8 aom_sad4x8_sse2
+
+unsigned int aom_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x8_avg aom_sad4x8_avg_sse2
+
+void aom_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x4d aom_sad4x8x4d_sse2
+
+void aom_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad4x8x8 aom_sad4x8x8_c
+
+unsigned int aom_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x3 aom_sad64x64x3_c
+
+void aom_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad64x64x8 aom_sad64x64x8_c
+
+unsigned int aom_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x16 aom_sad8x16_sse2
+
+unsigned int aom_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x16_avg aom_sad8x16_avg_sse2
+
+void aom_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x16x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x16x4d aom_sad8x16x4d_sse2
+
+void aom_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x16x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x4 aom_sad8x4_sse2
+
+unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x4_avg aom_sad8x4_avg_sse2
+
+void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x4d aom_sad8x4x4d_sse2
+
+void aom_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_sad8x4x8 aom_sad8x4x8_c
+
+unsigned int aom_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x8 aom_sad8x8_sse2
+
+unsigned int aom_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x8_avg aom_sad8x8_avg_sse2
+
+void aom_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x8x3)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+void aom_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x8x4d aom_sad8x8x4d_sse2
+
+void aom_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void aom_sad8x8x8_sse4_1(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_sad8x8x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+
+int aom_satd_c(const int16_t *coeff, int length);
+int aom_satd_sse2(const int16_t *coeff, int length);
+#define aom_satd aom_satd_sse2
+
+void aom_scaled_2d_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_scaled_2d_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);
+RTCD_EXTERN void (*aom_scaled_2d)(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_scaled_avg_2d_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_scaled_avg_2d aom_scaled_avg_2d_c
+
+void aom_scaled_avg_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_scaled_avg_horiz aom_scaled_avg_horiz_c
+
+void aom_scaled_avg_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_scaled_avg_vert aom_scaled_avg_vert_c
+
+void aom_scaled_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_scaled_horiz aom_scaled_horiz_c
+
+void aom_scaled_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_scaled_vert aom_scaled_vert_c
+
+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_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_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_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_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
+
+uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoff