Update aom to commit id f5bdeac22930ff4c6b219be49c843db35970b918

4 files changed, 978 insertions, 332 deletions

diff --git a/media/libaom/config/mac/x64/aom_config.asm b/media/libaom/config/mac/x64/aom_config.asm
index fec4de39c..2821f150f 100644
--- a/media/libaom/config/mac/x64/aom_config.asm
+++ b/media/libaom/config/mac/x64/aom_config.asm
@@ -1,146 +1,163 @@
-%define ARCH_ARM 0
-%define ARCH_MIPS 0
-%define ARCH_X86 0
-%define ARCH_X86_64 1
-%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_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 1
-%define CONFIG_EXPERIMENTAL 0
-%define CONFIG_SIZE_LIMIT 1
-%define CONFIG_FP_MB_STATS 0
-%define CONFIG_CDEF 1
-%define CONFIG_VAR_TX 1
-%define CONFIG_RECT_TX 1
-%define CONFIG_RECT_TX_EXT 0
-%define CONFIG_TPL_MV 0
-%define CONFIG_DUAL_FILTER 1
-%define CONFIG_CONVOLVE_ROUND 0
-%define CONFIG_COMPOUND_ROUND 0
-%define CONFIG_EXT_TX 1
-%define CONFIG_DPCM_INTRA 0
-%define CONFIG_TX64X64 0
-%define CONFIG_EXT_INTRA 1
-%define CONFIG_INTRA_INTERP 0
-%define CONFIG_FILTER_INTRA 0
-%define CONFIG_INTRA_EDGE 0
-%define CONFIG_INTRABC 0
-%define CONFIG_EXT_INTER 1
-%define CONFIG_INTERINTRA 1
-%define CONFIG_WEDGE 1
-%define CONFIG_COMPOUND_SEGMENT 1
-%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_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_BITSTREAM_DEBUG 0
-%define CONFIG_ALT_INTRA 1
-%define CONFIG_PALETTE 1
-%define CONFIG_PALETTE_DELTA_ENCODING 0
-%define CONFIG_RAWBITS 0
-%define CONFIG_EC_SMALLMUL 1
-%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 1
-%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 1
-%define CONFIG_TILE_GROUPS 1
-%define CONFIG_EC_ADAPT 1
-%define CONFIG_TEMPMV_SIGNALING 1
-%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_ONE_SIDED_COMPOUND 1
-%define CONFIG_SMOOTH_HV 0
-%define CONFIG_VAR_REFS 0
-%define CONFIG_ANALYZER 0
+ARCH_ARM equ 0
+ARCH_MIPS equ 0
+ARCH_X86 equ 0
+ARCH_X86_64 equ 1
+HAVE_NEON equ 0
+HAVE_NEON_ASM equ 0
+HAVE_MIPS32 equ 0
+HAVE_DSPR2 equ 0
+HAVE_MSA equ 0
+HAVE_MIPS64 equ 0
+HAVE_MMX equ 1
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 1
+HAVE_SSSE3 equ 1
+HAVE_SSE4_1 equ 1
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_AOM_PORTS equ 1
+HAVE_FEXCEPT equ 1
+HAVE_PTHREAD_H equ 1
+HAVE_WXWIDGETS equ 0
+CONFIG_DEPENDENCY_TRACKING equ 1
+CONFIG_EXTERNAL_BUILD equ 1
+CONFIG_INSTALL_DOCS equ 1
+CONFIG_INSTALL_BINS equ 1
+CONFIG_INSTALL_LIBS equ 1
+CONFIG_INSTALL_SRCS equ 0
+CONFIG_DEBUG equ 0
+CONFIG_GPROF equ 0
+CONFIG_GCOV equ 0
+CONFIG_RVCT equ 0
+CONFIG_GCC equ 1
+CONFIG_MSVS equ 0
+CONFIG_PIC equ 1
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_CODEC_SRCS equ 0
+CONFIG_DEBUG_LIBS equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_POSTPROC equ 1
+CONFIG_MULTITHREAD equ 1
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_AV1_ENCODER equ 1
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1 equ 1
+CONFIG_STATIC_MSVCRT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_ONTHEFLY_BITPACKING equ 0
+CONFIG_ERROR_CONCEALMENT equ 0
+CONFIG_SHARED equ 0
+CONFIG_STATIC equ 1
+CONFIG_SMALL equ 0
+CONFIG_POSTPROC_VISUALIZER equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_UNIT_TESTS equ 0
+CONFIG_WEBM_IO equ 1
+CONFIG_LIBYUV equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_DECODE_PERF_TESTS equ 0
+CONFIG_ENCODE_PERF_TESTS equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_LOWBITDEPTH equ 1
+CONFIG_HIGHBITDEPTH equ 1
+CONFIG_EXPERIMENTAL equ 0
+CONFIG_SIZE_LIMIT equ 1
+CONFIG_COLORSPACE_HEADERS equ 0
+CONFIG_FP_MB_STATS equ 0
+CONFIG_CDEF equ 1
+CONFIG_VAR_TX equ 1
+CONFIG_RECT_TX equ 1
+CONFIG_RECT_TX_EXT equ 0
+CONFIG_TPL_MV equ 0
+CONFIG_DUAL_FILTER equ 1
+CONFIG_CONVOLVE_ROUND equ 0
+CONFIG_COMPOUND_ROUND equ 0
+CONFIG_EXT_TX equ 1
+CONFIG_DPCM_INTRA equ 0
+CONFIG_TX64X64 equ 0
+CONFIG_EXT_INTRA equ 1
+CONFIG_INTRA_INTERP equ 0
+CONFIG_FILTER_INTRA equ 0
+CONFIG_INTRA_EDGE equ 0
+CONFIG_INTRABC equ 0
+CONFIG_EXT_INTER equ 1
+CONFIG_INTERINTRA equ 1
+CONFIG_WEDGE equ 1
+CONFIG_COMPOUND_SEGMENT equ 1
+CONFIG_EXT_REFS equ 1
+CONFIG_ALTREF2 equ 0
+CONFIG_SPEED_REFS equ 0
+CONFIG_GF_GROUPS equ 0
+CONFIG_FLEX_REFS equ 0
+CONFIG_GLOBAL_MOTION equ 1
+CONFIG_NEW_QUANT equ 0
+CONFIG_SUPERTX equ 0
+CONFIG_ANS equ 0
+CONFIG_LOOP_RESTORATION equ 0
+CONFIG_EXT_PARTITION equ 0
+CONFIG_EXT_PARTITION_TYPES equ 0
+CONFIG_UNPOISON_PARTITION_CTX equ 0
+CONFIG_EXT_TILE equ 0
+CONFIG_MOTION_VAR equ 1
+CONFIG_NCOBMC equ 0
+CONFIG_WARPED_MOTION equ 1
+CONFIG_Q_ADAPT_PROBS equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_ALT_INTRA equ 1
+CONFIG_PALETTE equ 1
+CONFIG_PALETTE_DELTA_ENCODING equ 0
+CONFIG_RAWBITS equ 0
+CONFIG_EC_SMALLMUL equ 1
+CONFIG_PVQ equ 0
+CONFIG_CFL equ 0
+CONFIG_XIPHRC equ 0
+CONFIG_DCT_ONLY equ 0
+CONFIG_DAALA_DCT4 equ 0
+CONFIG_DAALA_DCT8 equ 0
+CONFIG_CB4X4 equ 1
+CONFIG_CHROMA_2X2 equ 0
+CONFIG_CHROMA_SUB8X8 equ 1
+CONFIG_FRAME_SIZE equ 0
+CONFIG_DELTA_Q equ 1
+CONFIG_EXT_DELTA_Q equ 1
+CONFIG_ADAPT_SCAN equ 0
+CONFIG_FILTER_7BIT equ 1
+CONFIG_PARALLEL_DEBLOCKING equ 1
+CONFIG_LOOPFILTERING_ACROSS_TILES equ 1
+CONFIG_TEMPMV_SIGNALING equ 1
+CONFIG_RD_DEBUG equ 0
+CONFIG_REFERENCE_BUFFER equ 1
+CONFIG_COEF_INTERLEAVE equ 0
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_MASKED_TX equ 0
+CONFIG_DEPENDENT_HORZTILES equ 0
+CONFIG_DIST_8X8 equ 0
+CONFIG_DAALA_DIST equ 0
+CONFIG_TRIPRED equ 0
+CONFIG_PALETTE_THROUGHPUT equ 1
+CONFIG_REF_ADAPT equ 0
+CONFIG_LV_MAP equ 0
+CONFIG_TXK_SEL equ 0
+CONFIG_MV_COMPRESS equ 1
+CONFIG_FRAME_SUPERRES equ 0
+CONFIG_NEW_MULTISYMBOL equ 0
+CONFIG_COMPOUND_SINGLEREF equ 0
+CONFIG_AOM_QM equ 0
+CONFIG_ONE_SIDED_COMPOUND equ 1
+CONFIG_EXT_COMP_REFS equ 0
+CONFIG_SMOOTH_HV equ 1
+CONFIG_VAR_REFS equ 0
+CONFIG_RECT_INTRA_PRED equ 1
+CONFIG_LGT equ 0
+CONFIG_SBL_SYMBOL equ 0
+CONFIG_NCOBMC_ADAPT_WEIGHT equ 0
+CONFIG_BGSPRITE equ 0
+CONFIG_VAR_TX_NO_TX_MODE equ 0
+CONFIG_MRC_TX equ 0
+CONFIG_LPF_DIRECT equ 0
+CONFIG_UV_LVL equ 0
+CONFIG_ANALYZER equ 0
diff --git a/media/libaom/config/mac/x64/aom_config.h b/media/libaom/config/mac/x64/aom_config.h
index 31e1af025..f9f2caadd 100644
--- a/media/libaom/config/mac/x64/aom_config.h
+++ b/media/libaom/config/mac/x64/aom_config.h
@@ -9,7 +9,7 @@
 /* This file automatically generated by configure. Do not edit! */
 #ifndef AOM_CONFIG_H
 #define AOM_CONFIG_H
-#define RESTRICT    
+#define RESTRICT
 #define INLINE      inline
 #define ARCH_ARM 0
 #define ARCH_MIPS 0
@@ -78,6 +78,7 @@
 #define CONFIG_HIGHBITDEPTH 1
 #define CONFIG_EXPERIMENTAL 0
 #define CONFIG_SIZE_LIMIT 1
+#define CONFIG_COLORSPACE_HEADERS 0
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_CDEF 1
 #define CONFIG_VAR_TX 1
@@ -100,6 +101,10 @@
 #define CONFIG_WEDGE 1
 #define CONFIG_COMPOUND_SEGMENT 1
 #define CONFIG_EXT_REFS 1
+#define CONFIG_ALTREF2 0
+#define CONFIG_SPEED_REFS 0
+#define CONFIG_GF_GROUPS 0
+#define CONFIG_FLEX_REFS 0
 #define CONFIG_GLOBAL_MOTION 1
 #define CONFIG_NEW_QUANT 0
 #define CONFIG_SUPERTX 0
@@ -114,6 +119,7 @@
 #define CONFIG_WARPED_MOTION 1
 #define CONFIG_Q_ADAPT_PROBS 0
 #define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_ALT_INTRA 1
 #define CONFIG_PALETTE 1
 #define CONFIG_PALETTE_DELTA_ENCODING 0
@@ -122,6 +128,9 @@
 #define CONFIG_PVQ 0
 #define CONFIG_CFL 0
 #define CONFIG_XIPHRC 0
+#define CONFIG_DCT_ONLY 0
+#define CONFIG_DAALA_DCT4 0
+#define CONFIG_DAALA_DCT8 0
 #define CONFIG_CB4X4 1
 #define CONFIG_CHROMA_2X2 0
 #define CONFIG_CHROMA_SUB8X8 1
@@ -130,11 +139,8 @@
 #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_PARALLEL_DEBLOCKING 1
 #define CONFIG_LOOPFILTERING_ACROSS_TILES 1
-#define CONFIG_TILE_GROUPS 1
-#define CONFIG_EC_ADAPT 1
 #define CONFIG_TEMPMV_SIGNALING 1
 #define CONFIG_RD_DEBUG 0
 #define CONFIG_REFERENCE_BUFFER 1
@@ -142,6 +148,7 @@
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_MASKED_TX 0
 #define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DIST_8X8 0
 #define CONFIG_DAALA_DIST 0
 #define CONFIG_TRIPRED 0
 #define CONFIG_PALETTE_THROUGHPUT 1
@@ -154,8 +161,18 @@
 #define CONFIG_COMPOUND_SINGLEREF 0
 #define CONFIG_AOM_QM 0
 #define CONFIG_ONE_SIDED_COMPOUND 1
-#define CONFIG_SMOOTH_HV 0
+#define CONFIG_EXT_COMP_REFS 0
+#define CONFIG_SMOOTH_HV 1
 #define CONFIG_VAR_REFS 0
+#define CONFIG_RECT_INTRA_PRED 1
+#define CONFIG_LGT 0
+#define CONFIG_SBL_SYMBOL 0
+#define CONFIG_NCOBMC_ADAPT_WEIGHT 0
+#define CONFIG_BGSPRITE 0
+#define CONFIG_VAR_TX_NO_TX_MODE 0
+#define CONFIG_MRC_TX 0
+#define CONFIG_LPF_DIRECT 0
+#define CONFIG_UV_LVL 0
 #define CONFIG_ANALYZER 0
 #define DECODE_WIDTH_LIMIT 8192
 #define DECODE_HEIGHT_LIMIT 4608
diff --git a/media/libaom/config/mac/x64/aom_dsp_rtcd.h b/media/libaom/config/mac/x64/aom_dsp_rtcd.h
index 0bfb89de5..6ee856f1d 100644
--- a/media/libaom/config/mac/x64/aom_dsp_rtcd.h
+++ b/media/libaom/config/mac/x64/aom_dsp_rtcd.h
@@ -35,14 +35,14 @@ RTCD_EXTERN void (*aom_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const
 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);
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, 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, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
 
 void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_pred aom_comp_mask_pred_c
 
-void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_upsampled_pred aom_comp_mask_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);
@@ -61,23 +61,41 @@ void aom_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint
 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_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_horiz_scale aom_convolve8_avg_horiz_scale_c
+
+void aom_convolve8_avg_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_scale aom_convolve8_avg_scale_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);
 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_avg_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_vert_scale aom_convolve8_avg_vert_scale_c
+
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 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_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_horiz_scale aom_convolve8_horiz_scale_c
+
+void aom_convolve8_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_scale aom_convolve8_scale_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);
 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_convolve8_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_vert_scale aom_convolve8_vert_scale_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);
 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
@@ -89,30 +107,66 @@ void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *d
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x32 aom_d117_predictor_16x32_c
+
+void aom_d117_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x8 aom_d117_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x16 aom_d117_predictor_32x16_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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x8 aom_d117_predictor_4x8_c
+
+void aom_d117_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x16 aom_d117_predictor_8x16_c
+
+void aom_d117_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x4 aom_d117_predictor_8x4_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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x32 aom_d135_predictor_16x32_c
+
+void aom_d135_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x8 aom_d135_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x16 aom_d135_predictor_32x16_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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x8 aom_d135_predictor_4x8_c
+
+void aom_d135_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x16 aom_d135_predictor_8x16_c
+
+void aom_d135_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x4 aom_d135_predictor_8x4_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
 
@@ -120,9 +174,18 @@ void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x32 aom_d153_predictor_16x32_c
+
+void aom_d153_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x8 aom_d153_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_32x16 aom_d153_predictor_32x16_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);
@@ -131,6 +194,15 @@ void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_4x8 aom_d153_predictor_4x8_c
+
+void aom_d153_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x16 aom_d153_predictor_8x16_c
+
+void aom_d153_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x4 aom_d153_predictor_8x4_c
+
 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);
@@ -138,39 +210,84 @@ RTCD_EXTERN void (*aom_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, con
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x32 aom_d207e_predictor_16x32_c
+
+void aom_d207e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x8 aom_d207e_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x16 aom_d207e_predictor_32x16_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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x8 aom_d207e_predictor_4x8_c
+
+void aom_d207e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x16 aom_d207e_predictor_8x16_c
+
+void aom_d207e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x4 aom_d207e_predictor_8x4_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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x32 aom_d45e_predictor_16x32_c
+
+void aom_d45e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x8 aom_d45e_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x16 aom_d45e_predictor_32x16_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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x8 aom_d45e_predictor_4x8_c
+
+void aom_d45e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x16 aom_d45e_predictor_8x16_c
+
+void aom_d45e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x4 aom_d45e_predictor_8x4_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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x32 aom_d63e_predictor_16x32_c
+
+void aom_d63e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x8 aom_d63e_predictor_16x8_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_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x16 aom_d63e_predictor_32x16_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
 
@@ -178,6 +295,15 @@ void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_4x8 aom_d63e_predictor_4x8_c
+
+void aom_d63e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x16 aom_d63e_predictor_8x16_c
+
+void aom_d63e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x4 aom_d63e_predictor_8x4_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
 
@@ -185,9 +311,18 @@ void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
+
 void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
 
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c
+
 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
@@ -196,6 +331,15 @@ void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_c
+
 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
@@ -204,9 +348,18 @@ void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
+
 void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
 
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c
+
 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
@@ -215,6 +368,15 @@ void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_c
+
 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
@@ -223,9 +385,18 @@ void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *a
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
+
 void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
 
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c
+
 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
@@ -234,6 +405,15 @@ void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abo
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_c
+
 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
@@ -242,9 +422,18 @@ void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
+
 void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
 
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c
+
 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
@@ -253,6 +442,15 @@ void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_c
+
 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
@@ -261,21 +459,11 @@ 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);
@@ -294,10 +482,6 @@ 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);
@@ -318,9 +502,18 @@ void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_c
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_c
+
 void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_2x2 aom_h_predictor_2x2_c
 
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_c
+
 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
@@ -329,6 +522,15 @@ void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abov
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_c
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_c
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_c
+
 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
@@ -1238,14 +1440,14 @@ RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride
 void aom_highbd_comp_avg_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
 #define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
 
-void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
-void aom_highbd_comp_avg_upsampled_pred_sse2(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
+void aom_highbd_comp_avg_upsampled_pred_sse2(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
 #define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_sse2
 
 void aom_highbd_comp_mask_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_highbd_comp_mask_pred aom_highbd_comp_mask_pred_c
 
-void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, int bd);
 #define aom_highbd_comp_mask_upsampled_pred aom_highbd_comp_mask_upsampled_pred_c
 
 void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
@@ -1291,120 +1493,264 @@ RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_s
 void aom_highbd_d117_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_16x16 aom_highbd_d117_predictor_16x16_c
 
+void aom_highbd_d117_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x32 aom_highbd_d117_predictor_16x32_c
+
+void aom_highbd_d117_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x8 aom_highbd_d117_predictor_16x8_c
+
 void aom_highbd_d117_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_2x2 aom_highbd_d117_predictor_2x2_c
 
+void aom_highbd_d117_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_32x16 aom_highbd_d117_predictor_32x16_c
+
 void aom_highbd_d117_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_32x32 aom_highbd_d117_predictor_32x32_c
 
 void aom_highbd_d117_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_4x4 aom_highbd_d117_predictor_4x4_c
 
+void aom_highbd_d117_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_4x8 aom_highbd_d117_predictor_4x8_c
+
+void aom_highbd_d117_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x16 aom_highbd_d117_predictor_8x16_c
+
+void aom_highbd_d117_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x4 aom_highbd_d117_predictor_8x4_c
+
 void aom_highbd_d117_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_8x8 aom_highbd_d117_predictor_8x8_c
 
 void aom_highbd_d135_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_16x16 aom_highbd_d135_predictor_16x16_c
 
+void aom_highbd_d135_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x32 aom_highbd_d135_predictor_16x32_c
+
+void aom_highbd_d135_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x8 aom_highbd_d135_predictor_16x8_c
+
 void aom_highbd_d135_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_2x2 aom_highbd_d135_predictor_2x2_c
 
+void aom_highbd_d135_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_32x16 aom_highbd_d135_predictor_32x16_c
+
 void aom_highbd_d135_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_32x32 aom_highbd_d135_predictor_32x32_c
 
 void aom_highbd_d135_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_4x4 aom_highbd_d135_predictor_4x4_c
 
+void aom_highbd_d135_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_4x8 aom_highbd_d135_predictor_4x8_c
+
+void aom_highbd_d135_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x16 aom_highbd_d135_predictor_8x16_c
+
+void aom_highbd_d135_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x4 aom_highbd_d135_predictor_8x4_c
+
 void aom_highbd_d135_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_8x8 aom_highbd_d135_predictor_8x8_c
 
 void aom_highbd_d153_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_16x16 aom_highbd_d153_predictor_16x16_c
 
+void aom_highbd_d153_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x32 aom_highbd_d153_predictor_16x32_c
+
+void aom_highbd_d153_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x8 aom_highbd_d153_predictor_16x8_c
+
 void aom_highbd_d153_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_2x2 aom_highbd_d153_predictor_2x2_c
 
+void aom_highbd_d153_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_32x16 aom_highbd_d153_predictor_32x16_c
+
 void aom_highbd_d153_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_32x32 aom_highbd_d153_predictor_32x32_c
 
 void aom_highbd_d153_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_4x4 aom_highbd_d153_predictor_4x4_c
 
+void aom_highbd_d153_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_4x8 aom_highbd_d153_predictor_4x8_c
+
+void aom_highbd_d153_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x16 aom_highbd_d153_predictor_8x16_c
+
+void aom_highbd_d153_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x4 aom_highbd_d153_predictor_8x4_c
+
 void aom_highbd_d153_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_8x8 aom_highbd_d153_predictor_8x8_c
 
 void aom_highbd_d207e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_16x16 aom_highbd_d207e_predictor_16x16_c
 
+void aom_highbd_d207e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x32 aom_highbd_d207e_predictor_16x32_c
+
+void aom_highbd_d207e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x8 aom_highbd_d207e_predictor_16x8_c
+
 void aom_highbd_d207e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_2x2 aom_highbd_d207e_predictor_2x2_c
 
+void aom_highbd_d207e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_32x16 aom_highbd_d207e_predictor_32x16_c
+
 void aom_highbd_d207e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_32x32 aom_highbd_d207e_predictor_32x32_c
 
 void aom_highbd_d207e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_4x4 aom_highbd_d207e_predictor_4x4_c
 
+void aom_highbd_d207e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_4x8 aom_highbd_d207e_predictor_4x8_c
+
+void aom_highbd_d207e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x16 aom_highbd_d207e_predictor_8x16_c
+
+void aom_highbd_d207e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x4 aom_highbd_d207e_predictor_8x4_c
+
 void aom_highbd_d207e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_8x8 aom_highbd_d207e_predictor_8x8_c
 
 void aom_highbd_d45e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_16x16 aom_highbd_d45e_predictor_16x16_c
 
+void aom_highbd_d45e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x32 aom_highbd_d45e_predictor_16x32_c
+
+void aom_highbd_d45e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x8 aom_highbd_d45e_predictor_16x8_c
+
 void aom_highbd_d45e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_2x2 aom_highbd_d45e_predictor_2x2_c
 
+void aom_highbd_d45e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_32x16 aom_highbd_d45e_predictor_32x16_c
+
 void aom_highbd_d45e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_32x32 aom_highbd_d45e_predictor_32x32_c
 
 void aom_highbd_d45e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_4x4 aom_highbd_d45e_predictor_4x4_c
 
+void aom_highbd_d45e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_4x8 aom_highbd_d45e_predictor_4x8_c
+
+void aom_highbd_d45e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x16 aom_highbd_d45e_predictor_8x16_c
+
+void aom_highbd_d45e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x4 aom_highbd_d45e_predictor_8x4_c
+
 void aom_highbd_d45e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_8x8 aom_highbd_d45e_predictor_8x8_c
 
 void aom_highbd_d63e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_16x16 aom_highbd_d63e_predictor_16x16_c
 
+void aom_highbd_d63e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x32 aom_highbd_d63e_predictor_16x32_c
+
+void aom_highbd_d63e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x8 aom_highbd_d63e_predictor_16x8_c
+
 void aom_highbd_d63e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_2x2 aom_highbd_d63e_predictor_2x2_c
 
+void aom_highbd_d63e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_32x16 aom_highbd_d63e_predictor_32x16_c
+
 void aom_highbd_d63e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_32x32 aom_highbd_d63e_predictor_32x32_c
 
 void aom_highbd_d63e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_4x4 aom_highbd_d63e_predictor_4x4_c
 
+void aom_highbd_d63e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_4x8 aom_highbd_d63e_predictor_4x8_c
+
+void aom_highbd_d63e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x16 aom_highbd_d63e_predictor_8x16_c
+
+void aom_highbd_d63e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x4 aom_highbd_d63e_predictor_8x4_c
+
 void aom_highbd_d63e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_8x8 aom_highbd_d63e_predictor_8x8_c
 
 void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_c
 
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c
+
 void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c
 
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c
+
 void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_c
 
 void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_c
 
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_c
+
 void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_c
 
 void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_c
 
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c
+
 void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c
 
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c
+
 void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_c
 
 void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_c
 
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_c
+
 void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_c
 
@@ -1412,9 +1758,18 @@ void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const ui
 void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_sse2
 
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c
+
 void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c
 
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c
+
 void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_sse2
@@ -1423,6 +1778,15 @@ void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint
 void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_sse2
 
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_c
+
 void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_sse2
@@ -1430,15 +1794,33 @@ void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const u
 void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_c
 
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c
+
 void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c
 
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c
+
 void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_c
 
 void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_c
 
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_c
+
 void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_c
 
@@ -1446,16 +1828,10 @@ void aom_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride
 void aom_highbd_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct16x16 aom_highbd_fdct16x16_sse2
 
-void aom_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct16x16_1 aom_highbd_fdct16x16_1_c
-
 void aom_highbd_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
 void aom_highbd_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct32x32 aom_highbd_fdct32x32_sse2
 
-void aom_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct32x32_1 aom_highbd_fdct32x32_1_c
-
 void aom_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
 void aom_highbd_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct32x32_rd aom_highbd_fdct32x32_rd_sse2
@@ -1468,45 +1844,38 @@ void aom_highbd_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
 void aom_highbd_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct8x8 aom_highbd_fdct8x8_sse2
 
-void aom_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct8x8_1 aom_highbd_fdct8x8_1_c
-
 void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_c
 
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c
+
 void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c
 
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c
+
 void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_c
 
 void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_c
 
-void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c
-
-void aom_highbd_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct16x16_1_add aom_highbd_idct16x16_1_add_c
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_c
 
-void aom_highbd_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_1024_add aom_highbd_idct32x32_1024_add_c
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_c
 
-void aom_highbd_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_1_add aom_highbd_idct32x32_1_add_c
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_c
 
-void aom_highbd_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_34_add aom_highbd_idct32x32_34_add_c
-
-void aom_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-void aom_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct4x4_16_add aom_highbd_idct4x4_16_add_sse2
-
-void aom_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct4x4_1_add aom_highbd_idct4x4_1_add_c
-
-void aom_highbd_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct8x8_1_add aom_highbd_idct8x8_1_add_c
+void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c
 
 void aom_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define aom_highbd_iwht4x4_16_add aom_highbd_iwht4x4_16_add_c
@@ -1763,21 +2132,40 @@ RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance8x8)(const uint8_t *pre, int
 void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
 
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
 void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c
 
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
 void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
 
 void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
 
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
 void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
 
 void aom_highbd_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_highbd_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);
-#define aom_highbd_quantize_b aom_highbd_quantize_b_sse2
+void aom_highbd_quantize_b_avx2(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_highbd_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_highbd_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_highbd_quantize_b_32x32_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);
@@ -2010,36 +2398,129 @@ void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uin
 void aom_highbd_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x8x8 aom_highbd_sad8x8x8_c
 
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
 void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
 
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
 void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c
 
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
 void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
 
 void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
 
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
 void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
 
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
 void aom_highbd_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, int bd);
 void aom_highbd_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, int bd);
 #define aom_highbd_subtract_block aom_highbd_subtract_block_sse2
 
-void aom_highbd_upsampled_pred_c(uint16_t *comp_pred, int width, int height, const uint8_t *ref8, int ref_stride);
-void aom_highbd_upsampled_pred_sse2(uint16_t *comp_pred, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_upsampled_pred_c(uint16_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
+void aom_highbd_upsampled_pred_sse2(uint16_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
 #define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_sse2
 
 void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_sse2
 
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c
+
 void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c
 
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c
+
 void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_sse2
@@ -2048,6 +2529,15 @@ void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint1
 void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_sse2
 
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_c
+
 void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_sse2
@@ -2452,15 +2942,33 @@ RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_str
 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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+
 void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
 
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+
 void aom_paeth_predictor_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
 
@@ -2731,21 +3239,105 @@ void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
 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_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+
+void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+
 void aom_smooth_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_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
+
 void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
 
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c
+
 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_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x4 aom_smooth_predictor_8x4_c
+
 void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_8x8 aom_smooth_predictor_8x8_c
 
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+
+void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+
 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);
@@ -2892,17 +3484,26 @@ 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);
 #define aom_sum_squares_i16 aom_sum_squares_i16_sse2
 
-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);
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
+void aom_upsampled_pred_sse2(uint8_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
 #define aom_upsampled_pred aom_upsampled_pred_sse2
 
 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);
 #define aom_v_predictor_16x16 aom_v_predictor_16x16_sse2
 
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_c
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_c
+
 void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_2x2 aom_v_predictor_2x2_c
 
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x16 aom_v_predictor_32x16_c
+
 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);
 #define aom_v_predictor_32x32 aom_v_predictor_32x32_sse2
@@ -2911,6 +3512,15 @@ void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *abov
 void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_4x4 aom_v_predictor_4x4_sse2
 
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_c
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_c
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_c
+
 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);
 #define aom_v_predictor_8x8 aom_v_predictor_8x8_sse2
@@ -3038,12 +3648,8 @@ static void setup_rtcd_internal(void)
     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_fdct16x16_1 = aom_fdct16x16_1_sse2;
-    if (flags & HAS_AVX2) aom_fdct16x16_1 = aom_fdct16x16_1_avx2;
     aom_fdct32x32 = aom_fdct32x32_sse2;
     if (flags & HAS_AVX2) aom_fdct32x32 = aom_fdct32x32_avx2;
-    aom_fdct32x32_1 = aom_fdct32x32_1_sse2;
-    if (flags & HAS_AVX2) aom_fdct32x32_1 = aom_fdct32x32_1_avx2;
     aom_fdct32x32_rd = aom_fdct32x32_rd_sse2;
     if (flags & HAS_AVX2) aom_fdct32x32_rd = aom_fdct32x32_rd_avx2;
     aom_fdct8x8 = aom_fdct8x8_sse2;
@@ -3300,6 +3906,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_highbd_obmc_variance8x4 = aom_highbd_obmc_variance8x4_sse4_1;
     aom_highbd_obmc_variance8x8 = aom_highbd_obmc_variance8x8_c;
     if (flags & HAS_SSE4_1) aom_highbd_obmc_variance8x8 = aom_highbd_obmc_variance8x8_sse4_1;
+    aom_highbd_quantize_b = aom_highbd_quantize_b_sse2;
+    if (flags & HAS_AVX2) aom_highbd_quantize_b = aom_highbd_quantize_b_avx2;
     aom_highbd_sad16x16 = aom_highbd_sad16x16_sse2;
     if (flags & HAS_AVX2) aom_highbd_sad16x16 = aom_highbd_sad16x16_avx2;
     aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_sse2;
diff --git a/media/libaom/config/mac/x64/av1_rtcd.h b/media/libaom/config/mac/x64/av1_rtcd.h
index 9dcf96f84..d4fc99c10 100644
--- a/media/libaom/config/mac/x64/av1_rtcd.h
+++ b/media/libaom/config/mac/x64/av1_rtcd.h
@@ -12,6 +12,7 @@
  */
 
 #include "aom/aom_integer.h"
+#include "aom_dsp/txfm_common.h"
 #include "av1/common/common.h"
 #include "av1/common/enums.h"
 #include "av1/common/quant_common.h"
@@ -24,6 +25,7 @@ struct macroblockd;
 
 /* Encoder forward decls */
 struct macroblock;
+struct txfm_param;
 struct aom_variance_vtable;
 struct search_site_config;
 struct mv;
@@ -60,7 +62,8 @@ void aom_clpf_hblock_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride,
 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_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);
 
 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);
@@ -76,58 +79,58 @@ int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_conf
 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);
-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_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht16x16_avx2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+RTCD_EXTERN void (*av1_fht16x16)(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 
-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);
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht16x32_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x32 av1_fht16x32_sse2
 
-void av1_fht16x4_c(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, struct txfm_param *param);
 #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);
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht16x8_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x8 av1_fht16x8_sse2
 
-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);
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht32x16_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht32x16 av1_fht32x16_sse2
 
-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_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht32x32_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht32x32_avx2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+RTCD_EXTERN void (*av1_fht32x32)(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 
-void av1_fht32x8_c(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, struct txfm_param *param);
 #define av1_fht32x8 av1_fht32x8_c
 
-void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #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);
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht4x4 av1_fht4x4_sse2
 
-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);
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht4x8_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht4x8 av1_fht4x8_sse2
 
-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);
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht8x16_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x16 av1_fht8x16_sse2
 
-void av1_fht8x32_c(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, struct txfm_param *param);
 #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);
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht8x4_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x4 av1_fht8x4_sse2
 
-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);
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
+void av1_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x8 av1_fht8x8_sse2
 
 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);
@@ -145,6 +148,15 @@ void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, i
 void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_16x32 av1_fwd_txfm2d_16x32_c
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_16x8 av1_fwd_txfm2d_16x8_c
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_32x16 av1_fwd_txfm2d_32x16_c
+
 void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
@@ -153,10 +165,19 @@ void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, int
 void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_4x8 av1_fwd_txfm2d_4x8_c
+
 void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_8x16 av1_fwd_txfm2d_8x16_c
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_8x4 av1_fwd_txfm2d_8x4_c
+
 void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
@@ -210,89 +231,46 @@ void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride, uint16_t *d
 void av1_highbd_convolve_vert_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
 RTCD_EXTERN void (*av1_highbd_convolve_vert)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
 
-void av1_highbd_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x16 av1_highbd_fht16x16_c
-
-void av1_highbd_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x32 av1_highbd_fht16x32_c
-
-void av1_highbd_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x4 av1_highbd_fht16x4_c
-
-void av1_highbd_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x8 av1_highbd_fht16x8_c
-
-void av1_highbd_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x16 av1_highbd_fht32x16_c
-
-void av1_highbd_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x32 av1_highbd_fht32x32_c
-
-void av1_highbd_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x8 av1_highbd_fht32x8_c
-
-void av1_highbd_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht4x16 av1_highbd_fht4x16_c
-
-void av1_highbd_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-void av1_highbd_fht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-RTCD_EXTERN void (*av1_highbd_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-
-void av1_highbd_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht4x8 av1_highbd_fht4x8_c
-
-void av1_highbd_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x16 av1_highbd_fht8x16_c
-
-void av1_highbd_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x32 av1_highbd_fht8x32_c
-
-void av1_highbd_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x4 av1_highbd_fht8x4_c
-
-void av1_highbd_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x8 av1_highbd_fht8x8_c
-
 void av1_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define av1_highbd_fwht4x4 av1_highbd_fwht4x4_c
 
-void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd);
+void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 #define av1_highbd_iht16x16_256_add av1_highbd_iht16x16_256_add_c
 
-void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x32_512_add av1_highbd_iht16x32_512_add_c
 
-void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x4_64_add av1_highbd_iht16x4_64_add_c
 
-void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x8_128_add av1_highbd_iht16x8_128_add_c
 
-void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht32x16_512_add av1_highbd_iht32x16_512_add_c
 
-void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht32x8_256_add av1_highbd_iht32x8_256_add_c
 
-void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x16_64_add av1_highbd_iht4x16_64_add_c
 
-void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x4_16_add av1_highbd_iht4x4_16_add_c
 
-void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x8_32_add av1_highbd_iht4x8_32_add_c
 
-void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x16_128_add av1_highbd_iht8x16_128_add_c
 
-void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x32_256_add av1_highbd_iht8x32_256_add_c
 
-void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x4_32_add av1_highbd_iht8x4_32_add_c
 
-void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x8_64_add av1_highbd_iht8x8_64_add_c
 
 void av1_highbd_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);
@@ -300,71 +278,81 @@ void av1_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int
 
 void av1_highbd_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, int log_scale);
 void av1_highbd_quantize_fp_sse4_1(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);
+void av1_highbd_quantize_fp_avx2(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);
 RTCD_EXTERN void (*av1_highbd_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, int log_scale);
 
 void av1_highbd_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_highbd_temporal_filter_apply av1_highbd_temporal_filter_apply_c
 
-void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-void av1_highbd_warp_affine_ssse3(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_ssse3(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 
-void av1_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_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
+void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
+void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
+RTCD_EXTERN void (*av1_iht16x16_256_add)(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 
-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);
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht16x32_512_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht16x32_512_add av1_iht16x32_512_add_sse2
 
-void av1_iht16x4_64_add_c(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, const struct txfm_param *param);
 #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);
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht16x8_128_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht16x8_128_add av1_iht16x8_128_add_sse2
 
-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);
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht32x16_512_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht32x16_512_add av1_iht32x16_512_add_sse2
 
-void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 #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);
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #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);
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #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);
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht4x4_16_add av1_iht4x4_16_add_sse2
 
-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);
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht4x8_32_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht4x8_32_add av1_iht4x8_32_add_sse2
 
-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);
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht8x16_128_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x16_128_add av1_iht8x16_128_add_sse2
 
-void av1_iht8x32_256_add_c(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, const struct txfm_param *param);
 #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);
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht8x4_32_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x4_32_add av1_iht8x4_32_add_sse2
 
-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);
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
+void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x8_64_add av1_iht8x8_64_add_sse2
 
 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 void av1_inv_txfm2d_add_16x16_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_inv_txfm2d_add_16x16)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 void av1_inv_txfm2d_add_32x32_avx2(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
@@ -373,9 +361,18 @@ void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride
 void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
 
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
 void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
@@ -389,19 +386,21 @@ void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_b
 
 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);
-#define av1_quantize_fp av1_quantize_fp_sse2
+void av1_quantize_fp_avx2(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_quantize_fp_32x32_avx2(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_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 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);
 #define av1_temporal_filter_apply av1_temporal_filter_apply_sse2
 
-void av1_warp_affine_c(const int32_t *mat, const 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 comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-void av1_warp_affine_sse2(const int32_t *mat, const 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 comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-void av1_warp_affine_ssse3(const int32_t *mat, const 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 comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const 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 comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_c(const int32_t *mat, const 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, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_sse2(const int32_t *mat, const 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, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_ssse3(const int32_t *mat, const 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, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const 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, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 
 void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
 void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
@@ -495,6 +494,8 @@ static void setup_rtcd_internal(void)
     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_AVX2) av1_block_error = av1_block_error_avx2;
     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;
@@ -522,10 +523,9 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) av1_highbd_convolve_init = av1_highbd_convolve_init_sse4_1;
     av1_highbd_convolve_vert = av1_highbd_convolve_vert_c;
     if (flags & HAS_SSE4_1) av1_highbd_convolve_vert = av1_highbd_convolve_vert_sse4_1;
-    av1_highbd_fht4x4 = av1_highbd_fht4x4_c;
-    if (flags & HAS_SSE4_1) av1_highbd_fht4x4 = av1_highbd_fht4x4_sse4_1;
     av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
     if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
     av1_highbd_warp_affine = av1_highbd_warp_affine_c;
     if (flags & HAS_SSSE3) av1_highbd_warp_affine = av1_highbd_warp_affine_ssse3;
     av1_iht16x16_256_add = av1_iht16x16_256_add_sse2;
@@ -540,6 +540,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
     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_sse2;
+    if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
+    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
+    if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
     av1_warp_affine = av1_warp_affine_sse2;
     if (flags & HAS_SSSE3) av1_warp_affine = av1_warp_affine_ssse3;
     compute_cross_correlation = compute_cross_correlation_c;