Update aom to slightly newer commit ID

4 files changed, 2470 insertions, 82 deletions

diff --git a/media/libaom/config/win/x64/aom_config.asm b/media/libaom/config/win/x64/aom_config.asm
index 9cb036ce3..fbbc9882a 100644
--- a/media/libaom/config/win/x64/aom_config.asm
+++ b/media/libaom/config/win/x64/aom_config.asm
@@ -2,8 +2,6 @@
 %define ARCH_MIPS 0
 %define ARCH_X86 0
 %define ARCH_X86_64 1
-%define HAVE_EDSP 0
-%define HAVE_MEDIA 0
 %define HAVE_NEON 0
 %define HAVE_NEON_ASM 0
 %define HAVE_MIPS32 0
@@ -45,8 +43,6 @@
 %define CONFIG_AV1_ENCODER 1
 %define CONFIG_AV1_DECODER 1
 %define CONFIG_AV1 1
-%define CONFIG_ENCODERS 1
-%define CONFIG_DECODERS 1
 %define CONFIG_STATIC_MSVCRT 0
 %define CONFIG_SPATIAL_RESAMPLING 1
 %define CONFIG_REALTIME_ONLY 0
@@ -66,36 +62,35 @@
 %define CONFIG_ENCODE_PERF_TESTS 0
 %define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 %define CONFIG_LOWBITDEPTH 1
-%define CONFIG_HIGHBITDEPTH 0
+%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 0
+%define CONFIG_VAR_TX 1
 %define CONFIG_RECT_TX 1
-%define CONFIG_REF_MV 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 0
+%define CONFIG_EXT_TX 1
+%define CONFIG_DPCM_INTRA 0
 %define CONFIG_TX64X64 0
-%define CONFIG_SUB8X8_MC 0
 %define CONFIG_EXT_INTRA 1
 %define CONFIG_INTRA_INTERP 0
 %define CONFIG_FILTER_INTRA 0
+%define CONFIG_INTRA_EDGE 0
 %define CONFIG_INTRABC 0
-%define CONFIG_EXT_INTER 0
-%define CONFIG_INTERINTRA 0
-%define CONFIG_WEDGE 0
-%define CONFIG_COMPOUND_SEGMENT 0
+%define CONFIG_EXT_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_EC_MULTISYMBOL 1
-%define CONFIG_NEW_TOKENSET 1
 %define CONFIG_LOOP_RESTORATION 0
 %define CONFIG_EXT_PARTITION 0
 %define CONFIG_EXT_PARTITION_TYPES 0
@@ -105,20 +100,18 @@
 %define CONFIG_NCOBMC 0
 %define CONFIG_WARPED_MOTION 1
 %define CONFIG_Q_ADAPT_PROBS 0
-%define CONFIG_SUBFRAME_PROB_UPDATE 0
 %define CONFIG_BITSTREAM_DEBUG 0
 %define CONFIG_ALT_INTRA 1
 %define CONFIG_PALETTE 1
 %define CONFIG_PALETTE_DELTA_ENCODING 0
-%define CONFIG_DAALA_EC 1
 %define CONFIG_RAWBITS 0
-%define CONFIG_EC_SMALLMUL 0
+%define CONFIG_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 0
+%define CONFIG_CHROMA_SUB8X8 1
 %define CONFIG_FRAME_SIZE 0
 %define CONFIG_DELTA_Q 1
 %define CONFIG_EXT_DELTA_Q 1
@@ -126,10 +119,10 @@
 %define CONFIG_FILTER_7BIT 1
 %define CONFIG_PARALLEL_DEBLOCKING 0
 %define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
-%define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+%define CONFIG_LOOPFILTERING_ACROSS_TILES 1
 %define CONFIG_TILE_GROUPS 1
 %define CONFIG_EC_ADAPT 1
-%define CONFIG_TEMPMV_SIGNALING 0
+%define CONFIG_TEMPMV_SIGNALING 1
 %define CONFIG_RD_DEBUG 0
 %define CONFIG_REFERENCE_BUFFER 1
 %define CONFIG_COEF_INTERLEAVE 0
@@ -147,5 +140,7 @@
 %define CONFIG_NEW_MULTISYMBOL 0
 %define CONFIG_COMPOUND_SINGLEREF 0
 %define CONFIG_AOM_QM 0
-%define CONFIG_LOWDELAY_COMPOUND 0
+%define CONFIG_ONE_SIDED_COMPOUND 1
+%define CONFIG_SMOOTH_HV 0
+%define CONFIG_VAR_REFS 0
 %define CONFIG_ANALYZER 0
diff --git a/media/libaom/config/win/x64/aom_config.h b/media/libaom/config/win/x64/aom_config.h
index 92aee324f..aaeea05a9 100644
--- a/media/libaom/config/win/x64/aom_config.h
+++ b/media/libaom/config/win/x64/aom_config.h
@@ -15,8 +15,6 @@
 #define ARCH_MIPS 0
 #define ARCH_X86 0
 #define ARCH_X86_64 1
-#define HAVE_EDSP 0
-#define HAVE_MEDIA 0
 #define HAVE_NEON 0
 #define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 0
@@ -58,8 +56,6 @@
 #define CONFIG_AV1_ENCODER 1
 #define CONFIG_AV1_DECODER 1
 #define CONFIG_AV1 1
-#define CONFIG_ENCODERS 1
-#define CONFIG_DECODERS 1
 #define CONFIG_STATIC_MSVCRT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
 #define CONFIG_REALTIME_ONLY 0
@@ -79,36 +75,35 @@
 #define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_LOWBITDEPTH 1
-#define CONFIG_HIGHBITDEPTH 0
+#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 0
+#define CONFIG_VAR_TX 1
 #define CONFIG_RECT_TX 1
-#define CONFIG_REF_MV 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 0
+#define CONFIG_EXT_TX 1
+#define CONFIG_DPCM_INTRA 0
 #define CONFIG_TX64X64 0
-#define CONFIG_SUB8X8_MC 0
 #define CONFIG_EXT_INTRA 1
 #define CONFIG_INTRA_INTERP 0
 #define CONFIG_FILTER_INTRA 0
+#define CONFIG_INTRA_EDGE 0
 #define CONFIG_INTRABC 0
-#define CONFIG_EXT_INTER 0
-#define CONFIG_INTERINTRA 0
-#define CONFIG_WEDGE 0
-#define CONFIG_COMPOUND_SEGMENT 0
+#define CONFIG_EXT_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_EC_MULTISYMBOL 1
-#define CONFIG_NEW_TOKENSET 1
 #define CONFIG_LOOP_RESTORATION 0
 #define CONFIG_EXT_PARTITION 0
 #define CONFIG_EXT_PARTITION_TYPES 0
@@ -118,20 +113,18 @@
 #define CONFIG_NCOBMC 0
 #define CONFIG_WARPED_MOTION 1
 #define CONFIG_Q_ADAPT_PROBS 0
-#define CONFIG_SUBFRAME_PROB_UPDATE 0
 #define CONFIG_BITSTREAM_DEBUG 0
 #define CONFIG_ALT_INTRA 1
 #define CONFIG_PALETTE 1
 #define CONFIG_PALETTE_DELTA_ENCODING 0
-#define CONFIG_DAALA_EC 1
 #define CONFIG_RAWBITS 0
-#define CONFIG_EC_SMALLMUL 0
+#define CONFIG_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 0
+#define CONFIG_CHROMA_SUB8X8 1
 #define CONFIG_FRAME_SIZE 0
 #define CONFIG_DELTA_Q 1
 #define CONFIG_EXT_DELTA_Q 1
@@ -139,10 +132,10 @@
 #define CONFIG_FILTER_7BIT 1
 #define CONFIG_PARALLEL_DEBLOCKING 0
 #define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
-#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
+#define CONFIG_LOOPFILTERING_ACROSS_TILES 1
 #define CONFIG_TILE_GROUPS 1
 #define CONFIG_EC_ADAPT 1
-#define CONFIG_TEMPMV_SIGNALING 0
+#define CONFIG_TEMPMV_SIGNALING 1
 #define CONFIG_RD_DEBUG 0
 #define CONFIG_REFERENCE_BUFFER 1
 #define CONFIG_COEF_INTERLEAVE 0
@@ -160,7 +153,9 @@
 #define CONFIG_NEW_MULTISYMBOL 0
 #define CONFIG_COMPOUND_SINGLEREF 0
 #define CONFIG_AOM_QM 0
-#define CONFIG_LOWDELAY_COMPOUND 0
+#define CONFIG_ONE_SIDED_COMPOUND 1
+#define CONFIG_SMOOTH_HV 0
+#define CONFIG_VAR_REFS 0
 #define CONFIG_ANALYZER 0
 #define DECODE_WIDTH_LIMIT 8192
 #define DECODE_HEIGHT_LIMIT 4608
diff --git a/media/libaom/config/win/x64/aom_dsp_rtcd.h b/media/libaom/config/win/x64/aom_dsp_rtcd.h
index 6fda772d8..0bfb89de5 100644
--- a/media/libaom/config/win/x64/aom_dsp_rtcd.h
+++ b/media/libaom/config/win/x64/aom_dsp_rtcd.h
@@ -20,14 +20,6 @@
 extern "C" {
 #endif
 
-unsigned int aom_avg_4x4_c(const uint8_t *, int p);
-unsigned int aom_avg_4x4_sse2(const uint8_t *, int p);
-#define aom_avg_4x4 aom_avg_4x4_sse2
-
-unsigned int aom_avg_8x8_c(const uint8_t *, int p);
-unsigned int aom_avg_8x8_sse2(const uint8_t *, int p);
-#define aom_avg_8x8 aom_avg_8x8_sse2
-
 void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
@@ -47,6 +39,12 @@ void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int
 void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_sse2
 
+void aom_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);
+#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);
 void aom_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -344,38 +342,1756 @@ void aom_hadamard_8x8_sse2(const int16_t *src_diff, int src_stride, int16_t *coe
 void aom_hadamard_8x8_ssse3(const int16_t *src_diff, int src_stride, int16_t *coeff);
 RTCD_EXTERN void (*aom_hadamard_8x8)(const int16_t *src_diff, int src_stride, int16_t *coeff);
 
+void aom_highbd_10_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_10_get16x16var aom_highbd_10_get16x16var_c
+
+void aom_highbd_10_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_10_get8x8var aom_highbd_10_get8x8var_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+
+unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
+
+unsigned int aom_highbd_10_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x16 aom_highbd_10_mse8x16_c
+
+unsigned int aom_highbd_10_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_sse2
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_10_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_10_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_highbd_10_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);
+#define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x16 aom_highbd_10_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x32 aom_highbd_10_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x8 aom_highbd_10_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x16 aom_highbd_10_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x32 aom_highbd_10_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x64 aom_highbd_10_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_10_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x32 aom_highbd_10_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance64x64 aom_highbd_10_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x16 aom_highbd_10_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_10_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x8 aom_highbd_10_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance16x16 aom_highbd_10_variance16x16_sse2
+
+unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance16x32 aom_highbd_10_variance16x32_sse2
+
+unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance16x8 aom_highbd_10_variance16x8_sse2
+
+unsigned int aom_highbd_10_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance2x2 aom_highbd_10_variance2x2_c
+
+unsigned int aom_highbd_10_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance2x4 aom_highbd_10_variance2x4_c
+
+unsigned int aom_highbd_10_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance32x16 aom_highbd_10_variance32x16_sse2
+
+unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance32x32 aom_highbd_10_variance32x32_sse2
+
+unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance32x64 aom_highbd_10_variance32x64_sse2
+
+unsigned int aom_highbd_10_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance4x2 aom_highbd_10_variance4x2_c
+
+unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
+
+unsigned int aom_highbd_10_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance64x32 aom_highbd_10_variance64x32_sse2
+
+unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance64x64 aom_highbd_10_variance64x64_sse2
+
+unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance8x16 aom_highbd_10_variance8x16_sse2
+
+unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
+
+unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_10_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_10_variance8x8 aom_highbd_10_variance8x8_sse2
+
+void aom_highbd_12_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_12_get16x16var aom_highbd_12_get16x16var_c
+
+void aom_highbd_12_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_12_get8x8var aom_highbd_12_get8x8var_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_sse2
+
+unsigned int aom_highbd_12_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse16x8 aom_highbd_12_mse16x8_c
+
+unsigned int aom_highbd_12_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x16 aom_highbd_12_mse8x16_c
+
+unsigned int aom_highbd_12_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_sse2
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_12_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+uint32_t aom_highbd_12_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_highbd_12_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);
+#define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_12_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_12_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_sse2
+
+unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_sse2
+
+unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_sse2
+
+unsigned int aom_highbd_12_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance2x2 aom_highbd_12_variance2x2_c
+
+unsigned int aom_highbd_12_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance2x4 aom_highbd_12_variance2x4_c
+
+unsigned int aom_highbd_12_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_sse2
+
+unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_sse2
+
+unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_sse2
+
+unsigned int aom_highbd_12_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance4x2 aom_highbd_12_variance4x2_c
+
+unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_12_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
+
+unsigned int aom_highbd_12_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_sse2
+
+unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_sse2
+
+unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_sse2
+
+unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
+
+unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_12_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_sse2
+
+void aom_highbd_8_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_8_get16x16var aom_highbd_8_get16x16var_c
+
+void aom_highbd_8_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define aom_highbd_8_get8x8var aom_highbd_8_get8x8var_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_sse2
+
+unsigned int aom_highbd_8_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse16x8 aom_highbd_8_mse16x8_c
+
+unsigned int aom_highbd_8_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x16 aom_highbd_8_mse8x16_c
+
+unsigned int aom_highbd_8_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_sse2
+
+uint32_t aom_highbd_8_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_highbd_8_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);
+#define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+RTCD_EXTERN uint32_t (*aom_highbd_8_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+
+uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_sse2
+
+uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t aom_highbd_8_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_sse2
+
+unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_sse2
+
+unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_sse2
+
+unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_sse2
+
+unsigned int aom_highbd_8_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance2x2 aom_highbd_8_variance2x2_c
+
+unsigned int aom_highbd_8_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance2x4 aom_highbd_8_variance2x4_c
+
+unsigned int aom_highbd_8_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_sse2
+
+unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_sse2
+
+unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_sse2
+
+unsigned int aom_highbd_8_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance4x2 aom_highbd_8_variance4x2_c
+
+unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance4x4_sse4_1(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_8_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
+
+unsigned int aom_highbd_8_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_sse2
+
+unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_sse2
+
+unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_sse2
+
+unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
+
+unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int aom_highbd_8_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_sse2
+
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+void aom_highbd_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+
+void aom_highbd_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd);
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd);
+
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+void aom_highbd_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+RTCD_EXTERN void (*aom_highbd_blend_a64_vmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
+
+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);
+#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);
+#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);
+void aom_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_avg_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_avg_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, int bps);
+RTCD_EXTERN void (*aom_highbd_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, int bps);
+
+void aom_highbd_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, int bps);
+void aom_highbd_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, int bps);
+void aom_highbd_convolve8_avg_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_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, int bps);
+
+void aom_highbd_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, int bps);
+void aom_highbd_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, int bps);
+void aom_highbd_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, int bps);
+RTCD_EXTERN void (*aom_highbd_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, int bps);
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+void aom_highbd_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, int bps);
+void aom_highbd_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, int bps);
+void aom_highbd_convolve_avg_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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
+
+void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+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_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_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
+
+void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+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_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
+
+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_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_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_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
+
+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
+
+void aom_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void aom_highbd_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define aom_highbd_fdct4x4 aom_highbd_fdct4x4_sse2
+
+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_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_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_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_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_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_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
+
+void aom_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
+#define aom_highbd_iwht4x4_1_add aom_highbd_iwht4x4_1_add_c
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_sse2
+
+void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_4_dual aom_highbd_lpf_horizontal_4_dual_sse2
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_8 aom_highbd_lpf_horizontal_8_sse2
+
+void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_horizontal_8_dual aom_highbd_lpf_horizontal_8_dual_sse2
+
+void aom_highbd_lpf_horizontal_edge_16_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_edge_16_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_edge_16 aom_highbd_lpf_horizontal_edge_16_sse2
+
+void aom_highbd_lpf_horizontal_edge_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_horizontal_edge_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_horizontal_edge_8 aom_highbd_lpf_horizontal_edge_8_sse2
+
+void aom_highbd_lpf_vertical_16_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_16_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_16 aom_highbd_lpf_vertical_16_sse2
+
+void aom_highbd_lpf_vertical_16_dual_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_16_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_16_dual aom_highbd_lpf_vertical_16_dual_sse2
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_4 aom_highbd_lpf_vertical_4_sse2
+
+void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_4_dual aom_highbd_lpf_vertical_4_dual_sse2
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+#define aom_highbd_lpf_vertical_8 aom_highbd_lpf_vertical_8_sse2
+
+void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+#define aom_highbd_lpf_vertical_8_dual aom_highbd_lpf_vertical_8_dual_sse2
+
+unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad16x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad16x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad32x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad32x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad4x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad4x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x32_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x32)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad64x64_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad64x64)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x16)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x4_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x4)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_highbd_masked_sad8x8_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_highbd_masked_sad8x8)(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
+
+unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+unsigned int aom_highbd_obmc_sad8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_sad8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+
+unsigned int aom_highbd_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance16x16 aom_highbd_obmc_sub_pixel_variance16x16_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance16x32 aom_highbd_obmc_sub_pixel_variance16x32_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance16x8 aom_highbd_obmc_sub_pixel_variance16x8_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance32x16 aom_highbd_obmc_sub_pixel_variance32x16_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance32x32 aom_highbd_obmc_sub_pixel_variance32x32_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance32x64 aom_highbd_obmc_sub_pixel_variance32x64_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance4x4 aom_highbd_obmc_sub_pixel_variance4x4_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance4x8 aom_highbd_obmc_sub_pixel_variance4x8_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance64x32 aom_highbd_obmc_sub_pixel_variance64x32_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance64x64 aom_highbd_obmc_sub_pixel_variance64x64_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance8x16 aom_highbd_obmc_sub_pixel_variance8x16_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance8x4 aom_highbd_obmc_sub_pixel_variance8x4_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance8x8 aom_highbd_obmc_sub_pixel_variance8x8_c
+
+unsigned int aom_highbd_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance16x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance16x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance16x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance16x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance16x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance16x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance32x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance32x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance32x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance32x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance32x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance32x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance4x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance4x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance4x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance4x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance64x32_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance64x32)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance64x64_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance64x64)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance8x16_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance8x16)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance8x4_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance8x4)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+unsigned int aom_highbd_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+unsigned int aom_highbd_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+
+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_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_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_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_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);
+#define aom_highbd_quantize_b_32x32 aom_highbd_quantize_b_32x32_sse2
+
+void aom_highbd_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define aom_highbd_quantize_b_64x64 aom_highbd_quantize_b_64x64_c
+
+unsigned int aom_highbd_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad16x16x3 aom_highbd_sad16x16x3_c
+
+void aom_highbd_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad16x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_highbd_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad16x16x8 aom_highbd_sad16x16x8_c
+
+unsigned int aom_highbd_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad16x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad16x8_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad16x8x3 aom_highbd_sad16x8x3_c
+
+void aom_highbd_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad16x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_highbd_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad16x8x8 aom_highbd_sad16x8x8_c
+
+unsigned int aom_highbd_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x16_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_highbd_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad32x32x3 aom_highbd_sad32x32x3_c
+
+void aom_highbd_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad32x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_highbd_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad32x32x8 aom_highbd_sad32x32x8_c
+
+unsigned int aom_highbd_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad32x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad32x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x4 aom_highbd_sad4x4_c
+
+unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_c
+
+void aom_highbd_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x4x3 aom_highbd_sad4x4x3_c
+
+void aom_highbd_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x4x4d aom_highbd_sad4x4x4d_sse2
+
+void aom_highbd_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x4x8 aom_highbd_sad4x4x8_c
+
+unsigned int aom_highbd_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x8 aom_highbd_sad4x8_c
+
+unsigned int aom_highbd_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x8_avg aom_highbd_sad4x8_avg_c
+
+void aom_highbd_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x8x4d aom_highbd_sad4x8x4d_sse2
+
+void aom_highbd_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x8x8 aom_highbd_sad4x8x8_c
+
+unsigned int aom_highbd_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x32_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+unsigned int aom_highbd_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int aom_highbd_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*aom_highbd_sad64x64_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+
+void aom_highbd_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad64x64x3 aom_highbd_sad64x64x3_c
+
+void aom_highbd_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+RTCD_EXTERN void (*aom_highbd_sad64x64x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+
+void aom_highbd_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad64x64x8 aom_highbd_sad64x64x8_c
+
+unsigned int aom_highbd_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x16 aom_highbd_sad8x16_sse2
+
+unsigned int aom_highbd_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x16_avg aom_highbd_sad8x16_avg_sse2
+
+void aom_highbd_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x16x3 aom_highbd_sad8x16x3_c
+
+void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_sse2
+
+void aom_highbd_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x16x8 aom_highbd_sad8x16x8_c
+
+unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x4 aom_highbd_sad8x4_sse2
+
+unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_sse2
+
+void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_sse2
+
+void aom_highbd_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x4x8 aom_highbd_sad8x4x8_c
+
+unsigned int aom_highbd_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int aom_highbd_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x8 aom_highbd_sad8x8_sse2
+
+unsigned int aom_highbd_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int aom_highbd_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x8_avg aom_highbd_sad8x8_avg_sse2
+
+void aom_highbd_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x8x3 aom_highbd_sad8x8x3_c
+
+void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void aom_highbd_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_sse2
+
+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_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_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_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_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_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);
+#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_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_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
+
+void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+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_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
+
 void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define aom_idct16x16_10_add aom_idct16x16_10_add_sse2
+void aom_idct16x16_10_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_10_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define aom_idct16x16_1_add aom_idct16x16_1_add_sse2
+void aom_idct16x16_1_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define aom_idct16x16_256_add aom_idct16x16_256_add_sse2
+void aom_idct16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_256_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define aom_idct16x16_38_add aom_idct16x16_38_add_c
+void aom_idct16x16_38_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct16x16_38_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_1024_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 RTCD_EXTERN void (*aom_idct32x32_1024_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_135_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 RTCD_EXTERN void (*aom_idct32x32_135_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
-#define aom_idct32x32_1_add aom_idct32x32_1_add_sse2
+void aom_idct32x32_1_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+RTCD_EXTERN void (*aom_idct32x32_1_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 void aom_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void aom_idct32x32_34_add_avx2(const tran_low_t *input, uint8_t *dest, int dest_stride);
 RTCD_EXTERN void (*aom_idct32x32_34_add)(const tran_low_t *input, uint8_t *dest, int dest_stride);
 
 void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
@@ -465,6 +2181,110 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, co
 void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
 #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2
 
+unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad16x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad16x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad32x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad32x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad4x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad4x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x32_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x32)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad64x64_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad64x64)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x16)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x4_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x4)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+unsigned int aom_masked_sad8x8_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+RTCD_EXTERN unsigned int (*aom_masked_sad8x8)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+
+unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance16x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance16x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance32x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance32x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance4x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance4x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x32_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x32)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance64x64_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance64x64)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x16_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x16)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x4_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x4)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
+unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+unsigned int aom_masked_sub_pixel_variance8x8_ssse3(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_masked_sub_pixel_variance8x8)(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+
 void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 #define aom_minmax_8x8 aom_minmax_8x8_sse2
@@ -1232,12 +3052,321 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_get16x16var = aom_get16x16var_avx2;
     aom_hadamard_8x8 = aom_hadamard_8x8_sse2;
     if (flags & HAS_SSSE3) aom_hadamard_8x8 = aom_hadamard_8x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x32 = aom_highbd_10_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance16x8 = aom_highbd_10_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x16 = aom_highbd_10_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x32 = aom_highbd_10_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance32x64 = aom_highbd_10_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x4 = aom_highbd_10_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance4x8 = aom_highbd_10_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x32 = aom_highbd_10_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance64x64 = aom_highbd_10_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x16 = aom_highbd_10_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x16 = aom_highbd_10_obmc_variance16x16_sse4_1;
+    aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x32 = aom_highbd_10_obmc_variance16x32_sse4_1;
+    aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance16x8 = aom_highbd_10_obmc_variance16x8_sse4_1;
+    aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x16 = aom_highbd_10_obmc_variance32x16_sse4_1;
+    aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x32 = aom_highbd_10_obmc_variance32x32_sse4_1;
+    aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance32x64 = aom_highbd_10_obmc_variance32x64_sse4_1;
+    aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x4 = aom_highbd_10_obmc_variance4x4_sse4_1;
+    aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance4x8 = aom_highbd_10_obmc_variance4x8_sse4_1;
+    aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x32 = aom_highbd_10_obmc_variance64x32_sse4_1;
+    aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance64x64 = aom_highbd_10_obmc_variance64x64_sse4_1;
+    aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x16 = aom_highbd_10_obmc_variance8x16_sse4_1;
+    aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x4 = aom_highbd_10_obmc_variance8x4_sse4_1;
+    aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance8x8 = aom_highbd_10_obmc_variance8x8_sse4_1;
+    aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_avg_variance4x4 = aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_sub_pixel_variance4x4 = aom_highbd_10_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_10_variance4x4 = aom_highbd_10_variance4x4_sse4_1;
+    aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x16 = aom_highbd_12_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x32 = aom_highbd_12_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance16x8 = aom_highbd_12_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x16 = aom_highbd_12_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x32 = aom_highbd_12_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance32x64 = aom_highbd_12_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x4 = aom_highbd_12_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance4x8 = aom_highbd_12_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x32 = aom_highbd_12_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance64x64 = aom_highbd_12_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x16 = aom_highbd_12_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x4 = aom_highbd_12_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_12_masked_sub_pixel_variance8x8 = aom_highbd_12_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x16 = aom_highbd_12_obmc_variance16x16_sse4_1;
+    aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x32 = aom_highbd_12_obmc_variance16x32_sse4_1;
+    aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance16x8 = aom_highbd_12_obmc_variance16x8_sse4_1;
+    aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x16 = aom_highbd_12_obmc_variance32x16_sse4_1;
+    aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x32 = aom_highbd_12_obmc_variance32x32_sse4_1;
+    aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance32x64 = aom_highbd_12_obmc_variance32x64_sse4_1;
+    aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x4 = aom_highbd_12_obmc_variance4x4_sse4_1;
+    aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance4x8 = aom_highbd_12_obmc_variance4x8_sse4_1;
+    aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x32 = aom_highbd_12_obmc_variance64x32_sse4_1;
+    aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance64x64 = aom_highbd_12_obmc_variance64x64_sse4_1;
+    aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x16 = aom_highbd_12_obmc_variance8x16_sse4_1;
+    aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x4 = aom_highbd_12_obmc_variance8x4_sse4_1;
+    aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_obmc_variance8x8 = aom_highbd_12_obmc_variance8x8_sse4_1;
+    aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_avg_variance4x4 = aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_sub_pixel_variance4x4 = aom_highbd_12_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_12_variance4x4 = aom_highbd_12_variance4x4_sse4_1;
+    aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x16 = aom_highbd_8_masked_sub_pixel_variance16x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x32 = aom_highbd_8_masked_sub_pixel_variance16x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance16x8 = aom_highbd_8_masked_sub_pixel_variance16x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x16 = aom_highbd_8_masked_sub_pixel_variance32x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x32 = aom_highbd_8_masked_sub_pixel_variance32x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance32x64 = aom_highbd_8_masked_sub_pixel_variance32x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x4 = aom_highbd_8_masked_sub_pixel_variance4x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance4x8 = aom_highbd_8_masked_sub_pixel_variance4x8_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x32 = aom_highbd_8_masked_sub_pixel_variance64x32_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance64x64 = aom_highbd_8_masked_sub_pixel_variance64x64_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x16 = aom_highbd_8_masked_sub_pixel_variance8x16_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x4 = aom_highbd_8_masked_sub_pixel_variance8x4_ssse3;
+    aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_8_masked_sub_pixel_variance8x8 = aom_highbd_8_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_avg_variance4x4 = aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1;
+    aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_sub_pixel_variance4x4 = aom_highbd_8_sub_pixel_variance4x4_sse4_1;
+    aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_8_variance4x4 = aom_highbd_8_variance4x4_sse4_1;
+    aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_hmask = aom_highbd_blend_a64_hmask_sse4_1;
+    aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_mask = aom_highbd_blend_a64_mask_sse4_1;
+    aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_c;
+    if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
+    aom_highbd_convolve8 = aom_highbd_convolve8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8 = aom_highbd_convolve8_avx2;
+    aom_highbd_convolve8_avg = aom_highbd_convolve8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_avg = aom_highbd_convolve8_avg_avx2;
+    aom_highbd_convolve8_avg_horiz = aom_highbd_convolve8_avg_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_avg_horiz = aom_highbd_convolve8_avg_horiz_avx2;
+    aom_highbd_convolve8_avg_vert = aom_highbd_convolve8_avg_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_avg_vert = aom_highbd_convolve8_avg_vert_avx2;
+    aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_horiz = aom_highbd_convolve8_horiz_avx2;
+    aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve8_vert = aom_highbd_convolve8_vert_avx2;
+    aom_highbd_convolve_avg = aom_highbd_convolve_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_avg = aom_highbd_convolve_avg_avx2;
+    aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
+    if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
+    aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_ssse3;
+    aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad16x8 = aom_highbd_masked_sad16x8_ssse3;
+    aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x16 = aom_highbd_masked_sad32x16_ssse3;
+    aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x32 = aom_highbd_masked_sad32x32_ssse3;
+    aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad32x64 = aom_highbd_masked_sad32x64_ssse3;
+    aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x4 = aom_highbd_masked_sad4x4_ssse3;
+    aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad4x8 = aom_highbd_masked_sad4x8_ssse3;
+    aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x32 = aom_highbd_masked_sad64x32_ssse3;
+    aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad64x64 = aom_highbd_masked_sad64x64_ssse3;
+    aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x16 = aom_highbd_masked_sad8x16_ssse3;
+    aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x4 = aom_highbd_masked_sad8x4_ssse3;
+    aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_masked_sad8x8 = aom_highbd_masked_sad8x8_ssse3;
+    aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x16 = aom_highbd_obmc_sad16x16_sse4_1;
+    aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x32 = aom_highbd_obmc_sad16x32_sse4_1;
+    aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad16x8 = aom_highbd_obmc_sad16x8_sse4_1;
+    aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x16 = aom_highbd_obmc_sad32x16_sse4_1;
+    aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x32 = aom_highbd_obmc_sad32x32_sse4_1;
+    aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad32x64 = aom_highbd_obmc_sad32x64_sse4_1;
+    aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x4 = aom_highbd_obmc_sad4x4_sse4_1;
+    aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad4x8 = aom_highbd_obmc_sad4x8_sse4_1;
+    aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x32 = aom_highbd_obmc_sad64x32_sse4_1;
+    aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad64x64 = aom_highbd_obmc_sad64x64_sse4_1;
+    aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x16 = aom_highbd_obmc_sad8x16_sse4_1;
+    aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x4 = aom_highbd_obmc_sad8x4_sse4_1;
+    aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_sad8x8 = aom_highbd_obmc_sad8x8_sse4_1;
+    aom_highbd_obmc_variance16x16 = aom_highbd_obmc_variance16x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance16x16 = aom_highbd_obmc_variance16x16_sse4_1;
+    aom_highbd_obmc_variance16x32 = aom_highbd_obmc_variance16x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance16x32 = aom_highbd_obmc_variance16x32_sse4_1;
+    aom_highbd_obmc_variance16x8 = aom_highbd_obmc_variance16x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance16x8 = aom_highbd_obmc_variance16x8_sse4_1;
+    aom_highbd_obmc_variance32x16 = aom_highbd_obmc_variance32x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance32x16 = aom_highbd_obmc_variance32x16_sse4_1;
+    aom_highbd_obmc_variance32x32 = aom_highbd_obmc_variance32x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance32x32 = aom_highbd_obmc_variance32x32_sse4_1;
+    aom_highbd_obmc_variance32x64 = aom_highbd_obmc_variance32x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance32x64 = aom_highbd_obmc_variance32x64_sse4_1;
+    aom_highbd_obmc_variance4x4 = aom_highbd_obmc_variance4x4_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance4x4 = aom_highbd_obmc_variance4x4_sse4_1;
+    aom_highbd_obmc_variance4x8 = aom_highbd_obmc_variance4x8_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance4x8 = aom_highbd_obmc_variance4x8_sse4_1;
+    aom_highbd_obmc_variance64x32 = aom_highbd_obmc_variance64x32_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance64x32 = aom_highbd_obmc_variance64x32_sse4_1;
+    aom_highbd_obmc_variance64x64 = aom_highbd_obmc_variance64x64_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance64x64 = aom_highbd_obmc_variance64x64_sse4_1;
+    aom_highbd_obmc_variance8x16 = aom_highbd_obmc_variance8x16_c;
+    if (flags & HAS_SSE4_1) aom_highbd_obmc_variance8x16 = aom_highbd_obmc_variance8x16_sse4_1;
+    aom_highbd_obmc_variance8x4 = aom_highbd_obmc_variance8x4_c;
+    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_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;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16_avg = aom_highbd_sad16x16_avg_avx2;
+    aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x16x4d = aom_highbd_sad16x16x4d_avx2;
+    aom_highbd_sad16x32 = aom_highbd_sad16x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32 = aom_highbd_sad16x32_avx2;
+    aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32_avg = aom_highbd_sad16x32_avg_avx2;
+    aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x32x4d = aom_highbd_sad16x32x4d_avx2;
+    aom_highbd_sad16x8 = aom_highbd_sad16x8_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8 = aom_highbd_sad16x8_avx2;
+    aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8_avg = aom_highbd_sad16x8_avg_avx2;
+    aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad16x8x4d = aom_highbd_sad16x8x4d_avx2;
+    aom_highbd_sad32x16 = aom_highbd_sad32x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16 = aom_highbd_sad32x16_avx2;
+    aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16_avg = aom_highbd_sad32x16_avg_avx2;
+    aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x16x4d = aom_highbd_sad32x16x4d_avx2;
+    aom_highbd_sad32x32 = aom_highbd_sad32x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32 = aom_highbd_sad32x32_avx2;
+    aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32_avg = aom_highbd_sad32x32_avg_avx2;
+    aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x32x4d = aom_highbd_sad32x32x4d_avx2;
+    aom_highbd_sad32x64 = aom_highbd_sad32x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64 = aom_highbd_sad32x64_avx2;
+    aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64_avg = aom_highbd_sad32x64_avg_avx2;
+    aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad32x64x4d = aom_highbd_sad32x64x4d_avx2;
+    aom_highbd_sad64x32 = aom_highbd_sad64x32_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32 = aom_highbd_sad64x32_avx2;
+    aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32_avg = aom_highbd_sad64x32_avg_avx2;
+    aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x32x4d = aom_highbd_sad64x32x4d_avx2;
+    aom_highbd_sad64x64 = aom_highbd_sad64x64_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64 = aom_highbd_sad64x64_avx2;
+    aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64_avg = aom_highbd_sad64x64_avg_avx2;
+    aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) aom_highbd_sad64x64x4d = aom_highbd_sad64x64x4d_avx2;
+    aom_idct16x16_10_add = aom_idct16x16_10_add_sse2;
+    if (flags & HAS_AVX2) aom_idct16x16_10_add = aom_idct16x16_10_add_avx2;
+    aom_idct16x16_1_add = aom_idct16x16_1_add_sse2;
+    if (flags & HAS_AVX2) aom_idct16x16_1_add = aom_idct16x16_1_add_avx2;
+    aom_idct16x16_256_add = aom_idct16x16_256_add_sse2;
+    if (flags & HAS_AVX2) aom_idct16x16_256_add = aom_idct16x16_256_add_avx2;
+    aom_idct16x16_38_add = aom_idct16x16_38_add_c;
+    if (flags & HAS_AVX2) aom_idct16x16_38_add = aom_idct16x16_38_add_avx2;
     aom_idct32x32_1024_add = aom_idct32x32_1024_add_sse2;
     if (flags & HAS_SSSE3) aom_idct32x32_1024_add = aom_idct32x32_1024_add_ssse3;
+    if (flags & HAS_AVX2) aom_idct32x32_1024_add = aom_idct32x32_1024_add_avx2;
     aom_idct32x32_135_add = aom_idct32x32_1024_add_sse2;
     if (flags & HAS_SSSE3) aom_idct32x32_135_add = aom_idct32x32_135_add_ssse3;
+    if (flags & HAS_AVX2) aom_idct32x32_135_add = aom_idct32x32_135_add_avx2;
+    aom_idct32x32_1_add = aom_idct32x32_1_add_sse2;
+    if (flags & HAS_AVX2) aom_idct32x32_1_add = aom_idct32x32_1_add_avx2;
     aom_idct32x32_34_add = aom_idct32x32_34_add_sse2;
     if (flags & HAS_SSSE3) aom_idct32x32_34_add = aom_idct32x32_34_add_ssse3;
+    if (flags & HAS_AVX2) aom_idct32x32_34_add = aom_idct32x32_34_add_avx2;
     aom_idct8x8_12_add = aom_idct8x8_12_add_sse2;
     if (flags & HAS_SSSE3) aom_idct8x8_12_add = aom_idct8x8_12_add_ssse3;
     aom_idct8x8_64_add = aom_idct8x8_64_add_sse2;
@@ -1246,6 +3375,58 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_avx2;
     aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_sse2;
     if (flags & HAS_AVX2) aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_avx2;
+    aom_masked_sad16x16 = aom_masked_sad16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
+    aom_masked_sad16x32 = aom_masked_sad16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x32 = aom_masked_sad16x32_ssse3;
+    aom_masked_sad16x8 = aom_masked_sad16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad16x8 = aom_masked_sad16x8_ssse3;
+    aom_masked_sad32x16 = aom_masked_sad32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x16 = aom_masked_sad32x16_ssse3;
+    aom_masked_sad32x32 = aom_masked_sad32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x32 = aom_masked_sad32x32_ssse3;
+    aom_masked_sad32x64 = aom_masked_sad32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad32x64 = aom_masked_sad32x64_ssse3;
+    aom_masked_sad4x4 = aom_masked_sad4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x4 = aom_masked_sad4x4_ssse3;
+    aom_masked_sad4x8 = aom_masked_sad4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad4x8 = aom_masked_sad4x8_ssse3;
+    aom_masked_sad64x32 = aom_masked_sad64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x32 = aom_masked_sad64x32_ssse3;
+    aom_masked_sad64x64 = aom_masked_sad64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sad64x64 = aom_masked_sad64x64_ssse3;
+    aom_masked_sad8x16 = aom_masked_sad8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x16 = aom_masked_sad8x16_ssse3;
+    aom_masked_sad8x4 = aom_masked_sad8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x4 = aom_masked_sad8x4_ssse3;
+    aom_masked_sad8x8 = aom_masked_sad8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sad8x8 = aom_masked_sad8x8_ssse3;
+    aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x16 = aom_masked_sub_pixel_variance16x16_ssse3;
+    aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x32 = aom_masked_sub_pixel_variance16x32_ssse3;
+    aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance16x8 = aom_masked_sub_pixel_variance16x8_ssse3;
+    aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x16 = aom_masked_sub_pixel_variance32x16_ssse3;
+    aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x32 = aom_masked_sub_pixel_variance32x32_ssse3;
+    aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance32x64 = aom_masked_sub_pixel_variance32x64_ssse3;
+    aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x4 = aom_masked_sub_pixel_variance4x4_ssse3;
+    aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance4x8 = aom_masked_sub_pixel_variance4x8_ssse3;
+    aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x32 = aom_masked_sub_pixel_variance64x32_ssse3;
+    aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance64x64 = aom_masked_sub_pixel_variance64x64_ssse3;
+    aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x16 = aom_masked_sub_pixel_variance8x16_ssse3;
+    aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x4 = aom_masked_sub_pixel_variance8x4_ssse3;
+    aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSSE3) aom_masked_sub_pixel_variance8x8 = aom_masked_sub_pixel_variance8x8_ssse3;
     aom_mse16x16 = aom_mse16x16_sse2;
     if (flags & HAS_AVX2) aom_mse16x16 = aom_mse16x16_avx2;
     aom_obmc_sad16x16 = aom_obmc_sad16x16_c;
diff --git a/media/libaom/config/win/x64/av1_rtcd.h b/media/libaom/config/win/x64/av1_rtcd.h
index eacb5a555..9dcf96f84 100644
--- a/media/libaom/config/win/x64/av1_rtcd.h
+++ b/media/libaom/config/win/x64/av1_rtcd.h
@@ -60,13 +60,7 @@ 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);
-int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
-int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
-RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
-
-int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
-int64_t av1_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
-#define av1_block_error_fp av1_block_error_fp_sse2
+#define av1_block_error av1_block_error_c
 
 void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 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);
@@ -80,9 +74,7 @@ int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_conf
 #define av1_diamond_search_sad av1_diamond_search_sad_c
 
 void av1_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void av1_fdct8x8_quant_sse2(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-void av1_fdct8x8_quant_ssse3(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-RTCD_EXTERN void (*av1_fdct8x8_quant)(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#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);
@@ -149,9 +141,174 @@ RTCD_EXTERN int (*av1_full_search_sad)(const struct macroblock *x, const struct
 void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
 #define av1_fwd_idtx av1_fwd_idtx_c
 
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+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_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);
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+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_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_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);
+
 void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define av1_fwht4x4 av1_fwht4x4_c
 
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
+#define av1_highbd_block_error av1_highbd_block_error_sse2
+
+void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8 av1_highbd_convolve8_sse2
+
+void av1_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_avg av1_highbd_convolve8_avg_sse2
+
+void av1_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_avg_horiz av1_highbd_convolve8_avg_horiz_sse2
+
+void av1_highbd_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, int bps);
+void av1_highbd_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, int bps);
+#define av1_highbd_convolve8_avg_vert av1_highbd_convolve8_avg_vert_sse2
+
+void av1_highbd_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, int bps);
+void av1_highbd_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, int bps);
+#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
+
+void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
+
+void av1_highbd_convolve_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, int bps);
+#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
+
+void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
+
+void av1_highbd_convolve_horiz_c(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_convolve_horiz_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_horiz)(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_convolve_init_c(void);
+void av1_highbd_convolve_init_sse4_1(void);
+RTCD_EXTERN void (*av1_highbd_convolve_init)(void);
+
+void av1_highbd_convolve_vert_c(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_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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#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);
+#define av1_highbd_quantize_b av1_highbd_quantize_b_c
+
+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);
+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_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);
@@ -204,6 +361,25 @@ void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride
 void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 #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_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);
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+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_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_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);
+
 void av1_lowbd_convolve_init_c(void);
 void av1_lowbd_convolve_init_ssse3(void);
 RTCD_EXTERN void (*av1_lowbd_convolve_init)(void);
@@ -213,20 +389,35 @@ 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);
-void av1_quantize_fp_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-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);
+#define av1_quantize_fp av1_quantize_fp_sse2
 
 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);
-void av1_quantize_fp_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
-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);
+#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
 
 void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 void av1_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define av1_temporal_filter_apply av1_temporal_filter_apply_sse2
 
-void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-void av1_warp_affine_sse2(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-#define av1_warp_affine av1_warp_affine_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_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);
+#define av1_wedge_compute_delta_squares av1_wedge_compute_delta_squares_sse2
+
+int av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+int av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+#define av1_wedge_sign_from_residuals av1_wedge_sign_from_residuals_sse2
+
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
+#define av1_wedge_sse_from_residuals av1_wedge_sse_from_residuals_sse2
+
+double compute_cross_correlation_c(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
+double compute_cross_correlation_sse4_1(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
+RTCD_EXTERN double (*compute_cross_correlation)(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
 
 void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
 void copy_4x4_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
@@ -304,14 +495,10 @@ 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_sse2;
-    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;
     if (flags & HAS_SSSE3) av1_convolve_vert = av1_convolve_vert_ssse3;
-    av1_fdct8x8_quant = av1_fdct8x8_quant_sse2;
-    if (flags & HAS_SSSE3) av1_fdct8x8_quant = av1_fdct8x8_quant_ssse3;
     av1_fht16x16 = av1_fht16x16_sse2;
     if (flags & HAS_AVX2) av1_fht16x16 = av1_fht16x16_avx2;
     av1_fht32x32 = av1_fht32x32_sse2;
@@ -319,14 +506,44 @@ static void setup_rtcd_internal(void)
     av1_full_search_sad = av1_full_search_sad_c;
     if (flags & HAS_SSE3) av1_full_search_sad = av1_full_search_sadx3;
     if (flags & HAS_SSE4_1) av1_full_search_sad = av1_full_search_sadx8;
+    av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
+    av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
+    av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
+    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
+    av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
+    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_sse4_1;
+    av1_highbd_convolve_init = av1_highbd_convolve_init_c;
+    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;
+    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;
     if (flags & HAS_AVX2) av1_iht16x16_256_add = av1_iht16x16_256_add_avx2;
+    av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_sse4_1;
+    av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c;
+    if (flags & HAS_AVX2) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_avx2;
+    av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
+    av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
+    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_SSSE3) av1_quantize_fp = av1_quantize_fp_ssse3;
-    av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
-    if (flags & HAS_SSSE3) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_ssse3;
+    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;
+    if (flags & HAS_SSE4_1) compute_cross_correlation = compute_cross_correlation_sse4_1;
     copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse4_1;