Enable AVX2 optimizations for libvpx

source/config/linux/x64/vp9_rtcd.h

 #ifndef VP9_RTCD_H_
 #define VP9_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
 #else
 #define RTCD_EXTERN extern
 #endif
 
 #ifdef __cplusplus
@@ skipping 27 matching lines @@
 void vp9_blend_mb_outer_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, int v1, int alpha, int stride);
 #define vp9_blend_mb_outer vp9_blend_mb_outer_c
 
 int64_t vp9_block_error_c(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 int64_t vp9_block_error_sse2(const int16_t *coeff, const int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 #define vp9_block_error vp9_block_error_sse2
 
 void vp9_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 vp9_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 vp9_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vp9_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vp9_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vp9_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*vp9_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define vp9_convolve_avg vp9_convolve_avg_sse2
 
 void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void vp9_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define vp9_convolve_copy vp9_convolve_copy_sse2
 
@@ skipping 135 matching lines @@
 
 void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
 
 int vp9_diamond_search_sad_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
 int vp9_diamond_search_sadx4(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
 RTCD_EXTERN int (*vp9_diamond_search_sad)(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
 
 void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
 void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct16x16 vp9_fdct16x16_sse2
+void vp9_fdct16x16_avx2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct16x16)(const int16_t *input, int16_t *output, int stride);
 
 void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
 void vp9_fdct32x32_sse2(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct32x32 vp9_fdct32x32_sse2
+void vp9_fdct32x32_avx2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32)(const int16_t *input, int16_t *output, int stride);
 
 void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
 void vp9_fdct32x32_rd_sse2(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct32x32_rd vp9_fdct32x32_rd_sse2
+void vp9_fdct32x32_rd_avx2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32_rd)(const int16_t *input, int16_t *output, int stride);
 
 void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
 void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct4x4 vp9_fdct4x4_sse2
+void vp9_fdct4x4_avx2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct4x4)(const int16_t *input, int16_t *output, int stride);
 
 void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
 void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride);
-#define vp9_fdct8x8 vp9_fdct8x8_sse2
+void vp9_fdct8x8_avx2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct8x8)(const int16_t *input, int16_t *output, int stride);
 
 void vp9_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 void vp9_fht16x16_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
-#define vp9_fht16x16 vp9_fht16x16_sse2
+void vp9_fht16x16_avx2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_fht16x16)(const int16_t *input, int16_t *output, int stride, int tx_type);
 
 void vp9_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 void vp9_fht4x4_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
-#define vp9_fht4x4 vp9_fht4x4_sse2
+void vp9_fht4x4_avx2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_fht4x4)(const int16_t *input, int16_t *output, int stride, int tx_type);
 
 void vp9_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
 void vp9_fht8x8_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
-#define vp9_fht8x8 vp9_fht8x8_sse2
+void vp9_fht8x8_avx2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_fht8x8)(const int16_t *input, int16_t *output, int stride, int tx_type);
 
 int vp9_full_range_search_c(const struct macroblock *x, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv);
 #define vp9_full_range_search vp9_full_range_search_c
 
 int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
 int vp9_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
 int vp9_full_search_sadx8(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
 RTCD_EXTERN int (*vp9_full_search_sad)(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, const struct mv *center_mv, struct mv *best_mv);
 
 void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
@@ skipping 81 matching lines @@
 #define vp9_iht8x8_64_add vp9_iht8x8_64_add_sse2
 
 void vp9_iwht4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_iwht4x4_16_add vp9_iwht4x4_16_add_c
 
 void vp9_iwht4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
 #define vp9_iwht4x4_1_add vp9_iwht4x4_1_add_c
 
 void vp9_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 void vp9_lpf_horizontal_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
-#define vp9_lpf_horizontal_16 vp9_lpf_horizontal_16_sse2
+void vp9_lpf_horizontal_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_lpf_horizontal_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 
 void vp9_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 void vp9_lpf_horizontal_4_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 #define vp9_lpf_horizontal_4 vp9_lpf_horizontal_4_mmx
 
 void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
 void vp9_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
 #define vp9_lpf_horizontal_4_dual vp9_lpf_horizontal_4_dual_sse2
 
 void vp9_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
@@ skipping 24 matching lines @@
 void vp9_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
 #define vp9_lpf_vertical_8 vp9_lpf_vertical_8_sse2
 
 void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
 void vp9_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 vp9_lpf_vertical_8_dual vp9_lpf_vertical_8_dual_sse2
 
 unsigned int vp9_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vp9_mse16x16_mmx(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int vp9_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
-#define vp9_mse16x16 vp9_mse16x16_sse2
+unsigned int vp9_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int vp9_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vp9_mse16x8 vp9_mse16x8_c
 
 unsigned int vp9_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vp9_mse8x16 vp9_mse8x16_c
 
 unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define vp9_mse8x8 vp9_mse8x8_c
 
@@ skipping 81 matching lines @@
 
 unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
 unsigned int vp9_sad32x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
 #define vp9_sad32x32_avg vp9_sad32x32_avg_sse2
 
 void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
 #define vp9_sad32x32x3 vp9_sad32x32x3_c
 
 void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 void vp9_sad32x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
-#define vp9_sad32x32x4d vp9_sad32x32x4d_sse2
+void vp9_sad32x32x4d_avx2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad32x32x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 
 void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad32x32x8 vp9_sad32x32x8_c
 
 unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
 unsigned int vp9_sad32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
 #define vp9_sad32x64 vp9_sad32x64_sse2
 
 unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
 unsigned int vp9_sad32x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
@@ skipping 56 matching lines @@
 
 unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
 unsigned int vp9_sad64x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
 #define vp9_sad64x64_avg vp9_sad64x64_avg_sse2
 
 void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
 #define vp9_sad64x64x3 vp9_sad64x64x3_c
 
 void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 void vp9_sad64x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
-#define vp9_sad64x64x4d vp9_sad64x64x4d_sse2
+void vp9_sad64x64x4d_avx2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad64x64x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 
 void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
 #define vp9_sad64x64x8 vp9_sad64x64x8_c
 
 unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
 unsigned int vp9_sad8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
 unsigned int vp9_sad8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
 #define vp9_sad8x16 vp9_sad8x16_sse2
 
 unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
@@ skipping 62 matching lines @@
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
 
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
 unsigned int vp9_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
@@ skipping 26 matching lines @@
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
@@ skipping 40 matching lines @@
 void vp9_v_predictor_4x4_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_v_predictor_4x4 vp9_v_predictor_4x4_sse
 
 void vp9_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void vp9_v_predictor_8x8_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define vp9_v_predictor_8x8 vp9_v_predictor_8x8_sse
 
 unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance16x16 vp9_variance16x16_sse2
+unsigned int vp9_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance16x32 vp9_variance16x32_sse2
 
 unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance16x8 vp9_variance16x8_sse2
 
 unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance32x16 vp9_variance32x16_sse2
+unsigned int vp9_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance32x32 vp9_variance32x32_sse2
+unsigned int vp9_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance32x64 vp9_variance32x64_sse2
 
 unsigned int vp9_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance4x4 vp9_variance4x4_sse2
 
 unsigned int vp9_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance4x8 vp9_variance4x8_sse2
 
 unsigned int vp9_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance64x32 vp9_variance64x32_sse2
+unsigned int vp9_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
-#define vp9_variance64x64 vp9_variance64x64_sse2
+unsigned int vp9_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 
 unsigned int vp9_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x16 vp9_variance8x16_sse2
 
 unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define vp9_variance8x4 vp9_variance8x4_sse2
 
@@ skipping 37 matching lines @@
 #ifdef RTCD_C
 #include "vpx_ports/x86.h"
 static void setup_rtcd_internal(void)
 {
     int flags = x86_simd_caps();
 
     (void)flags;
 
     vp9_convolve8 = vp9_convolve8_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8 = vp9_convolve8_ssse3;
+    if (flags & HAS_AVX2) vp9_convolve8 = vp9_convolve8_avx2;
     vp9_convolve8_avg = vp9_convolve8_avg_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_avg = vp9_convolve8_avg_ssse3;
     vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_ssse3;
     vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_ssse3;
     vp9_convolve8_horiz = vp9_convolve8_horiz_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_horiz = vp9_convolve8_horiz_ssse3;
+    if (flags & HAS_AVX2) vp9_convolve8_horiz = vp9_convolve8_horiz_avx2;
     vp9_convolve8_vert = vp9_convolve8_vert_sse2;
     if (flags & HAS_SSSE3) vp9_convolve8_vert = vp9_convolve8_vert_ssse3;
+    if (flags & HAS_AVX2) vp9_convolve8_vert = vp9_convolve8_vert_avx2;
     vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_c;
     if (flags & HAS_SSSE3) vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_ssse3;
     vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_c;
     if (flags & HAS_SSSE3) vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_ssse3;
     vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_c;
     if (flags & HAS_SSSE3) vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_ssse3;
     vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_c;
     if (flags & HAS_SSSE3) vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_ssse3;
     vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_c;
     if (flags & HAS_SSSE3) vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_ssse3;
@@ skipping 12 matching lines @@
     vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_c;
     if (flags & HAS_SSSE3) vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_ssse3;
     vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_c;
     if (flags & HAS_SSSE3) vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_ssse3;
     vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_c;
     if (flags & HAS_SSSE3) vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_ssse3;
     vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_c;
     if (flags & HAS_SSSE3) vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_ssse3;
     vp9_diamond_search_sad = vp9_diamond_search_sad_c;
     if (flags & HAS_SSE3) vp9_diamond_search_sad = vp9_diamond_search_sadx4;
+    vp9_fdct16x16 = vp9_fdct16x16_sse2;
+    if (flags & HAS_AVX2) vp9_fdct16x16 = vp9_fdct16x16_avx2;
+    vp9_fdct32x32 = vp9_fdct32x32_sse2;
+    if (flags & HAS_AVX2) vp9_fdct32x32 = vp9_fdct32x32_avx2;
+    vp9_fdct32x32_rd = vp9_fdct32x32_rd_sse2;
+    if (flags & HAS_AVX2) vp9_fdct32x32_rd = vp9_fdct32x32_rd_avx2;
+    vp9_fdct4x4 = vp9_fdct4x4_sse2;
+    if (flags & HAS_AVX2) vp9_fdct4x4 = vp9_fdct4x4_avx2;
+    vp9_fdct8x8 = vp9_fdct8x8_sse2;
+    if (flags & HAS_AVX2) vp9_fdct8x8 = vp9_fdct8x8_avx2;
+    vp9_fht16x16 = vp9_fht16x16_sse2;
+    if (flags & HAS_AVX2) vp9_fht16x16 = vp9_fht16x16_avx2;
+    vp9_fht4x4 = vp9_fht4x4_sse2;
+    if (flags & HAS_AVX2) vp9_fht4x4 = vp9_fht4x4_avx2;
+    vp9_fht8x8 = vp9_fht8x8_sse2;
+    if (flags & HAS_AVX2) vp9_fht8x8 = vp9_fht8x8_avx2;
     vp9_full_search_sad = vp9_full_search_sad_c;
     if (flags & HAS_SSE3) vp9_full_search_sad = vp9_full_search_sadx3;
     if (flags & HAS_SSE4_1) vp9_full_search_sad = vp9_full_search_sadx8;
     vp9_h_predictor_16x16 = vp9_h_predictor_16x16_c;
     if (flags & HAS_SSSE3) vp9_h_predictor_16x16 = vp9_h_predictor_16x16_ssse3;
     vp9_h_predictor_32x32 = vp9_h_predictor_32x32_c;
     if (flags & HAS_SSSE3) vp9_h_predictor_32x32 = vp9_h_predictor_32x32_ssse3;
     vp9_h_predictor_4x4 = vp9_h_predictor_4x4_c;
     if (flags & HAS_SSSE3) vp9_h_predictor_4x4 = vp9_h_predictor_4x4_ssse3;
     vp9_h_predictor_8x8 = vp9_h_predictor_8x8_c;
     if (flags & HAS_SSSE3) vp9_h_predictor_8x8 = vp9_h_predictor_8x8_ssse3;
+    vp9_lpf_horizontal_16 = vp9_lpf_horizontal_16_sse2;
+    if (flags & HAS_AVX2) vp9_lpf_horizontal_16 = vp9_lpf_horizontal_16_avx2;
+    vp9_mse16x16 = vp9_mse16x16_sse2;
+    if (flags & HAS_AVX2) vp9_mse16x16 = vp9_mse16x16_avx2;
     vp9_quantize_b = vp9_quantize_b_c;
     if (flags & HAS_SSSE3) vp9_quantize_b = vp9_quantize_b_ssse3;
     vp9_quantize_b_32x32 = vp9_quantize_b_32x32_c;
     if (flags & HAS_SSSE3) vp9_quantize_b_32x32 = vp9_quantize_b_32x32_ssse3;
     vp9_refining_search_sad = vp9_refining_search_sad_c;
     if (flags & HAS_SSE3) vp9_refining_search_sad = vp9_refining_search_sadx4;
     vp9_sad16x16x3 = vp9_sad16x16x3_c;
     if (flags & HAS_SSE3) vp9_sad16x16x3 = vp9_sad16x16x3_sse3;
     if (flags & HAS_SSSE3) vp9_sad16x16x3 = vp9_sad16x16x3_ssse3;
     vp9_sad16x8x3 = vp9_sad16x8x3_c;
     if (flags & HAS_SSE3) vp9_sad16x8x3 = vp9_sad16x8x3_sse3;
     if (flags & HAS_SSSE3) vp9_sad16x8x3 = vp9_sad16x8x3_ssse3;
+    vp9_sad32x32x4d = vp9_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x32x4d = vp9_sad32x32x4d_avx2;
     vp9_sad4x4x3 = vp9_sad4x4x3_c;
     if (flags & HAS_SSE3) vp9_sad4x4x3 = vp9_sad4x4x3_sse3;
+    vp9_sad64x64x4d = vp9_sad64x64x4d_sse2;
+    if (flags & HAS_AVX2) vp9_sad64x64x4d = vp9_sad64x64x4d_avx2;
     vp9_sad8x16x3 = vp9_sad8x16x3_c;
     if (flags & HAS_SSE3) vp9_sad8x16x3 = vp9_sad8x16x3_sse3;
     vp9_sad8x8x3 = vp9_sad8x8x3_c;
     if (flags & HAS_SSE3) vp9_sad8x8x3 = vp9_sad8x8x3_sse3;
     vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_ssse3;
     vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_ssse3;
     vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_ssse3;
     vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_ssse3;
     vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_ssse3;
+    if (flags & HAS_AVX2) vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_avx2;
     vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_ssse3;
     vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_sse;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_ssse3;
     vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_sse;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_ssse3;
     vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_ssse3;
     vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_ssse3;
+    if (flags & HAS_AVX2) vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_avx2;
     vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_ssse3;
     vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_ssse3;
     vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_ssse3;
     vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_ssse3;
     vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_ssse3;
     vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_ssse3;
     vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_ssse3;
     vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_ssse3;
+    if (flags & HAS_AVX2) vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_avx2;
     vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_ssse3;
     vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_sse;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_ssse3;
     vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_sse;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_ssse3;
     vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_ssse3;
     vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_ssse3;
+    if (flags & HAS_AVX2) vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_avx2;
     vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_ssse3;
     vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_ssse3;
     vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_sse2;
     if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_ssse3;
+    vp9_variance16x16 = vp9_variance16x16_sse2;
+    if (flags & HAS_AVX2) vp9_variance16x16 = vp9_variance16x16_avx2;
+    vp9_variance32x16 = vp9_variance32x16_sse2;
+    if (flags & HAS_AVX2) vp9_variance32x16 = vp9_variance32x16_avx2;
+    vp9_variance32x32 = vp9_variance32x32_sse2;
+    if (flags & HAS_AVX2) vp9_variance32x32 = vp9_variance32x32_avx2;
+    vp9_variance64x32 = vp9_variance64x32_sse2;
+    if (flags & HAS_AVX2) vp9_variance64x32 = vp9_variance64x32_avx2;
+    vp9_variance64x64 = vp9_variance64x64_sse2;
+    if (flags & HAS_AVX2) vp9_variance64x64 = vp9_variance64x64_avx2;
 }
 #endif
 
 #ifdef __cplusplus
 }  // extern "C"
 #endif
 
 #endif