Enable AVX2 optimizations in libvpx

source/config/win/x64/vp9_rtcd.h

Index: source/config/win/x64/vp9_rtcd.h
diff --git a/source/config/win/x64/vp9_rtcd.h b/source/config/win/x64/vp9_rtcd.h
index 2e7d8a51eccedd460a972444b74e0a693060c858..235c390c159ef9c9c79ba9467345d5e8159d3cae 100644
--- a/source/config/win/x64/vp9_rtcd.h
+++ b/source/config/win/x64/vp9_rtcd.h
@@ -39,11 +39,13 @@ unsigned int vp9_avg_8x8_sse2(const uint8_t *, int p);
 
 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
-#define vp9_block_error vp9_block_error_sse2
+int64_t vp9_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 
 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);
@@ -64,11 +66,13 @@ RTCD_EXTERN void (*vp9_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_str
 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);
@@ -235,7 +239,8 @@ void vp9_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride);
 
 void vp9_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
 void vp9_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
-#define vp9_fdct32x32 vp9_fdct32x32_sse2
+void vp9_fdct32x32_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32)(const int16_t *input, tran_low_t *output, int stride);
 
 void vp9_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
 void vp9_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
@@ -243,7 +248,8 @@ void vp9_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
 
 void vp9_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
 void vp9_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
-#define vp9_fdct32x32_rd vp9_fdct32x32_rd_sse2
+void vp9_fdct32x32_rd_avx2(const int16_t *input, tran_low_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32_rd)(const int16_t *input, tran_low_t *output, int stride);
 
 void vp9_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
 void vp9_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
@@ -293,7 +299,8 @@ void vp9_fwht4x4_mmx(const int16_t *input, tran_low_t *output, int stride);
 
 void vp9_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 void vp9_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
-#define vp9_get16x16var vp9_get16x16var_sse2
+void vp9_get16x16var_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*vp9_get16x16var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 
 void vp9_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 void vp9_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
@@ -387,7 +394,8 @@ void vp9_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride
 
 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);
@@ -431,7 +439,8 @@ void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0,
 
 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_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);
 unsigned int vp9_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
@@ -517,11 +526,13 @@ void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref
 
 unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int vp9_sad32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
-#define vp9_sad32x16 vp9_sad32x16_sse2
+unsigned int vp9_sad32x16_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 
 unsigned int vp9_sad32x16_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 vp9_sad32x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
-#define vp9_sad32x16_avg vp9_sad32x16_avg_sse2
+unsigned int vp9_sad32x16_avg_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sad32x16_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 
 void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 void vp9_sad32x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
@@ -529,29 +540,34 @@ void vp9_sad32x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t
 
 unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 unsigned int vp9_sad32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
-#define vp9_sad32x32 vp9_sad32x32_sse2
+unsigned int vp9_sad32x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 
 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 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);
-#define vp9_sad32x32_avg vp9_sad32x32_avg_sse2
+unsigned int vp9_sad32x32_avg_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sad32x32_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 
 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 vp9_sad32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
-#define vp9_sad32x64 vp9_sad32x64_sse2
+unsigned int vp9_sad32x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 
 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 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);
-#define vp9_sad32x64_avg vp9_sad32x64_avg_sse2
+unsigned int vp9_sad32x64_avg_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sad32x64_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 
 void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 void vp9_sad32x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
@@ -593,11 +609,13 @@ void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_p
 
 unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int vp9_sad64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
-#define vp9_sad64x32 vp9_sad64x32_sse2
+unsigned int vp9_sad64x32_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride);
 
 unsigned int vp9_sad64x32_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 vp9_sad64x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
-#define vp9_sad64x32_avg vp9_sad64x32_avg_sse2
+unsigned int vp9_sad64x32_avg_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sad64x32_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 
 void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
 void vp9_sad64x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
@@ -605,18 +623,21 @@ void vp9_sad64x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t
 
 unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 unsigned int vp9_sad64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
-#define vp9_sad64x64 vp9_sad64x64_sse2
+unsigned int vp9_sad64x64_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride);
 
 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 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);
-#define vp9_sad64x64_avg vp9_sad64x64_avg_sse2
+unsigned int vp9_sad64x64_avg_avx2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sad64x64_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred);
 
 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
@@ -697,6 +718,7 @@ RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x16)(const uint8_t *src_p
 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);
@@ -722,6 +744,7 @@ RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x32)(const uint8_t *src_p
 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);
@@ -762,6 +785,7 @@ RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x16)(const uint8_t *src_ptr,
 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);
@@ -787,6 +811,7 @@ RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x32)(const uint8_t *src_ptr,
 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);
@@ -846,7 +871,8 @@ void vp9_v_predictor_8x8_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *ab
 
 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_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);
@@ -858,11 +884,13 @@ unsigned int vp9_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, co
 
 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);
@@ -878,11 +906,13 @@ unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, con
 
 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_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
@@ -906,8 +936,11 @@ static void setup_rtcd_internal(void)
 
     (void)flags;
 
+    vp9_block_error = vp9_block_error_sse2;
+    if (flags & HAS_AVX2) vp9_block_error = vp9_block_error_avx2;
     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;
@@ -916,8 +949,10 @@ static void setup_rtcd_internal(void)
     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;
@@ -948,6 +983,10 @@ static void setup_rtcd_internal(void)
     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_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_fdct8x8 = vp9_fdct8x8_sse2;
     if (flags & HAS_SSSE3) vp9_fdct8x8 = vp9_fdct8x8_ssse3;
     vp9_fdct8x8_quant = vp9_fdct8x8_quant_sse2;
@@ -955,6 +994,8 @@ static void setup_rtcd_internal(void)
     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_get16x16var = vp9_get16x16var_sse2;
+    if (flags & HAS_AVX2) vp9_get16x16var = vp9_get16x16var_avx2;
     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;
@@ -971,6 +1012,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) vp9_idct8x8_12_add = vp9_idct8x8_12_add_ssse3;
     vp9_idct8x8_64_add = vp9_idct8x8_64_add_sse2;
     if (flags & HAS_SSSE3) vp9_idct8x8_64_add = vp9_idct8x8_64_add_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_sse2;
     if (flags & HAS_SSSE3) vp9_quantize_b = vp9_quantize_b_ssse3;
     vp9_quantize_b_32x32 = vp9_quantize_b_32x32_c;
@@ -985,8 +1030,32 @@ static void setup_rtcd_internal(void)
     vp9_sad16x8x3 = vp9_sad16x8x3_c;
     if (flags & HAS_SSE3) vp9_sad16x8x3 = vp9_sad16x8x3_sse3;
     if (flags & HAS_SSSE3) vp9_sad16x8x3 = vp9_sad16x8x3_ssse3;
+    vp9_sad32x16 = vp9_sad32x16_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x16 = vp9_sad32x16_avx2;
+    vp9_sad32x16_avg = vp9_sad32x16_avg_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x16_avg = vp9_sad32x16_avg_avx2;
+    vp9_sad32x32 = vp9_sad32x32_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x32 = vp9_sad32x32_avx2;
+    vp9_sad32x32_avg = vp9_sad32x32_avg_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x32_avg = vp9_sad32x32_avg_avx2;
+    vp9_sad32x32x4d = vp9_sad32x32x4d_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x32x4d = vp9_sad32x32x4d_avx2;
+    vp9_sad32x64 = vp9_sad32x64_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x64 = vp9_sad32x64_avx2;
+    vp9_sad32x64_avg = vp9_sad32x64_avg_sse2;
+    if (flags & HAS_AVX2) vp9_sad32x64_avg = vp9_sad32x64_avg_avx2;
     vp9_sad4x4x3 = vp9_sad4x4x3_c;
     if (flags & HAS_SSE3) vp9_sad4x4x3 = vp9_sad4x4x3_sse3;
+    vp9_sad64x32 = vp9_sad64x32_sse2;
+    if (flags & HAS_AVX2) vp9_sad64x32 = vp9_sad64x32_avx2;
+    vp9_sad64x32_avg = vp9_sad64x32_avg_sse2;
+    if (flags & HAS_AVX2) vp9_sad64x32_avg = vp9_sad64x32_avg_avx2;
+    vp9_sad64x64 = vp9_sad64x64_sse2;
+    if (flags & HAS_AVX2) vp9_sad64x64 = vp9_sad64x64_avx2;
+    vp9_sad64x64_avg = vp9_sad64x64_avg_sse2;
+    if (flags & HAS_AVX2) vp9_sad64x64_avg = vp9_sad64x64_avg_avx2;
+    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;
@@ -1001,6 +1070,7 @@ static void setup_rtcd_internal(void)
     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;
@@ -1011,6 +1081,7 @@ static void setup_rtcd_internal(void)
     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;
@@ -1027,6 +1098,7 @@ static void setup_rtcd_internal(void)
     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;
@@ -1037,12 +1109,23 @@ static void setup_rtcd_internal(void)
     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