Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(583)

Unified Diff: source/config/linux/ia32/vp9_rtcd.h

Issue 863143002: Enable AVX2 optimizations in libvpx (Closed) Base URL: https://chromium.googlesource.com/chromium/deps/libvpx.git@master
Patch Set: Created 5 years, 11 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « source/config/linux/generic/vpx_config.c ('k') | source/config/linux/ia32/vpx_config.h » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: source/config/linux/ia32/vp9_rtcd.h
diff --git a/source/config/linux/ia32/vp9_rtcd.h b/source/config/linux/ia32/vp9_rtcd.h
index 45668bf59953b2e87734c7997d4dcdd7723e3418..59908f79cfb49926b5fd8152b71a29471555ddc1 100644
--- a/source/config/linux/ia32/vp9_rtcd.h
+++ b/source/config/linux/ia32/vp9_rtcd.h
@@ -39,11 +39,13 @@ RTCD_EXTERN unsigned int (*vp9_avg_8x8)(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);
+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);
@@ -235,6 +239,7 @@ RTCD_EXTERN void (*vp9_fdct16x16_1)(const int16_t *input, tran_low_t *output, in
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);
+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);
@@ -243,6 +248,7 @@ RTCD_EXTERN void (*vp9_fdct32x32_1)(const int16_t *input, tran_low_t *output, in
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);
+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);
@@ -292,6 +298,7 @@ RTCD_EXTERN void (*vp9_fwht4x4)(const int16_t *input, tran_low_t *output, int st
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);
+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);
@@ -384,6 +391,7 @@ 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);
+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);
@@ -428,6 +436,7 @@ RTCD_EXTERN void (*vp9_lpf_vertical_8_dual)(uint8_t *s, int pitch, const uint8_t
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);
+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);
@@ -510,10 +519,12 @@ 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);
+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);
+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);
@@ -522,10 +533,12 @@ RTCD_EXTERN void (*vp9_sad32x16x4d)(const uint8_t *src_ptr, int src_stride, con
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);
+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);
+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);
@@ -533,6 +546,7 @@ void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *
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);
+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);
@@ -540,10 +554,12 @@ void vp9_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *re
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);
+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);
+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);
@@ -586,10 +602,12 @@ 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);
+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);
+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);
@@ -598,10 +616,12 @@ RTCD_EXTERN void (*vp9_sad64x32x4d)(const uint8_t *src_ptr, int src_stride, con
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);
+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);
+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);
@@ -609,6 +629,7 @@ void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *
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);
+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);
@@ -690,6 +711,7 @@ 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);
@@ -715,6 +737,7 @@ 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);
@@ -755,6 +778,7 @@ 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);
@@ -780,6 +804,7 @@ 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);
@@ -838,6 +863,7 @@ RTCD_EXTERN void (*vp9_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const
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);
+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);
@@ -850,10 +876,12 @@ RTCD_EXTERN unsigned int (*vp9_variance16x8)(const uint8_t *src_ptr, int source_
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);
+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);
+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);
@@ -870,10 +898,12 @@ RTCD_EXTERN unsigned int (*vp9_variance4x8)(const uint8_t *src_ptr, int source_s
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);
+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);
+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);
@@ -904,9 +934,11 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_avg_8x8 = vp9_avg_8x8_sse2;
vp9_block_error = vp9_block_error_c;
if (flags & HAS_SSE2) vp9_block_error = vp9_block_error_sse2;
+ if (flags & HAS_AVX2) vp9_block_error = vp9_block_error_avx2;
vp9_convolve8 = vp9_convolve8_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) vp9_convolve8_avg = vp9_convolve8_avg_sse2;
if (flags & HAS_SSSE3) vp9_convolve8_avg = vp9_convolve8_avg_ssse3;
@@ -919,9 +951,11 @@ static void setup_rtcd_internal(void)
vp9_convolve8_horiz = vp9_convolve8_horiz_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) 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_convolve_avg = vp9_convolve_avg_c;
if (flags & HAS_SSE2) vp9_convolve_avg = vp9_convolve_avg_sse2;
vp9_convolve_copy = vp9_convolve_copy_c;
@@ -972,10 +1006,12 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_fdct16x16_1 = vp9_fdct16x16_1_sse2;
vp9_fdct32x32 = vp9_fdct32x32_c;
if (flags & HAS_SSE2) vp9_fdct32x32 = vp9_fdct32x32_sse2;
+ if (flags & HAS_AVX2) vp9_fdct32x32 = vp9_fdct32x32_avx2;
vp9_fdct32x32_1 = vp9_fdct32x32_1_c;
if (flags & HAS_SSE2) vp9_fdct32x32_1 = vp9_fdct32x32_1_sse2;
vp9_fdct32x32_rd = vp9_fdct32x32_rd_c;
if (flags & HAS_SSE2) vp9_fdct32x32_rd = vp9_fdct32x32_rd_sse2;
+ if (flags & HAS_AVX2) vp9_fdct32x32_rd = vp9_fdct32x32_rd_avx2;
vp9_fdct4x4 = vp9_fdct4x4_c;
if (flags & HAS_SSE2) vp9_fdct4x4 = vp9_fdct4x4_sse2;
vp9_fdct4x4_1 = vp9_fdct4x4_1_c;
@@ -1000,6 +1036,7 @@ static void setup_rtcd_internal(void)
if (flags & HAS_MMX) vp9_fwht4x4 = vp9_fwht4x4_mmx;
vp9_get16x16var = vp9_get16x16var_c;
if (flags & HAS_SSE2) vp9_get16x16var = vp9_get16x16var_sse2;
+ if (flags & HAS_AVX2) vp9_get16x16var = vp9_get16x16var_avx2;
vp9_get8x8var = vp9_get8x8var_c;
if (flags & HAS_SSE2) vp9_get8x8var = vp9_get8x8var_sse2;
vp9_get_mb_ss = vp9_get_mb_ss_c;
@@ -1044,6 +1081,7 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_iht8x8_64_add = vp9_iht8x8_64_add_sse2;
vp9_lpf_horizontal_16 = vp9_lpf_horizontal_16_c;
if (flags & HAS_SSE2) vp9_lpf_horizontal_16 = vp9_lpf_horizontal_16_sse2;
+ if (flags & HAS_AVX2) vp9_lpf_horizontal_16 = vp9_lpf_horizontal_16_avx2;
vp9_lpf_horizontal_4 = vp9_lpf_horizontal_4_c;
if (flags & HAS_MMX) vp9_lpf_horizontal_4 = vp9_lpf_horizontal_4_mmx;
vp9_lpf_horizontal_4_dual = vp9_lpf_horizontal_4_dual_c;
@@ -1066,6 +1104,7 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_lpf_vertical_8_dual = vp9_lpf_vertical_8_dual_sse2;
vp9_mse16x16 = vp9_mse16x16_c;
if (flags & HAS_SSE2) vp9_mse16x16 = vp9_mse16x16_sse2;
+ if (flags & HAS_AVX2) vp9_mse16x16 = vp9_mse16x16_avx2;
vp9_mse16x8 = vp9_mse16x8_c;
if (flags & HAS_SSE2) vp9_mse16x8 = vp9_mse16x8_sse2;
vp9_mse8x16 = vp9_mse8x16_c;
@@ -1102,20 +1141,27 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_sad16x8x4d = vp9_sad16x8x4d_sse2;
vp9_sad32x16 = vp9_sad32x16_c;
if (flags & HAS_SSE2) vp9_sad32x16 = vp9_sad32x16_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x16 = vp9_sad32x16_avx2;
vp9_sad32x16_avg = vp9_sad32x16_avg_c;
if (flags & HAS_SSE2) vp9_sad32x16_avg = vp9_sad32x16_avg_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x16_avg = vp9_sad32x16_avg_avx2;
vp9_sad32x16x4d = vp9_sad32x16x4d_c;
if (flags & HAS_SSE2) vp9_sad32x16x4d = vp9_sad32x16x4d_sse2;
vp9_sad32x32 = vp9_sad32x32_c;
if (flags & HAS_SSE2) vp9_sad32x32 = vp9_sad32x32_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x32 = vp9_sad32x32_avx2;
vp9_sad32x32_avg = vp9_sad32x32_avg_c;
if (flags & HAS_SSE2) vp9_sad32x32_avg = vp9_sad32x32_avg_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x32_avg = vp9_sad32x32_avg_avx2;
vp9_sad32x32x4d = vp9_sad32x32x4d_c;
if (flags & HAS_SSE2) vp9_sad32x32x4d = vp9_sad32x32x4d_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x32x4d = vp9_sad32x32x4d_avx2;
vp9_sad32x64 = vp9_sad32x64_c;
if (flags & HAS_SSE2) vp9_sad32x64 = vp9_sad32x64_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x64 = vp9_sad32x64_avx2;
vp9_sad32x64_avg = vp9_sad32x64_avg_c;
if (flags & HAS_SSE2) vp9_sad32x64_avg = vp9_sad32x64_avg_sse2;
+ if (flags & HAS_AVX2) vp9_sad32x64_avg = vp9_sad32x64_avg_avx2;
vp9_sad32x64x4d = vp9_sad32x64x4d_c;
if (flags & HAS_SSE2) vp9_sad32x64x4d = vp9_sad32x64x4d_sse2;
vp9_sad4x4 = vp9_sad4x4_c;
@@ -1134,16 +1180,21 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE) vp9_sad4x8x4d = vp9_sad4x8x4d_sse;
vp9_sad64x32 = vp9_sad64x32_c;
if (flags & HAS_SSE2) vp9_sad64x32 = vp9_sad64x32_sse2;
+ if (flags & HAS_AVX2) vp9_sad64x32 = vp9_sad64x32_avx2;
vp9_sad64x32_avg = vp9_sad64x32_avg_c;
if (flags & HAS_SSE2) vp9_sad64x32_avg = vp9_sad64x32_avg_sse2;
+ if (flags & HAS_AVX2) vp9_sad64x32_avg = vp9_sad64x32_avg_avx2;
vp9_sad64x32x4d = vp9_sad64x32x4d_c;
if (flags & HAS_SSE2) vp9_sad64x32x4d = vp9_sad64x32x4d_sse2;
vp9_sad64x64 = vp9_sad64x64_c;
if (flags & HAS_SSE2) vp9_sad64x64 = vp9_sad64x64_sse2;
+ if (flags & HAS_AVX2) vp9_sad64x64 = vp9_sad64x64_avx2;
vp9_sad64x64_avg = vp9_sad64x64_avg_c;
if (flags & HAS_SSE2) vp9_sad64x64_avg = vp9_sad64x64_avg_sse2;
+ if (flags & HAS_AVX2) vp9_sad64x64_avg = vp9_sad64x64_avg_avx2;
vp9_sad64x64x4d = vp9_sad64x64x4d_c;
if (flags & HAS_SSE2) vp9_sad64x64x4d = vp9_sad64x64x4d_sse2;
+ if (flags & HAS_AVX2) vp9_sad64x64x4d = vp9_sad64x64x4d_avx2;
vp9_sad8x16 = vp9_sad8x16_c;
if (flags & HAS_SSE2) vp9_sad8x16 = vp9_sad8x16_sse2;
vp9_sad8x16_avg = vp9_sad8x16_avg_c;
@@ -1181,6 +1232,7 @@ static void setup_rtcd_internal(void)
vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) 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;
@@ -1196,6 +1248,7 @@ static void setup_rtcd_internal(void)
vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) 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;
@@ -1220,6 +1273,7 @@ static void setup_rtcd_internal(void)
vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_sse2;
if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_ssse3;
@@ -1235,6 +1289,7 @@ static void setup_rtcd_internal(void)
vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_c;
if (flags & HAS_SSE2) 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_c;
if (flags & HAS_SSE2) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_sse2;
if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_ssse3;
@@ -1264,14 +1319,17 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE) vp9_v_predictor_8x8 = vp9_v_predictor_8x8_sse;
vp9_variance16x16 = vp9_variance16x16_c;
if (flags & HAS_SSE2) vp9_variance16x16 = vp9_variance16x16_sse2;
+ if (flags & HAS_AVX2) vp9_variance16x16 = vp9_variance16x16_avx2;
vp9_variance16x32 = vp9_variance16x32_c;
if (flags & HAS_SSE2) vp9_variance16x32 = vp9_variance16x32_sse2;
vp9_variance16x8 = vp9_variance16x8_c;
if (flags & HAS_SSE2) vp9_variance16x8 = vp9_variance16x8_sse2;
vp9_variance32x16 = vp9_variance32x16_c;
if (flags & HAS_SSE2) vp9_variance32x16 = vp9_variance32x16_sse2;
+ if (flags & HAS_AVX2) vp9_variance32x16 = vp9_variance32x16_avx2;
vp9_variance32x32 = vp9_variance32x32_c;
if (flags & HAS_SSE2) vp9_variance32x32 = vp9_variance32x32_sse2;
+ if (flags & HAS_AVX2) vp9_variance32x32 = vp9_variance32x32_avx2;
vp9_variance32x64 = vp9_variance32x64_c;
if (flags & HAS_SSE2) vp9_variance32x64 = vp9_variance32x64_sse2;
vp9_variance4x4 = vp9_variance4x4_c;
@@ -1280,8 +1338,10 @@ static void setup_rtcd_internal(void)
if (flags & HAS_SSE2) vp9_variance4x8 = vp9_variance4x8_sse2;
vp9_variance64x32 = vp9_variance64x32_c;
if (flags & HAS_SSE2) vp9_variance64x32 = vp9_variance64x32_sse2;
+ if (flags & HAS_AVX2) vp9_variance64x32 = vp9_variance64x32_avx2;
vp9_variance64x64 = vp9_variance64x64_c;
if (flags & HAS_SSE2) vp9_variance64x64 = vp9_variance64x64_sse2;
+ if (flags & HAS_AVX2) vp9_variance64x64 = vp9_variance64x64_avx2;
vp9_variance8x16 = vp9_variance8x16_c;
if (flags & HAS_SSE2) vp9_variance8x16 = vp9_variance8x16_sse2;
vp9_variance8x4 = vp9_variance8x4_c;
« no previous file with comments | « source/config/linux/generic/vpx_config.c ('k') | source/config/linux/ia32/vpx_config.h » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698