Merge "AVX2 SubPixel Variance Optimization"

vp9/common/vp9_rtcd_defs.sh

@@ -386,7 +386,7 @@ prototype unsigned int vp9_variance4x4 "const uint8_t *src_ptr, int source_strid
 specialize vp9_variance4x4 mmx $sse2_x86inc
 
 prototype 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"
-specialize vp9_sub_pixel_variance64x64 $sse2_x86inc $ssse3_x86inc
+specialize vp9_sub_pixel_variance64x64 $sse2_x86inc $ssse3_x86inc avx2
 
 prototype 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"
 specialize vp9_sub_pixel_avg_variance64x64 $sse2_x86inc $ssse3_x86inc
@@ -416,7 +416,7 @@ prototype unsigned int vp9_sub_pixel_avg_variance16x32 "const uint8_t *src_ptr,
 specialize vp9_sub_pixel_avg_variance16x32 $sse2_x86inc $ssse3_x86inc
 
 prototype 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"
-specialize vp9_sub_pixel_variance32x32 $sse2_x86inc $ssse3_x86inc
+specialize vp9_sub_pixel_variance32x32 $sse2_x86inc $ssse3_x86inc avx2
 
 prototype 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"
 specialize vp9_sub_pixel_avg_variance32x32 $sse2_x86inc $ssse3_x86inc

vp9/encoder/x86/vp9_variance_avx2.c

@@ -42,6 +42,18 @@ void vp9_get32x32var_avx2
   int *Sum
 );
 
+unsigned int vp9_sub_pixel_variance32xh_avx2
+(
+  const uint8_t *src,
+  int src_stride,
+  int x_offset,
+  int y_offset,
+  const uint8_t *dst,
+  int dst_stride,
+  int height,
+  unsigned int *sse
+);
+
 static void variance_avx2(const unsigned char *src_ptr, int  source_stride,
                         const unsigned char *ref_ptr, int  recon_stride,
                         int  w, int  h, unsigned int *sse, int *sum,
@@ -155,3 +167,43 @@ unsigned int vp9_variance64x32_avx2(const uint8_t *src_ptr,
   *sse = var;
   return (var - (((int64_t)avg * avg) >> 11));
 }
+
+unsigned int vp9_sub_pixel_variance64x64_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse_ptr) {
+  // processing 32 elements in parallel
+  unsigned int sse;
+  int se = vp9_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                           y_offset, dst, dst_stride,
+                                           64, &sse);
+  // processing the next 32 elements in parallel
+  unsigned int sse2;
+  int se2 = vp9_sub_pixel_variance32xh_avx2(src + 32, src_stride,
+                                            x_offset, y_offset,
+                                            dst + 32, dst_stride,
+                                            64, &sse2);
+  se += se2;
+  sse += sse2;
+  *sse_ptr = sse;
+  return sse - (((int64_t)se * se) >> 12);
+}
+
+unsigned int vp9_sub_pixel_variance32x32_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse_ptr) {
+  // processing 32 element in parallel
+  unsigned int sse;
+  int se = vp9_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                           y_offset, dst, dst_stride,
+                                           32, &sse);
+  *sse_ptr = sse;
+  return sse - (((int64_t)se * se) >> 10);
+}

vp9/vp9cx.mk

@@ -86,6 +86,7 @@ VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_variance_impl_sse2.asm
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_impl_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_sad4d_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_subpel_variance_impl_sse2.asm
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
 VP9_CX_SRCS-$(HAVE_SSE3) += encoder/x86/vp9_sad_sse3.asm