ref: f4fcfe3075cabc47b47db9df74720e24016ff200
parent: 90a7723f8c172ce130e5d7d824dfc639a4e56a43
author: Yunqing Wang <yunqingwang@google.com>
date: Wed May 22 05:07:30 EDT 2013
Optimize variance functions Added SSE2 version of variance functions for super blocks. Change-Id: Ibeaae8771ca21c99d41dd74067574a51e97b412d
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -188,11 +188,11 @@
#endif
#if HAVE_SSE2
-const vp9_variance_fn_t variance4x4_wmt = vp9_variance4x4_wmt;
-const vp9_variance_fn_t variance8x8_wmt = vp9_variance8x8_wmt;
-const vp9_variance_fn_t variance8x16_wmt = vp9_variance8x16_wmt;
-const vp9_variance_fn_t variance16x8_wmt = vp9_variance16x8_wmt;
-const vp9_variance_fn_t variance16x16_wmt = vp9_variance16x16_wmt;
+const vp9_variance_fn_t variance4x4_wmt = vp9_variance4x4_sse2;
+const vp9_variance_fn_t variance8x8_wmt = vp9_variance8x8_sse2;
+const vp9_variance_fn_t variance8x16_wmt = vp9_variance8x16_sse2;
+const vp9_variance_fn_t variance16x8_wmt = vp9_variance16x8_sse2;
+const vp9_variance_fn_t variance16x16_wmt = vp9_variance16x16_sse2;
INSTANTIATE_TEST_CASE_P(
SSE2, VP9VarianceTest,
::testing::Values(make_tuple(4, 4, variance4x4_wmt),
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -236,53 +236,43 @@
[ $arch = "x86_64" ] && mmx_x86_64=mmx && sse2_x86_64=sse2
prototype unsigned int vp9_variance32x16 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance32x16
+specialize vp9_variance32x16 sse2
prototype unsigned int vp9_variance16x32 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance16x32
+specialize vp9_variance16x32 sse2
prototype unsigned int vp9_variance64x32 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance64x32
+specialize vp9_variance64x32 sse2
prototype unsigned int vp9_variance32x64 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance32x64
+specialize vp9_variance32x64 sse2
prototype unsigned int vp9_variance32x32 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance32x32
+specialize vp9_variance32x32 sse2
prototype unsigned int vp9_variance64x64 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance64x64
+specialize vp9_variance64x64 sse2
prototype unsigned int vp9_variance16x16 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_variance16x16 mmx sse2
-vp9_variance16x16_sse2=vp9_variance16x16_wmt
-vp9_variance16x16_mmx=vp9_variance16x16_mmx
prototype unsigned int vp9_variance16x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_variance16x8 mmx sse2
-vp9_variance16x8_sse2=vp9_variance16x8_wmt
-vp9_variance16x8_mmx=vp9_variance16x8_mmx
prototype unsigned int vp9_variance8x16 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_variance8x16 mmx sse2
-vp9_variance8x16_sse2=vp9_variance8x16_wmt
-vp9_variance8x16_mmx=vp9_variance8x16_mmx
prototype unsigned int vp9_variance8x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_variance8x8 mmx sse2
-vp9_variance8x8_sse2=vp9_variance8x8_wmt
-vp9_variance8x8_mmx=vp9_variance8x8_mmx
prototype unsigned int vp9_variance8x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance8x4
+specialize vp9_variance8x4 sse2
prototype unsigned int vp9_variance4x8 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
-specialize vp9_variance4x8
+specialize vp9_variance4x8 sse2
prototype unsigned int vp9_variance4x4 "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse"
specialize vp9_variance4x4 mmx sse2
-vp9_variance4x4_sse2=vp9_variance4x4_wmt
-vp9_variance4x4_mmx=vp9_variance4x4_mmx
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
--- a/vp9/encoder/x86/vp9_variance_sse2.c
+++ b/vp9/encoder/x86/vp9_variance_sse2.c
@@ -139,8 +139,38 @@
DECLARE_ALIGNED(16, extern const short, vp9_bilinear_filters_mmx[16][8]);
-unsigned int vp9_variance4x4_wmt(
+typedef unsigned int (*get_var_sse2) (
const unsigned char *src_ptr,
+ int source_stride,
+ const unsigned char *ref_ptr,
+ int recon_stride,
+ unsigned int *SSE,
+ int *Sum
+);
+
+static void variance_sse2(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,
+ get_var_sse2 var_fn, int block_size) {+ unsigned int sse0;
+ int sum0;
+ int i, j;
+
+ *sse = 0;
+ *sum = 0;
+
+ for (i = 0; i < h; i += block_size) {+ for (j = 0; j < w; j += block_size) {+ var_fn(src_ptr + source_stride * i + j, source_stride,
+ ref_ptr + recon_stride * i + j, recon_stride, &sse0, &sum0);
+ *sse += sse0;
+ *sum += sum0;
+ }
+ }
+}
+
+unsigned int vp9_variance4x4_sse2(
+ const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
@@ -148,13 +178,41 @@
unsigned int var;
int avg;
- vp9_get4x4var_mmx(src_ptr, source_stride, ref_ptr, recon_stride, &var, &avg);
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 4, 4,
+ &var, &avg, vp9_get4x4var_mmx, 4);
*sse = var;
return (var - (((unsigned int)avg * avg) >> 4));
+}
+unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 4,
+ &var, &avg, vp9_get4x4var_mmx, 4);
+ *sse = var;
+ return (var - (((unsigned int)avg * avg) >> 5));
}
-unsigned int vp9_variance8x8_wmt
+unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 4, 8,
+ &var, &avg, vp9_get4x4var_mmx, 4);
+ *sse = var;
+ return (var - (((unsigned int)avg * avg) >> 5));
+}
+
+unsigned int vp9_variance8x8_sse2
(
const unsigned char *src_ptr,
int source_stride,
@@ -164,14 +222,13 @@
unsigned int var;
int avg;
- vp9_get8x8var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &var, &avg);
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 8,
+ &var, &avg, vp9_get8x8var_sse2, 8);
*sse = var;
return (var - (((unsigned int)avg * avg) >> 6));
-
}
-
-unsigned int vp9_variance16x16_wmt
+unsigned int vp9_variance16x8_sse2
(
const unsigned char *src_ptr,
int source_stride,
@@ -178,32 +235,32 @@
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {- unsigned int sse0;
- int sum0;
+ unsigned int var;
+ int avg;
-
- vp9_get16x16var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0, &sum0);
- *sse = sse0;
- return (sse0 - (((unsigned int)sum0 * sum0) >> 8));
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 8,
+ &var, &avg, vp9_get8x8var_sse2, 8);
+ *sse = var;
+ return (var - (((unsigned int)avg * avg) >> 7));
}
-unsigned int vp9_mse16x16_wmt(
+unsigned int vp9_variance8x16_sse2
+(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {+ unsigned int var;
+ int avg;
- unsigned int sse0;
- int sum0;
- vp9_get16x16var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0, &sum0);
- *sse = sse0;
- return sse0;
-
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 8, 16,
+ &var, &avg, vp9_get8x8var_sse2, 8);
+ *sse = var;
+ return (var - (((unsigned int)avg * avg) >> 7));
}
-
-unsigned int vp9_variance16x8_wmt
+unsigned int vp9_variance16x16_sse2
(
const unsigned char *src_ptr,
int source_stride,
@@ -210,37 +267,112 @@
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {- unsigned int sse0, sse1, var;
- int sum0, sum1, avg;
+ unsigned int var;
+ int avg;
- vp9_get8x8var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0, &sum0);
- vp9_get8x8var_sse2(src_ptr + 8, source_stride, ref_ptr + 8, recon_stride, &sse1, &sum1);
-
- var = sse0 + sse1;
- avg = sum0 + sum1;
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 16,
+ &var, &avg, vp9_get16x16var_sse2, 16);
*sse = var;
- return (var - (((unsigned int)avg * avg) >> 7));
-
+ return (var - (((unsigned int)avg * avg) >> 8));
}
-unsigned int vp9_variance8x16_wmt
-(
+unsigned int vp9_mse16x16_wmt(
const unsigned char *src_ptr,
int source_stride,
const unsigned char *ref_ptr,
int recon_stride,
unsigned int *sse) {- unsigned int sse0, sse1, var;
- int sum0, sum1, avg;
- vp9_get8x8var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0, &sum0);
- vp9_get8x8var_sse2(src_ptr + 8 * source_stride, source_stride, ref_ptr + 8 * recon_stride, recon_stride, &sse1, &sum1);
+ unsigned int sse0;
+ int sum0;
+ vp9_get16x16var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
+ &sum0);
+ *sse = sse0;
+ return sse0;
+}
- var = sse0 + sse1;
- avg = sum0 + sum1;
+unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 32,
+ &var, &avg, vp9_get16x16var_sse2, 16);
*sse = var;
- return (var - (((unsigned int)avg * avg) >> 7));
+ return (var - (((int64_t)avg * avg) >> 10));
+}
+unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 16,
+ &var, &avg, vp9_get16x16var_sse2, 16);
+ *sse = var;
+ return (var - (((int64_t)avg * avg) >> 9));
+}
+
+unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 16, 32,
+ &var, &avg, vp9_get16x16var_sse2, 16);
+ *sse = var;
+ return (var - (((int64_t)avg * avg) >> 9));
+}
+
+unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 64, 64,
+ &var, &avg, vp9_get16x16var_sse2, 16);
+ *sse = var;
+ return (var - (((int64_t)avg * avg) >> 12));
+}
+
+unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 64, 32,
+ &var, &avg, vp9_get16x16var_sse2, 16);
+ *sse = var;
+ return (var - (((int64_t)avg * avg) >> 11));
+}
+
+unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr,
+ int source_stride,
+ const uint8_t *ref_ptr,
+ int recon_stride,
+ unsigned int *sse) {+ unsigned int var;
+ int avg;
+
+ variance_sse2(src_ptr, source_stride, ref_ptr, recon_stride, 32, 64,
+ &var, &avg, vp9_get16x16var_sse2, 16);
+ *sse = var;
+ return (var - (((int64_t)avg * avg) >> 11));
}
unsigned int vp9_sub_pixel_variance4x4_wmt
--- a/vp9/encoder/x86/vp9_variance_ssse3.c
+++ b/vp9/encoder/x86/vp9_variance_ssse3.c
@@ -15,15 +15,6 @@
#define HALFNDX 8
-extern unsigned int vp9_get16x16var_sse2
-(
- const unsigned char *src_ptr,
- int source_stride,
- const unsigned char *ref_ptr,
- int recon_stride,
- unsigned int *SSE,
- int *Sum
-);
extern void vp9_half_horiz_vert_variance16x_h_sse2
(
const unsigned char *ref_ptr,