aboutsummaryrefslogtreecommitdiff
path: root/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
diff options
context:
space:
mode:
Diffstat (limited to 'libvpx/vp9/encoder/x86/vp9_variance_sse2.c')
-rw-r--r--libvpx/vp9/encoder/x86/vp9_variance_sse2.c445
1 files changed, 114 insertions, 331 deletions
diff --git a/libvpx/vp9/encoder/x86/vp9_variance_sse2.c b/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
index 9e65694a8..e935a233a 100644
--- a/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
+++ b/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
@@ -11,113 +11,29 @@
#include "./vpx_config.h"
#include "vp9/encoder/vp9_variance.h"
-#include "vp9/common/vp9_pragmas.h"
#include "vpx_ports/mem.h"
-extern unsigned int vp9_get4x4var_mmx
-(
- const unsigned char *src_ptr,
- int source_stride,
- const unsigned char *ref_ptr,
- int recon_stride,
- unsigned int *SSE,
- int *Sum
-);
-
-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
-);
-unsigned int vp9_get8x8var_sse2
-(
- const unsigned char *src_ptr,
- int source_stride,
- const unsigned char *ref_ptr,
- int recon_stride,
- unsigned int *SSE,
- int *Sum
-);
-void vp9_half_horiz_vert_variance8x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-void vp9_half_horiz_vert_variance16x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-void vp9_half_horiz_variance8x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-void vp9_half_horiz_variance16x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-void vp9_half_vert_variance8x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-void vp9_half_vert_variance16x_h_sse2
-(
- const unsigned char *ref_ptr,
- int ref_pixels_per_line,
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- unsigned int Height,
- int *sum,
- unsigned int *sumsquared
-);
-
-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;
+typedef unsigned int (*variance_fn_t) (const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse, int *sum);
+
+unsigned int vp9_get4x4var_mmx(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse, int *sum);
+
+
+unsigned int vp9_get8x8var_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse, int *sum);
+
+unsigned int vp9_get16x16var_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse, int *sum);
+
+static void variance_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ int w, int h, unsigned int *sse, int *sum,
+ variance_fn_t var_fn, int block_size) {
int i, j;
*sse = 0;
@@ -125,217 +41,139 @@ static void variance_sse2(const unsigned char *src_ptr, int source_stride,
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);
+ unsigned int sse0;
+ int sum0;
+ var_fn(src + src_stride * i + j, src_stride,
+ ref + ref_stride * i + j, ref_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,
- unsigned int *sse) {
- unsigned int var;
- int 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_variance4x4_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 4, 4,
+ sse, &sum, vp9_get4x4var_mmx, 4);
+ return *sse - (((unsigned int)sum * sum) >> 4);
}
-unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance8x4_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 8, 4,
+ sse, &sum, vp9_get4x4var_mmx, 4);
+ return *sse - (((unsigned int)sum * sum) >> 5);
}
-unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance4x8_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 4, 8,
+ sse, &sum, vp9_get4x4var_mmx, 4);
+ return *sse - (((unsigned int)sum * sum) >> 5);
}
-unsigned int vp9_variance8x8_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;
-
- 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_variance8x8_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 8, 8,
+ sse, &sum, vp9_get8x8var_sse2, 8);
+ return *sse - (((unsigned int)sum * sum) >> 6);
}
-unsigned int vp9_variance16x8_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;
-
- 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_variance16x8_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 16, 8,
+ sse, &sum, vp9_get8x8var_sse2, 8);
+ return *sse - (((unsigned int)sum * sum) >> 7);
}
-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;
-
- 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_variance8x16_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 8, 16,
+ sse, &sum, vp9_get8x8var_sse2, 8);
+ return *sse - (((unsigned int)sum * sum) >> 7);
}
-unsigned int vp9_variance16x16_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;
-
- 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) >> 8));
+unsigned int vp9_variance16x16_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 16, 16,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((unsigned int)sum * sum) >> 8);
}
-unsigned int vp9_mse16x16_sse2(
- const unsigned char *src_ptr,
- int source_stride,
- const unsigned char *ref_ptr,
- int recon_stride,
- unsigned int *sse) {
- unsigned int sse0;
- int sum0;
- vp9_get16x16var_sse2(src_ptr, source_stride, ref_ptr, recon_stride, &sse0,
- &sum0);
- *sse = sse0;
- return sse0;
+unsigned int vp9_mse16x16_sse2(const unsigned char *src, int src_stride,
+ const unsigned char *ref, int ref_stride,
+ unsigned int *sse) {
+ int sum;
+ vp9_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+ return *sse;
}
-unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance32x32_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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 - (((int64_t)avg * avg) >> 10));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 32, 32,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 10);
}
-unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance32x16_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 32, 16,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 9);
}
-unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance16x32_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 16, 32,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 9);
}
-unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance64x64_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 64, 64,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 12);
}
-unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance64x32_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 64, 32,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 11);
}
-unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr,
- int source_stride,
- const uint8_t *ref_ptr,
- int recon_stride,
+unsigned int vp9_variance32x64_sse2(const uint8_t *src, int src_stride,
+ const uint8_t *ref, int ref_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));
+ int sum;
+ variance_sse2(src, src_stride, ref, ref_stride, 32, 64,
+ sse, &sum, vp9_get16x16var_sse2, 16);
+ return *sse - (((int64_t)sum * sum) >> 11);
}
#define DECL(w, opt) \
@@ -494,58 +332,3 @@ FNS(ssse3, ssse3);
#undef FNS
#undef FN
-
-unsigned int vp9_variance_halfpixvar16x16_h_sse2(
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- const unsigned char *dst_ptr,
- int dst_pixels_per_line,
- unsigned int *sse) {
- int xsum0;
- unsigned int xxsum0;
-
- vp9_half_horiz_variance16x_h_sse2(
- src_ptr, src_pixels_per_line,
- dst_ptr, dst_pixels_per_line, 16,
- &xsum0, &xxsum0);
-
- *sse = xxsum0;
- return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
-
-
-unsigned int vp9_variance_halfpixvar16x16_v_sse2(
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- const unsigned char *dst_ptr,
- int dst_pixels_per_line,
- unsigned int *sse) {
- int xsum0;
- unsigned int xxsum0;
- vp9_half_vert_variance16x_h_sse2(
- src_ptr, src_pixels_per_line,
- dst_ptr, dst_pixels_per_line, 16,
- &xsum0, &xxsum0);
-
- *sse = xxsum0;
- return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
-
-
-unsigned int vp9_variance_halfpixvar16x16_hv_sse2(
- const unsigned char *src_ptr,
- int src_pixels_per_line,
- const unsigned char *dst_ptr,
- int dst_pixels_per_line,
- unsigned int *sse) {
- int xsum0;
- unsigned int xxsum0;
-
- vp9_half_horiz_vert_variance16x_h_sse2(
- src_ptr, src_pixels_per_line,
- dst_ptr, dst_pixels_per_line, 16,
- &xsum0, &xxsum0);
-
- *sse = xxsum0;
- return (xxsum0 - (((unsigned int)xsum0 * xsum0) >> 8));
-}
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-20 19:36:27 +0000