Improve the SSSE3 scaler.

This commit is contained in:
wolfbeast 2018-06-07 15:53:37 +02:00 committed by Roy Tam
commit a1fee5622b
3 changed files with 25 additions and 9 deletions

View file

@ -37,6 +37,7 @@
#include <tmmintrin.h>
#include <stdint.h>
#include <assert.h>
#include "ssse3-scaler.h"
typedef int32_t pixman_fixed_16_16_t;
typedef pixman_fixed_16_16_t pixman_fixed_t;
@ -44,6 +45,8 @@ typedef pixman_fixed_16_16_t pixman_fixed_t;
#define pixman_fixed_to_int(f) ((int) ((f) >> 16))
#define pixman_int_to_fixed(i) ((pixman_fixed_t) ((i) << 16))
#define pixman_double_to_fixed(d) ((pixman_fixed_t) ((d) * 65536.0))
#define PIXMAN_FIXED_INT_MAX 32767
#define PIXMAN_FIXED_INT_MIN -32768
typedef struct pixman_vector pixman_vector_t;
typedef int pixman_bool_t;
@ -463,6 +466,12 @@ ssse3_bilinear_cover_iter_init (pixman_iter_t *iter)
bilinear_info_t *info;
pixman_vector_t v;
if (iter->x > PIXMAN_FIXED_INT_MAX ||
iter->x < PIXMAN_FIXED_INT_MIN ||
iter->y > PIXMAN_FIXED_INT_MAX ||
iter->y < PIXMAN_FIXED_INT_MIN)
goto fail;
/* Reference point is the center of the pixel */
v.vector[0] = pixman_int_to_fixed (iter->x) + pixman_fixed_1 / 2;
v.vector[1] = pixman_int_to_fixed (iter->y) + pixman_fixed_1 / 2;
@ -505,7 +514,7 @@ fail:
/* scale the src from src_width/height to dest_width/height drawn
* into the rectangle x,y width,height
* src_stride and dst_stride are 4 byte units */
void ssse3_scale_data(uint32_t *src, int src_width, int src_height, int src_stride,
bool ssse3_scale_data(uint32_t *src, int src_width, int src_height, int src_stride,
uint32_t *dest, int dest_width, int dest_height,
int dest_stride,
int x, int y,
@ -551,6 +560,10 @@ void ssse3_scale_data(uint32_t *src, int src_width, int src_height, int src_stri
iter.data = NULL;
ssse3_bilinear_cover_iter_init(&iter);
if (!iter.fini)
return false;
if (iter.data) {
for (int iy = 0; iy < height; iy++) {
ssse3_fetch_bilinear_cover(&iter, NULL);
@ -558,4 +571,5 @@ void ssse3_scale_data(uint32_t *src, int src_width, int src_height, int src_stri
}
ssse3_bilinear_cover_iter_fini(&iter);
}
return true;
}

View file

@ -6,10 +6,12 @@
#ifndef MOZILLA_GFX_2D_SSSE3_SCALER_H_
#define MOZILLA_GFX_2D_SSSE3_SCALER_H_
#include <stdbool.h>
#ifdef __cplusplus
extern "C" {
#endif
void ssse3_scale_data(uint32_t *src, int src_width, int src_height,
bool ssse3_scale_data(uint32_t *src, int src_width, int src_height,
int src_stride,
uint32_t *dest, int dest_width, int dest_height,
int dest_rowstride,

View file

@ -470,15 +470,15 @@ AttemptVideoScale(TextureSourceBasic* aSource, const SourceSurface* aSourceMask,
RefPtr<DataSourceSurface> srcSource = aSource->GetSurface(aDest)->GetDataSurface();
DataSourceSurface::ScopedMap mapSrc(srcSource, DataSourceSurface::READ);
ssse3_scale_data((uint32_t*)mapSrc.GetData(), srcSource->GetSize().width, srcSource->GetSize().height,
mapSrc.GetStride()/4,
((uint32_t*)dstData) + fillRect.x + (dstStride / 4) * fillRect.y, dstRect.width, dstRect.height,
dstStride / 4,
offset.x, offset.y,
fillRect.width, fillRect.height);
bool success = ssse3_scale_data((uint32_t*)mapSrc.GetData(), srcSource->GetSize().width, srcSource->GetSize().height,
mapSrc.GetStride()/4,
((uint32_t*)dstData) + fillRect.x + (dstStride / 4) * fillRect.y, dstRect.width, dstRect.height,
dstStride / 4,
offset.x, offset.y,
fillRect.width, fillRect.height);
aDest->ReleaseBits(dstData);
return true;
return success;
} else
#endif // MOZILLA_SSE_HAVE_CPUID_DETECTION
return false;