Add vp8_sub_pixel_variance16x8_ssse3 function

Added SSSE3 function

Change-Id: I8c304c92458618d93fda3a2f62bd09ccb63e75ad
This commit is contained in:
Yunqing Wang 2011-03-09 11:16:30 -05:00
parent 4561109a69
commit 7b8e7f0f3a
3 changed files with 56 additions and 2 deletions

View File

@ -76,8 +76,8 @@ unsigned int vp8_sub_pixel_variance16x16_ssse3
unsigned int *sse unsigned int *sse
) )
{ {
int xsum0, xsum1; int xsum0;
unsigned int xxsum0, xxsum1; unsigned int xxsum0;
// note we could avoid these if statements if the calling function // note we could avoid these if statements if the calling function
// just called the appropriate functions inside. // just called the appropriate functions inside.
@ -114,3 +114,52 @@ unsigned int vp8_sub_pixel_variance16x16_ssse3
*sse = xxsum0; *sse = xxsum0;
return (xxsum0 - ((xsum0 * xsum0) >> 8)); return (xxsum0 - ((xsum0 * xsum0) >> 8));
} }
unsigned int vp8_sub_pixel_variance16x8_ssse3
(
const unsigned char *src_ptr,
int src_pixels_per_line,
int xoffset,
int yoffset,
const unsigned char *dst_ptr,
int dst_pixels_per_line,
unsigned int *sse
)
{
int xsum0;
unsigned int xxsum0;
if (xoffset == 4 && yoffset == 0)
{
vp8_half_horiz_variance16x_h_sse2(
src_ptr, src_pixels_per_line,
dst_ptr, dst_pixels_per_line, 8,
&xsum0, &xxsum0);
}
else if (xoffset == 0 && yoffset == 4)
{
vp8_half_vert_variance16x_h_sse2(
src_ptr, src_pixels_per_line,
dst_ptr, dst_pixels_per_line, 8,
&xsum0, &xxsum0);
}
else if (xoffset == 4 && yoffset == 4)
{
vp8_half_horiz_vert_variance16x_h_sse2(
src_ptr, src_pixels_per_line,
dst_ptr, dst_pixels_per_line, 8,
&xsum0, &xxsum0);
}
else
{
vp8_filter_block2d_bil_var_ssse3(
src_ptr, src_pixels_per_line,
dst_ptr, dst_pixels_per_line, 8,
xoffset, yoffset,
&xsum0, &xxsum0);
}
*sse = xxsum0;
return (xxsum0 - ((xsum0 * xsum0) >> 7));
}

View File

@ -286,6 +286,7 @@ extern prototype_sad_multi_dif_address(vp8_sad4x4x4d_sse3);
#if HAVE_SSSE3 #if HAVE_SSSE3
extern prototype_sad_multi_same_address(vp8_sad16x16x3_ssse3); extern prototype_sad_multi_same_address(vp8_sad16x16x3_ssse3);
extern prototype_sad_multi_same_address(vp8_sad16x8x3_ssse3); extern prototype_sad_multi_same_address(vp8_sad16x8x3_ssse3);
extern prototype_subpixvariance(vp8_sub_pixel_variance16x8_ssse3);
extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3); extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3);
#if !CONFIG_RUNTIME_CPU_DETECT #if !CONFIG_RUNTIME_CPU_DETECT
@ -295,6 +296,9 @@ extern prototype_subpixvariance(vp8_sub_pixel_variance16x16_ssse3);
#undef vp8_variance_sad16x8x3 #undef vp8_variance_sad16x8x3
#define vp8_variance_sad16x8x3 vp8_sad16x8x3_ssse3 #define vp8_variance_sad16x8x3 vp8_sad16x8x3_ssse3
#undef vp8_variance_subpixvar16x8
#define vp8_variance_subpixvar16x8 vp8_sub_pixel_variance16x8_ssse3
#undef vp8_variance_subpixvar16x16 #undef vp8_variance_subpixvar16x16
#define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_ssse3 #define vp8_variance_subpixvar16x16 vp8_sub_pixel_variance16x16_ssse3

View File

@ -334,6 +334,7 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
cpi->rtcd.variance.sad16x16x3 = vp8_sad16x16x3_ssse3; cpi->rtcd.variance.sad16x16x3 = vp8_sad16x16x3_ssse3;
cpi->rtcd.variance.sad16x8x3 = vp8_sad16x8x3_ssse3; cpi->rtcd.variance.sad16x8x3 = vp8_sad16x8x3_ssse3;
cpi->rtcd.variance.subpixvar16x8 = vp8_sub_pixel_variance16x8_ssse3;
cpi->rtcd.variance.subpixvar16x16 = vp8_sub_pixel_variance16x16_ssse3; cpi->rtcd.variance.subpixvar16x16 = vp8_sub_pixel_variance16x16_ssse3;
cpi->rtcd.quantize.fastquantb = vp8_fast_quantize_b_ssse3; cpi->rtcd.quantize.fastquantb = vp8_fast_quantize_b_ssse3;