generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Adjust half-pixel only search

Browse Source 
Changed motion search in vp8_find_best_half_pixel_step() to be the
same as in vp8_find_best_sub_pixel_step(), which checks 5 points
instead of 8 points. This only affects real-time mode with
cpu-used >=9. Tests showed it gives 2% encoding speedup with
a quality loss(psnr) of up to 0.5%.

Change-Id: I16049cad1535002346d46cfdfad345bfc3dc5146
This commit is contained in:
Yunqing Wang 2011-08-03 11:51:07 -04:00
parent 56b06aef6d
commit b9f19f8917
1 changed files with 14 additions and 64 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

82
vp8/encoder/mcomp.c
View File

@ -661,6 +661,7 @@ int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
unsigned char *z = (*(b->base_src) + b->src);
int left, right, up, down, diag;
unsigned int sse;
int whichdir ;
int thismse;
int y_stride;
@ -740,8 +741,6 @@ int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
*sse1 = sse;
}
// somewhat strangely not doing all the diagonals for half pel is slower than doing them.
#if 0
// now check 1 more diagonal -
whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
this_mv = startmv;
@ -749,76 +748,28 @@ int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
switch (whichdir)
{
case 0:
this_mv.col = (this_mv.col - 8) | 4;
this_mv.row = (this_mv.row - 8) | 4;
diag = vfp->svf(y - 1 - y_stride, y_stride, 4, 4, z, b->src_stride, &sse);
break;
case 1:
this_mv.col += 4;
this_mv.row = (this_mv.row - 8) | 4;
diag = vfp->svf(y - y_stride, y_stride, 4, 4, z, b->src_stride, &sse);
break;
case 2:
this_mv.col = (this_mv.col - 8) | 4;
this_mv.row += 4;
diag = vfp->svf(y - 1, y_stride, 4, 4, z, b->src_stride, &sse);
break;
case 3:
this_mv.col += 4;
this_mv.row += 4;
diag = vfp->svf(y, y_stride, 4, 4, z, b->src_stride, &sse);
break;
}
diag += mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
}
#else
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 8;
break;
case 1:
this_mv.as_mv.col += 4;
this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
break;
case 2:
this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
this_mv.as_mv.row = startmv.as_mv.row + 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
{
*bestmv = this_mv;
bestmse = diag;
*distortion = thismse;
*sse1 = sse;
}
this_mv.as_mv.col += 8;
break;
case 3:
default:
this_mv.as_mv.col += 4;
this_mv.as_mv.row += 4;
thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
break;
}
diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
if (diag < bestmse)
@ -829,7 +780,6 @@ int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
*sse1 = sse;
}
#endif
return bestmse;
}