Mercurial > libavcodec.hg
changeset 4974:a2e489e40ea3 libavcodec
tweak mmx2 sad.
6% faster on core2 and k8, no change on p4.
| author | lorenm |
|---|---|
| date | Thu, 10 May 2007 22:24:19 +0000 |
| parents | 91015dc624ad |
| children | 9a6a0818e93f |
| files | i386/motion_est_mmx.c |
| diffstat | 1 files changed, 21 insertions(+), 33 deletions(-) [+] |
line wrap: on
line diff
--- a/i386/motion_est_mmx.c Thu May 10 18:44:58 2007 +0000 +++ b/i386/motion_est_mmx.c Thu May 10 22:24:19 2007 +0000 @@ -75,13 +75,11 @@ ASMALIGN(4) "1: \n\t" "movq (%1, %%"REG_a"), %%mm0 \n\t" - "movq (%2, %%"REG_a"), %%mm2 \n\t" - "psadbw %%mm2, %%mm0 \n\t" + "psadbw (%2, %%"REG_a"), %%mm0 \n\t" "add %3, %%"REG_a" \n\t" "movq (%1, %%"REG_a"), %%mm1 \n\t" - "movq (%2, %%"REG_a"), %%mm3 \n\t" - "psadbw %%mm1, %%mm3 \n\t" - "paddw %%mm3, %%mm0 \n\t" + "psadbw (%2, %%"REG_a"), %%mm1 \n\t" + "paddw %%mm1, %%mm0 \n\t" "paddw %%mm0, %%mm6 \n\t" "add %3, %%"REG_a" \n\t" " js 1b \n\t" @@ -97,17 +95,13 @@ ASMALIGN(4) "1: \n\t" "movq (%1, %%"REG_a"), %%mm0 \n\t" - "movq (%2, %%"REG_a"), %%mm2 \n\t" - "pavgb %%mm2, %%mm0 \n\t" - "movq (%3, %%"REG_a"), %%mm2 \n\t" - "psadbw %%mm2, %%mm0 \n\t" + "pavgb (%2, %%"REG_a"), %%mm0 \n\t" + "psadbw (%3, %%"REG_a"), %%mm0 \n\t" "add %4, %%"REG_a" \n\t" "movq (%1, %%"REG_a"), %%mm1 \n\t" - "movq (%2, %%"REG_a"), %%mm3 \n\t" - "pavgb %%mm1, %%mm3 \n\t" - "movq (%3, %%"REG_a"), %%mm1 \n\t" - "psadbw %%mm1, %%mm3 \n\t" - "paddw %%mm3, %%mm0 \n\t" + "pavgb (%2, %%"REG_a"), %%mm1 \n\t" + "psadbw (%3, %%"REG_a"), %%mm1 \n\t" + "paddw %%mm1, %%mm0 \n\t" "paddw %%mm0, %%mm6 \n\t" "add %4, %%"REG_a" \n\t" " js 1b \n\t" @@ -120,30 +114,24 @@ { //FIXME reuse src long len= -(stride*h); asm volatile( + "movq "MANGLE(bone)", %%mm5 \n\t" ASMALIGN(4) - "movq "MANGLE(bone)", %%mm5 \n\t" "1: \n\t" "movq (%1, %%"REG_a"), %%mm0 \n\t" - "movq (%2, %%"REG_a"), %%mm2 \n\t" "movq 1(%1, %%"REG_a"), %%mm1 \n\t" - "movq 1(%2, %%"REG_a"), %%mm3 \n\t" - "pavgb %%mm2, %%mm0 \n\t" - "pavgb %%mm1, %%mm3 \n\t" - "psubusb %%mm5, %%mm3 \n\t" - "pavgb %%mm3, %%mm0 \n\t" - "movq (%3, %%"REG_a"), %%mm2 \n\t" - "psadbw %%mm2, %%mm0 \n\t" + "pavgb (%2, %%"REG_a"), %%mm0 \n\t" + "pavgb 1(%2, %%"REG_a"), %%mm1 \n\t" + "psubusb %%mm5, %%mm1 \n\t" + "pavgb %%mm1, %%mm0 \n\t" + "psadbw (%3, %%"REG_a"), %%mm0 \n\t" "add %4, %%"REG_a" \n\t" - "movq (%1, %%"REG_a"), %%mm1 \n\t" - "movq (%2, %%"REG_a"), %%mm3 \n\t" - "movq 1(%1, %%"REG_a"), %%mm2 \n\t" - "movq 1(%2, %%"REG_a"), %%mm4 \n\t" - "pavgb %%mm3, %%mm1 \n\t" - "pavgb %%mm4, %%mm2 \n\t" - "psubusb %%mm5, %%mm2 \n\t" - "pavgb %%mm1, %%mm2 \n\t" - "movq (%3, %%"REG_a"), %%mm1 \n\t" - "psadbw %%mm1, %%mm2 \n\t" + "movq (%1, %%"REG_a"), %%mm2 \n\t" + "movq 1(%1, %%"REG_a"), %%mm3 \n\t" + "pavgb (%2, %%"REG_a"), %%mm2 \n\t" + "pavgb 1(%2, %%"REG_a"), %%mm3 \n\t" + "psubusb %%mm5, %%mm3 \n\t" + "pavgb %%mm3, %%mm2 \n\t" + "psadbw (%3, %%"REG_a"), %%mm2 \n\t" "paddw %%mm2, %%mm0 \n\t" "paddw %%mm0, %%mm6 \n\t" "add %4, %%"REG_a" \n\t"