summaryrefslogtreecommitdiffstats
path: root/libvo
diff options
context:
space:
mode:
authornick <nick@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-11-11 15:35:11 +0000
committernick <nick@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-11-11 15:35:11 +0000
commit1ffb6bfa2be8d8f3b5e527d7950b6d2b115c1cda (patch)
tree61d19f843d579ec1af6cd3c1216c0773845a4474 /libvo
parentb66e9843f8d864fb22a5ab39ba8776dbcaa4cf4c (diff)
downloadmpv-1ffb6bfa2be8d8f3b5e527d7950b6d2b115c1cda.tar.bz2
mpv-1ffb6bfa2be8d8f3b5e527d7950b6d2b115c1cda.tar.xz
Minor speedup for K6-2, K7, P3
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@2836 b3059339-0415-0410-9bf9-f77b7e298cf2
Diffstat (limited to 'libvo')
-rw-r--r--libvo/osd.c67
-rw-r--r--libvo/osd_template.c67
2 files changed, 68 insertions, 66 deletions
diff --git a/libvo/osd.c b/libvo/osd.c
index e32fc6bc03..41d239f563 100644
--- a/libvo/osd.c
+++ b/libvo/osd.c
@@ -85,40 +85,41 @@ PROFILE_START();
register int x;
#ifdef ARCH_X86
#ifdef HAVE_MMX
+ asm volatile(
+ PREFETCHW" %0\n\t"
+ PREFETCH" %1\n\t"
+ PREFETCH" %2\n\t"
+ "pxor %%mm7, %%mm7\n\t"
+ "pcmpeqb %%mm6, %%mm6\n\t" // F..F
+ ::"m"(dstbase),"m"(srca),"m"(src):"memory");
+ for(x=0;x<w;x+=2){
asm volatile(
- "pxor %%mm7, %%mm7 \n\t"
- "xorl %%eax, %%eax \n\t"
- "pcmpeqb %%mm6, %%mm6 \n\t" // F..F
- ".balign 16\n\t"
- "1: \n\t"
- "movq (%0, %%eax, 4), %%mm0 \n\t" // dstbase
- "movq %%mm0, %%mm1 \n\t"
- "punpcklbw %%mm7, %%mm0 \n\t"
- "punpckhbw %%mm7, %%mm1 \n\t"
- "movd (%1, %%eax), %%mm2 \n\t" // srca ABCD0000
- "paddb %%mm6, %%mm2 \n\t"
- "punpcklbw %%mm2, %%mm2 \n\t" // srca AABBCCDD
- "punpcklbw %%mm2, %%mm2 \n\t" // srca AAAABBBB
- "movq %%mm2, %%mm3 \n\t"
- "punpcklbw %%mm7, %%mm2 \n\t" // srca 0A0A0A0A
- "punpckhbw %%mm7, %%mm3 \n\t" // srca 0B0B0B0B
- "pmullw %%mm2, %%mm0 \n\t"
- "pmullw %%mm3, %%mm1 \n\t"
- "psrlw $8, %%mm0 \n\t"
- "psrlw $8, %%mm1 \n\t"
- "packuswb %%mm1, %%mm0 \n\t"
- "movd (%2, %%eax), %%mm2 \n\t" // src ABCD0000
- "punpcklbw %%mm2, %%mm2 \n\t" // src AABBCCDD
- "punpcklbw %%mm2, %%mm2 \n\t" // src AAAABBBB
- "paddb %%mm2, %%mm0 \n\t"
- "movq %%mm0, (%0, %%eax, 4) \n\t"
- "addl $2, %%eax \n\t"
- "cmpl %3, %%eax \n\t"
- " jb 1b \n\t"
-
- :: "r" (dstbase), "r" (srca), "r" (src), "r" (w)
- : "%eax"
- );
+ PREFETCHW" 32%0\n\t"
+ PREFETCH" 32%1\n\t"
+ PREFETCH" 32%2\n\t"
+ "movq %0, %%mm0\n\t" // dstbase
+ "movq %%mm0, %%mm1\n\t"
+ "punpcklbw %%mm7, %%mm0\n\t"
+ "punpckhbw %%mm7, %%mm1\n\t"
+ "movd %1, %%mm2\n\t" // srca ABCD0000
+ "paddb %%mm6, %%mm2\n\t"
+ "punpcklbw %%mm2, %%mm2\n\t" // srca AABBCCDD
+ "punpcklbw %%mm2, %%mm2\n\t" // srca AAAABBBB
+ "movq %%mm2, %%mm3\n\t"
+ "punpcklbw %%mm7, %%mm2\n\t" // srca 0A0A0A0A
+ "punpckhbw %%mm7, %%mm3\n\t" // srca 0B0B0B0B
+ "pmullw %%mm2, %%mm0\n\t"
+ "pmullw %%mm3, %%mm1\n\t"
+ "psrlw $8, %%mm0\n\t"
+ "psrlw $8, %%mm1\n\t"
+ "packuswb %%mm1, %%mm0\n\t"
+ "movd %2, %%mm2 \n\t" // src ABCD0000
+ "punpcklbw %%mm2, %%mm2\n\t" // src AABBCCDD
+ "punpcklbw %%mm2, %%mm2\n\t" // src AAAABBBB
+ "paddb %%mm2, %%mm0\n\t"
+ "movq %%mm0, %0\n\t"
+ :: "m" (dstbase[4*x]), "m" (srca[x]), "m" (src[x]));
+ }
#else /* 0 HAVE_MMX2*/
for(x=0;x<w;x++){
if(srca[x]){
diff --git a/libvo/osd_template.c b/libvo/osd_template.c
index e32fc6bc03..41d239f563 100644
--- a/libvo/osd_template.c
+++ b/libvo/osd_template.c
@@ -85,40 +85,41 @@ PROFILE_START();
register int x;
#ifdef ARCH_X86
#ifdef HAVE_MMX
+ asm volatile(
+ PREFETCHW" %0\n\t"
+ PREFETCH" %1\n\t"
+ PREFETCH" %2\n\t"
+ "pxor %%mm7, %%mm7\n\t"
+ "pcmpeqb %%mm6, %%mm6\n\t" // F..F
+ ::"m"(dstbase),"m"(srca),"m"(src):"memory");
+ for(x=0;x<w;x+=2){
asm volatile(
- "pxor %%mm7, %%mm7 \n\t"
- "xorl %%eax, %%eax \n\t"
- "pcmpeqb %%mm6, %%mm6 \n\t" // F..F
- ".balign 16\n\t"
- "1: \n\t"
- "movq (%0, %%eax, 4), %%mm0 \n\t" // dstbase
- "movq %%mm0, %%mm1 \n\t"
- "punpcklbw %%mm7, %%mm0 \n\t"
- "punpckhbw %%mm7, %%mm1 \n\t"
- "movd (%1, %%eax), %%mm2 \n\t" // srca ABCD0000
- "paddb %%mm6, %%mm2 \n\t"
- "punpcklbw %%mm2, %%mm2 \n\t" // srca AABBCCDD
- "punpcklbw %%mm2, %%mm2 \n\t" // srca AAAABBBB
- "movq %%mm2, %%mm3 \n\t"
- "punpcklbw %%mm7, %%mm2 \n\t" // srca 0A0A0A0A
- "punpckhbw %%mm7, %%mm3 \n\t" // srca 0B0B0B0B
- "pmullw %%mm2, %%mm0 \n\t"
- "pmullw %%mm3, %%mm1 \n\t"
- "psrlw $8, %%mm0 \n\t"
- "psrlw $8, %%mm1 \n\t"
- "packuswb %%mm1, %%mm0 \n\t"
- "movd (%2, %%eax), %%mm2 \n\t" // src ABCD0000
- "punpcklbw %%mm2, %%mm2 \n\t" // src AABBCCDD
- "punpcklbw %%mm2, %%mm2 \n\t" // src AAAABBBB
- "paddb %%mm2, %%mm0 \n\t"
- "movq %%mm0, (%0, %%eax, 4) \n\t"
- "addl $2, %%eax \n\t"
- "cmpl %3, %%eax \n\t"
- " jb 1b \n\t"
-
- :: "r" (dstbase), "r" (srca), "r" (src), "r" (w)
- : "%eax"
- );
+ PREFETCHW" 32%0\n\t"
+ PREFETCH" 32%1\n\t"
+ PREFETCH" 32%2\n\t"
+ "movq %0, %%mm0\n\t" // dstbase
+ "movq %%mm0, %%mm1\n\t"
+ "punpcklbw %%mm7, %%mm0\n\t"
+ "punpckhbw %%mm7, %%mm1\n\t"
+ "movd %1, %%mm2\n\t" // srca ABCD0000
+ "paddb %%mm6, %%mm2\n\t"
+ "punpcklbw %%mm2, %%mm2\n\t" // srca AABBCCDD
+ "punpcklbw %%mm2, %%mm2\n\t" // srca AAAABBBB
+ "movq %%mm2, %%mm3\n\t"
+ "punpcklbw %%mm7, %%mm2\n\t" // srca 0A0A0A0A
+ "punpckhbw %%mm7, %%mm3\n\t" // srca 0B0B0B0B
+ "pmullw %%mm2, %%mm0\n\t"
+ "pmullw %%mm3, %%mm1\n\t"
+ "psrlw $8, %%mm0\n\t"
+ "psrlw $8, %%mm1\n\t"
+ "packuswb %%mm1, %%mm0\n\t"
+ "movd %2, %%mm2 \n\t" // src ABCD0000
+ "punpcklbw %%mm2, %%mm2\n\t" // src AABBCCDD
+ "punpcklbw %%mm2, %%mm2\n\t" // src AAAABBBB
+ "paddb %%mm2, %%mm0\n\t"
+ "movq %%mm0, %0\n\t"
+ :: "m" (dstbase[4*x]), "m" (srca[x]), "m" (src[x]));
+ }
#else /* 0 HAVE_MMX2*/
for(x=0;x<w;x++){
if(srca[x]){