# HG changeset patch # User michael # Date 1003881930 0 # Node ID de434f02dee632bba2c3e90cb628c07ac0b8c46c # Parent 18c12656c816feff5aa8771adb3bc5dfd7c6180d more speed diff -r 18c12656c816 -r de434f02dee6 postproc/postprocess.c --- a/postproc/postprocess.c Tue Oct 23 20:15:00 2001 +0000 +++ b/postproc/postprocess.c Wed Oct 24 00:05:30 2001 +0000 @@ -2603,6 +2603,7 @@ after watching a black picture for 5 hours*/ static uint64_t *yHistogram= NULL; int black=0, white=255; // blackest black and whitest white in the picture + int QPCorrecture= 256; /* Temporary buffers for handling the last row(s) */ static uint8_t *tempDst= NULL; @@ -2693,6 +2694,9 @@ packedYOffset= 0; } + if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; + else QPCorrecture= 256; + /* copy first row of 8x8 blocks */ for(x=0; x>3)*QPStride] :&QPs[(y>>4)*QPStride]; + int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); + int QPFrac= QPDelta; +#endif /* can we mess with a 8x16 block from srcBlock/dstBlock downwards, if not than use a temporary buffer */ if(y+15 >= height) @@ -2734,16 +2742,26 @@ for(x=0; x>3)*QPStride + (x>>3)]: + QPs[(y>>4)*QPStride + (x>>4)]; +#endif + if(!isColor) { - QP=QPs[(y>>3)*QPStride + (x>>3)]; - } - else - { - QP= QPs[(y>>4)*QPStride + (x>>4)]; - if(mode & LEVEL_FIX) QP= (QP* (packedYScale &0xFFFF))>>8; - yHistogram[ srcBlock[srcStride*5] ]++; + QP= (QP* QPCorrecture)>>8; + yHistogram[ srcBlock[srcStride*4 + 4] ]++; } #ifdef HAVE_MMX asm volatile( @@ -2761,10 +2779,38 @@ #endif #ifdef HAVE_MMX2 +/* prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); +*/ +/* + prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); + prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); + prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); + prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); +*/ + + asm( + "movl %4, %%eax \n\t" + "shrl $2, %%eax \n\t" + "andl $6, %%eax \n\t" + "addl $5, %%eax \n\t" + "movl %%eax, %%ebx \n\t" + "imul %1, %%eax \n\t" + "imul %3, %%ebx \n\t" + "prefetchnta 32(%%eax, %0) \n\t" + "prefetcht0 32(%%ebx, %2) \n\t" + "addl %1, %%eax \n\t" + "addl %3, %%ebx \n\t" + "prefetchnta 32(%%eax, %0) \n\t" + "prefetcht0 32(%%ebx, %2) \n\t" + :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), + "m" (x) + : "%eax", "%ebx" + ); + #elif defined(HAVE_3DNOW) //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); diff -r 18c12656c816 -r de434f02dee6 postproc/postprocess_template.c --- a/postproc/postprocess_template.c Tue Oct 23 20:15:00 2001 +0000 +++ b/postproc/postprocess_template.c Wed Oct 24 00:05:30 2001 +0000 @@ -2603,6 +2603,7 @@ after watching a black picture for 5 hours*/ static uint64_t *yHistogram= NULL; int black=0, white=255; // blackest black and whitest white in the picture + int QPCorrecture= 256; /* Temporary buffers for handling the last row(s) */ static uint8_t *tempDst= NULL; @@ -2693,6 +2694,9 @@ packedYOffset= 0; } + if(mode & LEVEL_FIX) QPCorrecture= packedYScale &0xFFFF; + else QPCorrecture= 256; + /* copy first row of 8x8 blocks */ for(x=0; x>3)*QPStride] :&QPs[(y>>4)*QPStride]; + int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4); + int QPFrac= QPDelta; +#endif /* can we mess with a 8x16 block from srcBlock/dstBlock downwards, if not than use a temporary buffer */ if(y+15 >= height) @@ -2734,16 +2742,26 @@ for(x=0; x>3)*QPStride + (x>>3)]: + QPs[(y>>4)*QPStride + (x>>4)]; +#endif + if(!isColor) { - QP=QPs[(y>>3)*QPStride + (x>>3)]; - } - else - { - QP= QPs[(y>>4)*QPStride + (x>>4)]; - if(mode & LEVEL_FIX) QP= (QP* (packedYScale &0xFFFF))>>8; - yHistogram[ srcBlock[srcStride*5] ]++; + QP= (QP* QPCorrecture)>>8; + yHistogram[ srcBlock[srcStride*4 + 4] ]++; } #ifdef HAVE_MMX asm volatile( @@ -2761,10 +2779,38 @@ #endif #ifdef HAVE_MMX2 +/* prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); +*/ +/* + prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); + prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); + prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); + prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); +*/ + + asm( + "movl %4, %%eax \n\t" + "shrl $2, %%eax \n\t" + "andl $6, %%eax \n\t" + "addl $5, %%eax \n\t" + "movl %%eax, %%ebx \n\t" + "imul %1, %%eax \n\t" + "imul %3, %%ebx \n\t" + "prefetchnta 32(%%eax, %0) \n\t" + "prefetcht0 32(%%ebx, %2) \n\t" + "addl %1, %%eax \n\t" + "addl %3, %%ebx \n\t" + "prefetchnta 32(%%eax, %0) \n\t" + "prefetcht0 32(%%ebx, %2) \n\t" + :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), + "m" (x) + : "%eax", "%ebx" + ); + #elif defined(HAVE_3DNOW) //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);