split this huge file
optimize c versions
try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
-put fastmemcpy back
...
*/
uint64_t __attribute__((aligned(8))) pQPb;
uint64_t __attribute__((aligned(8))) pQPb2;
-
- uint64_t __attribute__((aligned(8))) mmxDcOffset;
- uint64_t __attribute__((aligned(8))) mmxDcThreshold;
+ uint64_t __attribute__((aligned(8))) mmxDcOffset[32];
+ uint64_t __attribute__((aligned(8))) mmxDcThreshold[32];
+
+ QP_STORE_T *nonBQPTable;
+
int QP;
- int dcOffset;
- int dcThreshold;
+ int nonBQP;
int frameNum;
{
int numEq= 0;
int y;
- const int dcOffset= c->dcOffset;
- const int dcThreshold= c->dcThreshold;
+ const int dcOffset= ((c->QP*c->ppMode.baseDcDiff)>>8) + 1;
+ const int dcThreshold= dcOffset*2 + 1;
for(y=0; y<BLOCK_SIZE; y++)
{
if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
int numEq= 0;
int y;
- const int dcOffset= c->dcOffset;
- const int dcThreshold= c->dcThreshold;
+ const int dcOffset= ((c->QP*c->ppMode.baseDcDiff)>>8) + 1;
+ const int dcThreshold= dcOffset*2 + 1;
src+= stride*4; // src points to begin of the 8x8 Block
for(y=0; y<BLOCK_SIZE-1; y++)
{
else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK)
{
int o;
- ppMode.maxDcDiff=1;
+ ppMode.baseDcDiff=256/4;
// hFlatnessThreshold= 40;
// vFlatnessThreshold= 40;
if(tail==options[o]) break;
numOfUnknownOptions--;
- if(o==0) ppMode.maxDcDiff= val;
+ if(o==0) ppMode.baseDcDiff= val;
else ppMode.flatnessThreshold= val;
}
}
void *getPPContext(int width, int height){
PPContext *c= memalign(32, sizeof(PPContext));
int i;
+ int mbWidth = (width+15)>>4;
+ int mbHeight= (height+15)>>4;
c->tempBlocks= (uint8_t*)memalign(8, 2*16*8);
c->yHistogram= (uint64_t*)memalign(8, 256*sizeof(uint64_t));
c->tempDstBlock= (uint8_t*)memalign(8, 1024*24);
c->tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
c->deintTemp= (uint8_t*)memalign(8, width+16);
+ c->nonBQPTable= (QP_STORE_T*)memalign(8, mbWidth*mbHeight*sizeof(QP_STORE_T));
+ memset(c->nonBQPTable, 0, mbWidth*mbHeight*sizeof(QP_STORE_T));
c->frameNum=-1;
free(c->tempDstBlock);
free(c->tempSrcBlock);
free(c->deintTemp);
+ free(c->nonBQPTable);
free(c);
}
void postprocess(uint8_t * src[3], int srcStride[3],
uint8_t * dst[3], int dstStride[3],
- int horizontalSize, int verticalSize,
+ int width, int height,
QP_STORE_T *QP_store, int QPStride,
- PPMode *mode, void *c)
+ PPMode *mode, void *vc, int pict_type)
{
-
+ int mbWidth = (width+15)>>4;
+ int mbHeight= (height+15)>>4;
QP_STORE_T quantArray[2048/8];
+ PPContext *c = (PPContext*)vc;
if(QP_store==NULL || (mode->lumMode & FORCE_QUANT))
{
else
for(i=0; i<2048/8; i++) quantArray[i]= 1;
}
+if(0){
+int x,y;
+for(y=0; y<mbHeight; y++){
+ for(x=0; x<mbWidth; x++){
+ printf("%2d ", QP_store[x + y*QPStride]);
+ }
+ printf("\n");
+}
+ printf("\n");
+}
+//printf("pict_type:%d\n", pict_type);
+ if(pict_type!=3)
+ {
+ int x,y;
+ for(y=0; y<mbHeight; y++){
+ for(x=0; x<mbWidth; x++){
+ int qscale= QP_store[x + y*QPStride];
+ if(qscale&~31)
+ qscale=31;
+ c->nonBQPTable[y*mbWidth + x]= qscale;
+ }
+ }
+ }
if(firstTime2 && verbose)
{
}
postProcess(src[0], srcStride[0], dst[0], dstStride[0],
- horizontalSize, verticalSize, QP_store, QPStride, 0, mode, c);
+ width, height, QP_store, QPStride, 0, mode, c);
- horizontalSize = (horizontalSize+1)>> 1;
- verticalSize = (verticalSize+1)>>1;
+ width = (width +1)>>1;
+ height = (height+1)>>1;
if(mode->chromMode)
{
postProcess(src[1], srcStride[1], dst[1], dstStride[1],
- horizontalSize, verticalSize, QP_store, QPStride, 1, mode, c);
+ width, height, QP_store, QPStride, 1, mode, c);
postProcess(src[2], srcStride[2], dst[2], dstStride[2],
- horizontalSize, verticalSize, QP_store, QPStride, 2, mode, c);
+ width, height, QP_store, QPStride, 2, mode, c);
}
else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
{
- memcpy(dst[1], src[1], srcStride[1]*verticalSize);
- memcpy(dst[2], src[2], srcStride[2]*verticalSize);
+ memcpy(dst[1], src[1], srcStride[1]*height);
+ memcpy(dst[2], src[2], srcStride[2]*height);
}
else
{
int y;
- for(y=0; y<verticalSize; y++)
+ for(y=0; y<height; y++)
{
- memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), horizontalSize);
- memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), horizontalSize);
+ memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
+ memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
}
}
}
"leal (%1, %2), %%eax \n\t"
// 0 1 2 3 4 5 6 7 8 9
// %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
- "movq %3, %%mm7 \n\t" // mm7 = 0x7F
- "movq %4, %%mm6 \n\t" // mm6 = 0x7D
+ "movq %3, %%mm7 \n\t"
+ "movq %4, %%mm6 \n\t"
+
"movq (%1), %%mm0 \n\t"
"movq (%%eax), %%mm1 \n\t"
"psubb %%mm1, %%mm0 \n\t" // mm0 = differnece
#endif
"movd %%mm0, %0 \n\t"
: "=r" (numEq)
- : "r" (src), "r" (stride), "m" (c->mmxDcOffset), "m" (c->mmxDcThreshold)
+ : "r" (src), "r" (stride), "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
: "%eax"
);
numEq= (-numEq) &0xFF;
);
return isOk==0;
#else
+#if 1
int x;
const int QP= c->QP;
src+= stride*3;
}
return 1;
+#else
+ int x;
+ const int QP= c->QP;
+ src+= stride*4;
+ for(x=0; x<BLOCK_SIZE; x++)
+ {
+ int min=255;
+ int max=0;
+ int y;
+ for(y=0; y<8; y++){
+ int v= src[x + y*stride];
+ if(v>max) max=v;
+ if(v<min) min=v;
+ }
+ if(max-min > 2*QP) return 0;
+ }
+ return 1;
+#endif
#endif
}
int black=0, white=255; // blackest black and whitest white in the picture
int QPCorrecture= 256*256;
- int copyAhead;
+ int copyAhead, i;
//FIXME remove
uint64_t * const yHistogram= c.yHistogram;
uint8_t * const tempSrc= c.tempSrc;
uint8_t * const tempDst= c.tempDst;
-
- c.dcOffset= c.ppMode.maxDcDiff;
- c.dcThreshold= c.ppMode.maxDcDiff*2 + 1;
+ const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
#ifdef HAVE_MMX
- c.mmxDcOffset= 0x7F - c.dcOffset;
- c.mmxDcThreshold= 0x7F - c.dcThreshold;
-
- c.mmxDcOffset*= 0x0101010101010101LL;
- c.mmxDcThreshold*= 0x0101010101010101LL;
+ for(i=0; i<32; i++){
+ int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
+ int threshold= offset*2 + 1;
+ c.mmxDcOffset[i]= 0x7F - offset;
+ c.mmxDcThreshold[i]= 0x7F - threshold;
+ c.mmxDcOffset[i]*= 0x0101010101010101LL;
+ c.mmxDcThreshold[i]*= 0x0101010101010101LL;
+ }
#endif
if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
uint8_t *tempBlock1= c.tempBlocks;
uint8_t *tempBlock2= c.tempBlocks + 8;
#endif
-#ifdef ARCH_X86
int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
- int QPDelta= isColor ? (-1) : 1<<31;
- int QPFrac= 1<<30;
-#endif
+ int *nonBQPptr= isColor ? &c.nonBQPTable[(y>>3)*mbWidth] :&c.nonBQPTable[(y>>4)*mbWidth];
int QP=0;
/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
if not than use a temporary buffer */
#ifdef HAVE_MMX
uint8_t *tmpXchg;
#endif
-#ifdef ARCH_X86
- QP= *QPptr;
- asm volatile(
- "addl %2, %1 \n\t"
- "sbbl %%eax, %%eax \n\t"
- "shll $2, %%eax \n\t"
- "subl %%eax, %0 \n\t"
- : "+r" (QPptr), "+m" (QPFrac)
- : "r" (QPDelta)
- : "%eax"
- );
-#else
- QP= isColor ?
- QPs[(y>>3)*QPStride + (x>>3)]:
- QPs[(y>>4)*QPStride + (x>>4)];
-#endif
- if(!isColor)
+ if(isColor)
+ {
+ QP= QPptr[x>>3];
+ c.nonBQP= nonBQPptr[x>>3];
+ }
+ else
{
+ QP= QPptr[x>>4];
QP= (QP* QPCorrecture + 256*128)>>16;
+ c.nonBQP= nonBQPptr[x>>4];
+ c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
yHistogram[ srcBlock[srcStride*12 + 4] ]++;
}
-//printf("%d ", QP);
c.QP= QP;
#ifdef HAVE_MMX
asm volatile(