postprocess.c

	for(y=0; y<height; y+=BLOCK_SIZE)
	{
		//1% speedup if these are here instead of the inner loop
		uint8_t *srcBlock= &(src[y*srcStride]);
		uint8_t *dstBlock= &(dst[y*dstStride]);
#ifdef ARCH_X86
		int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
		int QPDelta= isColor ? 1<<(32-3) : 1<<(32-4);
		int QPFrac= QPDelta;
		uint8_t *tempBlock1= tempBlocks;
		uint8_t *tempBlock2= tempBlocks + 8;
#endif
		/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
		   if not than use a temporary buffer */
		if(y+15 >= height)
		{
			/* copy from line 5 to 12 of src, these will be copied with
			   blockcopy to dst later */
			memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5,
				srcStride*MAX(height-y-5, 0) );

			/* duplicate last line to fill the void upto line 12 */
			if(y+12 >= height)
			{
				int i;
				for(i=height-y; i<=12; i++)
					memcpy(tempSrc + srcStride*i,
						src + srcStride*(height-1), srcStride);
			}


			/* copy up to 6 lines of dst */
			memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 6) );
			dstBlock= tempDst + dstStride;
			srcBlock= tempSrc;
		}

		// From this point on it is guranteed that we can read and write 16 lines downward
		// finish 1 block before the next otherwise well might have a problem
		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
		for(x=0; x<width; x+=BLOCK_SIZE)
		{
			const int stride= dstStride;
			uint8_t *tmpXchg;
#ifdef ARCH_X86
			int 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
			int QP= isColor ?
                                QPs[(y>>3)*QPStride + (x>>3)]:
                                QPs[(y>>4)*QPStride + (x>>4)];
#endif
			if(!isColor)
			{
				QP= (QP* QPCorrecture)>>8;
				yHistogram[ srcBlock[srcStride*4 + 4] ]++;
			}
#ifdef HAVE_MMX
			asm volatile(
				"movd %0, %%mm7					\n\t"
				"packuswb %%mm7, %%mm7				\n\t" // 0, 0, 0, QP, 0, 0, 0, QP
				"packuswb %%mm7, %%mm7				\n\t" // 0,QP, 0, QP, 0,QP, 0, QP
				"packuswb %%mm7, %%mm7				\n\t" // QP,..., QP
				"movq %%mm7, pQPb				\n\t"
				: : "r" (QP)
			);
#endif

#ifdef MORE_TIMING
			T0= rdtsc();
#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);
			prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
			prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
			prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
*/
#endif

#ifdef PP_FUNNY_STRIDE
			//can we mess with a 8x16 block, if not use a temp buffer, yes again
			if(x+7 >= width)
			{
				int i;
				dstBlockPtrBackup= dstBlock;
				srcBlockPtrBackup= srcBlock;

				for(i=0;i<BLOCK_SIZE*2; i++)
				{
					memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x);
					memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x);
				}

				dstBlock= tempDstBlock;
				srcBlock= tempSrcBlock;
			}
#endif

			blockCopy(dstBlock + dstStride*5, dstStride,
				srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX);

			if(mode & LINEAR_IPOL_DEINT_FILTER)
				deInterlaceInterpolateLinear(dstBlock, dstStride);
			else if(mode & LINEAR_BLEND_DEINT_FILTER)
				deInterlaceBlendLinear(dstBlock, dstStride);
			else if(mode & MEDIAN_DEINT_FILTER)
				deInterlaceMedian(dstBlock, dstStride);
			else if(mode & CUBIC_IPOL_DEINT_FILTER)
				deInterlaceInterpolateCubic(dstBlock, dstStride);
/*			else if(mode & CUBIC_BLEND_DEINT_FILTER)
				deInterlaceBlendCubic(dstBlock, dstStride);
*/

			/* only deblock if we have 2 blocks */
			if(y + 8 < height)
			{
#ifdef MORE_TIMING
				T1= rdtsc();
				memcpyTime+= T1-T0;
				T0=T1;
#endif
				if(mode & V_RK1_FILTER)
					vertRK1Filter(dstBlock, stride, QP);
				else if(mode & V_X1_FILTER)
					vertX1Filter(dstBlock, stride, QP);
				else if(mode & V_DEBLOCK)
				{
					if( isVertDC(dstBlock, stride))
					{
						if(isVertMinMaxOk(dstBlock, stride, QP))
							doVertLowPass(dstBlock, stride, QP);
					}
					else
						doVertDefFilter(dstBlock, stride, QP);
				}
#ifdef MORE_TIMING
				T1= rdtsc();
				vertTime+= T1-T0;
				T0=T1;
#endif
			}

#ifdef HAVE_MMX
			transpose1(tempBlock1, tempBlock2, dstBlock, dstStride);
#endif
			/* check if we have a previous block to deblock it with dstBlock */
			if(x - 8 >= 0)
			{
#ifdef MORE_TIMING
				T0= rdtsc();
#endif
#ifdef HAVE_MMX
				if(mode & H_RK1_FILTER)
					vertRK1Filter(tempBlock1, 16, QP);
				else if(mode & H_X1_FILTER)
					vertX1Filter(tempBlock1, 16, QP);
				else if(mode & H_DEBLOCK)
				{
					if( isVertDC(tempBlock1, 16))
					{
						if(isVertMinMaxOk(tempBlock1, 16, QP))
							doVertLowPass(tempBlock1, 16, QP);
					}
					else
						doVertDefFilter(tempBlock1, 16, QP);
				}

				transpose2(dstBlock-4, dstStride, tempBlock1 + 4*16);

#else
				if(mode & H_X1_FILTER)
					horizX1Filter(dstBlock-4, stride, QP);
				else if(mode & H_DEBLOCK)
				{
					if( isHorizDC(dstBlock-4, stride))
					{
						if(isHorizMinMaxOk(dstBlock-4, stride, QP))
							doHorizLowPass(dstBlock-4, stride, QP);
					}
					else
						doHorizDefFilter(dstBlock-4, stride, QP);
				}
#endif
#ifdef MORE_TIMING
				T1= rdtsc();
				horizTime+= T1-T0;
				T0=T1;
#endif
				if(mode & DERING)
				{
				//FIXME filter first line
					if(y>0) dering(dstBlock - stride - 8, stride, QP);
				}
			}
			else if(mode & DERING)
			{
			 //FIXME y+15 is required cuz of the tempBuffer thing -> bottom right block isnt filtered
					if(y > 8 && y+15 < height) dering(dstBlock - stride*9 + width - 8, stride, QP);
			}


#ifdef PP_FUNNY_STRIDE
			/* did we use a tmp-block buffer */
			if(x+7 >= width)
			{
				int i;
				dstBlock= dstBlockPtrBackup;
				srcBlock= srcBlockPtrBackup;

				for(i=0;i<BLOCK_SIZE*2; i++)
				{
					memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x);
				}
			}
#endif

			dstBlock+=8;
			srcBlock+=8;

#ifdef HAVE_MMX
			tmpXchg= tempBlock1;
			tempBlock1= tempBlock2;
			tempBlock2 = tmpXchg;
#endif
		}

		/* did we use a tmp buffer */
		if(y+15 >= height)
		{
			uint8_t *dstBlock= &(dst[y*dstStride]);
			memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
		}
	}
#ifdef HAVE_3DNOW
	asm volatile("femms");
#elif defined (HAVE_MMX)
	asm volatile("emms");
#endif

#ifdef TIMING
	// FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)
	sumTime= rdtsc() - sumTime;
	if(!isColor)
		printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d    \r",
			(int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000),
			(int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000)
			, black, white);
#endif
}