postprocess.c

	qual++;
	qual%=7;
	printf("\n%X %X %X %X :%d: %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error,
		qual, ppMode.maxTmpNoise[0], ppMode.maxTmpNoise[1], ppMode.maxTmpNoise[2]);
	postprocess2(src, src_stride, dst, dst_stride,
                 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode);

	return;
*/
	if(QP_store==NULL)
	{
		QP_store= zeroArray;
		QP_stride= 0;
	}

	ppMode.lumMode= mode;
	mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00);
	ppMode.chromMode= mode;
	ppMode.maxTmpNoise[0]= 700;
	ppMode.maxTmpNoise[1]= 1500;
	ppMode.maxTmpNoise[2]= 3000;

#ifdef HAVE_ODIVX_POSTPROCESS
// Note: I could make this shit outside of this file, but it would mean one
// more function call...
	if(use_old_pp){
	    odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode);
	    return;
	}
#endif

	postProcess(src[0], src_stride, dst[0], dst_stride,
		horizontal_size, vertical_size, QP_store, QP_stride, 0, &ppMode);

	horizontal_size >>= 1;
	vertical_size   >>= 1;
	src_stride      >>= 1;
	dst_stride      >>= 1;

	if(ppMode.chromMode)
	{
		postProcess(src[1], src_stride, dst[1], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode);
		postProcess(src[2], src_stride, dst[2], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 2, &ppMode);
	}
	else if(src_stride == dst_stride)
	{
		memcpy(dst[1], src[1], src_stride*vertical_size);
		memcpy(dst[2], src[2], src_stride*vertical_size);
	}
	else
	{
		int y;
		for(y=0; y<vertical_size; y++)
		{
			memcpy(&(dst[1][y*dst_stride]), &(src[1][y*src_stride]), horizontal_size);
			memcpy(&(dst[2][y*dst_stride]), &(src[2][y*src_stride]), horizontal_size);
		}
	}

#if 0
		memset(dst[1], 128, dst_stride*vertical_size);
		memset(dst[2], 128, dst_stride*vertical_size);
#endif
}

void  postprocess2(unsigned char * src[], int src_stride,
                 unsigned char * dst[], int dst_stride,
                 int horizontal_size,   int vertical_size,
                 QP_STORE_T *QP_store,  int QP_stride,
		 struct PPMode *mode)
{

	static QP_STORE_T zeroArray[2048/8];
	if(QP_store==NULL)
	{
		QP_store= zeroArray;
		QP_stride= 0;
	}

#ifdef HAVE_ODIVX_POSTPROCESS
// Note: I could make this shit outside of this file, but it would mean one
// more function call...
	if(use_old_pp){
	    odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,
	    mode->oldMode);
	    return;
	}
#endif

	postProcess(src[0], src_stride, dst[0], dst_stride,
		horizontal_size, vertical_size, QP_store, QP_stride, 0, mode);

	horizontal_size >>= 1;
	vertical_size   >>= 1;
	src_stride      >>= 1;
	dst_stride      >>= 1;

	if(mode->chromMode)
	{
		postProcess(src[1], src_stride, dst[1], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 1, mode);
		postProcess(src[2], src_stride, dst[2], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 2, mode);
	}
	else if(src_stride == dst_stride)
	{
		memcpy(dst[1], src[1], src_stride*vertical_size);
		memcpy(dst[2], src[2], src_stride*vertical_size);
	}
	else
	{
		int y;
		for(y=0; y<vertical_size; y++)
		{
			memcpy(&(dst[1][y*dst_stride]), &(src[1][y*src_stride]), horizontal_size);
			memcpy(&(dst[2][y*dst_stride]), &(src[2][y*src_stride]), horizontal_size);
		}
	}
}


/**
 * gets the mode flags for a given quality (larger values mean slower but better postprocessing)
 * 0 <= quality <= 6
 */
int getPpModeForQuality(int quality){
	int modes[1+GET_PP_QUALITY_MAX]= {
		0,
#if 1
		// horizontal filters first
		LUM_H_DEBLOCK,
		LUM_H_DEBLOCK | LUM_V_DEBLOCK,
		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK,
		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK,
		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING,
		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING
#else
		// vertical filters first
		LUM_V_DEBLOCK,
		LUM_V_DEBLOCK | LUM_H_DEBLOCK,
		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK,
		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK,
		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING,
		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING
#endif
	};

#ifdef HAVE_ODIVX_POSTPROCESS
	int odivx_modes[1+GET_PP_QUALITY_MAX]= {
		0,
		PP_DEBLOCK_Y_H,
		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V,
		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H,
		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V,
		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y,
		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C
	};
	if(use_old_pp) return odivx_modes[quality];
#endif
	return modes[quality];
}

/**
 * Copies a block from src to dst and fixes the blacklevel
 * numLines must be a multiple of 4
 * levelFix == 0 -> dont touch the brighness & contrast
 */
static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
	int levelFix)
{
#ifndef HAVE_MMX
	int i;
#endif
	if(levelFix)
	{
#ifdef HAVE_MMX
					asm volatile(
						"leal (%0,%2), %%eax	\n\t"
						"leal (%1,%3), %%ebx	\n\t"
						"movq packedYOffset, %%mm2	\n\t"
						"movq packedYScale, %%mm3	\n\t"
						"pxor %%mm4, %%mm4	\n\t"

#define SCALED_CPY(src1, src2, dst1, dst2)					\
						"movq " #src1 ", %%mm0	\n\t"\
						"movq " #src1 ", %%mm5	\n\t"\
						"punpcklbw %%mm4, %%mm0 \n\t"\
						"punpckhbw %%mm4, %%mm5 \n\t"\
						"psubw %%mm2, %%mm0	\n\t"\
						"psubw %%mm2, %%mm5	\n\t"\
						"movq " #src2 ", %%mm1	\n\t"\
						"psllw $6, %%mm0	\n\t"\
						"psllw $6, %%mm5	\n\t"\
						"pmulhw %%mm3, %%mm0	\n\t"\
						"movq " #src2 ", %%mm6	\n\t"\
						"pmulhw %%mm3, %%mm5	\n\t"\
						"punpcklbw %%mm4, %%mm1 \n\t"\
						"punpckhbw %%mm4, %%mm6 \n\t"\
						"psubw %%mm2, %%mm1	\n\t"\
						"psubw %%mm2, %%mm6	\n\t"\
						"psllw $6, %%mm1	\n\t"\
						"psllw $6, %%mm6	\n\t"\
						"pmulhw %%mm3, %%mm1	\n\t"\
						"pmulhw %%mm3, %%mm6	\n\t"\
						"packuswb %%mm5, %%mm0	\n\t"\
						"packuswb %%mm6, %%mm1	\n\t"\
						"movq %%mm0, " #dst1 "	\n\t"\
						"movq %%mm1, " #dst2 "	\n\t"\

SCALED_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
SCALED_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
SCALED_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
						"leal (%%eax,%2,4), %%eax	\n\t"
						"leal (%%ebx,%3,4), %%ebx	\n\t"
SCALED_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))


						: : "r"(src),
						"r"(dst),
						"r" (srcStride),
						"r" (dstStride)
						: "%eax", "%ebx"
					);
#else
				for(i=0; i<8; i++)
					memcpy(	&(dst[dstStride*i]),
						&(src[srcStride*i]), BLOCK_SIZE);
#endif
	}
	else
	{
#ifdef HAVE_MMX
					asm volatile(
						"leal (%0,%2), %%eax	\n\t"
						"leal (%1,%3), %%ebx	\n\t"

#define SIMPLE_CPY(src1, src2, dst1, dst2)				\
						"movq " #src1 ", %%mm0	\n\t"\
						"movq " #src2 ", %%mm1	\n\t"\
						"movq %%mm0, " #dst1 "	\n\t"\
						"movq %%mm1, " #dst2 "	\n\t"\

SIMPLE_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
						"leal (%%eax,%2,4), %%eax	\n\t"
						"leal (%%ebx,%3,4), %%ebx	\n\t"
SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))

						: : "r" (src),
						"r" (dst),
						"r" (srcStride),
						"r" (dstStride)
						: "%eax", "%ebx"
					);
#else
				for(i=0; i<8; i++)
					memcpy(	&(dst[dstStride*i]),
						&(src[srcStride*i]), BLOCK_SIZE);
#endif
	}
}


/**
 * Filters array of bytes (Y or U or V values)
 */
static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
	QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode)
{
	int x,y;
	const int mode= isColor ? ppMode->chromMode : ppMode->lumMode;

	/* we need 64bit here otherwise well going to have a problem
	   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;
	static uint8_t *tempSrc= NULL;

	/* Temporary buffers for handling the last block */
	static uint8_t *tempDstBlock= NULL;
	static uint8_t *tempSrcBlock= NULL;

	/* Temporal noise reducing buffers */
	static uint8_t *tempBlured[3]= {NULL,NULL,NULL};
	static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL};

	int copyAhead;

#ifdef PP_FUNNY_STRIDE
	uint8_t *dstBlockPtrBackup;
	uint8_t *srcBlockPtrBackup;
#endif

#ifdef MORE_TIMING
	long long T0, T1, diffTime=0;
#endif
#ifdef TIMING
	long long memcpyTime=0, vertTime=0, horizTime=0, sumTime;
	sumTime= rdtsc();
#endif
//mode= 0x7F;
#ifdef HAVE_MMX
	maxTmpNoise[0]= ppMode->maxTmpNoise[0];
	maxTmpNoise[1]= ppMode->maxTmpNoise[1];
	maxTmpNoise[2]= ppMode->maxTmpNoise[2];
#endif

	if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
	else if(mode & LINEAR_BLEND_DEINT_FILTER) copyAhead=14;
	else if(   (mode & V_DEBLOCK)
		|| (mode & LINEAR_IPOL_DEINT_FILTER)
		|| (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
	else if(mode & V_X1_FILTER) copyAhead=11;
	else if(mode & V_RK1_FILTER) copyAhead=10;
	else if(mode & DERING) copyAhead=9;
	else copyAhead=8;

	copyAhead-= 8;

	if(tempDst==NULL)
	{
		tempDst= (uint8_t*)memalign(8, 1024*24);
		tempSrc= (uint8_t*)memalign(8, 1024*24);
		tempDstBlock= (uint8_t*)memalign(8, 1024*24);
		tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
	}

	if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER))
	{
//		printf("%d %d %d\n", isColor, dstStride, height);
		//FIXME works only as long as the size doesnt increase
		//Note:the +17*1024 is just there so i dont have to worry about r/w over te end
		tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024);
		tempBluredPast[isColor]= (uint32_t*)memalign(8, 256*((height+7)&(~7))/2 + 17*1024);

		memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024);
		memset(tempBluredPast[isColor], 0, 256*((height+7)&(~7))/2 + 17*1024);
	}

	if(!yHistogram)
	{
		int i;
		yHistogram= (uint64_t*)malloc(8*256);
		for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;

		if(mode & FULL_Y_RANGE)
		{
			maxAllowedY=255;
			minAllowedY=0;
		}
	}

	if(!isColor)
	{
		uint64_t sum= 0;
		int i;
		static int framenum= -1;
		uint64_t maxClipped;
		uint64_t clipped;
		double scale;

		framenum++;
		if(framenum == 1) yHistogram[0]= width*height/64*15/256;

		for(i=0; i<256; i++)
		{
			sum+= yHistogram[i];
//			printf("%d ", yHistogram[i]);
		}
//		printf("\n\n");

		/* we allways get a completly black picture first */
		maxClipped= (uint64_t)(sum * maxClippedThreshold);

		clipped= sum;
		for(black=255; black>0; black--)
		{
			if(clipped < maxClipped) break;
			clipped-= yHistogram[black];
		}

		clipped= sum;
		for(white=0; white<256; white++)
		{
			if(clipped < maxClipped) break;
			clipped-= yHistogram[white];
		}

		packedYOffset= (black - minAllowedY) & 0xFFFF;
		packedYOffset|= packedYOffset<<32;
		packedYOffset|= packedYOffset<<16;

		scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black);

		packedYScale= (uint16_t)(scale*1024.0 + 0.5);
		packedYScale|= packedYScale<<32;
		packedYScale|= packedYScale<<16;
	}
	else
	{
		packedYScale= 0x0100010001000100LL;
		packedYOffset= 0;
	}

	if(mode & LEVEL_FIX)	QPCorrecture= packedYScale &0xFFFF;
	else			QPCorrecture= 256;

	/* copy & deinterlace first row of blocks */
	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]);

		dstBlock= tempDst + dstStride;

		// 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)
		{

#ifdef HAVE_MMX2
/*
			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), "m" (copyAhead)
			: "%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

			blockCopy(dstBlock + dstStride*copyAhead, dstStride,
				srcBlock + srcStride*copyAhead, srcStride, 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);
*/
			dstBlock+=8;
			srcBlock+=8;
		}
		memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, copyAhead*dstStride );
	}

	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
		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 */
		if(y+15 >= height)
		{
			int i;
			/* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
			   blockcopy to dst later */
			memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
				srcStride*MAX(height-y-copyAhead, 0) );

			/* duplicate last line of src to fill the void upto line (copyAhead+7) */
			for(i=MAX(height-y, 8); i<copyAhead+8; i++)
				memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);

			/* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
			memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );

			/* duplicate last line of dst to fill the void upto line (copyAhead) */
			for(i=height-y+1; i<=copyAhead; i++)
				memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);

			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
			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)
			{
				QP= (QP* QPCorrecture)>>8;
				yHistogram[ srcBlock[srcStride*12 + 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>>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), "m" (copyAhead)
			: "%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*copyAhead, dstStride,
				srcBlock + srcStride*copyAhead, srcStride, 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);
				}

				if(mode & TEMP_NOISE_FILTER)
				{
					tempNoiseReducer(dstBlock-8, stride,
						tempBlured[isColor] + y*dstStride + x,
						tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
						ppMode->maxTmpNoise);
				}
			}

#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
		}

		if(mode & DERING)
		{
				if(y > 0) dering(dstBlock - dstStride - 8, dstStride, QP);
		}

		if((mode & TEMP_NOISE_FILTER))
		{
			tempNoiseReducer(dstBlock-8, dstStride,
				tempBlured[isColor] + y*dstStride + x,
				tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
				ppMode->maxTmpNoise);
		}

		/* did we use a tmp buffer for the last lines*/
		if(y+15 >= height)
		{
			uint8_t *dstBlock= &(dst[y*dstStride]);
			memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
		}
/*
		for(x=0; x<width; x+=32)
		{
			volatile int i;
			i+=	+ dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
				+ dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
				+ dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
//				+ dstBlock[x +13*dstStride]
//				+ dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
		}*/
	}
#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
#ifdef DEBUG_BRIGHTNESS
	if(!isColor)
	{
		int max=1;
		int i;
		for(i=0; i<256; i++)
			if(yHistogram[i] > max) max=yHistogram[i];

		for(i=1; i<256; i++)
		{
			int x;
			int start=yHistogram[i-1]/(max/256+1);
			int end=yHistogram[i]/(max/256+1);
			int inc= end > start ? 1 : -1;
			for(x=start; x!=end+inc; x+=inc)
				dst[ i*dstStride + x]+=128;
		}

		for(i=0; i<100; i+=2)
		{
			dst[ (white)*dstStride + i]+=128;
			dst[ (black)*dstStride + i]+=128;
		}

	}
#endif

}