postprocess.c

}

/**
 * Deinterlaces the given block
 * will be called for every 8x8 block, and can read & write into an 8x16 block
 * will shift the image up by 1 line (FIXME if this is a problem)
 */
static inline void deInterlaceBlendLinear(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1

		"movq (%0), %%mm0				\n\t" // L0
		"movq (%%eax, %1), %%mm1			\n\t" // L2
		PAVGB(%%mm1, %%mm0)				      // L0+L2
		"movq (%%eax), %%mm2				\n\t" // L1
		PAVGB(%%mm2, %%mm0)
		"movq %%mm0, (%0)				\n\t"
		"movq (%%eax, %1, 2), %%mm0			\n\t" // L3
		PAVGB(%%mm0, %%mm2)				      // L1+L3
		PAVGB(%%mm1, %%mm2)				      // 2L2 + L1 + L3
		"movq %%mm2, (%%eax)				\n\t"
		"movq (%0, %1, 4), %%mm2			\n\t" // L4
		PAVGB(%%mm2, %%mm1)				      // L2+L4
		PAVGB(%%mm0, %%mm1)				      // 2L3 + L2 + L4
		"movq %%mm1, (%%eax, %1)			\n\t"
		"movq (%%ebx), %%mm1				\n\t" // L5
		PAVGB(%%mm1, %%mm0)				      // L3+L5
		PAVGB(%%mm2, %%mm0)				      // 2L4 + L3 + L5
		"movq %%mm0, (%%eax, %1, 2)			\n\t"
		"movq (%%ebx, %1), %%mm0			\n\t" // L6
		PAVGB(%%mm0, %%mm2)				      // L4+L6
		PAVGB(%%mm1, %%mm2)				      // 2L5 + L4 + L6
		"movq %%mm2, (%0, %1, 4)			\n\t"
		"movq (%%ebx, %1, 2), %%mm2			\n\t" // L7
		PAVGB(%%mm2, %%mm1)				      // L5+L7
		PAVGB(%%mm0, %%mm1)				      // 2L6 + L5 + L7
		"movq %%mm1, (%%ebx)				\n\t"
		"movq (%0, %1, 8), %%mm1			\n\t" // L8
		PAVGB(%%mm1, %%mm0)				      // L6+L8
		PAVGB(%%mm2, %%mm0)				      // 2L7 + L6 + L8
		"movq %%mm0, (%%ebx, %1)			\n\t"
		"movq (%%ebx, %1, 4), %%mm0			\n\t" // L9
		PAVGB(%%mm0, %%mm2)				      // L7+L9
		PAVGB(%%mm1, %%mm2)				      // 2L8 + L7 + L9
		"movq %%mm2, (%%ebx, %1, 2)			\n\t"


		: : "r" (src), "r" (stride)
		: "%eax", "%ebx"
	);
#else
	int x;
	for(x=0; x<8; x++)
	{
		src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
		src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
		src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
		src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
		src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
		src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
		src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
		src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
		src++;
	}
#endif
}

/**
 * Deinterlaces the given block
 * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block
 */
static inline void deInterlaceMedian(uint8_t src[], int stride)
{
#ifdef HAVE_MMX
#ifdef HAVE_MMX2
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1

		"movq (%0), %%mm0				\n\t" //
		"movq (%%eax, %1), %%mm2			\n\t" //
		"movq (%%eax), %%mm1				\n\t" //
		"movq %%mm0, %%mm3				\n\t"
		"pmaxub %%mm1, %%mm0				\n\t" //
		"pminub %%mm3, %%mm1				\n\t" //
		"pmaxub %%mm2, %%mm1				\n\t" //
		"pminub %%mm1, %%mm0				\n\t"
		"movq %%mm0, (%%eax)				\n\t"

		"movq (%0, %1, 4), %%mm0			\n\t" //
		"movq (%%eax, %1, 2), %%mm1			\n\t" //
		"movq %%mm2, %%mm3				\n\t"
		"pmaxub %%mm1, %%mm2				\n\t" //
		"pminub %%mm3, %%mm1				\n\t" //
		"pmaxub %%mm0, %%mm1				\n\t" //
		"pminub %%mm1, %%mm2				\n\t"
		"movq %%mm2, (%%eax, %1, 2)			\n\t"

		"movq (%%ebx), %%mm2				\n\t" //
		"movq (%%ebx, %1), %%mm1			\n\t" //
		"movq %%mm2, %%mm3				\n\t"
		"pmaxub %%mm0, %%mm2				\n\t" //
		"pminub %%mm3, %%mm0				\n\t" //
		"pmaxub %%mm1, %%mm0				\n\t" //
		"pminub %%mm0, %%mm2				\n\t"
		"movq %%mm2, (%%ebx)				\n\t"

		"movq (%%ebx, %1, 2), %%mm2			\n\t" //
		"movq (%0, %1, 8), %%mm0			\n\t" //
		"movq %%mm2, %%mm3				\n\t"
		"pmaxub %%mm0, %%mm2				\n\t" //
		"pminub %%mm3, %%mm0				\n\t" //
		"pmaxub %%mm1, %%mm0				\n\t" //
		"pminub %%mm0, %%mm2				\n\t"
		"movq %%mm2, (%%ebx, %1, 2)			\n\t"


		: : "r" (src), "r" (stride)
		: "%eax", "%ebx"
	);

#else // MMX without MMX2
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1
		"pxor %%mm7, %%mm7				\n\t"

#define MEDIAN(a,b,c)\
		"movq " #a ", %%mm0				\n\t"\
		"movq " #b ", %%mm2				\n\t"\
		"movq " #c ", %%mm1				\n\t"\
		"movq %%mm0, %%mm3				\n\t"\
		"movq %%mm1, %%mm4				\n\t"\
		"movq %%mm2, %%mm5				\n\t"\
		"psubusb %%mm1, %%mm3				\n\t"\
		"psubusb %%mm2, %%mm4				\n\t"\
		"psubusb %%mm0, %%mm5				\n\t"\
		"pcmpeqb %%mm7, %%mm3				\n\t"\
		"pcmpeqb %%mm7, %%mm4				\n\t"\
		"pcmpeqb %%mm7, %%mm5				\n\t"\
		"movq %%mm3, %%mm6				\n\t"\
		"pxor %%mm4, %%mm3				\n\t"\
		"pxor %%mm5, %%mm4				\n\t"\
		"pxor %%mm6, %%mm5				\n\t"\
		"por %%mm3, %%mm1				\n\t"\
		"por %%mm4, %%mm2				\n\t"\
		"por %%mm5, %%mm0				\n\t"\
		"pand %%mm2, %%mm0				\n\t"\
		"pand %%mm1, %%mm0				\n\t"\
		"movq %%mm0, " #b "				\n\t"

MEDIAN((%0), (%%eax), (%%eax, %1))
MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1))
MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8))

		: : "r" (src), "r" (stride)
		: "%eax", "%ebx"
	);
#endif // MMX
#else
	//FIXME
	int x;
	for(x=0; x<8; x++)
	{
		src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
		src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
		src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
		src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
		src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
		src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
		src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
		src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
		src++;
	}
#endif
}

#ifdef HAVE_ODIVX_POSTPROCESS
#include "../opendivx/postprocess.h"
int use_old_pp=0;
#endif

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, int mode);

/* -pp Command line Help
NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)?

-pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]...

long form example:
-pp vdeblock:autoq,hdeblock:autoq,linblenddeint		-pp default,-vdeblock
short form example:
-pp vb:a,hb:a,lb					-pp de,-vb

Filters			Options
short	long name	short	long option	Description
*	*		a	autoq		cpu power dependant enabler
			c	chrom		chrominance filtring enabled
			y	nochrom		chrominance filtring disabled
hb	hdeblock				horizontal deblocking filter
vb	vdeblock				vertical deblocking filter
vr	rkvdeblock
h1	x1hdeblock				Experimental horizontal deblock filter 1
v1	x1vdeblock				Experimental vertical deblock filter 1
dr	dering					not implemented yet
al	autolevels				automatic brightness / contrast fixer
			f	fullyrange	stretch luminance range to (0..255)
lb	linblenddeint				linear blend deinterlacer
li	linipoldeint				linear interpolating deinterlacer
ci	cubicipoldeint				cubic interpolating deinterlacer
md	mediandeint				median deinterlacer
de	default					hdeblock:a,vdeblock:a,dering:a,autolevels
fa	fast					x1hdeblock:a,x1vdeblock:a,dering:a,autolevels
*/

/**
 * returns a PPMode struct which will have a non 0 error variable if an error occured
 * name is the string after "-pp" on the command line
 * quality is a number from 0 to GET_PP_QUALITY_MAX
 */
struct PPMode getPPModeByNameAndQuality(char *name, int quality)
{
	char temp[GET_MODE_BUFFER_SIZE];
	char *p= temp;
	char *filterDelimiters= ",";
	char *optionDelimiters= ":";
	struct PPMode ppMode= {0,0,0,0,0,0};
	char *filterToken;

	strncpy(temp, name, GET_MODE_BUFFER_SIZE);

	for(;;){
		char *p2;
		char *filterName;
		int q= GET_PP_QUALITY_MAX;
		int chrom=-1;
		char *option;
		char *options[OPTIONS_ARRAY_SIZE];
		int i;
		int filterNameOk=0;
		int numOfUnknownOptions=0;
		int enable=1; //does the user want us to enabled or disabled the filter

		filterToken= strtok(p, filterDelimiters);
		if(filterToken == NULL) break;
		p+= strlen(filterToken) + 1;
		filterName= strtok(filterToken, optionDelimiters);
		printf("%s::%s\n", filterToken, filterName);

		if(*filterName == '-')
		{
			enable=0;
			filterName++;
		}
		for(;;){ //for all options
			option= strtok(NULL, optionDelimiters);
			if(option == NULL) break;

			printf("%s\n", option);
			if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
			else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
			else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
			else
			{
				options[numOfUnknownOptions] = option;
				numOfUnknownOptions++;
				options[numOfUnknownOptions] = NULL;
			}
			if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
		}

		/* replace stuff from the replace Table */
		for(i=0; replaceTable[2*i]!=NULL; i++)
		{
			if(!strcmp(replaceTable[2*i], filterName))
			{
				int newlen= strlen(replaceTable[2*i + 1]);
				int plen;
				int spaceLeft;

				if(p==NULL) p= temp, *p=0; 	//last filter
				else p--, *p=',';		//not last filter

				plen= strlen(p);
				spaceLeft= (int)p - (int)temp + plen;
				if(spaceLeft + newlen  >= GET_MODE_BUFFER_SIZE)
				{
					ppMode.error++;
					break;
				}
				memmove(p + newlen, p, plen+1);
				memcpy(p, replaceTable[2*i + 1], newlen);
				filterNameOk=1;
			}
		}

		for(i=0; filters[i].shortName!=NULL; i++)
		{
			if(   !strcmp(filters[i].longName, filterName)
			   || !strcmp(filters[i].shortName, filterName))
			{
				ppMode.lumMode &= ~filters[i].mask;
				ppMode.chromMode &= ~filters[i].mask;

				filterNameOk=1;
				if(!enable) break; // user wants to disable it

				if(q >= filters[i].minLumQuality)
					ppMode.lumMode|= filters[i].mask;
				if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
					if(q >= filters[i].minChromQuality)
						ppMode.chromMode|= filters[i].mask;

				if(filters[i].mask == LEVEL_FIX)
				{
					int o;
					ppMode.minAllowedY= 16;
					ppMode.maxAllowedY= 234;
					for(o=0; options[o]!=NULL; o++)
						if(  !strcmp(options[o],"fullyrange")
						   ||!strcmp(options[o],"f"))
						{
							ppMode.minAllowedY= 0;
							ppMode.maxAllowedY= 255;
							numOfUnknownOptions--;
						}
				}
			}
		}
		if(!filterNameOk) ppMode.error++;
		ppMode.error += numOfUnknownOptions;
	}

	if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H;
	if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V;
	if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H;
	if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V;
	if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y;
	if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C;

	return ppMode;
}

/**
 * ...
 */
void  postprocess(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,
					  int mode)
{
/*
	static int qual=0;

	struct PPMode ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock", qual);
	qual++;
	qual%=7;
	printf("\n%d %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error);
	postprocess2(src, src_stride, dst, dst_stride,
                 horizontal_size, vertical_size, QP_store, QP_stride, &ppMode);

	return;
*/

#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, mode);

	horizontal_size >>= 1;
	vertical_size   >>= 1;
	src_stride      >>= 1;
	dst_stride      >>= 1;
	mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00);

	if(1)
	{
		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
	{
		memcpy(dst[1], src[1], src_stride*horizontal_size);
		memcpy(dst[2], src[2], src_stride*horizontal_size);
	}
}

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)
{

#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->lumMode);

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

	postProcess(src[1], src_stride, dst[1], dst_stride,
		horizontal_size, vertical_size, QP_store, QP_stride, 1, mode->chromMode);
	postProcess(src[2], src_stride, dst[2], dst_stride,
		horizontal_size, vertical_size, QP_store, QP_stride, 2, mode->chromMode);
}


/**
 * 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 numLines, int levelFix)
{
	int i;
	if(levelFix)
	{
#ifdef HAVE_MMX
					asm volatile(
						"leal (%2,%2), %%eax	\n\t"
						"leal (%3,%3), %%ebx	\n\t"
						"movq packedYOffset, %%mm2	\n\t"
						"movq packedYScale, %%mm3	\n\t"
						"pxor %%mm4, %%mm4	\n\t"

#define SCALED_CPY					\
						"movq (%0), %%mm0	\n\t"\
						"movq (%0), %%mm5	\n\t"\
						"punpcklbw %%mm4, %%mm0 \n\t"\
						"punpckhbw %%mm4, %%mm5 \n\t"\
						"psubw %%mm2, %%mm0	\n\t"\
						"psubw %%mm2, %%mm5	\n\t"\
						"movq (%0,%2), %%mm1	\n\t"\
						"psllw $6, %%mm0	\n\t"\
						"psllw $6, %%mm5	\n\t"\
						"pmulhw %%mm3, %%mm0	\n\t"\
						"movq (%0,%2), %%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"\
						"addl %%eax, %0		\n\t"\
						"packuswb %%mm5, %%mm0	\n\t"\
						"packuswb %%mm6, %%mm1	\n\t"\
						"movq %%mm0, (%1)	\n\t"\
						"movq %%mm1, (%1, %3)	\n\t"\

SCALED_CPY
						"addl %%ebx, %1		\n\t"
SCALED_CPY
						"addl %%ebx, %1		\n\t"
SCALED_CPY
						"addl %%ebx, %1		\n\t"
SCALED_CPY

						: "+r"(src),
						"+r"(dst)
						:"r" (srcStride),
						"r" (dstStride)
						: "%eax", "%ebx"
					);
#else
				for(i=0; i<numLines; i++)
					memcpy(	&(dst[dstStride*i]),
						&(src[srcStride*i]), BLOCK_SIZE);
#endif
	}
	else
	{
#ifdef HAVE_MMX
					asm volatile(
						"movl %4, %%eax \n\t"
						"movl %%eax, temp0\n\t"
						"pushl %0 \n\t"
						"pushl %1 \n\t"
						"leal (%2,%2), %%eax	\n\t"
						"leal (%3,%3), %%ebx	\n\t"
						"movq packedYOffset, %%mm2	\n\t"
						"movq packedYScale, %%mm3	\n\t"

#define SIMPLE_CPY					\
						"movq (%0), %%mm0	\n\t"\
						"movq (%0,%2), %%mm1	\n\t"\
						"movq %%mm0, (%1)	\n\t"\
						"movq %%mm1, (%1, %3)	\n\t"\

						"1:			\n\t"
SIMPLE_CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
SIMPLE_CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
						"decl temp0		\n\t"
						"jnz 1b			\n\t"

						"popl %1 \n\t"
						"popl %0 \n\t"
						: : "r" (src),
						"r" (dst),
						"r" (srcStride),
						"r" (dstStride),
						"m" (numLines>>2)
						: "%eax", "%ebx"
					);
#else
				for(i=0; i<numLines; 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, int mode)
{
	int x,y;
	/* 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;

	uint8_t *dstBlockPtrBackup;
	uint8_t *srcBlockPtrBackup;

#ifdef TIMING
	long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0;
	sumTime= rdtsc();
#endif

	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(!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 first row of 8x8 blocks */
	for(x=0; x<width; x+=BLOCK_SIZE)
		blockCopy(dst + x, dstStride, src + x, srcStride, 8, mode & LEVEL_FIX);

	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;
#endif
		/* can we mess with a 8x16 block from srcBlock/dstBlock downwards, if not
		   than use a temporary buffer */
		if(y+15 >= height)
		{
			/* copy from line 5 to 12 of src, these will e 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 5 lines of dst */
			memcpy(tempDst, dstBlock, dstStride*MIN(height-y, 5) );
			dstBlock= tempDst;
			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;
#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
			}

			/* check if we have a previous block to deblock it with dstBlock */
			if(x - 8 >= 0)
			{
#ifdef MORE_TIMING
				T0= rdtsc();
#endif
				if(mode & H_X1_FILTER)
					horizX1Filter(dstBlock-4, stride, QP);
				else if(mode & H_DEBLOCK)
				{
					if( isHorizDCAndCopy2Temp(dstBlock-4, stride))
					{
						if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP))
							doHorizLowPassAndCopyBack(dstBlock-4, stride, QP);
					}
					else
						doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP);
				}
#ifdef MORE_TIMING
				T1= rdtsc();
				horizTime+= T1-T0;
				T0=T1;
#endif
				dering(dstBlock - 9 - stride, stride, QP);
			}
			else if(y!=0)
				dering(dstBlock - stride*9 + width-9, stride, QP);
			//FIXME dering filter will not be applied to last block (bottom right)

#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;
		}

		/* did we use a tmp buffer */
		if(y+15 >= height)
		{
			uint8_t *dstBlock= &(dst[y*dstStride]);
			memcpy(dstBlock, tempDst, 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
}