comparison libpostproc/postprocess_template.c @ 97:e57b1d38d71f libavcodec

bugfixes: last 3 lines not brightness/contrast corrected brightness statistics messed up with initial black pic changed initial values of the brightness statistics C++ -> C conversation QP range question solved (very likely 1<=QP<=32 according to arpi) new experimental vertical deblocking filter RK filter has 3dNow support now (untested)

96:29ac11dc53d3

isHorizDC		Ec	Ec
isHorizMinMaxOk		a
doHorizLowPass		E		a	a*
doHorizDefFilter	E	ac	ac
deRing

* i dont have a 3dnow CPU -> its untested
E = Exact implementation
e = allmost exact implementation
a = alternative / approximate impl

TODO:
verify that everything workes as it should
reduce the time wasted on the mem transfer
implement dering
implement everything in C at least (done at the moment but ...)
figure range of QP out (assuming <256 for now)
unroll stuff if instructions depend too much on the prior one
we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4?
move YScale thing to the end instead of fixing QP
write a faster and higher quality deblocking filter :)
...

Notes:

*/

/*
Changelog:
0.1.2
	fixed a bug in the horizontal default filter
	3dnow version of the Horizontal & Vertical Lowpass filters
	mmx version of the Horizontal Default filter
	mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar

*/


#include <inttypes.h>
#include <stdio.h>
#include "../config.h"
//#undef HAVE_MMX2
//#define HAVE_3DNOW
//#undef HAVE_MMX
#include "postprocess.h"

//FIXME? |255-0| = 1 (shouldnt be a problem ...)
/**
 * Check if the middle 8x8 Block in the given 8x10 block is flat
 */
static inline bool isVertDC(uint8_t src[], int stride){
//	return true;
	int numEq= 0;
	src+= stride; // src points to begin of the 8x8 Block
#ifdef HAVE_MMX
	asm volatile(
//		"int $3 \n\t"
		"pushl %1\n\t"

//	int asmEq= numEq;
//	numEq=0;
//	uint8_t *temp= src;

#else
	for(int y=0; y<BLOCK_SIZE-1; y++)
	{
		if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++;

			}
			printf("\n");
		}
	}
*/
	return numEq > vFlatnessThreshold;
}

static inline bool isVertMinMaxOk(uint8_t src[], int stride, int QP)
{
#ifdef HAVE_MMX
	int isOk;
	asm volatile(
//		"int $3 \n\t"

//		"movd %%mm0, (%1, %2, 4)\n\t"
		"movd %%mm0, %0					\n\t"
		: "=r" (isOk)
		: "r" (src), "r" (stride)
		);
	return isOk;
#else

	int isOk2= true;
	for(int x=0; x<BLOCK_SIZE; x++)
	{
		if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=false;
	}
/*	if(isOk && !isOk2 || !isOk && isOk2)
	{
		printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);
		for(int y=0; y<9; y++)

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
	const int l9= stride + l8;

	for(int x=0; x<BLOCK_SIZE; x++)
	{
		const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];
		const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];

		int sums[9];

	x/8 = 1
	1 12 12 23
 */
static inline void vertRKFilter(uint8_t *src, int stride, int QP)
{
#ifdef HAVE_MMX2
// FIXME rounding
	asm volatile(
		"pxor %%mm7, %%mm7				\n\t" // 0
		"movq b80, %%mm6				\n\t" // MIN_SIGNED_BYTE
		"leal (%0, %1), %%eax				\n\t"

		"movq (%0, %1, 4), %%mm2			\n\t" // line 4
		"movq (%%ebx), %%mm3				\n\t" // line 5
		"movq %%mm2, %%mm4				\n\t" // line 4
		"pcmpeqb %%mm5, %%mm5				\n\t" // -1
		"pxor %%mm2, %%mm5				\n\t" // -line 4 - 1
		"pavgb %%mm3, %%mm5				\n\t"
		"paddb %%mm6, %%mm5				\n\t" // (l5-l4)/2
		"psubusb %%mm3, %%mm4				\n\t"
		"psubusb %%mm2, %%mm3				\n\t"
		"por %%mm3, %%mm4				\n\t" // |l4 - l5|
		"psubusb %%mm0, %%mm4				\n\t"

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
	const int l9= stride + l8;
	for(int x=0; x<BLOCK_SIZE; x++)
	{
		if(ABS(src[l4]-src[l5]) < QP + QP/4)
		{
			int x = src[l5] - src[l4];

			src[l3] +=x/8;
			src[l4] +=x/2;
			src[l5] -=x/2;
			src[l6] -=x/8;
		}
		src++;
	}

#endif
}

/**
 * Experimental Filter 1
 */
static inline void vertX1Filter(uint8_t *src, int stride, int QP)
{
#ifdef HAVE_MMX2X
// FIXME
	asm volatile(

		:
		: "r" (src), "r" (stride)
		: "%eax", "%ebx"
	);
#else
 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;

			src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16;
			src[l6] = (       1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16;
		}
		src++;
	}

#endif
}


static inline void doVertDefFilter(uint8_t src[], int stride, int QP)

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
//	const int l9= stride + l8;

	for(int x=0; x<BLOCK_SIZE; x++)
	{
		const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
		if(ABS(middleEnergy) < 8*QP)
		{
			const int q=(src[l4] - src[l5])/2;

//FIXME?  |255-0| = 1
/**
 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock.
 */
static inline bool isHorizDCAndCopy2Temp(uint8_t src[], int stride)
{
//	src++;
	int numEq= 0;
#ifdef HAVE_MMX
asm volatile (

		: "%eax"
		);
//	printf("%d\n", numEq);
	numEq= (256 - (numEq & 0xFF)) &0xFF;
#else
	for(int y=0; y<BLOCK_SIZE; y++)
	{
		if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;

*/
//	printf("%d\n", numEq);
	return numEq > hFlatnessThreshold;
}

static inline bool isHorizMinMaxOk(uint8_t src[], int stride, int QP)
{
#ifdef MMX_FIXME
FIXME
	int isOk;
	asm volatile(

		: "=r" (isOk)
		: "r" (src), "r" (stride)
		);
	return isOk;
#else
	if(abs(src[0] - src[7]) > 2*QP) return false;

	return true;
#endif
}

static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP)
{

		: "%eax"
	);
#else
	uint8_t *src= tempBlock;

	for(int y=0; y<BLOCK_SIZE; y++)
	{
		dst[0] = src[0];
		dst[1] = src[1];
		dst[2] = src[2];
		dst[3] = src[3];

		: "%eax", "%ebx"
	);

#else
	uint8_t *temp= tempBlock;
	for(int y=0; y<BLOCK_SIZE; y++)
	{
		const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];
		const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];

		int sums[9];

	while( (rdtsc() - T)/1000 < 4000);

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

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

	if(1)
	{
		postProcess(src[1], src_stride, dst[1], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, true, mode >>4);
		postProcess(src[2], src_stride, dst[2], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, true, mode >>4);
	}
	else
	{
		memcpy(dst[1], src[1], src_stride*horizontal_size);
		memcpy(dst[2], src[2], src_stride*horizontal_size);

} // extern "C"

/**
 * Copies a block from src to dst and fixes the blacklevel
 */
static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride)
{
#ifdef HAVE_MMX
					asm volatile(
						"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"\
						"psubusb %%mm2, %%mm0	\n\t"\
						"psubusb %%mm2, %%mm1	\n\t"\
						"movq %%mm0, (%1)	\n\t"\
						"movq %%mm1, (%1, %3)	\n\t"\

#define SCALED_CPY					\
						"movq (%0), %%mm0	\n\t"\
						"movq (%0,%2), %%mm1	\n\t"\
						"psubusb %%mm2, %%mm0	\n\t"\

						"pmulhuw %%mm3, %%mm4	\n\t"\
						"pmulhuw %%mm3, %%mm5	\n\t"\
						"packuswb %%mm5, %%mm4	\n\t"\
						"movq %%mm4, (%1, %3)	\n\t"\


#define CPY SCALED_CPY
//#define CPY SIMPLE_CPY
//				"prefetchnta 8(%0)\n\t"
CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
CPY
						"popl %1 \n\t"
						"popl %0 \n\t"
						: : "r" (src),
						"r" (dst),
						"r" (srcStride),
						"r" (dstStride)
						: "%eax", "%ebx"
					);
#else
				for(int i=0; i<BLOCK_SIZE; i++) // last 10x8 Block is copied allready so +2
					memcpy(	&(dst[dstStride*i]),
						&(src[srcStride*i]), BLOCK_SIZE);
#endif
}


/**
 * Filters array of bytes (Y or U or V values)
 */
void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
	QP_STORE_T QPs[], int QPStride, bool isColor, int mode)
{

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

	/* we need 64bit here otherwise weŽll going to have a problem
	   after watching a black picture for 5 hours*/
	static uint64_t *yHistogram= NULL;
	if(!yHistogram)
	{
		yHistogram= new uint64_t[256];
		for(int i=0; i<256; i++) yHistogram[i]= width*height/64/256;
	}

	int black=0, white=255; // blackest black and whitest white in the picture
	if(!isColor)
	{
		uint64_t sum= 0;
		for(int i=0; i<256; i++)
			sum+= yHistogram[i];

		uint64_t maxClipped= (uint64_t)(sum * maxClippedThreshold);

		uint64_t clipped= sum;
		for(black=255; black>0; black--)
		{
			if(clipped < maxClipped) break;
			clipped-= yHistogram[black];
		}

		packedYOffset= MAX(black - minAllowedY, 0);
		packedYOffset|= packedYOffset<<32;
		packedYOffset|= packedYOffset<<16;
		packedYOffset|= packedYOffset<<8;

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

		packedYScale= uint16_t(scale*256.0 + 0.5);
		packedYScale|= packedYScale<<32;
		packedYScale|= packedYScale<<16;
	}
	else
	{
		packedYScale= 0x0100010001000100LL;
		packedYOffset= 0;
	}

	for(int x=0; x<width; x+=BLOCK_SIZE)
		blockCopy(dst + x, dstStride, src + x, srcStride);

	for(int 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]);
		uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start
		uint8_t *vertBlock= &(dstBlock[dstStride*3]);

		// finish 1 block before the next otherwise weŽll might have a problem
		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
		for(int x=0; x<width; x+=BLOCK_SIZE)
		{
			int QP= isColor ?
				QPs[(y>>3)*QPStride + (x>>3)]:
				(QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8;
#ifdef HAVE_MMX
		asm volatile(

			: : "r" (QP)
		);
#endif


			const int stride= dstStride;
			if(y + 12 < height)
			{
#ifdef MORE_TIMEING
				T0= rdtsc();
#endif

*/
#endif
				if(!isColor) yHistogram[ srcBlock[0] ]++;

				blockCopy(vertBlock + dstStride*2, dstStride,
					vertSrcBlock + srcStride*2, srcStride);


#ifdef MORE_TIMEING
				T1= rdtsc();
				memcpyTime+= T1-T0;

#endif
				if(mode & V_DEBLOCK)
				{
					if(mode & RK_FILTER)
						vertRKFilter(vertBlock, stride, QP);
					else if(0)
						vertX1Filter(vertBlock, stride, QP);
					else
					{
						if( isVertDC(vertBlock, stride))
						{

				vertTime+= T1-T0;
				T0=T1;
#endif
			}
			else
			{
				for(int i=2; i<BLOCK_SIZE/2+1; i++) // last 10x8 Block is copied allready so +2
					memcpy(	&(vertBlock[dstStride*i]),
						&(vertSrcBlock[srcStride*i]), BLOCK_SIZE);

			}

			if(x - 8 >= 0 && x<width)
			{
#ifdef MORE_TIMEING
				T0= rdtsc();

#ifdef TIMEING
	// 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
}

97:e57b1d38d71f

isHorizDC		Ec	Ec
isHorizMinMaxOk		a
doHorizLowPass		E		a	a*
doHorizDefFilter	E	ac	ac
deRing
RKAlgo1			E		a	a*
X1			a		E	E*


* i dont have a 3dnow CPU -> its untested
E = Exact implementation
e = allmost exact implementation
a = alternative / approximate impl

TODO:
verify that everything workes as it should
reduce the time wasted on the mem transfer
implement dering
implement everything in C at least (done at the moment but ...)
unroll stuff if instructions depend too much on the prior one
we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4?
move YScale thing to the end instead of fixing QP
write a faster and higher quality deblocking filter :)
do something about the speed of the horizontal filters
make the mainloop more flexible (variable number of blocks at once
	(the if/else stuff per block is slowing things down)
...

Notes:

*/

/*
Changelog:
0.1.3
	bugfixes: last 3 lines not brightness/contrast corrected
		brightness statistics messed up with initial black pic
	changed initial values of the brightness statistics
	C++ -> C conversation
	QP range question solved (very likely 1<=QP<=32 according to arpi)
	new experimental vertical deblocking filter
	RK filter has 3dNow support now (untested)
0.1.2
	fixed a bug in the horizontal default filter
	3dnow version of the Horizontal & Vertical Lowpass filters
	mmx version of the Horizontal Default filter
	mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar

*/


#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>
#include "../config.h"
//#undef HAVE_MMX2
//#define HAVE_3DNOW
//#undef HAVE_MMX
#include "postprocess.h"

//FIXME? |255-0| = 1 (shouldnt be a problem ...)
/**
 * Check if the middle 8x8 Block in the given 8x10 block is flat
 */
static inline int isVertDC(uint8_t src[], int stride){
//	return true;
	int numEq= 0;
	int y;
	src+= stride; // src points to begin of the 8x8 Block
#ifdef HAVE_MMX
	asm volatile(
//		"int $3 \n\t"
		"pushl %1\n\t"

//	int asmEq= numEq;
//	numEq=0;
//	uint8_t *temp= src;

#else
	for(y=0; y<BLOCK_SIZE-1; y++)
	{
		if(((src[0] - src[0+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[1] - src[1+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[2] - src[2+stride] + 1)&0xFFFF) < 3) numEq++;
		if(((src[3] - src[3+stride] + 1)&0xFFFF) < 3) numEq++;

			}
			printf("\n");
		}
	}
*/
//	for(int i=0; i<numEq/8; i++) src[i]=255;
	return (numEq > vFlatnessThreshold) ? 1 : 0;
}

static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP)
{
#ifdef HAVE_MMX
	int isOk;
	asm volatile(
//		"int $3 \n\t"

//		"movd %%mm0, (%1, %2, 4)\n\t"
		"movd %%mm0, %0					\n\t"
		: "=r" (isOk)
		: "r" (src), "r" (stride)
		);
	return isOk ? 1 : 0;
#else

	int isOk2= 1;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0;
	}
/*	if(isOk && !isOk2 || !isOk && isOk2)
	{
		printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);
		for(int y=0; y<9; y++)

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
	const int l9= stride + l8;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];
		const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];

		int sums[9];

	x/8 = 1
	1 12 12 23
 */
static inline void vertRKFilter(uint8_t *src, int stride, int QP)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
// FIXME rounding
	asm volatile(
		"pxor %%mm7, %%mm7				\n\t" // 0
		"movq b80, %%mm6				\n\t" // MIN_SIGNED_BYTE
		"leal (%0, %1), %%eax				\n\t"

		"movq (%0, %1, 4), %%mm2			\n\t" // line 4
		"movq (%%ebx), %%mm3				\n\t" // line 5
		"movq %%mm2, %%mm4				\n\t" // line 4
		"pcmpeqb %%mm5, %%mm5				\n\t" // -1
		"pxor %%mm2, %%mm5				\n\t" // -line 4 - 1
		PAVGB(%%mm3, %%mm5)
		"paddb %%mm6, %%mm5				\n\t" // (l5-l4)/2
		"psubusb %%mm3, %%mm4				\n\t"
		"psubusb %%mm2, %%mm3				\n\t"
		"por %%mm3, %%mm4				\n\t" // |l4 - l5|
		"psubusb %%mm0, %%mm4				\n\t"

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
	const int l9= stride + l8;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		if(ABS(src[l4]-src[l5]) < QP + QP/4)
		{
			int v = (src[l5] - src[l4]);

			src[l3] +=v/8;
			src[l4] +=v/2;
			src[l5] -=v/2;
			src[l6] -=v/8;

		}
		src++;
	}

#endif
}

/**
 * Experimental Filter 1
 * will nor damage linear gradients
 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
 * MMX2 version does correct clipping C version doesnt
 */
static inline void vertX1Filter(uint8_t *src, int stride, int QP)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"pxor %%mm7, %%mm7				\n\t" // 0
//		"movq b80, %%mm6				\n\t" // MIN_SIGNED_BYTE
		"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 (%%eax, %1, 2), %%mm0			\n\t" // line 3
		"movq (%0, %1, 4), %%mm1			\n\t" // line 4
		"movq %%mm1, %%mm2				\n\t" // line 4
		"psubusb %%mm0, %%mm1				\n\t"
		"psubusb %%mm2, %%mm0				\n\t"
		"por %%mm1, %%mm0				\n\t" // |l2 - l3|
		"movq (%%ebx), %%mm3				\n\t" // line 5
		"movq (%%ebx, %1), %%mm4				\n\t" // line 6
		"movq %%mm3, %%mm5				\n\t" // line 5
		"psubusb %%mm4, %%mm3				\n\t"
		"psubusb %%mm5, %%mm4				\n\t"
		"por %%mm4, %%mm3				\n\t" // |l5 - l6|
		PAVGB(%%mm3, %%mm0)				      // (|l2 - l3| + |l5 - l6|)/2
		"movq %%mm2, %%mm1				\n\t" // line 4
		"psubusb %%mm5, %%mm2				\n\t"
		"movq %%mm2, %%mm4				\n\t"
		"pcmpeqb %%mm7, %%mm2				\n\t" // (l4 - l5) <= 0 ? -1 : 0
		"psubusb %%mm1, %%mm5				\n\t"
		"por %%mm5, %%mm4				\n\t" // |l4 - l5|
		"psubusb %%mm0, %%mm4		\n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
		"movq %%mm4, %%mm3				\n\t" // d
		"psubusb pQPb, %%mm4				\n\t"
		"pcmpeqb %%mm7, %%mm4				\n\t" // d <= QP ? -1 : 0
		"pand %%mm4, %%mm3				\n\t" // d <= QP ? d : 0

		PAVGB(%%mm7, %%mm3)				      // d/2

		"movq (%0, %1, 4), %%mm0			\n\t" // line 4
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
		"psubusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%0, %1, 4)			\n\t" // line 4

		"movq (%%ebx), %%mm0				\n\t" // line 5
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
		"paddusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%ebx)				\n\t" // line 5

		PAVGB(%%mm7, %%mm3)				      // d/4

		"movq (%%eax, %1, 2), %%mm0			\n\t" // line 3
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
		"psubusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%eax, %1, 2)			\n\t" // line 3

		"movq (%%ebx, %1), %%mm0			\n\t" // line 6
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
		"paddusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%ebx, %1)			\n\t" // line 6

		PAVGB(%%mm7, %%mm3)				      // d/8

		"movq (%%eax, %1), %%mm0			\n\t" // line 2
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
		"psubusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%eax, %1)			\n\t" // line 2

		"movq (%%ebx, %1, 2), %%mm0			\n\t" // line 7
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
		"paddusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%ebx, %1, 2)			\n\t" // line 7

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

 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
	const int l9= stride + l8;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		int a= src[l3] - src[l4];
		int b= src[l4] - src[l5];
		int c= src[l6] - src[l7];

		int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);

		if(d < QP)
		{
			int v = d * SIGN(-b);

			src[l2] +=v/8;
			src[l3] +=v/4;
			src[l4] +=v/2;
			src[l5] -=v/2;
			src[l6] -=v/4;
			src[l7] -=v/8;

		}
		src++;
	}
	/*
 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;

			src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16;
			src[l6] = (       1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16;
		}
		src++;
	}
*/
#endif
}


static inline void doVertDefFilter(uint8_t src[], int stride, int QP)

	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
	const int l8= stride + l7;
//	const int l9= stride + l8;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
		if(ABS(middleEnergy) < 8*QP)
		{
			const int q=(src[l4] - src[l5])/2;

//FIXME?  |255-0| = 1
/**
 * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock.
 */
static inline int isHorizDCAndCopy2Temp(uint8_t src[], int stride)
{
//	src++;
	int numEq= 0;
#ifdef HAVE_MMX
asm volatile (

		: "%eax"
		);
//	printf("%d\n", numEq);
	numEq= (256 - (numEq & 0xFF)) &0xFF;
#else
	int y;
	for(y=0; y<BLOCK_SIZE; y++)
	{
		if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;

*/
//	printf("%d\n", numEq);
	return numEq > hFlatnessThreshold;
}

static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)
{
#ifdef MMX_FIXME
FIXME
	int isOk;
	asm volatile(

		: "=r" (isOk)
		: "r" (src), "r" (stride)
		);
	return isOk;
#else
	if(abs(src[0] - src[7]) > 2*QP) return 0;

	return 1;
#endif
}

static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP)
{

		: "%eax"
	);
#else
	uint8_t *src= tempBlock;

	int y;
	for(y=0; y<BLOCK_SIZE; y++)
	{
		dst[0] = src[0];
		dst[1] = src[1];
		dst[2] = src[2];
		dst[3] = src[3];

		: "%eax", "%ebx"
	);

#else
	uint8_t *temp= tempBlock;
	int y;
	for(y=0; y<BLOCK_SIZE; y++)
	{
		const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];
		const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];

		int sums[9];

	while( (rdtsc() - T)/1000 < 4000);

	return;
*/
	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(1)
	{
		postProcess(src[1], src_stride, dst[1], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 1, mode >>4);
		postProcess(src[2], src_stride, dst[2], dst_stride,
			horizontal_size, vertical_size, QP_store, QP_stride, 1, mode >>4);
	}
	else
	{
		memcpy(dst[1], src[1], src_stride*horizontal_size);
		memcpy(dst[2], src[2], src_stride*horizontal_size);

} // extern "C"

/**
 * 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(
						"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 SCALED_CPY					\
						"movq (%0), %%mm0	\n\t"\
						"movq (%0,%2), %%mm1	\n\t"\
						"psubusb %%mm2, %%mm0	\n\t"\

						"pmulhuw %%mm3, %%mm4	\n\t"\
						"pmulhuw %%mm3, %%mm5	\n\t"\
						"packuswb %%mm5, %%mm4	\n\t"\
						"movq %%mm4, (%1, %3)	\n\t"\

						"1:			\n\t"
SCALED_CPY
						"addl %%eax, %0		\n\t"
						"addl %%ebx, %1		\n\t"
SCALED_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
	}
	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)
 */
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 weŽll 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

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

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

	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];
		}

		packedYOffset= MAX(black - minAllowedY, 0);
		packedYOffset|= packedYOffset<<32;
		packedYOffset|= packedYOffset<<16;
		packedYOffset|= packedYOffset<<8;

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

		packedYScale= (uint16_t)(scale*256.0 + 0.5);
		packedYScale|= packedYScale<<32;
		packedYScale|= packedYScale<<16;
	}
	else
	{
		packedYScale= 0x0100010001000100LL;
		packedYOffset= 0;
	}

	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]);
		uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start
		uint8_t *vertBlock= &(dstBlock[dstStride*3]);

		// finish 1 block before the next otherwise weŽll 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;
			int QP= isColor ?
				QPs[(y>>3)*QPStride + (x>>3)]:
				(QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8;
#ifdef HAVE_MMX
		asm volatile(

			: : "r" (QP)
		);
#endif


			if(y + 12 < height)
			{
#ifdef MORE_TIMEING
				T0= rdtsc();
#endif

*/
#endif
				if(!isColor) yHistogram[ srcBlock[0] ]++;

				blockCopy(vertBlock + dstStride*2, dstStride,
					vertSrcBlock + srcStride*2, srcStride, 8, mode & LEVEL_FIX);


#ifdef MORE_TIMEING
				T1= rdtsc();
				memcpyTime+= T1-T0;

#endif
				if(mode & V_DEBLOCK)
				{
					if(mode & RK_FILTER)
						vertRKFilter(vertBlock, stride, QP);
					else if(mode & X1_FILTER)
						vertX1Filter(vertBlock, stride, QP);
					else
					{
						if( isVertDC(vertBlock, stride))
						{

				vertTime+= T1-T0;
				T0=T1;
#endif
			}
			else
				blockCopy(vertBlock + dstStride*1, dstStride,
					vertSrcBlock + srcStride*1, srcStride, 4, mode & LEVEL_FIX);


			if(x - 8 >= 0 && x<width)
			{
#ifdef MORE_TIMEING
				T0= rdtsc();

#ifdef TIMEING
	// 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
}