postprocess.c

		"pminsw %%mm0, %%mm4				\n\t"
		"pminsw %%mm1, %%mm5				\n\t"
#else
		"movq %%mm4, %%mm2				\n\t"
		"psubusw %%mm0, %%mm2				\n\t"
		"psubw %%mm2, %%mm4				\n\t"
		"movq %%mm5, %%mm2				\n\t"
		"psubusw %%mm1, %%mm2				\n\t"
		"psubw %%mm2, %%mm5				\n\t"
#endif
		"pxor %%mm6, %%mm4				\n\t"
		"pxor %%mm7, %%mm5				\n\t"
		"psubw %%mm6, %%mm4				\n\t"
		"psubw %%mm7, %%mm5				\n\t"
		"packsswb %%mm5, %%mm4				\n\t"
		"movq (%%eax, %1, 2), %%mm0			\n\t"
		"paddb   %%mm4, %%mm0				\n\t"
		"movq %%mm0, (%%eax, %1, 2) 			\n\t"
		"movq (%0, %1, 4), %%mm0			\n\t"
		"psubb %%mm4, %%mm0				\n\t"
//		"pxor %%mm0, %%mm0 \n\t"
		"movq %%mm0, (%0, %1, 4) 			\n\t"

		:
		: "r" (src), "r" (stride), "r" (QP)
		: "%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++)
	{
		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;
			const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
			const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);

			int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
			d= MAX(d, 0);

			d= (5*d + 32) >> 6;
			d*= SIGN(-middleEnergy);

			if(q>0)
			{
				d= d<0 ? 0 : d;
				d= d>q ? q : d;
			}
			else
			{
				d= d>0 ? 0 : d;
				d= d<q ? q : d;
			}

        		src[l4]-= d;
	        	src[l5]+= d;
		}
		src++;
	}
#endif
}

//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 (
//		"int $3 \n\t"
		"pushl %1\n\t"
		"movq b7E, %%mm7				\n\t" // mm7 = 0x7F
		"movq b7C, %%mm6				\n\t" // mm6 = 0x7D
		"leal tempBlock, %%eax				\n\t"
		"pxor %%mm0, %%mm0				\n\t"

#define HDC_CHECK_AND_CPY(i) \
		"movq -4(%1), %%mm2				\n\t"\
		"psrlq $32, %%mm2				\n\t"\
		"punpckldq 4(%1), %%mm2				\n\t" /* (%1) */\
		"movq %%mm2, %%mm1				\n\t"\
		"psrlq $8, %%mm2				\n\t"\
		"psubb %%mm1, %%mm2				\n\t"\
		"paddb %%mm7, %%mm2				\n\t"\
		"pcmpgtb %%mm6, %%mm2				\n\t"\
		"paddb %%mm2, %%mm0				\n\t"\
		"movq %%mm1," #i "(%%eax)			\n\t"

		HDC_CHECK_AND_CPY(0)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(8)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(16)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(24)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(32)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(40)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(48)
		"addl %2, %1					\n\t"
		HDC_CHECK_AND_CPY(56)

		"psllq $8, %%mm0				\n\t" // remove dummy value
		"movq %%mm0, %%mm1				\n\t"
		"psrlw $8, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"psrlq $16, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"psrlq $32, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"popl %1\n\t"
		"movd %%mm0, %0					\n\t"
		: "=r" (numEq)
		: "r" (src), "r" (stride)
		: "%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++;
		if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++;
		if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++;
		tempBlock[0 + y*TEMP_STRIDE] = src[0];
		tempBlock[1 + y*TEMP_STRIDE] = src[1];
		tempBlock[2 + y*TEMP_STRIDE] = src[2];
		tempBlock[3 + y*TEMP_STRIDE] = src[3];
		tempBlock[4 + y*TEMP_STRIDE] = src[4];
		tempBlock[5 + y*TEMP_STRIDE] = src[5];
		tempBlock[6 + y*TEMP_STRIDE] = src[6];
		tempBlock[7 + y*TEMP_STRIDE] = src[7];
		src+= stride;
	}
#endif
/*	if(abs(numEq - asmEq) > 0)
	{
//		printf("\nasm:%d  c:%d\n", asmEq, numEq);
		for(int y=0; y<8; y++)
		{
			for(int x=0; x<8; x++)
			{
				printf("%d ", src[x + y*stride]);
			}
			printf("\n");
		}
	}
*/
//	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(
//		"int $3 \n\t"
		"movq (%1, %2), %%mm0				\n\t"
		"movq (%1, %2, 8), %%mm1			\n\t"
		"movq %%mm0, %%mm2				\n\t"
		"psubusb %%mm1, %%mm0				\n\t"
		"psubusb %%mm2, %%mm1				\n\t"
		"por %%mm1, %%mm0				\n\t" // ABS Diff

		"movq pQPb, %%mm7				\n\t" // QP,..., QP
		"paddusb %%mm7, %%mm7				\n\t" // 2QP ... 2QP
		"psubusb %%mm7, %%mm0				\n\t" // Diff <= 2QP -> 0
		"pcmpeqd b00, %%mm0				\n\t"
		"psrlq $16, %%mm0				\n\t"
		"pcmpeqd bFF, %%mm0				\n\t"
//		"movd %%mm0, (%1, %2, 4)\n\t"
		"movd %%mm0, %0					\n\t"
		: "=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)
{
#ifdef HAVE_MMX
	asm volatile(
		"pushl %0					\n\t"
		"pxor %%mm7, %%mm7				\n\t"
		"movq bm00001000, %%mm6				\n\t"
		"movd %2, %%mm5					\n\t" // QP
		"movq %%mm5, %%mm4				\n\t"
		"paddusb %%mm5, %%mm5				\n\t" // 2QP
		"paddusb %%mm5, %%mm4				\n\t" // 3QP
		"psllq $24, %%mm4				\n\t"
		"pxor %%mm5, %%mm5				\n\t" // 0
		"psubb %%mm4, %%mm5				\n\t" // -QP
		"leal tempBlock, %%eax				\n\t"

//FIXME? "unroll by 2" and mix
#ifdef HAVE_MMX2
#define HDF(i)	\
		"movq " #i "(%%eax), %%mm0			\n\t"\
		"movq %%mm0, %%mm1				\n\t"\
		"movq %%mm0, %%mm2				\n\t"\
		"psrlq $8, %%mm1				\n\t"\
		"psubusb %%mm1, %%mm2				\n\t"\
		"psubusb %%mm0, %%mm1				\n\t"\
		"por %%mm2, %%mm1				\n\t" /* px = |px - p(x+1)| */\
		"pcmpeqb %%mm7, %%mm2				\n\t" /* px = sgn[px - p(x+1)] */\
		"pshufw $0x00, %%mm1, %%mm3			\n\t" /* p5 = |p1 - p2| */\
		"pminub %%mm1, %%mm3				\n\t" /* p5 = min(|p2-p1|, |p6-p5|)*/\
		"psrlq $16, %%mm3				\n\t" /* p3 = min(|p2-p1|, |p6-p5|)*/\
		"psubusb %%mm3, %%mm1			\n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\
		"paddb %%mm5, %%mm1				\n\t"\
		"psubusb %%mm5, %%mm1				\n\t"\
		"psrlw $2, %%mm1				\n\t"\
		"pxor %%mm2, %%mm1				\n\t"\
		"psubb %%mm2, %%mm1				\n\t"\
		"pand %%mm6, %%mm1				\n\t"\
		"psubb %%mm1, %%mm0				\n\t"\
		"psllq $8, %%mm1				\n\t"\
		"paddb %%mm1, %%mm0				\n\t"\
		"movd %%mm0, (%0)				\n\t"\
		"psrlq $32, %%mm0				\n\t"\
		"movd %%mm0, 4(%0)				\n\t"
#else
#define HDF(i)\
		"movq " #i "(%%eax), %%mm0			\n\t"\
		"movq %%mm0, %%mm1				\n\t"\
		"movq %%mm0, %%mm2				\n\t"\
		"psrlq $8, %%mm1				\n\t"\
		"psubusb %%mm1, %%mm2				\n\t"\
		"psubusb %%mm0, %%mm1				\n\t"\
		"por %%mm2, %%mm1				\n\t" /* px = |px - p(x+1)| */\
		"pcmpeqb %%mm7, %%mm2				\n\t" /* px = sgn[px - p(x+1)] */\
		"movq %%mm1, %%mm3				\n\t"\
		"psllq $32, %%mm3				\n\t"\
		"movq %%mm3, %%mm4				\n\t"\
		"psubusb %%mm1, %%mm4				\n\t"\
		"psubb %%mm4, %%mm3				\n\t"\
		"psrlq $16, %%mm3				\n\t" /* p3 = min(|p2-p1|, |p6-p5|)*/\
		"psubusb %%mm3, %%mm1			\n\t" /* |p3-p4|-min(|p1-p2|,|p5,6|) */\
		"paddb %%mm5, %%mm1				\n\t"\
		"psubusb %%mm5, %%mm1				\n\t"\
		"psrlw $2, %%mm1				\n\t"\
		"pxor %%mm2, %%mm1				\n\t"\
		"psubb %%mm2, %%mm1				\n\t"\
		"pand %%mm6, %%mm1				\n\t"\
		"psubb %%mm1, %%mm0				\n\t"\
		"psllq $8, %%mm1				\n\t"\
		"paddb %%mm1, %%mm0				\n\t"\
		"movd %%mm0, (%0)				\n\t"\
		"psrlq $32, %%mm0				\n\t"\
		"movd %%mm0, 4(%0)				\n\t"
#endif
		HDF(0)
		"addl %1, %0					\n\t"
		HDF(8)
		"addl %1, %0					\n\t"
		HDF(16)
		"addl %1, %0					\n\t"
		HDF(24)
		"addl %1, %0					\n\t"
		HDF(32)
		"addl %1, %0					\n\t"
		HDF(40)
		"addl %1, %0					\n\t"
		HDF(48)
		"addl %1, %0					\n\t"
		HDF(56)
		"popl %0					\n\t"
		:
		: "r" (dst), "r" (stride), "r" (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];
		dst[4] = src[4];
		dst[5] = src[5];
		dst[6] = src[6];
		dst[7] = src[7];

		const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]);
		if(ABS(middleEnergy) < 8*QP)
		{
			const int q=(src[3] - src[4])/2;
			const int leftEnergy=  5*(src[2] - src[1]) + 2*(src[0] - src[3]);
			const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]);

			int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
			d= MAX(d, 0);

			d= (5*d + 32) >> 6;
			d*= SIGN(-middleEnergy);

			if(q>0)
			{
				d= d<0 ? 0 : d;
				d= d>q ? q : d;
			}
			else
			{
				d= d>0 ? 0 : d;
				d= d<q ? q : d;
			}

        		dst[3]-= d;
	        	dst[4]+= d;
		}
		dst+= stride;
		src+= TEMP_STRIDE;
	}
#endif
}

/**
 * Do a horizontal low pass filter on the 8x8 block
 * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
 * useing approximately the 7-Tap Filter (1,2,3,4,3,2,1)/16 (MMX2/3DNOW version)
 */
static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP)
{
//return;
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(	//"movv %0 %1 %2\n\t"
		"pushl %0\n\t"
		"pxor %%mm7, %%mm7					\n\t"
		"leal tempBlock, %%eax					\n\t"
/*
#define HLP1	"movq (%0), %%mm0					\n\t"\
		"movq %%mm0, %%mm1					\n\t"\
		"psllq $8, %%mm0					\n\t"\
		PAVGB(%%mm1, %%mm0)\
		"psrlw $8, %%mm0					\n\t"\
		"pxor %%mm1, %%mm1					\n\t"\
		"packuswb %%mm1, %%mm0					\n\t"\
		"movq %%mm0, %%mm1					\n\t"\
		"movq %%mm0, %%mm2					\n\t"\
		"psllq $32, %%mm0					\n\t"\
		"paddb %%mm0, %%mm1					\n\t"\
		"psllq $16, %%mm2					\n\t"\
		PAVGB(%%mm2, %%mm0)\
		"movq %%mm0, %%mm3					\n\t"\
		"pand bm11001100, %%mm0					\n\t"\
		"paddusb %%mm0, %%mm3					\n\t"\
		"psrlq $8, %%mm3					\n\t"\
		PAVGB(%%mm1, %%mm4)\
		PAVGB(%%mm3, %%mm2)\
		"psrlq $16, %%mm2					\n\t"\
		"punpcklbw %%mm2, %%mm2					\n\t"\
		"movq %%mm2, (%0)					\n\t"\

#define HLP2	"movq (%0), %%mm0					\n\t"\
		"movq %%mm0, %%mm1					\n\t"\
		"psllq $8, %%mm0					\n\t"\
		PAVGB(%%mm1, %%mm0)\
		"psrlw $8, %%mm0					\n\t"\
		"pxor %%mm1, %%mm1					\n\t"\
		"packuswb %%mm1, %%mm0					\n\t"\
		"movq %%mm0, %%mm2					\n\t"\
		"psllq $32, %%mm0					\n\t"\
		"psllq $16, %%mm2					\n\t"\
		PAVGB(%%mm2, %%mm0)\
		"movq %%mm0, %%mm3					\n\t"\
		"pand bm11001100, %%mm0					\n\t"\
		"paddusb %%mm0, %%mm3					\n\t"\
		"psrlq $8, %%mm3					\n\t"\
		PAVGB(%%mm3, %%mm2)\
		"psrlq $16, %%mm2					\n\t"\
		"punpcklbw %%mm2, %%mm2					\n\t"\
		"movq %%mm2, (%0)					\n\t"\
*/
// approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16
/*
 31
 121
  121
   121
    121
     121
      121
       13
Implemented	Exact 7-Tap
 9421		A321
 36421		64321
 334321		=
 1234321	=
  1234321	=
   123433	=
    12463	  12346
     1249	   123A

*/
#ifdef HAVE_MMX2
#define HLP3(i)	"movq " #i "(%%eax), %%mm0				\n\t"\
		"movq %%mm0, %%mm1					\n\t"\
		"movq %%mm0, %%mm2					\n\t"\
		"movq %%mm0, %%mm3					\n\t"\
		"movq %%mm0, %%mm4					\n\t"\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"pand bm00000001, %%mm3					\n\t"\
		"pand bm10000000, %%mm4					\n\t"\
		"por %%mm3, %%mm1					\n\t"\
		"por %%mm4, %%mm2					\n\t"\
		PAVGB(%%mm2, %%mm1)\
		PAVGB(%%mm1, %%mm0)\
\
		"pshufw $0xF9, %%mm0, %%mm3				\n\t"\
		"pshufw $0x90, %%mm0, %%mm4				\n\t"\
		PAVGB(%%mm3, %%mm4)\
		PAVGB(%%mm4, %%mm0)\
		"movd %%mm0, (%0)					\n\t"\
		"psrlq $32, %%mm0					\n\t"\
		"movd %%mm0, 4(%0)					\n\t"
#else
#define HLP3(i)	"movq " #i "(%%eax), %%mm0				\n\t"\
		"movq %%mm0, %%mm1					\n\t"\
		"movq %%mm0, %%mm2					\n\t"\
		"movq %%mm0, %%mm3					\n\t"\
		"movq %%mm0, %%mm4					\n\t"\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"pand bm00000001, %%mm3					\n\t"\
		"pand bm10000000, %%mm4					\n\t"\
		"por %%mm3, %%mm1					\n\t"\
		"por %%mm4, %%mm2					\n\t"\
		PAVGB(%%mm2, %%mm1)\
		PAVGB(%%mm1, %%mm0)\
\
		"movq %%mm0, %%mm3					\n\t"\
		"movq %%mm0, %%mm4					\n\t"\
		"movq %%mm0, %%mm5					\n\t"\
		"psrlq $16, %%mm3					\n\t"\
		"psllq $16, %%mm4					\n\t"\
		"pand bm11000000, %%mm5					\n\t"\
		"por %%mm5, %%mm3					\n\t"\
		"movq %%mm0, %%mm5					\n\t"\
		"pand bm00000011, %%mm5					\n\t"\
		"por %%mm5, %%mm4					\n\t"\
		PAVGB(%%mm3, %%mm4)\
		PAVGB(%%mm4, %%mm0)\
		"movd %%mm0, (%0)					\n\t"\
		"psrlq $32, %%mm0					\n\t"\
		"movd %%mm0, 4(%0)					\n\t"
#endif

#define HLP(i) HLP3(i)

		HLP(0)
		"addl %1, %0						\n\t"
		HLP(8)
		"addl %1, %0						\n\t"
		HLP(16)
		"addl %1, %0						\n\t"
		HLP(24)
		"addl %1, %0						\n\t"
		HLP(32)
		"addl %1, %0						\n\t"
		HLP(40)
		"addl %1, %0						\n\t"
		HLP(48)
		"addl %1, %0						\n\t"
		HLP(56)

		"popl %0\n\t"
		:
		: "r" (dst), "r" (stride)
		: "%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];
		sums[0] = first + temp[0];
		sums[1] = temp[0] + temp[1];
		sums[2] = temp[1] + temp[2];
		sums[3] = temp[2] + temp[3];
		sums[4] = temp[3] + temp[4];
		sums[5] = temp[4] + temp[5];
		sums[6] = temp[5] + temp[6];
		sums[7] = temp[6] + temp[7];
		sums[8] = temp[7] + last;

		dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
		dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4;
		dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4;
		dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4;
		dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4;
		dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4;
		dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4;
		dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4;

		dst+= stride;
		temp+= TEMP_STRIDE;
	}
#endif
}


static inline void dering(uint8_t src[], int stride, int QP)
{
//FIXME

#ifdef HAVE_MMX2X
	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

		"pcmpeq %%mm6, %%mm6				\n\t"
		"pxor %%mm7, %%mm7				\n\t"

#define FIND_MIN_MAX(addr)\
		"movq (" #addr "), %%mm0,			\n\t"\
		"pminub %%mm0, %%mm6				\n\t"\
		"pmaxub %%mm0, %%mm7				\n\t"

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

		"movq %%mm6, %%mm4				\n\t"
		"psrlq $32, %%mm6				\n\t"
		"pminub %%mm4, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $16, %%mm6				\n\t"
		"pminub %%mm4, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $8, %%mm6				\n\t"
		"pminub %%mm4, %%mm6				\n\t" // min of pixels

		"movq %%mm7, %%mm4				\n\t"
		"psrlq $32, %%mm7				\n\t"
		"pmaxub %%mm4, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $16, %%mm7				\n\t"
		"pmaxub %%mm4, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $8, %%mm7				\n\t"
		"pmaxub %%mm4, %%mm7				\n\t" // max of pixels
		PAVGB(%%mm6, %%mm7)				      // (max + min)/2


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

//FIXME
#endif
}




/**
 * ...
 * the mode value is interpreted as a quality value if its negative, its range is then (-1 ... -63)
 * -63 is best quality -1 is worst
 */
//extern "C"{
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)
{

	if(mode<0) mode= getModeForQuality(-mode);

/*
	long long T= rdtsc();
	for(int y=vertical_size-1; y>=0 ; y--)
		memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride);
//	memcpy(dst[0], src[0],src_stride*vertical_size);
	printf("%4dk\r", (rdtsc()-T)/1000);

	return;
*/
/*
	long long T= rdtsc();
	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);
	}
}
/**
 * gets the mode flags for a given quality (larger values mean slower but better postprocessing)
 * 0 <= quality < 64
 */
int getModeForQuality(int quality){
	int modes[6]= {
		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
		};

	return modes[ (quality*6) >>6 ];
}

//} // 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"\
						"psubusb %%mm2, %%mm1	\n\t"\
						"pxor %%mm4, %%mm4	\n\t"\
						"pxor %%mm5, %%mm5	\n\t"\
						"punpcklbw %%mm0, %%mm4 \n\t"\
						"punpckhbw %%mm0, %%mm5 \n\t"\
						"pmulhuw %%mm3, %%mm4	\n\t"\
						"pmulhuw %%mm3, %%mm5	\n\t"\
						"packuswb %%mm5, %%mm4	\n\t"\
						"movq %%mm4, (%1)	\n\t"\
						"pxor %%mm4, %%mm4	\n\t"\
						"pxor %%mm5, %%mm5	\n\t"\
						"punpcklbw %%mm1, %%mm4 \n\t"\
						"punpckhbw %%mm1, %%mm5 \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 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

#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];
		}

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

		// we cant handle negative correctures
		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 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;
			int QP= isColor ?
				QPs[(y>>3)*QPStride + (x>>3)]:
				(QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8;
#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


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

#ifdef HAVE_MMX2
				prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32);
				prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32);
				prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32);
				prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32);
#elif defined(HAVE_3DNOW)
//FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
/*				prefetch(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32);
				prefetch(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32);
				prefetchw(vertBlock + (((x>>3)&3) + 2)*dstStride + 32);
				prefetchw(vertBlock + (((x>>3)&3) + 6)*dstStride + 32);
*/
#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;
				T0=T1;
#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))
						{
							if(isVertMinMaxOk(vertBlock, stride, QP))
								doVertLowPass(vertBlock, stride, QP);
						}
						else
							doVertDefFilter(vertBlock, stride, QP);
					}
				}
#ifdef MORE_TIMEING
				T1= rdtsc();
				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();
#endif
				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_TIMEING
				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)


			dstBlock+=8;
			srcBlock+=8;
			vertBlock+=8;
			vertSrcBlock+=8;
		}
	}
#ifdef HAVE_3DNOW
	asm volatile("femms");
#elif defined (HAVE_MMX)
	asm volatile("emms");
#endif

#ifdef TIMEING
	// FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)
	sumTime= rdtsc() - sumTime;
	if(!isColor)