postprocess_template.c

		"pcmpgtw %%mm0, %%mm6				\n\t"
		"pxor %%mm6, %%mm0				\n\t"
		"psubw %%mm6, %%mm0				\n\t" // |2L4 - 5L5 + 5L6 - 2L7|
		"movq %%mm7, %%mm6				\n\t" // 0
		"pcmpgtw %%mm1, %%mm6				\n\t"
		"pxor %%mm6, %%mm1				\n\t"
		"psubw %%mm6, %%mm1				\n\t" // |2H4 - 5H5 + 5H6 - 2H7|
		"movq %%mm7, %%mm6				\n\t" // 0
		"pcmpgtw %%mm2, %%mm6				\n\t"
		"pxor %%mm6, %%mm2				\n\t"
		"psubw %%mm6, %%mm2				\n\t" // |2L0 - 5L1 + 5L2 - 2L3|
		"movq %%mm7, %%mm6				\n\t" // 0
		"pcmpgtw %%mm3, %%mm6				\n\t"
		"pxor %%mm6, %%mm3				\n\t"
		"psubw %%mm6, %%mm3				\n\t" // |2H0 - 5H1 + 5H2 - 2H3|
#endif

#ifdef HAVE_MMX2
		"pminsw %%mm2, %%mm0				\n\t"
		"pminsw %%mm3, %%mm1				\n\t"
#else
		"movq %%mm0, %%mm6				\n\t"
		"psubusw %%mm2, %%mm6				\n\t"
		"psubw %%mm6, %%mm0				\n\t"
		"movq %%mm1, %%mm6				\n\t"
		"psubusw %%mm3, %%mm6				\n\t"
		"psubw %%mm6, %%mm1				\n\t"
#endif

		"movq %%mm7, %%mm6				\n\t" // 0
		"pcmpgtw %%mm4, %%mm6				\n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
		"pxor %%mm6, %%mm4				\n\t"
		"psubw %%mm6, %%mm4				\n\t" // |2L2 - 5L3 + 5L4 - 2L5|
		"pcmpgtw %%mm5, %%mm7				\n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
		"pxor %%mm7, %%mm5				\n\t"
		"psubw %%mm7, %%mm5				\n\t" // |2H2 - 5H3 + 5H4 - 2H5|
// 100 opcodes
		"movd %2, %%mm2					\n\t" // QP
		"psllw $3, %%mm2				\n\t" // 8QP
		"movq %%mm2, %%mm3				\n\t" // 8QP
		"pcmpgtw %%mm4, %%mm2				\n\t"
		"pcmpgtw %%mm5, %%mm3				\n\t"
		"pand %%mm2, %%mm4				\n\t"
		"pand %%mm3, %%mm5				\n\t"


		"psubusw %%mm0, %%mm4				\n\t" // hd
		"psubusw %%mm1, %%mm5				\n\t" // ld


		"movq "MANGLE(w05)", %%mm2			\n\t" // 5
		"pmullw %%mm2, %%mm4				\n\t"
		"pmullw %%mm2, %%mm5				\n\t"
		"movq "MANGLE(w20)", %%mm2			\n\t" // 32
		"paddw %%mm2, %%mm4				\n\t"
		"paddw %%mm2, %%mm5				\n\t"
		"psrlw $6, %%mm4				\n\t"
		"psrlw $6, %%mm5				\n\t"

		"movq 16(%%ecx), %%mm0				\n\t" // L3 - L4
		"movq 24(%%ecx), %%mm1				\n\t" // H3 - H4

		"pxor %%mm2, %%mm2				\n\t"
		"pxor %%mm3, %%mm3				\n\t"

		"pcmpgtw %%mm0, %%mm2				\n\t" // sign (L3-L4)
		"pcmpgtw %%mm1, %%mm3				\n\t" // sign (H3-H4)
		"pxor %%mm2, %%mm0				\n\t"
		"pxor %%mm3, %%mm1				\n\t"
		"psubw %%mm2, %%mm0				\n\t" // |L3-L4|
		"psubw %%mm3, %%mm1				\n\t" // |H3-H4|
		"psrlw $1, %%mm0				\n\t" // |L3 - L4|/2
		"psrlw $1, %%mm1				\n\t" // |H3 - H4|/2

		"pxor %%mm6, %%mm2				\n\t"
		"pxor %%mm7, %%mm3				\n\t"
		"pand %%mm2, %%mm4				\n\t"
		"pand %%mm3, %%mm5				\n\t"

#ifdef HAVE_MMX2
		"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 (%0), %%mm0				\n\t"
		"paddb   %%mm4, %%mm0				\n\t"
		"movq %%mm0, (%0)				\n\t"
		"movq (%0, %1), %%mm0				\n\t"
		"psubb %%mm4, %%mm0				\n\t"
		"movq %%mm0, (%0, %1)				\n\t"

		: "+r" (src)
		: "r" (stride), "m" (c->pQPb)
		: "%eax", "%ecx"
	);
#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;
	src+= stride*3;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
		if(ABS(middleEnergy) < 8*c->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
}

static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"pxor %%mm6, %%mm6				\n\t"
		"pcmpeqb %%mm7, %%mm7				\n\t"
		"movq %2, %%mm0					\n\t"
		"punpcklbw %%mm6, %%mm0				\n\t"
		"psrlw $1, %%mm0				\n\t"
		"psubw %%mm7, %%mm0				\n\t"
		"packuswb %%mm0, %%mm0				\n\t"
		"movq %%mm0, %3					\n\t"

		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
		
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

#undef FIND_MIN_MAX
#ifdef HAVE_MMX2
#define FIND_MIN_MAX(addr)\
		"movq " #addr ", %%mm0				\n\t"\
		"pminub %%mm0, %%mm7				\n\t"\
		"pmaxub %%mm0, %%mm6				\n\t"
#else
#define FIND_MIN_MAX(addr)\
		"movq " #addr ", %%mm0				\n\t"\
		"movq %%mm7, %%mm1				\n\t"\
		"psubusb %%mm0, %%mm6				\n\t"\
		"paddb %%mm0, %%mm6				\n\t"\
		"psubusb %%mm0, %%mm1				\n\t"\
		"psubb %%mm1, %%mm7				\n\t"
#endif

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

		"movq %%mm7, %%mm4				\n\t"
		"psrlq $8, %%mm7				\n\t"
#ifdef HAVE_MMX2
		"pminub %%mm4, %%mm7				\n\t" // min of pixels
		"pshufw $0xF9, %%mm7, %%mm4			\n\t"
		"pminub %%mm4, %%mm7				\n\t" // min of pixels
		"pshufw $0xFE, %%mm7, %%mm4			\n\t"
		"pminub %%mm4, %%mm7				\n\t"
#else
		"movq %%mm7, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $16, %%mm7				\n\t"
		"movq %%mm7, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $32, %%mm7				\n\t"
		"movq %%mm7, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm7				\n\t"
#endif


		"movq %%mm6, %%mm4				\n\t"
		"psrlq $8, %%mm6				\n\t"
#ifdef HAVE_MMX2
		"pmaxub %%mm4, %%mm6				\n\t" // max of pixels
		"pshufw $0xF9, %%mm6, %%mm4			\n\t"
		"pmaxub %%mm4, %%mm6				\n\t"
		"pshufw $0xFE, %%mm6, %%mm4			\n\t"
		"pmaxub %%mm4, %%mm6				\n\t"
#else
		"psubusb %%mm4, %%mm6				\n\t"
		"paddb %%mm4, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $16, %%mm6				\n\t"
		"psubusb %%mm4, %%mm6				\n\t"
		"paddb %%mm4, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $32, %%mm6				\n\t"
		"psubusb %%mm4, %%mm6				\n\t"
		"paddb %%mm4, %%mm6				\n\t"
#endif
		"movq %%mm6, %%mm0				\n\t" // max
		"psubb %%mm7, %%mm6				\n\t" // max - min
		"movd %%mm6, %%ecx				\n\t"
		"cmpb "MANGLE(deringThreshold)", %%cl		\n\t"
		" jb 1f						\n\t"
		"leal -24(%%esp), %%ecx				\n\t"
		"andl $0xFFFFFFF8, %%ecx			\n\t" 
		PAVGB(%%mm0, %%mm7)				      // a=(max + min)/2
		"punpcklbw %%mm7, %%mm7				\n\t"
		"punpcklbw %%mm7, %%mm7				\n\t"
		"punpcklbw %%mm7, %%mm7				\n\t"
		"movq %%mm7, (%%ecx)				\n\t"

		"movq (%0), %%mm0				\n\t" // L10
		"movq %%mm0, %%mm1				\n\t" // L10
		"movq %%mm0, %%mm2				\n\t" // L10
		"psllq $8, %%mm1				\n\t"
		"psrlq $8, %%mm2				\n\t"
		"movd -4(%0), %%mm3				\n\t"
		"movd 8(%0), %%mm4				\n\t"
		"psrlq $24, %%mm3				\n\t"
		"psllq $56, %%mm4				\n\t"
		"por %%mm3, %%mm1				\n\t" // L00
		"por %%mm4, %%mm2				\n\t" // L20
		"movq %%mm1, %%mm3				\n\t" // L00
		PAVGB(%%mm2, %%mm1)				      // (L20 + L00)/2
		PAVGB(%%mm0, %%mm1)				      // (L20 + L00 + 2L10)/4
		"psubusb %%mm7, %%mm0				\n\t"
		"psubusb %%mm7, %%mm2				\n\t"
		"psubusb %%mm7, %%mm3				\n\t"
		"pcmpeqb "MANGLE(b00)", %%mm0			\n\t" // L10 > a ? 0 : -1
		"pcmpeqb "MANGLE(b00)", %%mm2			\n\t" // L20 > a ? 0 : -1
		"pcmpeqb "MANGLE(b00)", %%mm3			\n\t" // L00 > a ? 0 : -1
		"paddb %%mm2, %%mm0				\n\t"
		"paddb %%mm3, %%mm0				\n\t"

		"movq (%%eax), %%mm2				\n\t" // L11
		"movq %%mm2, %%mm3				\n\t" // L11
		"movq %%mm2, %%mm4				\n\t" // L11
		"psllq $8, %%mm3				\n\t"
		"psrlq $8, %%mm4				\n\t"
		"movd -4(%%eax), %%mm5				\n\t"
		"movd 8(%%eax), %%mm6				\n\t"
		"psrlq $24, %%mm5				\n\t"
		"psllq $56, %%mm6				\n\t"
		"por %%mm5, %%mm3				\n\t" // L01
		"por %%mm6, %%mm4				\n\t" // L21
		"movq %%mm3, %%mm5				\n\t" // L01
		PAVGB(%%mm4, %%mm3)				      // (L21 + L01)/2
		PAVGB(%%mm2, %%mm3)				      // (L21 + L01 + 2L11)/4
		"psubusb %%mm7, %%mm2				\n\t"
		"psubusb %%mm7, %%mm4				\n\t"
		"psubusb %%mm7, %%mm5				\n\t"
		"pcmpeqb "MANGLE(b00)", %%mm2			\n\t" // L11 > a ? 0 : -1
		"pcmpeqb "MANGLE(b00)", %%mm4			\n\t" // L21 > a ? 0 : -1
		"pcmpeqb "MANGLE(b00)", %%mm5			\n\t" // L01 > a ? 0 : -1
		"paddb %%mm4, %%mm2				\n\t"
		"paddb %%mm5, %%mm2				\n\t"
// 0, 2, 3, 1
#define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
		"movq " #src ", " #sx "				\n\t" /* src[0] */\
		"movq " #sx ", " #lx "				\n\t" /* src[0] */\
		"movq " #sx ", " #t0 "				\n\t" /* src[0] */\
		"psllq $8, " #lx "				\n\t"\
		"psrlq $8, " #t0 "				\n\t"\
		"movd -4" #src ", " #t1 "			\n\t"\
		"psrlq $24, " #t1 "				\n\t"\
		"por " #t1 ", " #lx "				\n\t" /* src[-1] */\
		"movd 8" #src ", " #t1 "			\n\t"\
		"psllq $56, " #t1 "				\n\t"\
		"por " #t1 ", " #t0 "				\n\t" /* src[+1] */\
		"movq " #lx ", " #t1 "				\n\t" /* src[-1] */\
		PAVGB(t0, lx)				              /* (src[-1] + src[+1])/2 */\
		PAVGB(sx, lx)				      /* (src[-1] + 2src[0] + src[+1])/4 */\
		PAVGB(lx, pplx)					     \
		"movq " #lx ", 8(%%ecx)				\n\t"\
		"movq (%%ecx), " #lx "				\n\t"\
		"psubusb " #lx ", " #t1 "			\n\t"\
		"psubusb " #lx ", " #t0 "			\n\t"\
		"psubusb " #lx ", " #sx "			\n\t"\
		"movq "MANGLE(b00)", " #lx "			\n\t"\
		"pcmpeqb " #lx ", " #t1 "			\n\t" /* src[-1] > a ? 0 : -1*/\
		"pcmpeqb " #lx ", " #t0 "			\n\t" /* src[+1] > a ? 0 : -1*/\
		"pcmpeqb " #lx ", " #sx "			\n\t" /* src[0]  > a ? 0 : -1*/\
		"paddb " #t1 ", " #t0 "				\n\t"\
		"paddb " #t0 ", " #sx "				\n\t"\
\
		PAVGB(plx, pplx)				      /* filtered */\
		"movq " #dst ", " #t0 "				\n\t" /* dst */\
		"movq " #t0 ", " #t1 "				\n\t" /* dst */\
		"psubusb %3, " #t0 "				\n\t"\
		"paddusb %3, " #t1 "				\n\t"\
		PMAXUB(t0, pplx)\
		PMINUB(t1, pplx, t0)\
		"paddb " #sx ", " #ppsx "			\n\t"\
		"paddb " #psx ", " #ppsx "			\n\t"\
		"#paddb "MANGLE(b02)", " #ppsx "		\n\t"\
		"pand "MANGLE(b08)", " #ppsx "			\n\t"\
		"pcmpeqb " #lx ", " #ppsx "			\n\t"\
		"pand " #ppsx ", " #pplx "			\n\t"\
		"pandn " #dst ", " #ppsx "			\n\t"\
		"por " #pplx ", " #ppsx "			\n\t"\
		"movq " #ppsx ", " #dst "			\n\t"\
		"movq 8(%%ecx), " #lx "				\n\t"

/*
0000000
1111111

1111110
1111101
1111100
1111011
1111010
1111001

1111000
1110111

*/
//DERING_CORE(dst,src                  ,ppsx ,psx  ,sx   ,pplx ,plx  ,lx   ,t0   ,t1)
DERING_CORE((%%eax),(%%eax, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
DERING_CORE((%0, %1, 4),(%%edx)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
DERING_CORE((%%edx),(%%edx, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
DERING_CORE((%%edx, %1), (%%edx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
DERING_CORE((%%edx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
DERING_CORE((%0, %1, 8),(%%edx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)

		"1:			\n\t"
		: : "r" (src), "r" (stride), "m" (c->pQPb), "m"(c->pQPb2)
		: "%eax", "%edx", "%ecx"
	);
#else
	int y;
	int min=255;
	int max=0;
	int avg;
	uint8_t *p;
	int s[10];
	const int QP2= c->QP/2 + 1;

	for(y=1; y<9; y++)
	{
		int x;
		p= src + stride*y;
		for(x=1; x<9; x++)
		{
			p++;
			if(*p > max) max= *p;
			if(*p < min) min= *p;
		}
	}
	avg= (min + max + 1)>>1;

	if(max - min <deringThreshold) return;

	for(y=0; y<10; y++)
	{
		int t = 0;

		if(src[stride*y + 0] > avg) t+= 1;
		if(src[stride*y + 1] > avg) t+= 2;
		if(src[stride*y + 2] > avg) t+= 4;
		if(src[stride*y + 3] > avg) t+= 8;
		if(src[stride*y + 4] > avg) t+= 16;
		if(src[stride*y + 5] > avg) t+= 32;
		if(src[stride*y + 6] > avg) t+= 64;
		if(src[stride*y + 7] > avg) t+= 128;
		if(src[stride*y + 8] > avg) t+= 256;
		if(src[stride*y + 9] > avg) t+= 512;
		
		t |= (~t)<<16;
		t &= (t<<1) & (t>>1);
		s[y] = t;
	}
	
	for(y=1; y<9; y++)
	{
		int t = s[y-1] & s[y] & s[y+1];
		t|= t>>16;
		s[y-1]= t;
	}

	for(y=1; y<9; y++)
	{
		int x;
		int t = s[y-1];

		p= src + stride*y;
		for(x=1; x<9; x++)
		{
			p++;
			if(t & (1<<x))
			{
				int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
				      +2*(*(p     -1)) + 4*(*p         ) + 2*(*(p     +1))
				      +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
				f= (f + 8)>>4;

#ifdef DEBUG_DERING_THRESHOLD
				asm volatile("emms\n\t":);
				{
				static long long numPixels=0;
				if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
//				if((max-min)<20 || (max-min)*QP<200)
//				if((max-min)*QP < 500)
//				if(max-min<QP/2)
				if(max-min < 20)
				{
					static int numSkiped=0;
					static int errorSum=0;
					static int worstQP=0;
					static int worstRange=0;
					static int worstDiff=0;
					int diff= (f - *p);
					int absDiff= ABS(diff);
					int error= diff*diff;

					if(x==1 || x==8 || y==1 || y==8) continue;

					numSkiped++;
					if(absDiff > worstDiff)
					{
						worstDiff= absDiff;
						worstQP= QP;
						worstRange= max-min;
					}
					errorSum+= error;

					if(1024LL*1024LL*1024LL % numSkiped == 0)
					{
						printf( "sum:%1.3f, skip:%d, wQP:%d, "
							"wRange:%d, wDiff:%d, relSkip:%1.3f\n",
							(float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
							worstDiff, (float)numSkiped/numPixels);
					}
				}
				}
#endif
				if     (*p + QP2 < f) *p= *p + QP2;
				else if(*p - QP2 > f) *p= *p - QP2;
				else *p=f;
			}
		}
	}
#ifdef DEBUG_DERING_THRESHOLD
	if(max-min < 20)
	{
		for(y=1; y<9; y++)
		{
			int x;
			int t = 0;
			p= src + stride*y;
			for(x=1; x<9; x++)
			{
				p++;
				*p = MIN(*p + 20, 255);
			}
		}
//		src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
	}
#endif
#endif
}

/**
 * Deinterlaces the given block by linearly interpolating every second line.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 */
static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= 4*stride;
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ecx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1	%0+8%1	ecx+4%1

		"movq (%0), %%mm0				\n\t"
		"movq (%%eax, %1), %%mm1			\n\t"
		PAVGB(%%mm1, %%mm0)
		"movq %%mm0, (%%eax)				\n\t"
		"movq (%0, %1, 4), %%mm0			\n\t"
		PAVGB(%%mm0, %%mm1)
		"movq %%mm1, (%%eax, %1, 2)			\n\t"
		"movq (%%ecx, %1), %%mm1			\n\t"
		PAVGB(%%mm1, %%mm0)
		"movq %%mm0, (%%ecx)				\n\t"
		"movq (%0, %1, 8), %%mm0			\n\t"
		PAVGB(%%mm0, %%mm1)
		"movq %%mm1, (%%ecx, %1, 2)			\n\t"

		: : "r" (src), "r" (stride)
		: "%eax", "%ecx"
	);
#else
	int a, b, x;
	src+= 4*stride;

	for(x=0; x<2; x++){
		a= *(uint32_t*)&src[stride*0];
		b= *(uint32_t*)&src[stride*2];
		*(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
		a= *(uint32_t*)&src[stride*4];
		*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
		b= *(uint32_t*)&src[stride*6];
		*(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
		a= *(uint32_t*)&src[stride*8];
		*(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);
		src += 4;
	}
#endif
}

/**
 * Deinterlaces the given block by cubic interpolating every second line.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 * this filter will read lines 3-15 and write 7-13
 */
static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*3;
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
		"leal (%%edx, %1, 4), %%ecx			\n\t"
		"addl %1, %%ecx					\n\t"
		"pxor %%mm7, %%mm7				\n\t"
//	0	1	2	3	4	5	6	7	8	9	10
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

#define DEINT_CUBIC(a,b,c,d,e)\
		"movq " #a ", %%mm0				\n\t"\
		"movq " #b ", %%mm1				\n\t"\
		"movq " #d ", %%mm2				\n\t"\
		"movq " #e ", %%mm3				\n\t"\
		PAVGB(%%mm2, %%mm1)					/* (b+d) /2 */\
		PAVGB(%%mm3, %%mm0)					/* a(a+e) /2 */\
		"movq %%mm0, %%mm2				\n\t"\
		"punpcklbw %%mm7, %%mm0				\n\t"\
		"punpckhbw %%mm7, %%mm2				\n\t"\
		"movq %%mm1, %%mm3				\n\t"\
		"punpcklbw %%mm7, %%mm1				\n\t"\
		"punpckhbw %%mm7, %%mm3				\n\t"\
		"psubw %%mm1, %%mm0				\n\t"	/* L(a+e - (b+d))/2 */\
		"psubw %%mm3, %%mm2				\n\t"	/* H(a+e - (b+d))/2 */\
		"psraw $3, %%mm0				\n\t"	/* L(a+e - (b+d))/16 */\
		"psraw $3, %%mm2				\n\t"	/* H(a+e - (b+d))/16 */\
		"psubw %%mm0, %%mm1				\n\t"	/* L(9b + 9d - a - e)/16 */\
		"psubw %%mm2, %%mm3				\n\t"	/* H(9b + 9d - a - e)/16 */\
		"packuswb %%mm3, %%mm1				\n\t"\
		"movq %%mm1, " #c "				\n\t"

DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx, %1))
DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%edx), (%%edx, %1), (%0, %1, 8))
DEINT_CUBIC((%0, %1, 4), (%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%ecx))
DEINT_CUBIC((%%edx, %1), (%0, %1, 8), (%%edx, %1, 4), (%%ecx), (%%ecx, %1, 2))

		: : "r" (src), "r" (stride)
		: "%eax", "%edx", "ecx"
	);
#else
	int x;
	src+= stride*3;
	for(x=0; x<8; x++)
	{
		src[stride*3] = CLIP((-src[0]        + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);
		src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);
		src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);
		src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);
		src++;
	}
#endif
}

/**
 * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 * this filter will read lines 4-13 and write 5-11
 */
static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*4;
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
		"pxor %%mm7, %%mm7				\n\t"
		"movq (%2), %%mm0				\n\t"
//	0	1	2	3	4	5	6	7	8	9	10
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

#define DEINT_FF(a,b,c,d)\
		"movq " #a ", %%mm1				\n\t"\
		"movq " #b ", %%mm2				\n\t"\
		"movq " #c ", %%mm3				\n\t"\
		"movq " #d ", %%mm4				\n\t"\
		PAVGB(%%mm3, %%mm1)					\
		PAVGB(%%mm4, %%mm0)					\
		"movq %%mm0, %%mm3				\n\t"\
		"punpcklbw %%mm7, %%mm0				\n\t"\
		"punpckhbw %%mm7, %%mm3				\n\t"\
		"movq %%mm1, %%mm4				\n\t"\
		"punpcklbw %%mm7, %%mm1				\n\t"\
		"punpckhbw %%mm7, %%mm4				\n\t"\
		"psllw $2, %%mm1				\n\t"\
		"psllw $2, %%mm4				\n\t"\
		"psubw %%mm0, %%mm1				\n\t"\
		"psubw %%mm3, %%mm4				\n\t"\
		"movq %%mm2, %%mm5				\n\t"\
		"movq %%mm2, %%mm0				\n\t"\
		"punpcklbw %%mm7, %%mm2				\n\t"\
		"punpckhbw %%mm7, %%mm5				\n\t"\
		"paddw %%mm2, %%mm1				\n\t"\
		"paddw %%mm5, %%mm4				\n\t"\
		"psraw $2, %%mm1				\n\t"\
		"psraw $2, %%mm4				\n\t"\
		"packuswb %%mm4, %%mm1				\n\t"\
		"movq %%mm1, " #b "				\n\t"\

DEINT_FF((%0)       , (%%eax)       , (%%eax, %1), (%%eax, %1, 2))
DEINT_FF((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx)       )
DEINT_FF((%0, %1, 4), (%%edx)       , (%%edx, %1), (%%edx, %1, 2))
DEINT_FF((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%edx, %1, 4))

		"movq %%mm0, (%2)				\n\t"
		: : "r" (src), "r" (stride), "r"(tmp)
		: "%eax", "%edx"
	);
#else
	int x;
	src+= stride*4;
	for(x=0; x<8; x++)
	{
		int t1= tmp[x];
		int t2= src[stride*1];

		src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);
		t1= src[stride*4];
		src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);
		t2= src[stride*6];
		src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);
		t1= src[stride*8];
		src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);
		tmp[x]= t1;

		src++;
	}
#endif
}

/**
 * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 * this filter will read lines 4-13 and write 4-11
 */
static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*4;
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
		"pxor %%mm7, %%mm7				\n\t"
		"movq (%2), %%mm0				\n\t"
		"movq (%3), %%mm1				\n\t"
//	0	1	2	3	4	5	6	7	8	9	10
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

#define DEINT_L5(t1,t2,a,b,c)\
		"movq " #a ", %%mm2				\n\t"\
		"movq " #b ", %%mm3				\n\t"\
		"movq " #c ", %%mm4				\n\t"\
		PAVGB(t2, %%mm3)					\
		PAVGB(t1, %%mm4)					\
		"movq %%mm2, %%mm5				\n\t"\
		"movq %%mm2, " #t1 "				\n\t"\
		"punpcklbw %%mm7, %%mm2				\n\t"\
		"punpckhbw %%mm7, %%mm5				\n\t"\
		"movq %%mm2, %%mm6				\n\t"\
		"paddw %%mm2, %%mm2				\n\t"\
		"paddw %%mm6, %%mm2				\n\t"\
		"movq %%mm5, %%mm6				\n\t"\
		"paddw %%mm5, %%mm5				\n\t"\
		"paddw %%mm6, %%mm5				\n\t"\
		"movq %%mm3, %%mm6				\n\t"\
		"punpcklbw %%mm7, %%mm3				\n\t"\
		"punpckhbw %%mm7, %%mm6				\n\t"\
		"paddw %%mm3, %%mm3				\n\t"\
		"paddw %%mm6, %%mm6				\n\t"\
		"paddw %%mm3, %%mm2				\n\t"\
		"paddw %%mm6, %%mm5				\n\t"\
		"movq %%mm4, %%mm6				\n\t"\
		"punpcklbw %%mm7, %%mm4				\n\t"\
		"punpckhbw %%mm7, %%mm6				\n\t"\
		"psubw %%mm4, %%mm2				\n\t"\
		"psubw %%mm6, %%mm5				\n\t"\
		"psraw $2, %%mm2				\n\t"\
		"psraw $2, %%mm5				\n\t"\
		"packuswb %%mm5, %%mm2				\n\t"\
		"movq %%mm2, " #a "				\n\t"\

DEINT_L5(%%mm0, %%mm1, (%0)          , (%%eax)       , (%%eax, %1)   )
DEINT_L5(%%mm1, %%mm0, (%%eax)       , (%%eax, %1)   , (%%eax, %1, 2))
DEINT_L5(%%mm0, %%mm1, (%%eax, %1)   , (%%eax, %1, 2), (%0, %1, 4)   )
DEINT_L5(%%mm1, %%mm0, (%%eax, %1, 2), (%0, %1, 4)   , (%%edx)       )
DEINT_L5(%%mm0, %%mm1, (%0, %1, 4)   , (%%edx)       , (%%edx, %1)   )  
DEINT_L5(%%mm1, %%mm0, (%%edx)       , (%%edx, %1)   , (%%edx, %1, 2))
DEINT_L5(%%mm0, %%mm1, (%%edx, %1)   , (%%edx, %1, 2), (%0, %1, 8)   )
DEINT_L5(%%mm1, %%mm0, (%%edx, %1, 2), (%0, %1, 8)   , (%%edx, %1, 4))

		"movq %%mm0, (%2)				\n\t"
		"movq %%mm1, (%3)				\n\t"
		: : "r" (src), "r" (stride), "r"(tmp), "r"(tmp2)
		: "%eax", "%edx"
	);
#else
	int x;
	src+= stride*4;
	for(x=0; x<8; x++)
	{
		int t1= tmp[x];
		int t2= tmp2[x];
		int t3= src[0];

		src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);
		t1= src[stride*1];
		src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);
		t2= src[stride*2];
		src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);
		t3= src[stride*3];
		src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);
		t1= src[stride*4];
		src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);
		t2= src[stride*5];
		src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);
		t3= src[stride*6];
		src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);
		t1= src[stride*7];
		src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);

		tmp[x]= t3;
		tmp2[x]= t1;

		src++;
	}
#endif
}

/**
 * Deinterlaces the given block by filtering all lines with a (1 2 1) filter.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 * this filter will read lines 4-13 and write 4-11
 */
static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= 4*stride;
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

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

		: : "r" (src), "r" (stride), "r" (tmp)
		: "%eax", "%edx"
	);
#else
	int a, b, c, x;
	src+= 4*stride;

	for(x=0; x<2; x++){
		a= *(uint32_t*)&tmp[stride*0];
		b= *(uint32_t*)&src[stride*0];
		c= *(uint32_t*)&src[stride*1];
		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

		a= *(uint32_t*)&src[stride*2];
		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

		b= *(uint32_t*)&src[stride*3];
		c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);

		c= *(uint32_t*)&src[stride*4];
		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

		a= *(uint32_t*)&src[stride*5];
		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

		b= *(uint32_t*)&src[stride*6];
		c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);

		c= *(uint32_t*)&src[stride*7];
		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

		a= *(uint32_t*)&src[stride*8];
		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);
		*(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

		*(uint32_t*)&tmp[stride*0]= c;
		src += 4;
		tmp += 4;
	}
#endif
}

/**
 * Deinterlaces the given block by applying a median filter to every second line.
 * will be called for every 8x8 block and can read & write from line 4-15,
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
 * lines 4-12 will be read into the deblocking filter and should be deinterlaced
 */
static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
{
#ifdef HAVE_MMX
	src+= 4*stride;
#ifdef HAVE_MMX2
	asm volatile(
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%edx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+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 (%%edx), %%mm2				\n\t" //
		"movq (%%edx, %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, (%%edx)				\n\t"

		"movq (%%edx, %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, (%%edx, %1, 2)			\n\t"


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

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