postprocess.c

		PAVGB(%%mm3, %%mm1)				      /* p5 = (|p2-p1| + |p6-p5|)/2 */\
		"paddb %%mm6, %%mm4				\n\t"\
		"psrlq $16, %%mm1				\n\t" /* p3 = (|p2-p1| + |p6-p5|)/2 */\
		"psubusb %%mm1, %%mm3			\n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
		"paddb %%mm5, %%mm3				\n\t"\
		"psubusb %%mm5, %%mm3				\n\t"\
		PAVGB(%%mm7, %%mm3)\
		"pxor %%mm2, %%mm3				\n\t"\
		"psubb %%mm2, %%mm3				\n\t"\
		"psrlq $24, %%mm3				\n\t"\
		"movd " #c ", %%mm0				\n\t"\
		"movd 4" #c ", %%mm1				\n\t"\
		"punpckldq %%mm1, %%mm0				\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"paddsb (%2, %%ecx, 8), %%mm0			\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"movq %%mm0, " #c "				\n\t"\
		"movd %%mm3, %%ecx				\n\t"\
		"movd " #d ", %%mm0				\n\t"\
		"paddsb (%2, %%ecx, 8), %%mm4			\n\t"\
		"movd 4" #d ", %%mm1				\n\t"\
		"paddb %%mm6, %%mm4				\n\t"\
		"punpckldq %%mm1, %%mm0				\n\t"\
		"movq %%mm4, " #b "				\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"paddsb (%2, %%ecx, 8), %%mm0			\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"movq %%mm0, " #d "				\n\t"\

HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2))
HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2))


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

//FIXME (has little in common with the mmx2 version)
	for(y=0; y<BLOCK_SIZE; y++)
	{
		int a= src[1] - src[2];
		int b= src[3] - src[4];
		int c= src[5] - src[6];

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

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

			src[1] +=v/8;
			src[2] +=v/4;
			src[3] +=3*v/8;
			src[4] -=3*v/8;
			src[5] -=v/4;
			src[6] -=v/8;

		}
		src+=stride;
	}
#endif
}


static inline void doVertDefFilter(uint8_t src[], int stride, int QP)
{
#ifdef HAVE_MMX
	src+= stride*4;
	//FIXME try pmul for *5 stuff
//	src[0]=0;
	asm volatile(
		"pxor %%mm7, %%mm7				\n\t"
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7
//	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ebx+%1	ebx+2%1
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1

		"movq (%0), %%mm0				\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"punpcklbw %%mm7, %%mm0				\n\t" // low part of line 0
		"punpckhbw %%mm7, %%mm1				\n\t" // high part of line 0

		"movq (%%eax), %%mm2				\n\t"
		"movq %%mm2, %%mm3				\n\t"
		"punpcklbw %%mm7, %%mm2				\n\t" // low part of line 1
		"punpckhbw %%mm7, %%mm3				\n\t" // high part of line 1

		"movq (%%eax, %1), %%mm4			\n\t"
		"movq %%mm4, %%mm5				\n\t"
		"punpcklbw %%mm7, %%mm4				\n\t" // low part of line 2
		"punpckhbw %%mm7, %%mm5				\n\t" // high part of line 2

		"paddw %%mm0, %%mm0				\n\t" // 2L0
		"paddw %%mm1, %%mm1				\n\t" // 2H0
		"psubw %%mm4, %%mm2				\n\t" // L1 - L2
		"psubw %%mm5, %%mm3				\n\t" // H1 - H2
		"psubw %%mm2, %%mm0				\n\t" // 2L0 - L1 + L2
		"psubw %%mm3, %%mm1				\n\t" // 2H0 - H1 + H2

		"psllw $2, %%mm2				\n\t" // 4L1 - 4L2
		"psllw $2, %%mm3				\n\t" // 4H1 - 4H2
		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2
		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2

		"movq (%%eax, %1, 2), %%mm2			\n\t"
		"movq %%mm2, %%mm3				\n\t"
		"punpcklbw %%mm7, %%mm2				\n\t" // L3
		"punpckhbw %%mm7, %%mm3				\n\t" // H3

		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2 - L3
		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2 - H3
		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2 - 2L3
		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2 - 2H3
		"movq %%mm0, temp0				\n\t" // 2L0 - 5L1 + 5L2 - 2L3
		"movq %%mm1, temp1				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

		"movq (%0, %1, 4), %%mm0			\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"punpcklbw %%mm7, %%mm0				\n\t" // L4
		"punpckhbw %%mm7, %%mm1				\n\t" // H4

		"psubw %%mm0, %%mm2				\n\t" // L3 - L4
		"psubw %%mm1, %%mm3				\n\t" // H3 - H4
		"movq %%mm2, temp2				\n\t" // L3 - L4
		"movq %%mm3, temp3				\n\t" // H3 - H4
		"paddw %%mm4, %%mm4				\n\t" // 2L2
		"paddw %%mm5, %%mm5				\n\t" // 2H2
		"psubw %%mm2, %%mm4				\n\t" // 2L2 - L3 + L4
		"psubw %%mm3, %%mm5				\n\t" // 2H2 - H3 + H4

		"psllw $2, %%mm2				\n\t" // 4L3 - 4L4
		"psllw $2, %%mm3				\n\t" // 4H3 - 4H4
		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4
		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4
//50 opcodes so far
		"movq (%%ebx), %%mm2				\n\t"
		"movq %%mm2, %%mm3				\n\t"
		"punpcklbw %%mm7, %%mm2				\n\t" // L5
		"punpckhbw %%mm7, %%mm3				\n\t" // H5
		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4 - L5
		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4 - H5
		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4 - 2L5
		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4 - 2H5

		"movq (%%ebx, %1), %%mm6			\n\t"
		"punpcklbw %%mm7, %%mm6				\n\t" // L6
		"psubw %%mm6, %%mm2				\n\t" // L5 - L6
		"movq (%%ebx, %1), %%mm6			\n\t"
		"punpckhbw %%mm7, %%mm6				\n\t" // H6
		"psubw %%mm6, %%mm3				\n\t" // H5 - H6

		"paddw %%mm0, %%mm0				\n\t" // 2L4
		"paddw %%mm1, %%mm1				\n\t" // 2H4
		"psubw %%mm2, %%mm0				\n\t" // 2L4 - L5 + L6
		"psubw %%mm3, %%mm1				\n\t" // 2H4 - H5 + H6

		"psllw $2, %%mm2				\n\t" // 4L5 - 4L6
		"psllw $2, %%mm3				\n\t" // 4H5 - 4H6
		"psubw %%mm2, %%mm0				\n\t" // 2L4 - 5L5 + 5L6
		"psubw %%mm3, %%mm1				\n\t" // 2H4 - 5H5 + 5H6

		"movq (%%ebx, %1, 2), %%mm2			\n\t"
		"movq %%mm2, %%mm3				\n\t"
		"punpcklbw %%mm7, %%mm2				\n\t" // L7
		"punpckhbw %%mm7, %%mm3				\n\t" // H7

		"paddw %%mm2, %%mm2				\n\t" // 2L7
		"paddw %%mm3, %%mm3				\n\t" // 2H7
		"psubw %%mm2, %%mm0				\n\t" // 2L4 - 5L5 + 5L6 - 2L7
		"psubw %%mm3, %%mm1				\n\t" // 2H4 - 5H5 + 5H6 - 2H7

		"movq temp0, %%mm2				\n\t" // 2L0 - 5L1 + 5L2 - 2L3
		"movq temp1, %%mm3				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

#ifdef HAVE_MMX2
		"movq %%mm7, %%mm6				\n\t" // 0
		"psubw %%mm0, %%mm6				\n\t"
		"pmaxsw %%mm6, %%mm0				\n\t" // |2L4 - 5L5 + 5L6 - 2L7|
		"movq %%mm7, %%mm6				\n\t" // 0
		"psubw %%mm1, %%mm6				\n\t"
		"pmaxsw %%mm6, %%mm1				\n\t" // |2H4 - 5H5 + 5H6 - 2H7|
		"movq %%mm7, %%mm6				\n\t" // 0
		"psubw %%mm2, %%mm6				\n\t"
		"pmaxsw %%mm6, %%mm2				\n\t" // |2L0 - 5L1 + 5L2 - 2L3|
		"movq %%mm7, %%mm6				\n\t" // 0
		"psubw %%mm3, %%mm6				\n\t"
		"pmaxsw %%mm6, %%mm3				\n\t" // |2H0 - 5H1 + 5H2 - 2H3|
#else
		"movq %%mm7, %%mm6				\n\t" // 0
		"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
		"punpcklwd %%mm2, %%mm2				\n\t"
		"punpcklwd %%mm2, %%mm2				\n\t"
		"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 w05, %%mm2				\n\t" // 5
		"pmullw %%mm2, %%mm4				\n\t"
		"pmullw %%mm2, %%mm5				\n\t"
		"movq 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 w06, %%mm2				\n\t" // 6
		"paddw %%mm2, %%mm4				\n\t"
		"paddw %%mm2, %%mm5				\n\t"
		"movq w1400, %%mm2				\n\t" // 1400h = 5120 = 5/64*2^16
//FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120
		"pmulhw %%mm2, %%mm4				\n\t" // hd/13
		"pmulhw %%mm2, %%mm5				\n\t" // ld/13
*/

		"movq temp2, %%mm0				\n\t" // L3 - L4
		"movq temp3, %%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 (%%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"
		"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;
	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*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"
 */
static inline int isHorizDC(uint8_t src[], int stride)
{
//	src++;
	int numEq= 0;
#if 0
asm volatile (
//		"int $3 \n\t"
		"leal (%1, %2), %%ecx				\n\t"
		"leal (%%ecx, %2, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%1	ecx	ecx+%2	ecx+2%2	%1+4%2	ebx	ebx+%2	ebx+2%2	%1+8%2	ebx+4%2
		"movq b7E, %%mm7				\n\t" // mm7 = 0x7F
		"movq b7C, %%mm6				\n\t" // mm6 = 0x7D
		"pxor %%mm0, %%mm0				\n\t"
		"movl %1, %%eax					\n\t"
		"andl $0x1F, %%eax				\n\t"
		"cmpl $24, %%eax				\n\t"
		"leal tempBlock, %%eax				\n\t"
		"jb 1f						\n\t"

#define HDC_CHECK_AND_CPY(src, dst) \
		"movd " #src ", %%mm2				\n\t"\
		"punpckldq 4" #src ", %%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," #dst "(%%eax)			\n\t"

		HDC_CHECK_AND_CPY((%1),0)
		HDC_CHECK_AND_CPY((%%ecx),8)
		HDC_CHECK_AND_CPY((%%ecx, %2),16)
		HDC_CHECK_AND_CPY((%%ecx, %2, 2),24)
		HDC_CHECK_AND_CPY((%1, %2, 4),32)
		HDC_CHECK_AND_CPY((%%ebx),40)
		HDC_CHECK_AND_CPY((%%ebx, %2),48)
		HDC_CHECK_AND_CPY((%%ebx, %2, 2),56)
		"jmp 2f						\n\t"
		"1:						\n\t"
// src does not cross a 32 byte cache line so dont waste time with alignment
#define HDC_CHECK_AND_CPY2(src, dst) \
		"movq " #src ", %%mm2				\n\t"\
		"movq " #src ", %%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," #dst "(%%eax)			\n\t"

		HDC_CHECK_AND_CPY2((%1),0)
		HDC_CHECK_AND_CPY2((%%ecx),8)
		HDC_CHECK_AND_CPY2((%%ecx, %2),16)
		HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24)
		HDC_CHECK_AND_CPY2((%1, %2, 4),32)
		HDC_CHECK_AND_CPY2((%%ebx),40)
		HDC_CHECK_AND_CPY2((%%ebx, %2),48)
		HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56)
		"2:						\n\t"
		"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"
		"movd %%mm0, %0					\n\t"
		: "=r" (numEq)
		: "r" (src), "r" (stride)
		: "%eax", "%ebx", "%ecx"
		);
//	printf("%d\n", numEq);
	numEq= (256 - numEq) &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++;
		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)
{
	if(abs(src[0] - src[7]) > 2*QP) return 0;

	return 1;
}

static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP)
{
#if 0
	asm volatile(
		"leal (%0, %1), %%ecx				\n\t"
		"leal (%%ecx, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	ecx	ecx+%1	ecx+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1
		"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(src, dst)	\
		"movq " #src "(%%eax), %%mm0			\n\t"\
		"movq " #src "(%%eax), %%mm1			\n\t"\
		"movq " #src "(%%eax), %%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, " #dst"				\n\t"\
		"psrlq $32, %%mm0				\n\t"\
		"movd %%mm0, 4" #dst"				\n\t"
#else
#define HDF(src, dst)\
		"movq " #src "(%%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, " #dst "				\n\t"\
		"psrlq $32, %%mm0				\n\t"\
		"movd %%mm0, 4" #dst "				\n\t"
#endif
		HDF(0,(%0))
		HDF(8,(%%ecx))
		HDF(16,(%%ecx, %1))
		HDF(24,(%%ecx, %1, 2))
		HDF(32,(%0, %1, 4))
		HDF(40,(%%ebx))
		HDF(48,(%%ebx, %1))
		HDF(56,(%%ebx, %1, 2))
		:
		: "r" (dst), "r" (stride), "r" (QP)
		: "%eax", "%ebx", "%ecx"
	);
#else
	int y;
	for(y=0; y<BLOCK_SIZE; y++)
	{
		const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]);

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

/**
 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
 * using the 7-Tap Filter   (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version)
 */
static inline void doHorizLowPass(uint8_t dst[], int stride, int QP)
{

#if 0
	asm volatile(
		"leal (%0, %1), %%ecx				\n\t"
		"leal (%%ecx, %1, 4), %%ebx			\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	ecx	ecx+%1	ecx+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1
		"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
/*
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"\
		"movd -4(%0), %%mm3					\n\t" /*0001000*/\
		"movd 8(%0), %%mm4					\n\t" /*0001000*/\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"psrlq $24, %%mm3					\n\t"\
		"psllq $56, %%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

/* uses the 7-Tap Filter: 1112111 */
#define NEW_HLP(src, dst)\
		"movq " #src "(%%eax), %%mm1				\n\t"\
		"movq " #src "(%%eax), %%mm2				\n\t"\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"movd -4" #dst ", %%mm3					\n\t" /*0001000*/\
		"movd 8" #dst ", %%mm4					\n\t" /*0001000*/\
		"psrlq $24, %%mm3					\n\t"\
		"psllq $56, %%mm4					\n\t"\
		"por %%mm3, %%mm1					\n\t"\
		"por %%mm4, %%mm2					\n\t"\
		"movq %%mm1, %%mm5					\n\t"\
		PAVGB(%%mm2, %%mm1)\
		"movq " #src "(%%eax), %%mm0				\n\t"\
		PAVGB(%%mm1, %%mm0)\
		"psllq $8, %%mm5					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"por %%mm3, %%mm5					\n\t"\
		"por %%mm4, %%mm2					\n\t"\
		"movq %%mm5, %%mm1					\n\t"\
		PAVGB(%%mm2, %%mm5)\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"por %%mm3, %%mm1					\n\t"\
		"por %%mm4, %%mm2					\n\t"\
		PAVGB(%%mm2, %%mm1)\
		PAVGB(%%mm1, %%mm5)\
		PAVGB(%%mm5, %%mm0)\
		"movd %%mm0, " #dst "					\n\t"\
		"psrlq $32, %%mm0					\n\t"\
		"movd %%mm0, 4" #dst "					\n\t"

/* uses the 9-Tap Filter: 112242211 */
#define NEW_HLP2(i)\
		"movq " #i "(%%eax), %%mm0				\n\t" /*0001000*/\
		"movq %%mm0, %%mm1					\n\t" /*0001000*/\
		"movq %%mm0, %%mm2					\n\t" /*0001000*/\
		"movd -4(%0), %%mm3					\n\t" /*0001000*/\
		"movd 8(%0), %%mm4					\n\t" /*0001000*/\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"psrlq $24, %%mm3					\n\t"\
		"psllq $56, %%mm4					\n\t"\
		"por %%mm3, %%mm1					\n\t" /*0010000*/\
		"por %%mm4, %%mm2					\n\t" /*0000100*/\
		"movq %%mm1, %%mm5					\n\t" /*0010000*/\
		PAVGB(%%mm2, %%mm1)					      /*0010100*/\
		PAVGB(%%mm1, %%mm0)					      /*0012100*/\
		"psllq $8, %%mm5					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"por %%mm3, %%mm5					\n\t" /*0100000*/\
		"por %%mm4, %%mm2					\n\t" /*0000010*/\
		"movq %%mm5, %%mm1					\n\t" /*0100000*/\
		PAVGB(%%mm2, %%mm5)					      /*0100010*/\
		"psllq $8, %%mm1					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"por %%mm3, %%mm1					\n\t" /*1000000*/\
		"por %%mm4, %%mm2					\n\t" /*0000001*/\
		"movq %%mm1, %%mm6					\n\t" /*1000000*/\
		PAVGB(%%mm2, %%mm1)					      /*1000001*/\
		"psllq $8, %%mm6					\n\t"\
		"psrlq $8, %%mm2					\n\t"\
		"por %%mm3, %%mm6					\n\t"/*100000000*/\
		"por %%mm4, %%mm2					\n\t"/*000000001*/\
		PAVGB(%%mm2, %%mm6)					     /*100000001*/\
		PAVGB(%%mm6, %%mm1)					     /*110000011*/\
		PAVGB(%%mm1, %%mm5)					     /*112000211*/\
		PAVGB(%%mm5, %%mm0)					     /*112242211*/\
		"movd %%mm0, (%0)					\n\t"\
		"psrlq $32, %%mm0					\n\t"\
		"movd %%mm0, 4(%0)					\n\t"

#define HLP(src, dst) NEW_HLP(src, dst)

		HLP(0, (%0))
		HLP(8, (%%ecx))
		HLP(16, (%%ecx, %1))
		HLP(24, (%%ecx, %1, 2))
		HLP(32, (%0, %1, 4))
		HLP(40, (%%ebx))
		HLP(48, (%%ebx, %1))
		HLP(56, (%%ebx, %1, 2))

		:
		: "r" (dst), "r" (stride)
		: "%eax", "%ebx", "%ecx"
	);

#else
	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 + dst[0];
		sums[1] = dst[0] + dst[1];
		sums[2] = dst[1] + dst[2];
		sums[3] = dst[2] + dst[3];
		sums[4] = dst[3] + dst[4];
		sums[5] = dst[4] + dst[5];
		sums[6] = dst[5] + dst[6];
		sums[7] = dst[6] + dst[7];
		sums[8] = dst[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;
	}
#endif
}

static inline void dering(uint8_t src[], int stride, int QP)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"movq pQPb, %%mm0				\n\t"
		"paddusb %%mm0, %%mm0				\n\t"
		"movq %%mm0, pQPb2				\n\t"

		"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

		"pcmpeqb %%mm6, %%mm6				\n\t"
		"pxor %%mm7, %%mm7				\n\t"
#ifdef HAVE_MMX2
#define FIND_MIN_MAX(addr)\
		"movq " #addr ", %%mm0				\n\t"\
		"pminub %%mm0, %%mm6				\n\t"\
		"pmaxub %%mm0, %%mm7				\n\t"
#else
#define FIND_MIN_MAX(addr)\
		"movq " #addr ", %%mm0				\n\t"\
		"movq %%mm6, %%mm1				\n\t"\
		"psubusb %%mm0, %%mm7				\n\t"\
		"paddb %%mm0, %%mm7				\n\t"\
		"psubusb %%mm0, %%mm1				\n\t"\
		"psubb %%mm1, %%mm6				\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((%%ebx))
FIND_MIN_MAX((%%ebx, %1))
FIND_MIN_MAX((%%ebx, %1, 2))
FIND_MIN_MAX((%0, %1, 8))

		"movq %%mm6, %%mm4				\n\t"
		"psrlq $8, %%mm6				\n\t"
#ifdef HAVE_MMX2
		"pminub %%mm4, %%mm6				\n\t" // min of pixels
		"pshufw $0xF9, %%mm6, %%mm4			\n\t"
		"pminub %%mm4, %%mm6				\n\t" // min of pixels
		"pshufw $0xFE, %%mm6, %%mm4			\n\t"
		"pminub %%mm4, %%mm6				\n\t"
#else
		"movq %%mm6, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $16, %%mm6				\n\t"
		"movq %%mm6, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm6				\n\t"
		"movq %%mm6, %%mm4				\n\t"
		"psrlq $32, %%mm6				\n\t"
		"movq %%mm6, %%mm1				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"psubb %%mm1, %%mm6				\n\t"
#endif


		"movq %%mm7, %%mm4				\n\t"
		"psrlq $8, %%mm7				\n\t"
#ifdef HAVE_MMX2
		"pmaxub %%mm4, %%mm7				\n\t" // max of pixels
		"pshufw $0xF9, %%mm7, %%mm4			\n\t"
		"pmaxub %%mm4, %%mm7				\n\t"
		"pshufw $0xFE, %%mm7, %%mm4			\n\t"
		"pmaxub %%mm4, %%mm7				\n\t"
#else
		"psubusb %%mm4, %%mm7				\n\t"
		"paddb %%mm4, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $16, %%mm7				\n\t"
		"psubusb %%mm4, %%mm7				\n\t"
		"paddb %%mm4, %%mm7				\n\t"
		"movq %%mm7, %%mm4				\n\t"
		"psrlq $32, %%mm7				\n\t"
		"psubusb %%mm4, %%mm7				\n\t"
		"paddb %%mm4, %%mm7				\n\t"
#endif
		PAVGB(%%mm6, %%mm7)				      // a=(max + min)/2
		"punpcklbw %%mm7, %%mm7				\n\t"
		"punpcklbw %%mm7, %%mm7				\n\t"
		"punpcklbw %%mm7, %%mm7				\n\t"
		"movq %%mm7, temp0				\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 b00, %%mm0				\n\t" // L10 > a ? 0 : -1
		"pcmpeqb b00, %%mm2				\n\t" // L20 > a ? 0 : -1
		"pcmpeqb 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 b00, %%mm2				\n\t" // L11 > a ? 0 : -1
		"pcmpeqb b00, %%mm4				\n\t" // L21 > a ? 0 : -1
		"pcmpeqb 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 */\