runtime cpu detection

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@3100 b3059339-0415-0410-9bf9-f77b7e298cf2

2 files changed, 174 insertions, 3865 deletions

diff --git a/postproc/postprocess.c b/postproc/postprocess.c
index a2e9174b70..e54bcf05e7 100644
--- a/postproc/postprocess.c
+++ b/postproc/postprocess.c
@@ -62,6 +62,8 @@ optimize c versions
 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
 smart blur
 commandline option for   the deblock / dering thresholds
+put fastmemcpy back
+dont use #ifdef ARCH_X86 for the asm stuff ... cross compilers? (note cpudetect uses ARCH_X86)
 ...
 */
 
@@ -78,43 +80,25 @@ commandline option for   the deblock / dering thresholds
 //#undef HAVE_MMX2
 //#define HAVE_3DNOW
 //#undef HAVE_MMX
+//#undef ARCH_X86
 //#define DEBUG_BRIGHTNESS
-#include "../libvo/fastmemcpy.h"
+//#include "../libvo/fastmemcpy.h"
 #include "postprocess.h"
+#include "../cpudetect.h"
 
 #define MIN(a,b) ((a) > (b) ? (b) : (a))
 #define MAX(a,b) ((a) < (b) ? (b) : (a))
 #define ABS(a) ((a) > 0 ? (a) : (-(a)))
 #define SIGN(a) ((a) > 0 ? 1 : -1)
 
-#ifdef HAVE_MMX2
-#define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
-#elif defined (HAVE_3DNOW)
-#define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
-#endif
-
-#ifdef HAVE_MMX2
-#define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
-#elif defined (HAVE_MMX)
-#define PMINUB(b,a,t) \
-	"movq " #a ", " #t " \n\t"\
-	"psubusb " #b ", " #t " \n\t"\
-	"psubb " #t ", " #a " \n\t"
-#endif
-
-#ifdef HAVE_MMX2
-#define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
-#elif defined (HAVE_MMX)
-#define PMAXUB(a,b) \
-	"psubusb " #a ", " #b " \n\t"\
-	"paddb " #a ", " #b " \n\t"
-#endif
-
-
 #define GET_MODE_BUFFER_SIZE 500
 #define OPTIONS_ARRAY_SIZE 10
 
-#ifdef HAVE_MMX
+#ifdef ARCH_X86
+#define CAN_COMPILE_X86_ASM
+#endif
+
+#ifdef CAN_COMPILE_X86_ASM
 static volatile uint64_t __attribute__((aligned(8))) packedYOffset=	0x0000000000000000LL;
 static volatile uint64_t __attribute__((aligned(8))) packedYScale=	0x0100010001000100LL;
 static uint64_t __attribute__((aligned(8))) w05=		0x0005000500050005LL;
@@ -157,7 +141,6 @@ static uint32_t __attribute__((aligned(4))) maxTmpNoise[4];
 #else
 static uint64_t packedYOffset=	0x0000000000000000LL;
 static uint64_t packedYScale=	0x0100010001000100LL;
-static uint8_t tempBlocks[8*16*2]; //used for the horizontal code
 #endif
 
 int hFlatnessThreshold= 56 - 16;
@@ -196,7 +179,7 @@ static char *replaceTable[]=
 	NULL //End Marker
 };
 
-#ifdef HAVE_MMX
+#ifdef CAN_COMPILE_X86_ASM
 static inline void unusedVariableWarningFixer()
 {
 if(
@@ -220,7 +203,7 @@ static inline long long rdtsc()
 }
 #endif
 
-#ifdef HAVE_MMX2
+#ifdef CAN_COMPILE_X86_ASM
 static inline void prefetchnta(void *p)
 {
 	asm volatile(	"prefetchnta (%0)\n\t"
@@ -250,1229 +233,7 @@ static inline void prefetcht2(void *p)
 }
 #endif
 
-//FIXME? |255-0| = 1 (shouldnt be a problem ...)
-/**
- * Check if the middle 8x8 Block in the given 8x16 block is flat
- */
-static inline int isVertDC(uint8_t src[], int stride){
-	int numEq= 0;
-#ifndef HAVE_MMX
-	int y;
-#endif
-	src+= stride*4; // src points to begin of the 8x8 Block
-#ifdef HAVE_MMX
-asm volatile(
-		"leal (%1, %2), %%eax				\n\t"
-		"leal (%%eax, %2, 4), %%ebx			\n\t"
-//	0	1	2	3	4	5	6	7	8	9
-//	%1	eax	eax+%2	eax+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
-		"movq (%1), %%mm0				\n\t"
-		"movq (%%eax), %%mm1				\n\t"
-		"psubb %%mm1, %%mm0				\n\t" // mm0 = differnece
-		"paddb %%mm7, %%mm0				\n\t"
-		"pcmpgtb %%mm6, %%mm0				\n\t"
-
-		"movq (%%eax,%2), %%mm2				\n\t"
-		"psubb %%mm2, %%mm1				\n\t"
-		"paddb %%mm7, %%mm1				\n\t"
-		"pcmpgtb %%mm6, %%mm1				\n\t"
-		"paddb %%mm1, %%mm0				\n\t"
-
-		"movq (%%eax, %2, 2), %%mm1			\n\t"
-		"psubb %%mm1, %%mm2				\n\t"
-		"paddb %%mm7, %%mm2				\n\t"
-		"pcmpgtb %%mm6, %%mm2				\n\t"
-		"paddb %%mm2, %%mm0				\n\t"
-
-		"movq (%1, %2, 4), %%mm2			\n\t"
-		"psubb %%mm2, %%mm1				\n\t"
-		"paddb %%mm7, %%mm1				\n\t"
-		"pcmpgtb %%mm6, %%mm1				\n\t"
-		"paddb %%mm1, %%mm0				\n\t"
-
-		"movq (%%ebx), %%mm1				\n\t"
-		"psubb %%mm1, %%mm2				\n\t"
-		"paddb %%mm7, %%mm2				\n\t"
-		"pcmpgtb %%mm6, %%mm2				\n\t"
-		"paddb %%mm2, %%mm0				\n\t"
-
-		"movq (%%ebx, %2), %%mm2			\n\t"
-		"psubb %%mm2, %%mm1				\n\t"
-		"paddb %%mm7, %%mm1				\n\t"
-		"pcmpgtb %%mm6, %%mm1				\n\t"
-		"paddb %%mm1, %%mm0				\n\t"
-
-		"movq (%%ebx, %2, 2), %%mm1			\n\t"
-		"psubb %%mm1, %%mm2				\n\t"
-		"paddb %%mm7, %%mm2				\n\t"
-		"pcmpgtb %%mm6, %%mm2				\n\t"
-		"paddb %%mm2, %%mm0				\n\t"
-
-		"						\n\t"
-#ifdef HAVE_MMX2
-		"pxor %%mm7, %%mm7				\n\t"
-		"psadbw %%mm7, %%mm0				\n\t"
-#else
-		"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"
-#endif
-		"movd %%mm0, %0					\n\t"
-		: "=r" (numEq)
-		: "r" (src), "r" (stride)
-		: "%ebx"
-		);
-	numEq= (-numEq) &0xFF;
-
-#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++;
-		if(((src[4] - src[4+stride] + 1)&0xFFFF) < 3) numEq++;
-		if(((src[5] - src[5+stride] + 1)&0xFFFF) < 3) numEq++;
-		if(((src[6] - src[6+stride] + 1)&0xFFFF) < 3) numEq++;
-		if(((src[7] - src[7+stride] + 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 ", temp[x + y*stride]);
-			}
-			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;
-	src+= stride*3;
-	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
-
-	int isOk2= 1;
-	int x;
-	src+= stride*3;
-	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++)
-		{
-			for(int x=0; x<8; x++)
-			{
-				printf("%d ", src[x + y*stride]);
-			}
-			printf("\n");
-		}
-	} */
-
-	return isOk2;
-#endif
-
-}
-
-/**
- * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
- * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
- */
-static inline void doVertLowPass(uint8_t *src, int stride, int QP)
-{
-#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
-	src+= stride*3;
-	asm volatile(	//"movv %0 %1 %2\n\t"
-		"movq pQPb, %%mm0				\n\t"  // QP,..., QP
-
-		"movq (%0), %%mm6				\n\t"
-		"movq (%0, %1), %%mm5				\n\t"
-		"movq %%mm5, %%mm1				\n\t"
-		"movq %%mm6, %%mm2				\n\t"
-		"psubusb %%mm6, %%mm5				\n\t"
-		"psubusb %%mm1, %%mm2				\n\t"
-		"por %%mm5, %%mm2				\n\t" // ABS Diff of lines
-		"psubusb %%mm0, %%mm2				\n\t" // diff <= QP -> 0
-		"pcmpeqb b00, %%mm2				\n\t" // diff <= QP -> FF
-
-		"pand %%mm2, %%mm6				\n\t"
-		"pandn %%mm1, %%mm2				\n\t"
-		"por %%mm2, %%mm6				\n\t"// First Line to Filter
-
-		"movq (%0, %1, 8), %%mm5			\n\t"
-		"leal (%0, %1, 4), %%eax			\n\t"
-		"leal (%0, %1, 8), %%ebx			\n\t"
-		"subl %1, %%ebx					\n\t"
-		"addl %1, %0					\n\t" // %0 points to line 1 not 0
-		"movq (%0, %1, 8), %%mm7			\n\t"
-		"movq %%mm5, %%mm1				\n\t"
-		"movq %%mm7, %%mm2				\n\t"
-		"psubusb %%mm7, %%mm5				\n\t"
-		"psubusb %%mm1, %%mm2				\n\t"
-		"por %%mm5, %%mm2				\n\t" // ABS Diff of lines
-		"psubusb %%mm0, %%mm2				\n\t" // diff <= QP -> 0
-		"pcmpeqb b00, %%mm2				\n\t" // diff <= QP -> FF
-
-		"pand %%mm2, %%mm7				\n\t"
-		"pandn %%mm1, %%mm2				\n\t"
-		"por %%mm2, %%mm7				\n\t" // First Line to Filter
-
-
-		// 	1	2	3	4	5	6	7	8
-		//	%0	%0+%1	%0+2%1	eax	%0+4%1	eax+2%1	ebx	eax+4%1
-		// 6 4 2 2 1 1
-		// 6 4 4 2
-		// 6 8 2
-
-		"movq (%0, %1), %%mm0				\n\t" //  1
-		"movq %%mm0, %%mm1				\n\t" //  1
-		PAVGB(%%mm6, %%mm0)				      //1 1	/2
-		PAVGB(%%mm6, %%mm0)				      //3 1	/4
-
-		"movq (%0, %1, 4), %%mm2			\n\t" //     1
-		"movq %%mm2, %%mm5				\n\t" //     1
-		PAVGB((%%eax), %%mm2)				      //    11	/2
-		PAVGB((%0, %1, 2), %%mm2)			      //   211	/4
-		"movq %%mm2, %%mm3				\n\t" //   211	/4
-		"movq (%0), %%mm4				\n\t" // 1
-		PAVGB(%%mm4, %%mm3)				      // 4 211	/8
-		PAVGB(%%mm0, %%mm3)				      //642211	/16
-		"movq %%mm3, (%0)				\n\t" // X
-		// mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
-		"movq %%mm1, %%mm0				\n\t" //  1
-		PAVGB(%%mm6, %%mm0)				      //1 1	/2
-		"movq %%mm4, %%mm3				\n\t" // 1
-		PAVGB((%0,%1,2), %%mm3)				      // 1 1	/2
-		PAVGB((%%eax,%1,2), %%mm5)			      //     11	/2
-		PAVGB((%%eax), %%mm5)				      //    211 /4
-		PAVGB(%%mm5, %%mm3)				      // 2 2211 /8
-		PAVGB(%%mm0, %%mm3)				      //4242211 /16
-		"movq %%mm3, (%0,%1)				\n\t" //  X
-		// mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
-		PAVGB(%%mm4, %%mm6)				      //11	/2
-		"movq (%%ebx), %%mm0				\n\t" //       1
-		PAVGB((%%eax, %1, 2), %%mm0)			      //      11/2
-		"movq %%mm0, %%mm3				\n\t" //      11/2
-		PAVGB(%%mm1, %%mm0)				      //  2   11/4
-		PAVGB(%%mm6, %%mm0)				      //222   11/8
-		PAVGB(%%mm2, %%mm0)				      //22242211/16
-		"movq (%0, %1, 2), %%mm2			\n\t" //   1
-		"movq %%mm0, (%0, %1, 2)			\n\t" //   X
-		// mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
-		"movq (%%eax, %1, 4), %%mm0			\n\t" //        1
-		PAVGB((%%ebx), %%mm0)				      //       11	/2
-		PAVGB(%%mm0, %%mm6)				      //11     11	/4
-		PAVGB(%%mm1, %%mm4)				      // 11		/2
-		PAVGB(%%mm2, %%mm1)				      //  11		/2
-		PAVGB(%%mm1, %%mm6)				      //1122   11	/8
-		PAVGB(%%mm5, %%mm6)				      //112242211	/16
-		"movq (%%eax), %%mm5				\n\t" //    1
-		"movq %%mm6, (%%eax)				\n\t" //    X
-		// mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
-		"movq (%%eax, %1, 4), %%mm6			\n\t" //        1
-		PAVGB(%%mm7, %%mm6)				      //        11	/2
-		PAVGB(%%mm4, %%mm6)				      // 11     11	/4
-		PAVGB(%%mm3, %%mm6)				      // 11   2211	/8
-		PAVGB(%%mm5, %%mm2)				      //   11		/2
-		"movq (%0, %1, 4), %%mm4			\n\t" //     1
-		PAVGB(%%mm4, %%mm2)				      //   112		/4
-		PAVGB(%%mm2, %%mm6)				      // 112242211	/16
-		"movq %%mm6, (%0, %1, 4)			\n\t" //     X
-		// mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
-		PAVGB(%%mm7, %%mm1)				      //  11     2	/4
-		PAVGB(%%mm4, %%mm5)				      //    11		/2
-		PAVGB(%%mm5, %%mm0)				      //    11 11	/4
-		"movq (%%eax, %1, 2), %%mm6			\n\t" //      1
-		PAVGB(%%mm6, %%mm1)				      //  11  4  2	/8
-		PAVGB(%%mm0, %%mm1)				      //  11224222	/16
-		"movq %%mm1, (%%eax, %1, 2)			\n\t" //      X
-		// mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
-		PAVGB((%%ebx), %%mm2)				      //   112 4	/8
-		"movq (%%eax, %1, 4), %%mm0			\n\t" //        1
-		PAVGB(%%mm0, %%mm6)				      //      1 1	/2
-		PAVGB(%%mm7, %%mm6)				      //      1 12	/4
-		PAVGB(%%mm2, %%mm6)				      //   1122424	/4
-		"movq %%mm6, (%%ebx)				\n\t" //       X
-		// mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
-		PAVGB(%%mm7, %%mm5)				      //    11   2	/4
-		PAVGB(%%mm7, %%mm5)				      //    11   6	/8
-
-		PAVGB(%%mm3, %%mm0)				      //      112	/4
-		PAVGB(%%mm0, %%mm5)				      //    112246	/16
-		"movq %%mm5, (%%eax, %1, 4)			\n\t" //        X
-		"subl %1, %0					\n\t"
-
-		:
-		: "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;
-	src+= stride*3;
-	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];
-		sums[0] = first + src[l1];
-		sums[1] = src[l1] + src[l2];
-		sums[2] = src[l2] + src[l3];
-		sums[3] = src[l3] + src[l4];
-		sums[4] = src[l4] + src[l5];
-		sums[5] = src[l5] + src[l6];
-		sums[6] = src[l6] + src[l7];
-		sums[7] = src[l7] + src[l8];
-		sums[8] = src[l8] + last;
-
-		src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
-		src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
-		src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
-		src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
-		src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
-		src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
-		src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
-		src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
-
-		src++;
-	}
-
-#endif
-}
-
-/**
- * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
- * values are correctly clipped (MMX2)
- * values are wraparound (C)
- * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient
-	0 8 16 24
-	x = 8
-	x/2 = 4
-	x/8 = 1
-	1 12 12 23
- */
-static inline void vertRK1Filter(uint8_t *src, int stride, int QP)
-{
-#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
-	src+= stride*3;
-// FIXME rounding
-	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 pQPb, %%mm0				\n\t" // QP,..., QP
-		"movq %%mm0, %%mm1				\n\t" // QP,..., QP
-		"paddusb b02, %%mm0				\n\t"
-		"psrlw $2, %%mm0				\n\t"
-		"pand b3F, %%mm0				\n\t" // QP/4,..., QP/4
-		"paddusb %%mm1, %%mm0				\n\t" // QP*1.25 ...
-		"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"
-		"pcmpeqb %%mm7, %%mm4				\n\t"
-		"pand %%mm4, %%mm5				\n\t" // d/2
-
-//		"paddb %%mm6, %%mm2				\n\t" // line 4 + 0x80
-		"paddb %%mm5, %%mm2				\n\t"
-//		"psubb %%mm6, %%mm2				\n\t"
-		"movq %%mm2, (%0,%1, 4)				\n\t"
-
-		"movq (%%ebx), %%mm2				\n\t"
-//		"paddb %%mm6, %%mm2				\n\t" // line 5 + 0x80
-		"psubb %%mm5, %%mm2				\n\t"
-//		"psubb %%mm6, %%mm2				\n\t"
-		"movq %%mm2, (%%ebx)				\n\t"
-
-		"paddb %%mm6, %%mm5				\n\t"
-		"psrlw $2, %%mm5				\n\t"
-		"pand b3F, %%mm5				\n\t"
-		"psubb b20, %%mm5				\n\t" // (l5-l4)/8
-
-		"movq (%%eax, %1, 2), %%mm2			\n\t"
-		"paddb %%mm6, %%mm2				\n\t" // line 3 + 0x80
-		"paddsb %%mm5, %%mm2				\n\t"
-		"psubb %%mm6, %%mm2				\n\t"
-		"movq %%mm2, (%%eax, %1, 2)			\n\t"
-
-		"movq (%%ebx, %1), %%mm2			\n\t"
-		"paddb %%mm6, %%mm2				\n\t" // line 6 + 0x80
-		"psubsb %%mm5, %%mm2				\n\t"
-		"psubb %%mm6, %%mm2				\n\t"
-		"movq %%mm2, (%%ebx, %1)			\n\t"
-
-		:
-		: "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;
-	const int QP15= QP + (QP>>2);
-	src+= stride*3;
-	for(x=0; x<BLOCK_SIZE; x++)
-	{
-		const int v = (src[x+l5] - src[x+l4]);
-		if(ABS(v) < QP15)
-		{
-			src[x+l3] +=v>>3;
-			src[x+l4] +=v>>1;
-			src[x+l5] -=v>>1;
-			src[x+l6] -=v>>3;
-
-		}
-	}
-
-#endif
-}
-
-/**
- * Experimental Filter 1
- * will not damage linear gradients
- * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
- * 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)
-	src+= stride*3;
-
-	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
-		"psubusb b01, %%mm3				\n\t"
-		"pand %%mm4, %%mm3				\n\t" // d <= QP ? d : 0
-
-		PAVGB(%%mm7, %%mm3)				      // d/2
-		"movq %%mm3, %%mm1				\n\t" // d/2
-		PAVGB(%%mm7, %%mm3)				      // d/4
-		PAVGB(%%mm1, %%mm3)				      // 3*d/8
-
-		"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, %%mm1)				      // d/4
-
-		"movq (%%eax, %1, 2), %%mm0			\n\t" // line 3
-		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
-		"psubusb %%mm1, %%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 %%mm1, %%mm0				\n\t"
-		"pxor %%mm2, %%mm0				\n\t"
-		"movq %%mm0, (%%ebx, %1)			\n\t" // line 6
-
-		PAVGB(%%mm7, %%mm1)				      // d/8
-
-		"movq (%%eax, %1), %%mm0			\n\t" // line 2
-		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
-		"psubusb %%mm1, %%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 %%mm1, %%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;
-
-	src+= stride*3;
-	for(x=0; x<BLOCK_SIZE; x++)
-	{
-		int a= src[l3] - src[l4];
-		int b= src[l4] - src[l5];
-		int c= src[l5] - src[l6];
-
-		int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
-		d= MAX(d, 0);
-
-		if(d < QP)
-		{
-			int v = d * SIGN(-b);
-
-			src[l2] +=v>>3;
-			src[l3] +=v>>2;
-			src[l4] +=(3*v)>>3;
-			src[l5] -=(3*v)>>3;
-			src[l6] -=v>>2;
-			src[l7] -=v>>3;
-
-		}
-		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;
-	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++)
-	{
-		int v2= src[l2];
-		int v3= src[l3];
-		int v4= src[l4];
-		int v5= src[l5];
-		int v6= src[l6];
-		int v7= src[l7];
-
-		if(ABS(v4-v5)<QP &&  ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 )
-		{
-			src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6         )/16;
-			src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7  )/16;
-			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
-}
-
-/**
- * Experimental Filter 1 (Horizontal)
- * will not damage linear gradients
- * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
- * can only smooth blocks at the expected locations (it cant smooth them if they did move)
- * MMX2 version does correct clipping C version doesnt
- * not identical with the vertical one
- */
-static inline void horizX1Filter(uint8_t *src, int stride, int QP)
-{
-	int y;
-//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;
-	}
-}
-
-
-static inline void doVertDefFilter(uint8_t src[], int stride, int QP)
-{
-#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
-/*
-	uint8_t tmp[16];
-	const int l1= stride;
-	const int l2= stride + l1;
-	const int l3= stride + l2;
-	const int l4= (int)tmp - (int)src - stride*3;
-	const int l5= (int)tmp - (int)src - stride*3 + 8;
-	const int l6= stride*3 + l3;
-	const int l7= stride + l6;
-	const int l8= stride + l7;
-
-	memcpy(tmp, src+stride*7, 8);
-	memcpy(tmp+8, src+stride*8, 8);
-*/
-	src+= stride*4;
-	asm volatile(
-
-#if 0 //sligtly more accurate and slightly slower
-		"pxor %%mm7, %%mm7				\n\t" // 0
-		"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, %1, 2), %%mm0			\n\t" // l2
-		"movq (%0), %%mm1				\n\t" // l0
-		"movq %%mm0, %%mm2				\n\t" // l2
-		PAVGB(%%mm7, %%mm0)				      // ~l2/2
-		PAVGB(%%mm1, %%mm0)				      // ~(l2 + 2l0)/4
-		PAVGB(%%mm2, %%mm0)				      // ~(5l2 + 2l0)/8
-
-		"movq (%%eax), %%mm1				\n\t" // l1
-		"movq (%%eax, %1, 2), %%mm3			\n\t" // l3
-		"movq %%mm1, %%mm4				\n\t" // l1
-		PAVGB(%%mm7, %%mm1)				      // ~l1/2
-		PAVGB(%%mm3, %%mm1)				      // ~(l1 + 2l3)/4
-		PAVGB(%%mm4, %%mm1)				      // ~(5l1 + 2l3)/8
-
-		"movq %%mm0, %%mm4				\n\t" // ~(5l2 + 2l0)/8
-		"psubusb %%mm1, %%mm0				\n\t"
-		"psubusb %%mm4, %%mm1				\n\t"
-		"por %%mm0, %%mm1				\n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
-// mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
-
-		"movq (%0, %1, 4), %%mm0			\n\t" // l4
-		"movq %%mm0, %%mm4				\n\t" // l4
-		PAVGB(%%mm7, %%mm0)				      // ~l4/2
-		PAVGB(%%mm2, %%mm0)				      // ~(l4 + 2l2)/4
-		PAVGB(%%mm4, %%mm0)				      // ~(5l4 + 2l2)/8
-
-		"movq (%%ebx), %%mm2				\n\t" // l5
-		"movq %%mm3, %%mm5				\n\t" // l3
-		PAVGB(%%mm7, %%mm3)				      // ~l3/2
-		PAVGB(%%mm2, %%mm3)				      // ~(l3 + 2l5)/4
-		PAVGB(%%mm5, %%mm3)				      // ~(5l3 + 2l5)/8
-
-		"movq %%mm0, %%mm6				\n\t" // ~(5l4 + 2l2)/8
-		"psubusb %%mm3, %%mm0				\n\t"
-		"psubusb %%mm6, %%mm3				\n\t"
-		"por %%mm0, %%mm3				\n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
-		"pcmpeqb %%mm7, %%mm0				\n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
-// mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
-
-		"movq (%%ebx, %1), %%mm6			\n\t" // l6
-		"movq %%mm6, %%mm5				\n\t" // l6
-		PAVGB(%%mm7, %%mm6)				      // ~l6/2
-		PAVGB(%%mm4, %%mm6)				      // ~(l6 + 2l4)/4
-		PAVGB(%%mm5, %%mm6)				      // ~(5l6 + 2l4)/8
-
-		"movq (%%ebx, %1, 2), %%mm5			\n\t" // l7
-		"movq %%mm2, %%mm4				\n\t" // l5
-		PAVGB(%%mm7, %%mm2)				      // ~l5/2
-		PAVGB(%%mm5, %%mm2)				      // ~(l5 + 2l7)/4
-		PAVGB(%%mm4, %%mm2)				      // ~(5l5 + 2l7)/8
-
-		"movq %%mm6, %%mm4				\n\t" // ~(5l6 + 2l4)/8
-		"psubusb %%mm2, %%mm6				\n\t"
-		"psubusb %%mm4, %%mm2				\n\t"
-		"por %%mm6, %%mm2				\n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
-// mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
-
-
-		PMINUB(%%mm2, %%mm1, %%mm4)			      // MIN(|lenergy|,|renergy|)/8
-		"movq pQPb, %%mm4				\n\t" // QP //FIXME QP+1 ?
-		"paddusb b01, %%mm4				\n\t"
-		"pcmpgtb %%mm3, %%mm4				\n\t" // |menergy|/8 < QP
-		"psubusb %%mm1, %%mm3				\n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
-		"pand %%mm4, %%mm3				\n\t"
-
-		"movq %%mm3, %%mm1				\n\t"
-//		"psubusb b01, %%mm3				\n\t"
-		PAVGB(%%mm7, %%mm3)
-		PAVGB(%%mm7, %%mm3)
-		"paddusb %%mm1, %%mm3				\n\t"
-//		"paddusb b01, %%mm3				\n\t"
-
-		"movq (%%eax, %1, 2), %%mm6			\n\t" //l3
-		"movq (%0, %1, 4), %%mm5			\n\t" //l4
-		"movq (%0, %1, 4), %%mm4			\n\t" //l4
-		"psubusb %%mm6, %%mm5				\n\t"
-		"psubusb %%mm4, %%mm6				\n\t"
-		"por %%mm6, %%mm5				\n\t" // |l3-l4|
-		"pcmpeqb %%mm7, %%mm6				\n\t" // SIGN(l3-l4)
-		"pxor %%mm6, %%mm0				\n\t"
-		"pand %%mm0, %%mm3				\n\t"
-		PMINUB(%%mm5, %%mm3, %%mm0)
-
-		"psubusb b01, %%mm3				\n\t"
-		PAVGB(%%mm7, %%mm3)
-
-		"movq (%%eax, %1, 2), %%mm0			\n\t"
-		"movq (%0, %1, 4), %%mm2			\n\t"
-		"pxor %%mm6, %%mm0				\n\t"
-		"pxor %%mm6, %%mm2				\n\t"
-		"psubb %%mm3, %%mm0				\n\t"
-		"paddb %%mm3, %%mm2				\n\t"
-		"pxor %%mm6, %%mm0				\n\t"
-		"pxor %%mm6, %%mm2				\n\t"
-		"movq %%mm0, (%%eax, %1, 2)			\n\t"
-		"movq %%mm2, (%0, %1, 4)			\n\t"
-#endif
-
-		"leal (%0, %1), %%eax				\n\t"
-		"pcmpeqb %%mm6, %%mm6				\n\t" // -1
-//	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 (%%eax, %1, 2), %%mm1			\n\t" // l3
-		"movq (%0, %1, 4), %%mm0			\n\t" // l4
-		"pxor %%mm6, %%mm1				\n\t" // -l3-1
-		PAVGB(%%mm1, %%mm0)				      // -q+128 = (l4-l3+256)/2
-// mm1=-l3-1, mm0=128-q
-
-		"movq (%%eax, %1, 4), %%mm2			\n\t" // l5
-		"movq (%%eax, %1), %%mm3			\n\t" // l2
-		"pxor %%mm6, %%mm2				\n\t" // -l5-1
-		"movq %%mm2, %%mm5				\n\t" // -l5-1
-		"movq b80, %%mm4				\n\t" // 128
-		"leal (%%eax, %1, 4), %%ebx			\n\t"
-		PAVGB(%%mm3, %%mm2)				      // (l2-l5+256)/2
-		PAVGB(%%mm0, %%mm4)				      // ~(l4-l3)/4 + 128
-		PAVGB(%%mm2, %%mm4)				      // ~(l2-l5)/4 +(l4-l3)/8 + 128
-		PAVGB(%%mm0, %%mm4)				      // ~(l2-l5)/8 +5(l4-l3)/16 + 128
-// mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
-
-		"movq (%%eax), %%mm2				\n\t" // l1
-		"pxor %%mm6, %%mm2				\n\t" // -l1-1
-		PAVGB(%%mm3, %%mm2)				      // (l2-l1+256)/2
-		PAVGB((%0), %%mm1)				      // (l0-l3+256)/2
-		"movq b80, %%mm3				\n\t" // 128
-		PAVGB(%%mm2, %%mm3)				      // ~(l2-l1)/4 + 128
-		PAVGB(%%mm1, %%mm3)				      // ~(l0-l3)/4 +(l2-l1)/8 + 128
-		PAVGB(%%mm2, %%mm3)				      // ~(l0-l3)/8 +5(l2-l1)/16 + 128
-// mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
-
-		PAVGB((%%ebx, %1), %%mm5)			      // (l6-l5+256)/2
-		"movq (%%ebx, %1, 2), %%mm1			\n\t" // l7
-		"pxor %%mm6, %%mm1				\n\t" // -l7-1
-		PAVGB((%0, %1, 4), %%mm1)			      // (l4-l7+256)/2
-		"movq b80, %%mm2				\n\t" // 128
-		PAVGB(%%mm5, %%mm2)				      // ~(l6-l5)/4 + 128
-		PAVGB(%%mm1, %%mm2)				      // ~(l4-l7)/4 +(l6-l5)/8 + 128
-		PAVGB(%%mm5, %%mm2)				      // ~(l4-l7)/8 +5(l6-l5)/16 + 128
-// mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
-
-		"movq b00, %%mm1				\n\t" // 0
-		"movq b00, %%mm5				\n\t" // 0
-		"psubb %%mm2, %%mm1				\n\t" // 128 - renergy/16
-		"psubb %%mm3, %%mm5				\n\t" // 128 - lenergy/16
-		PMAXUB(%%mm1, %%mm2)				      // 128 + |renergy/16|
- 		PMAXUB(%%mm5, %%mm3)				      // 128 + |lenergy/16|
-		PMINUB(%%mm2, %%mm3, %%mm1)			      // 128 + MIN(|lenergy|,|renergy|)/16
-
-// mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
-
-		"movq b00, %%mm7				\n\t" // 0
-		"movq pQPb, %%mm2				\n\t" // QP
-		PAVGB(%%mm6, %%mm2)				      // 128 + QP/2
-		"psubb %%mm6, %%mm2				\n\t"
-
-		"movq %%mm4, %%mm1				\n\t"
-		"pcmpgtb %%mm7, %%mm1				\n\t" // SIGN(menergy)
-		"pxor %%mm1, %%mm4				\n\t"
-		"psubb %%mm1, %%mm4				\n\t" // 128 + |menergy|/16
-		"pcmpgtb %%mm4, %%mm2				\n\t" // |menergy|/16 < QP/2
-		"psubusb %%mm3, %%mm4				\n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
-// mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
-
-		"movq %%mm4, %%mm3				\n\t" // d
-		"psubusb b01, %%mm4				\n\t"
-		PAVGB(%%mm7, %%mm4)				      // d/32
-		PAVGB(%%mm7, %%mm4)				      // (d + 32)/64
-		"paddb %%mm3, %%mm4				\n\t" // 5d/64
-		"pand %%mm2, %%mm4				\n\t"
-
-		"movq b80, %%mm5				\n\t" // 128
-		"psubb %%mm0, %%mm5				\n\t" // q
-		"paddsb %%mm6, %%mm5				\n\t" // fix bad rounding
-		"pcmpgtb %%mm5, %%mm7				\n\t" // SIGN(q)
-		"pxor %%mm7, %%mm5				\n\t"
-
-		PMINUB(%%mm5, %%mm4, %%mm3)			      // MIN(|q|, 5d/64)
-		"pxor %%mm1, %%mm7				\n\t" // SIGN(d*q)
-
-		"pand %%mm7, %%mm4				\n\t"
-		"movq (%%eax, %1, 2), %%mm0			\n\t"
-		"movq (%0, %1, 4), %%mm2			\n\t"
-		"pxor %%mm1, %%mm0				\n\t"
-		"pxor %%mm1, %%mm2				\n\t"
-		"paddb %%mm4, %%mm0				\n\t"
-		"psubb %%mm4, %%mm2				\n\t"
-		"pxor %%mm1, %%mm0				\n\t"
-		"pxor %%mm1, %%mm2				\n\t"
-		"movq %%mm0, (%%eax, %1, 2)			\n\t"
-		"movq %%mm2, (%0, %1, 4)			\n\t"
-
-		:
-		: "r" (src), "r" (stride)
-		: "%eax", "%ebx"
-	);
-
-/*
-	{
-	int x;
-	src-= stride;
-	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++;
-	}
-src-=8;
-	for(x=0; x<8; x++)
-	{
-		int y;
-		for(y=4; y<6; y++)
-		{
-			int d= src[x+y*stride] - tmp[x+(y-4)*8];
-			int ad= ABS(d);
-			static int max=0;
-			static int sum=0;
-			static int num=0;
-			static int bias=0;
-
-			if(max<ad) max=ad;
-			sum+= ad>3 ? 1 : 0;
-			if(ad>3)
-			{
-				src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
-			}
-			if(y==4) bias+=d;
-			num++;
-			if(num%1000000 == 0)
-			{
-				printf(" %d %d %d %d\n", num, sum, max, bias);
-			}
-		}
-	}
-}
-*/
-#elif defined (HAVE_MMX)
-	src+= stride*4;
-
-	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
-