postprocess.c

/*
    Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at)

    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation; either version 2 of the License, or
    (at your option) any later version.

    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.

    You should have received a copy of the GNU General Public License
    along with this program; if not, write to the Free Software
    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
*/

/*
			C	MMX	MMX2	3DNow*
isVertDC		Ec	Ec
isVertMinMaxOk		Ec	Ec
doVertLowPass		E		e	e*
doVertDefFilter		Ec	Ec	Ec
isHorizDC		Ec	Ec
isHorizMinMaxOk		a
doHorizLowPass		E		a	a*
doHorizDefFilter	E	ac	ac
deRing
Vertical RKAlgo1	E		a	a*
Vertical X1		a		E	E*
Horizontal X1		a		E	E*


* i dont have a 3dnow CPU -> its untested
E = Exact implementation
e = allmost exact implementation
a = alternative / approximate impl
c = checked against the other implementations (-vo md5)
*/

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

Notes:

*/

/*
Changelog: use the CVS log
0.1.3
	bugfixes: last 3 lines not brightness/contrast corrected
		brightness statistics messed up with initial black pic
	changed initial values of the brightness statistics
	C++ -> C conversation
	QP range question solved (very likely 1<=QP<=32 according to arpi)
	new experimental vertical deblocking filter
	RK filter has 3dNow support now (untested)
0.1.2
	fixed a bug in the horizontal default filter
	3dnow version of the Horizontal & Vertical Lowpass filters
	mmx version of the Horizontal Default filter
	mmx2 & C versions of a simple filter described in a paper from ramkishor & karandikar
	added mode flags & quality2mode function
0.1.1
*/


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


static uint64_t packedYOffset=	0x0000000000000000LL;
static uint64_t packedYScale=	0x0100010001000100LL;
static uint64_t w05=		0x0005000500050005LL;
static uint64_t w20=		0x0020002000200020LL;
static uint64_t w1400=		0x1400140014001400LL;
static uint64_t bm00000001=	0x00000000000000FFLL;
static uint64_t bm00010000=	0x000000FF00000000LL;
static uint64_t bm00001000=	0x00000000FF000000LL;
static uint64_t bm10000000=	0xFF00000000000000LL;
static uint64_t bm10000001=	0xFF000000000000FFLL;
static uint64_t bm11000011=	0xFFFF00000000FFFFLL;
static uint64_t bm00000011=	0x000000000000FFFFLL;
static uint64_t bm11111110=	0xFFFFFFFFFFFFFF00LL;
static uint64_t bm11000000=	0xFFFF000000000000LL;
static uint64_t bm00011000=	0x000000FFFF000000LL;
static uint64_t bm00110011=	0x0000FFFF0000FFFFLL;
static uint64_t bm11001100=	0xFFFF0000FFFF0000LL;
static uint64_t b00= 		0x0000000000000000LL;
static uint64_t b01= 		0x0101010101010101LL;
static uint64_t b02= 		0x0202020202020202LL;
static uint64_t b0F= 		0x0F0F0F0F0F0F0F0FLL;
static uint64_t bFF= 		0xFFFFFFFFFFFFFFFFLL;
static uint64_t b20= 		0x2020202020202020LL;
static uint64_t b80= 		0x8080808080808080LL;
static uint64_t b7E= 		0x7E7E7E7E7E7E7E7ELL;
static uint64_t b7C= 		0x7C7C7C7C7C7C7C7CLL;
static uint64_t b3F= 		0x3F3F3F3F3F3F3F3FLL;
static uint64_t temp0=0;
static uint64_t temp1=0;
static uint64_t temp2=0;
static uint64_t temp3=0;
static uint64_t temp4=0;
static uint64_t temp5=0;
static uint64_t pQPb=0;
static uint8_t tempBlock[16*16];

int hFlatnessThreshold= 56 - 16;
int vFlatnessThreshold= 56 - 16;

//amount of "black" u r willing to loose to get a brightness corrected picture
double maxClippedThreshold= 0.01;

int maxAllowedY=255;
//FIXME can never make a movies black brighter (anyone needs that?)
int minAllowedY=0;


static inline long long rdtsc()
{
	long long l;
	asm volatile(	"rdtsc\n\t"
		: "=A" (l)
	);
//	printf("%d\n", int(l/1000));
	return l;
}

static inline void prefetchnta(void *p)
{
	asm volatile(	"prefetchnta (%0)\n\t"
		: : "r" (p)
	);
}

static inline void prefetcht0(void *p)
{
	asm volatile(	"prefetcht0 (%0)\n\t"
		: : "r" (p)
	);
}

static inline void prefetcht1(void *p)
{
	asm volatile(	"prefetcht1 (%0)\n\t"
		: : "r" (p)
	);
}

static inline void prefetcht2(void *p)
{
	asm volatile(	"prefetcht2 (%0)\n\t"
		: : "r" (p)
	);
}

//FIXME? |255-0| = 1 (shouldnt be a problem ...)
/**
 * Check if the middle 8x8 Block in the given 8x10 block is flat
 */
static inline int isVertDC(uint8_t src[], int stride){
//	return true;
	int numEq= 0;
	int y;
	src+= stride; // src points to begin of the 8x8 Block
#ifdef HAVE_MMX
	asm volatile(
//		"int $3 \n\t"
		"pushl %1\n\t"
		"movq b7E, %%mm7					\n\t" // mm7 = 0x7F
		"movq b7C, %%mm6					\n\t" // mm6 = 0x7D
		"movq (%1), %%mm0				\n\t"
		"addl %2, %1					\n\t"
		"movq (%1), %%mm1				\n\t"
		"psubb %%mm1, %%mm0				\n\t" // mm0 = differnece
		"paddb %%mm7, %%mm0				\n\t"
		"pcmpgtb %%mm6, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%mm2				\n\t"
		"psubb %%mm2, %%mm1				\n\t"
		"paddb %%mm7, %%mm1				\n\t"
		"pcmpgtb %%mm6, %%mm1				\n\t"
		"paddb %%mm1, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%mm1				\n\t"
		"psubb %%mm1, %%mm2				\n\t"
		"paddb %%mm7, %%mm2				\n\t"
		"pcmpgtb %%mm6, %%mm2				\n\t"
		"paddb %%mm2, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%mm2				\n\t"
		"psubb %%mm2, %%mm1				\n\t"
		"paddb %%mm7, %%mm1				\n\t"
		"pcmpgtb %%mm6, %%mm1				\n\t"
		"paddb %%mm1, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%mm1				\n\t"
		"psubb %%mm1, %%mm2				\n\t"
		"paddb %%mm7, %%mm2				\n\t"
		"pcmpgtb %%mm6, %%mm2				\n\t"
		"paddb %%mm2, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%mm2				\n\t"
		"psubb %%mm2, %%mm1				\n\t"
		"paddb %%mm7, %%mm1				\n\t"
		"pcmpgtb %%mm6, %%mm1				\n\t"
		"paddb %%mm1, %%mm0				\n\t"

		"addl %2, %1					\n\t"
		"movq (%1), %%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"
		"movq %%mm0, %%mm1				\n\t"
		"psrlw $8, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"psrlq $16, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"movq %%mm0, %%mm1				\n\t"
		"psrlq $32, %%mm0				\n\t"
		"paddb %%mm1, %%mm0				\n\t"
		"popl %1\n\t"
		"movd %%mm0, %0					\n\t"
		: "=r" (numEq)
		: "r" (src), "r" (stride)
		);
//	printf("%d\n", numEq);
	numEq= (256 - (numEq & 0xFF)) &0xFF;

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

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

	int isOk2= 1;
	int x;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0;
	}
/*	if(isOk && !isOk2 || !isOk && isOk2)
	{
		printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);
		for(int y=0; y<9; y++)
		{
			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 8x10 block (only write to the 8x8 block in the middle)
 * useing 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)
{
//	QP= 64;

#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
//#ifdef HAVE_MMX2
	asm volatile(	//"movv %0 %1 %2\n\t"
		"pushl %0 \n\t"
		"movq pQPb, %%mm0				\n\t"  // QP,..., QP
//		"movq bFF  , %%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 %%mm6, %%mm2				\n\t" //1
		"movq %%mm6, %%mm3				\n\t" //1
		"paddusb b02, %%mm3 				\n\t"
		"psrlw $2, %%mm3				\n\t" //1	/4
		"pand b3F, %%mm3				\n\t"
		"psubb %%mm3, %%mm2				\n\t"
		"movq (%0, %1), %%mm0				\n\t" //  1
		"movq %%mm0, %%mm1				\n\t" //  1
		"paddusb b02, %%mm0				\n\t"
		"psrlw $2, %%mm0				\n\t" //  1	/4
		"pand b3F, %%mm0				\n\t"
		"paddusb %%mm2, %%mm0				\n\t" //3 1	/4
*/
		"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
//		"pxor %%mm1, %%mm1 \n\t"
		"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
//		"pxor %%mm6, %%mm6 \n\t"
		"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
//		"pxor %%mm5, %%mm5 \n\t"
//		"movq pQPb, %%mm5 \n\t"
		"movq %%mm5, (%%eax, %1, 4)			\n\t" //        X
		"popl %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;
	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)
// 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;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		if(ABS(src[l4]-src[l5]) < QP + QP/4)
		{
			int v = (src[l5] - src[l4]);

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

		}
		src++;
	}

#endif
}

/**
 * Experimental Filter 1
 * will 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)
	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;
	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= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);

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

			src[l2] +=v/8;
			src[l3] +=v/4;
			src[l4] +=3*v/8;
			src[l5] -=3*v/8;
			src[l6] -=v/4;
			src[l7] -=v/8;

		}
		src++;
	}
	/*
 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;
	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;
	static uint64_t *lut= NULL;
	if(lut==NULL)
	{
		int i;
		lut= (uint64_t*)memalign(8, 256*8);
		for(i=0; i<256; i++)
		{
			int v= i < 128 ? 2*i : 2*(i-256);
/*
//Simulate 112242211 9-Tap filter
			uint64_t a= (v/16) & 0xFF;
			uint64_t b= (v/8) & 0xFF;
			uint64_t c= (v/4) & 0xFF;
			uint64_t d= (3*v/8) & 0xFF;
*/
//Simulate piecewise linear interpolation
			uint64_t a= (v/16) & 0xFF;
			uint64_t b= (v*3/16) & 0xFF;
			uint64_t c= (v*5/16) & 0xFF;
			uint64_t d= (7*v/16) & 0xFF;
			uint64_t A= (0x100 - a)&0xFF;
			uint64_t B= (0x100 - b)&0xFF;
			uint64_t C= (0x100 - c)&0xFF;
			uint64_t D= (0x100 - c)&0xFF;

			lut[i]   = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
				(D<<24) | (C<<16) | (B<<8) | (A);
			//lut[i] = (v<<32) | (v<<24);
		}
	}

#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"pxor %%mm7, %%mm7				\n\t" // 0
//		"movq b80, %%mm6				\n\t" // MIN_SIGNED_BYTE
		"leal (%0, %1), %%eax				\n\t"
		"leal (%%eax, %1, 4), %%ebx			\n\t"

		"movq b80, %%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
		"pxor %%mm5, %%mm5				\n\t" // 0
		"psubb %%mm4, %%mm5				\n\t" // -3QP
		"por bm11111110, %%mm5				\n\t" // ...,FF,FF,-3QP
		"psllq $24, %%mm5				\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

#define HX1old(a) \
		"movd " #a ", %%mm0				\n\t"\
		"movd 4" #a ", %%mm1				\n\t"\
		"punpckldq %%mm1, %%mm0				\n\t"\
		"movq %%mm0, %%mm1				\n\t"\
		"movq %%mm0, %%mm2				\n\t"\
		"psrlq $8, %%mm1				\n\t"\
		"psubusb %%mm1, %%mm2				\n\t"\
		"psubusb %%mm0, %%mm1				\n\t"\
		"por %%mm2, %%mm1				\n\t" /* px = |px - p(x+1)| */\
		"pcmpeqb %%mm7, %%mm2				\n\t" /* px = sgn[px - p(x+1)] */\
		"pshufw $0x00, %%mm1, %%mm3			\n\t" /* p5 = |p1 - p2| */\
		PAVGB(%%mm1, %%mm3)				      /* p5 = (|p2-p1| + |p6-p5|)/2 */\
		"psrlq $16, %%mm3				\n\t" /* p3 = (|p2-p1| + |p6-p5|)/2 */\
		"psubusb %%mm3, %%mm1			\n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
		"paddb %%mm5, %%mm1				\n\t"\
		"psubusb %%mm5, %%mm1				\n\t"\
		PAVGB(%%mm7, %%mm1)\
		"pxor %%mm2, %%mm1				\n\t"\
		"psubb %%mm2, %%mm1				\n\t"\
		"psrlq $24, %%mm1				\n\t"\
		"movd %%mm1, %%ecx				\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"paddsb (%3, %%ecx, 8), %%mm0			\n\t"\
		"paddb %%mm6, %%mm0				\n\t"\
		"movq %%mm0, " #a "				\n\t"\

/*
HX1old((%0))
HX1old((%%eax))
HX1old((%%eax, %1))
HX1old((%%eax, %1, 2))
HX1old((%0, %1, 4))
HX1old((%%ebx))
HX1old((%%ebx, %1))
HX1old((%%ebx, %1, 2))
*/

//FIXME add some comments, its unreadable ...
#define HX1b(a, c, b, d) \
		"movd " #a ", %%mm0				\n\t"\
		"movd 4" #a ", %%mm1				\n\t"\
		"punpckldq %%mm1, %%mm0				\n\t"\
		"movd " #b ", %%mm4				\n\t"\
		"movq %%mm0, %%mm1				\n\t"\
		"movq %%mm0, %%mm2				\n\t"\
		"psrlq $8, %%mm1				\n\t"\
		"movd 4" #b ", %%mm3				\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)] */\
		"punpckldq %%mm3, %%mm4				\n\t"\
		"movq %%mm1, %%mm3				\n\t"\
		"psllq $32, %%mm3				\n\t" /* p5 = |p1 - p2| */\
		PAVGB(%%mm1, %%mm3)				      /* p5 = (|p2-p1| + |p6-p5|)/2 */\
		"paddb %%mm6, %%mm0				\n\t"\
		"psrlq $16, %%mm3				\n\t" /* p3 = (|p2-p1| + |p6-p5|)/2 */\
		"psubusb %%mm3, %%mm1			\n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
		"movq %%mm4, %%mm3				\n\t"\
		"paddb %%mm5, %%mm1				\n\t"\
		"psubusb %%mm5, %%mm1				\n\t"\
		"psrlq $8, %%mm3				\n\t"\
		PAVGB(%%mm7, %%mm1)\
		"pxor %%mm2, %%mm1				\n\t"\
		"psubb %%mm2, %%mm1				\n\t"\
		"movq %%mm4, %%mm2				\n\t"\
		"psrlq $24, %%mm1				\n\t"\
		"psubusb %%mm3, %%mm2				\n\t"\
		"movd %%mm1, %%ecx				\n\t"\
		"psubusb %%mm4, %%mm3				\n\t"\
		"paddsb (%3, %%ecx, 8), %%mm0			\n\t"\
		"por %%mm2, %%mm3				\n\t" /* px = |px - p(x+1)| */\
		"paddb %%mm6, %%mm0				\n\t"\
		"pcmpeqb %%mm7, %%mm2				\n\t" /* px = sgn[px - p(x+1)] */\
		"movq %%mm3, %%mm1				\n\t"\
		"psllq $32, %%mm1				\n\t" /* p5 = |p1 - p2| */\
		"movq %%mm0, " #a "				\n\t"\
		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 (%3, %%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 (%3, %%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 (%3, %%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" (QP), "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;