/*
- Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at)
+ Copyright (C) 2001-2002 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
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
-LinIpolDeinterlace e E E*
-CubicIpolDeinterlace a e e*
-LinBlendDeinterlace e E E*
-MedianDeinterlace Ec Ec
-
-
-* i dont have a 3dnow CPU -> its untested
-E = Exact implementation
-e = allmost exact implementation (slightly different rounding,...)
-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
-split this huge file
-fix warnings (unused vars, ...)
-noise reduction filters
-...
-
-Notes:
-
-*/
-
-//Changelog: use the CVS log
-
-#include <inttypes.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include "../config.h"
-//#undef HAVE_MMX2
-//#define HAVE_3DNOW
-//#undef HAVE_MMX
-#include "postprocess.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)
+#undef PAVGB
+#undef PMINUB
+#undef PMAXUB
#ifdef HAVE_MMX2
#define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
#define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
#endif
-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]; //used so the horizontal code gets aligned data
-
-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 movieĀ“s black brighter (anyone needs that?)
-int minAllowedY=16;
-
-#ifdef TIMING
-static inline long long rdtsc()
-{
- long long l;
- asm volatile( "rdtsc\n\t"
- : "=A" (l)
- );
-// printf("%d\n", int(l/1000));
- return l;
-}
+#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
-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)
- );
-}
+#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
+
//FIXME? |255-0| = 1 (shouldnt be a problem ...)
+#ifdef HAVE_MMX
/**
* Check if the middle 8x8 Block in the given 8x16 block is flat
*/
-static inline int isVertDC(uint8_t src[], int stride){
+static inline int RENAME(isVertDC)(uint8_t src[], int stride, PPContext *c){
int numEq= 0;
- int y;
src+= stride*4; // src points to begin of the 8x8 Block
-#ifdef HAVE_MMX
- asm volatile(
- "pushl %1\n\t"
- "movq b7E, %%mm7 \n\t" // mm7 = 0x7F
- "movq b7C, %%mm6 \n\t" // mm6 = 0x7D
+asm volatile(
+ "leal (%1, %2), %%eax \n\t"
+// 0 1 2 3 4 5 6 7 8 9
+// %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2
+ "movq %3, %%mm7 \n\t"
+ "movq %4, %%mm6 \n\t"
+
"movq (%1), %%mm0 \n\t"
- "addl %2, %1 \n\t"
- "movq (%1), %%mm1 \n\t"
+ "movq (%%eax), %%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"
+ "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"
- "addl %2, %1 \n\t"
- "movq (%1), %%mm1 \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"
+
+ "leal (%%eax, %2, 4), %%eax \n\t"
- "addl %2, %1 \n\t"
- "movq (%1), %%mm2 \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"
- "addl %2, %1 \n\t"
- "movq (%1), %%mm1 \n\t"
+ "movq (%%eax), %%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"
+ "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"
- "addl %2, %1 \n\t"
- "movq (%1), %%mm1 \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"
" \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 $32, %%mm0 \n\t"
"paddb %%mm1, %%mm0 \n\t"
- "popl %1\n\t"
+#endif
"movd %%mm0, %0 \n\t"
: "=r" (numEq)
- : "r" (src), "r" (stride)
+ : "r" (src), "r" (stride), "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP])
+ : "%eax"
);
-// 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;
+ numEq= (-numEq) &0xFF;
+ return numEq > c->ppMode.flatnessThreshold;
}
+#endif
-static inline int isVertMinMaxOk(uint8_t src[], int stride, int QP)
+static inline int RENAME(isVertMinMaxOk)(uint8_t src[], int stride, PPContext *c)
{
#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 %%mm2, %%mm1 \n\t"
"por %%mm1, %%mm0 \n\t" // ABS Diff
- "movq pQPb, %%mm7 \n\t" // QP,..., QP
+ "movq %3, %%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"
+ "packssdw %%mm0, %%mm0 \n\t"
"movd %%mm0, %0 \n\t"
: "=r" (isOk)
- : "r" (src), "r" (stride)
+ : "r" (src), "r" (stride), "m" (c->pQPb)
);
- return isOk ? 1 : 0;
+ return isOk==0;
#else
-
- int isOk2= 1;
+#if 1
int x;
+ const int QP= c->QP;
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((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
}
-/* if(isOk && !isOk2 || !isOk && isOk2)
+
+ return 1;
+#else
+ int x;
+ const int QP= c->QP;
+ src+= stride*4;
+ for(x=0; x<BLOCK_SIZE; x++)
{
- 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");
+ int min=255;
+ int max=0;
+ int y;
+ for(y=0; y<8; y++){
+ int v= src[x + y*stride];
+ if(v>max) max=v;
+ if(v<min) min=v;
}
- } */
-
- return isOk2;
+ if(max-min > 2*QP) return 0;
+ }
+ return 1;
+#endif
#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)
+static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
src+= stride*3;
asm volatile( //"movv %0 %1 %2\n\t"
- "pushl %0 \n\t"
- "movq pQPb, %%mm0 \n\t" // QP,..., QP
+ "movq %2, %%mm0 \n\t" // QP,..., QP
+ "pxor %%mm4, %%mm4 \n\t"
"movq (%0), %%mm6 \n\t"
"movq (%0, %1), %%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
+ "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
"pand %%mm2, %%mm6 \n\t"
"pandn %%mm1, %%mm2 \n\t"
"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"
+ "leal (%0, %1, 8), %%ecx \n\t"
+ "subl %1, %%ecx \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"
"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
+ "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF
"pand %%mm2, %%mm7 \n\t"
"pandn %%mm1, %%mm2 \n\t"
// 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
+ // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1
// 6 4 2 2 1 1
// 6 4 4 2
// 6 8 2
"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
+ "movq (%%ecx), %%mm0 \n\t" // 1
PAVGB((%%eax, %1, 2), %%mm0) // 11/2
"movq %%mm0, %%mm3 \n\t" // 11/2
PAVGB(%%mm1, %%mm0) // 2 11/4
"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((%%ecx), %%mm0) // 11 /2
PAVGB(%%mm0, %%mm6) //11 11 /4
PAVGB(%%mm1, %%mm4) // 11 /2
PAVGB(%%mm2, %%mm1) // 11 /2
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
+ PAVGB((%%ecx), %%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
+ "movq %%mm6, (%%ecx) \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
- "popl %0\n\t"
+ "subl %1, %0 \n\t"
:
- : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "r" (src), "r" (stride), "m" (c->pQPb)
+ : "%eax", "%ecx"
);
#else
const int l1= stride;
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];
+ const int first= ABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
+ const int last= ABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];
int sums[9];
sums[0] = first + src[l1];
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[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
}
+#if 0
/**
* Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
* values are correctly clipped (MMX2)
x/8 = 1
1 12 12 23
*/
-static inline void vertRK1Filter(uint8_t *src, int stride, int QP)
+static inline void RENAME(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
+ "movq "MANGLE(b80)", %%mm6 \n\t" // MIN_SIGNED_BYTE
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
+ "leal (%%eax, %1, 4), %%ecx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
- "movq pQPb, %%mm0 \n\t" // QP,..., QP
+// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
+ "movq "MANGLE(pQPb)", %%mm0 \n\t" // QP,..., QP
"movq %%mm0, %%mm1 \n\t" // QP,..., QP
- "paddusb b02, %%mm0 \n\t"
+ "paddusb "MANGLE(b02)", %%mm0 \n\t"
"psrlw $2, %%mm0 \n\t"
- "pand b3F, %%mm0 \n\t" // QP/4,..., QP/4
+ "pand "MANGLE(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 (%%ecx), %%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
// "psubb %%mm6, %%mm2 \n\t"
"movq %%mm2, (%0,%1, 4) \n\t"
- "movq (%%ebx), %%mm2 \n\t"
+ "movq (%%ecx), %%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"
+ "movq %%mm2, (%%ecx) \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
+ "pand "MANGLE(b3F)", %%mm5 \n\t"
+ "psubb "MANGLE(b20)", %%mm5 \n\t" // (l5-l4)/8
"movq (%%eax, %1, 2), %%mm2 \n\t"
"paddb %%mm6, %%mm2 \n\t" // line 3 + 0x80
"psubb %%mm6, %%mm2 \n\t"
"movq %%mm2, (%%eax, %1, 2) \n\t"
- "movq (%%ebx, %1), %%mm2 \n\t"
+ "movq (%%ecx, %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"
+ "movq %%mm2, (%%ecx, %1) \n\t"
:
: "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "%eax", "%ecx"
);
#else
const int l1= stride;
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;
+// 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++)
{
- if(ABS(src[l4]-src[l5]) < QP + QP/4)
+ const int v = (src[x+l5] - src[x+l4]);
+ if(ABS(v) < QP15)
{
- int v = (src[l5] - src[l4]);
-
- src[l3] +=v/8;
- src[l4] +=v/2;
- src[l5] -=v/2;
- src[l6] -=v/8;
+ src[x+l3] +=v>>3;
+ src[x+l4] +=v>>1;
+ src[x+l5] -=v>>1;
+ src[x+l6] -=v>>3;
}
- src++;
}
#endif
}
+#endif
/**
* Experimental Filter 1
* 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)
+static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
{
#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"
+ "leal (%%eax, %1, 4), %%ecx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+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 (%%ecx), %%mm3 \n\t" // line 5
+ "movq (%%ecx, %1), %%mm4 \n\t" // line 6
"movq %%mm3, %%mm5 \n\t" // line 5
"psubusb %%mm4, %%mm3 \n\t"
"psubusb %%mm5, %%mm4 \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"
+ "movq %2, %%mm0 \n\t"
+ "paddusb %%mm0, %%mm0 \n\t"
+ "psubusb %%mm0, %%mm4 \n\t"
"pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0
- "psubusb b01, %%mm3 \n\t"
+ "psubusb "MANGLE(b01)", %%mm3 \n\t"
"pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0
PAVGB(%%mm7, %%mm3) // d/2
"pxor %%mm2, %%mm0 \n\t"
"movq %%mm0, (%0, %1, 4) \n\t" // line 4
- "movq (%%ebx), %%mm0 \n\t" // line 5
+ "movq (%%ecx), %%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
+ "movq %%mm0, (%%ecx) \n\t" // line 5
PAVGB(%%mm7, %%mm1) // d/4
"pxor %%mm2, %%mm0 \n\t"
"movq %%mm0, (%%eax, %1, 2) \n\t" // line 3
- "movq (%%ebx, %1), %%mm0 \n\t" // line 6
+ "movq (%%ecx, %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
+ "movq %%mm0, (%%ecx, %1) \n\t" // line 6
PAVGB(%%mm7, %%mm1) // d/8
"pxor %%mm2, %%mm0 \n\t"
"movq %%mm0, (%%eax, %1) \n\t" // line 2
- "movq (%%ebx, %1, 2), %%mm0 \n\t" // line 7
+ "movq (%%ecx, %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
+ "movq %%mm0, (%%ecx, %1, 2) \n\t" // line 7
:
- : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "r" (src), "r" (stride), "m" (co->pQPb)
+ : "%eax", "%ecx"
);
#else
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;
+// const int l8= stride + l7;
+// const int l9= stride + l8;
int x;
src+= stride*3;
int b= src[l4] - src[l5];
int c= src[l5] - src[l6];
- int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
+ int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
+ d= MAX(d, 0);
- if(d < QP)
+ if(d < co->QP*2)
{
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[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)
+static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
{
- 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);
+#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
/*
-//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);
- }
- }
+ 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;
-#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ 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
-// "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 pQPb, %%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"
+ "leal (%%eax, %1, 4), %%ecx \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 ecx+%1 ecx+2%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1
-// 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" /* pĀ“x = |px - p(x+1)| */\
- "pcmpeqb %%mm7, %%mm2 \n\t" /* pĀ“x = sgn[px - p(x+1)] */\
- "pshufw $0x00, %%mm1, %%mm3 \n\t" /* pĀ“5 = |p1 - p2| */\
- PAVGB(%%mm1, %%mm3) /* pĀ“5 = (|p2-p1| + |p6-p5|)/2 */\
- "psrlq $16, %%mm3 \n\t" /* pĀ“3 = (|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"\
+ "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
-/*
-HX1old((%0))
-HX1old((%%eax))
-HX1old((%%eax, %1))
-HX1old((%%eax, %1, 2))
-HX1old((%0, %1, 4))
-HX1old((%%ebx))
-HX1old((%%ebx, %1))
-HX1old((%%ebx, %1, 2))
-*/
+ "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
-//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" /* pĀ“x = |px - p(x+1)| */\
- "pcmpeqb %%mm7, %%mm2 \n\t" /* pĀ“x = sgn[px - p(x+1)] */\
- "punpckldq %%mm3, %%mm4 \n\t"\
- "movq %%mm1, %%mm3 \n\t"\
- "psllq $32, %%mm3 \n\t" /* pĀ“5 = |p1 - p2| */\
- PAVGB(%%mm1, %%mm3) /* pĀ“5 = (|p2-p1| + |p6-p5|)/2 */\
- "paddb %%mm6, %%mm0 \n\t"\
- "psrlq $16, %%mm3 \n\t" /* pĀ“3 = (|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 (%2, %%ecx, 8), %%mm0 \n\t"\
- "por %%mm2, %%mm3 \n\t" /* pĀ“x = |px - p(x+1)| */\
- "paddb %%mm6, %%mm0 \n\t"\
- "pcmpeqb %%mm7, %%mm2 \n\t" /* pĀ“x = sgn[px - p(x+1)] */\
- "movq %%mm3, %%mm1 \n\t"\
- "psllq $32, %%mm1 \n\t" /* pĀ“5 = |p1 - p2| */\
- "movq %%mm0, " #a " \n\t"\
- PAVGB(%%mm3, %%mm1) /* pĀ“5 = (|p2-p1| + |p6-p5|)/2 */\
- "paddb %%mm6, %%mm4 \n\t"\
- "psrlq $16, %%mm1 \n\t" /* pĀ“3 = (|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))
+ "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 (%%ecx), %%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 (%%ecx, %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 (%%ecx, %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 %2, %%mm4 \n\t" // QP //FIXME QP+1 ?
+ "paddusb "MANGLE(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 "MANGLE(b01)", %%mm3 \n\t"
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm7, %%mm3)
+ "paddusb %%mm1, %%mm3 \n\t"
+// "paddusb "MANGLE(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 "MANGLE(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 ecx+%1 ecx+2%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+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 "MANGLE(b80)", %%mm4 \n\t" // 128
+ "leal (%%eax, %1, 4), %%ecx \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 "MANGLE(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((%%ecx, %1), %%mm5) // (l6-l5+256)/2
+ "movq (%%ecx, %1, 2), %%mm1 \n\t" // l7
+ "pxor %%mm6, %%mm1 \n\t" // -l7-1
+ PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2
+ "movq "MANGLE(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 "MANGLE(b00)", %%mm1 \n\t" // 0
+ "movq "MANGLE(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 "MANGLE(b00)", %%mm7 \n\t" // 0
+ "movq %2, %%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 "MANGLE(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 "MANGLE(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), "r" (lut)
- : "%eax", "%ebx", "%ecx"
+ : "r" (src), "r" (stride), "m" (c->pQPb)
+ : "%eax", "%ecx"
);
-#else
-//FIXME (has little in common with the mmx2 version)
- for(y=0; y<BLOCK_SIZE; y++)
+/*
+ {
+ int x;
+ src-= stride;
+ for(x=0; x<BLOCK_SIZE; x++)
{
- int a= src[1] - src[2];
- int b= src[3] - src[4];
- int c= src[5] - src[6];
+ 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= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
+ int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
+ d= MAX(d, 0);
- if(d < QP)
- {
- int v = d * SIGN(-b);
+ d= (5*d + 32) >> 6;
+ d*= SIGN(-middleEnergy);
- 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;
+ 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);
+ }
}
- src+=stride;
}
-#endif
}
-
-
-static inline void doVertDefFilter(uint8_t src[], int stride, int QP)
-{
-#ifdef HAVE_MMX
+*/
+#elif defined (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"
+ "leal -40(%%esp), %%ecx \n\t" // make space for 4 8-byte vars
+ "andl $0xFFFFFFF8, %%ecx \n\t" // align
// 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
+// %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+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 (%0, %1), %%mm2 \n\t"
+ "leal (%0, %1, 2), %%eax \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 (%%eax), %%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
"psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
"psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
- "movq (%%eax, %1, 2), %%mm2 \n\t"
+ "movq (%%eax, %1), %%mm2 \n\t"
"movq %%mm2, %%mm3 \n\t"
"punpcklbw %%mm7, %%mm2 \n\t" // L3
"punpckhbw %%mm7, %%mm3 \n\t" // H3
"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 %%mm0, (%%ecx) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
+ "movq %%mm1, 8(%%ecx) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
- "movq (%0, %1, 4), %%mm0 \n\t"
+ "movq (%%eax, %1, 2), %%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
+ "movq %%mm2, 16(%%ecx) \n\t" // L3 - L4
+ "movq %%mm3, 24(%%ecx) \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
+ "leal (%%eax, %1), %0 \n\t"
"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 (%0, %1, 2), %%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 - 2L5
"psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
- "movq (%%ebx, %1), %%mm6 \n\t"
+ "movq (%%eax, %1, 4), %%mm6 \n\t"
"punpcklbw %%mm7, %%mm6 \n\t" // L6
"psubw %%mm6, %%mm2 \n\t" // L5 - L6
- "movq (%%ebx, %1), %%mm6 \n\t"
+ "movq (%%eax, %1, 4), %%mm6 \n\t"
"punpckhbw %%mm7, %%mm6 \n\t" // H6
"psubw %%mm6, %%mm3 \n\t" // H5 - H6
"psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
"psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
- "movq (%%ebx, %1, 2), %%mm2 \n\t"
+ "movq (%0, %1, 4), %%mm2 \n\t"
"movq %%mm2, %%mm3 \n\t"
"punpcklbw %%mm7, %%mm2 \n\t" // L7
"punpckhbw %%mm7, %%mm3 \n\t" // H7
"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
-//FIXME pxor, psubw, pmax for abs
+ "movq (%%ecx), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
+ "movq 8(%%ecx), %%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"
"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"
"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"
"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"
"psubusw %%mm1, %%mm5 \n\t" // ld
- "movq w05, %%mm2 \n\t" // 5
+ "movq "MANGLE(w05)", %%mm2 \n\t" // 5
"pmullw %%mm2, %%mm4 \n\t"
"pmullw %%mm2, %%mm5 \n\t"
- "movq w20, %%mm2 \n\t" // 32
+ "movq "MANGLE(w20)", %%mm2 \n\t" // 32
"paddw %%mm2, %%mm4 \n\t"
"paddw %%mm2, %%mm5 \n\t"
"psrlw $6, %%mm4 \n\t"
"psrlw $6, %%mm5 \n\t"
-/*
- "movq 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
+ "movq 16(%%ecx), %%mm0 \n\t" // L3 - L4
+ "movq 24(%%ecx), %%mm1 \n\t" // H3 - H4
"pxor %%mm2, %%mm2 \n\t"
"pxor %%mm3, %%mm3 \n\t"
- // FIXME rounding error
- "psraw $1, %%mm0 \n\t" // (L3 - L4)/2
- "psraw $1, %%mm1 \n\t" // (H3 - H4)/2
"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
+ "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"
"psubw %%mm6, %%mm4 \n\t"
"psubw %%mm7, %%mm5 \n\t"
"packsswb %%mm5, %%mm4 \n\t"
- "movq (%%eax, %1, 2), %%mm0 \n\t"
+ "movq (%0), %%mm0 \n\t"
"paddb %%mm4, %%mm0 \n\t"
- "movq %%mm0, (%%eax, %1, 2) \n\t"
- "movq (%0, %1, 4), %%mm0 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq (%0, %1), %%mm0 \n\t"
"psubb %%mm4, %%mm0 \n\t"
- "movq %%mm0, (%0, %1, 4) \n\t"
+ "movq %%mm0, (%0, %1) \n\t"
- :
- : "r" (src), "r" (stride), "r" (QP)
- : "%eax", "%ebx"
+ : "+r" (src)
+ : "r" (stride), "m" (c->pQPb)
+ : "%eax", "%ecx"
);
#else
const int l1= 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)
+ if(ABS(middleEnergy) < 8*c->QP)
{
const int q=(src[l4] - src[l5])/2;
const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
#endif
}
-//FIXME? |255-0| = 1
-/**
- * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock.
- */
-static inline int isHorizDCAndCopy2Temp(uint8_t src[], int stride)
+static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
{
-// src++;
- int numEq= 0;
-#ifdef HAVE_MMX
-asm volatile (
-// "int $3 \n\t"
- "pushl %1\n\t"
- "movq b7E, %%mm7 \n\t" // mm7 = 0x7F
- "movq b7C, %%mm6 \n\t" // mm6 = 0x7D
- "leal tempBlock, %%eax \n\t"
- "pxor %%mm0, %%mm0 \n\t"
+#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ asm volatile(
+ "pxor %%mm6, %%mm6 \n\t"
+ "pcmpeqb %%mm7, %%mm7 \n\t"
+ "movq %2, %%mm0 \n\t"
+ "punpcklbw %%mm6, %%mm0 \n\t"
+ "psrlw $1, %%mm0 \n\t"
+ "psubw %%mm7, %%mm0 \n\t"
+ "packuswb %%mm0, %%mm0 \n\t"
+ "movq %%mm0, %3 \n\t"
-#define HDC_CHECK_AND_CPY(i) \
- "movq -4(%1), %%mm2 \n\t"\
- "psrlq $32, %%mm2 \n\t"\
- "punpckldq 4(%1), %%mm2 \n\t" /* (%1) */\
- "movq %%mm2, %%mm1 \n\t"\
- "psrlq $8, %%mm2 \n\t"\
- "psubb %%mm1, %%mm2 \n\t"\
- "paddb %%mm7, %%mm2 \n\t"\
- "pcmpgtb %%mm6, %%mm2 \n\t"\
- "paddb %%mm2, %%mm0 \n\t"\
- "movq %%mm1," #i "(%%eax) \n\t"
-
- HDC_CHECK_AND_CPY(0)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(8)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(16)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(24)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(32)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(40)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(48)
- "addl %2, %1 \n\t"
- HDC_CHECK_AND_CPY(56)
-
- "psllq $8, %%mm0 \n\t" // remove dummy value
- "movq %%mm0, %%mm1 \n\t"
- "psrlw $8, %%mm0 \n\t"
- "paddb %%mm1, %%mm0 \n\t"
- "movq %%mm0, %%mm1 \n\t"
- "psrlq $16, %%mm0 \n\t"
- "paddb %%mm1, %%mm0 \n\t"
- "movq %%mm0, %%mm1 \n\t"
- "psrlq $32, %%mm0 \n\t"
- "paddb %%mm1, %%mm0 \n\t"
- "popl %1\n\t"
- "movd %%mm0, %0 \n\t"
- : "=r" (numEq)
- : "r" (src), "r" (stride)
- : "%eax"
- );
-// printf("%d\n", numEq);
- numEq= (256 - (numEq & 0xFF)) &0xFF;
-#else
- int y;
- for(y=0; y<BLOCK_SIZE; y++)
- {
- if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++;
- if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++;
- tempBlock[0 + y*TEMP_STRIDE] = src[0];
- tempBlock[1 + y*TEMP_STRIDE] = src[1];
- tempBlock[2 + y*TEMP_STRIDE] = src[2];
- tempBlock[3 + y*TEMP_STRIDE] = src[3];
- tempBlock[4 + y*TEMP_STRIDE] = src[4];
- tempBlock[5 + y*TEMP_STRIDE] = src[5];
- tempBlock[6 + y*TEMP_STRIDE] = src[6];
- tempBlock[7 + y*TEMP_STRIDE] = src[7];
- src+= stride;
- }
-#endif
-/* if(abs(numEq - asmEq) > 0)
- {
-// printf("\nasm:%d c:%d\n", asmEq, numEq);
- for(int y=0; y<8; y++)
- {
- for(int x=0; x<8; x++)
- {
- printf("%d ", src[x + y*stride]);
- }
- printf("\n");
- }
- }
-*/
-// printf("%d\n", numEq);
- return numEq > hFlatnessThreshold;
-}
-
-static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)
-{
-#ifdef MMX_FIXME
-FIXME
- int isOk;
- asm volatile(
-// "int $3 \n\t"
- "movq (%1, %2), %%mm0 \n\t"
- "movq (%1, %2, 8), %%mm1 \n\t"
- "movq %%mm0, %%mm2 \n\t"
- "psubusb %%mm1, %%mm0 \n\t"
- "psubusb %%mm2, %%mm1 \n\t"
- "por %%mm1, %%mm0 \n\t" // ABS Diff
-
- "movq pQPb, %%mm7 \n\t" // QP,..., QP
- "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP
- "psubusb %%mm7, %%mm0 \n\t" // Diff <= 2QP -> 0
- "pcmpeqd b00, %%mm0 \n\t"
- "psrlq $16, %%mm0 \n\t"
- "pcmpeqd bFF, %%mm0 \n\t"
-// "movd %%mm0, (%1, %2, 4)\n\t"
- "movd %%mm0, %0 \n\t"
- : "=r" (isOk)
- : "r" (src), "r" (stride)
- );
- return isOk;
-#else
- if(abs(src[0] - src[7]) > 2*QP) return 0;
-
- return 1;
-#endif
-}
+ "leal (%0, %1), %%eax \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
+
+// 0 1 2 3 4 5 6 7 8 9
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
-static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP)
-{
-#ifdef HAVE_MMX
- asm volatile(
- "pushl %0 \n\t"
- "pxor %%mm7, %%mm7 \n\t"
- "movq bm00001000, %%mm6 \n\t"
- "movd %2, %%mm5 \n\t" // QP
- "movq %%mm5, %%mm4 \n\t"
- "paddusb %%mm5, %%mm5 \n\t" // 2QP
- "paddusb %%mm5, %%mm4 \n\t" // 3QP
- "psllq $24, %%mm4 \n\t"
- "pxor %%mm5, %%mm5 \n\t" // 0
- "psubb %%mm4, %%mm5 \n\t" // -QP
- "leal tempBlock, %%eax \n\t"
-
-//FIXME? "unroll by 2" and mix
+#undef FIND_MIN_MAX
#ifdef HAVE_MMX2
-#define HDF(i) \
- "movq " #i "(%%eax), %%mm0 \n\t"\
- "movq %%mm0, %%mm1 \n\t"\
- "movq %%mm0, %%mm2 \n\t"\
- "psrlq $8, %%mm1 \n\t"\
- "psubusb %%mm1, %%mm2 \n\t"\
- "psubusb %%mm0, %%mm1 \n\t"\
- "por %%mm2, %%mm1 \n\t" /* pĀ“x = |px - p(x+1)| */\
- "pcmpeqb %%mm7, %%mm2 \n\t" /* pĀ“x = sgn[px - p(x+1)] */\
- "pshufw $0x00, %%mm1, %%mm3 \n\t" /* pĀ“5 = |p1 - p2| */\
- "pminub %%mm1, %%mm3 \n\t" /* pĀ“5 = min(|p2-p1|, |p6-p5|)*/\
- "psrlq $16, %%mm3 \n\t" /* pĀ“3 = min(|p2-p1|, |p6-p5|)*/\
- "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\
- "paddb %%mm5, %%mm1 \n\t"\
- "psubusb %%mm5, %%mm1 \n\t"\
- "psrlw $2, %%mm1 \n\t"\
- "pxor %%mm2, %%mm1 \n\t"\
- "psubb %%mm2, %%mm1 \n\t"\
- "pand %%mm6, %%mm1 \n\t"\
- "psubb %%mm1, %%mm0 \n\t"\
- "psllq $8, %%mm1 \n\t"\
- "paddb %%mm1, %%mm0 \n\t"\
- "movd %%mm0, (%0) \n\t"\
- "psrlq $32, %%mm0 \n\t"\
- "movd %%mm0, 4(%0) \n\t"
+#define FIND_MIN_MAX(addr)\
+ "movq " #addr ", %%mm0 \n\t"\
+ "pminub %%mm0, %%mm7 \n\t"\
+ "pmaxub %%mm0, %%mm6 \n\t"
#else
-#define HDF(i)\
- "movq " #i "(%%eax), %%mm0 \n\t"\
- "movq %%mm0, %%mm1 \n\t"\
- "movq %%mm0, %%mm2 \n\t"\
- "psrlq $8, %%mm1 \n\t"\
- "psubusb %%mm1, %%mm2 \n\t"\
+#define FIND_MIN_MAX(addr)\
+ "movq " #addr ", %%mm0 \n\t"\
+ "movq %%mm7, %%mm1 \n\t"\
+ "psubusb %%mm0, %%mm6 \n\t"\
+ "paddb %%mm0, %%mm6 \n\t"\
"psubusb %%mm0, %%mm1 \n\t"\
- "por %%mm2, %%mm1 \n\t" /* pĀ“x = |px - p(x+1)| */\
- "pcmpeqb %%mm7, %%mm2 \n\t" /* pĀ“x = 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" /* pĀ“3 = min(|p2-p1|, |p6-p5|)*/\
- "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,Ć¼6|) */\
- "paddb %%mm5, %%mm1 \n\t"\
- "psubusb %%mm5, %%mm1 \n\t"\
- "psrlw $2, %%mm1 \n\t"\
- "pxor %%mm2, %%mm1 \n\t"\
- "psubb %%mm2, %%mm1 \n\t"\
- "pand %%mm6, %%mm1 \n\t"\
- "psubb %%mm1, %%mm0 \n\t"\
- "psllq $8, %%mm1 \n\t"\
- "paddb %%mm1, %%mm0 \n\t"\
- "movd %%mm0, (%0) \n\t"\
- "psrlq $32, %%mm0 \n\t"\
- "movd %%mm0, 4(%0) \n\t"
+ "psubb %%mm1, %%mm7 \n\t"
#endif
- HDF(0)
- "addl %1, %0 \n\t"
- HDF(8)
- "addl %1, %0 \n\t"
- HDF(16)
- "addl %1, %0 \n\t"
- HDF(24)
- "addl %1, %0 \n\t"
- HDF(32)
- "addl %1, %0 \n\t"
- HDF(40)
- "addl %1, %0 \n\t"
- HDF(48)
- "addl %1, %0 \n\t"
- HDF(56)
- "popl %0 \n\t"
- :
- : "r" (dst), "r" (stride), "r" (QP)
- : "%eax"
- );
-#else
- uint8_t *src= tempBlock;
- int y;
- for(y=0; y<BLOCK_SIZE; y++)
- {
- const int middleEnergy= 5*(src[4] - src[5]) + 2*(src[2] - src[5]);
-
- dst[0] = src[0];
- dst[1] = src[1];
- dst[2] = src[2];
- dst[3] = src[3];
- dst[4] = src[4];
- dst[5] = src[5];
- dst[6] = src[6];
- dst[7] = src[7];
-
- if(ABS(middleEnergy) < 8*QP)
- {
- const int q=(src[3] - src[4])/2;
- const int leftEnergy= 5*(src[2] - src[1]) + 2*(src[0] - src[3]);
- const int rightEnergy= 5*(src[6] - src[5]) + 2*(src[4] - src[7]);
+FIND_MIN_MAX((%%eax))
+FIND_MIN_MAX((%%eax, %1))
+FIND_MIN_MAX((%%eax, %1, 2))
+FIND_MIN_MAX((%0, %1, 4))
+FIND_MIN_MAX((%%edx))
+FIND_MIN_MAX((%%edx, %1))
+FIND_MIN_MAX((%%edx, %1, 2))
+FIND_MIN_MAX((%0, %1, 8))
- int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
- d= MAX(d, 0);
+ "movq %%mm7, %%mm4 \n\t"
+ "psrlq $8, %%mm7 \n\t"
+#ifdef HAVE_MMX2
+ "pminub %%mm4, %%mm7 \n\t" // min of pixels
+ "pshufw $0xF9, %%mm7, %%mm4 \n\t"
+ "pminub %%mm4, %%mm7 \n\t" // min of pixels
+ "pshufw $0xFE, %%mm7, %%mm4 \n\t"
+ "pminub %%mm4, %%mm7 \n\t"
+#else
+ "movq %%mm7, %%mm1 \n\t"
+ "psubusb %%mm4, %%mm1 \n\t"
+ "psubb %%mm1, %%mm7 \n\t"
+ "movq %%mm7, %%mm4 \n\t"
+ "psrlq $16, %%mm7 \n\t"
+ "movq %%mm7, %%mm1 \n\t"
+ "psubusb %%mm4, %%mm1 \n\t"
+ "psubb %%mm1, %%mm7 \n\t"
+ "movq %%mm7, %%mm4 \n\t"
+ "psrlq $32, %%mm7 \n\t"
+ "movq %%mm7, %%mm1 \n\t"
+ "psubusb %%mm4, %%mm1 \n\t"
+ "psubb %%mm1, %%mm7 \n\t"
+#endif
- d= (5*d + 32) >> 6;
- d*= SIGN(-middleEnergy);
- if(q>0)
- {
- d= d<0 ? 0 : d;
- d= d>q ? q : d;
- }
- else
- {
- d= d>0 ? 0 : d;
- d= d<q ? q : d;
- }
-
- dst[3]-= d;
- dst[4]+= d;
- }
- dst+= stride;
- src+= TEMP_STRIDE;
- }
+ "movq %%mm6, %%mm4 \n\t"
+ "psrlq $8, %%mm6 \n\t"
+#ifdef HAVE_MMX2
+ "pmaxub %%mm4, %%mm6 \n\t" // max of pixels
+ "pshufw $0xF9, %%mm6, %%mm4 \n\t"
+ "pmaxub %%mm4, %%mm6 \n\t"
+ "pshufw $0xFE, %%mm6, %%mm4 \n\t"
+ "pmaxub %%mm4, %%mm6 \n\t"
+#else
+ "psubusb %%mm4, %%mm6 \n\t"
+ "paddb %%mm4, %%mm6 \n\t"
+ "movq %%mm6, %%mm4 \n\t"
+ "psrlq $16, %%mm6 \n\t"
+ "psubusb %%mm4, %%mm6 \n\t"
+ "paddb %%mm4, %%mm6 \n\t"
+ "movq %%mm6, %%mm4 \n\t"
+ "psrlq $32, %%mm6 \n\t"
+ "psubusb %%mm4, %%mm6 \n\t"
+ "paddb %%mm4, %%mm6 \n\t"
#endif
-}
+ "movq %%mm6, %%mm0 \n\t" // max
+ "psubb %%mm7, %%mm6 \n\t" // max - min
+ "movd %%mm6, %%ecx \n\t"
+ "cmpb "MANGLE(deringThreshold)", %%cl \n\t"
+ " jb 1f \n\t"
+ "leal -24(%%esp), %%ecx \n\t"
+ "andl $0xFFFFFFF8, %%ecx \n\t"
+ PAVGB(%%mm0, %%mm7) // a=(max + min)/2
+ "punpcklbw %%mm7, %%mm7 \n\t"
+ "punpcklbw %%mm7, %%mm7 \n\t"
+ "punpcklbw %%mm7, %%mm7 \n\t"
+ "movq %%mm7, (%%ecx) \n\t"
+
+ "movq (%0), %%mm0 \n\t" // L10
+ "movq %%mm0, %%mm1 \n\t" // L10
+ "movq %%mm0, %%mm2 \n\t" // L10
+ "psllq $8, %%mm1 \n\t"
+ "psrlq $8, %%mm2 \n\t"
+ "movd -4(%0), %%mm3 \n\t"
+ "movd 8(%0), %%mm4 \n\t"
+ "psrlq $24, %%mm3 \n\t"
+ "psllq $56, %%mm4 \n\t"
+ "por %%mm3, %%mm1 \n\t" // L00
+ "por %%mm4, %%mm2 \n\t" // L20
+ "movq %%mm1, %%mm3 \n\t" // L00
+ PAVGB(%%mm2, %%mm1) // (L20 + L00)/2
+ PAVGB(%%mm0, %%mm1) // (L20 + L00 + 2L10)/4
+ "psubusb %%mm7, %%mm0 \n\t"
+ "psubusb %%mm7, %%mm2 \n\t"
+ "psubusb %%mm7, %%mm3 \n\t"
+ "pcmpeqb "MANGLE(b00)", %%mm0 \n\t" // L10 > a ? 0 : -1
+ "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1
+ "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1
+ "paddb %%mm2, %%mm0 \n\t"
+ "paddb %%mm3, %%mm0 \n\t"
+
+ "movq (%%eax), %%mm2 \n\t" // L11
+ "movq %%mm2, %%mm3 \n\t" // L11
+ "movq %%mm2, %%mm4 \n\t" // L11
+ "psllq $8, %%mm3 \n\t"
+ "psrlq $8, %%mm4 \n\t"
+ "movd -4(%%eax), %%mm5 \n\t"
+ "movd 8(%%eax), %%mm6 \n\t"
+ "psrlq $24, %%mm5 \n\t"
+ "psllq $56, %%mm6 \n\t"
+ "por %%mm5, %%mm3 \n\t" // L01
+ "por %%mm6, %%mm4 \n\t" // L21
+ "movq %%mm3, %%mm5 \n\t" // L01
+ PAVGB(%%mm4, %%mm3) // (L21 + L01)/2
+ PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4
+ "psubusb %%mm7, %%mm2 \n\t"
+ "psubusb %%mm7, %%mm4 \n\t"
+ "psubusb %%mm7, %%mm5 \n\t"
+ "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1
+ "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1
+ "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1
+ "paddb %%mm4, %%mm2 \n\t"
+ "paddb %%mm5, %%mm2 \n\t"
+// 0, 2, 3, 1
+#define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
+ "movq " #src ", " #sx " \n\t" /* src[0] */\
+ "movq " #sx ", " #lx " \n\t" /* src[0] */\
+ "movq " #sx ", " #t0 " \n\t" /* src[0] */\
+ "psllq $8, " #lx " \n\t"\
+ "psrlq $8, " #t0 " \n\t"\
+ "movd -4" #src ", " #t1 " \n\t"\
+ "psrlq $24, " #t1 " \n\t"\
+ "por " #t1 ", " #lx " \n\t" /* src[-1] */\
+ "movd 8" #src ", " #t1 " \n\t"\
+ "psllq $56, " #t1 " \n\t"\
+ "por " #t1 ", " #t0 " \n\t" /* src[+1] */\
+ "movq " #lx ", " #t1 " \n\t" /* src[-1] */\
+ PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\
+ PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\
+ PAVGB(lx, pplx) \
+ "movq " #lx ", 8(%%ecx) \n\t"\
+ "movq (%%ecx), " #lx " \n\t"\
+ "psubusb " #lx ", " #t1 " \n\t"\
+ "psubusb " #lx ", " #t0 " \n\t"\
+ "psubusb " #lx ", " #sx " \n\t"\
+ "movq "MANGLE(b00)", " #lx " \n\t"\
+ "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\
+ "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\
+ "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\
+ "paddb " #t1 ", " #t0 " \n\t"\
+ "paddb " #t0 ", " #sx " \n\t"\
+\
+ PAVGB(plx, pplx) /* filtered */\
+ "movq " #dst ", " #t0 " \n\t" /* dst */\
+ "movq " #t0 ", " #t1 " \n\t" /* dst */\
+ "psubusb %3, " #t0 " \n\t"\
+ "paddusb %3, " #t1 " \n\t"\
+ PMAXUB(t0, pplx)\
+ PMINUB(t1, pplx, t0)\
+ "paddb " #sx ", " #ppsx " \n\t"\
+ "paddb " #psx ", " #ppsx " \n\t"\
+ "#paddb "MANGLE(b02)", " #ppsx " \n\t"\
+ "pand "MANGLE(b08)", " #ppsx " \n\t"\
+ "pcmpeqb " #lx ", " #ppsx " \n\t"\
+ "pand " #ppsx ", " #pplx " \n\t"\
+ "pandn " #dst ", " #ppsx " \n\t"\
+ "por " #pplx ", " #ppsx " \n\t"\
+ "movq " #ppsx ", " #dst " \n\t"\
+ "movq 8(%%ecx), " #lx " \n\t"
-/**
- * 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 doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP)
-{
-//return;
-#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
- asm volatile( //"movv %0 %1 %2\n\t"
- "pushl %0\n\t"
- "pxor %%mm7, %%mm7 \n\t"
- "leal tempBlock, %%eax \n\t"
-/*
-#define HLP1 "movq (%0), %%mm0 \n\t"\
- "movq %%mm0, %%mm1 \n\t"\
- "psllq $8, %%mm0 \n\t"\
- PAVGB(%%mm1, %%mm0)\
- "psrlw $8, %%mm0 \n\t"\
- "pxor %%mm1, %%mm1 \n\t"\
- "packuswb %%mm1, %%mm0 \n\t"\
- "movq %%mm0, %%mm1 \n\t"\
- "movq %%mm0, %%mm2 \n\t"\
- "psllq $32, %%mm0 \n\t"\
- "paddb %%mm0, %%mm1 \n\t"\
- "psllq $16, %%mm2 \n\t"\
- PAVGB(%%mm2, %%mm0)\
- "movq %%mm0, %%mm3 \n\t"\
- "pand bm11001100, %%mm0 \n\t"\
- "paddusb %%mm0, %%mm3 \n\t"\
- "psrlq $8, %%mm3 \n\t"\
- PAVGB(%%mm1, %%mm4)\
- PAVGB(%%mm3, %%mm2)\
- "psrlq $16, %%mm2 \n\t"\
- "punpcklbw %%mm2, %%mm2 \n\t"\
- "movq %%mm2, (%0) \n\t"\
-
-#define HLP2 "movq (%0), %%mm0 \n\t"\
- "movq %%mm0, %%mm1 \n\t"\
- "psllq $8, %%mm0 \n\t"\
- PAVGB(%%mm1, %%mm0)\
- "psrlw $8, %%mm0 \n\t"\
- "pxor %%mm1, %%mm1 \n\t"\
- "packuswb %%mm1, %%mm0 \n\t"\
- "movq %%mm0, %%mm2 \n\t"\
- "psllq $32, %%mm0 \n\t"\
- "psllq $16, %%mm2 \n\t"\
- PAVGB(%%mm2, %%mm0)\
- "movq %%mm0, %%mm3 \n\t"\
- "pand bm11001100, %%mm0 \n\t"\
- "paddusb %%mm0, %%mm3 \n\t"\
- "psrlq $8, %%mm3 \n\t"\
- PAVGB(%%mm3, %%mm2)\
- "psrlq $16, %%mm2 \n\t"\
- "punpcklbw %%mm2, %%mm2 \n\t"\
- "movq %%mm2, (%0) \n\t"\
-*/
-// approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16
/*
-Implemented Exact 7-Tap
- 9421 A321
- 36421 64321
- 334321 =
- 1234321 =
- 1234321 =
- 123433 =
- 12463 12346
- 1249 123A
+0000000
+1111111
-*/
+1111110
+1111101
+1111100
+1111011
+1111010
+1111001
-#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
+1111000
+1110111
-/* uses the 7-Tap Filter: 1112111 */
-#define NEW_HLP(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"\
- "movq %%mm1, %%mm5 \n\t"\
- PAVGB(%%mm2, %%mm1)\
- 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, (%0) \n\t"\
- "psrlq $32, %%mm0 \n\t"\
- "movd %%mm0, 4(%0) \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(i) NEW_HLP(i)
-
- HLP(0)
- "addl %1, %0 \n\t"
- HLP(8)
- "addl %1, %0 \n\t"
- HLP(16)
- "addl %1, %0 \n\t"
- HLP(24)
- "addl %1, %0 \n\t"
- HLP(32)
- "addl %1, %0 \n\t"
- HLP(40)
- "addl %1, %0 \n\t"
- HLP(48)
- "addl %1, %0 \n\t"
- HLP(56)
-
- "popl %0\n\t"
- :
- : "r" (dst), "r" (stride)
- : "%eax", "%ebx"
+*/
+//DERING_CORE(dst,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1)
+DERING_CORE((%%eax),(%%eax, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
+DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
+DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
+DERING_CORE((%0, %1, 4),(%%edx) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
+DERING_CORE((%%edx),(%%edx, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
+DERING_CORE((%%edx, %1), (%%edx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
+DERING_CORE((%%edx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
+DERING_CORE((%0, %1, 8),(%%edx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
+
+ "1: \n\t"
+ : : "r" (src), "r" (stride), "m" (c->pQPb), "m"(c->pQPb2)
+ : "%eax", "%edx", "%ecx"
);
-
#else
- uint8_t *temp= tempBlock;
int y;
- for(y=0; y<BLOCK_SIZE; y++)
+ int min=255;
+ int max=0;
+ int avg;
+ uint8_t *p;
+ int s[10];
+ const int QP2= c->QP/2 + 1;
+
+ for(y=1; y<9; y++)
{
- const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];
- const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];
-
- int sums[9];
- sums[0] = first + temp[0];
- sums[1] = temp[0] + temp[1];
- sums[2] = temp[1] + temp[2];
- sums[3] = temp[2] + temp[3];
- sums[4] = temp[3] + temp[4];
- sums[5] = temp[4] + temp[5];
- sums[6] = temp[5] + temp[6];
- sums[7] = temp[6] + temp[7];
- sums[8] = temp[7] + last;
-
- dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
- dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4;
- dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4;
- dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4;
- dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4;
- dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4;
- dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4;
- dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4;
-
- dst+= stride;
- temp+= TEMP_STRIDE;
+ int x;
+ p= src + stride*y;
+ for(x=1; x<9; x++)
+ {
+ p++;
+ if(*p > max) max= *p;
+ if(*p < min) min= *p;
+ }
}
-#endif
-}
-
-
-static inline void dering(uint8_t src[], int stride, int QP)
-{
-//FIXME
-
-#ifdef HAVE_MMX2X
- asm volatile(
- "leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
-// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
-
- "pcmpeq %%mm6, %%mm6 \n\t"
- "pxor %%mm7, %%mm7 \n\t"
+ avg= (min + max + 1)>>1;
-#define FIND_MIN_MAX(addr)\
- "movq (" #addr "), %%mm0, \n\t"\
- "pminub %%mm0, %%mm6 \n\t"\
- "pmaxub %%mm0, %%mm7 \n\t"
-
-FIND_MIN_MAX(%0)
-FIND_MIN_MAX(%%eax)
-FIND_MIN_MAX(%%eax, %1)
-FIND_MIN_MAX(%%eax, %1, 2)
-FIND_MIN_MAX(%0, %1, 4)
-FIND_MIN_MAX(%%ebx)
-FIND_MIN_MAX(%%ebx, %1)
-FIND_MIN_MAX(%%ebx, %1, 2)
-FIND_MIN_MAX(%0, %1, 8)
-FIND_MIN_MAX(%%ebx, %1, 2)
+ if(max - min <deringThreshold) return;
- "movq %%mm6, %%mm4 \n\t"
- "psrlq $32, %%mm6 \n\t"
- "pminub %%mm4, %%mm6 \n\t"
- "movq %%mm6, %%mm4 \n\t"
- "psrlq $16, %%mm6 \n\t"
- "pminub %%mm4, %%mm6 \n\t"
- "movq %%mm6, %%mm4 \n\t"
- "psrlq $8, %%mm6 \n\t"
- "pminub %%mm4, %%mm6 \n\t" // min of pixels
+ for(y=0; y<10; y++)
+ {
+ int t = 0;
+
+ if(src[stride*y + 0] > avg) t+= 1;
+ if(src[stride*y + 1] > avg) t+= 2;
+ if(src[stride*y + 2] > avg) t+= 4;
+ if(src[stride*y + 3] > avg) t+= 8;
+ if(src[stride*y + 4] > avg) t+= 16;
+ if(src[stride*y + 5] > avg) t+= 32;
+ if(src[stride*y + 6] > avg) t+= 64;
+ if(src[stride*y + 7] > avg) t+= 128;
+ if(src[stride*y + 8] > avg) t+= 256;
+ if(src[stride*y + 9] > avg) t+= 512;
+
+ t |= (~t)<<16;
+ t &= (t<<1) & (t>>1);
+ s[y] = t;
+ }
+
+ for(y=1; y<9; y++)
+ {
+ int t = s[y-1] & s[y] & s[y+1];
+ t|= t>>16;
+ s[y-1]= t;
+ }
- "movq %%mm7, %%mm4 \n\t"
- "psrlq $32, %%mm7 \n\t"
- "pmaxub %%mm4, %%mm7 \n\t"
- "movq %%mm7, %%mm4 \n\t"
- "psrlq $16, %%mm7 \n\t"
- "pmaxub %%mm4, %%mm7 \n\t"
- "movq %%mm7, %%mm4 \n\t"
- "psrlq $8, %%mm7 \n\t"
- "pmaxub %%mm4, %%mm7 \n\t" // max of pixels
- PAVGB(%%mm6, %%mm7) // (max + min)/2
+ for(y=1; y<9; y++)
+ {
+ int x;
+ int t = s[y-1];
+ p= src + stride*y;
+ for(x=1; x<9; x++)
+ {
+ p++;
+ if(t & (1<<x))
+ {
+ int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
+ +2*(*(p -1)) + 4*(*p ) + 2*(*(p +1))
+ +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
+ f= (f + 8)>>4;
- : : "r" (src), "r" (stride), "r" (QP)
- : "%eax", "%ebx"
- );
-#else
+#ifdef DEBUG_DERING_THRESHOLD
+ asm volatile("emms\n\t":);
+ {
+ static long long numPixels=0;
+ if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
+// if((max-min)<20 || (max-min)*QP<200)
+// if((max-min)*QP < 500)
+// if(max-min<QP/2)
+ if(max-min < 20)
+ {
+ static int numSkiped=0;
+ static int errorSum=0;
+ static int worstQP=0;
+ static int worstRange=0;
+ static int worstDiff=0;
+ int diff= (f - *p);
+ int absDiff= ABS(diff);
+ int error= diff*diff;
+
+ if(x==1 || x==8 || y==1 || y==8) continue;
+
+ numSkiped++;
+ if(absDiff > worstDiff)
+ {
+ worstDiff= absDiff;
+ worstQP= QP;
+ worstRange= max-min;
+ }
+ errorSum+= error;
-//FIXME
+ if(1024LL*1024LL*1024LL % numSkiped == 0)
+ {
+ printf( "sum:%1.3f, skip:%d, wQP:%d, "
+ "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
+ (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
+ worstDiff, (float)numSkiped/numPixels);
+ }
+ }
+ }
+#endif
+ if (*p + QP2 < f) *p= *p + QP2;
+ else if(*p - QP2 > f) *p= *p - QP2;
+ else *p=f;
+ }
+ }
+ }
+#ifdef DEBUG_DERING_THRESHOLD
+ if(max-min < 20)
+ {
+ for(y=1; y<9; y++)
+ {
+ int x;
+ int t = 0;
+ p= src + stride*y;
+ for(x=1; x<9; x++)
+ {
+ p++;
+ *p = MIN(*p + 20, 255);
+ }
+ }
+// src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
+ }
+#endif
#endif
}
/**
* Deinterlaces the given block
- * will be called for every 8x8 block, and can read & write into an 8x16 block
+ * will be called for every 8x8 block and can read & write from line 4-15
+ * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
+ * lines 4-12 will be read into the deblocking filter and should be deinterlaced
*/
-static inline void deInterlaceInterpolateLinear(uint8_t src[], int stride)
+static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ src+= 4*stride;
asm volatile(
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
+ "leal (%%eax, %1, 4), %%ecx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1
"movq (%0), %%mm0 \n\t"
"movq (%%eax, %1), %%mm1 \n\t"
"movq (%0, %1, 4), %%mm0 \n\t"
PAVGB(%%mm0, %%mm1)
"movq %%mm1, (%%eax, %1, 2) \n\t"
- "movq (%%ebx, %1), %%mm1 \n\t"
+ "movq (%%ecx, %1), %%mm1 \n\t"
PAVGB(%%mm1, %%mm0)
- "movq %%mm0, (%%ebx) \n\t"
+ "movq %%mm0, (%%ecx) \n\t"
"movq (%0, %1, 8), %%mm0 \n\t"
PAVGB(%%mm0, %%mm1)
- "movq %%mm1, (%%ebx, %1, 2) \n\t"
+ "movq %%mm1, (%%ecx, %1, 2) \n\t"
: : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "%eax", "%ecx"
);
#else
int x;
+ src+= 4*stride;
for(x=0; x<8; x++)
{
src[stride] = (src[0] + src[stride*2])>>1;
/**
* Deinterlaces the given block
- * will be called for every 8x8 block, and can read & write into an 8x16 block
+ * will be called for every 8x8 block and can read & write from line 4-15
+ * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
+ * lines 4-12 will be read into the deblocking filter and should be deinterlaced
+ * this filter will read lines 3-15 and write 7-13
* no cliping in C version
*/
-static inline void deInterlaceInterpolateCubic(uint8_t src[], int stride)
+static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ src+= stride*3;
asm volatile(
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
- "leal (%%ebx, %1, 4), %%ecx \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
+ "leal (%%edx, %1, 4), %%ecx \n\t"
"addl %1, %%ecx \n\t"
"pxor %%mm7, %%mm7 \n\t"
// 0 1 2 3 4 5 6 7 8 9 10
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1 ecx
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
#define DEINT_CUBIC(a,b,c,d,e)\
"movq " #a ", %%mm0 \n\t"\
"packuswb %%mm3, %%mm1 \n\t"\
"movq %%mm1, " #c " \n\t"
-DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1))
-DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8))
-DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx))
-DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2))
+DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx, %1))
+DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%edx), (%%edx, %1), (%0, %1, 8))
+DEINT_CUBIC((%0, %1, 4), (%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%ecx))
+DEINT_CUBIC((%%edx, %1), (%0, %1, 8), (%%edx, %1, 4), (%%ecx), (%%ecx, %1, 2))
: : "r" (src), "r" (stride)
- : "%eax", "%ebx", "ecx"
+ : "%eax", "%edx", "ecx"
);
#else
int x;
+ src+= stride*3;
for(x=0; x<8; x++)
{
src[stride*3] = (-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4;
/**
* Deinterlaces the given block
- * will be called for every 8x8 block, and can read & write into an 8x16 block
+ * will be called for every 8x8 block and can read & write from line 4-15
+ * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
+ * lines 4-12 will be read into the deblocking filter and should be deinterlaced
+ * this filter will read lines 4-13 and write 5-11
+ * no cliping in C version
+ */
+static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)
+{
+#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ src+= stride*4;
+ asm volatile(
+ "leal (%0, %1), %%eax \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
+ "pxor %%mm7, %%mm7 \n\t"
+ "movq (%2), %%mm0 \n\t"
+// 0 1 2 3 4 5 6 7 8 9 10
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx
+
+#define DEINT_FF(a,b,c,d)\
+ "movq " #a ", %%mm1 \n\t"\
+ "movq " #b ", %%mm2 \n\t"\
+ "movq " #c ", %%mm3 \n\t"\
+ "movq " #d ", %%mm4 \n\t"\
+ PAVGB(%%mm3, %%mm1) \
+ PAVGB(%%mm4, %%mm0) \
+ "movq %%mm0, %%mm3 \n\t"\
+ "punpcklbw %%mm7, %%mm0 \n\t"\
+ "punpckhbw %%mm7, %%mm3 \n\t"\
+ "movq %%mm1, %%mm4 \n\t"\
+ "punpcklbw %%mm7, %%mm1 \n\t"\
+ "punpckhbw %%mm7, %%mm4 \n\t"\
+ "psllw $2, %%mm1 \n\t"\
+ "psllw $2, %%mm4 \n\t"\
+ "psubw %%mm0, %%mm1 \n\t"\
+ "psubw %%mm3, %%mm4 \n\t"\
+ "movq %%mm2, %%mm5 \n\t"\
+ "movq %%mm2, %%mm0 \n\t"\
+ "punpcklbw %%mm7, %%mm2 \n\t"\
+ "punpckhbw %%mm7, %%mm5 \n\t"\
+ "paddw %%mm2, %%mm1 \n\t"\
+ "paddw %%mm5, %%mm4 \n\t"\
+ "psraw $2, %%mm1 \n\t"\
+ "psraw $2, %%mm4 \n\t"\
+ "packuswb %%mm4, %%mm1 \n\t"\
+ "movq %%mm1, " #b " \n\t"\
+
+DEINT_FF((%0) , (%%eax) , (%%eax, %1), (%%eax, %1, 2))
+DEINT_FF((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx) )
+DEINT_FF((%0, %1, 4), (%%edx) , (%%edx, %1), (%%edx, %1, 2))
+DEINT_FF((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%edx, %1, 4))
+
+ "movq %%mm0, (%2) \n\t"
+ : : "r" (src), "r" (stride), "r"(tmp)
+ : "%eax", "%edx"
+ );
+#else
+ int x;
+ src+= stride*4;
+ for(x=0; x<8; x++)
+ {
+ int t1= tmp[x];
+ int t2= src[stride*1];
+
+ src[stride*1]= (-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3;
+ t1= src[stride*4];
+ src[stride*3]= (-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3;
+ t2= src[stride*6];
+ src[stride*5]= (-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3;
+ t1= src[stride*8];
+ src[stride*7]= (-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3;
+ tmp[x]= t1;
+
+ src++;
+ }
+#endif
+}
+
+/**
+ * Deinterlaces the given block
+ * will be called for every 8x8 block and can read & write from line 4-15
+ * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
+ * lines 4-12 will be read into the deblocking filter and should be deinterlaced
* will shift the image up by 1 line (FIXME if this is a problem)
+ * this filter will read lines 4-13 and write 4-11
*/
-static inline void deInterlaceBlendLinear(uint8_t src[], int stride)
+static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ src+= 4*stride;
asm volatile(
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
"movq (%0), %%mm0 \n\t" // L0
"movq (%%eax, %1), %%mm1 \n\t" // L2
PAVGB(%%mm2, %%mm1) // L2+L4
PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4
"movq %%mm1, (%%eax, %1) \n\t"
- "movq (%%ebx), %%mm1 \n\t" // L5
+ "movq (%%edx), %%mm1 \n\t" // L5
PAVGB(%%mm1, %%mm0) // L3+L5
PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5
"movq %%mm0, (%%eax, %1, 2) \n\t"
- "movq (%%ebx, %1), %%mm0 \n\t" // L6
+ "movq (%%edx, %1), %%mm0 \n\t" // L6
PAVGB(%%mm0, %%mm2) // L4+L6
PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6
"movq %%mm2, (%0, %1, 4) \n\t"
- "movq (%%ebx, %1, 2), %%mm2 \n\t" // L7
+ "movq (%%edx, %1, 2), %%mm2 \n\t" // L7
PAVGB(%%mm2, %%mm1) // L5+L7
PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7
- "movq %%mm1, (%%ebx) \n\t"
+ "movq %%mm1, (%%edx) \n\t"
"movq (%0, %1, 8), %%mm1 \n\t" // L8
PAVGB(%%mm1, %%mm0) // L6+L8
PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8
- "movq %%mm0, (%%ebx, %1) \n\t"
- "movq (%%ebx, %1, 4), %%mm0 \n\t" // L9
+ "movq %%mm0, (%%edx, %1) \n\t"
+ "movq (%%edx, %1, 4), %%mm0 \n\t" // L9
PAVGB(%%mm0, %%mm2) // L7+L9
PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9
- "movq %%mm2, (%%ebx, %1, 2) \n\t"
+ "movq %%mm2, (%%edx, %1, 2) \n\t"
: : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "%eax", "%edx"
);
#else
int x;
+ src+= 4*stride;
for(x=0; x<8; x++)
{
src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
/**
* Deinterlaces the given block
- * will be called for every 8x8 block, except the last row, and can read & write into an 8x16 block
+ * will be called for every 8x8 block and can read & write from line 4-15,
+ * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
+ * lines 4-12 will be read into the deblocking filter and should be deinterlaced
*/
-static inline void deInterlaceMedian(uint8_t src[], int stride)
+static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
{
#ifdef HAVE_MMX
+ src+= 4*stride;
#ifdef HAVE_MMX2
asm volatile(
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
"movq (%0), %%mm0 \n\t" //
"movq (%%eax, %1), %%mm2 \n\t" //
"pminub %%mm1, %%mm2 \n\t"
"movq %%mm2, (%%eax, %1, 2) \n\t"
- "movq (%%ebx), %%mm2 \n\t" //
- "movq (%%ebx, %1), %%mm1 \n\t" //
+ "movq (%%edx), %%mm2 \n\t" //
+ "movq (%%edx, %1), %%mm1 \n\t" //
"movq %%mm2, %%mm3 \n\t"
"pmaxub %%mm0, %%mm2 \n\t" //
"pminub %%mm3, %%mm0 \n\t" //
"pmaxub %%mm1, %%mm0 \n\t" //
"pminub %%mm0, %%mm2 \n\t"
- "movq %%mm2, (%%ebx) \n\t"
+ "movq %%mm2, (%%edx) \n\t"
- "movq (%%ebx, %1, 2), %%mm2 \n\t" //
+ "movq (%%edx, %1, 2), %%mm2 \n\t" //
"movq (%0, %1, 8), %%mm0 \n\t" //
"movq %%mm2, %%mm3 \n\t"
"pmaxub %%mm0, %%mm2 \n\t" //
"pminub %%mm3, %%mm0 \n\t" //
"pmaxub %%mm1, %%mm0 \n\t" //
"pminub %%mm0, %%mm2 \n\t"
- "movq %%mm2, (%%ebx, %1, 2) \n\t"
+ "movq %%mm2, (%%edx, %1, 2) \n\t"
: : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "%eax", "%edx"
);
#else // MMX without MMX2
asm volatile(
"leal (%0, %1), %%eax \n\t"
- "leal (%%eax, %1, 4), %%ebx \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
// 0 1 2 3 4 5 6 7 8 9
-// %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
"pxor %%mm7, %%mm7 \n\t"
#define MEDIAN(a,b,c)\
MEDIAN((%0), (%%eax), (%%eax, %1))
MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
-MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1))
-MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8))
+MEDIAN((%0, %1, 4), (%%edx), (%%edx, %1))
+MEDIAN((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8))
: : "r" (src), "r" (stride)
- : "%eax", "%ebx"
+ : "%eax", "%edx"
);
#endif // MMX
#else
//FIXME
int x;
+ src+= 4*stride;
for(x=0; x<8; x++)
{
src[0 ] = (src[0 ] + 2*src[stride ] + src[stride*2])>>2;
#endif
}
-#ifdef HAVE_ODIVX_POSTPROCESS
-#include "../opendivx/postprocess.h"
-int use_old_pp=0;
-#endif
-
-static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
- QP_STORE_T QPs[], int QPStride, int isColor, int mode);
-
+#ifdef HAVE_MMX
/**
- * ...
+ * transposes and shift the given 8x8 Block into dst1 and dst2
*/
-void postprocess(unsigned char * src[], int src_stride,
- unsigned char * dst[], int dst_stride,
- int horizontal_size, int vertical_size,
- QP_STORE_T *QP_store, int QP_stride,
- int mode)
+static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
{
+ asm(
+ "leal (%0, %1), %%eax \n\t"
+// 0 1 2 3 4 5 6 7 8 9
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
+ "movq (%0), %%mm0 \n\t" // 12345678
+ "movq (%%eax), %%mm1 \n\t" // abcdefgh
+ "movq %%mm0, %%mm2 \n\t" // 12345678
+ "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
+ "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
+
+ "movq (%%eax, %1), %%mm1 \n\t"
+ "movq (%%eax, %1, 2), %%mm3 \n\t"
+ "movq %%mm1, %%mm4 \n\t"
+ "punpcklbw %%mm3, %%mm1 \n\t"
+ "punpckhbw %%mm3, %%mm4 \n\t"
+
+ "movq %%mm0, %%mm3 \n\t"
+ "punpcklwd %%mm1, %%mm0 \n\t"
+ "punpckhwd %%mm1, %%mm3 \n\t"
+ "movq %%mm2, %%mm1 \n\t"
+ "punpcklwd %%mm4, %%mm2 \n\t"
+ "punpckhwd %%mm4, %%mm1 \n\t"
+
+ "movd %%mm0, 128(%2) \n\t"
+ "psrlq $32, %%mm0 \n\t"
+ "movd %%mm0, 144(%2) \n\t"
+ "movd %%mm3, 160(%2) \n\t"
+ "psrlq $32, %%mm3 \n\t"
+ "movd %%mm3, 176(%2) \n\t"
+ "movd %%mm3, 48(%3) \n\t"
+ "movd %%mm2, 192(%2) \n\t"
+ "movd %%mm2, 64(%3) \n\t"
+ "psrlq $32, %%mm2 \n\t"
+ "movd %%mm2, 80(%3) \n\t"
+ "movd %%mm1, 96(%3) \n\t"
+ "psrlq $32, %%mm1 \n\t"
+ "movd %%mm1, 112(%3) \n\t"
+
+ "leal (%%eax, %1, 4), %%eax \n\t"
+
+ "movq (%0, %1, 4), %%mm0 \n\t" // 12345678
+ "movq (%%eax), %%mm1 \n\t" // abcdefgh
+ "movq %%mm0, %%mm2 \n\t" // 12345678
+ "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
+ "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
+
+ "movq (%%eax, %1), %%mm1 \n\t"
+ "movq (%%eax, %1, 2), %%mm3 \n\t"
+ "movq %%mm1, %%mm4 \n\t"
+ "punpcklbw %%mm3, %%mm1 \n\t"
+ "punpckhbw %%mm3, %%mm4 \n\t"
+
+ "movq %%mm0, %%mm3 \n\t"
+ "punpcklwd %%mm1, %%mm0 \n\t"
+ "punpckhwd %%mm1, %%mm3 \n\t"
+ "movq %%mm2, %%mm1 \n\t"
+ "punpcklwd %%mm4, %%mm2 \n\t"
+ "punpckhwd %%mm4, %%mm1 \n\t"
+
+ "movd %%mm0, 132(%2) \n\t"
+ "psrlq $32, %%mm0 \n\t"
+ "movd %%mm0, 148(%2) \n\t"
+ "movd %%mm3, 164(%2) \n\t"
+ "psrlq $32, %%mm3 \n\t"
+ "movd %%mm3, 180(%2) \n\t"
+ "movd %%mm3, 52(%3) \n\t"
+ "movd %%mm2, 196(%2) \n\t"
+ "movd %%mm2, 68(%3) \n\t"
+ "psrlq $32, %%mm2 \n\t"
+ "movd %%mm2, 84(%3) \n\t"
+ "movd %%mm1, 100(%3) \n\t"
+ "psrlq $32, %%mm1 \n\t"
+ "movd %%mm1, 116(%3) \n\t"
+
+
+ :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
+ : "%eax"
+ );
+}
-#ifdef HAVE_ODIVX_POSTPROCESS
-// Note: I could make this shit outside of this file, but it would mean one
-// more function call...
- if(use_old_pp){
- odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode);
- return;
- }
+/**
+ * transposes the given 8x8 block
+ */
+static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
+{
+ asm(
+ "leal (%0, %1), %%eax \n\t"
+ "leal (%%eax, %1, 4), %%edx \n\t"
+// 0 1 2 3 4 5 6 7 8 9
+// %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1
+ "movq (%2), %%mm0 \n\t" // 12345678
+ "movq 16(%2), %%mm1 \n\t" // abcdefgh
+ "movq %%mm0, %%mm2 \n\t" // 12345678
+ "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
+ "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
+
+ "movq 32(%2), %%mm1 \n\t"
+ "movq 48(%2), %%mm3 \n\t"
+ "movq %%mm1, %%mm4 \n\t"
+ "punpcklbw %%mm3, %%mm1 \n\t"
+ "punpckhbw %%mm3, %%mm4 \n\t"
+
+ "movq %%mm0, %%mm3 \n\t"
+ "punpcklwd %%mm1, %%mm0 \n\t"
+ "punpckhwd %%mm1, %%mm3 \n\t"
+ "movq %%mm2, %%mm1 \n\t"
+ "punpcklwd %%mm4, %%mm2 \n\t"
+ "punpckhwd %%mm4, %%mm1 \n\t"
+
+ "movd %%mm0, (%0) \n\t"
+ "psrlq $32, %%mm0 \n\t"
+ "movd %%mm0, (%%eax) \n\t"
+ "movd %%mm3, (%%eax, %1) \n\t"
+ "psrlq $32, %%mm3 \n\t"
+ "movd %%mm3, (%%eax, %1, 2) \n\t"
+ "movd %%mm2, (%0, %1, 4) \n\t"
+ "psrlq $32, %%mm2 \n\t"
+ "movd %%mm2, (%%edx) \n\t"
+ "movd %%mm1, (%%edx, %1) \n\t"
+ "psrlq $32, %%mm1 \n\t"
+ "movd %%mm1, (%%edx, %1, 2) \n\t"
+
+
+ "movq 64(%2), %%mm0 \n\t" // 12345678
+ "movq 80(%2), %%mm1 \n\t" // abcdefgh
+ "movq %%mm0, %%mm2 \n\t" // 12345678
+ "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d
+ "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h
+
+ "movq 96(%2), %%mm1 \n\t"
+ "movq 112(%2), %%mm3 \n\t"
+ "movq %%mm1, %%mm4 \n\t"
+ "punpcklbw %%mm3, %%mm1 \n\t"
+ "punpckhbw %%mm3, %%mm4 \n\t"
+
+ "movq %%mm0, %%mm3 \n\t"
+ "punpcklwd %%mm1, %%mm0 \n\t"
+ "punpckhwd %%mm1, %%mm3 \n\t"
+ "movq %%mm2, %%mm1 \n\t"
+ "punpcklwd %%mm4, %%mm2 \n\t"
+ "punpckhwd %%mm4, %%mm1 \n\t"
+
+ "movd %%mm0, 4(%0) \n\t"
+ "psrlq $32, %%mm0 \n\t"
+ "movd %%mm0, 4(%%eax) \n\t"
+ "movd %%mm3, 4(%%eax, %1) \n\t"
+ "psrlq $32, %%mm3 \n\t"
+ "movd %%mm3, 4(%%eax, %1, 2) \n\t"
+ "movd %%mm2, 4(%0, %1, 4) \n\t"
+ "psrlq $32, %%mm2 \n\t"
+ "movd %%mm2, 4(%%edx) \n\t"
+ "movd %%mm1, 4(%%edx, %1) \n\t"
+ "psrlq $32, %%mm1 \n\t"
+ "movd %%mm1, 4(%%edx, %1, 2) \n\t"
+
+ :: "r" (dst), "r" (dstStride), "r" (src)
+ : "%eax", "%edx"
+ );
+}
#endif
+//static int test=0;
-/*
- long long T= rdtsc();
- for(int y=vertical_size-1; y>=0 ; y--)
- memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride);
-// memcpy(dst[0], src[0],src_stride*vertical_size);
- printf("%4dk\r", (rdtsc()-T)/1000);
+static void inline RENAME(tempNoiseReducer)(uint8_t *src, int stride,
+ uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
+{
+ // to save a register (FIXME do this outside of the loops)
+ tempBluredPast[127]= maxNoise[0];
+ tempBluredPast[128]= maxNoise[1];
+ tempBluredPast[129]= maxNoise[2];
+
+#define FAST_L2_DIFF
+//#define L1_DIFF //u should change the thresholds too if u try that one
+#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
+ asm volatile(
+ "leal (%2, %2, 2), %%eax \n\t" // 3*stride
+ "leal (%2, %2, 4), %%edx \n\t" // 5*stride
+ "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
+// 0 1 2 3 4 5 6 7 8 9
+// %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2
+//FIXME reorder?
+#ifdef L1_DIFF //needs mmx2
+ "movq (%0), %%mm0 \n\t" // L0
+ "psadbw (%1), %%mm0 \n\t" // |L0-R0|
+ "movq (%0, %2), %%mm1 \n\t" // L1
+ "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1|
+ "movq (%0, %2, 2), %%mm2 \n\t" // L2
+ "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2|
+ "movq (%0, %%eax), %%mm3 \n\t" // L3
+ "psadbw (%1, %%eax), %%mm3 \n\t" // |L3-R3|
+
+ "movq (%0, %2, 4), %%mm4 \n\t" // L4
+ "paddw %%mm1, %%mm0 \n\t"
+ "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4|
+ "movq (%0, %%edx), %%mm5 \n\t" // L5
+ "paddw %%mm2, %%mm0 \n\t"
+ "psadbw (%1, %%edx), %%mm5 \n\t" // |L5-R5|
+ "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
+ "paddw %%mm3, %%mm0 \n\t"
+ "psadbw (%1, %%eax, 2), %%mm6 \n\t" // |L6-R6|
+ "movq (%0, %%ecx), %%mm7 \n\t" // L7
+ "paddw %%mm4, %%mm0 \n\t"
+ "psadbw (%1, %%ecx), %%mm7 \n\t" // |L7-R7|
+ "paddw %%mm5, %%mm6 \n\t"
+ "paddw %%mm7, %%mm6 \n\t"
+ "paddw %%mm6, %%mm0 \n\t"
+#elif defined (FAST_L2_DIFF)
+ "pcmpeqb %%mm7, %%mm7 \n\t"
+ "movq "MANGLE(b80)", %%mm6 \n\t"
+ "pxor %%mm0, %%mm0 \n\t"
+#define L2_DIFF_CORE(a, b)\
+ "movq " #a ", %%mm5 \n\t"\
+ "movq " #b ", %%mm2 \n\t"\
+ "pxor %%mm7, %%mm2 \n\t"\
+ PAVGB(%%mm2, %%mm5)\
+ "paddb %%mm6, %%mm5 \n\t"\
+ "movq %%mm5, %%mm2 \n\t"\
+ "psllw $8, %%mm5 \n\t"\
+ "pmaddwd %%mm5, %%mm5 \n\t"\
+ "pmaddwd %%mm2, %%mm2 \n\t"\
+ "paddd %%mm2, %%mm5 \n\t"\
+ "psrld $14, %%mm5 \n\t"\
+ "paddd %%mm5, %%mm0 \n\t"
+
+L2_DIFF_CORE((%0), (%1))
+L2_DIFF_CORE((%0, %2), (%1, %2))
+L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
+L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
+L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
+L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
+L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
+L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
- return;
-*/
-/*
- long long T= rdtsc();
- while( (rdtsc() - T)/1000 < 4000);
+#else
+ "pxor %%mm7, %%mm7 \n\t"
+ "pxor %%mm0, %%mm0 \n\t"
+#define L2_DIFF_CORE(a, b)\
+ "movq " #a ", %%mm5 \n\t"\
+ "movq " #b ", %%mm2 \n\t"\
+ "movq %%mm5, %%mm1 \n\t"\
+ "movq %%mm2, %%mm3 \n\t"\
+ "punpcklbw %%mm7, %%mm5 \n\t"\
+ "punpckhbw %%mm7, %%mm1 \n\t"\
+ "punpcklbw %%mm7, %%mm2 \n\t"\
+ "punpckhbw %%mm7, %%mm3 \n\t"\
+ "psubw %%mm2, %%mm5 \n\t"\
+ "psubw %%mm3, %%mm1 \n\t"\
+ "pmaddwd %%mm5, %%mm5 \n\t"\
+ "pmaddwd %%mm1, %%mm1 \n\t"\
+ "paddd %%mm1, %%mm5 \n\t"\
+ "paddd %%mm5, %%mm0 \n\t"
+
+L2_DIFF_CORE((%0), (%1))
+L2_DIFF_CORE((%0, %2), (%1, %2))
+L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
+L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
+L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
+L2_DIFF_CORE((%0, %%edx), (%1, %%edx))
+L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
+L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
- return;
-*/
- postProcess(src[0], src_stride, dst[0], dst_stride,
- horizontal_size, vertical_size, QP_store, QP_stride, 0, mode);
+#endif
- horizontal_size >>= 1;
- vertical_size >>= 1;
- src_stride >>= 1;
- dst_stride >>= 1;
- mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00);
+ "movq %%mm0, %%mm4 \n\t"
+ "psrlq $32, %%mm0 \n\t"
+ "paddd %%mm0, %%mm4 \n\t"
+ "movd %%mm4, %%ecx \n\t"
+ "shll $2, %%ecx \n\t"
+ "movl %3, %%edx \n\t"
+ "addl -4(%%edx), %%ecx \n\t"
+ "addl 4(%%edx), %%ecx \n\t"
+ "addl -1024(%%edx), %%ecx \n\t"
+ "addl $4, %%ecx \n\t"
+ "addl 1024(%%edx), %%ecx \n\t"
+ "shrl $3, %%ecx \n\t"
+ "movl %%ecx, (%%edx) \n\t"
+
+// "movl %3, %%ecx \n\t"
+// "movl %%ecx, test \n\t"
+// "jmp 4f \n\t"
+ "cmpl 512(%%edx), %%ecx \n\t"
+ " jb 2f \n\t"
+ "cmpl 516(%%edx), %%ecx \n\t"
+ " jb 1f \n\t"
+
+ "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
+ "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
+ "movq (%0), %%mm0 \n\t" // L0
+ "movq (%0, %2), %%mm1 \n\t" // L1
+ "movq (%0, %2, 2), %%mm2 \n\t" // L2
+ "movq (%0, %%eax), %%mm3 \n\t" // L3
+ "movq (%0, %2, 4), %%mm4 \n\t" // L4
+ "movq (%0, %%edx), %%mm5 \n\t" // L5
+ "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
+ "movq (%0, %%ecx), %%mm7 \n\t" // L7
+ "movq %%mm0, (%1) \n\t" // L0
+ "movq %%mm1, (%1, %2) \n\t" // L1
+ "movq %%mm2, (%1, %2, 2) \n\t" // L2
+ "movq %%mm3, (%1, %%eax) \n\t" // L3
+ "movq %%mm4, (%1, %2, 4) \n\t" // L4
+ "movq %%mm5, (%1, %%edx) \n\t" // L5
+ "movq %%mm6, (%1, %%eax, 2) \n\t" // L6
+ "movq %%mm7, (%1, %%ecx) \n\t" // L7
+ "jmp 4f \n\t"
+
+ "1: \n\t"
+ "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
+ "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
+ "movq (%0), %%mm0 \n\t" // L0
+ PAVGB((%1), %%mm0) // L0
+ "movq (%0, %2), %%mm1 \n\t" // L1
+ PAVGB((%1, %2), %%mm1) // L1
+ "movq (%0, %2, 2), %%mm2 \n\t" // L2
+ PAVGB((%1, %2, 2), %%mm2) // L2
+ "movq (%0, %%eax), %%mm3 \n\t" // L3
+ PAVGB((%1, %%eax), %%mm3) // L3
+ "movq (%0, %2, 4), %%mm4 \n\t" // L4
+ PAVGB((%1, %2, 4), %%mm4) // L4
+ "movq (%0, %%edx), %%mm5 \n\t" // L5
+ PAVGB((%1, %%edx), %%mm5) // L5
+ "movq (%0, %%eax, 2), %%mm6 \n\t" // L6
+ PAVGB((%1, %%eax, 2), %%mm6) // L6
+ "movq (%0, %%ecx), %%mm7 \n\t" // L7
+ PAVGB((%1, %%ecx), %%mm7) // L7
+ "movq %%mm0, (%1) \n\t" // R0
+ "movq %%mm1, (%1, %2) \n\t" // R1
+ "movq %%mm2, (%1, %2, 2) \n\t" // R2
+ "movq %%mm3, (%1, %%eax) \n\t" // R3
+ "movq %%mm4, (%1, %2, 4) \n\t" // R4
+ "movq %%mm5, (%1, %%edx) \n\t" // R5
+ "movq %%mm6, (%1, %%eax, 2) \n\t" // R6
+ "movq %%mm7, (%1, %%ecx) \n\t" // R7
+ "movq %%mm0, (%0) \n\t" // L0
+ "movq %%mm1, (%0, %2) \n\t" // L1
+ "movq %%mm2, (%0, %2, 2) \n\t" // L2
+ "movq %%mm3, (%0, %%eax) \n\t" // L3
+ "movq %%mm4, (%0, %2, 4) \n\t" // L4
+ "movq %%mm5, (%0, %%edx) \n\t" // L5
+ "movq %%mm6, (%0, %%eax, 2) \n\t" // L6
+ "movq %%mm7, (%0, %%ecx) \n\t" // L7
+ "jmp 4f \n\t"
+
+ "2: \n\t"
+ "cmpl 508(%%edx), %%ecx \n\t"
+ " jb 3f \n\t"
+
+ "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
+ "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
+ "movq (%0), %%mm0 \n\t" // L0
+ "movq (%0, %2), %%mm1 \n\t" // L1
+ "movq (%0, %2, 2), %%mm2 \n\t" // L2
+ "movq (%0, %%eax), %%mm3 \n\t" // L3
+ "movq (%1), %%mm4 \n\t" // R0
+ "movq (%1, %2), %%mm5 \n\t" // R1
+ "movq (%1, %2, 2), %%mm6 \n\t" // R2
+ "movq (%1, %%eax), %%mm7 \n\t" // R3
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ "movq %%mm0, (%1) \n\t" // R0
+ "movq %%mm1, (%1, %2) \n\t" // R1
+ "movq %%mm2, (%1, %2, 2) \n\t" // R2
+ "movq %%mm3, (%1, %%eax) \n\t" // R3
+ "movq %%mm0, (%0) \n\t" // L0
+ "movq %%mm1, (%0, %2) \n\t" // L1
+ "movq %%mm2, (%0, %2, 2) \n\t" // L2
+ "movq %%mm3, (%0, %%eax) \n\t" // L3
+
+ "movq (%0, %2, 4), %%mm0 \n\t" // L4
+ "movq (%0, %%edx), %%mm1 \n\t" // L5
+ "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
+ "movq (%0, %%ecx), %%mm3 \n\t" // L7
+ "movq (%1, %2, 4), %%mm4 \n\t" // R4
+ "movq (%1, %%edx), %%mm5 \n\t" // R5
+ "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
+ "movq (%1, %%ecx), %%mm7 \n\t" // R7
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ "movq %%mm0, (%1, %2, 4) \n\t" // R4
+ "movq %%mm1, (%1, %%edx) \n\t" // R5
+ "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
+ "movq %%mm3, (%1, %%ecx) \n\t" // R7
+ "movq %%mm0, (%0, %2, 4) \n\t" // L4
+ "movq %%mm1, (%0, %%edx) \n\t" // L5
+ "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
+ "movq %%mm3, (%0, %%ecx) \n\t" // L7
+ "jmp 4f \n\t"
+
+ "3: \n\t"
+ "leal (%%eax, %2, 2), %%edx \n\t" // 5*stride
+ "leal (%%edx, %2, 2), %%ecx \n\t" // 7*stride
+ "movq (%0), %%mm0 \n\t" // L0
+ "movq (%0, %2), %%mm1 \n\t" // L1
+ "movq (%0, %2, 2), %%mm2 \n\t" // L2
+ "movq (%0, %%eax), %%mm3 \n\t" // L3
+ "movq (%1), %%mm4 \n\t" // R0
+ "movq (%1, %2), %%mm5 \n\t" // R1
+ "movq (%1, %2, 2), %%mm6 \n\t" // R2
+ "movq (%1, %%eax), %%mm7 \n\t" // R3
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ "movq %%mm0, (%1) \n\t" // R0
+ "movq %%mm1, (%1, %2) \n\t" // R1
+ "movq %%mm2, (%1, %2, 2) \n\t" // R2
+ "movq %%mm3, (%1, %%eax) \n\t" // R3
+ "movq %%mm0, (%0) \n\t" // L0
+ "movq %%mm1, (%0, %2) \n\t" // L1
+ "movq %%mm2, (%0, %2, 2) \n\t" // L2
+ "movq %%mm3, (%0, %%eax) \n\t" // L3
+
+ "movq (%0, %2, 4), %%mm0 \n\t" // L4
+ "movq (%0, %%edx), %%mm1 \n\t" // L5
+ "movq (%0, %%eax, 2), %%mm2 \n\t" // L6
+ "movq (%0, %%ecx), %%mm3 \n\t" // L7
+ "movq (%1, %2, 4), %%mm4 \n\t" // R4
+ "movq (%1, %%edx), %%mm5 \n\t" // R5
+ "movq (%1, %%eax, 2), %%mm6 \n\t" // R6
+ "movq (%1, %%ecx), %%mm7 \n\t" // R7
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ PAVGB(%%mm4, %%mm0)
+ PAVGB(%%mm5, %%mm1)
+ PAVGB(%%mm6, %%mm2)
+ PAVGB(%%mm7, %%mm3)
+ "movq %%mm0, (%1, %2, 4) \n\t" // R4
+ "movq %%mm1, (%1, %%edx) \n\t" // R5
+ "movq %%mm2, (%1, %%eax, 2) \n\t" // R6
+ "movq %%mm3, (%1, %%ecx) \n\t" // R7
+ "movq %%mm0, (%0, %2, 4) \n\t" // L4
+ "movq %%mm1, (%0, %%edx) \n\t" // L5
+ "movq %%mm2, (%0, %%eax, 2) \n\t" // L6
+ "movq %%mm3, (%0, %%ecx) \n\t" // L7
+
+ "4: \n\t"
+
+ :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)
+ : "%eax", "%edx", "%ecx", "memory"
+ );
+//printf("%d\n", test);
+#else
+{
+ int y;
+ int d=0;
+ int sysd=0;
+ int i;
- if(1)
+ for(y=0; y<8; y++)
{
- postProcess(src[1], src_stride, dst[1], dst_stride,
- horizontal_size, vertical_size, QP_store, QP_stride, 1, mode);
- postProcess(src[2], src_stride, dst[2], dst_stride,
- horizontal_size, vertical_size, QP_store, QP_stride, 1, mode);
+ int x;
+ for(x=0; x<8; x++)
+ {
+ int ref= tempBlured[ x + y*stride ];
+ int cur= src[ x + y*stride ];
+ int d1=ref - cur;
+// if(x==0 || x==7) d1+= d1>>1;
+// if(y==0 || y==7) d1+= d1>>1;
+// d+= ABS(d1);
+ d+= d1*d1;
+ sysd+= d1;
+ }
+ }
+ i=d;
+ d= (
+ 4*d
+ +(*(tempBluredPast-256))
+ +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
+ +(*(tempBluredPast+256))
+ +4)>>3;
+ *tempBluredPast=i;
+// ((*tempBluredPast)*3 + d + 2)>>2;
+
+//printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);
+/*
+Switch between
+ 1 0 0 0 0 0 0 (0)
+64 32 16 8 4 2 1 (1)
+64 48 36 27 20 15 11 (33) (approx)
+64 56 49 43 37 33 29 (200) (approx)
+*/
+ if(d > maxNoise[1])
+ {
+ if(d < maxNoise[2])
+ {
+ for(y=0; y<8; y++)
+ {
+ int x;
+ for(x=0; x<8; x++)
+ {
+ int ref= tempBlured[ x + y*stride ];
+ int cur= src[ x + y*stride ];
+ tempBlured[ x + y*stride ]=
+ src[ x + y*stride ]=
+ (ref + cur + 1)>>1;
+ }
+ }
+ }
+ else
+ {
+ for(y=0; y<8; y++)
+ {
+ int x;
+ for(x=0; x<8; x++)
+ {
+ tempBlured[ x + y*stride ]= src[ x + y*stride ];
+ }
+ }
+ }
}
else
{
- memcpy(dst[1], src[1], src_stride*horizontal_size);
- memcpy(dst[2], src[2], src_stride*horizontal_size);
+ if(d < maxNoise[0])
+ {
+ for(y=0; y<8; y++)
+ {
+ int x;
+ for(x=0; x<8; x++)
+ {
+ int ref= tempBlured[ x + y*stride ];
+ int cur= src[ x + y*stride ];
+ tempBlured[ x + y*stride ]=
+ src[ x + y*stride ]=
+ (ref*7 + cur + 4)>>3;
+ }
+ }
+ }
+ else
+ {
+ for(y=0; y<8; y++)
+ {
+ int x;
+ for(x=0; x<8; x++)
+ {
+ int ref= tempBlured[ x + y*stride ];
+ int cur= src[ x + y*stride ];
+ tempBlured[ x + y*stride ]=
+ src[ x + y*stride ]=
+ (ref*3 + cur + 2)>>2;
+ }
+ }
+ }
}
}
-
-/**
- * gets the mode flags for a given quality (larger values mean slower but better postprocessing)
- * 0 <= quality <= 6
- */
-int getPpModeForQuality(int quality){
- int modes[1+GET_PP_QUALITY_MAX]= {
- 0,
-#if 1
- // horizontal filters first
- LUM_H_DEBLOCK,
- LUM_H_DEBLOCK | LUM_V_DEBLOCK,
- LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK,
- LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK,
- LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING,
- LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING
-#else
- // vertical filters first
- LUM_V_DEBLOCK,
- LUM_V_DEBLOCK | LUM_H_DEBLOCK,
- LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK,
- LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK,
- LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING,
- LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING
#endif
- };
-
-#ifdef HAVE_ODIVX_POSTPROCESS
- int odivx_modes[1+GET_PP_QUALITY_MAX]= {
- 0,
- PP_DEBLOCK_Y_H,
- PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V,
- PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H,
- PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V,
- PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y,
- PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C
- };
- if(use_old_pp) return odivx_modes[quality];
-#endif
- return modes[quality];
}
-//} // extern "C"
+static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
+ QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);
/**
* Copies a block from src to dst and fixes the blacklevel
- * numLines must be a multiple of 4
* levelFix == 0 -> dont touch the brighness & contrast
*/
-static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
- int numLines, int levelFix)
+#undef SCALED_CPY
+
+static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
+ int levelFix, int64_t *packedOffsetAndScale)
{
+#ifndef HAVE_MMX
int i;
+#endif
if(levelFix)
{
#ifdef HAVE_MMX
asm volatile(
- "movl %4, %%eax \n\t"
- "movl %%eax, temp0\n\t"
- "pushl %0 \n\t"
- "pushl %1 \n\t"
- "leal (%2,%2), %%eax \n\t"
- "leal (%3,%3), %%ebx \n\t"
- "movq packedYOffset, %%mm2 \n\t"
- "movq packedYScale, %%mm3 \n\t"
+ "movq (%%eax), %%mm2 \n\t" // packedYOffset
+ "movq 8(%%eax), %%mm3 \n\t" // packedYScale
+ "leal (%2,%4), %%eax \n\t"
+ "leal (%3,%5), %%edx \n\t"
"pxor %%mm4, %%mm4 \n\t"
-
-#define SCALED_CPY \
- "movq (%0), %%mm0 \n\t"\
- "movq (%0,%2), %%mm1 \n\t"\
- "psubusb %%mm2, %%mm0 \n\t"\
- "psubusb %%mm2, %%mm1 \n\t"\
- "movq %%mm0, %%mm5 \n\t"\
+#ifdef HAVE_MMX2
+#define SCALED_CPY(src1, src2, dst1, dst2) \
+ "movq " #src1 ", %%mm0 \n\t"\
+ "movq " #src1 ", %%mm5 \n\t"\
+ "movq " #src2 ", %%mm1 \n\t"\
+ "movq " #src2 ", %%mm6 \n\t"\
+ "punpcklbw %%mm0, %%mm0 \n\t"\
+ "punpckhbw %%mm5, %%mm5 \n\t"\
+ "punpcklbw %%mm1, %%mm1 \n\t"\
+ "punpckhbw %%mm6, %%mm6 \n\t"\
+ "pmulhuw %%mm3, %%mm0 \n\t"\
+ "pmulhuw %%mm3, %%mm5 \n\t"\
+ "pmulhuw %%mm3, %%mm1 \n\t"\
+ "pmulhuw %%mm3, %%mm6 \n\t"\
+ "psubw %%mm2, %%mm0 \n\t"\
+ "psubw %%mm2, %%mm5 \n\t"\
+ "psubw %%mm2, %%mm1 \n\t"\
+ "psubw %%mm2, %%mm6 \n\t"\
+ "packuswb %%mm5, %%mm0 \n\t"\
+ "packuswb %%mm6, %%mm1 \n\t"\
+ "movq %%mm0, " #dst1 " \n\t"\
+ "movq %%mm1, " #dst2 " \n\t"\
+
+#else //HAVE_MMX2
+#define SCALED_CPY(src1, src2, dst1, dst2) \
+ "movq " #src1 ", %%mm0 \n\t"\
+ "movq " #src1 ", %%mm5 \n\t"\
"punpcklbw %%mm4, %%mm0 \n\t"\
"punpckhbw %%mm4, %%mm5 \n\t"\
- "psllw $7, %%mm0 \n\t"\
- "psllw $7, %%mm5 \n\t"\
+ "psubw %%mm2, %%mm0 \n\t"\
+ "psubw %%mm2, %%mm5 \n\t"\
+ "movq " #src2 ", %%mm1 \n\t"\
+ "psllw $6, %%mm0 \n\t"\
+ "psllw $6, %%mm5 \n\t"\
"pmulhw %%mm3, %%mm0 \n\t"\
+ "movq " #src2 ", %%mm6 \n\t"\
"pmulhw %%mm3, %%mm5 \n\t"\
- "packuswb %%mm5, %%mm0 \n\t"\
- "movq %%mm0, (%1) \n\t"\
- "movq %%mm1, %%mm5 \n\t"\
"punpcklbw %%mm4, %%mm1 \n\t"\
- "punpckhbw %%mm4, %%mm5 \n\t"\
- "psllw $7, %%mm1 \n\t"\
- "psllw $7, %%mm5 \n\t"\
+ "punpckhbw %%mm4, %%mm6 \n\t"\
+ "psubw %%mm2, %%mm1 \n\t"\
+ "psubw %%mm2, %%mm6 \n\t"\
+ "psllw $6, %%mm1 \n\t"\
+ "psllw $6, %%mm6 \n\t"\
"pmulhw %%mm3, %%mm1 \n\t"\
- "pmulhw %%mm3, %%mm5 \n\t"\
- "packuswb %%mm5, %%mm1 \n\t"\
- "movq %%mm1, (%1, %3) \n\t"\
-
- "1: \n\t"
-SCALED_CPY
- "addl %%eax, %0 \n\t"
- "addl %%ebx, %1 \n\t"
-SCALED_CPY
- "addl %%eax, %0 \n\t"
- "addl %%ebx, %1 \n\t"
- "decl temp0 \n\t"
- "jnz 1b \n\t"
-
- "popl %1 \n\t"
- "popl %0 \n\t"
- : : "r" (src),
- "r" (dst),
+ "pmulhw %%mm3, %%mm6 \n\t"\
+ "packuswb %%mm5, %%mm0 \n\t"\
+ "packuswb %%mm6, %%mm1 \n\t"\
+ "movq %%mm0, " #dst1 " \n\t"\
+ "movq %%mm1, " #dst2 " \n\t"\
+
+#endif //!HAVE_MMX2
+
+SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5))
+SCALED_CPY((%2, %4, 2), (%%eax, %4, 2), (%3, %5, 2), (%%edx, %5, 2))
+SCALED_CPY((%2, %4, 4), (%%eax, %4, 4), (%3, %5, 4), (%%edx, %5, 4))
+ "leal (%%eax,%4,4), %%eax \n\t"
+ "leal (%%edx,%5,4), %%edx \n\t"
+SCALED_CPY((%%eax, %4), (%%eax, %4, 2), (%%edx, %5), (%%edx, %5, 2))
+
+
+ : "=&a" (packedOffsetAndScale)
+ : "0" (packedOffsetAndScale),
+ "r"(src),
+ "r"(dst),
"r" (srcStride),
- "r" (dstStride),
- "m" (numLines>>2)
- : "%eax", "%ebx"
+ "r" (dstStride)
+ : "%edx"
);
#else
- for(i=0; i<numLines; i++)
+ for(i=0; i<8; i++)
memcpy( &(dst[dstStride*i]),
&(src[srcStride*i]), BLOCK_SIZE);
#endif
{
#ifdef HAVE_MMX
asm volatile(
- "movl %4, %%eax \n\t"
- "movl %%eax, temp0\n\t"
- "pushl %0 \n\t"
- "pushl %1 \n\t"
- "leal (%2,%2), %%eax \n\t"
- "leal (%3,%3), %%ebx \n\t"
- "movq packedYOffset, %%mm2 \n\t"
- "movq packedYScale, %%mm3 \n\t"
-
-#define SIMPLE_CPY \
- "movq (%0), %%mm0 \n\t"\
- "movq (%0,%2), %%mm1 \n\t"\
- "movq %%mm0, (%1) \n\t"\
- "movq %%mm1, (%1, %3) \n\t"\
-
- "1: \n\t"
-SIMPLE_CPY
- "addl %%eax, %0 \n\t"
- "addl %%ebx, %1 \n\t"
-SIMPLE_CPY
- "addl %%eax, %0 \n\t"
- "addl %%ebx, %1 \n\t"
- "decl temp0 \n\t"
- "jnz 1b \n\t"
-
- "popl %1 \n\t"
- "popl %0 \n\t"
+ "leal (%0,%2), %%eax \n\t"
+ "leal (%1,%3), %%edx \n\t"
+
+#define SIMPLE_CPY(src1, src2, dst1, dst2) \
+ "movq " #src1 ", %%mm0 \n\t"\
+ "movq " #src2 ", %%mm1 \n\t"\
+ "movq %%mm0, " #dst1 " \n\t"\
+ "movq %%mm1, " #dst2 " \n\t"\
+
+SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3))
+SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%edx, %3, 2))
+SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%edx, %3, 4))
+ "leal (%%eax,%2,4), %%eax \n\t"
+ "leal (%%edx,%3,4), %%edx \n\t"
+SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%edx, %3), (%%edx, %3, 2))
+
: : "r" (src),
"r" (dst),
"r" (srcStride),
- "r" (dstStride),
- "m" (numLines>>2)
- : "%eax", "%ebx"
+ "r" (dstStride)
+ : "%eax", "%edx"
);
#else
- for(i=0; i<numLines; i++)
+ for(i=0; i<8; i++)
memcpy( &(dst[dstStride*i]),
&(src[srcStride*i]), BLOCK_SIZE);
#endif
}
}
+/**
+ * Duplicates the given 8 src pixels ? times upward
+ */
+static inline void RENAME(duplicate)(uint8_t src[], int stride)
+{
+#ifdef HAVE_MMX
+ asm volatile(
+ "movq (%0), %%mm0 \n\t"
+ "addl %1, %0 \n\t"
+ "movq %%mm0, (%0) \n\t"
+ "movq %%mm0, (%0, %1) \n\t"
+ "movq %%mm0, (%0, %1, 2) \n\t"
+ : "+r" (src)
+ : "r" (-stride)
+ );
+#else
+ int i;
+ uint8_t *p=src;
+ for(i=0; i<3; i++)
+ {
+ p-= stride;
+ memcpy(p, src, 8);
+ }
+#endif
+}
/**
* Filters array of bytes (Y or U or V values)
*/
-static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
- QP_STORE_T QPs[], int QPStride, int isColor, int mode)
+static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
+ QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)
{
+ PPContext __attribute__((aligned(8))) c= *c2; //copy to stack for faster access
int x,y;
- /* we need 64bit here otherwise weĀ“ll going to have a problem
- after watching a black picture for 5 hours*/
- static uint64_t *yHistogram= NULL;
+#ifdef COMPILE_TIME_MODE
+ const int mode= COMPILE_TIME_MODE;
+#else
+ const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;
+#endif
int black=0, white=255; // blackest black and whitest white in the picture
+ int QPCorrecture= 256*256;
- /* Temporary buffers for handling the last row(s) */
- static uint8_t *tempDst= NULL;
- static uint8_t *tempSrc= NULL;
-
- /* Temporary buffers for handling the last block */
- static uint8_t *tempDstBlock= NULL;
- static uint8_t *tempSrcBlock= NULL;
-
- uint8_t *dstBlockPtrBackup;
- uint8_t *srcBlockPtrBackup;
-
-#ifdef TIMING
- long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0;
- sumTime= rdtsc();
+ int copyAhead;
+#ifdef HAVE_MMX
+ int i;
#endif
- if(tempDst==NULL)
- {
- tempDst= (uint8_t*)memalign(8, 1024*24);
- tempSrc= (uint8_t*)memalign(8, 1024*24);
- tempDstBlock= (uint8_t*)memalign(8, 1024*24);
- tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
+ //FIXME remove
+ uint64_t * const yHistogram= c.yHistogram;
+ uint8_t * const tempSrc= c.tempSrc;
+ uint8_t * const tempDst= c.tempDst;
+ const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;
+
+#ifdef HAVE_MMX
+ for(i=0; i<32; i++){
+ int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;
+ int threshold= offset*2 + 1;
+ c.mmxDcOffset[i]= 0x7F - offset;
+ c.mmxDcThreshold[i]= 0x7F - threshold;
+ c.mmxDcOffset[i]*= 0x0101010101010101LL;
+ c.mmxDcThreshold[i]*= 0x0101010101010101LL;
}
+#endif
- if(!yHistogram)
- {
- int i;
- yHistogram= (uint64_t*)malloc(8*256);
- for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;
+ if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
+ else if( (mode & LINEAR_BLEND_DEINT_FILTER)
+ || (mode & FFMPEG_DEINT_FILTER)) copyAhead=14;
+ else if( (mode & V_DEBLOCK)
+ || (mode & LINEAR_IPOL_DEINT_FILTER)
+ || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
+ else if(mode & V_X1_FILTER) copyAhead=11;
+// else if(mode & V_RK1_FILTER) copyAhead=10;
+ else if(mode & DERING) copyAhead=9;
+ else copyAhead=8;
- if(mode & FULL_Y_RANGE)
- {
- maxAllowedY=255;
- minAllowedY=0;
- }
- }
+ copyAhead-= 8;
if(!isColor)
{
uint64_t sum= 0;
int i;
- static int framenum= -1;
uint64_t maxClipped;
uint64_t clipped;
double scale;
- framenum++;
- if(framenum == 1) yHistogram[0]= width*height/64*15/256;
+ c.frameNum++;
+ // first frame is fscked so we ignore it
+ if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;
for(i=0; i<256; i++)
{
// printf("\n\n");
/* we allways get a completly black picture first */
- maxClipped= (uint64_t)(sum * maxClippedThreshold);
+ maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);
clipped= sum;
for(black=255; black>0; black--)
clipped-= yHistogram[white];
}
- // we cant handle negative correctures
- packedYOffset= MAX(black - minAllowedY, 0);
- packedYOffset|= packedYOffset<<32;
- packedYOffset|= packedYOffset<<16;
- packedYOffset|= packedYOffset<<8;
+ scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);
+
+#ifdef HAVE_MMX2
+ c.packedYScale= (uint16_t)(scale*256.0 + 0.5);
+ c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;
+#else
+ c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);
+ c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;
+#endif
- scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black);
+ c.packedYOffset|= c.packedYOffset<<32;
+ c.packedYOffset|= c.packedYOffset<<16;
- packedYScale= (uint16_t)(scale*512.0 + 0.5);
- packedYScale|= packedYScale<<32;
- packedYScale|= packedYScale<<16;
+ c.packedYScale|= c.packedYScale<<32;
+ c.packedYScale|= c.packedYScale<<16;
+
+ if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5);
+ else QPCorrecture= 256*256;
}
else
{
- packedYScale= 0x0100010001000100LL;
- packedYOffset= 0;
+ c.packedYScale= 0x0100010001000100LL;
+ c.packedYOffset= 0;
+ QPCorrecture= 256*256;
}
- /* copy first row of 8x8 blocks */
- for(x=0; x<width; x+=BLOCK_SIZE)
- blockCopy(dst + x, dstStride, src + x, srcStride, 8, mode & LEVEL_FIX);
+ /* copy & deinterlace first row of blocks */
+ y=-BLOCK_SIZE;
+ {
+ uint8_t *srcBlock= &(src[y*srcStride]);
+ uint8_t *dstBlock= tempDst + dstStride;
+
+ // From this point on it is guranteed that we can read and write 16 lines downward
+ // finish 1 block before the next otherwise weĀ“ll might have a problem
+ // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
+ for(x=0; x<width; x+=BLOCK_SIZE)
+ {
+
+#ifdef HAVE_MMX2
+/*
+ prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
+ prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
+ prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
+ prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
+*/
+
+ asm(
+ "movl %4, %%eax \n\t"
+ "shrl $2, %%eax \n\t"
+ "andl $6, %%eax \n\t"
+ "addl %5, %%eax \n\t"
+ "movl %%eax, %%edx \n\t"
+ "imul %1, %%eax \n\t"
+ "imul %3, %%edx \n\t"
+ "prefetchnta 32(%%eax, %0) \n\t"
+ "prefetcht0 32(%%edx, %2) \n\t"
+ "addl %1, %%eax \n\t"
+ "addl %3, %%edx \n\t"
+ "prefetchnta 32(%%eax, %0) \n\t"
+ "prefetcht0 32(%%edx, %2) \n\t"
+ :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
+ "m" (x), "m" (copyAhead)
+ : "%eax", "%edx"
+ );
+
+#elif defined(HAVE_3DNOW)
+//FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
+/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
+ prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
+ prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
+ prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
+*/
+#endif
+
+ RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
+ srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
+ RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
+
+ if(mode & LINEAR_IPOL_DEINT_FILTER)
+ RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
+ else if(mode & LINEAR_BLEND_DEINT_FILTER)
+ RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
+ else if(mode & MEDIAN_DEINT_FILTER)
+ RENAME(deInterlaceMedian)(dstBlock, dstStride);
+ else if(mode & CUBIC_IPOL_DEINT_FILTER)
+ RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
+ else if(mode & FFMPEG_DEINT_FILTER)
+ RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
+/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
+ RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
+*/
+ dstBlock+=8;
+ srcBlock+=8;
+ }
+ memcpy(dst, tempDst + 9*dstStride, copyAhead*dstStride );
+ }
+
+//printf("\n");
for(y=0; y<height; y+=BLOCK_SIZE)
{
//1% speedup if these are here instead of the inner loop
uint8_t *srcBlock= &(src[y*srcStride]);
uint8_t *dstBlock= &(dst[y*dstStride]);
-
- /* can we mess with a 8x16 block from srcBlock/dstBlock downwards, if not
- than use a temporary buffer */
+#ifdef HAVE_MMX
+ uint8_t *tempBlock1= c.tempBlocks;
+ uint8_t *tempBlock2= c.tempBlocks + 8;
+#endif
+ int8_t *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
+ int8_t *nonBQPptr= isColor ? &c.nonBQPTable[(y>>3)*mbWidth] :&c.nonBQPTable[(y>>4)*mbWidth];
+ int QP=0;
+ /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
+ if not than use a temporary buffer */
if(y+15 >= height)
{
- /* copy from line 5 to 12 of src, these will e copied with
+ int i;
+ /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
blockcopy to dst later */
- memcpy(tempSrc + srcStride*5, srcBlock + srcStride*5,
- srcStride*MAX(height-y-5, 0) );
+ memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
+ srcStride*MAX(height-y-copyAhead, 0) );
- /* duplicate last line to fill the void upto line 12 */
- if(y+12 >= height)
- {
- int i;
- for(i=height-y; i<=12; i++)
- memcpy(tempSrc + srcStride*i,
- src + srcStride*(height-1), srcStride);
- }
+ /* duplicate last line of src to fill the void upto line (copyAhead+7) */
+ for(i=MAX(height-y, 8); i<copyAhead+8; i++)
+ memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
+
+ /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
+ memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );
+ /* duplicate last line of dst to fill the void upto line (copyAhead) */
+ for(i=height-y+1; i<=copyAhead; i++)
+ memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
- /* copy up to 5 lines of dst */
- memcpy(tempDst, dstBlock, dstStride*MIN(height-y, 5) );
- dstBlock= tempDst;
+ dstBlock= tempDst + dstStride;
srcBlock= tempSrc;
}
+//printf("\n");
// From this point on it is guranteed that we can read and write 16 lines downward
// finish 1 block before the next otherwise weĀ“ll might have a problem
for(x=0; x<width; x+=BLOCK_SIZE)
{
const int stride= dstStride;
- int QP= isColor ?
- QPs[(y>>3)*QPStride + (x>>3)]:
- QPs[(y>>4)*QPStride + (x>>4)];
- if(!isColor && (mode & LEVEL_FIX)) QP= (QP* (packedYScale &0xFFFF))>>8;
+#ifdef HAVE_MMX
+ uint8_t *tmpXchg;
+#endif
+ if(isColor)
+ {
+ QP= QPptr[x>>3];
+ c.nonBQP= nonBQPptr[x>>3];
+ }
+ else
+ {
+ QP= QPptr[x>>4];
+ QP= (QP* QPCorrecture + 256*128)>>16;
+ c.nonBQP= nonBQPptr[x>>4];
+ c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;
+ yHistogram[ srcBlock[srcStride*12 + 4] ]++;
+ }
+ c.QP= QP;
#ifdef HAVE_MMX
asm volatile(
- "movd %0, %%mm7 \n\t"
+ "movd %1, %%mm7 \n\t"
"packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
"packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
"packuswb %%mm7, %%mm7 \n\t" // QP,..., QP
- "movq %%mm7, pQPb \n\t"
- : : "r" (QP)
+ "movq %%mm7, %0 \n\t"
+ : "=m" (c.pQPb)
+ : "r" (QP)
);
#endif
-#ifdef MORE_TIMING
- T0= rdtsc();
-#endif
#ifdef HAVE_MMX2
- prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
- prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
- prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
- prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
+/*
+ prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
+ prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
+ prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
+ prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
+*/
+
+ asm(
+ "movl %4, %%eax \n\t"
+ "shrl $2, %%eax \n\t"
+ "andl $6, %%eax \n\t"
+ "addl %5, %%eax \n\t"
+ "movl %%eax, %%edx \n\t"
+ "imul %1, %%eax \n\t"
+ "imul %3, %%edx \n\t"
+ "prefetchnta 32(%%eax, %0) \n\t"
+ "prefetcht0 32(%%edx, %2) \n\t"
+ "addl %1, %%eax \n\t"
+ "addl %3, %%edx \n\t"
+ "prefetchnta 32(%%eax, %0) \n\t"
+ "prefetcht0 32(%%edx, %2) \n\t"
+ :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
+ "m" (x), "m" (copyAhead)
+ : "%eax", "%edx"
+ );
+
#elif defined(HAVE_3DNOW)
//FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
/* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
*/
#endif
- if(!isColor) yHistogram[ srcBlock[srcStride*5] ]++;
-
-#ifdef PP_FUNNY_STRIDE
- //can we mess with a 8x16 block, if not use a temp buffer, yes again
- if(x+7 >= width)
- {
- int i;
- dstBlockPtrBackup= dstBlock;
- srcBlockPtrBackup= srcBlock;
-
- for(i=0;i<BLOCK_SIZE*2; i++)
- {
- memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x);
- memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x);
- }
-
- dstBlock= tempDstBlock;
- srcBlock= tempSrcBlock;
- }
-#endif
-
- blockCopy(dstBlock + dstStride*5, dstStride,
- srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX);
+ RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
+ srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);
if(mode & LINEAR_IPOL_DEINT_FILTER)
- deInterlaceInterpolateLinear(dstBlock, dstStride);
+ RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
else if(mode & LINEAR_BLEND_DEINT_FILTER)
- deInterlaceBlendLinear(dstBlock, dstStride);
+ RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
else if(mode & MEDIAN_DEINT_FILTER)
- deInterlaceMedian(dstBlock, dstStride);
+ RENAME(deInterlaceMedian)(dstBlock, dstStride);
else if(mode & CUBIC_IPOL_DEINT_FILTER)
- deInterlaceInterpolateCubic(dstBlock, dstStride);
+ RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
+ else if(mode & FFMPEG_DEINT_FILTER)
+ RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);
/* else if(mode & CUBIC_BLEND_DEINT_FILTER)
- deInterlaceBlendCubic(dstBlock, dstStride);
+ RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
*/
/* only deblock if we have 2 blocks */
if(y + 8 < height)
{
-#ifdef MORE_TIMING
- T1= rdtsc();
- memcpyTime+= T1-T0;
- T0=T1;
-#endif
- if(mode & V_RK1_FILTER)
- vertRK1Filter(dstBlock, stride, QP);
- else if(mode & V_X1_FILTER)
- vertX1Filter(dstBlock, stride, QP);
+ if(mode & V_X1_FILTER)
+ RENAME(vertX1Filter)(dstBlock, stride, &c);
else if(mode & V_DEBLOCK)
{
- if( isVertDC(dstBlock, stride))
+ if( RENAME(isVertDC)(dstBlock, stride, &c))
{
- if(isVertMinMaxOk(dstBlock, stride, QP))
- doVertLowPass(dstBlock, stride, QP);
+ if(RENAME(isVertMinMaxOk)(dstBlock, stride, &c))
+ RENAME(doVertLowPass)(dstBlock, stride, &c);
}
else
- doVertDefFilter(dstBlock, stride, QP);
+ RENAME(doVertDefFilter)(dstBlock, stride, &c);
}
-#ifdef MORE_TIMING
- T1= rdtsc();
- vertTime+= T1-T0;
- T0=T1;
-#endif
}
+#ifdef HAVE_MMX
+ RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
+#endif
/* check if we have a previous block to deblock it with dstBlock */
if(x - 8 >= 0)
{
-#ifdef MORE_TIMING
- T0= rdtsc();
-#endif
+#ifdef HAVE_MMX
+ if(mode & H_X1_FILTER)
+ RENAME(vertX1Filter)(tempBlock1, 16, &c);
+ else if(mode & H_DEBLOCK)
+ {
+ if( RENAME(isVertDC)(tempBlock1, 16, &c))
+ {
+ if(RENAME(isVertMinMaxOk)(tempBlock1, 16, &c))
+ RENAME(doVertLowPass)(tempBlock1, 16, &c);
+ }
+ else
+ RENAME(doVertDefFilter)(tempBlock1, 16, &c);
+ }
+
+ RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
+
+#else
if(mode & H_X1_FILTER)
horizX1Filter(dstBlock-4, stride, QP);
else if(mode & H_DEBLOCK)
{
- if( isHorizDCAndCopy2Temp(dstBlock-4, stride))
+ if( isHorizDC(dstBlock-4, stride, &c))
{
- if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP))
- doHorizLowPassAndCopyBack(dstBlock-4, stride, QP);
+ if(isHorizMinMaxOk(dstBlock-4, stride, QP))
+ doHorizLowPass(dstBlock-4, stride, QP);
}
else
- doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP);
+ doHorizDefFilter(dstBlock-4, stride, QP);
}
-#ifdef MORE_TIMING
- T1= rdtsc();
- horizTime+= T1-T0;
- T0=T1;
#endif
- dering(dstBlock - 9 - stride, stride, QP);
- }
- else if(y!=0)
- dering(dstBlock - stride*9 + width-9, stride, QP);
- //FIXME dering filter will not be applied to last block (bottom right)
-
-#ifdef PP_FUNNY_STRIDE
- /* did we use a tmp-block buffer */
- if(x+7 >= width)
- {
- int i;
- dstBlock= dstBlockPtrBackup;
- srcBlock= srcBlockPtrBackup;
+ if(mode & DERING)
+ {
+ //FIXME filter first line
+ if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);
+ }
- for(i=0;i<BLOCK_SIZE*2; i++)
+ if(mode & TEMP_NOISE_FILTER)
{
- memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x);
+ RENAME(tempNoiseReducer)(dstBlock-8, stride,
+ c.tempBlured[isColor] + y*dstStride + x,
+ c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
+ c.ppMode.maxTmpNoise);
}
}
-#endif
dstBlock+=8;
srcBlock+=8;
+
+#ifdef HAVE_MMX
+ tmpXchg= tempBlock1;
+ tempBlock1= tempBlock2;
+ tempBlock2 = tmpXchg;
+#endif
+ }
+
+ if(mode & DERING)
+ {
+ if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);
}
- /* did we use a tmp buffer */
+ if((mode & TEMP_NOISE_FILTER))
+ {
+ RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
+ c.tempBlured[isColor] + y*dstStride + x,
+ c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
+ c.ppMode.maxTmpNoise);
+ }
+
+ /* did we use a tmp buffer for the last lines*/
if(y+15 >= height)
{
uint8_t *dstBlock= &(dst[y*dstStride]);
- memcpy(dstBlock, tempDst, dstStride*(height-y) );
+ memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
}
+/*
+ for(x=0; x<width; x+=32)
+ {
+ volatile int i;
+ i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
+ + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
+ + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
+// + dstBlock[x +13*dstStride]
+// + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
+ }*/
}
#ifdef HAVE_3DNOW
asm volatile("femms");
asm volatile("emms");
#endif
-#ifdef TIMING
- // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)
- sumTime= rdtsc() - sumTime;
+#ifdef DEBUG_BRIGHTNESS
if(!isColor)
- printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d \r",
- (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000),
- (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000)
- , black, white);
+ {
+ int max=1;
+ int i;
+ for(i=0; i<256; i++)
+ if(yHistogram[i] > max) max=yHistogram[i];
+
+ for(i=1; i<256; i++)
+ {
+ int x;
+ int start=yHistogram[i-1]/(max/256+1);
+ int end=yHistogram[i]/(max/256+1);
+ int inc= end > start ? 1 : -1;
+ for(x=start; x!=end+inc; x+=inc)
+ dst[ i*dstStride + x]+=128;
+ }
+
+ for(i=0; i<100; i+=2)
+ {
+ dst[ (white)*dstStride + i]+=128;
+ dst[ (black)*dstStride + i]+=128;
+ }
+
+ }
#endif
+
+ *c2= c; //copy local context back
+
}