2 |
* |
* |
3 |
* Modifications: |
* Modifications: |
4 |
* |
* |
5 |
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
6 |
|
* 14.04.2002 added MotionEstimationBVOP() |
7 |
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
8 |
|
* EPZS and EPZS^2 |
9 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
10 |
* PMVFast_Refine to support multiple searches with different start points |
* PMVFast_Refine to support multiple searches with different start points |
11 |
* 07.01.2002 uv-block-based interpolation |
* 07.01.2002 uv-block-based interpolation |
35 |
|
|
36 |
#include <assert.h> |
#include <assert.h> |
37 |
#include <stdio.h> |
#include <stdio.h> |
38 |
|
#include <stdlib.h> |
39 |
|
|
40 |
#include "../encoder.h" |
#include "../encoder.h" |
41 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
42 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
43 |
#include "../global.h" |
#include "../global.h" |
44 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
45 |
|
#include "motion.h" |
46 |
#include "sad.h" |
#include "sad.h" |
47 |
|
|
48 |
// very large value |
// very large value |
71 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
72 |
|
|
73 |
|
|
74 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
int32_t PMVfastSearch16( |
75 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
const uint8_t * const pRef, |
76 |
#define ABS(X) (((X)>0)?(X):-(X)) |
const uint8_t * const pRefH, |
77 |
#define SIGN(X) (((X)>0)?1:-1) |
const uint8_t * const pRefV, |
78 |
|
const uint8_t * const pRefHV, |
79 |
|
const IMAGE * const pCur, |
80 |
|
const int x, const int y, |
81 |
|
const uint32_t MotionFlags, |
82 |
|
const MBParam * const pParam, |
83 |
|
MACROBLOCK * const pMBs, |
84 |
|
VECTOR * const currMV, |
85 |
|
VECTOR * const currPMV); |
86 |
|
|
87 |
|
int32_t EPZSSearch16( |
88 |
|
const uint8_t * const pRef, |
89 |
|
const uint8_t * const pRefH, |
90 |
|
const uint8_t * const pRefV, |
91 |
|
const uint8_t * const pRefHV, |
92 |
|
const IMAGE * const pCur, |
93 |
|
const int x, const int y, |
94 |
|
const uint32_t MotionFlags, |
95 |
|
const MBParam * const pParam, |
96 |
|
MACROBLOCK * const pMBs, |
97 |
|
VECTOR * const currMV, |
98 |
|
VECTOR * const currPMV); |
99 |
|
|
100 |
|
|
101 |
int32_t PMVfastSearch8( |
int32_t PMVfastSearch8( |
106 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
107 |
const int x, const int y, |
const int x, const int y, |
108 |
const int start_x, int start_y, |
const int start_x, int start_y, |
109 |
const uint32_t iQuality, |
const uint32_t MotionFlags, |
110 |
MBParam * const pParam, |
const MBParam * const pParam, |
111 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
112 |
VECTOR * const currMV, |
VECTOR * const currMV, |
113 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
114 |
|
|
115 |
int32_t PMVfastSearch16( |
int32_t EPZSSearch8( |
116 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
117 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
118 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
119 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
120 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
121 |
const int x, const int y, |
const int x, const int y, |
122 |
const uint32_t iQuality, |
const int start_x, int start_y, |
123 |
MBParam * const pParam, |
const uint32_t MotionFlags, |
124 |
|
const MBParam * const pParam, |
125 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
126 |
VECTOR * const currMV, |
VECTOR * const currMV, |
127 |
VECTOR * const currPMV); |
VECTOR * const currPMV); |
128 |
|
|
129 |
|
|
130 |
|
typedef int32_t (MainSearch16Func)( |
131 |
|
const uint8_t * const pRef, |
132 |
|
const uint8_t * const pRefH, |
133 |
|
const uint8_t * const pRefV, |
134 |
|
const uint8_t * const pRefHV, |
135 |
|
const uint8_t * const cur, |
136 |
|
const int x, const int y, |
137 |
|
int32_t startx, int32_t starty, |
138 |
|
int32_t iMinSAD, |
139 |
|
VECTOR * const currMV, |
140 |
|
const VECTOR * const pmv, |
141 |
|
const int32_t min_dx, const int32_t max_dx, |
142 |
|
const int32_t min_dy, const int32_t max_dy, |
143 |
|
const int32_t iEdgedWidth, |
144 |
|
const int32_t iDiamondSize, |
145 |
|
const int32_t iFcode, |
146 |
|
const int32_t iQuant, |
147 |
|
int iFound); |
148 |
|
|
149 |
/* |
typedef MainSearch16Func* MainSearch16FuncPtr; |
|
* diamond search stuff |
|
|
* keep the the sequence in circular order (so optimization works) |
|
|
*/ |
|
|
|
|
|
typedef struct |
|
|
{ |
|
|
int32_t dx; |
|
|
int32_t dy; |
|
|
} |
|
|
DPOINT; |
|
|
|
|
|
|
|
|
static const DPOINT diamond_small[4] = |
|
|
{ |
|
|
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
|
|
}; |
|
150 |
|
|
151 |
|
|
152 |
static const DPOINT diamond_large[8] = |
typedef int32_t (MainSearch8Func)( |
153 |
{ |
const uint8_t * const pRef, |
154 |
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
const uint8_t * const pRefH, |
155 |
}; |
const uint8_t * const pRefV, |
156 |
|
const uint8_t * const pRefHV, |
157 |
|
const uint8_t * const cur, |
158 |
|
const int x, const int y, |
159 |
|
int32_t startx, int32_t starty, |
160 |
|
int32_t iMinSAD, |
161 |
|
VECTOR * const currMV, |
162 |
|
const VECTOR * const pmv, |
163 |
|
const int32_t min_dx, const int32_t max_dx, |
164 |
|
const int32_t min_dy, const int32_t max_dy, |
165 |
|
const int32_t iEdgedWidth, |
166 |
|
const int32_t iDiamondSize, |
167 |
|
const int32_t iFcode, |
168 |
|
const int32_t iQuant, |
169 |
|
int iFound); |
170 |
|
|
171 |
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
172 |
|
|
173 |
// mv.length table |
// mv.length table |
174 |
static const uint32_t mvtab[33] = { |
static const uint32_t mvtab[33] = { |
219 |
|
|
220 |
|
|
221 |
|
|
|
/* calculate the min/max range (in halfpixels) |
|
|
relative to the _MACROBLOCK_ position |
|
|
*/ |
|
|
|
|
|
static void __inline get_range( |
|
|
int32_t * const min_dx, int32_t * const max_dx, |
|
|
int32_t * const min_dy, int32_t * const max_dy, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block_sz, // block dimension, 8 or 16 |
|
|
const uint32_t width, const uint32_t height, |
|
|
const uint32_t fcode) |
|
|
{ |
|
|
|
|
|
const int search_range = 32 << (fcode - 1); |
|
|
const int high = search_range - 1; |
|
|
const int low = -search_range; |
|
|
|
|
|
// convert full-pixel measurements to half pixel |
|
|
const int hp_width = 2 * width; |
|
|
const int hp_height = 2 * height; |
|
|
const int hp_edge = 2 * block_sz; |
|
|
const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate |
|
|
const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ |
|
|
|
|
|
*max_dx = MIN(high, hp_width - hp_x); |
|
|
*max_dy = MIN(high, hp_height - hp_y); |
|
|
*min_dx = MAX(low, -(hp_edge + hp_x)); |
|
|
*min_dy = MAX(low, -(hp_edge + hp_y)); |
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
/* |
|
|
* getref: calculate reference image pointer |
|
|
* the decision to use interpolation h/v/hv or the normal image is |
|
|
* based on dx & dy. |
|
|
*/ |
|
|
|
|
|
static __inline const uint8_t * get_ref( |
|
|
const uint8_t * const refn, |
|
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const int32_t dx, const int32_t dy, |
|
|
const uint32_t stride) |
|
|
{ |
|
|
|
|
|
switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) |
|
|
{ |
|
|
case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; |
|
|
case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; |
|
|
case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; |
|
|
} |
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
|
|
|
|
|
static __inline const uint8_t * get_ref_mv( |
|
|
const uint8_t * const refn, |
|
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const VECTOR* mv, // measured in half-pel! |
|
|
const uint32_t stride) |
|
|
{ |
|
|
|
|
|
switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
|
|
{ |
|
|
case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; |
|
|
case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
|
case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; |
|
|
default : |
|
|
case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
|
} |
|
|
|
|
|
} |
|
222 |
|
|
223 |
#ifndef SEARCH16 |
#ifndef SEARCH16 |
224 |
#define SEARCH16 PMVfastSearch16 |
#define SEARCH16 PMVfastSearch16 |
225 |
|
//#define SEARCH16 FullSearch16 |
226 |
|
//#define SEARCH16 EPZSSearch16 |
227 |
#endif |
#endif |
228 |
|
|
229 |
#ifndef SEARCH8 |
#ifndef SEARCH8 |
230 |
#define SEARCH8 PMVfastSearch8 |
#define SEARCH8 PMVfastSearch8 |
231 |
|
//#define SEARCH8 EPZSSearch8 |
232 |
#endif |
#endif |
233 |
|
|
234 |
bool MotionEstimation( |
bool MotionEstimation( |
254 |
int32_t sad16; |
int32_t sad16; |
255 |
int32_t deviation; |
int32_t deviation; |
256 |
|
|
257 |
|
if (sadInit) |
258 |
|
(*sadInit)(); |
259 |
|
|
260 |
// note: i==horizontal, j==vertical |
// note: i==horizontal, j==vertical |
261 |
for (i = 0; i < iHcount; i++) |
for (i = 0; i < iHcount; i++) |
262 |
for (j = 0; j < iWcount; j++) |
for (j = 0; j < iWcount; j++) |
359 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
360 |
} |
} |
361 |
|
|
362 |
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
363 |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
364 |
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
365 |
|
if (iSAD < iMinSAD) \ |
366 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
367 |
|
} |
368 |
|
|
369 |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
370 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
404 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
405 |
} |
} |
406 |
|
|
407 |
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
408 |
|
{ \ |
409 |
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
410 |
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
411 |
|
if (iSAD < iMinSAD) \ |
412 |
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
413 |
|
} |
414 |
|
|
415 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
416 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
481 |
} |
} |
482 |
*/ |
*/ |
483 |
|
|
484 |
int32_t PMVfastSearch16_MainSearch( |
int32_t Diamond16_MainSearch( |
485 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
486 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
487 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
538 |
return iMinSAD; |
return iMinSAD; |
539 |
} |
} |
540 |
|
|
541 |
int32_t PMVfastSearch16_Refine( |
int32_t Square16_MainSearch( |
542 |
|
const uint8_t * const pRef, |
543 |
|
const uint8_t * const pRefH, |
544 |
|
const uint8_t * const pRefV, |
545 |
|
const uint8_t * const pRefHV, |
546 |
|
const uint8_t * const cur, |
547 |
|
const int x, const int y, |
548 |
|
int32_t startx, int32_t starty, |
549 |
|
int32_t iMinSAD, |
550 |
|
VECTOR * const currMV, |
551 |
|
const VECTOR * const pmv, |
552 |
|
const int32_t min_dx, const int32_t max_dx, |
553 |
|
const int32_t min_dy, const int32_t max_dy, |
554 |
|
const int32_t iEdgedWidth, |
555 |
|
const int32_t iDiamondSize, |
556 |
|
const int32_t iFcode, |
557 |
|
const int32_t iQuant, |
558 |
|
int iFound) |
559 |
|
{ |
560 |
|
/* Do a square search around given starting point, return SAD of best */ |
561 |
|
|
562 |
|
int32_t iDirection=0; |
563 |
|
int32_t iSAD; |
564 |
|
VECTOR backupMV; |
565 |
|
backupMV.x = startx; |
566 |
|
backupMV.y = starty; |
567 |
|
|
568 |
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
569 |
|
|
570 |
|
/* new direction are extra, so 1-4 is normal diamond |
571 |
|
537 |
572 |
|
1*2 |
573 |
|
648 |
574 |
|
*/ |
575 |
|
|
576 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
577 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
578 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
579 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
580 |
|
|
581 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
582 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
583 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
584 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
585 |
|
|
586 |
|
|
587 |
|
if (iDirection) |
588 |
|
while (!iFound) |
589 |
|
{ |
590 |
|
iFound = 1; |
591 |
|
backupMV=*currMV; |
592 |
|
|
593 |
|
switch (iDirection) |
594 |
|
{ |
595 |
|
case 1: |
596 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
597 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
598 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
599 |
|
break; |
600 |
|
case 2: |
601 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
602 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
603 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
604 |
|
break; |
605 |
|
|
606 |
|
case 3: |
607 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
608 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
609 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
610 |
|
break; |
611 |
|
|
612 |
|
case 4: |
613 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
614 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
615 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
616 |
|
break; |
617 |
|
|
618 |
|
case 5: |
619 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
620 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
621 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
622 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
623 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
624 |
|
break; |
625 |
|
|
626 |
|
case 6: |
627 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
628 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
629 |
|
|
630 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
631 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
632 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
633 |
|
|
634 |
|
break; |
635 |
|
|
636 |
|
case 7: |
637 |
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
638 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
639 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
640 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
641 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
642 |
|
break; |
643 |
|
|
644 |
|
case 8: |
645 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
646 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
647 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
648 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
649 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
650 |
|
break; |
651 |
|
default: |
652 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
653 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
654 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
655 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
656 |
|
|
657 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
658 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
659 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
660 |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
661 |
|
break; |
662 |
|
} |
663 |
|
} |
664 |
|
else |
665 |
|
{ |
666 |
|
currMV->x = startx; |
667 |
|
currMV->y = starty; |
668 |
|
} |
669 |
|
return iMinSAD; |
670 |
|
} |
671 |
|
|
672 |
|
|
673 |
|
int32_t Full16_MainSearch( |
674 |
|
const uint8_t * const pRef, |
675 |
|
const uint8_t * const pRefH, |
676 |
|
const uint8_t * const pRefV, |
677 |
|
const uint8_t * const pRefHV, |
678 |
|
const uint8_t * const cur, |
679 |
|
const int x, const int y, |
680 |
|
int32_t startx, int32_t starty, |
681 |
|
int32_t iMinSAD, |
682 |
|
VECTOR * const currMV, |
683 |
|
const VECTOR * const pmv, |
684 |
|
const int32_t min_dx, const int32_t max_dx, |
685 |
|
const int32_t min_dy, const int32_t max_dy, |
686 |
|
const int32_t iEdgedWidth, |
687 |
|
const int32_t iDiamondSize, |
688 |
|
const int32_t iFcode, |
689 |
|
const int32_t iQuant, |
690 |
|
int iFound) |
691 |
|
{ |
692 |
|
int32_t iSAD; |
693 |
|
int32_t dx,dy; |
694 |
|
VECTOR backupMV; |
695 |
|
backupMV.x = startx; |
696 |
|
backupMV.y = starty; |
697 |
|
|
698 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
699 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
700 |
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
701 |
|
|
702 |
|
return iMinSAD; |
703 |
|
} |
704 |
|
|
705 |
|
int32_t Full8_MainSearch( |
706 |
|
const uint8_t * const pRef, |
707 |
|
const uint8_t * const pRefH, |
708 |
|
const uint8_t * const pRefV, |
709 |
|
const uint8_t * const pRefHV, |
710 |
|
const uint8_t * const cur, |
711 |
|
const int x, const int y, |
712 |
|
int32_t startx, int32_t starty, |
713 |
|
int32_t iMinSAD, |
714 |
|
VECTOR * const currMV, |
715 |
|
const VECTOR * const pmv, |
716 |
|
const int32_t min_dx, const int32_t max_dx, |
717 |
|
const int32_t min_dy, const int32_t max_dy, |
718 |
|
const int32_t iEdgedWidth, |
719 |
|
const int32_t iDiamondSize, |
720 |
|
const int32_t iFcode, |
721 |
|
const int32_t iQuant, |
722 |
|
int iFound) |
723 |
|
{ |
724 |
|
int32_t iSAD; |
725 |
|
int32_t dx,dy; |
726 |
|
VECTOR backupMV; |
727 |
|
backupMV.x = startx; |
728 |
|
backupMV.y = starty; |
729 |
|
|
730 |
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
731 |
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
732 |
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
733 |
|
|
734 |
|
return iMinSAD; |
735 |
|
} |
736 |
|
|
737 |
|
|
738 |
|
|
739 |
|
int32_t Halfpel16_Refine( |
740 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
741 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
742 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
771 |
|
|
772 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
773 |
|
|
774 |
|
|
775 |
int32_t PMVfastSearch16( |
int32_t PMVfastSearch16( |
776 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
777 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
780 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
781 |
const int x, const int y, |
const int x, const int y, |
782 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
783 |
MBParam * const pParam, |
const MBParam * const pParam, |
784 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
785 |
VECTOR * const currMV, |
VECTOR * const currMV, |
786 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
907 |
iEdgedWidth, MV_MAX_ERROR); |
iEdgedWidth, MV_MAX_ERROR); |
908 |
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
909 |
|
|
910 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < pMB->sad16) ) ) |
911 |
{ |
{ |
912 |
|
|
913 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
914 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
915 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
916 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
917 |
} |
} |
918 |
|
|
919 |
/* |
/* |
950 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
951 |
|
|
952 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
953 |
if (x != (iWcount-1)) |
if ((uint32_t)x != (iWcount-1)) |
954 |
{ |
{ |
955 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
956 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
964 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
965 |
*/ |
*/ |
966 |
|
|
967 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD < pMB->sad16) ) ) |
968 |
{ |
{ |
969 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
970 |
goto step10b; |
goto PMVfast16_Terminate_without_Refine; |
971 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (MotionFlags & PMV_EARLYSTOP16) |
972 |
goto step10; |
goto PMVfast16_Terminate_with_Refine; |
973 |
} |
} |
974 |
|
|
975 |
|
|
986 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
987 |
|
|
988 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
989 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
990 |
x, y, |
x, y, |
991 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
992 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1002 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1003 |
|
|
1004 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1005 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1006 |
x, y, |
x, y, |
1007 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1008 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1015 |
} |
} |
1016 |
|
|
1017 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1018 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1019 |
x, y, |
x, y, |
1020 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1021 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1032 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
1033 |
*/ |
*/ |
1034 |
|
|
1035 |
step10: |
PMVfast16_Terminate_with_Refine: |
1036 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1037 |
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1038 |
x, y, |
x, y, |
1039 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1040 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1041 |
|
|
1042 |
step10b: |
PMVfast16_Terminate_without_Refine: |
1043 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1044 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1045 |
return iMinSAD; |
return iMinSAD; |
1050 |
|
|
1051 |
|
|
1052 |
|
|
1053 |
int32_t PMVfastSearch8_MainSearch( |
int32_t Diamond8_MainSearch( |
1054 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1055 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1056 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1107 |
return iMinSAD; |
return iMinSAD; |
1108 |
} |
} |
1109 |
|
|
1110 |
int32_t PMVfastSearch8_Refine( |
int32_t Halfpel8_Refine( |
1111 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
1112 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1113 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1152 |
const int x, const int y, |
const int x, const int y, |
1153 |
const int start_x, int start_y, |
const int start_x, int start_y, |
1154 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1155 |
MBParam * const pParam, |
const MBParam * const pParam, |
1156 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMBs, |
1157 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1158 |
VECTOR * const currPMV) |
VECTOR * const currPMV) |
1258 |
iEdgedWidth); |
iEdgedWidth); |
1259 |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
1260 |
|
|
1261 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < pMB->sad8[iSubBlock]) ) ) |
1262 |
{ |
{ |
1263 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1264 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1265 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1266 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1267 |
} |
} |
1268 |
|
|
1269 |
|
|
1270 |
/* |
/* |
1271 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1272 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1317 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
1318 |
*/ |
*/ |
1319 |
|
|
1320 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < pMB->sad8[iSubBlock]) ) ) |
1321 |
{ |
{ |
1322 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP16) |
1323 |
goto step10_8b; |
goto PMVfast8_Terminate_without_Refine; |
1324 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (MotionFlags & PMV_EARLYSTOP16) |
1325 |
goto step10_8; |
goto PMVfast8_Terminate_with_Refine; |
1326 |
} |
} |
1327 |
|
|
1328 |
/************ (Diamond Search) **************/ |
/************ (Diamond Search) **************/ |
1338 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1339 |
|
|
1340 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1341 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1342 |
x, y, |
x, y, |
1343 |
currMV->x, currMV->y, iMinSAD, &newMV, |
currMV->x, currMV->y, iMinSAD, &newMV, |
1344 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1354 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
1355 |
|
|
1356 |
if (!(MVequal(pmv[0],backupMV)) ) |
if (!(MVequal(pmv[0],backupMV)) ) |
1357 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1358 |
x, y, |
x, y, |
1359 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1360 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1367 |
} |
} |
1368 |
|
|
1369 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1370 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
1371 |
x, y, |
x, y, |
1372 |
0, 0, iMinSAD, &newMV, |
0, 0, iMinSAD, &newMV, |
1373 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
1384 |
By performing an optional local half-pixel search, we can refine this result even further. |
By performing an optional local half-pixel search, we can refine this result even further. |
1385 |
*/ |
*/ |
1386 |
|
|
1387 |
step10_8: |
PMVfast8_Terminate_with_Refine: |
1388 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1389 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1390 |
x, y, |
x, y, |
1391 |
currMV, iMinSAD, |
currMV, iMinSAD, |
1392 |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1393 |
|
|
|
step10_8b: |
|
1394 |
|
|
1395 |
|
PMVfast8_Terminate_without_Refine: |
1396 |
currPMV->x = currMV->x - pmv[0].x; |
currPMV->x = currMV->x - pmv[0].x; |
1397 |
currPMV->y = currMV->y - pmv[0].y; |
currPMV->y = currMV->y - pmv[0].y; |
1398 |
|
|
1399 |
return iMinSAD; |
return iMinSAD; |
1400 |
} |
} |
1401 |
|
|
1402 |
|
int32_t EPZSSearch16( |
1403 |
|
const uint8_t * const pRef, |
1404 |
|
const uint8_t * const pRefH, |
1405 |
|
const uint8_t * const pRefV, |
1406 |
|
const uint8_t * const pRefHV, |
1407 |
|
const IMAGE * const pCur, |
1408 |
|
const int x, const int y, |
1409 |
|
const uint32_t MotionFlags, |
1410 |
|
const MBParam * const pParam, |
1411 |
|
MACROBLOCK * const pMBs, |
1412 |
|
VECTOR * const currMV, |
1413 |
|
VECTOR * const currPMV) |
1414 |
|
{ |
1415 |
|
const uint32_t iWcount = pParam->mb_width; |
1416 |
|
const uint32_t iHcount = pParam->mb_height; |
1417 |
|
const int32_t iFcode = pParam->fixed_code; |
1418 |
|
const int32_t iQuant = pParam->quant; |
1419 |
|
|
1420 |
|
const int32_t iWidth = pParam->width; |
1421 |
|
const int32_t iHeight = pParam->height; |
1422 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1423 |
|
|
1424 |
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
1425 |
|
|
1426 |
|
int32_t min_dx; |
1427 |
|
int32_t max_dx; |
1428 |
|
int32_t min_dy; |
1429 |
|
int32_t max_dy; |
1430 |
|
|
1431 |
|
VECTOR newMV; |
1432 |
|
VECTOR backupMV; |
1433 |
|
|
1434 |
|
VECTOR pmv[4]; |
1435 |
|
int32_t psad[8]; |
1436 |
|
|
1437 |
|
static MACROBLOCK * oldMBs = NULL; |
1438 |
|
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1439 |
|
MACROBLOCK * oldMB = NULL; |
1440 |
|
|
1441 |
|
static int32_t thresh2; |
1442 |
|
int32_t bPredEq; |
1443 |
|
int32_t iMinSAD,iSAD=9999; |
1444 |
|
|
1445 |
|
MainSearch16FuncPtr EPZSMainSearchPtr; |
1446 |
|
|
1447 |
|
if (oldMBs == NULL) |
1448 |
|
{ oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); |
1449 |
|
fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
1450 |
|
} |
1451 |
|
oldMB = oldMBs + x + y * iWcount; |
1452 |
|
|
1453 |
|
/* Get maximum range */ |
1454 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1455 |
|
x, y, 16, iWidth, iHeight, iFcode); |
1456 |
|
|
1457 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1458 |
|
|
1459 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1460 |
|
{ min_dx = EVEN(min_dx); |
1461 |
|
max_dx = EVEN(max_dx); |
1462 |
|
min_dy = EVEN(min_dy); |
1463 |
|
max_dy = EVEN(max_dy); |
1464 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1465 |
|
|
1466 |
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
1467 |
|
|
1468 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1469 |
|
MinSAD=SAD |
1470 |
|
If Motion Vector equal to Previous frame motion vector |
1471 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1472 |
|
If SAD<=256 goto Step 10. |
1473 |
|
*/ |
1474 |
|
|
1475 |
|
// Prepare for main loop |
1476 |
|
|
1477 |
|
*currMV=pmv[0]; /* current best := median prediction */ |
1478 |
|
if (!(MotionFlags & PMV_HALFPEL16)) |
1479 |
|
{ |
1480 |
|
currMV->x = EVEN(currMV->x); |
1481 |
|
currMV->y = EVEN(currMV->y); |
1482 |
|
} |
1483 |
|
|
1484 |
|
if (currMV->x > max_dx) |
1485 |
|
currMV->x=max_dx; |
1486 |
|
if (currMV->x < min_dx) |
1487 |
|
currMV->x=min_dx; |
1488 |
|
if (currMV->y > max_dy) |
1489 |
|
currMV->y=max_dy; |
1490 |
|
if (currMV->y < min_dy) |
1491 |
|
currMV->y=min_dy; |
1492 |
|
|
1493 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1494 |
|
|
1495 |
|
iMinSAD = sad16( cur, |
1496 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
1497 |
|
iEdgedWidth, MV_MAX_ERROR); |
1498 |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
1499 |
|
|
1500 |
|
// thresh1 is fixed to 256 |
1501 |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < pMB->sad16) ) ) |
1502 |
|
{ |
1503 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1504 |
|
goto EPZS16_Terminate_without_Refine; |
1505 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1506 |
|
goto EPZS16_Terminate_with_Refine; |
1507 |
|
} |
1508 |
|
|
1509 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1510 |
|
|
1511 |
|
// previous frame MV |
1512 |
|
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
1513 |
|
|
1514 |
|
// set threshhold based on Min of Prediction and SAD of collocated block |
1515 |
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
1516 |
|
|
1517 |
|
if ((x==0) && (y==0) ) |
1518 |
|
{ |
1519 |
|
thresh2 = 512; |
1520 |
|
} |
1521 |
|
else |
1522 |
|
{ |
1523 |
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
1524 |
|
|
1525 |
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
1526 |
|
} |
1527 |
|
|
1528 |
|
// MV=(0,0) is often a good choice |
1529 |
|
|
1530 |
|
CHECK_MV16_ZERO; |
1531 |
|
|
1532 |
|
|
1533 |
|
// left neighbour, if allowed |
1534 |
|
if (x != 0) |
1535 |
|
{ |
1536 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1537 |
|
{ pmv[1].x = EVEN(pmv[1].x); |
1538 |
|
pmv[1].y = EVEN(pmv[1].y); |
1539 |
|
} |
1540 |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1541 |
|
} |
1542 |
|
|
1543 |
|
// top neighbour, if allowed |
1544 |
|
if (y != 0) |
1545 |
|
{ |
1546 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1547 |
|
{ pmv[2].x = EVEN(pmv[2].x); |
1548 |
|
pmv[2].y = EVEN(pmv[2].y); |
1549 |
|
} |
1550 |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1551 |
|
|
1552 |
|
// top right neighbour, if allowed |
1553 |
|
if ((uint32_t)x != (iWcount-1)) |
1554 |
|
{ |
1555 |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
1556 |
|
{ pmv[3].x = EVEN(pmv[3].x); |
1557 |
|
pmv[3].y = EVEN(pmv[3].y); |
1558 |
|
} |
1559 |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
1560 |
|
} |
1561 |
|
} |
1562 |
|
|
1563 |
|
/* Terminate if MinSAD <= T_2 |
1564 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1565 |
|
*/ |
1566 |
|
|
1567 |
|
if ( (iMinSAD <= thresh2) |
1568 |
|
|| ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= pMB->sad16) ) ) |
1569 |
|
{ |
1570 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1571 |
|
goto EPZS16_Terminate_without_Refine; |
1572 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1573 |
|
goto EPZS16_Terminate_with_Refine; |
1574 |
|
} |
1575 |
|
|
1576 |
|
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
1577 |
|
|
1578 |
|
backupMV = pMB->mvs[0]; // last MV |
1579 |
|
backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
1580 |
|
backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
1581 |
|
|
1582 |
|
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
1583 |
|
|
1584 |
|
// left neighbour |
1585 |
|
if (x != 0) |
1586 |
|
CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); |
1587 |
|
|
1588 |
|
// top neighbour |
1589 |
|
if (y != 0) |
1590 |
|
CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); |
1591 |
|
|
1592 |
|
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
1593 |
|
|
1594 |
|
if ((uint32_t)x != iWcount-1) |
1595 |
|
CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); |
1596 |
|
|
1597 |
|
// bottom neighbour, dito |
1598 |
|
if ((uint32_t)y != iHcount-1) |
1599 |
|
CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); |
1600 |
|
|
1601 |
|
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
1602 |
|
if (iMinSAD <= thresh2) |
1603 |
|
{ |
1604 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1605 |
|
goto EPZS16_Terminate_without_Refine; |
1606 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1607 |
|
goto EPZS16_Terminate_with_Refine; |
1608 |
|
} |
1609 |
|
|
1610 |
|
/************ (if Diamond Search) **************/ |
1611 |
|
|
1612 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1613 |
|
|
1614 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1615 |
|
|
1616 |
|
if (MotionFlags & PMV_USESQUARES16) |
1617 |
|
EPZSMainSearchPtr = Square16_MainSearch; |
1618 |
|
else |
1619 |
|
EPZSMainSearchPtr = Diamond16_MainSearch; |
1620 |
|
|
1621 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1622 |
|
x, y, |
1623 |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1624 |
|
2, iFcode, iQuant, 0); |
1625 |
|
|
1626 |
|
if (iSAD < iMinSAD) |
1627 |
|
{ |
1628 |
|
*currMV = newMV; |
1629 |
|
iMinSAD = iSAD; |
1630 |
|
} |
1631 |
|
|
1632 |
|
|
1633 |
|
if (MotionFlags & PMV_EXTSEARCH16) |
1634 |
|
{ |
1635 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1636 |
|
|
1637 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1638 |
|
{ |
1639 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1640 |
|
x, y, |
1641 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1642 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
1643 |
|
} |
1644 |
|
|
1645 |
|
if (iSAD < iMinSAD) |
1646 |
|
{ |
1647 |
|
*currMV = newMV; |
1648 |
|
iMinSAD = iSAD; |
1649 |
|
} |
1650 |
|
|
1651 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1652 |
|
{ |
1653 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1654 |
|
x, y, |
1655 |
|
0, 0, iMinSAD, &newMV, |
1656 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); |
1657 |
|
|
1658 |
|
if (iSAD < iMinSAD) |
1659 |
|
{ |
1660 |
|
*currMV = newMV; |
1661 |
|
iMinSAD = iSAD; |
1662 |
|
} |
1663 |
|
} |
1664 |
|
} |
1665 |
|
|
1666 |
|
/*************** Choose best MV found **************/ |
1667 |
|
|
1668 |
|
EPZS16_Terminate_with_Refine: |
1669 |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
1670 |
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1671 |
|
x, y, |
1672 |
|
currMV, iMinSAD, |
1673 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1674 |
|
|
1675 |
|
EPZS16_Terminate_without_Refine: |
1676 |
|
|
1677 |
|
*oldMB = *pMB; |
1678 |
|
|
1679 |
|
currPMV->x = currMV->x - pmv[0].x; |
1680 |
|
currPMV->y = currMV->y - pmv[0].y; |
1681 |
|
return iMinSAD; |
1682 |
|
} |
1683 |
|
|
1684 |
|
|
1685 |
|
int32_t EPZSSearch8( |
1686 |
|
const uint8_t * const pRef, |
1687 |
|
const uint8_t * const pRefH, |
1688 |
|
const uint8_t * const pRefV, |
1689 |
|
const uint8_t * const pRefHV, |
1690 |
|
const IMAGE * const pCur, |
1691 |
|
const int x, const int y, |
1692 |
|
const int start_x, const int start_y, |
1693 |
|
const uint32_t MotionFlags, |
1694 |
|
const MBParam * const pParam, |
1695 |
|
MACROBLOCK * const pMBs, |
1696 |
|
VECTOR * const currMV, |
1697 |
|
VECTOR * const currPMV) |
1698 |
|
{ |
1699 |
|
const uint32_t iWcount = pParam->mb_width; |
1700 |
|
const int32_t iFcode = pParam->fixed_code; |
1701 |
|
const int32_t iQuant = pParam->quant; |
1702 |
|
|
1703 |
|
const int32_t iWidth = pParam->width; |
1704 |
|
const int32_t iHeight = pParam->height; |
1705 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1706 |
|
|
1707 |
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
1708 |
|
|
1709 |
|
int32_t iDiamondSize=1; |
1710 |
|
|
1711 |
|
int32_t min_dx; |
1712 |
|
int32_t max_dx; |
1713 |
|
int32_t min_dy; |
1714 |
|
int32_t max_dy; |
1715 |
|
|
1716 |
|
VECTOR newMV; |
1717 |
|
VECTOR backupMV; |
1718 |
|
|
1719 |
|
VECTOR pmv[4]; |
1720 |
|
int32_t psad[8]; |
1721 |
|
|
1722 |
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
1723 |
|
|
1724 |
|
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1725 |
|
|
1726 |
|
int32_t bPredEq; |
1727 |
|
int32_t iMinSAD,iSAD=9999; |
1728 |
|
|
1729 |
|
MainSearch8FuncPtr EPZSMainSearchPtr; |
1730 |
|
|
1731 |
|
/* Get maximum range */ |
1732 |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1733 |
|
x, y, 8, iWidth, iHeight, iFcode); |
1734 |
|
|
1735 |
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
1736 |
|
|
1737 |
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
1738 |
|
{ min_dx = EVEN(min_dx); |
1739 |
|
max_dx = EVEN(max_dx); |
1740 |
|
min_dy = EVEN(min_dy); |
1741 |
|
max_dy = EVEN(max_dy); |
1742 |
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
1743 |
|
|
1744 |
|
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
1745 |
|
|
1746 |
|
|
1747 |
|
/* Step 4: Calculate SAD around the Median prediction. |
1748 |
|
MinSAD=SAD |
1749 |
|
If Motion Vector equal to Previous frame motion vector |
1750 |
|
and MinSAD<PrevFrmSAD goto Step 10. |
1751 |
|
If SAD<=256 goto Step 10. |
1752 |
|
*/ |
1753 |
|
|
1754 |
|
// Prepare for main loop |
1755 |
|
|
1756 |
|
|
1757 |
|
if (!(MotionFlags & PMV_HALFPEL8)) |
1758 |
|
{ |
1759 |
|
currMV->x = EVEN(currMV->x); |
1760 |
|
currMV->y = EVEN(currMV->y); |
1761 |
|
} |
1762 |
|
|
1763 |
|
if (currMV->x > max_dx) |
1764 |
|
currMV->x=max_dx; |
1765 |
|
if (currMV->x < min_dx) |
1766 |
|
currMV->x=min_dx; |
1767 |
|
if (currMV->y > max_dy) |
1768 |
|
currMV->y=max_dy; |
1769 |
|
if (currMV->y < min_dy) |
1770 |
|
currMV->y=min_dy; |
1771 |
|
|
1772 |
|
/***************** This is predictor SET A: only median prediction ******************/ |
1773 |
|
|
1774 |
|
|
1775 |
|
iMinSAD = sad8( cur, |
1776 |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1777 |
|
iEdgedWidth); |
1778 |
|
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
1779 |
|
|
1780 |
|
|
1781 |
|
// thresh1 is fixed to 256 |
1782 |
|
if (iMinSAD < 256/4 ) |
1783 |
|
{ |
1784 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1785 |
|
goto EPZS8_Terminate_without_Refine; |
1786 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1787 |
|
goto EPZS8_Terminate_with_Refine; |
1788 |
|
} |
1789 |
|
|
1790 |
|
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
1791 |
|
|
1792 |
|
// previous frame MV |
1793 |
|
CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
1794 |
|
|
1795 |
|
// MV=(0,0) is often a good choice |
1796 |
|
|
1797 |
|
CHECK_MV8_ZERO; |
1798 |
|
|
1799 |
|
/* Terminate if MinSAD <= T_2 |
1800 |
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
1801 |
|
*/ |
1802 |
|
|
1803 |
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
1804 |
|
{ |
1805 |
|
if (MotionFlags & PMV_QUICKSTOP8) |
1806 |
|
goto EPZS8_Terminate_without_Refine; |
1807 |
|
if (MotionFlags & PMV_EARLYSTOP8) |
1808 |
|
goto EPZS8_Terminate_with_Refine; |
1809 |
|
} |
1810 |
|
|
1811 |
|
/************ (if Diamond Search) **************/ |
1812 |
|
|
1813 |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
1814 |
|
|
1815 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
1816 |
|
iDiamondSize *= 2; |
1817 |
|
|
1818 |
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
1819 |
|
|
1820 |
|
// if (MotionFlags & PMV_USESQUARES8) |
1821 |
|
// EPZSMainSearchPtr = Square8_MainSearch; |
1822 |
|
// else |
1823 |
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
1824 |
|
|
1825 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1826 |
|
x, y, |
1827 |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
1828 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
1829 |
|
iDiamondSize, iFcode, iQuant, 00); |
1830 |
|
|
1831 |
|
|
1832 |
|
if (iSAD < iMinSAD) |
1833 |
|
{ |
1834 |
|
*currMV = newMV; |
1835 |
|
iMinSAD = iSAD; |
1836 |
|
} |
1837 |
|
|
1838 |
|
if (MotionFlags & PMV_EXTSEARCH8) |
1839 |
|
{ |
1840 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
1841 |
|
|
1842 |
|
if (!(MVequal(pmv[0],backupMV)) ) |
1843 |
|
{ |
1844 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1845 |
|
x, y, |
1846 |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
1847 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
1848 |
|
|
1849 |
|
if (iSAD < iMinSAD) |
1850 |
|
{ |
1851 |
|
*currMV = newMV; |
1852 |
|
iMinSAD = iSAD; |
1853 |
|
} |
1854 |
|
} |
1855 |
|
|
1856 |
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
1857 |
|
{ |
1858 |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
1859 |
|
x, y, |
1860 |
|
0, 0, iMinSAD, &newMV, |
1861 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
1862 |
|
|
1863 |
|
if (iSAD < iMinSAD) |
1864 |
|
{ |
1865 |
|
*currMV = newMV; |
1866 |
|
iMinSAD = iSAD; |
1867 |
|
} |
1868 |
|
} |
1869 |
|
} |
1870 |
|
|
1871 |
|
/*************** Choose best MV found **************/ |
1872 |
|
|
1873 |
|
EPZS8_Terminate_with_Refine: |
1874 |
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
1875 |
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
1876 |
|
x, y, |
1877 |
|
currMV, iMinSAD, |
1878 |
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
1879 |
|
|
1880 |
|
EPZS8_Terminate_without_Refine: |
1881 |
|
|
1882 |
|
currPMV->x = currMV->x - pmv[0].x; |
1883 |
|
currPMV->y = currMV->y - pmv[0].y; |
1884 |
|
return iMinSAD; |
1885 |
|
} |
1886 |
|
|
1887 |
|
|
1888 |
|
|
1889 |
|
|
1890 |
|
|
1891 |
|
/* *********************************************************** |
1892 |
|
bvop motion estimation |
1893 |
|
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
1894 |
|
***************************************************************/ |
1895 |
|
|
1896 |
|
/* |
1897 |
|
void MotionEstimationBVOP( |
1898 |
|
MBParam * const pParam, |
1899 |
|
FRAMEINFO * const frame, |
1900 |
|
|
1901 |
|
// forward (past) reference |
1902 |
|
const MACROBLOCK * const f_mbs, |
1903 |
|
const IMAGE * const f_ref, |
1904 |
|
const IMAGE * const f_refH, |
1905 |
|
const IMAGE * const f_refV, |
1906 |
|
const IMAGE * const f_refHV, |
1907 |
|
// backward (future) reference |
1908 |
|
const MACROBLOCK * const b_mbs, |
1909 |
|
const IMAGE * const b_ref, |
1910 |
|
const IMAGE * const b_refH, |
1911 |
|
const IMAGE * const b_refV, |
1912 |
|
const IMAGE * const b_refHV) |
1913 |
|
{ |
1914 |
|
const uint32_t mb_width = pParam->mb_width; |
1915 |
|
const uint32_t mb_height = pParam->mb_height; |
1916 |
|
const int32_t edged_width = pParam->edged_width; |
1917 |
|
|
1918 |
|
int32_t i,j; |
1919 |
|
|
1920 |
|
int32_t f_sad16; |
1921 |
|
int32_t b_sad16; |
1922 |
|
int32_t i_sad16; |
1923 |
|
int32_t d_sad16; |
1924 |
|
int32_t best_sad; |
1925 |
|
|
1926 |
|
VECTOR pmv_dontcare; |
1927 |
|
|
1928 |
|
// note: i==horizontal, j==vertical |
1929 |
|
for (j = 0; j < mb_height; j++) |
1930 |
|
{ |
1931 |
|
for (i = 0; i < mb_width; i++) |
1932 |
|
{ |
1933 |
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
1934 |
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
1935 |
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
1936 |
|
|
1937 |
|
if (b_mb->mode == MODE_INTER |
1938 |
|
&& b_mb->cbp == 0 |
1939 |
|
&& b_mb->mvs[0].x == 0 |
1940 |
|
&& b_mb->mvs[0].y == 0) |
1941 |
|
{ |
1942 |
|
mb->mode = MB_IGNORE; |
1943 |
|
mb->mvs[0].x = 0; |
1944 |
|
mb->mvs[0].y = 0; |
1945 |
|
mb->b_mvs[0].x = 0; |
1946 |
|
mb->b_mvs[0].y = 0; |
1947 |
|
continue; |
1948 |
|
} |
1949 |
|
|
1950 |
|
|
1951 |
|
// forward search |
1952 |
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1953 |
|
&frame->image, |
1954 |
|
i, j, |
1955 |
|
frame->motion_flags, frame->quant, frame->fcode, |
1956 |
|
pParam, |
1957 |
|
f_mbs, |
1958 |
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
1959 |
|
|
1960 |
|
// backward search |
1961 |
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1962 |
|
&frame->image, |
1963 |
|
i, j, |
1964 |
|
frame->motion_flags, frame->quant, frame->bcode, |
1965 |
|
pParam, |
1966 |
|
b_mbs, |
1967 |
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
1968 |
|
|
1969 |
|
// interpolate search (simple, but effective) |
1970 |
|
i_sad16 = sad16bi_c( |
1971 |
|
frame->image.y + i*16 + j*16*edged_width, |
1972 |
|
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1973 |
|
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
1974 |
|
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1975 |
|
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
1976 |
|
edged_width); |
1977 |
|
|
1978 |
|
// TODO: direct search |
1979 |
|
// predictor + range of [-32,32] |
1980 |
|
d_sad16 = 65535; |
1981 |
|
|
1982 |
|
|
1983 |
|
if (f_sad16 < b_sad16) |
1984 |
|
{ |
1985 |
|
best_sad = f_sad16; |
1986 |
|
mb->mode = MB_FORWARD; |
1987 |
|
} |
1988 |
|
else |
1989 |
|
{ |
1990 |
|
best_sad = b_sad16; |
1991 |
|
mb->mode = MB_BACKWARD; |
1992 |
|
} |
1993 |
|
|
1994 |
|
if (i_sad16 < best_sad) |
1995 |
|
{ |
1996 |
|
best_sad = i_sad16; |
1997 |
|
mb->mode = MB_INTERPOLATE; |
1998 |
|
} |
1999 |
|
|
2000 |
|
if (d_sad16 < best_sad) |
2001 |
|
{ |
2002 |
|
best_sad = d_sad16; |
2003 |
|
mb->mode = MB_DIRECT; |
2004 |
|
} |
2005 |
|
|
2006 |
|
} |
2007 |
|
} |
2008 |
|
} |
2009 |
|
|
2010 |
|
*/ |