1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
4 |
* motion estimation |
* - Motion Estimation module - |
5 |
* |
* |
6 |
* This program is an implementation of a part of one or more MPEG-4 |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* 2002 Michael Militzer <michael@xvid.org> |
|
* to use this software module in hardware or software products are |
|
|
* advised that its use may infringe existing patents or copyrights, and |
|
|
* any such use would be at such party's own risk. The original |
|
|
* developer of this software module and his/her company, and subsequent |
|
|
* editors and their companies, will have no liability for use of this |
|
|
* software or modifications or derivatives thereof. |
|
8 |
* |
* |
9 |
* This program is free software; you can redistribute it and/or modify |
* This file is part of XviD, a free MPEG-4 video encoder/decoder |
10 |
* it under the terms of the GNU General Public License as published by |
* |
11 |
|
* XviD is free software; you can redistribute it and/or modify it |
12 |
|
* under the terms of the GNU General Public License as published by |
13 |
* the Free Software Foundation; either version 2 of the License, or |
* the Free Software Foundation; either version 2 of the License, or |
14 |
* (at your option) any later version. |
* (at your option) any later version. |
15 |
* |
* |
20 |
* |
* |
21 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
22 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
23 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
24 |
* |
* |
25 |
*************************************************************************/ |
* Under section 8 of the GNU General Public License, the copyright |
26 |
|
* holders of XVID explicitly forbid distribution in the following |
27 |
/************************************************************************** |
* countries: |
28 |
|
* |
29 |
|
* - Japan |
30 |
|
* - United States of America |
31 |
* |
* |
32 |
* Modifications: |
* Linking XviD statically or dynamically with other modules is making a |
33 |
|
* combined work based on XviD. Thus, the terms and conditions of the |
34 |
|
* GNU General Public License cover the whole combination. |
35 |
* |
* |
36 |
* 01.05.2002 updated MotionEstimationBVOP |
* As a special exception, the copyright holders of XviD give you |
37 |
* 25.04.2002 partial prevMB conversion |
* permission to link XviD with independent modules that communicate with |
38 |
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
* XviD solely through the VFW1.1 and DShow interfaces, regardless of the |
39 |
* 14.04.2002 added MotionEstimationBVOP() |
* license terms of these independent modules, and to copy and distribute |
40 |
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
* the resulting combined work under terms of your choice, provided that |
41 |
* EPZS and EPZS^2 |
* every copy of the combined work is accompanied by a complete copy of |
42 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* the source code of XviD (the version of XviD used to produce the |
43 |
* PMVFast_Refine to support multiple searches with different start points |
* combined work), being distributed under the terms of the GNU General |
44 |
* 07.01.2002 uv-block-based interpolation |
* Public License plus this exception. An independent module is a module |
45 |
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
* which is not derived from or based on XviD. |
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
46 |
* |
* |
47 |
* Michael Militzer <isibaar@videocoding.de> |
* Note that people who make modified versions of XviD are not obligated |
48 |
|
* to grant this special exception for their modified versions; it is |
49 |
|
* their choice whether to do so. The GNU General Public License gives |
50 |
|
* permission to release a modified version without this exception; this |
51 |
|
* exception also makes it possible to release a modified version which |
52 |
|
* carries forward this exception. |
53 |
* |
* |
54 |
**************************************************************************/ |
* $Id$ |
55 |
|
* |
56 |
|
*************************************************************************/ |
57 |
|
|
58 |
#include <assert.h> |
#include <assert.h> |
59 |
#include <stdio.h> |
#include <stdio.h> |
89 |
|
|
90 |
|
|
91 |
|
|
92 |
// mv.length table |
/* mv.length table */ |
93 |
static const uint32_t mvtab[33] = { |
static const uint32_t mvtab[33] = { |
94 |
1, 2, 3, 4, 6, 7, 7, 7, |
1, 2, 3, 4, 6, 7, 7, 7, |
95 |
9, 9, 9, 10, 10, 10, 10, 10, |
9, 9, 9, 10, 10, 10, 10, 10, |
164 |
static const VECTOR zeroMV = { 0, 0 }; |
static const VECTOR zeroMV = { 0, 0 }; |
165 |
VECTOR predMV; |
VECTOR predMV; |
166 |
|
|
167 |
int32_t x, y; |
uint32_t x, y; |
168 |
int32_t iIntra = 0; |
uint32_t iIntra = 0; |
169 |
VECTOR pmv; |
VECTOR pmv; |
170 |
|
|
171 |
if (sadInit) |
if (sadInit) |
379 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
380 |
} |
} |
381 |
|
|
382 |
|
#if 0 |
383 |
/* too slow and not fully functional at the moment */ |
/* too slow and not fully functional at the moment */ |
|
/* |
|
384 |
int32_t ZeroSearch16( |
int32_t ZeroSearch16( |
385 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
386 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
419 |
return iSAD; |
return iSAD; |
420 |
|
|
421 |
} |
} |
422 |
*/ |
#endif /* 0 */ |
423 |
|
|
424 |
int32_t |
int32_t |
425 |
Diamond16_MainSearch(const uint8_t * const pRef, |
Diamond16_MainSearch(const uint8_t * const pRef, |
713 |
const uint8_t * const cur, |
const uint8_t * const cur, |
714 |
const int x, |
const int x, |
715 |
const int y, |
const int y, |
716 |
int start_x, |
const int start_xi, |
717 |
int start_y, |
const int start_yi, |
718 |
int iMinSAD, |
int iMinSAD, |
719 |
VECTOR * const currMV, |
VECTOR * const currMV, |
720 |
const int center_x, |
const int center_x, |
731 |
{ |
{ |
732 |
|
|
733 |
int32_t iSAD; |
int32_t iSAD; |
734 |
|
int start_x = start_xi, start_y = start_yi; |
735 |
|
|
736 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
737 |
|
|
745 |
|
|
746 |
do { |
do { |
747 |
iDirection = 0; |
iDirection = 0; |
748 |
if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
if (bDirection & 1) /*we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) */ |
749 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
750 |
|
|
751 |
if (bDirection & 2) |
if (bDirection & 2) |
759 |
|
|
760 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
761 |
|
|
762 |
if (iDirection) //checking if anything found |
if (iDirection) /*checking if anything found */ |
763 |
{ |
{ |
764 |
bDirection = iDirection; |
bDirection = iDirection; |
765 |
iDirection = 0; |
iDirection = 0; |
766 |
start_x = currMV->x; |
start_x = currMV->x; |
767 |
start_y = currMV->y; |
start_y = currMV->y; |
768 |
if (bDirection & 3) //our candidate is left or right |
if (bDirection & 3) /*our candidate is left or right */ |
769 |
{ |
{ |
770 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
771 |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
772 |
} else // what remains here is up or down |
} else /* what remains here is up or down */ |
773 |
{ |
{ |
774 |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
775 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
780 |
start_x = currMV->x; |
start_x = currMV->x; |
781 |
start_y = currMV->y; |
start_y = currMV->y; |
782 |
} |
} |
783 |
} else //about to quit, eh? not so fast.... |
} else /*about to quit, eh? not so fast.... */ |
784 |
{ |
{ |
785 |
switch (bDirection) { |
switch (bDirection) { |
786 |
case 2: |
case 2: |
840 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
841 |
start_y + iDiamondSize, 1 + 8); |
start_y + iDiamondSize, 1 + 8); |
842 |
break; |
break; |
843 |
default: //1+2+4+8 == we didn't find anything at all |
default: /*1+2+4+8 == we didn't find anything at all */ |
844 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
845 |
start_y - iDiamondSize, 1 + 4); |
start_y - iDiamondSize, 1 + 4); |
846 |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, |
852 |
break; |
break; |
853 |
} |
} |
854 |
if (!iDirection) |
if (!iDirection) |
855 |
break; //ok, the end. really |
break; /*ok, the end. really */ |
856 |
else { |
else { |
857 |
bDirection = iDirection; |
bDirection = iDirection; |
858 |
start_x = currMV->x; |
start_x = currMV->x; |
860 |
} |
} |
861 |
} |
} |
862 |
} |
} |
863 |
while (1); //forever |
while (1); /*forever */ |
|
} |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL(X,Y) { \ |
|
|
if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ |
|
|
&& ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_F_INTERPOL_FOUND(X,Y) { \ |
|
|
if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ |
|
|
&& ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); iFound=0;} } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL(X,Y) { \ |
|
|
if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ |
|
|
&& ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_B_INTERPOL_FOUND(X,Y) { \ |
|
|
if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ |
|
|
&& ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16bi( cur, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ |
|
|
iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); iFound=0;} } \ |
|
864 |
} |
} |
865 |
|
|
|
int32_t |
|
|
Diamond16_InterpolMainSearch( |
|
|
const uint8_t * const f_pRef, |
|
|
const uint8_t * const f_pRefH, |
|
|
const uint8_t * const f_pRefV, |
|
|
const uint8_t * const f_pRefHV, |
|
|
|
|
|
const uint8_t * const cur, |
|
|
|
|
|
const uint8_t * const b_pRef, |
|
|
const uint8_t * const b_pRefH, |
|
|
const uint8_t * const b_pRefV, |
|
|
const uint8_t * const b_pRefHV, |
|
|
|
|
|
const int x, |
|
|
const int y, |
|
|
|
|
|
const int f_start_x, |
|
|
const int f_start_y, |
|
|
const int b_start_x, |
|
|
const int b_start_y, |
|
|
|
|
|
int iMinSAD, |
|
|
VECTOR * const f_currMV, |
|
|
VECTOR * const b_currMV, |
|
|
|
|
|
const int f_center_x, |
|
|
const int f_center_y, |
|
|
const int b_center_x, |
|
|
const int b_center_y, |
|
|
|
|
|
const int32_t f_min_dx, |
|
|
const int32_t f_max_dx, |
|
|
const int32_t f_min_dy, |
|
|
const int32_t f_max_dy, |
|
|
|
|
|
const int32_t b_min_dx, |
|
|
const int32_t b_max_dx, |
|
|
const int32_t b_min_dy, |
|
|
const int32_t b_max_dy, |
|
|
|
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
|
|
|
const int32_t f_iFcode, |
|
|
const int32_t b_iFcode, |
|
|
|
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iSAD; |
|
|
|
|
|
VECTOR f_backupMV; |
|
|
VECTOR b_backupMV; |
|
|
|
|
|
f_currMV->x = f_start_x; |
|
|
f_currMV->y = f_start_y; |
|
|
b_currMV->x = b_start_x; |
|
|
b_currMV->y = b_start_y; |
|
|
|
|
|
do |
|
|
{ |
|
|
iFound = 1; |
|
|
|
|
|
f_backupMV = *f_currMV; |
|
|
|
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x - iDiamondSize, f_backupMV.y); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x + iDiamondSize, f_backupMV.y); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y - iDiamondSize); |
|
|
CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y + iDiamondSize); |
|
|
|
|
|
b_backupMV = *b_currMV; |
|
|
|
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x - iDiamondSize, b_backupMV.y); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x + iDiamondSize, b_backupMV.y); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y - iDiamondSize); |
|
|
CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y + iDiamondSize); |
|
|
|
|
|
} while (!iFound); |
|
|
|
|
866 |
return iMinSAD; |
return iMinSAD; |
867 |
} |
} |
868 |
|
|
|
/* Sorry, these MACROS really got too large... I'll turn them into function soon! */ |
|
|
|
|
|
#define CHECK_MV16_DIRECT_FOUND(X,Y) \ |
|
|
if ( (X)>=(-32) && (X)<=(31) && ((Y)>=-32) && ((Y)<=31) ) \ |
|
|
{ int k;\ |
|
|
VECTOR mvs,b_mvs; \ |
|
|
iSAD = 0;\ |
|
|
for (k = 0; k < 4; k++) { \ |
|
|
mvs.x = (int32_t) ((TRB * directmv[k].x) / TRD + (X)); \ |
|
|
b_mvs.x = (int32_t) (((X) == 0) \ |
|
|
? ((TRB - TRD) * directmv[k].x) / TRD \ |
|
|
: mvs.x - directmv[k].x); \ |
|
|
\ |
|
|
mvs.y = (int32_t) ((TRB * directmv[k].y) / TRD + (Y)); \ |
|
|
b_mvs.y = (int32_t) (((Y) == 0) \ |
|
|
? ((TRB - TRD) * directmv[k].y) / TRD \ |
|
|
: mvs.y - directmv[k].y); \ |
|
|
\ |
|
|
if ( (mvs.x <= max_dx) && (mvs.x >= min_dx) \ |
|
|
&& (mvs.y <= max_dy) && (mvs.y >= min_dy) \ |
|
|
&& (b_mvs.x <= max_dx) && (b_mvs.x >= min_dx) \ |
|
|
&& (b_mvs.y <= max_dy) && (b_mvs.y >= min_dy) ) { \ |
|
|
iSAD += sad8bi( cur + 8*(k&1) + 8*(k>>1)*iEdgedWidth, \ |
|
|
get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ |
|
|
mvs.x, mvs.y, iEdgedWidth), \ |
|
|
get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ |
|
|
b_mvs.x, b_mvs.y, iEdgedWidth), \ |
|
|
iEdgedWidth); \ |
|
|
} \ |
|
|
else \ |
|
|
iSAD = 65535; \ |
|
|
} \ |
|
|
iSAD += calc_delta_16((X),(Y), 1, iQuant);\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iFound=0; } \ |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
Diamond16_DirectMainSearch( |
|
|
const uint8_t * const f_pRef, |
|
|
const uint8_t * const f_pRefH, |
|
|
const uint8_t * const f_pRefV, |
|
|
const uint8_t * const f_pRefHV, |
|
|
|
|
|
const uint8_t * const cur, |
|
|
|
|
|
const uint8_t * const b_pRef, |
|
|
const uint8_t * const b_pRefH, |
|
|
const uint8_t * const b_pRefV, |
|
|
const uint8_t * const b_pRefHV, |
|
|
|
|
|
const int x, |
|
|
const int y, |
|
|
|
|
|
const int TRB, |
|
|
const int TRD, |
|
|
|
|
|
const int start_x, |
|
|
const int start_y, |
|
|
|
|
|
int iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const directmv, |
|
|
|
|
|
const int32_t min_dx, |
|
|
const int32_t max_dx, |
|
|
const int32_t min_dy, |
|
|
const int32_t max_dy, |
|
|
|
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
|
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iSAD; |
|
|
|
|
|
VECTOR backupMV; |
|
|
|
|
|
currMV->x = start_x; |
|
|
currMV->y = start_y; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
do |
|
|
{ |
|
|
iFound = 1; |
|
|
|
|
|
backupMV = *currMV; |
|
|
|
|
|
CHECK_MV16_DIRECT_FOUND(backupMV.x - iDiamondSize, backupMV.y); |
|
|
CHECK_MV16_DIRECT_FOUND(backupMV.x + iDiamondSize, backupMV.y); |
|
|
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y - iDiamondSize); |
|
|
CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y + iDiamondSize); |
|
|
|
|
|
} while (!iFound); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
869 |
int32_t |
int32_t |
870 |
AdvDiamond8_MainSearch(const uint8_t * const pRef, |
AdvDiamond8_MainSearch(const uint8_t * const pRef, |
871 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
874 |
const uint8_t * const cur, |
const uint8_t * const cur, |
875 |
const int x, |
const int x, |
876 |
const int y, |
const int y, |
877 |
int start_x, |
const int start_xi, |
878 |
int start_y, |
const int start_yi, |
879 |
int iMinSAD, |
int iMinSAD, |
880 |
VECTOR * const currMV, |
VECTOR * const currMV, |
881 |
const int center_x, |
const int center_x, |
892 |
{ |
{ |
893 |
|
|
894 |
int32_t iSAD; |
int32_t iSAD; |
895 |
|
int start_x = start_xi, start_y = start_yi; |
896 |
|
|
897 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
898 |
|
|
906 |
|
|
907 |
do { |
do { |
908 |
iDirection = 0; |
iDirection = 0; |
909 |
if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
if (bDirection & 1) /*we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) */ |
910 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
911 |
|
|
912 |
if (bDirection & 2) |
if (bDirection & 2) |
920 |
|
|
921 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
922 |
|
|
923 |
if (iDirection) //checking if anything found |
if (iDirection) /*checking if anything found */ |
924 |
{ |
{ |
925 |
bDirection = iDirection; |
bDirection = iDirection; |
926 |
iDirection = 0; |
iDirection = 0; |
927 |
start_x = currMV->x; |
start_x = currMV->x; |
928 |
start_y = currMV->y; |
start_y = currMV->y; |
929 |
if (bDirection & 3) //our candidate is left or right |
if (bDirection & 3) /*our candidate is left or right */ |
930 |
{ |
{ |
931 |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); |
932 |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); |
933 |
} else // what remains here is up or down |
} else /* what remains here is up or down */ |
934 |
{ |
{ |
935 |
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); |
936 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); |
941 |
start_x = currMV->x; |
start_x = currMV->x; |
942 |
start_y = currMV->y; |
start_y = currMV->y; |
943 |
} |
} |
944 |
} else //about to quit, eh? not so fast.... |
} else /*about to quit, eh? not so fast.... */ |
945 |
{ |
{ |
946 |
switch (bDirection) { |
switch (bDirection) { |
947 |
case 2: |
case 2: |
1000 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
1001 |
start_y + iDiamondSize, 1 + 8); |
start_y + iDiamondSize, 1 + 8); |
1002 |
break; |
break; |
1003 |
default: //1+2+4+8 == we didn't find anything at all |
default: /*1+2+4+8 == we didn't find anything at all */ |
1004 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
1005 |
start_y - iDiamondSize, 1 + 4); |
start_y - iDiamondSize, 1 + 4); |
1006 |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, |
1012 |
break; |
break; |
1013 |
} |
} |
1014 |
if (!(iDirection)) |
if (!(iDirection)) |
1015 |
break; //ok, the end. really |
break; /*ok, the end. really */ |
1016 |
else { |
else { |
1017 |
bDirection = iDirection; |
bDirection = iDirection; |
1018 |
start_x = currMV->x; |
start_x = currMV->x; |
1020 |
} |
} |
1021 |
} |
} |
1022 |
} |
} |
1023 |
while (1); //forever |
while (1); /*forever */ |
1024 |
} |
} |
1025 |
return iMinSAD; |
return iMinSAD; |
1026 |
} |
} |
1172 |
} |
} |
1173 |
|
|
1174 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
1175 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
1176 |
|
|
1177 |
if ((x == 0) && (y == 0)) { |
if ((x == 0) && (y == 0)) { |
1229 |
if ((iMinSAD < 256) || |
if ((iMinSAD < 256) || |
1230 |
((MVequal(*currMV, prevMB->mvs[0])) && |
((MVequal(*currMV, prevMB->mvs[0])) && |
1231 |
((int32_t) iMinSAD < prevMB->sad16))) { |
((int32_t) iMinSAD < prevMB->sad16))) { |
1232 |
if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode |
if (iMinSAD < (int)(2 * iQuant)) /* high chances for SKIP-mode */ |
1233 |
{ |
{ |
1234 |
if (!MVzero(*currMV)) { |
if (!MVzero(*currMV)) { |
1235 |
iMinSAD += MV16_00_BIAS; |
iMinSAD += MV16_00_BIAS; |
1236 |
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
CHECK_MV16_ZERO; /* (0,0) saves space for letterboxed pictures */ |
1237 |
iMinSAD -= MV16_00_BIAS; |
iMinSAD -= MV16_00_BIAS; |
1238 |
} |
} |
1239 |
} |
} |
1258 |
*/ |
*/ |
1259 |
|
|
1260 |
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
1261 |
iDiamondSize = 1; // halfpel! |
iDiamondSize = 1; /* halfpel! */ |
1262 |
else |
else |
1263 |
iDiamondSize = 2; // halfpel! |
iDiamondSize = 2; /* halfpel! */ |
1264 |
|
|
1265 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16)) |
if (!(MotionFlags & PMV_HALFPELDIAMOND16)) |
1266 |
iDiamondSize *= 2; |
iDiamondSize *= 2; |
1272 |
If MV is (0,0) subtract offset. |
If MV is (0,0) subtract offset. |
1273 |
*/ |
*/ |
1274 |
|
|
1275 |
// (0,0) is always possible |
/* (0,0) is always possible */ |
1276 |
|
|
1277 |
if (!MVzero(pmv[0])) |
if (!MVzero(pmv[0])) |
1278 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
1279 |
|
|
1280 |
// previous frame MV is always possible |
/* previous frame MV is always possible */ |
1281 |
|
|
1282 |
if (!MVzero(prevMB->mvs[0])) |
if (!MVzero(prevMB->mvs[0])) |
1283 |
if (!MVequal(prevMB->mvs[0], pmv[0])) |
if (!MVequal(prevMB->mvs[0], pmv[0])) |
1284 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
1285 |
|
|
1286 |
// left neighbour, if allowed |
/* left neighbour, if allowed */ |
1287 |
|
|
1288 |
if (!MVzero(pmv[1])) |
if (!MVzero(pmv[1])) |
1289 |
if (!MVequal(pmv[1], prevMB->mvs[0])) |
if (!MVequal(pmv[1], prevMB->mvs[0])) |
1295 |
|
|
1296 |
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
1297 |
} |
} |
1298 |
// top neighbour, if allowed |
/* top neighbour, if allowed */ |
1299 |
if (!MVzero(pmv[2])) |
if (!MVzero(pmv[2])) |
1300 |
if (!MVequal(pmv[2], prevMB->mvs[0])) |
if (!MVequal(pmv[2], prevMB->mvs[0])) |
1301 |
if (!MVequal(pmv[2], pmv[0])) |
if (!MVequal(pmv[2], pmv[0])) |
1306 |
} |
} |
1307 |
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
1308 |
|
|
1309 |
// top right neighbour, if allowed |
/* top right neighbour, if allowed */ |
1310 |
if (!MVzero(pmv[3])) |
if (!MVzero(pmv[3])) |
1311 |
if (!MVequal(pmv[3], prevMB->mvs[0])) |
if (!MVequal(pmv[3], prevMB->mvs[0])) |
1312 |
if (!MVequal(pmv[3], pmv[0])) |
if (!MVequal(pmv[3], pmv[0])) |
1409 |
*/ |
*/ |
1410 |
|
|
1411 |
PMVfast16_Terminate_with_Refine: |
PMVfast16_Terminate_with_Refine: |
1412 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) /* perform final half-pel step */ |
1413 |
iMinSAD = |
iMinSAD = |
1414 |
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
1415 |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
1434 |
const uint8_t * const cur, |
const uint8_t * const cur, |
1435 |
const int x, |
const int x, |
1436 |
const int y, |
const int y, |
1437 |
int32_t start_x, |
const int32_t start_x, |
1438 |
int32_t start_y, |
const int32_t start_y, |
1439 |
int32_t iMinSAD, |
int32_t iMinSAD, |
1440 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1441 |
const int center_x, |
const int center_x, |
1470 |
if (iDirection) { |
if (iDirection) { |
1471 |
while (!iFound) { |
while (!iFound) { |
1472 |
iFound = 1; |
iFound = 1; |
1473 |
backupMV = *currMV; // since iDirection!=0, this is well defined! |
backupMV = *currMV; /* since iDirection!=0, this is well defined! */ |
1474 |
iDirectionBackup = iDirection; |
iDirectionBackup = iDirection; |
1475 |
|
|
1476 |
if (iDirectionBackup != 2) |
if (iDirectionBackup != 2) |
1504 |
const uint8_t * const cur, |
const uint8_t * const cur, |
1505 |
const int x, |
const int x, |
1506 |
const int y, |
const int y, |
1507 |
int32_t start_x, |
const int32_t start_x, |
1508 |
int32_t start_y, |
const int32_t start_y, |
1509 |
int32_t iMinSAD, |
int32_t iMinSAD, |
1510 |
VECTOR * const currMV, |
VECTOR * const currMV, |
1511 |
const int center_x, |
const int center_x, |
1759 |
VECTOR backupMV; |
VECTOR backupMV; |
1760 |
VECTOR startMV; |
VECTOR startMV; |
1761 |
|
|
1762 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
/* const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; */ |
1763 |
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
1764 |
|
|
1765 |
int32_t threshA, threshB; |
int32_t threshA, threshB; |
1786 |
} |
} |
1787 |
|
|
1788 |
/* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, (x >> 1), (y >> 1), iWcount, iSubBlock, pmv, psad); |
|
1789 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); |
1790 |
|
|
1791 |
if ((x == 0) && (y == 0)) { |
if ((x == 0) && (y == 0)) { |
1813 |
*/ |
*/ |
1814 |
|
|
1815 |
|
|
1816 |
// Prepare for main loop |
/* Prepare for main loop */ |
1817 |
|
|
1818 |
if (MotionFlags & PMV_USESQUARES8) |
if (MotionFlags & PMV_USESQUARES8) |
1819 |
MainSearchPtr = Square8_MainSearch; |
MainSearchPtr = Square8_MainSearch; |
1857 |
*/ |
*/ |
1858 |
|
|
1859 |
if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) |
if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) |
1860 |
iDiamondSize = 1; // 1 halfpel! |
iDiamondSize = 1; /* 1 halfpel! */ |
1861 |
else |
else |
1862 |
iDiamondSize = 2; // 2 halfpel = 1 full pixel! |
iDiamondSize = 2; /* 2 halfpel = 1 full pixel! */ |
1863 |
|
|
1864 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
1865 |
iDiamondSize *= 2; |
iDiamondSize *= 2; |
1872 |
If MV is (0,0) subtract offset. |
If MV is (0,0) subtract offset. |
1873 |
*/ |
*/ |
1874 |
|
|
1875 |
// the median prediction might be even better than mv16 |
/* the median prediction might be even better than mv16 */ |
1876 |
|
|
1877 |
if (!MVequal(pmv[0], startMV)) |
if (!MVequal(pmv[0], startMV)) |
1878 |
CHECK_MV8_CANDIDATE(center_x, center_y); |
CHECK_MV8_CANDIDATE(center_x, center_y); |
1879 |
|
|
1880 |
// (0,0) if needed |
/* (0,0) if needed */ |
1881 |
if (!MVzero(pmv[0])) |
if (!MVzero(pmv[0])) |
1882 |
if (!MVzero(startMV)) |
if (!MVzero(startMV)) |
1883 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
1884 |
|
|
1885 |
// previous frame MV if needed |
/* previous frame MV if needed */ |
1886 |
if (!MVzero(prevMB->mvs[iSubBlock])) |
if (!MVzero(prevMB->mvs[iSubBlock])) |
1887 |
if (!MVequal(prevMB->mvs[iSubBlock], startMV)) |
if (!MVequal(prevMB->mvs[iSubBlock], startMV)) |
1888 |
if (!MVequal(prevMB->mvs[iSubBlock], pmv[0])) |
if (!MVequal(prevMB->mvs[iSubBlock], pmv[0])) |
1898 |
goto PMVfast8_Terminate_with_Refine; |
goto PMVfast8_Terminate_with_Refine; |
1899 |
} |
} |
1900 |
|
|
1901 |
// left neighbour, if allowed and needed |
/* left neighbour, if allowed and needed */ |
1902 |
if (!MVzero(pmv[1])) |
if (!MVzero(pmv[1])) |
1903 |
if (!MVequal(pmv[1], startMV)) |
if (!MVequal(pmv[1], startMV)) |
1904 |
if (!MVequal(pmv[1], prevMB->mvs[iSubBlock])) |
if (!MVequal(pmv[1], prevMB->mvs[iSubBlock])) |
1909 |
} |
} |
1910 |
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
1911 |
} |
} |
1912 |
// top neighbour, if allowed and needed |
/* top neighbour, if allowed and needed */ |
1913 |
if (!MVzero(pmv[2])) |
if (!MVzero(pmv[2])) |
1914 |
if (!MVequal(pmv[2], startMV)) |
if (!MVequal(pmv[2], startMV)) |
1915 |
if (!MVequal(pmv[2], prevMB->mvs[iSubBlock])) |
if (!MVequal(pmv[2], prevMB->mvs[iSubBlock])) |
1921 |
} |
} |
1922 |
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
1923 |
|
|
1924 |
// top right neighbour, if allowed and needed |
/* top right neighbour, if allowed and needed */ |
1925 |
if (!MVzero(pmv[3])) |
if (!MVzero(pmv[3])) |
1926 |
if (!MVequal(pmv[3], startMV)) |
if (!MVequal(pmv[3], startMV)) |
1927 |
if (!MVequal(pmv[3], prevMB->mvs[iSubBlock])) |
if (!MVequal(pmv[3], prevMB->mvs[iSubBlock])) |
2016 |
*/ |
*/ |
2017 |
|
|
2018 |
PMVfast8_Terminate_with_Refine: |
PMVfast8_Terminate_with_Refine: |
2019 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) /* perform final half-pel step */ |
2020 |
iMinSAD = |
iMinSAD = |
2021 |
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
2022 |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
2073 |
|
|
2074 |
static MACROBLOCK *oldMBs = NULL; |
static MACROBLOCK *oldMBs = NULL; |
2075 |
|
|
2076 |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
/* const MACROBLOCK * const pMB = pMBs + x + y * iWcount; */ |
2077 |
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
2078 |
MACROBLOCK *oldMB = NULL; |
MACROBLOCK *oldMB = NULL; |
2079 |
|
|
2085 |
|
|
2086 |
if (oldMBs == NULL) { |
if (oldMBs == NULL) { |
2087 |
oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); |
oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); |
2088 |
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
/* fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); */ |
2089 |
} |
} |
2090 |
oldMB = oldMBs + x + y * iWcount; |
oldMB = oldMBs + x + y * iWcount; |
2091 |
|
|
2100 |
max_dy = EVEN(max_dy); |
max_dy = EVEN(max_dy); |
2101 |
} |
} |
2102 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
2103 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
2104 |
|
|
2105 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
2109 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
2110 |
*/ |
*/ |
2111 |
|
|
2112 |
// Prepare for main loop |
/* Prepare for main loop */ |
2113 |
|
|
2114 |
currMV->x = start_x; |
currMV->x = start_x; |
2115 |
currMV->y = start_y; |
currMV->y = start_y; |
2138 |
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
2139 |
(uint8_t) iFcode, iQuant); |
(uint8_t) iFcode, iQuant); |
2140 |
|
|
2141 |
// thresh1 is fixed to 256 |
/* thresh1 is fixed to 256 */ |
2142 |
if ((iMinSAD < 256) || |
if ((iMinSAD < 256) || |
2143 |
((MVequal(*currMV, prevMB->mvs[0])) && |
((MVequal(*currMV, prevMB->mvs[0])) && |
2144 |
((int32_t) iMinSAD < prevMB->sad16))) { |
((int32_t) iMinSAD < prevMB->sad16))) { |
2150 |
|
|
2151 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
2152 |
|
|
2153 |
// previous frame MV |
/* previous frame MV */ |
2154 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); |
2155 |
|
|
2156 |
// set threshhold based on Min of Prediction and SAD of collocated block |
/* set threshhold based on Min of Prediction and SAD of collocated block */ |
2157 |
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
/* CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want */ |
2158 |
|
|
2159 |
if ((x == 0) && (y == 0)) { |
if ((x == 0) && (y == 0)) { |
2160 |
thresh2 = 512; |
thresh2 = 512; |
2164 |
thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; |
thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; |
2165 |
} |
} |
2166 |
|
|
2167 |
// MV=(0,0) is often a good choice |
/* MV=(0,0) is often a good choice */ |
2168 |
|
|
2169 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
2170 |
|
|
2171 |
|
|
2172 |
// left neighbour, if allowed |
/* left neighbour, if allowed */ |
2173 |
if (x != 0) { |
if (x != 0) { |
2174 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
2175 |
pmv[1].x = EVEN(pmv[1].x); |
pmv[1].x = EVEN(pmv[1].x); |
2177 |
} |
} |
2178 |
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
2179 |
} |
} |
2180 |
// top neighbour, if allowed |
/* top neighbour, if allowed */ |
2181 |
if (y != 0) { |
if (y != 0) { |
2182 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
2183 |
pmv[2].x = EVEN(pmv[2].x); |
pmv[2].x = EVEN(pmv[2].x); |
2185 |
} |
} |
2186 |
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
2187 |
|
|
2188 |
// top right neighbour, if allowed |
/* top right neighbour, if allowed */ |
2189 |
if ((uint32_t) x != (iWcount - 1)) { |
if ((uint32_t) x != (iWcount - 1)) { |
2190 |
if (!(MotionFlags & PMV_HALFPEL16)) { |
if (!(MotionFlags & PMV_HALFPEL16)) { |
2191 |
pmv[3].x = EVEN(pmv[3].x); |
pmv[3].x = EVEN(pmv[3].x); |
2210 |
|
|
2211 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
2212 |
|
|
2213 |
backupMV = prevMB->mvs[0]; // collocated MV |
backupMV = prevMB->mvs[0]; /* collocated MV */ |
2214 |
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x); // acceleration X |
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x); /* acceleration X */ |
2215 |
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y); // acceleration Y |
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y); /* acceleration Y */ |
2216 |
|
|
2217 |
CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); |
CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); |
2218 |
|
|
2219 |
// left neighbour |
/* left neighbour */ |
2220 |
if (x != 0) |
if (x != 0) |
2221 |
CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); |
CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); |
2222 |
|
|
2223 |
// top neighbour |
/* top neighbour */ |
2224 |
if (y != 0) |
if (y != 0) |
2225 |
CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, |
CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, |
2226 |
(prevMB - iWcount)->mvs[0].y); |
(prevMB - iWcount)->mvs[0].y); |
2227 |
|
|
2228 |
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
/* right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs */ |
2229 |
|
|
2230 |
if ((uint32_t) x != iWcount - 1) |
if ((uint32_t) x != iWcount - 1) |
2231 |
CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); |
CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); |
2232 |
|
|
2233 |
// bottom neighbour, dito |
/* bottom neighbour, dito */ |
2234 |
if ((uint32_t) y != iHcount - 1) |
if ((uint32_t) y != iHcount - 1) |
2235 |
CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, |
CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, |
2236 |
(prevMB + iWcount)->mvs[0].y); |
(prevMB + iWcount)->mvs[0].y); |
2300 |
/*************** Choose best MV found **************/ |
/*************** Choose best MV found **************/ |
2301 |
|
|
2302 |
EPZS16_Terminate_with_Refine: |
EPZS16_Terminate_with_Refine: |
2303 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE16) /* perform final half-pel step */ |
2304 |
iMinSAD = |
iMinSAD = |
2305 |
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
2306 |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
2361 |
|
|
2362 |
const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); |
const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); |
2363 |
|
|
2364 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
/* const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; */ |
2365 |
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; |
2366 |
|
|
2367 |
int32_t bPredEq; |
int32_t bPredEq; |
2382 |
max_dy = EVEN(max_dy); |
max_dy = EVEN(max_dy); |
2383 |
} |
} |
2384 |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
/* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
//bPredEq = get_pmvdata(pMBs, x >> 1, y >> 1, iWcount, iSubBlock, pmv[0].x, pmv[0].y, psad); |
|
2385 |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); |
bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); |
2386 |
|
|
|
|
|
2387 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
2388 |
MinSAD=SAD |
MinSAD=SAD |
2389 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
2391 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
2392 |
*/ |
*/ |
2393 |
|
|
2394 |
// Prepare for main loop |
/* Prepare for main loop */ |
2395 |
|
|
2396 |
|
|
2397 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (!(MotionFlags & PMV_HALFPEL8)) { |
2420 |
(uint8_t) iFcode, iQuant); |
(uint8_t) iFcode, iQuant); |
2421 |
|
|
2422 |
|
|
2423 |
// thresh1 is fixed to 256 |
/* thresh1 is fixed to 256 */ |
2424 |
if (iMinSAD < 256 / 4) { |
if (iMinSAD < 256 / 4) { |
2425 |
if (MotionFlags & PMV_QUICKSTOP8) |
if (MotionFlags & PMV_QUICKSTOP8) |
2426 |
goto EPZS8_Terminate_without_Refine; |
goto EPZS8_Terminate_without_Refine; |
2431 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
2432 |
|
|
2433 |
|
|
2434 |
// MV=(0,0) is often a good choice |
/* MV=(0,0) is often a good choice */ |
2435 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
2436 |
|
|
2437 |
// previous frame MV |
/* previous frame MV */ |
2438 |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); |
2439 |
|
|
2440 |
// left neighbour, if allowed |
/* left neighbour, if allowed */ |
2441 |
if (psad[1] != MV_MAX_ERROR) { |
if (psad[1] != MV_MAX_ERROR) { |
2442 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (!(MotionFlags & PMV_HALFPEL8)) { |
2443 |
pmv[1].x = EVEN(pmv[1].x); |
pmv[1].x = EVEN(pmv[1].x); |
2445 |
} |
} |
2446 |
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); |
2447 |
} |
} |
2448 |
// top neighbour, if allowed |
/* top neighbour, if allowed */ |
2449 |
if (psad[2] != MV_MAX_ERROR) { |
if (psad[2] != MV_MAX_ERROR) { |
2450 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (!(MotionFlags & PMV_HALFPEL8)) { |
2451 |
pmv[2].x = EVEN(pmv[2].x); |
pmv[2].x = EVEN(pmv[2].x); |
2453 |
} |
} |
2454 |
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
CHECK_MV8_CANDIDATE(pmv[2].x, pmv[2].y); |
2455 |
|
|
2456 |
// top right neighbour, if allowed |
/* top right neighbour, if allowed */ |
2457 |
if (psad[3] != MV_MAX_ERROR) { |
if (psad[3] != MV_MAX_ERROR) { |
2458 |
if (!(MotionFlags & PMV_HALFPEL8)) { |
if (!(MotionFlags & PMV_HALFPEL8)) { |
2459 |
pmv[3].x = EVEN(pmv[3].x); |
pmv[3].x = EVEN(pmv[3].x); |
2463 |
} |
} |
2464 |
} |
} |
2465 |
|
|
2466 |
/* // this bias is zero anyway, at the moment! |
#if 0 |
2467 |
|
/* this bias is zero anyway, at the moment! */ |
2468 |
|
|
2469 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) /* && (iMinSAD <= iQuant * 96) */ |
2470 |
iMinSAD -= MV8_00_BIAS; |
iMinSAD -= MV8_00_BIAS; |
2471 |
|
|
2472 |
*/ |
#endif |
2473 |
|
|
2474 |
/* Terminate if MinSAD <= T_2 |
/* Terminate if MinSAD <= T_2 |
2475 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
2491 |
|
|
2492 |
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
2493 |
|
|
2494 |
// there is no EPZS^2 for inter4v at the moment |
/* there is no EPZS^2 for inter4v at the moment */ |
2495 |
|
|
2496 |
if (MotionFlags & PMV_USESQUARES8) |
if (MotionFlags & PMV_USESQUARES8) |
2497 |
MainSearchPtr = Square8_MainSearch; |
MainSearchPtr = Square8_MainSearch; |
2547 |
/*************** Choose best MV found **************/ |
/*************** Choose best MV found **************/ |
2548 |
|
|
2549 |
EPZS8_Terminate_with_Refine: |
EPZS8_Terminate_with_Refine: |
2550 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
if (MotionFlags & PMV_HALFPELREFINE8) /* perform final half-pel step */ |
2551 |
iMinSAD = |
iMinSAD = |
2552 |
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
2553 |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
2559 |
currPMV->y = currMV->y - center_y; |
currPMV->y = currMV->y - center_y; |
2560 |
return iMinSAD; |
return iMinSAD; |
2561 |
} |
} |
|
|
|
|
|
|
|
|
|
|
int32_t |
|
|
PMVfastIntSearch16(const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, |
|
|
const int y, |
|
|
const int start_x, /* start should be most likely vector */ |
|
|
const int start_y, |
|
|
const int center_x, /* center is from where length of MVs is measured */ |
|
|
const int center_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
|
|
const VECTOR zeroMV = { 0, 0 }; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
int32_t iFound; |
|
|
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
|
|
|
MainSearch16FuncPtr MainSearchPtr; |
|
|
|
|
|
const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK *const pMB = pMBs + x + y * iWcount; |
|
|
|
|
|
int32_t threshA, threshB; |
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD, iSAD; |
|
|
|
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, |
|
|
iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if ((x == 0) && (y == 0)) { |
|
|
threshA = 512; |
|
|
threshB = 1024; |
|
|
|
|
|
bPredEq = 0; |
|
|
psad[0] = psad[1] = psad[2] = psad[3] = 0; |
|
|
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
|
|
|
|
|
} else { |
|
|
threshA = psad[0]; |
|
|
threshB = threshA + 256; |
|
|
if (threshA < 512) |
|
|
threshA = 512; |
|
|
if (threshA > 1024) |
|
|
threshA = 1024; |
|
|
if (threshB > 1792) |
|
|
threshB = 1792; |
|
|
|
|
|
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
|
|
*currMV = pmv[0]; /* current best := prediction */ |
|
|
} |
|
|
|
|
|
iFound = 0; |
|
|
|
|
|
/* Step 4: Calculate SAD around the Median prediction. |
|
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
|
|
|
|
if (currMV->x > max_dx) { |
|
|
currMV->x = EVEN(max_dx); |
|
|
} |
|
|
if (currMV->x < min_dx) { |
|
|
currMV->x = EVEN(min_dx); |
|
|
} |
|
|
if (currMV->y > max_dy) { |
|
|
currMV->y = EVEN(max_dy); |
|
|
} |
|
|
if (currMV->y < min_dy) { |
|
|
currMV->y = EVEN(min_dy); |
|
|
} |
|
|
|
|
|
iMinSAD = |
|
|
sad16(cur, |
|
|
get_iref_mv(pRef, x, y, 16, currMV, |
|
|
iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += |
|
|
calc_delta_16(currMV->x - center_x, currMV->y - center_y, |
|
|
(uint8_t) iFcode, iQuant); |
|
|
|
|
|
if ((iMinSAD < 256) || |
|
|
((MVequal(*currMV, prevMB->i_mvs[0])) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
|
if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode |
|
|
{ |
|
|
if (!MVzero(*currMV)) { |
|
|
iMinSAD += MV16_00_BIAS; |
|
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
} |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) |
|
|
iFound = 2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) |
|
|
iDiamondSize = 2; // halfpel units! |
|
|
else |
|
|
iDiamondSize = 4; // halfpel units! |
|
|
|
|
|
/* |
|
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
|
|
|
|
// (0,0) is often a good choice |
|
|
|
|
|
if (!MVzero(pmv[0])) |
|
|
CHECK_MV16_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
|
|
|
if (!MVzero(prevMB->i_mvs[0])) |
|
|
if (!MVequal(prevMB->i_mvs[0], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(prevMB->i_mvs[0].x, prevMB->i_mvs[0].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
|
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[1], pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); |
|
|
|
|
|
// top neighbour, if allowed |
|
|
if (!MVzero(pmv[2])) |
|
|
if (!MVequal(pmv[2], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[2], pmv[0])) |
|
|
if (!MVequal(pmv[2], pmv[1])) |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); |
|
|
|
|
|
// top right neighbour, if allowed |
|
|
if (!MVzero(pmv[3])) |
|
|
if (!MVequal(pmv[3], prevMB->i_mvs[0])) |
|
|
if (!MVequal(pmv[3], pmv[0])) |
|
|
if (!MVequal(pmv[3], pmv[1])) |
|
|
if (!MVequal(pmv[3], pmv[2])) |
|
|
CHECK_MV16_CANDIDATE(pmv[3].x, |
|
|
pmv[3].y); |
|
|
|
|
|
if ((MVzero(*currMV)) && |
|
|
(!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
|
|
|
|
|
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
|
|
|
|
if ((iMinSAD <= threshA) || |
|
|
(MVequal(*currMV, prevMB->i_mvs[0]) && |
|
|
((int32_t) iMinSAD < prevMB->i_sad16))) { |
|
|
|
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfastInt16_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
|
|
|
/************ (Diamond Search) **************/ |
|
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
|
else |
|
|
MainSearchPtr = Diamond16_MainSearch; |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, |
|
|
currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, |
|
|
min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) { |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0], backupMV))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, |
|
|
min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
|
|
|
if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { |
|
|
iSAD = |
|
|
(*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, |
|
|
iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, |
|
|
max_dy, iEdgedWidth, iDiamondSize, iFcode, |
|
|
iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) { |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
|
|
|
|
PMVfastInt16_Terminate_with_Refine: |
|
|
|
|
|
pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; |
|
|
pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; |
|
|
|
|
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = |
|
|
Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, |
|
|
iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, |
|
|
iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) |
|
|
|
|
|
PMVfastInt16_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - center_x; |
|
|
currPMV->y = currMV->y - center_y; |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
***************************************************************/ |
|
|
|
|
|
void |
|
|
MotionEstimationBVOP(MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
|
const int32_t time_bp, |
|
|
const int32_t time_pp, |
|
|
// forward (past) reference |
|
|
const MACROBLOCK * const f_mbs, |
|
|
const IMAGE * const f_ref, |
|
|
const IMAGE * const f_refH, |
|
|
const IMAGE * const f_refV, |
|
|
const IMAGE * const f_refHV, |
|
|
// backward (future) reference |
|
|
const MACROBLOCK * const b_mbs, |
|
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
|
{ |
|
|
const int mb_width = pParam->mb_width; |
|
|
const int mb_height = pParam->mb_height; |
|
|
const int edged_width = pParam->edged_width; |
|
|
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
|
|
|
int i, j, k; |
|
|
|
|
|
static const VECTOR zeroMV={0,0}; |
|
|
|
|
|
int f_sad16; /* forward (as usual) search */ |
|
|
int b_sad16; /* backward (only in b-frames) search */ |
|
|
int i_sad16; /* interpolated (both direction, b-frames only) */ |
|
|
int d_sad16; /* direct mode (assume almost linear motion) */ |
|
|
|
|
|
int best_sad; |
|
|
|
|
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
|
|
VECTOR f_interpolMV, b_interpolMV; |
|
|
VECTOR pmv_dontcare; |
|
|
|
|
|
int min_dx, max_dx, min_dy, max_dy; |
|
|
int f_min_dx, f_max_dx, f_min_dy, f_max_dy; |
|
|
int b_min_dx, b_max_dx, b_min_dy, b_max_dy; |
|
|
|
|
|
int f_count=0; |
|
|
int b_count=0; |
|
|
int i_count=0; |
|
|
int d_count=0; |
|
|
|
|
|
const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; |
|
|
const int64_t TRD = (int32_t)time_pp; |
|
|
|
|
|
// fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); |
|
|
// note: i==horizontal, j==vertical |
|
|
for (j = 0; j < mb_height; j++) { |
|
|
|
|
|
f_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
|
b_predMV = zeroMV; |
|
|
|
|
|
for (i = 0; i < mb_width; i++) { |
|
|
MACROBLOCK *mb = &frame->mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; |
|
|
|
|
|
mb->deltamv=zeroMV; |
|
|
|
|
|
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
|
|
|
|
|
if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && |
|
|
b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->b_mvs[0] = mb->mvs[0] = zeroMV; |
|
|
continue; |
|
|
} |
|
|
|
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
{ |
|
|
d_sad16 = 0; |
|
|
/* same method of scaling as in decoder.c, so we copy from there */ |
|
|
for (k = 0; k < 4; k++) { |
|
|
|
|
|
mb->directmv[k] = b_mb->mvs[k]; |
|
|
|
|
|
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
|
|
: mb->mvs[k].x - mb->directmv[k].x); |
|
|
|
|
|
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
|
: mb->mvs[k].y - mb->directmv[k].y); |
|
|
|
|
|
d_sad16 += |
|
|
sad8bi(frame->image.y + (2*i+(k&1))*8 + (2*j+(k>>1))*8*edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
(2*i+(k&1)), (2*j+(k>>1)), 8, &mb->mvs[k], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
(2*i+(k&1)), (2*j+(k>>1)), 8, &mb->b_mvs[k], edged_width), |
|
|
edged_width); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = |
|
|
mb->directmv[0] = b_mb->mvs[0]; |
|
|
|
|
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
|
|
|
|
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
|
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
|
|
|
|
d_sad16 = sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
|
|
|
} |
|
|
d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); |
|
|
|
|
|
// forward search |
|
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, /* start point f_directMV */ |
|
|
f_predMV.x, f_predMV.y, /* center is f-prediction */ |
|
|
frame->motion_flags, |
|
|
frame->quant, frame->fcode, pParam, |
|
|
f_mbs, f_mbs, |
|
|
&mb->mvs[0], &pmv_dontcare); |
|
|
|
|
|
|
|
|
// backward search |
|
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, /* start point b_directMV */ |
|
|
b_predMV.x, b_predMV.y, /* center is b-prediction */ |
|
|
frame->motion_flags, |
|
|
frame->quant, frame->bcode, pParam, |
|
|
b_mbs, b_mbs, |
|
|
&mb->b_mvs[0], &pmv_dontcare); |
|
|
|
|
|
i_sad16 = |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
i_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, |
|
|
frame->fcode, frame->quant); |
|
|
i_sad16 += calc_delta_16(mb->b_mvs[0].x-b_predMV.x, mb->b_mvs[0].y-b_predMV.y, |
|
|
frame->bcode, frame->quant); |
|
|
|
|
|
get_range(&f_min_dx, &f_max_dx, &f_min_dy, &f_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->fcode); |
|
|
get_range(&b_min_dx, &b_max_dx, &b_min_dy, &b_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->bcode); |
|
|
|
|
|
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
|
|
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
|
|
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
|
|
as "optimal". */ |
|
|
|
|
|
i_sad16 = Diamond16_InterpolMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 2, |
|
|
frame->fcode, frame->bcode,frame->quant,0); |
|
|
|
|
|
i_sad16 = Diamond16_InterpolMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
f_interpolMV.x, f_interpolMV.y, |
|
|
b_interpolMV.x, b_interpolMV.y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 1, |
|
|
frame->fcode, frame->bcode,frame->quant,0); // equiv to halfpel refine |
|
|
|
|
|
|
|
|
/* DIRECT MODE DELTA VECTOR SEARCH. |
|
|
This has to be made more effective, but at the moment I'm happy it's running at all */ |
|
|
|
|
|
/* There are two range restrictions for direct mode: deltaMV is limited to [-32,31] in halfpel units, and |
|
|
absolute vector must not lie outside of image dimensions. Constraint one is dealt with by CHECK_MV16_DIRECT |
|
|
and for constraint two we need distance to boundary. This is done by get_range very large fcode (hack!) */ |
|
|
|
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, i, j, 16, iWidth, iHeight, 19); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
TRB,TRD, |
|
|
0,0, |
|
|
d_sad16, |
|
|
&mb->deltamv, |
|
|
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
edged_width, 2, frame->quant, 0); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
TRB,TRD, |
|
|
mb->deltamv.x, mb->deltamv.y, |
|
|
d_sad16, |
|
|
&mb->deltamv, |
|
|
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
edged_width, 1, frame->quant, 0); // equiv to halfpel refine |
|
|
|
|
|
|
|
|
// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ |
|
|
// f_sad16 = 65535; |
|
|
// b_sad16 = 65535; |
|
|
// d_sad16 = 65535; |
|
|
|
|
|
if (f_sad16 < b_sad16) { |
|
|
best_sad = f_sad16; |
|
|
mb->mode = MODE_FORWARD; |
|
|
} else { |
|
|
best_sad = b_sad16; |
|
|
mb->mode = MODE_BACKWARD; |
|
|
} |
|
|
|
|
|
if (i_sad16 < best_sad) { |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
|
} |
|
|
|
|
|
if (d_sad16 < best_sad) { |
|
|
|
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
{ |
|
|
|
|
|
/* how to calc vectors is defined in standard. mvs[] and b_mvs[] are only for motion compensation */ |
|
|
/* for the bitstream, the value mb->deltamv is read directly */ |
|
|
|
|
|
for (k = 0; k < 4; k++) { |
|
|
|
|
|
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
|
|
: mb->mvs[k].x - mb->directmv[k].x); |
|
|
|
|
|
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
|
: mb->mvs[k].y - mb->directmv[k].y); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
|
|
|
|
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
|
|
|
|
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
|
|
|
|
|
mb->b_mvs[0].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
|
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
|
|
|
|
mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; |
|
|
mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; |
|
|
} |
|
|
|
|
|
best_sad = d_sad16; |
|
|
mb->mode = MODE_DIRECT; |
|
|
} |
|
|
|
|
|
switch (mb->mode) |
|
|
{ |
|
|
case MODE_FORWARD: |
|
|
f_count++; |
|
|
f_predMV = mb->mvs[0]; |
|
|
break; |
|
|
case MODE_BACKWARD: |
|
|
b_count++; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
|
|
|
break; |
|
|
case MODE_INTERPOLATE: |
|
|
i_count++; |
|
|
mb->mvs[0] = f_interpolMV; |
|
|
mb->b_mvs[0] = b_interpolMV; |
|
|
f_predMV = mb->mvs[0]; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
break; |
|
|
case MODE_DIRECT: |
|
|
d_count++; |
|
|
break; |
|
|
default: |
|
|
break; |
|
|
} |
|
|
|
|
|
} |
|
|
} |
|
|
|
|
|
#ifdef _DEBUG_BFRAME_STAT |
|
|
fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", |
|
|
f_count,b_count,i_count,d_count); |
|
|
#endif |
|
|
|
|
|
} |
|