1 |
/************************************************************************** |
/************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* motion estimation |
5 |
* |
* |
6 |
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
* This program is an implementation of a part of one or more MPEG-4 |
7 |
* EPZS and EPZS^2 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
8 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* to use this software module in hardware or software products are |
9 |
* PMVFast_Refine to support multiple searches with different start points |
* advised that its use may infringe existing patents or copyrights, and |
10 |
* 07.01.2002 uv-block-based interpolation |
* any such use would be at such party's own risk. The original |
11 |
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
* developer of this software module and his/her company, and subsequent |
12 |
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
* editors and their companies, will have no liability for use of this |
13 |
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
* software or modifications or derivatives thereof. |
|
* 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. |
|
14 |
* |
* |
15 |
* Michael Militzer <isibaar@videocoding.de> |
* This program is free software; you can redistribute it and/or modify |
16 |
|
* it under the terms of the GNU General Public License as published by |
17 |
|
* the Free Software Foundation; either version 2 of the License, or |
18 |
|
* (at your option) any later version. |
19 |
* |
* |
20 |
**************************************************************************/ |
* This program is distributed in the hope that it will be useful, |
21 |
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
22 |
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
23 |
|
* GNU General Public License for more details. |
24 |
|
* |
25 |
|
* You should have received a copy of the GNU General Public License |
26 |
|
* along with this program; if not, write to the Free Software |
27 |
|
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
28 |
|
* |
29 |
|
*************************************************************************/ |
30 |
|
|
31 |
#include <assert.h> |
#include <assert.h> |
32 |
#include <stdio.h> |
#include <stdio.h> |
37 |
#include "../prediction/mbprediction.h" |
#include "../prediction/mbprediction.h" |
38 |
#include "../global.h" |
#include "../global.h" |
39 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
40 |
|
#include "motion_est.h" |
41 |
|
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
|
44 |
// very large value |
#define INITIAL_SKIP_THRESH (10) |
45 |
#define MV_MAX_ERROR (4096 * 256) |
#define FINAL_SKIP_THRESH (50) |
46 |
|
#define MAX_SAD00_FOR_SKIP (20) |
47 |
// stop search if sdelta < THRESHOLD |
#define MAX_CHROMA_SAD_FOR_SKIP (18) |
48 |
#define MV16_THRESHOLD 192 |
#define SKIP_THRESH_B (10) |
|
#define MV8_THRESHOLD 56 |
|
|
|
|
|
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
|
|
/* nb = vop pixels * 2^(bpp-8) */ |
|
|
#define MV16_00_BIAS (128+1) |
|
49 |
|
|
50 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
51 |
#define INTER_BIAS 512 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
52 |
|
|
53 |
/* Parameters which control inter/inter4v decision */ |
#define iDiamondSize 2 |
|
#define IMV16X16 5 |
|
54 |
|
|
55 |
/* vector map (vlc delta size) smoother parameters */ |
//FILE * debug; |
|
#define NEIGH_TEND_16X16 2 |
|
|
#define NEIGH_TEND_8X8 2 |
|
56 |
|
|
57 |
|
static __inline int |
58 |
|
d_mv_bits(int x, int y, const uint32_t iFcode) |
59 |
|
{ |
60 |
|
int xb, yb; |
61 |
|
|
62 |
// fast ((A)/2)*2 |
if (x == 0) xb = 1; |
63 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
else { |
64 |
|
if (x < 0) x = -x; |
65 |
|
x += (1 << (iFcode - 1)) - 1; |
66 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
x >>= (iFcode - 1); |
67 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
if (x > 32) x = 32; |
68 |
#define ABS(X) (((X)>0)?(X):-(X)) |
xb = mvtab[x] + iFcode; |
69 |
#define SIGN(X) (((X)>0)?1:-1) |
} |
|
|
|
|
int32_t PMVfastSearch16( |
|
|
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 uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
70 |
|
|
71 |
int32_t EPZSSearch16( |
if (y == 0) yb = 1; |
72 |
const uint8_t * const pRef, |
else { |
73 |
const uint8_t * const pRefH, |
if (y < 0) y = -y; |
74 |
const uint8_t * const pRefV, |
y += (1 << (iFcode - 1)) - 1; |
75 |
const uint8_t * const pRefHV, |
y >>= (iFcode - 1); |
76 |
const IMAGE * const pCur, |
if (y > 32) y = 32; |
77 |
const int x, const int y, |
yb = mvtab[y] + iFcode; |
78 |
const uint32_t MotionFlags, |
} |
79 |
const MBParam * const pParam, |
return xb + yb; |
80 |
MACROBLOCK * const pMBs, |
} |
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
81 |
|
|
82 |
|
/* CHACK_CANDIATE FUNCTIONS START */ |
83 |
|
|
84 |
int32_t PMVfastSearch8( |
static void |
85 |
const uint8_t * const pRef, |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
86 |
const uint8_t * const pRefH, |
{ |
87 |
const uint8_t * const pRefV, |
int32_t * const sad = data->temp; |
88 |
const uint8_t * const pRefHV, |
// static int32_t sad[5]; |
89 |
const IMAGE * const pCur, |
int t; |
90 |
const int x, const int y, |
const uint8_t * Reference; |
|
const int start_x, int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
91 |
|
|
92 |
int32_t EPZSSearch8( |
if (( x > data->max_dx) || ( x < data->min_dx) |
93 |
const uint8_t * const pRef, |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
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, int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
94 |
|
|
95 |
|
switch ( ((x&1)<<1) + (y&1) ) { |
96 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
97 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
98 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
99 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
100 |
|
} |
101 |
|
|
102 |
typedef int32_t (MainSearch16Func)( |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound); |
|
103 |
|
|
104 |
typedef MainSearch16Func* MainSearch16FuncPtr; |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
105 |
|
data->temp[0] += lambda_vec16[data->iQuant] * t; |
106 |
|
data->temp[1] += lambda_vec8[data->iQuant] * t; |
107 |
|
|
108 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
109 |
|
data->iMinSAD[0] = data->temp[0]; |
110 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
111 |
|
*dir = Direction; } |
112 |
|
|
113 |
typedef int32_t (MainSearch8Func)( |
if (data->temp[1] < data->iMinSAD[1]) { |
114 |
const uint8_t * const pRef, |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
115 |
const uint8_t * const pRefH, |
if (data->temp[2] < data->iMinSAD[2]) { |
116 |
const uint8_t * const pRefV, |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
117 |
const uint8_t * const pRefHV, |
if (data->temp[3] < data->iMinSAD[3]) { |
118 |
const uint8_t * const cur, |
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
119 |
const int x, const int y, |
if (data->temp[4] < data->iMinSAD[4]) { |
120 |
int32_t startx, int32_t starty, |
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound); |
|
|
|
|
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
|
|
|
|
|
// mv.length table |
|
|
static const uint32_t mvtab[33] = { |
|
|
1, 2, 3, 4, 6, 7, 7, 7, |
|
|
9, 9, 9, 10, 10, 10, 10, 10, |
|
|
10, 10, 10, 10, 10, 10, 10, 10, |
|
|
10, 11, 11, 11, 11, 11, 11, 12, 12 |
|
|
}; |
|
121 |
|
|
122 |
|
} |
123 |
|
|
124 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
static void |
125 |
|
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
126 |
{ |
{ |
127 |
if (component == 0) |
int32_t sad; |
128 |
return 1; |
const uint8_t * Reference; |
129 |
|
|
130 |
if (component < 0) |
if (( x > data->max_dx) || ( x < data->min_dx) |
131 |
component = -component; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
132 |
|
|
133 |
if (iFcode == 1) |
switch ( ((x&1)<<1) + (y&1) ) |
134 |
{ |
{ |
135 |
if (component > 32) |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
136 |
component = 32; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
137 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
138 |
return mvtab[component] + 1; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
139 |
} |
} |
140 |
|
|
141 |
component += (1 << (iFcode - 1)) - 1; |
sad = lambda_vec16[data->iQuant] * |
142 |
component >>= (iFcode - 1); |
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
143 |
|
sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
144 |
|
|
145 |
if (component > 32) |
if (sad < *(data->iMinSAD)) { |
146 |
component = 32; |
*(data->iMinSAD) = sad; |
147 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
148 |
return mvtab[component] + 1 + iFcode - 1; |
*dir = Direction; } |
149 |
} |
} |
150 |
|
|
151 |
|
static void |
152 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
153 |
{ |
{ |
154 |
return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int32_t sad; |
155 |
} |
const int xb = data->currentMV[1].x; |
156 |
|
const int yb = data->currentMV[1].y; |
157 |
|
const uint8_t *ReferenceF, *ReferenceB; |
158 |
|
|
159 |
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
160 |
|
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
161 |
|
|
162 |
{ |
switch ( ((xf&1)<<1) + (yf&1) ) { |
163 |
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
164 |
|
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
165 |
|
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
166 |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
167 |
} |
} |
168 |
|
|
169 |
|
switch ( ((xb&1)<<1) + (yb&1) ) { |
170 |
|
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
171 |
|
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
172 |
/* calculate the min/max range (in halfpixels) |
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
173 |
relative to the _MACROBLOCK_ position |
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
|
*/ |
|
|
|
|
|
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)); |
|
|
|
|
174 |
} |
} |
175 |
|
|
176 |
|
sad = lambda_vec16[data->iQuant] * |
177 |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
178 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
179 |
|
|
180 |
/* |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
* 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; |
|
|
} |
|
181 |
|
|
182 |
|
if (sad < *(data->iMinSAD)) { |
183 |
|
*(data->iMinSAD) = sad; |
184 |
|
data->currentMV->x = xf; data->currentMV->y = yf; |
185 |
|
*dir = Direction; } |
186 |
} |
} |
187 |
|
|
188 |
|
static void |
189 |
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
|
|
|
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) |
|
190 |
{ |
{ |
191 |
|
int32_t sad; |
192 |
|
int k; |
193 |
|
const uint8_t *ReferenceF; |
194 |
|
const uint8_t *ReferenceB; |
195 |
|
VECTOR mvs, b_mvs; |
196 |
|
|
197 |
switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
{ |
|
|
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; |
|
|
} |
|
198 |
|
|
199 |
} |
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
200 |
|
|
201 |
#ifndef SEARCH16 |
for (k = 0; k < 4; k++) { |
202 |
#define SEARCH16 PMVfastSearch16 |
mvs.x = data->directmvF[k].x + x; |
203 |
//#define SEARCH16 FullSearch16 |
b_mvs.x = ((x == 0) ? |
204 |
//#define SEARCH16 EPZSSearch16 |
data->directmvB[k].x |
205 |
#endif |
: mvs.x - data->referencemv[k].x); |
206 |
|
|
207 |
#ifndef SEARCH8 |
mvs.y = data->directmvF[k].y + y; |
208 |
#define SEARCH8 PMVfastSearch8 |
b_mvs.y = ((y == 0) ? |
209 |
//#define SEARCH8 EPZSSearch8 |
data->directmvB[k].y |
210 |
#endif |
: mvs.y - data->referencemv[k].y); |
|
|
|
|
bool MotionEstimation( |
|
|
MACROBLOCK * const pMBs, |
|
|
MBParam * const pParam, |
|
|
const IMAGE * const pRef, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
IMAGE * const pCurrent, |
|
|
const uint32_t iLimit) |
|
211 |
|
|
212 |
{ |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
213 |
const uint32_t iWcount = pParam->mb_width; |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
214 |
const uint32_t iHcount = pParam->mb_height; |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
215 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
216 |
|
|
217 |
uint32_t i, j, iIntra = 0; |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
218 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
219 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
220 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
221 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
222 |
|
} |
223 |
|
|
224 |
VECTOR mv16; |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
225 |
VECTOR pmv16; |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
226 |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
227 |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
228 |
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
229 |
|
} |
230 |
|
|
231 |
int32_t sad8 = 0; |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
232 |
int32_t sad16; |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
233 |
int32_t deviation; |
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
234 |
|
data->iEdgedWidth); |
235 |
|
if (sad > *(data->iMinSAD)) return; |
236 |
|
} |
237 |
|
|
238 |
if (sadInit) |
if (sad < *(data->iMinSAD)) { |
239 |
(*sadInit)(); |
*(data->iMinSAD) = sad; |
240 |
|
data->currentMV->x = x; data->currentMV->y = y; |
241 |
|
*dir = Direction; } |
242 |
|
} |
243 |
|
|
244 |
// note: i==horizontal, j==vertical |
static void |
245 |
for (i = 0; i < iHcount; i++) |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
for (j = 0; j < iWcount; j++) |
|
246 |
{ |
{ |
247 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
int32_t sad; |
248 |
|
const uint8_t *ReferenceF; |
249 |
|
const uint8_t *ReferenceB; |
250 |
|
VECTOR mvs, b_mvs; |
251 |
|
|
252 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; |
|
j, i, pParam->motion_flags, |
|
|
pParam, pMBs, &mv16, &pmv16); |
|
|
pMB->sad16=sad16; |
|
253 |
|
|
254 |
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
255 |
|
|
256 |
/* decide: MODE_INTER or MODE_INTRA |
mvs.x = data->directmvF[0].x + x; |
257 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
b_mvs.x = ((x == 0) ? |
258 |
*/ |
data->directmvB[0].x |
259 |
|
: mvs.x - data->referencemv[0].x); |
260 |
|
|
261 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
mvs.y = data->directmvF[0].y + y; |
262 |
|
b_mvs.y = ((y == 0) ? |
263 |
|
data->directmvB[0].y |
264 |
|
: mvs.y - data->referencemv[0].y); |
265 |
|
|
266 |
if (deviation < (sad16 - INTER_BIAS)) |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
267 |
{ |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
268 |
pMB->mode = MODE_INTRA; |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
269 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
|
270 |
|
|
271 |
iIntra++; |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
272 |
if(iIntra >= iLimit) |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
273 |
return 1; |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
274 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
275 |
continue; |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
276 |
} |
} |
277 |
|
|
278 |
if (pParam->global_flags & XVID_INTER4V) |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
279 |
{ |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
280 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
281 |
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
282 |
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
283 |
|
} |
|
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
|
|
|
|
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
284 |
|
|
285 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
|
286 |
|
|
287 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
if (sad < *(data->iMinSAD)) { |
288 |
|
*(data->iMinSAD) = sad; |
289 |
|
data->currentMV->x = x; data->currentMV->y = y; |
290 |
|
*dir = Direction; } |
291 |
} |
} |
292 |
|
|
293 |
|
static void |
294 |
|
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
295 |
|
{ |
296 |
|
int32_t sad; |
297 |
|
const uint8_t * Reference; |
298 |
|
|
299 |
/* decide: MODE_INTER or MODE_INTER4V |
if (( x > data->max_dx) || ( x < data->min_dx) |
300 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
*/ |
|
|
|
|
|
if (pMB->dquant == NO_CHANGE) { |
|
|
if (((pParam->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
|
301 |
|
|
302 |
sad8 = sad16; |
switch ( ((x&1)<<1) + (y&1) ) |
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
else |
|
|
pMB->mode = MODE_INTER4V; |
|
|
} |
|
|
else |
|
303 |
{ |
{ |
304 |
sad8 = sad16; |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
305 |
pMB->mode = MODE_INTER; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
306 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
307 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
} |
|
|
|
|
|
return 0; |
|
308 |
} |
} |
309 |
|
|
310 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
311 |
|
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
|
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
|
|
|
|
312 |
|
|
313 |
#define CHECK_MV16_ZERO {\ |
if (sad < *(data->iMinSAD)) { |
314 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
*(data->iMinSAD) = sad; |
315 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
data->currentMV->x = x; data->currentMV->y = y; |
316 |
{ \ |
*dir = Direction; } |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
|
|
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD <= iQuant * 96) \ |
|
|
iSAD -= MV16_00_BIAS; \ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
|
|
} |
|
|
|
|
|
#define NOCHECK_MV16_CANDIDATE(X,Y) { \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
|
} |
|
|
|
|
|
|
|
|
#define CHECK_MV8_ZERO {\ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
|
|
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
|
|
} |
|
|
|
|
|
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
|
|
|
|
#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
317 |
} |
} |
318 |
|
|
319 |
/* too slow and not fully functional at the moment */ |
/* CHACK_CANDIATE FUNCTIONS END */ |
|
/* |
|
|
int32_t ZeroSearch16( |
|
|
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 uint32_t MotionFlags, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
|
|
|
|
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
|
|
|
|
|
iSAD = sad16( cur, |
|
|
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
if (iSAD <= iQuant * 96) |
|
|
iSAD -= MV16_00_BIAS; |
|
|
|
|
|
currMV->x = 0; |
|
|
currMV->y = 0; |
|
|
currPMV->x = -pred_x; |
|
|
currPMV->y = -pred_y; |
|
|
|
|
|
return iSAD; |
|
320 |
|
|
321 |
} |
/* MAINSEARCH FUNCTIONS START */ |
|
*/ |
|
322 |
|
|
323 |
int32_t Diamond16_MainSearch( |
static void |
324 |
const uint8_t * const pRef, |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
325 |
{ |
{ |
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
326 |
|
|
327 |
int32_t Square16_MainSearch( |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a square search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full square pattern, and new parts for all following diamonds */ |
|
|
|
|
|
/* new direction are extra, so 1-4 is normal diamond |
|
|
537 |
|
|
1*2 |
|
|
648 |
|
|
*/ |
|
328 |
|
|
329 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
int iDirection; |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
330 |
|
|
331 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
do { |
332 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
iDirection = 0; |
333 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
334 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
335 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
336 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
337 |
|
|
338 |
|
/* now we're doing diagonal checks near our candidate */ |
339 |
|
|
340 |
if (iDirection) |
if (iDirection) { //checking if anything found |
341 |
while (!iFound) |
bDirection = iDirection; |
342 |
{ |
iDirection = 0; |
343 |
iFound = 1; |
x = data->currentMV->x; y = data->currentMV->y; |
344 |
backupMV=*currMV; |
if (bDirection & 3) { //our candidate is left or right |
345 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
346 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
347 |
|
} else { // what remains here is up or down |
348 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
349 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
350 |
|
|
351 |
switch (iDirection) |
if (iDirection) { |
352 |
{ |
bDirection += iDirection; |
353 |
case 1: |
x = data->currentMV->x; y = data->currentMV->y; } |
354 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
} else { //about to quit, eh? not so fast.... |
355 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
switch (bDirection) { |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
break; |
|
356 |
case 2: |
case 2: |
357 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
358 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
359 |
break; |
break; |
360 |
|
case 1: |
361 |
case 3: |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
362 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
363 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
break; |
364 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
case 2 + 4: |
365 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
366 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
367 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
368 |
break; |
break; |
|
|
|
369 |
case 4: |
case 4: |
370 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
371 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
372 |
break; |
break; |
373 |
|
case 8: |
374 |
case 5: |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
375 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
376 |
break; |
break; |
377 |
|
case 1 + 4: |
378 |
case 6: |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
379 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
380 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
|
|
381 |
break; |
break; |
382 |
|
case 2 + 8: |
383 |
case 7: |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
384 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
385 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
386 |
break; |
break; |
387 |
|
case 1 + 8: |
388 |
case 8: |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
389 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
390 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
391 |
break; |
break; |
392 |
default: |
default: //1+2+4+8 == we didn't find anything at all |
393 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
394 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
395 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
396 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
|
|
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
397 |
break; |
break; |
398 |
} |
} |
399 |
|
if (!iDirection) break; //ok, the end. really |
400 |
|
bDirection = iDirection; |
401 |
|
x = data->currentMV->x; y = data->currentMV->y; |
402 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
403 |
} |
} |
404 |
return iMinSAD; |
while (1); //forever |
405 |
} |
} |
406 |
|
|
407 |
|
static void |
408 |
|
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
409 |
|
{ |
410 |
|
int iDirection; |
411 |
|
|
412 |
int32_t Full16_MainSearch( |
do { |
413 |
const uint8_t * const pRef, |
iDirection = 0; |
414 |
const uint8_t * const pRefH, |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
415 |
const uint8_t * const pRefV, |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
416 |
const uint8_t * const pRefHV, |
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
417 |
const uint8_t * const cur, |
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
418 |
const int x, const int y, |
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
419 |
int32_t startx, int32_t starty, |
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
420 |
int32_t iMinSAD, |
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
421 |
VECTOR * const currMV, |
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
int32_t iSAD; |
|
|
int32_t dx,dy; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
|
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
|
|
NOCHECK_MV16_CANDIDATE(dx,dy); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
int32_t Full8_MainSearch( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
int32_t iSAD; |
|
|
int32_t dx,dy; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) |
|
|
for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) |
|
|
NOCHECK_MV8_CANDIDATE(dx,dy); |
|
422 |
|
|
423 |
return iMinSAD; |
bDirection = iDirection; |
424 |
|
x = data->currentMV->x; y = data->currentMV->y; |
425 |
|
} while (iDirection); |
426 |
} |
} |
427 |
|
|
428 |
|
static void |
429 |
|
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
|
int32_t Halfpel16_Refine( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
430 |
{ |
{ |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
431 |
|
|
432 |
int32_t iSAD; |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
VECTOR backupMV = *currMV; |
|
433 |
|
|
434 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
int iDirection; |
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
435 |
|
|
436 |
return iMinSAD; |
do { |
437 |
} |
iDirection = 0; |
438 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
439 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
440 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
441 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
442 |
|
|
443 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
/* now we're doing diagonal checks near our candidate */ |
444 |
|
|
445 |
|
if (iDirection) { //checking if anything found |
446 |
|
bDirection = iDirection; |
447 |
|
iDirection = 0; |
448 |
|
x = data->currentMV->x; y = data->currentMV->y; |
449 |
|
if (bDirection & 3) { //our candidate is left or right |
450 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
451 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
452 |
|
} else { // what remains here is up or down |
453 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
454 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
455 |
|
|
456 |
int32_t PMVfastSearch16( |
bDirection += iDirection; |
457 |
const uint8_t * const pRef, |
x = data->currentMV->x; y = data->currentMV->y; |
458 |
const uint8_t * const pRefH, |
} |
459 |
const uint8_t * const pRefV, |
} |
460 |
const uint8_t * const pRefHV, |
while (iDirection); |
461 |
const IMAGE * const pCur, |
} |
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
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; |
|
|
|
|
|
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; /* just for PMVFAST */ |
|
462 |
|
|
463 |
VECTOR pmv[4]; |
/* MAINSEARCH FUNCTIONS END */ |
|
int32_t psad[4]; |
|
464 |
|
|
465 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
466 |
|
|
467 |
static int32_t threshA,threshB; |
static void |
468 |
int32_t bPredEq; |
HalfpelRefine(const SearchData * const data) |
469 |
int32_t iMinSAD,iSAD; |
{ |
470 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
471 |
|
|
472 |
/* Get maximum range */ |
VECTOR backupMV = *(data->currentMV); |
473 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
int iDirection; //not needed |
|
x, y, 16, iWidth, iHeight, iFcode); |
|
474 |
|
|
475 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
476 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
477 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
478 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
479 |
|
|
480 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
481 |
{ min_dx = EVEN(min_dx); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
482 |
|
|
483 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
484 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
485 |
|
} |
486 |
|
|
487 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
static __inline int |
488 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
489 |
|
const int x, const int y, |
490 |
|
const uint32_t iEdgedWidth, const uint32_t iQuant) |
491 |
|
|
|
if ((x==0) && (y==0) ) |
|
492 |
{ |
{ |
493 |
threshA = 512; |
/* keep repeating checks for all b-frames before this P frame, |
494 |
threshB = 1024; |
to make sure that SKIP is possible (todo) |
495 |
|
how: if skip is not possible set sad00 to a very high value */ |
496 |
|
|
497 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
498 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
499 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
500 |
|
sadC += sad8(current->v + x*8 + y*(iEdgedWidth/2)*8, |
501 |
|
reference->v + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
502 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
503 |
|
|
504 |
|
return 1; |
505 |
} |
} |
506 |
else |
|
507 |
|
static __inline void |
508 |
|
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
509 |
{ |
{ |
510 |
threshA = psad[0]; |
pMB->mode = MODE_NOT_CODED; |
511 |
threshB = threshA+256; |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->mv16.x = 0; |
512 |
if (threshA< 512) threshA = 512; |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->mv16.y = 0; |
513 |
if (threshA>1024) threshA = 1024; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
|
if (threshB>1792) threshB = 1792; |
|
514 |
} |
} |
515 |
|
|
516 |
iFound=0; |
bool |
517 |
|
MotionEstimation(MBParam * const pParam, |
518 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
FRAMEINFO * const current, |
519 |
vector of the median. |
FRAMEINFO * const reference, |
520 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
const IMAGE * const pRefH, |
521 |
*/ |
const IMAGE * const pRefV, |
522 |
|
const IMAGE * const pRefHV, |
523 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) |
const uint32_t iLimit) |
524 |
iFound=2; |
{ |
525 |
|
MACROBLOCK *const pMBs = current->mbs; |
526 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
const IMAGE *const pCurrent = ¤t->image; |
527 |
Otherwise select large Diamond Search. |
const IMAGE *const pRef = &reference->image; |
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
528 |
|
|
529 |
/* Step 4: Calculate SAD around the Median prediction. |
const VECTOR zeroMV = { 0, 0 }; |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
530 |
|
|
531 |
|
uint32_t x, y; |
532 |
|
uint32_t iIntra = 0; |
533 |
|
int32_t InterBias; |
534 |
|
|
535 |
// Prepare for main loop |
if (sadInit) (*sadInit) (); |
536 |
|
|
537 |
*currMV=pmv[0]; /* current best := prediction */ |
for (y = 0; y < pParam->mb_height; y++) { |
538 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
for (x = 0; x < pParam->mb_width; x++) { |
|
{ /* This should NOT be necessary! */ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
539 |
|
|
540 |
if (currMV->x > max_dx) |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
541 |
{ |
int32_t sad00 = pMB->sad16 |
542 |
currMV->x=max_dx; |
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
543 |
} |
pRef->y + (x + y * pParam->edged_width) * 16, |
544 |
if (currMV->x < min_dx) |
pParam->edged_width, pMB->sad8 ); |
|
{ |
|
|
currMV->x=min_dx; |
|
|
} |
|
|
if (currMV->y > max_dy) |
|
|
{ |
|
|
currMV->y=max_dy; |
|
|
} |
|
|
if (currMV->y < min_dy) |
|
|
{ |
|
|
currMV->y=min_dy; |
|
|
} |
|
545 |
|
|
546 |
iMinSAD = sad16( cur, |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
547 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
pMB->dquant = NO_CHANGE; |
548 |
iEdgedWidth, MV_MAX_ERROR); |
pMB->quant = current->quant; } |
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
549 |
|
|
550 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
//initial skip decision |
|
{ |
|
551 |
|
|
552 |
if (MotionFlags & PMV_QUICKSTOP16) |
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
553 |
goto PMVfast16_Terminate_without_Refine; |
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
554 |
if (MotionFlags & PMV_EARLYSTOP16) |
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
555 |
goto PMVfast16_Terminate_with_Refine; |
SkipMacroblockP(pMB, sad00); |
556 |
|
continue; |
557 |
} |
} |
558 |
|
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
559 |
|
|
560 |
/* |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
561 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
y, current->motion_flags, pMB->quant, |
562 |
Also calculate (0,0) but do not subtract offset. |
current->fcode, pParam, pMBs, reference->mbs, |
563 |
Let MinSAD be the smallest SAD up to this point. |
current->global_flags & XVID_INTER4V, pMB); |
564 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
565 |
*/ |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
566 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
567 |
// (0,0) is always possible |
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
568 |
|
{ SkipMacroblockP(pMB, sad00); continue; } |
569 |
|
|
570 |
CHECK_MV16_ZERO; |
/* finally, intra decision */ |
571 |
|
|
572 |
// previous frame MV is always possible |
InterBias = MV16_INTER_BIAS; |
573 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
574 |
|
if (y != 0) |
575 |
// left neighbour, if allowed |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
576 |
if (x != 0) |
if (x != 0) |
577 |
{ |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
|
} |
|
578 |
|
|
579 |
// top neighbour, if allowed |
if (InterBias < pMB->sad16) { |
580 |
if (y != 0) |
const int32_t deviation = |
581 |
{ |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
582 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
pParam->edged_width); |
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
583 |
|
|
584 |
// top right neighbour, if allowed |
if (deviation < (pMB->sad16 - InterBias)) { |
585 |
if (x != (iWcount-1)) |
if (++iIntra >= iLimit) return 1; |
586 |
{ |
pMB->mode = MODE_INTRA; |
587 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
588 |
{ pmv[3].x = EVEN(pmv[3].x); |
pMB->mvs[3] = zeroMV; |
589 |
pmv[3].y = EVEN(pmv[3].y); |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
590 |
|
pMB->sad8[3] = 0; |
591 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
592 |
} |
} |
593 |
} |
} |
|
|
|
|
/* 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,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast16_Terminate_with_Refine; |
|
594 |
} |
} |
595 |
|
return 0; |
|
|
|
|
/************ (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. |
|
|
*/ |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
596 |
} |
} |
597 |
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) |
|
|
{ |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0],backupMV)) ) |
|
|
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
598 |
|
|
599 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
600 |
|
|
601 |
if (iSAD < iMinSAD) |
static __inline int |
602 |
|
make_mask(const VECTOR * const pmv, const int i) |
603 |
{ |
{ |
604 |
*currMV = newMV; |
int mask = 0xFF, j; |
605 |
iMinSAD = iSAD; |
for (j = 0; j < i; j++) { |
606 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
607 |
|
if (pmv[i].x == pmv[j].x) { |
608 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
609 |
|
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
610 |
|
} else |
611 |
|
if (pmv[i].y == pmv[j].y) { |
612 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
613 |
|
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
614 |
} |
} |
615 |
} |
} |
616 |
|
return mask; |
617 |
} |
} |
618 |
|
|
619 |
/* |
static __inline void |
620 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
621 |
*/ |
const int iHcount, const MACROBLOCK * const prevMB) |
622 |
|
{ |
623 |
|
|
624 |
PMVfast16_Terminate_with_Refine: |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
625 |
|
|
626 |
PMVfast16_Terminate_without_Refine: |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
627 |
currPMV->x = currMV->x - pmv[0].x; |
pmv[5].x = EVEN(pmv[3].x); |
628 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[5].y = EVEN(pmv[3].y); } |
629 |
return iMinSAD; |
else pmv[5].x = pmv[5].y = 0; |
|
} |
|
630 |
|
|
631 |
|
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
632 |
|
else pmv[3].x = pmv[3].y = 0; |
633 |
|
|
634 |
|
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
635 |
|
else pmv[4].x = pmv[4].y = 0; |
636 |
|
|
637 |
|
// [1] median prediction |
638 |
|
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
639 |
|
|
640 |
|
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
641 |
|
|
642 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
643 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
644 |
|
|
645 |
int32_t Diamond8_MainSearch( |
if ((x != iWcount-1) && (y != iHcount-1)) { |
646 |
const uint8_t * const pRef, |
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
647 |
const uint8_t * const pRefH, |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } |
648 |
const uint8_t * const pRefV, |
else pmv[6].x = pmv[6].y = 0; |
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
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 iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* Do a diamond search around given starting point, return SAD of best */ |
|
|
|
|
|
int32_t iDirection=0; |
|
|
int32_t iSAD; |
|
|
VECTOR backupMV; |
|
|
backupMV.x = startx; |
|
|
backupMV.y = starty; |
|
|
|
|
|
/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ |
|
|
|
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
|
|
|
if (iDirection) |
|
|
while (!iFound) |
|
|
{ |
|
|
iFound = 1; |
|
|
backupMV=*currMV; // since iDirection!=0, this is well defined! |
|
|
|
|
|
if ( iDirection != 2) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
|
|
if ( iDirection != 1) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
if ( iDirection != 4) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); |
|
|
if ( iDirection != 3) |
|
|
CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
649 |
} |
} |
650 |
|
|
651 |
int32_t Halfpel8_Refine( |
static void |
652 |
const uint8_t * const pRef, |
SearchP(const uint8_t * const pRef, |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
VECTOR * const currMV, |
|
|
int32_t iMinSAD, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
const int32_t iEdgedWidth) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
|
|
|
|
int32_t iSAD; |
|
|
VECTOR backupMV = *currMV; |
|
|
|
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); |
|
|
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
|
|
|
|
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
|
|
|
|
|
int32_t PMVfastSearch8( |
|
|
const uint8_t * const pRef, |
|
653 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
654 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
655 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
656 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
657 |
const int x, const int y, |
const int x, |
658 |
const int start_x, int start_y, |
const int y, |
659 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
660 |
|
const uint32_t iQuant, |
661 |
|
const uint32_t iFcode, |
662 |
const MBParam * const pParam, |
const MBParam * const pParam, |
663 |
MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
664 |
VECTOR * const currMV, |
const MACROBLOCK * const prevMBs, |
665 |
VECTOR * const currPMV) |
int inter4v, |
666 |
{ |
MACROBLOCK * const pMB) |
667 |
const uint32_t iWcount = pParam->mb_width; |
{ |
|
|
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
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*8 + y*8*iEdgedWidth; |
|
|
|
|
|
int32_t iDiamondSize; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
|
|
|
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
|
|
|
|
|
static int32_t threshA,threshB; |
|
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
668 |
|
|
669 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
const int32_t iEdgedWidth = pParam->edged_width; |
670 |
|
|
671 |
/* Get maximum range */ |
int i, iDirection = 255, mask, threshA; |
672 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
int32_t temp[5]; |
673 |
x, y, 8, iWidth, iHeight, iFcode); |
VECTOR currentMV[5], pmv[7]; |
674 |
|
int32_t psad[4], iMinSAD[5]; |
675 |
|
MainSearchFunc * MainSearchPtr; |
676 |
|
SearchData Data; |
677 |
|
|
678 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, psad); //has to be changed to get_pmv(2)() |
679 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
680 |
|
pParam->width, pParam->height, iFcode); |
681 |
|
|
682 |
|
Data.predMV = pmv[0]; |
683 |
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
684 |
|
Data.iEdgedWidth = iEdgedWidth; |
685 |
|
Data.currentMV = currentMV; |
686 |
|
Data.iMinSAD = iMinSAD; |
687 |
|
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
688 |
|
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
689 |
|
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
690 |
|
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
691 |
|
Data.temp = temp; |
692 |
|
|
693 |
|
Data.iQuant = iQuant; |
694 |
|
Data.iFcode = iFcode; |
695 |
|
|
696 |
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
697 |
|
Data.min_dx = EVEN(Data.min_dx); |
698 |
|
Data.max_dx = EVEN(Data.max_dx); |
699 |
|
Data.min_dy = EVEN(Data.min_dy); |
700 |
|
Data.max_dy = EVEN(Data.max_dy); } |
701 |
|
|
702 |
|
for(i = 0; i < 5; i++) currentMV[i].x = currentMV[i].y = 0; |
703 |
|
|
704 |
|
i = d_mv_bits(pmv[0].x, pmv[0].y, iFcode); |
705 |
|
|
706 |
|
iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
707 |
|
iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
708 |
|
iMinSAD[2] = pMB->sad8[1]; |
709 |
|
iMinSAD[3] = pMB->sad8[2]; |
710 |
|
iMinSAD[4] = pMB->sad8[3]; |
711 |
|
|
712 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
713 |
|
|
714 |
|
if ((x == 0) && (y == 0)) threshA = 512; |
715 |
|
else { |
716 |
|
threshA = psad[0] + 20; |
717 |
|
if (threshA < 512) threshA = 512; |
718 |
|
if (threshA > 1024) threshA = 1024; } |
719 |
|
|
720 |
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
721 |
|
prevMBs + x + y * pParam->mb_width); |
722 |
|
|
723 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
if (inter4v) CheckCandidate = CheckCandidate16; |
724 |
{ min_dx = EVEN(min_dx); |
else CheckCandidate = CheckCandidate16no4v; |
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
725 |
|
|
726 |
|
/* main loop. checking all predictions */ |
727 |
|
|
728 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
for (i = 1; i < 7; i++) { |
729 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
730 |
|
CheckCandidate16(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
731 |
|
if (iMinSAD[0] < threshA) break; |
732 |
|
} |
733 |
|
|
734 |
if ((x==0) && (y==0) ) |
if ((iMinSAD[0] <= threshA) || |
735 |
{ |
(MVequal(currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
736 |
threshA = 512/4; |
(iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
737 |
threshB = 1024/4; |
inter4v = 0; |
738 |
|
if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; |
739 |
|
if (MotionFlags & PMV_EARLYSTOP16) { |
740 |
|
CheckCandidate = CheckCandidate16no4v; // I sure hope it's faster |
741 |
|
goto PMVfast16_Terminate_with_Refine; |
742 |
|
} |
743 |
|
} |
744 |
|
|
745 |
|
if (MotionFlags & PMV_USESQUARES16) |
746 |
|
MainSearchPtr = SquareSearch; |
747 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
748 |
|
MainSearchPtr = AdvDiamondSearch; |
749 |
|
else MainSearchPtr = DiamondSearch; |
750 |
|
|
751 |
|
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, iDirection); |
752 |
|
|
753 |
|
/* extended search, diamond starting in 0,0 and in prediction. |
754 |
|
note that this search is/might be done in halfpel positions, |
755 |
|
which makes it more different than the diamond above */ |
756 |
|
|
757 |
|
if (MotionFlags & PMV_EXTSEARCH16) { |
758 |
|
int32_t bSAD; |
759 |
|
VECTOR startMV = Data.predMV, backupMV = currentMV[0]; |
760 |
|
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
761 |
|
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
762 |
|
if (!(MVequal(startMV, backupMV))) { |
763 |
|
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
764 |
|
|
765 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
766 |
|
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
767 |
|
if (bSAD < iMinSAD[0]) { |
768 |
|
currentMV[0] = backupMV; |
769 |
|
iMinSAD[0] = bSAD; } |
770 |
|
} |
771 |
|
|
772 |
|
backupMV = currentMV[0]; |
773 |
|
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
774 |
|
else startMV.x = startMV.y = 0; |
775 |
|
if (!(MVequal(startMV, backupMV))) { |
776 |
|
bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; |
777 |
|
|
778 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); |
779 |
|
(*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); |
780 |
|
if (bSAD < iMinSAD[0]) { |
781 |
|
currentMV[0] = backupMV; |
782 |
|
iMinSAD[0] = bSAD; } |
783 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]/4; /* good estimate */ |
|
|
threshB = threshA+256/4; |
|
|
if (threshA< 512/4) threshA = 512/4; |
|
|
if (threshA>1024/4) threshA = 1024/4; |
|
|
if (threshB>1792/4) threshB = 1792/4; |
|
784 |
} |
} |
785 |
|
|
786 |
iFound=0; |
PMVfast16_Terminate_with_Refine: |
|
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
787 |
|
|
788 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
|
iFound=2; |
|
789 |
|
|
790 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
PMVfast16_Terminate_without_Refine: |
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
791 |
|
|
792 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
if (inter4v) |
793 |
iDiamondSize=1; // 1 halfpel! |
for(i = 0; i < 4; i++) |
794 |
else |
Search8(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
795 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
|
796 |
|
if (!(inter4v) || |
797 |
|
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
798 |
|
// INTER MODE |
799 |
|
pMB->mode = MODE_INTER; |
800 |
|
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
801 |
|
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
802 |
|
|
803 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
804 |
iDiamondSize*=2; |
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
805 |
|
|
806 |
/* Step 4: Calculate SAD around the Median prediction. |
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
807 |
MinSAD=SAD |
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
808 |
If Motion Vector equal to Previous frame motion vector |
} else { |
809 |
and MinSAD<PrevFrmSAD goto Step 10. |
// INTER4V MODE; all other things are already set in Search8 |
810 |
If SAD<=256 goto Step 10. |
pMB->mode = MODE_INTER4V; |
811 |
*/ |
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
812 |
|
} |
813 |
|
|
814 |
|
} |
815 |
|
|
816 |
// Prepare for main loop |
static void |
817 |
|
Search8(const SearchData * const OldData, |
818 |
|
const int x, const int y, |
819 |
|
const uint32_t MotionFlags, |
820 |
|
const MBParam * const pParam, |
821 |
|
MACROBLOCK * const pMB, |
822 |
|
const MACROBLOCK * const pMBs, |
823 |
|
const int block) |
824 |
|
{ |
825 |
|
SearchData Data; |
826 |
|
|
827 |
currMV->x=start_x; /* start with mv16 */ |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
828 |
currMV->y=start_y; |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
829 |
|
Data.currentMV = OldData->currentMV+1+block; |
830 |
|
Data.iFcode = OldData->iFcode; |
831 |
|
Data.iQuant = OldData->iQuant; |
832 |
|
|
833 |
iMinSAD = sad8( cur, |
if (block != 0) |
834 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
835 |
iEdgedWidth); |
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
836 |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
Data.currentMV->y - Data.predMV.y, |
837 |
|
Data.iFcode); |
838 |
|
|
|
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
|
} |
|
839 |
|
|
840 |
|
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
841 |
|
|
842 |
/* |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
843 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
844 |
Also calculate (0,0) but do not subtract offset. |
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
845 |
Let MinSAD be the smallest SAD up to this point. |
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
846 |
|
|
847 |
// the prediction might be even better than mv16 |
Data.iEdgedWidth = pParam->edged_width; |
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
848 |
|
|
849 |
// (0,0) is always possible |
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
CHECK_MV8_ZERO; |
|
850 |
|
|
851 |
// previous frame MV is always possible |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
852 |
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
pParam->width, pParam->height, OldData->iFcode); |
853 |
|
|
854 |
// left neighbour, if allowed |
CheckCandidate = CheckCandidate8; |
|
if (psad[1] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
|
} |
|
855 |
|
|
856 |
// top neighbour, if allowed |
if (MotionFlags & PMV_EXTSEARCH8) { |
|
if (psad[2] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
857 |
|
|
858 |
// top right neighbour, if allowed |
MainSearchFunc *MainSearchPtr; |
859 |
if (psad[3] != MV_MAX_ERROR) |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
860 |
{ |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
861 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
else MainSearchPtr = DiamondSearch; |
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
|
} |
|
|
} |
|
862 |
|
|
863 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); } |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
864 |
|
|
865 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
866 |
} |
} |
867 |
|
|
868 |
/************ (Diamond Search) **************/ |
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
869 |
/* |
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
870 |
Step 7: Perform Diamond search, with either the small or large diamond. |
pMB->mvs[block] = *(Data.currentMV); |
871 |
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
872 |
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. |
|
|
*/ |
|
|
|
|
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
873 |
|
|
874 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* B-frames code starts here */ |
|
iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
875 |
|
|
876 |
if (iSAD < iMinSAD) |
static __inline VECTOR |
877 |
|
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
878 |
{ |
{ |
879 |
*currMV = newMV; |
/* the stupidiest function ever */ |
880 |
iMinSAD = iSAD; |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
881 |
|
else return pMB->b_mvs[0]; |
882 |
} |
} |
883 |
|
|
884 |
if (MotionFlags & PMV_EXTSEARCH8) |
static void __inline |
885 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
886 |
|
const uint32_t iWcount, |
887 |
|
const MACROBLOCK * const pMB, |
888 |
|
const uint32_t mode_curr) |
889 |
{ |
{ |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
890 |
|
|
891 |
if (!(MVequal(pmv[0],backupMV)) ) |
// [0] is prediction |
892 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
893 |
|
|
894 |
if (iSAD < iMinSAD) |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
895 |
|
|
896 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
pmv[2] = ChoosePred(pMB, mode_curr); |
897 |
{ iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
898 |
|
|
899 |
if (iSAD < iMinSAD) |
pmv[3].x = pmv[3].y = 0; |
900 |
{ |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
901 |
*currMV = newMV; |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
902 |
iMinSAD = iSAD; |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } |
|
} |
|
|
} |
|
|
} |
|
903 |
|
|
904 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
if (y != 0) { |
905 |
By performing an optional local half-pixel search, we can refine this result even further. |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
906 |
*/ |
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
907 |
|
} else pmv[4].x = pmv[4].y = 0; |
|
PMVfast8_Terminate_with_Refine: |
|
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
|
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
908 |
|
|
909 |
|
if (x != 0) { |
910 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
911 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
912 |
|
} else pmv[5].x = pmv[5].y = 0; |
913 |
|
|
914 |
PMVfast8_Terminate_without_Refine: |
if ((x != 0)&&(y != 0)) { |
915 |
currPMV->x = currMV->x - pmv[0].x; |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
916 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
917 |
|
} else pmv[6].x = pmv[6].y = 0; |
918 |
|
|
919 |
return iMinSAD; |
// more? |
920 |
} |
} |
921 |
|
|
922 |
int32_t EPZSSearch16( |
|
923 |
const uint8_t * const pRef, |
/* search backward or forward, for b-frames */ |
924 |
|
static void |
925 |
|
SearchBF( const uint8_t * const pRef, |
926 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
927 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
928 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
929 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
930 |
const int x, const int y, |
const int x, const int y, |
931 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
932 |
|
const uint32_t iQuant, |
933 |
|
const uint32_t iFcode, |
934 |
const MBParam * const pParam, |
const MBParam * const pParam, |
935 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
936 |
VECTOR * const currMV, |
const VECTOR * const predMV, |
937 |
VECTOR * const currPMV) |
int32_t * const best_sad, |
938 |
{ |
const int32_t mode_current) |
939 |
const uint32_t iWcount = pParam->mb_width; |
{ |
|
const uint32_t iHcount = pParam->mb_height; |
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
940 |
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
941 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
942 |
|
|
943 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
int i, iDirection, mask; |
944 |
|
VECTOR currentMV, pmv[7]; |
945 |
int32_t min_dx; |
MainSearchFunc *MainSearchPtr; |
946 |
int32_t max_dx; |
int32_t iMinSAD = MV_MAX_ERROR; |
947 |
int32_t min_dy; |
SearchData Data; |
948 |
int32_t max_dy; |
|
949 |
|
Data.iMinSAD = &iMinSAD; |
950 |
VECTOR newMV; |
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
951 |
VECTOR backupMV; |
Data.iEdgedWidth = iEdgedWidth; |
952 |
|
Data.currentMV = ¤tMV; |
953 |
VECTOR pmv[4]; |
Data.iMinSAD = &iMinSAD; |
954 |
int32_t psad[8]; |
Data.Ref = pRef + (x + y * iEdgedWidth) * 16; |
955 |
|
Data.RefH = pRefH + (x + y * iEdgedWidth) * 16; |
956 |
static MACROBLOCK * oldMBs = NULL; |
Data.RefV = pRefV + (x + y * iEdgedWidth) * 16; |
957 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
Data.RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
958 |
MACROBLOCK * oldMB = NULL; |
|
959 |
|
Data.iQuant = iQuant; |
960 |
static int32_t thresh2; |
Data.iFcode = iFcode; |
961 |
int32_t bPredEq; |
Data.predMV = *predMV; |
962 |
int32_t iMinSAD,iSAD=9999; |
|
963 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
964 |
MainSearch16FuncPtr EPZSMainSearchPtr; |
pParam->width, pParam->height, iFcode); |
965 |
|
|
966 |
if (oldMBs == NULL) |
if (!(MotionFlags & PMV_HALFPEL16)) { |
967 |
{ oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); |
Data.min_dx = EVEN(Data.min_dx); |
968 |
fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
Data.max_dx = EVEN(Data.max_dx); |
969 |
|
Data.min_dy = EVEN(Data.min_dy); |
970 |
|
Data.max_dy = EVEN(Data.max_dy); } // no-halpel and b-frames. do we need it? |
971 |
|
|
972 |
|
|
973 |
|
pmv[0] = Data.predMV; |
974 |
|
PreparePredictionsBF(pmv, x, y, pParam->mb_width, |
975 |
|
pMB, mode_current); |
976 |
|
|
977 |
|
currentMV.x = currentMV.y = 0; |
978 |
|
|
979 |
|
CheckCandidate = CheckCandidate16no4v; |
980 |
|
|
981 |
|
// main loop. checking all predictions |
982 |
|
for (i = 0; i < 8; i++) { |
983 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
984 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); |
985 |
} |
} |
|
oldMB = oldMBs + x + y * iWcount; |
|
|
|
|
|
/* 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 (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
|
|
|
|
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
|
|
|
|
|
/* 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. |
|
|
*/ |
|
|
|
|
|
// Prepare for main loop |
|
986 |
|
|
987 |
*currMV=pmv[0]; /* current best := median prediction */ |
if (MotionFlags & PMV_USESQUARES16) |
988 |
if (!(MotionFlags & PMV_HALFPEL16)) |
MainSearchPtr = SquareSearch; |
989 |
{ |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
990 |
currMV->x = EVEN(currMV->x); |
MainSearchPtr = AdvDiamondSearch; |
991 |
currMV->y = EVEN(currMV->y); |
else MainSearchPtr = DiamondSearch; |
|
} |
|
|
|
|
|
if (currMV->x > max_dx) |
|
|
currMV->x=max_dx; |
|
|
if (currMV->x < min_dx) |
|
|
currMV->x=min_dx; |
|
|
if (currMV->y > max_dy) |
|
|
currMV->y=max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y=min_dy; |
|
|
|
|
|
/***************** This is predictor SET A: only median prediction ******************/ |
|
|
|
|
|
iMinSAD = sad16( cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
992 |
|
|
993 |
// thresh1 is fixed to 256 |
(*MainSearchPtr)(currentMV.x, currentMV.y, &Data, 255); |
|
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
|
} |
|
994 |
|
|
995 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
996 |
|
|
997 |
// previous frame MV |
// three bits are needed to code backward mode. four for forward |
998 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
// we treat the bits just like they were vector's |
999 |
|
if (mode_current == MODE_FORWARD) iMinSAD += 4 * lambda_vec16[iQuant]; |
1000 |
|
else iMinSAD += 3 * lambda_vec16[iQuant]; |
1001 |
|
|
|
// set threshhold based on Min of Prediction and SAD of collocated block |
|
|
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
|
1002 |
|
|
1003 |
if ((x==0) && (y==0) ) |
if (iMinSAD < *best_sad) { |
1004 |
{ |
*best_sad = iMinSAD; |
1005 |
thresh2 = 512; |
pMB->mode = mode_current; |
1006 |
|
pMB->pmvs[0].x = currentMV.x - predMV->x; |
1007 |
|
pMB->pmvs[0].y = currentMV.y - predMV->y; |
1008 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = currentMV; |
1009 |
|
else pMB->b_mvs[0] = currentMV; |
1010 |
} |
} |
|
else |
|
|
{ |
|
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
|
1011 |
|
|
|
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
|
1012 |
} |
} |
1013 |
|
|
1014 |
// MV=(0,0) is often a good choice |
static int32_t |
1015 |
|
SearchDirect(const uint8_t * const f_Ref, |
1016 |
CHECK_MV16_ZERO; |
const uint8_t * const f_RefH, |
1017 |
|
const uint8_t * const f_RefV, |
1018 |
|
const uint8_t * const f_RefHV, |
1019 |
|
const uint8_t * const b_Ref, |
1020 |
|
const uint8_t * const b_RefH, |
1021 |
|
const uint8_t * const b_RefV, |
1022 |
|
const uint8_t * const b_RefHV, |
1023 |
|
const IMAGE * const pCur, |
1024 |
|
const int x, const int y, |
1025 |
|
const uint32_t MotionFlags, |
1026 |
|
const uint32_t iQuant, |
1027 |
|
const int32_t TRB, const int32_t TRD, |
1028 |
|
const MBParam * const pParam, |
1029 |
|
MACROBLOCK * const pMB, |
1030 |
|
const MACROBLOCK * const b_mb, |
1031 |
|
int32_t * const best_sad) |
1032 |
|
|
1033 |
|
{ |
1034 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
1035 |
|
int32_t iMinSAD = 0, skip_sad; |
1036 |
|
int k; |
1037 |
|
VECTOR currentMV; |
1038 |
|
MainSearchFunc *MainSearchPtr; |
1039 |
|
SearchData Data; |
1040 |
|
|
1041 |
|
Data.iMinSAD = &iMinSAD; |
1042 |
|
Data.Cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; |
1043 |
|
Data.iEdgedWidth = iEdgedWidth; |
1044 |
|
Data.currentMV = ¤tMV; |
1045 |
|
Data.iQuant = iQuant; |
1046 |
|
Data.referencemv = b_mb->mvs; |
1047 |
|
|
1048 |
|
Data.Ref= f_Ref + (x + iEdgedWidth*y) * 16; |
1049 |
|
Data.RefH = f_RefH + (x + iEdgedWidth*y) * 16; |
1050 |
|
Data.RefV = f_RefV + (x + iEdgedWidth*y) * 16; |
1051 |
|
Data.RefHV = f_RefHV + (x + iEdgedWidth*y) * 16; |
1052 |
|
Data.bRef = b_Ref + (x + iEdgedWidth*y) * 16; |
1053 |
|
Data.bRefH = b_RefH + (x + iEdgedWidth*y) * 16; |
1054 |
|
Data.bRefV = b_RefV + (x + iEdgedWidth*y) * 16; |
1055 |
|
Data.bRefHV = b_RefHV + (x + iEdgedWidth*y) * 16; |
1056 |
|
/* |
1057 |
|
//What we do here is a complicated version of CheckCandidateDirect(0,0); |
1058 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, pParam->width, pParam->height, 19); |
1059 |
|
|
1060 |
|
*/ |
1061 |
|
Data.max_dx = 2 * pParam->width - 2 * (x) * 16; |
1062 |
|
Data.max_dy = 2 * pParam->height - 2 * (y) * 16; |
1063 |
|
Data.min_dx = -(2 * 16 + 2 * (x) * 16); |
1064 |
|
Data.min_dy = -(2 * 16 + 2 * (y) * 16); |
1065 |
|
|
1066 |
|
for (k = 0; k < 4; k++) { |
1067 |
|
pMB->mvs[k].x = Data.directmvF[k].x = ((TRB * Data.referencemv[k].x) / TRD); |
1068 |
|
pMB->b_mvs[k].x = Data.directmvB[k].x = ((TRB - TRD) * Data.referencemv[k].x) / TRD; |
1069 |
|
pMB->mvs[k].y = Data.directmvF[k].y = ((TRB * Data.referencemv[k].y) / TRD); |
1070 |
|
pMB->b_mvs[k].y = Data.directmvB[k].y = ((TRB - TRD) * Data.referencemv[k].y) / TRD; |
1071 |
|
|
1072 |
|
if (( pMB->mvs[k].x > Data.max_dx ) || ( pMB->mvs[k].x < Data.min_dx ) |
1073 |
|
|| ( pMB->mvs[k].y > Data.max_dy ) || ( pMB->mvs[k].y < Data.min_dy ) |
1074 |
|
|| ( pMB->b_mvs[k].x > Data.max_dx ) || ( pMB->b_mvs[k].x < Data.min_dx ) |
1075 |
|
|| ( pMB->b_mvs[k].y > Data.max_dy ) || ( pMB->b_mvs[k].y < Data.min_dy )) { |
1076 |
|
/* |
1077 |
|
fprintf(debug, "\nERROR - out of range : vector %d,%d and %d,%d\n", pMB->mvs[k].x, pMB->mvs[k].y,pMB->b_mvs[k].x,pMB->b_mvs[k].y ); |
1078 |
|
fprintf(debug, " range is x: %d..%d y: %d..%d \n", Data.min_dx,Data.max_dx,Data.min_dy,Data.max_dy); |
1079 |
|
fprintf(debug,"macroblock %d, %d \n", x, y); |
1080 |
|
fprintf(debug, "direct MV is %d,%d \n", directmv[k].x, directmv[k].y); |
1081 |
|
*/ |
1082 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1083 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1084 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; /* because backwards and interpol might rely on this */ |
1085 |
|
return 0; } |
1086 |
|
|
1087 |
|
|
1088 |
|
if (b_mb->mode != MODE_INTER4V) { |
1089 |
|
iMinSAD = sad16bi(Data.Cur, |
1090 |
|
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
1091 |
|
x, y, 16, &pMB->mvs[0], iEdgedWidth), |
1092 |
|
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
1093 |
|
x, y, 16, &pMB->b_mvs[0], iEdgedWidth), iEdgedWidth); |
1094 |
|
|
1095 |
// left neighbour, if allowed |
Data.directmvF[1] = Data.directmvF[2] = Data.directmvF[3] = Data.directmvF[0]; |
1096 |
if (x != 0) |
Data.directmvB[1] = Data.directmvB[2] = Data.directmvB[3] = Data.directmvB[0]; |
1097 |
{ |
break; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1098 |
} |
} |
1099 |
|
iMinSAD += sad8bi(Data.Cur + (k&1)*8 + (k>>1)* 8 * iEdgedWidth, |
1100 |
// top neighbour, if allowed |
get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, |
1101 |
if (y != 0) |
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->mvs[k], iEdgedWidth), |
1102 |
{ |
get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, |
1103 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
(2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->b_mvs[k], iEdgedWidth), |
1104 |
{ pmv[2].x = EVEN(pmv[2].x); |
iEdgedWidth); |
|
pmv[2].y = EVEN(pmv[2].y); |
|
1105 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1106 |
|
|
1107 |
// top right neighbour, if allowed |
// skip decision |
1108 |
if (x != (iWcount-1)) |
if (iMinSAD < (int32_t)iQuant * SKIP_THRESH_B) { |
1109 |
{ |
pMB->mode = MODE_DIRECT_NONE_MV; |
1110 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
return iMinSAD; } |
1111 |
{ pmv[3].x = EVEN(pmv[3].x); |
|
1112 |
pmv[3].y = EVEN(pmv[3].y); |
skip_sad = iMinSAD; |
1113 |
|
iMinSAD += 2 * lambda_vec16[iQuant]; // 2 bits needed to code vector 0,0 |
1114 |
|
currentMV.x = currentMV.y = 0; |
1115 |
|
if (b_mb->mode == MODE_INTER4V) |
1116 |
|
CheckCandidate = CheckCandidateDirect; |
1117 |
|
else CheckCandidate = CheckCandidateDirectno4v; |
1118 |
|
|
1119 |
|
// DIRECT MODE DELTA VECTOR SEARCH. |
1120 |
|
// This has to be made more effective, but at the moment I'm happy it's running at all |
1121 |
|
|
1122 |
|
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1123 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1124 |
|
else MainSearchPtr = DiamondSearch; |
1125 |
|
|
1126 |
|
(*MainSearchPtr)(0, 0, &Data, 255); |
1127 |
|
|
1128 |
|
HalfpelRefine(&Data); |
1129 |
|
|
1130 |
|
iMinSAD += 1 * lambda_vec16[iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
1131 |
|
*best_sad = iMinSAD; |
1132 |
|
|
1133 |
|
if (b_mb->mode == MODE_INTER4V) |
1134 |
|
pMB->mode = MODE_DIRECT; |
1135 |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1136 |
|
|
1137 |
|
pMB->pmvs[3] = currentMV; |
1138 |
|
|
1139 |
|
for (k = 0; k < 4; k++) { |
1140 |
|
pMB->mvs[k].x = Data.directmvF[k].x + currentMV.x; |
1141 |
|
pMB->b_mvs[k].x = ((currentMV.x == 0) |
1142 |
|
? Data.directmvB[k].x |
1143 |
|
: pMB->mvs[k].x - Data.referencemv[k].x); |
1144 |
|
pMB->mvs[k].y = (Data.directmvF[k].y + currentMV.y); |
1145 |
|
pMB->b_mvs[k].y = ((currentMV.y == 0) |
1146 |
|
? Data.directmvB[k].y |
1147 |
|
: pMB->mvs[k].y - Data.referencemv[k].y); |
1148 |
|
if (b_mb->mode != MODE_INTER4V) { |
1149 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1150 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1151 |
|
break; |
1152 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1153 |
} |
} |
1154 |
|
return 0;//skip_sad; |
1155 |
} |
} |
1156 |
|
|
1157 |
/* Terminate if MinSAD <= T_2 |
static __inline void |
1158 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
SearchInterpolate(const uint8_t * const f_Ref, |
1159 |
*/ |
const uint8_t * const f_RefH, |
1160 |
|
const uint8_t * const f_RefV, |
1161 |
|
const uint8_t * const f_RefHV, |
1162 |
|
const uint8_t * const b_Ref, |
1163 |
|
const uint8_t * const b_RefH, |
1164 |
|
const uint8_t * const b_RefV, |
1165 |
|
const uint8_t * const b_RefHV, |
1166 |
|
const IMAGE * const pCur, |
1167 |
|
const int x, const int y, |
1168 |
|
const uint32_t fcode, |
1169 |
|
const uint32_t bcode, |
1170 |
|
const uint32_t MotionFlags, |
1171 |
|
const uint32_t iQuant, |
1172 |
|
const MBParam * const pParam, |
1173 |
|
const VECTOR * const f_predMV, |
1174 |
|
const VECTOR * const b_predMV, |
1175 |
|
MACROBLOCK * const pMB, |
1176 |
|
int32_t * const best_sad) |
1177 |
|
|
|
if ( (iMinSAD <= thresh2) |
|
|
|| ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD <= pMB->sad16) ) ) |
|
1178 |
{ |
{ |
1179 |
if (MotionFlags & PMV_QUICKSTOP16) |
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
1180 |
goto EPZS16_Terminate_without_Refine; |
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
1181 |
if (MotionFlags & PMV_EARLYSTOP16) |
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
1182 |
goto EPZS16_Terminate_with_Refine; |
as "optimal". */ |
|
} |
|
1183 |
|
|
1184 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
const int32_t iEdgedWidth = pParam->edged_width; |
1185 |
|
|
1186 |
backupMV = pMB->mvs[0]; // last MV |
int iDirection, i, j; |
1187 |
backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
int32_t iMinSAD = 256*4096; |
1188 |
backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
VECTOR currentMV[3]; |
1189 |
|
SearchData fData, bData; |
1190 |
|
|
1191 |
|
|
1192 |
|
fData.iMinSAD = bData.iMinSAD = &iMinSAD; |
1193 |
|
|
1194 |
|
fData.Cur = bData.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1195 |
|
fData.iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1196 |
|
fData.currentMV = currentMV; bData.currentMV = currentMV + 1; |
1197 |
|
fData.iQuant = bData.iQuant = iQuant; |
1198 |
|
fData.iFcode = bData.bFcode = fcode; fData.bFcode = bData.iFcode = bcode; |
1199 |
|
|
1200 |
|
|
1201 |
|
bData.bRef = fData.Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
1202 |
|
bData.bRefH = fData.RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
1203 |
|
bData.bRefV = fData.RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
1204 |
|
bData.bRefHV = fData.RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
1205 |
|
bData.Ref = fData.bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
1206 |
|
bData.RefH = fData.bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1207 |
|
bData.RefV = fData.bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1208 |
|
bData.RefHV = fData.bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
1209 |
|
|
1210 |
|
bData.bpredMV = fData.predMV = *f_predMV; |
1211 |
|
fData.bpredMV = bData.predMV = *b_predMV; |
1212 |
|
|
1213 |
|
|
1214 |
|
currentMV[0] = pMB->mvs[0]; |
1215 |
|
currentMV[1] = pMB->b_mvs[0]; |
1216 |
|
get_range(&fData.min_dx, &fData.max_dx, &fData.min_dy, &fData.max_dy, x, y, 16, pParam->width, pParam->height, fcode); |
1217 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); |
1218 |
|
|
1219 |
|
CheckCandidateInt(currentMV[0].x, currentMV[0].y, 255, &iDirection, &fData); |
1220 |
|
|
1221 |
|
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1222 |
|
|
1223 |
|
do { |
1224 |
|
iDirection = 255; |
1225 |
|
// forward MV moves |
1226 |
|
i = currentMV[0].x; j = currentMV[0].y; |
1227 |
|
|
1228 |
|
CheckCandidateInt(i + 2, j, 0, &iDirection, &fData); |
1229 |
|
CheckCandidateInt(i, j + 2, 0, &iDirection, &fData); |
1230 |
|
CheckCandidateInt(i - 2, j, 0, &iDirection, &fData); |
1231 |
|
CheckCandidateInt(i, j - 2, 0, &iDirection, &fData); |
1232 |
|
|
1233 |
|
// backward MV moves |
1234 |
|
i = currentMV[1].x; j = currentMV[1].y; |
1235 |
|
currentMV[2] = currentMV[0]; |
1236 |
|
|
1237 |
|
CheckCandidateInt(i + 2, j, 0, &iDirection, &bData); |
1238 |
|
CheckCandidateInt(i, j + 2, 0, &iDirection, &bData); |
1239 |
|
CheckCandidateInt(i - 2, j, 0, &iDirection, &bData); |
1240 |
|
CheckCandidateInt(i, j - 2, 0, &iDirection, &bData); |
1241 |
|
|
1242 |
|
} while (!(iDirection)); |
1243 |
|
|
1244 |
|
/* halfpel refinement. luckly we can use normal halfpel function for it */ |
1245 |
|
|
1246 |
|
if (MotionFlags & PMV_HALFPELREFINE16) { |
1247 |
|
CheckCandidate = CheckCandidateInt; |
1248 |
|
HalfpelRefine(&fData); |
1249 |
|
currentMV[2] = currentMV[0]; |
1250 |
|
HalfpelRefine(&bData); |
1251 |
|
} |
1252 |
|
|
1253 |
|
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
1254 |
|
iMinSAD += 2 * lambda_vec16[iQuant]; |
1255 |
|
if (iMinSAD < *best_sad) { |
1256 |
|
*best_sad = iMinSAD; |
1257 |
|
pMB->mvs[0] = currentMV[0]; |
1258 |
|
pMB->b_mvs[0] = currentMV[1]; |
1259 |
|
pMB->mode = MODE_INTERPOLATE; |
1260 |
|
|
1261 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1262 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1263 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1264 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1265 |
|
} |
1266 |
|
} |
1267 |
|
|
1268 |
|
void |
1269 |
|
MotionEstimationBVOP(MBParam * const pParam, |
1270 |
|
FRAMEINFO * const frame, |
1271 |
|
const int32_t time_bp, |
1272 |
|
const int32_t time_pp, |
1273 |
|
// forward (past) reference |
1274 |
|
const MACROBLOCK * const f_mbs, |
1275 |
|
const IMAGE * const f_ref, |
1276 |
|
const IMAGE * const f_refH, |
1277 |
|
const IMAGE * const f_refV, |
1278 |
|
const IMAGE * const f_refHV, |
1279 |
|
// backward (future) reference |
1280 |
|
const MACROBLOCK * const b_mbs, |
1281 |
|
const IMAGE * const b_ref, |
1282 |
|
const IMAGE * const b_refH, |
1283 |
|
const IMAGE * const b_refV, |
1284 |
|
const IMAGE * const b_refHV) |
1285 |
|
{ |
1286 |
|
uint32_t i, j; |
1287 |
|
int32_t best_sad, skip_sad; |
1288 |
|
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1289 |
|
static const VECTOR zeroMV={0,0}; |
1290 |
|
|
1291 |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1292 |
|
|
1293 |
// left neighbour |
const int32_t TRB = time_pp - time_bp; |
1294 |
if (x != 0) |
const int32_t TRD = time_pp; |
|
CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); |
|
1295 |
|
|
1296 |
// top neighbour |
// note: i==horizontal, j==vertical |
|
if (y != 0) |
|
|
CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); |
|
1297 |
|
|
1298 |
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
for (j = 0; j < pParam->mb_height; j++) { |
1299 |
|
|
1300 |
if (x != iWcount-1) |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); |
|
1301 |
|
|
1302 |
// bottom neighbour, dito |
for (i = 0; i < pParam->mb_width; i++) { |
1303 |
if (y != iHcount-1) |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1304 |
CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1305 |
|
|
1306 |
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
1307 |
if (iMinSAD <= thresh2) |
if (b_mb->mode == MODE_NOT_CODED) { |
1308 |
{ |
pMB->mode = MODE_NOT_CODED; |
1309 |
if (MotionFlags & PMV_QUICKSTOP16) |
continue; |
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
1310 |
} |
} |
1311 |
|
|
1312 |
/************ (if Diamond Search) **************/ |
/* direct search comes first, because it (1) checks for SKIP-mode |
1313 |
|
and (2) sets very good predictions for forward and backward search */ |
|
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
|
1314 |
|
|
1315 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
skip_sad = SearchDirect(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1316 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1317 |
|
&frame->image, |
1318 |
|
i, j, |
1319 |
|
frame->motion_flags, |
1320 |
|
frame->quant, |
1321 |
|
TRB, TRD, |
1322 |
|
pParam, |
1323 |
|
pMB, b_mb, |
1324 |
|
&best_sad); |
1325 |
|
|
1326 |
if (MotionFlags & PMV_USESQUARES16) |
if (!(frame->global_flags & XVID_HALFPEL)) best_sad = skip_sad = 256*4096; |
|
EPZSMainSearchPtr = Square16_MainSearch; |
|
1327 |
else |
else |
1328 |
EPZSMainSearchPtr = Diamond16_MainSearch; |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
|
|
|
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
2, iFcode, iQuant, 0); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
|
|
|
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) |
|
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
1329 |
|
|
1330 |
if (!(MVequal(pmv[0],backupMV)) ) |
// best_sad = 256*4096; //uncomment to disable Directsearch. |
1331 |
{ |
// To disable any other mode, just comment the function call |
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
|
} |
|
1332 |
|
|
1333 |
if (iSAD < iMinSAD) |
// forward search |
1334 |
{ |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1335 |
*currMV = newMV; |
&frame->image, i, j, |
1336 |
iMinSAD = iSAD; |
frame->motion_flags, |
1337 |
} |
frame->quant, frame->fcode, pParam, |
1338 |
|
pMB, &f_predMV, &best_sad, |
1339 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
MODE_FORWARD); |
1340 |
{ |
|
1341 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
// backward search |
1342 |
x, y, |
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1343 |
0, 0, iMinSAD, &newMV, |
&frame->image, i, j, |
1344 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); |
frame->motion_flags, |
1345 |
|
frame->quant, frame->bcode, pParam, |
1346 |
if (iSAD < iMinSAD) |
pMB, &b_predMV, &best_sad, |
1347 |
{ |
MODE_BACKWARD); |
1348 |
*currMV = newMV; |
|
1349 |
iMinSAD = iSAD; |
// interpolate search comes last, because it uses data from forward and backward as prediction |
1350 |
|
|
1351 |
|
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1352 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1353 |
|
&frame->image, |
1354 |
|
i, j, |
1355 |
|
frame->fcode, frame->bcode, |
1356 |
|
frame->motion_flags, |
1357 |
|
frame->quant, pParam, |
1358 |
|
&f_predMV, &b_predMV, |
1359 |
|
pMB, &best_sad); |
1360 |
|
|
1361 |
|
switch (pMB->mode) { |
1362 |
|
case MODE_FORWARD: |
1363 |
|
f_count++; |
1364 |
|
f_predMV = pMB->mvs[0]; |
1365 |
|
break; |
1366 |
|
case MODE_BACKWARD: |
1367 |
|
b_count++; |
1368 |
|
b_predMV = pMB->b_mvs[0]; |
1369 |
|
break; |
1370 |
|
case MODE_INTERPOLATE: |
1371 |
|
i_count++; |
1372 |
|
f_predMV = pMB->mvs[0]; |
1373 |
|
b_predMV = pMB->b_mvs[0]; |
1374 |
|
break; |
1375 |
|
case MODE_DIRECT: |
1376 |
|
case MODE_DIRECT_NO4V: |
1377 |
|
d_count++; |
1378 |
|
break; |
1379 |
|
default: |
1380 |
|
break; |
1381 |
} |
} |
1382 |
} |
} |
1383 |
} |
} |
1384 |
|
|
1385 |
/*************** Choose best MV found **************/ |
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
1386 |
|
// f_count,b_count,i_count,d_count,n_count); |
1387 |
|
|
|
EPZS16_Terminate_with_Refine: |
|
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
EPZS16_Terminate_without_Refine: |
|
|
|
|
|
*oldMB = *pMB; |
|
|
|
|
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
1388 |
} |
} |
1389 |
|
|
1390 |
|
/* Hinted ME starts here */ |
1391 |
|
|
1392 |
int32_t EPZSSearch8( |
static __inline void |
1393 |
const uint8_t * const pRef, |
Search8hinted( const SearchData * const OldData, |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
1394 |
const int x, const int y, |
const int x, const int y, |
|
const int start_x, const int start_y, |
|
1395 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1396 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1397 |
MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
1398 |
VECTOR * const currMV, |
const MACROBLOCK * const pMBs, |
1399 |
VECTOR * const currPMV) |
const int block) |
1400 |
{ |
{ |
1401 |
const uint32_t iWcount = pParam->mb_width; |
SearchData Data; |
1402 |
const int32_t iFcode = pParam->fixed_code; |
MainSearchFunc *MainSearchPtr; |
|
const int32_t iQuant = pParam->quant; |
|
|
|
|
|
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*8 + y*8*iEdgedWidth; |
|
|
|
|
|
int32_t iDiamondSize=1; |
|
|
|
|
|
int32_t min_dx; |
|
|
int32_t max_dx; |
|
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
|
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[8]; |
|
|
|
|
|
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
|
|
|
|
|
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
|
1403 |
|
|
1404 |
int32_t bPredEq; |
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1405 |
int32_t iMinSAD,iSAD=9999; |
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
1406 |
|
Data.currentMV = OldData->currentMV+1+block; |
1407 |
|
Data.iFcode = OldData->iFcode; |
1408 |
|
Data.iQuant = OldData->iQuant; |
1409 |
|
|
1410 |
MainSearch8FuncPtr EPZSMainSearchPtr; |
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1411 |
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1412 |
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1413 |
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1414 |
|
Data.iEdgedWidth = pParam->edged_width; |
1415 |
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1416 |
|
|
1417 |
/* Get maximum range */ |
CheckCandidate = CheckCandidate8; |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 8, iWidth, iHeight, iFcode); |
|
1418 |
|
|
1419 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
if (block != 0) |
1420 |
|
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
1421 |
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
1422 |
|
Data.currentMV->y - Data.predMV.y, |
1423 |
|
Data.iFcode); |
1424 |
|
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ min_dx = EVEN(min_dx); |
|
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
1425 |
|
|
1426 |
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
1427 |
|
pParam->width, pParam->height, OldData->iFcode); |
1428 |
|
|
1429 |
|
if (pMB->mode == MODE_INTER4V) { |
1430 |
|
int dummy; |
1431 |
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
1432 |
|
|
1433 |
/* Step 4: Calculate SAD around the Median prediction. |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1434 |
MinSAD=SAD |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1435 |
If Motion Vector equal to Previous frame motion vector |
else MainSearchPtr = DiamondSearch; |
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1436 |
|
|
1437 |
// Prepare for main loop |
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
1438 |
|
|
1439 |
|
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
1440 |
|
|
1441 |
if (!(MotionFlags & PMV_HALFPEL8)) |
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
1442 |
{ |
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
1443 |
currMV->x = EVEN(currMV->x); |
pMB->mvs[block] = *(Data.currentMV); |
1444 |
currMV->y = EVEN(currMV->y); |
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
1445 |
} |
} |
1446 |
|
|
|
if (currMV->x > max_dx) |
|
|
currMV->x=max_dx; |
|
|
if (currMV->x < min_dx) |
|
|
currMV->x=min_dx; |
|
|
if (currMV->y > max_dy) |
|
|
currMV->y=max_dy; |
|
|
if (currMV->y < min_dy) |
|
|
currMV->y=min_dy; |
|
1447 |
|
|
1448 |
/***************** This is predictor SET A: only median prediction ******************/ |
static void |
1449 |
|
SearchPhinted ( const uint8_t * const pRef, |
1450 |
|
const uint8_t * const pRefH, |
1451 |
iMinSAD = sad8( cur, |
const uint8_t * const pRefV, |
1452 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
const uint8_t * const pRefHV, |
1453 |
iEdgedWidth); |
const IMAGE * const pCur, |
1454 |
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
const int x, |
1455 |
|
const int y, |
1456 |
|
const uint32_t MotionFlags, |
1457 |
|
const uint32_t iQuant, |
1458 |
|
const uint32_t iFcode, |
1459 |
|
const MBParam * const pParam, |
1460 |
|
const MACROBLOCK * const pMBs, |
1461 |
|
int inter4v, |
1462 |
|
MACROBLOCK * const pMB) |
1463 |
|
{ |
1464 |
|
|
1465 |
|
const int32_t iEdgedWidth = pParam->edged_width; |
1466 |
|
|
1467 |
// thresh1 is fixed to 256 |
int i; |
1468 |
if (iMinSAD < 256/4 ) |
VECTOR currentMV[5]; |
1469 |
{ |
int32_t iMinSAD[5]; |
1470 |
if (MotionFlags & PMV_QUICKSTOP8) |
int32_t temp[5]; |
1471 |
goto EPZS8_Terminate_without_Refine; |
MainSearchFunc * MainSearchPtr; |
1472 |
if (MotionFlags & PMV_EARLYSTOP8) |
SearchData Data; |
1473 |
goto EPZS8_Terminate_with_Refine; |
|
1474 |
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1475 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, |
1476 |
|
pParam->width, pParam->height, iFcode); |
1477 |
|
|
1478 |
|
Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1479 |
|
Data.iEdgedWidth = iEdgedWidth; |
1480 |
|
Data.currentMV = currentMV; |
1481 |
|
Data.iMinSAD = iMinSAD; |
1482 |
|
Data.Ref = pRef + (x + iEdgedWidth*y)*16; |
1483 |
|
Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; |
1484 |
|
Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; |
1485 |
|
Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
1486 |
|
Data.temp = temp; |
1487 |
|
Data.iQuant = iQuant; |
1488 |
|
Data.iFcode = iFcode; |
1489 |
|
|
1490 |
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
1491 |
|
Data.min_dx = EVEN(Data.min_dx); |
1492 |
|
Data.max_dx = EVEN(Data.max_dx); |
1493 |
|
Data.min_dy = EVEN(Data.min_dy); |
1494 |
|
Data.max_dy = EVEN(Data.max_dy); |
1495 |
|
} |
1496 |
|
|
1497 |
|
for(i = 0; i < 5; i++) iMinSAD[i] = MV_MAX_ERROR; |
1498 |
|
|
1499 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
1500 |
|
|
1501 |
|
if (inter4v) |
1502 |
|
CheckCandidate = CheckCandidate16; |
1503 |
|
else CheckCandidate = CheckCandidate16no4v; |
1504 |
|
|
1505 |
|
|
1506 |
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1507 |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
1508 |
|
if (pMB->mvs[0].x > Data.max_dx) pMB->mvs[0].x = Data.max_dx; // this is in case iFcode changed |
1509 |
|
if (pMB->mvs[0].x < Data.min_dx) pMB->mvs[0].x = Data.min_dx; |
1510 |
|
if (pMB->mvs[0].y > Data.max_dy) pMB->mvs[0].y = Data.max_dy; |
1511 |
|
if (pMB->mvs[0].y < Data.min_dy) pMB->mvs[0].y = Data.min_dy; |
1512 |
|
|
1513 |
|
CheckCandidate16(pMB->mvs[0].x, pMB->mvs[0].y, 0, &i, &Data); |
1514 |
|
|
1515 |
|
if (pMB->mode == MODE_INTER4V) |
1516 |
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
1517 |
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
1518 |
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
1519 |
|
if (!(make_mask(pMB->mvs, i))) |
1520 |
|
CheckCandidate16(pMB->mvs[i].x, pMB->mvs[i].y, 0, &i, &Data); |
1521 |
} |
} |
1522 |
|
|
1523 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
if (MotionFlags & PMV_USESQUARES16) |
1524 |
|
MainSearchPtr = SquareSearch; |
1525 |
// previous frame MV |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1526 |
CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
MainSearchPtr = AdvDiamondSearch; |
1527 |
|
else MainSearchPtr = DiamondSearch; |
1528 |
|
|
1529 |
|
(*MainSearchPtr)(currentMV->x, currentMV->y, &Data, 255); |
1530 |
|
|
1531 |
|
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); |
1532 |
|
|
1533 |
|
if (inter4v) |
1534 |
|
for(i = 0; i < 4; i++) |
1535 |
|
Search8hinted(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
1536 |
|
|
1537 |
|
if (!(inter4v) || |
1538 |
|
(iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1539 |
|
// INTER MODE |
1540 |
|
|
1541 |
// MV=(0,0) is often a good choice |
pMB->mode = MODE_INTER; |
1542 |
|
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
1543 |
|
= pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; |
1544 |
|
|
1545 |
CHECK_MV8_ZERO; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1546 |
|
pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; |
1547 |
|
|
1548 |
/* Terminate if MinSAD <= T_2 |
pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; |
1549 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; |
1550 |
*/ |
} else { |
1551 |
|
// INTER4V MODE; all other things are already set in Search8hinted |
1552 |
|
pMB->mode = MODE_INTER4V; |
1553 |
|
pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; |
1554 |
|
} |
1555 |
|
|
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
1556 |
} |
} |
1557 |
|
|
1558 |
/************ (if Diamond Search) **************/ |
void |
1559 |
|
MotionEstimationHinted( MBParam * const pParam, |
1560 |
|
FRAMEINFO * const current, |
1561 |
|
FRAMEINFO * const reference, |
1562 |
|
const IMAGE * const pRefH, |
1563 |
|
const IMAGE * const pRefV, |
1564 |
|
const IMAGE * const pRefHV) |
1565 |
|
{ |
1566 |
|
MACROBLOCK *const pMBs = current->mbs; |
1567 |
|
const IMAGE *const pCurrent = ¤t->image; |
1568 |
|
const IMAGE *const pRef = &reference->image; |
1569 |
|
|
1570 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
uint32_t x, y; |
1571 |
|
|
1572 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
if (sadInit) (*sadInit) (); |
|
iDiamondSize *= 2; |
|
1573 |
|
|
1574 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
for (y = 0; y < pParam->mb_height; y++) { |
1575 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1576 |
|
int32_t sad00; |
1577 |
|
|
1578 |
// if (MotionFlags & PMV_USESQUARES8) |
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
|
// EPZSMainSearchPtr = Square8_MainSearch; |
|
|
// else |
|
|
EPZSMainSearchPtr = Diamond8_MainSearch; |
|
1579 |
|
|
1580 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
//intra mode is copied from the first pass. At least for the time being |
1581 |
x, y, |
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, 00); |
|
1582 |
|
|
1583 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1584 |
|
pMB->dquant = NO_CHANGE; |
1585 |
|
pMB->quant = current->quant; } |
1586 |
|
|
1587 |
if (iSAD < iMinSAD) |
if (pMB->dquant == NO_CHANGE) //no skip otherwise, anyway |
1588 |
{ |
sad00 = pMB->sad16 |
1589 |
*currMV = newMV; |
= sad16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
1590 |
iMinSAD = iSAD; |
pRef->y + (x + y * pParam->edged_width) * 16, |
1591 |
} |
pParam->edged_width, 256*4096 ); |
1592 |
|
else sad00 = 256*4096; |
1593 |
|
|
|
if (MotionFlags & PMV_EXTSEARCH8) |
|
|
{ |
|
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
1594 |
|
|
1595 |
if (!(MVequal(pmv[0],backupMV)) ) |
//initial skip decision |
|
{ |
|
|
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
1596 |
|
|
1597 |
if (iSAD < iMinSAD) |
if ( (pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
1598 |
{ |
&& ( //(pMB->mode == MODE_NOT_CODED) || |
1599 |
*currMV = newMV; |
(SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant) )) ) { |
1600 |
iMinSAD = iSAD; |
if (sad00 < pMB->quant * INITIAL_SKIP_THRESH) { |
1601 |
} |
SkipMacroblockP(pMB, sad00); |
1602 |
|
continue; } //skipped |
1603 |
} |
} |
1604 |
|
else sad00 = 256*4096; |
1605 |
|
|
1606 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
if (pMB->mode == MODE_NOT_CODED) |
1607 |
{ |
SearchP( pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1608 |
iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
y, current->motion_flags, pMB->quant, |
1609 |
x, y, |
current->fcode, pParam, pMBs, reference->mbs, |
1610 |
0, 0, iMinSAD, &newMV, |
current->global_flags & XVID_INTER4V, pMB); |
1611 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
1612 |
|
else |
1613 |
|
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1614 |
|
y, current->motion_flags, pMB->quant, |
1615 |
|
current->fcode, pParam, pMBs, |
1616 |
|
current->global_flags & XVID_INTER4V, pMB); |
1617 |
|
|
1618 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1619 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
1620 |
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
1621 |
|
SkipMacroblockP(pMB, sad00); |
1622 |
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
1623 |
} |
} |
1624 |
} |
} |
|
|
|
|
/*************** Choose best MV found **************/ |
|
|
|
|
|
EPZS8_Terminate_with_Refine: |
|
|
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
|
|
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
|
|
|
|
EPZS8_Terminate_without_Refine: |
|
|
|
|
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
1625 |
} |
} |
1626 |
|
|