1 |
/************************************************************************** |
/************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* motion estimation |
5 |
* |
* |
6 |
* 01.05.2002 updated MotionEstimationBVOP |
* This program is an implementation of a part of one or more MPEG-4 |
7 |
* 25.04.2002 partial prevMB conversion |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
8 |
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
* to use this software module in hardware or software products are |
9 |
* 14.04.2002 added MotionEstimationBVOP() |
* advised that its use may infringe existing patents or copyrights, and |
10 |
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
* any such use would be at such party's own risk. The original |
11 |
* EPZS and EPZS^2 |
* developer of this software module and his/her company, and subsequent |
12 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* editors and their companies, will have no liability for use of this |
13 |
* PMVFast_Refine to support multiple searches with different start points |
* software or modifications or derivatives thereof. |
|
* 07.01.2002 uv-block-based interpolation |
|
|
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
|
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
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 "../image/interpolate8x8.h" |
41 |
|
#include "motion_est.h" |
42 |
#include "motion.h" |
#include "motion.h" |
43 |
#include "sad.h" |
#include "sad.h" |
44 |
|
#include "../utils/emms.h" |
45 |
|
|
46 |
// very large value |
#define INITIAL_SKIP_THRESH (10) |
47 |
#define MV_MAX_ERROR (4096 * 256) |
#define FINAL_SKIP_THRESH (50) |
48 |
|
#define MAX_SAD00_FOR_SKIP (20) |
49 |
|
#define MAX_CHROMA_SAD_FOR_SKIP (22) |
50 |
|
#define SKIP_THRESH_B (25) |
51 |
|
|
52 |
// stop search if sdelta < THRESHOLD |
#define CHECK_CANDIDATE(X,Y,D) { \ |
53 |
#define MV16_THRESHOLD 192 |
(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } |
|
#define MV8_THRESHOLD 56 |
|
54 |
|
|
55 |
#define NEIGH_MOVE_THRESH 0 |
#define GET_REFERENCE(X, Y, REF) { \ |
56 |
// how much a block's MV must differ from his neighbour |
switch ( (((X)&1)<<1) + ((Y)&1) ) \ |
57 |
// to be search for INTER4V. The more, the faster... |
{ \ |
58 |
|
case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
59 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
60 |
/* nb = vop pixels * 2^(bpp-8) */ |
case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
61 |
#define MV16_00_BIAS (128+1) |
default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
62 |
#define MV8_00_BIAS (0) |
} \ |
63 |
|
} |
64 |
|
|
65 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
#define iDiamondSize 2 |
|
#define MV16_INTER_BIAS 512 |
|
66 |
|
|
67 |
/* Parameters which control inter/inter4v decision */ |
static __inline int |
68 |
#define IMV16X16 5 |
d_mv_bits(int x, int y, const uint32_t iFcode) |
69 |
|
{ |
70 |
|
int xb, yb; |
71 |
|
|
72 |
/* vector map (vlc delta size) smoother parameters */ |
if (x == 0) xb = 1; |
73 |
#define NEIGH_TEND_16X16 2 |
else { |
74 |
#define NEIGH_TEND_8X8 2 |
if (x < 0) x = -x; |
75 |
|
x += (1 << (iFcode - 1)) - 1; |
76 |
|
x >>= (iFcode - 1); |
77 |
|
if (x > 32) x = 32; |
78 |
|
xb = mvtab[x] + iFcode; |
79 |
|
} |
80 |
|
|
81 |
// fast ((A)/2)*2 |
if (y == 0) yb = 1; |
82 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
else { |
83 |
|
if (y < 0) y = -y; |
84 |
|
y += (1 << (iFcode - 1)) - 1; |
85 |
|
y >>= (iFcode - 1); |
86 |
|
if (y > 32) y = 32; |
87 |
|
yb = mvtab[y] + iFcode; |
88 |
|
} |
89 |
|
return xb + yb; |
90 |
|
} |
91 |
|
|
|
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
|
|
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
|
92 |
|
|
93 |
int32_t PMVfastSearch16( |
/* CHECK_CANDIATE FUNCTIONS START */ |
|
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 uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
|
|
|
|
int32_t EPZSSearch16( |
|
|
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 uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
94 |
|
|
95 |
|
static void |
96 |
|
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
97 |
|
{ |
98 |
|
int t; |
99 |
|
const uint8_t * Reference; |
100 |
|
|
101 |
int32_t PMVfastSearch8( |
if (( x > data->max_dx) || ( x < data->min_dx) |
102 |
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, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
103 |
|
|
104 |
int32_t EPZSSearch8( |
switch ( ((x&1)<<1) + (y&1) ) { |
105 |
const uint8_t * const pRef, |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
106 |
const uint8_t * const pRefH, |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
107 |
const uint8_t * const pRefV, |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
108 |
const uint8_t * const pRefHV, |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
109 |
const IMAGE * const pCur, |
} |
|
const int x, const int y, |
|
|
const int start_x, const int start_y, |
|
|
const uint32_t MotionFlags, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
110 |
|
|
111 |
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
112 |
|
|
113 |
typedef int32_t (MainSearch16Func)( |
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
114 |
const uint8_t * const pRef, |
data->temp[0] += data->lambda16 * t; |
115 |
const uint8_t * const pRefH, |
data->temp[1] += data->lambda8 * t; |
|
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); |
|
116 |
|
|
117 |
typedef MainSearch16Func* MainSearch16FuncPtr; |
if (data->temp[0] < data->iMinSAD[0]) { |
118 |
|
data->iMinSAD[0] = data->temp[0]; |
119 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
120 |
|
*dir = Direction; } |
121 |
|
|
122 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
123 |
|
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
124 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
125 |
|
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
126 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
127 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
128 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
129 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
130 |
|
|
131 |
typedef int32_t (MainSearch8Func)( |
} |
|
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); |
|
|
|
|
|
typedef MainSearch8Func* MainSearch8FuncPtr; |
|
|
|
|
|
static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ |
|
|
{ 0 ,(int)(1.00235+0.5), (int)(1.15582+0.5), (int)(1.31976+0.5), (int)(1.49591+0.5), (int)(1.68601+0.5), |
|
|
(int)(1.89187+0.5), (int)(2.11542+0.5), (int)(2.35878+0.5), (int)(2.62429+0.5), (int)(2.91455+0.5), |
|
|
(int)(3.23253+0.5), (int)(3.58158+0.5), (int)(3.96555+0.5), (int)(4.38887+0.5), (int)(4.85673+0.5), |
|
|
(int)(5.37519+0.5), (int)(5.95144+0.5), (int)(6.59408+0.5), (int)(7.31349+0.5), (int)(8.12242+0.5), |
|
|
(int)(9.03669+0.5), (int)(10.0763+0.5), (int)(11.2669+0.5), (int)(12.6426+0.5), (int)(14.2493+0.5), |
|
|
(int)(16.1512+0.5), (int)(18.442+0.5), (int)(21.2656+0.5), (int)(24.8580+0.5), (int)(29.6436+0.5), |
|
|
(int)(36.4949+0.5) }; |
|
|
|
|
|
static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now*/ |
|
132 |
|
|
133 |
|
static void |
134 |
|
CheckCandidate16Q(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
135 |
|
{ |
136 |
|
//this function expects X and Y in halfpel positions, but assumes that qpel positions are our goal |
137 |
|
int t; |
138 |
|
const uint8_t * Reference; |
139 |
|
|
140 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
141 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
142 |
|
|
143 |
// mv.length table |
switch ( ((x&1)<<1) + (y&1) ) { |
144 |
static const uint32_t mvtab[33] = { |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
145 |
1, 2, 3, 4, 6, 7, 7, 7, |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
146 |
9, 9, 9, 10, 10, 10, 10, 10, |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
147 |
10, 10, 10, 10, 10, 10, 10, 10, |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
148 |
10, 11, 11, 11, 11, 11, 11, 12, 12 |
} |
|
}; |
|
149 |
|
|
150 |
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); |
151 |
|
|
152 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
153 |
{ |
data->temp[0] += data->lambda16 * t; |
154 |
if (component == 0) |
data->temp[1] += data->lambda8 * t; |
|
return 1; |
|
155 |
|
|
156 |
if (component < 0) |
if (data->temp[0] < data->iMinSAD[0]) { |
157 |
component = -component; |
data->iMinSAD[0] = data->temp[0]; |
158 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
159 |
|
*dir = Direction; } |
160 |
|
|
161 |
if (iFcode == 1) |
if (data->temp[1] < data->iMinSAD[1]) { |
162 |
{ |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
163 |
if (component > 32) |
if (data->temp[2] < data->iMinSAD[2]) { |
164 |
component = 32; |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
165 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
166 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
167 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
168 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
169 |
|
|
|
return mvtab[component] + 1; |
|
170 |
} |
} |
171 |
|
|
172 |
component += (1 << (iFcode - 1)) - 1; |
static void |
173 |
component >>= (iFcode - 1); |
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
174 |
|
{ |
175 |
|
int32_t sad; |
176 |
|
const uint8_t * Reference; |
177 |
|
|
178 |
if (component > 32) |
if (( x > data->max_dx) || ( x < data->min_dx) |
179 |
component = 32; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
180 |
|
|
181 |
return mvtab[component] + 1 + iFcode - 1; |
switch ( ((x&1)<<1) + (y&1) ) |
182 |
|
{ |
183 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
184 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
185 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
186 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
187 |
} |
} |
188 |
|
|
189 |
|
sad = data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
190 |
|
sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
191 |
|
|
192 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
if (sad < *(data->iMinSAD)) { |
193 |
{ |
*(data->iMinSAD) = sad; |
194 |
return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
195 |
|
*dir = Direction; } |
196 |
} |
} |
197 |
|
|
198 |
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode, const uint32_t iQuant) |
static void |
199 |
|
CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
200 |
|
|
201 |
|
// CheckCandidate16 variant which expects x and y in quarter pixel resolution |
202 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
203 |
|
// around currentMV! |
204 |
{ |
{ |
205 |
return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int t; |
206 |
} |
uint8_t * Reference = (uint8_t *)data->RefQ; |
207 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
208 |
|
VECTOR halfpelMV = *(data->currentMV); |
209 |
|
|
210 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
211 |
|
uint32_t rounding = data->rounding; |
212 |
|
|
213 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
214 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
215 |
|
|
216 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases |
217 |
|
switch( ((x&1)<<1) + (y&1) ) |
218 |
|
{ |
219 |
|
case 0: // pure halfpel position - shouldn't happen during a refinement step |
220 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
221 |
|
break; |
222 |
|
|
223 |
#ifndef SEARCH16 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
224 |
#define SEARCH16 PMVfastSearch16 |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
225 |
//#define SEARCH16 FullSearch16 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
226 |
//#define SEARCH16 EPZSSearch16 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
227 |
#endif |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
228 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
229 |
|
break; |
230 |
|
|
231 |
#ifndef SEARCH8 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
232 |
#define SEARCH8 PMVfastSearch8 |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
233 |
//#define SEARCH8 EPZSSearch8 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
234 |
#endif |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
235 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
236 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
237 |
|
break; |
238 |
|
|
239 |
bool MotionEstimation( |
default: // x and y in qpel resolution - the "corners" (top left/right and |
240 |
MBParam * const pParam, |
// bottom left/right) during qpel refinement |
241 |
FRAMEINFO * const current, |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
242 |
FRAMEINFO * const reference, |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
243 |
const IMAGE * const pRefH, |
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
const uint32_t iLimit) |
|
244 |
|
|
245 |
{ |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
246 |
const uint32_t iWcount = pParam->mb_width; |
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
247 |
const uint32_t iHcount = pParam->mb_height; |
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
248 |
MACROBLOCK * const pMBs = current->mbs; |
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
249 |
MACROBLOCK * const prevMBs = reference->mbs; // previous frame |
break; |
250 |
|
} |
251 |
|
|
252 |
const IMAGE * const pCurrent = ¤t->image; |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); |
|
const IMAGE * const pRef = &reference->image; |
|
253 |
|
|
254 |
const VECTOR zeroMV = {0,0}; |
t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
255 |
|
data->temp[0] += data->lambda16 * t; |
256 |
|
data->temp[1] += data->lambda8 * t; |
257 |
|
|
258 |
int32_t x, y; |
if (data->temp[0] < data->iMinSAD[0]) { |
259 |
int32_t iIntra = 0; |
data->iMinSAD[0] = data->temp[0]; |
260 |
VECTOR pmv; |
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
261 |
|
/* *dir = Direction;*/ } |
262 |
|
|
263 |
if (sadInit) |
if (data->temp[1] < data->iMinSAD[1]) { |
264 |
(*sadInit)(); |
data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
265 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
266 |
|
data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
267 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
268 |
|
data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
269 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
270 |
|
data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
271 |
|
} |
272 |
|
|
273 |
for (y = 0; y < iHcount; y++) |
static void |
274 |
for (x = 0; x < iWcount; x++) |
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
275 |
{ |
{ |
276 |
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
int32_t sad; |
277 |
|
|
278 |
pMB->sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (( x > data->max_dx) || ( x < data->min_dx) |
279 |
x, y, current->motion_flags, current->quant, current->fcode, |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
pParam, pMBs, prevMBs, &pMB->mv16, &pMB->pmvs[0]); |
|
280 |
|
|
281 |
|
sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
282 |
|
data->iEdgedWidth, 256*4096); |
283 |
|
|
284 |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) |
if (sad < *(data->iMinSAD)) { |
285 |
{ |
*(data->iMinSAD) = sad; |
286 |
int32_t deviation; |
data->currentMV[0].x = x; data->currentMV[0].y = y; |
287 |
deviation = dev16(pCurrent->y + x*16 + y*16*pParam->edged_width, |
*dir = Direction; } |
288 |
pParam->edged_width); |
} |
289 |
|
|
290 |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) |
|
291 |
|
static void |
292 |
|
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
293 |
{ |
{ |
294 |
pMB->mode = MODE_INTRA; |
int32_t sad; |
295 |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
const int xb = data->currentMV[1].x; |
296 |
= pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
const int yb = data->currentMV[1].y; |
297 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] |
const uint8_t *ReferenceF, *ReferenceB; |
|
= pMB->sad8[2] = pMB->sad8[3] = 0; |
|
298 |
|
|
299 |
iIntra++; |
if (( xf > data->max_dx) || ( xf < data->min_dx) |
300 |
if (iIntra >= iLimit) |
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
|
return 1; |
|
301 |
|
|
302 |
continue; |
switch ( ((xf&1)<<1) + (yf&1) ) { |
303 |
} |
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
304 |
|
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
305 |
|
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
306 |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
307 |
} |
} |
308 |
pMB->mode = MODE_INTER; |
|
309 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
switch ( ((xb&1)<<1) + (yb&1) ) { |
310 |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = pMB->sad16; |
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
311 |
|
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
312 |
|
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
313 |
|
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
314 |
} |
} |
315 |
|
|
316 |
// we try to do as few INTER4V-searches as possible. So we split ME in two parts, normal |
sad = data->lambda16 * |
317 |
// SEARCH16 and only for special blocks SEARCH8. May this should be modified for quality |
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
318 |
// levels. |
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
319 |
|
|
320 |
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
321 |
|
|
322 |
|
if (sad < *(data->iMinSAD)) { |
323 |
|
*(data->iMinSAD) = sad; |
324 |
|
data->currentMV->x = xf; data->currentMV->y = yf; |
325 |
|
*dir = Direction; } |
326 |
|
} |
327 |
|
|
328 |
if (current->global_flags & XVID_INTER4V) |
static void |
329 |
for (y = 0; y < iHcount; y++) |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
for (x = 0; x < iWcount; x++) |
|
330 |
{ |
{ |
331 |
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
int32_t sad; |
332 |
|
int k; |
333 |
|
const uint8_t *ReferenceF; |
334 |
|
const uint8_t *ReferenceB; |
335 |
|
VECTOR mvs, b_mvs; |
336 |
|
|
337 |
if (pMB->mode == MODE_INTRA) |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
continue; |
|
338 |
|
|
339 |
|
sad = data->lambda16 * d_mv_bits(x, y, 1); |
340 |
|
|
341 |
if ( (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) ) |
for (k = 0; k < 4; k++) { |
342 |
{ |
mvs.x = data->directmvF[k].x + x; |
343 |
int32_t neigh=0; |
b_mvs.x = ((x == 0) ? |
344 |
|
data->directmvB[k].x |
345 |
|
: mvs.x - data->referencemv[k].x); |
346 |
|
|
347 |
if (x>0) |
mvs.y = data->directmvF[k].y + y; |
348 |
{ neigh += abs((pMB->mv16.x)-((pMB-1)->mv16.x)); |
b_mvs.y = ((y == 0) ? |
349 |
neigh += abs((pMB->mv16.y)-((pMB-1)->mv16.y)); |
data->directmvB[k].y |
350 |
} |
: mvs.y - data->referencemv[k].y); |
|
if (y>0) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB-iWcount)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB-iWcount)->mv16.y)); |
|
|
} |
|
|
if (x<(iWcount-1)) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB+1)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB+1)->mv16.y)); |
|
|
} |
|
|
if (y<(iHcount-1)) |
|
|
{ neigh += abs((pMB->mv16.x)-((pMB+iHcount)->mv16.x)); |
|
|
neigh += abs((pMB->mv16.y)-((pMB+iHcount)->mv16.y)); |
|
|
} |
|
|
|
|
|
if (neigh > NEIGH_MOVE_THRESH) |
|
|
{ |
|
|
int32_t sad8 = IMV16X16 * current->quant; |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[0] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x, 2*y, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[1] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x+1, 2*y, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[2] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
|
|
|
|
|
if (sad8 < pMB->sad16) |
|
|
sad8 += pMB->sad8[3] |
|
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
2*x+1, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
|
|
current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
|
|
|
|
|
/* decide: MODE_INTER or MODE_INTER4V |
|
|
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
|
|
*/ |
|
351 |
|
|
352 |
if (sad8 < pMB->sad16) |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
353 |
{ |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
354 |
pMB->mode = MODE_INTER4V; |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
355 |
pMB->sad8[0] *= 4; |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
pMB->sad8[1] *= 4; |
|
|
pMB->sad8[2] *= 4; |
|
|
pMB->sad8[3] *= 4; |
|
|
continue; |
|
|
} |
|
356 |
|
|
357 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
358 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
359 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
360 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
361 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
362 |
} |
} |
363 |
|
|
364 |
|
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
365 |
|
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
366 |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
367 |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
368 |
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
369 |
} |
} |
370 |
|
|
371 |
// get_pmv has to be called again, because inter4v changes predictors |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
372 |
|
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
373 |
pmv = get_pmv(pMBs, x, y, pParam->mb_width, 0); |
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
374 |
pMB->pmvs[0].x = pMB->mv16.x - pmv.x; /* the other pmvs are only needed in INTER4V-mode */ |
data->iEdgedWidth); |
375 |
pMB->pmvs[0].y = pMB->mv16.y - pmv.y; |
if (sad > *(data->iMinSAD)) return; |
|
|
|
376 |
} |
} |
377 |
|
|
378 |
return 0; |
if (sad < *(data->iMinSAD)) { |
379 |
|
*(data->iMinSAD) = sad; |
380 |
|
data->currentMV->x = x; data->currentMV->y = y; |
381 |
|
*dir = Direction; } |
382 |
} |
} |
383 |
|
|
384 |
#define CHECK_MV16_ZERO {\ |
static void |
385 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
386 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
{ |
387 |
{ \ |
int32_t sad; |
388 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
const uint8_t *ReferenceF; |
389 |
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ |
const uint8_t *ReferenceB; |
390 |
if (iSAD < iMinSAD) \ |
VECTOR mvs, b_mvs; |
|
{ 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); } } \ |
|
|
} |
|
391 |
|
|
392 |
#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
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); } } \ |
|
|
} |
|
393 |
|
|
394 |
#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ |
sad = data->lambda16 * d_mv_bits(x, y, 1); |
|
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; } } \ |
|
|
} |
|
395 |
|
|
396 |
|
mvs.x = data->directmvF[0].x + x; |
397 |
|
b_mvs.x = ((x == 0) ? |
398 |
|
data->directmvB[0].x |
399 |
|
: mvs.x - data->referencemv[0].x); |
400 |
|
|
401 |
#define CHECK_MV8_ZERO {\ |
mvs.y = data->directmvF[0].y + y; |
402 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
b_mvs.y = ((y == 0) ? |
403 |
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ |
data->directmvB[0].y |
404 |
if (iSAD < iMinSAD) \ |
: mvs.y - data->referencemv[0].y); |
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
|
|
} |
|
405 |
|
|
406 |
#define NOCHECK_MV8_CANDIDATE(X,Y) \ |
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
407 |
{ \ |
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
408 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
409 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ |
|
|
} |
|
410 |
|
|
411 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
412 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
413 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
414 |
{ \ |
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
415 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
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); } } \ |
|
416 |
} |
} |
417 |
|
|
418 |
#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ |
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
419 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
420 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
421 |
{ \ |
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
422 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
|
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); } } \ |
|
423 |
} |
} |
424 |
|
|
425 |
#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ |
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
426 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
427 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
if (sad < *(data->iMinSAD)) { |
428 |
{ \ |
*(data->iMinSAD) = sad; |
429 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
data->currentMV->x = x; data->currentMV->y = y; |
430 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ |
*dir = Direction; } |
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
431 |
} |
} |
432 |
|
|
433 |
/* too slow and not fully functional at the moment */ |
static void |
434 |
/* |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
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, |
|
|
const uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
435 |
{ |
{ |
436 |
const int32_t iEdgedWidth = pParam->edged_width; |
int32_t sad; |
437 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
const uint8_t * Reference; |
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
438 |
|
|
439 |
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
if (( x > data->max_dx) || ( x < data->min_dx) |
440 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
441 |
|
|
442 |
iSAD = sad16( cur, |
switch ( ((x&1)<<1) + (y&1) ) |
443 |
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
{ |
444 |
iEdgedWidth, MV_MAX_ERROR); |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
445 |
if (iSAD <= iQuant * 96) |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
446 |
iSAD -= MV16_00_BIAS; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
447 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
448 |
currMV->x = 0; |
} |
|
currMV->y = 0; |
|
|
currPMV->x = -pred_x; |
|
|
currPMV->y = -pred_y; |
|
449 |
|
|
450 |
return iSAD; |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
451 |
|
sad += data->lambda8 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
452 |
|
|
453 |
|
if (sad < *(data->iMinSAD)) { |
454 |
|
*(data->iMinSAD) = sad; |
455 |
|
data->currentMV->x = x; data->currentMV->y = y; |
456 |
|
*dir = Direction; } |
457 |
} |
} |
|
*/ |
|
458 |
|
|
459 |
int32_t Diamond16_MainSearch( |
static void |
460 |
const uint8_t * const pRef, |
CheckCandidate8Q(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
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 |
|
461 |
{ |
{ |
462 |
currMV->x = startx; |
int32_t sad; |
463 |
currMV->y = starty; |
const uint8_t * Reference; |
464 |
} |
|
465 |
return iMinSAD; |
if (( x > data->max_dx) || ( x < data->min_dx) |
466 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
467 |
|
|
468 |
|
switch ( ((x&1)<<1) + (y&1) ) |
469 |
|
{ |
470 |
|
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
471 |
|
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
472 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
473 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
474 |
} |
} |
475 |
|
|
476 |
int32_t Square16_MainSearch( |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
477 |
const uint8_t * const pRef, |
sad += data->lambda8 * d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); |
|
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 |
|
|
*/ |
|
|
|
|
|
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); |
|
|
|
|
|
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); |
|
478 |
|
|
479 |
|
if (sad < *(data->iMinSAD)) { |
480 |
|
*(data->iMinSAD) = sad; |
481 |
|
data->currentMV->x = x; data->currentMV->y = y; |
482 |
|
*dir = Direction; } |
483 |
|
} |
484 |
|
|
485 |
if (iDirection) |
static void |
486 |
while (!iFound) |
CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
487 |
{ |
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
488 |
iFound = 1; |
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
489 |
backupMV=*currMV; |
// around currentMV! |
490 |
|
|
|
switch (iDirection) |
|
491 |
{ |
{ |
492 |
case 1: |
int32_t sad; |
493 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
uint8_t *Reference = (uint8_t *) data->RefQ; |
494 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
const uint8_t *ref1, *ref2, *ref3, *ref4; |
495 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
VECTOR halfpelMV = *(data->currentMV); |
|
break; |
|
|
case 2: |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
break; |
|
496 |
|
|
497 |
case 3: |
int32_t iEdgedWidth = data->iEdgedWidth; |
498 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
uint32_t rounding = data->rounding; |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
|
break; |
|
499 |
|
|
500 |
case 4: |
if (( x > data->max_dx) || ( x < data->min_dx) |
501 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
|
|
break; |
|
502 |
|
|
503 |
case 5: |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
504 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
switch( ((x&1)<<1) + (y&1) ) |
505 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
{ |
506 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
507 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
508 |
break; |
break; |
509 |
|
|
510 |
case 6: |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
511 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
|
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,8); |
|
512 |
|
|
513 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
514 |
break; |
break; |
515 |
|
|
516 |
case 7: |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
517 |
CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
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); |
|
|
break; |
|
518 |
|
|
519 |
case 8: |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
|
|
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); |
|
520 |
break; |
break; |
521 |
default: |
|
522 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); |
default: // x and y in qpel resolution - the "corners" (top left/right and |
523 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); |
// bottom left/right) during qpel refinement |
524 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
525 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
526 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
527 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); |
|
528 |
CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); |
|
|
CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); |
|
529 |
break; |
break; |
530 |
} |
} |
|
} |
|
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
|
} |
|
|
|
|
531 |
|
|
532 |
int32_t Full16_MainSearch( |
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
533 |
const uint8_t * const pRef, |
sad += data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
|
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_MV16_CANDIDATE(dx,dy); |
|
534 |
|
|
535 |
return iMinSAD; |
if (sad < *(data->iMinSAD)) { |
536 |
|
*(data->iMinSAD) = sad; |
537 |
|
data->currentQMV->x = x; data->currentQMV->y = y; |
538 |
|
*dir = Direction; } |
539 |
} |
} |
540 |
|
|
541 |
int32_t AdvDiamond16_MainSearch( |
/* CHECK_CANDIATE FUNCTIONS END */ |
|
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 iDirection) |
|
|
{ |
|
|
|
|
|
int32_t iSAD; |
|
542 |
|
|
543 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
/* MAINSEARCH FUNCTIONS START */ |
544 |
|
|
545 |
if (iDirection) |
static void |
546 |
{ |
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
|
CHECK_MV16_CANDIDATE(startx-iDiamondSize, starty); |
|
|
CHECK_MV16_CANDIDATE(startx+iDiamondSize, starty); |
|
|
CHECK_MV16_CANDIDATE(startx, starty-iDiamondSize); |
|
|
CHECK_MV16_CANDIDATE(startx, starty+iDiamondSize); |
|
|
} |
|
|
else |
|
|
{ |
|
|
int bDirection = 1+2+4+8; |
|
|
do |
|
547 |
{ |
{ |
|
iDirection = 0; |
|
|
if (bDirection&1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
|
|
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize,starty,1); |
|
548 |
|
|
549 |
if (bDirection&2) |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize,starty,2); |
|
550 |
|
|
551 |
if (bDirection&4) |
int iDirection; |
|
CHECK_MV16_CANDIDATE_DIR(startx,starty-iDiamondSize,4); |
|
552 |
|
|
553 |
if (bDirection&8) |
do { |
554 |
CHECK_MV16_CANDIDATE_DIR(startx,starty+iDiamondSize,8); |
iDirection = 0; |
555 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
556 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
557 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
558 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
559 |
|
|
560 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
561 |
|
|
562 |
if (iDirection) //checking if anything found |
if (iDirection) { //checking if anything found |
|
{ |
|
563 |
bDirection = iDirection; |
bDirection = iDirection; |
564 |
iDirection = 0; |
iDirection = 0; |
565 |
startx=currMV->x; starty=currMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
566 |
if (bDirection & 3) //our candidate is left or right |
if (bDirection & 3) { //our candidate is left or right |
567 |
{ |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
568 |
CHECK_MV16_CANDIDATE_DIR(startx,starty+iDiamondSize, 8); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
569 |
CHECK_MV16_CANDIDATE_DIR(startx,starty-iDiamondSize, 4); |
} else { // what remains here is up or down |
570 |
} |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
571 |
else // what remains here is up or down |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
572 |
{ |
|
573 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty, 2); |
if (iDirection) { |
574 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty, 1); |
bDirection += iDirection; |
575 |
} |
x = data->currentMV->x; y = data->currentMV->y; } |
576 |
|
} else { //about to quit, eh? not so fast.... |
577 |
if (iDirection) |
switch (bDirection) { |
|
{ bDirection+=iDirection; |
|
|
startx=currMV->x; starty=currMV->y; |
|
|
} |
|
|
} |
|
|
else //about to quit, eh? not so fast.... |
|
|
{ |
|
|
switch (bDirection) |
|
|
{ |
|
578 |
case 2: |
case 2: |
579 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
580 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
581 |
break; |
break; |
582 |
case 1: |
case 1: |
583 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
584 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
585 |
break; |
break; |
586 |
case 2+4: |
case 2+4: |
587 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
588 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
589 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
590 |
break; |
break; |
591 |
case 4: |
case 4: |
592 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
593 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
594 |
break; |
break; |
595 |
case 8: |
case 8: |
596 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
597 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
598 |
break; |
break; |
599 |
case 1+4: |
case 1+4: |
600 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
601 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
602 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
603 |
break; |
break; |
604 |
case 2+8: |
case 2+8: |
605 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
606 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
607 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
608 |
break; |
break; |
609 |
case 1+8: |
case 1+8: |
610 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
611 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
612 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
613 |
break; |
break; |
614 |
default: //1+2+4+8 == we didn't find anything at all |
default: //1+2+4+8 == we didn't find anything at all |
615 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
616 |
CHECK_MV16_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
617 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
618 |
CHECK_MV16_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
619 |
break; |
break; |
620 |
} |
} |
621 |
if (!iDirection) break; //ok, the end. really |
if (!iDirection) break; //ok, the end. really |
622 |
else |
bDirection = iDirection; |
623 |
{ bDirection=iDirection; |
x = data->currentMV->x; y = data->currentMV->y; |
|
startx=currMV->x; starty=currMV->y; |
|
|
} |
|
624 |
} |
} |
625 |
} |
} |
626 |
while (1); //forever |
while (1); //forever |
627 |
} |
} |
|
return iMinSAD; |
|
|
} |
|
628 |
|
|
629 |
int32_t AdvDiamond8_MainSearch( |
static void |
630 |
const uint8_t * const pRef, |
SquareSearch(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 iDirection) |
|
631 |
{ |
{ |
632 |
|
int iDirection; |
633 |
|
|
634 |
int32_t iSAD; |
do { |
635 |
|
iDirection = 0; |
636 |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
637 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
638 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
639 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
640 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
641 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
642 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
643 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
644 |
|
|
645 |
if (iDirection) |
bDirection = iDirection; |
646 |
{ |
x = data->currentMV->x; y = data->currentMV->y; |
647 |
CHECK_MV8_CANDIDATE(startx-iDiamondSize, starty); |
} while (iDirection); |
|
CHECK_MV8_CANDIDATE(startx+iDiamondSize, starty); |
|
|
CHECK_MV8_CANDIDATE(startx, starty-iDiamondSize); |
|
|
CHECK_MV8_CANDIDATE(startx, starty+iDiamondSize); |
|
648 |
} |
} |
649 |
else |
|
650 |
{ |
static void |
651 |
int bDirection = 1+2+4+8; |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
|
do |
|
652 |
{ |
{ |
|
iDirection = 0; |
|
|
if (bDirection&1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize,starty,1); |
|
653 |
|
|
654 |
if (bDirection&2) |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize,starty,2); |
|
655 |
|
|
656 |
if (bDirection&4) |
int iDirection; |
|
CHECK_MV8_CANDIDATE_DIR(startx,starty-iDiamondSize,4); |
|
657 |
|
|
658 |
if (bDirection&8) |
do { |
659 |
CHECK_MV8_CANDIDATE_DIR(startx,starty+iDiamondSize,8); |
iDirection = 0; |
660 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
661 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
662 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
663 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
664 |
|
|
665 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
666 |
|
|
667 |
if (iDirection) //checking if anything found |
if (iDirection) { //checking if anything found |
|
{ |
|
668 |
bDirection = iDirection; |
bDirection = iDirection; |
669 |
iDirection = 0; |
iDirection = 0; |
670 |
startx=currMV->x; starty=currMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
671 |
if (bDirection & 3) //our candidate is left or right |
if (bDirection & 3) { //our candidate is left or right |
672 |
{ |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
673 |
CHECK_MV8_CANDIDATE_DIR(startx,starty+iDiamondSize, 8); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
674 |
CHECK_MV8_CANDIDATE_DIR(startx,starty-iDiamondSize, 4); |
} else { // what remains here is up or down |
675 |
} |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
676 |
else // what remains here is up or down |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
|
{ |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty, 2); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty, 1); |
|
|
} |
|
677 |
|
|
678 |
if (iDirection) |
bDirection += iDirection; |
679 |
{ bDirection+=iDirection; |
x = data->currentMV->x; y = data->currentMV->y; |
|
startx=currMV->x; starty=currMV->y; |
|
|
} |
|
|
} |
|
|
else //about to quit, eh? not so fast.... |
|
|
{ |
|
|
switch (bDirection) |
|
|
{ |
|
|
case 2: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
break; |
|
|
case 1: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
break; |
|
|
case 2+4: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
break; |
|
|
case 4: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
break; |
|
|
case 8: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
break; |
|
|
case 1+4: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
break; |
|
|
case 2+8: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
break; |
|
|
case 1+8: |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
break; |
|
|
default: //1+2+4+8 == we didn't find anything at all |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty-iDiamondSize, 1+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx-iDiamondSize, starty+iDiamondSize, 1+8); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty-iDiamondSize, 2+4); |
|
|
CHECK_MV8_CANDIDATE_DIR(startx+iDiamondSize, starty+iDiamondSize, 2+8); |
|
|
break; |
|
|
} |
|
|
if (!(iDirection)) break; //ok, the end. really |
|
|
else |
|
|
{ bDirection=iDirection; |
|
|
startx=currMV->x; starty=currMV->y; |
|
680 |
} |
} |
681 |
} |
} |
682 |
} |
while (iDirection); |
|
while (1); //forever |
|
|
} |
|
|
return iMinSAD; |
|
683 |
} |
} |
684 |
|
|
685 |
|
/* MAINSEARCH FUNCTIONS END */ |
686 |
|
|
687 |
int32_t Full8_MainSearch( |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
|
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); |
|
|
|
|
|
return iMinSAD; |
|
|
} |
|
688 |
|
|
689 |
|
static void |
690 |
|
HalfpelRefine(const SearchData * const data) |
|
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) |
|
691 |
{ |
{ |
692 |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
693 |
|
|
694 |
int32_t iSAD; |
VECTOR backupMV = *(data->currentMV); |
695 |
VECTOR backupMV = *currMV; |
int iDirection; //not needed |
696 |
|
|
697 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
698 |
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
699 |
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
700 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
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); |
|
701 |
|
|
702 |
return iMinSAD; |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
703 |
} |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
704 |
|
|
705 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
706 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
707 |
|
} |
708 |
|
|
709 |
|
|
710 |
int32_t PMVfastSearch16( |
static void |
711 |
const uint8_t * const pRef, |
QuarterpelRefine(const SearchData * const data) |
|
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 uint32_t iQuant, |
|
|
const uint32_t iFcode, |
|
|
const MBParam * const pParam, |
|
|
const MACROBLOCK * const pMBs, |
|
|
const MACROBLOCK * const prevMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
712 |
{ |
{ |
713 |
const uint32_t iWcount = pParam->mb_width; |
/* Perform quarter pixel refinement*/ |
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
714 |
|
|
715 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
VECTOR backupMV = *(data->currentQMV); |
716 |
|
int iDirection; //not needed |
717 |
|
|
718 |
int32_t iDiamondSize; |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
719 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
720 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
721 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
722 |
|
|
723 |
int32_t min_dx; |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
724 |
int32_t max_dx; |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
725 |
|
|
726 |
int32_t iFound; |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
727 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
VECTOR newMV; |
|
|
VECTOR backupMV; /* just for PMVFAST */ |
|
|
|
|
|
VECTOR pmv[4]; |
|
|
int32_t psad[4]; |
|
|
|
|
|
MainSearch16FuncPtr MainSearchPtr; |
|
|
|
|
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
|
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
|
|
|
|
|
static int32_t threshA,threshB; |
|
|
int32_t bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
|
|
|
|
/* Get maximum range */ |
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 16, iWidth, iHeight, iFcode); |
|
|
|
|
|
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
|
|
|
|
|
if (!(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; */ |
|
728 |
|
|
729 |
|
} |
730 |
|
|
731 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
static __inline int |
732 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
733 |
|
const int x, const int y, |
734 |
|
const uint32_t iEdgedWidth, const uint32_t iQuant) |
735 |
|
|
|
if ((x==0) && (y==0) ) |
|
736 |
{ |
{ |
737 |
threshA = 512; |
/* keep repeating checks for all b-frames before this P frame, |
738 |
threshB = 1024; |
to make sure that SKIP is possible (todo) |
739 |
|
how: if skip is not possible set sad00 to a very high value */ |
740 |
|
|
741 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
742 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
743 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
744 |
|
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
745 |
|
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
746 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
747 |
|
|
748 |
|
return 1; |
749 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]; |
|
|
threshB = threshA+256; |
|
|
if (threshA< 512) threshA = 512; |
|
|
if (threshA>1024) threshA = 1024; |
|
|
if (threshB>1792) threshB = 1792; |
|
|
} |
|
|
|
|
|
iFound=0; |
|
750 |
|
|
751 |
/* Step 4: Calculate SAD around the Median prediction. |
static __inline void |
752 |
MinSAD=SAD |
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
753 |
If Motion Vector equal to Previous frame motion vector |
{ |
754 |
and MinSAD<PrevFrmSAD goto Step 10. |
pMB->mode = MODE_NOT_CODED; |
755 |
If SAD<=256 goto Step 10. |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
756 |
*/ |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
757 |
|
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
758 |
|
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
759 |
|
|
760 |
*currMV=pmv[0]; /* current best := prediction */ |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ /* This should NOT be necessary! */ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
761 |
} |
} |
762 |
|
|
763 |
if (currMV->x > max_dx) |
bool |
764 |
{ |
MotionEstimation(MBParam * const pParam, |
765 |
currMV->x=max_dx; |
FRAMEINFO * const current, |
766 |
} |
FRAMEINFO * const reference, |
767 |
if (currMV->x < min_dx) |
const IMAGE * const pRefH, |
768 |
{ |
const IMAGE * const pRefV, |
769 |
currMV->x=min_dx; |
const IMAGE * const pRefHV, |
770 |
} |
const uint32_t iLimit) |
|
if (currMV->y > max_dy) |
|
|
{ |
|
|
currMV->y=max_dy; |
|
|
} |
|
|
if (currMV->y < min_dy) |
|
771 |
{ |
{ |
772 |
currMV->y=min_dy; |
MACROBLOCK *const pMBs = current->mbs; |
773 |
} |
const IMAGE *const pCurrent = ¤t->image; |
774 |
|
const IMAGE *const pRef = &reference->image; |
775 |
|
|
776 |
iMinSAD = sad16( cur, |
const VECTOR zeroMV = { 0, 0 }; |
|
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); |
|
777 |
|
|
778 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
uint32_t x, y; |
779 |
{ |
uint32_t iIntra = 0; |
780 |
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
int32_t InterBias, quant = current->quant; |
781 |
{ |
uint8_t *qimage; |
782 |
if (!MVzero(*currMV)) |
|
783 |
{ |
// some pre-initialized thingies for SearchP |
784 |
iMinSAD += MV16_00_BIAS; |
int32_t temp[5]; |
785 |
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
VECTOR currentMV[5]; |
786 |
iMinSAD -= MV16_00_BIAS; |
VECTOR currentQMV[5]; |
787 |
|
int32_t iMinSAD[5]; |
788 |
|
SearchData Data; |
789 |
|
Data.iEdgedWidth = pParam->edged_width; |
790 |
|
Data.currentMV = currentMV; |
791 |
|
Data.currentQMV = currentQMV; |
792 |
|
Data.iMinSAD = iMinSAD; |
793 |
|
Data.temp = temp; |
794 |
|
Data.iFcode = current->fcode; |
795 |
|
Data.rounding = pParam->m_rounding_type; |
796 |
|
|
797 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
798 |
|
return 1; // allocate some mem for qpel interpolated blocks |
799 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
800 |
|
// encoder_create() - so please fix me! |
801 |
|
|
802 |
|
if (sadInit) (*sadInit) (); |
803 |
|
|
804 |
|
for (y = 0; y < pParam->mb_height; y++) { |
805 |
|
for (x = 0; x < pParam->mb_width; x++) { |
806 |
|
|
807 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
808 |
|
int32_t sad00 = pMB->sad16 |
809 |
|
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
810 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
811 |
|
pParam->edged_width, pMB->sad8 ); |
812 |
|
|
813 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
814 |
|
pMB->dquant = NO_CHANGE; |
815 |
|
pMB->quant = current->quant; |
816 |
|
} else { |
817 |
|
if (pMB->dquant != NO_CHANGE) { |
818 |
|
quant += DQtab[pMB->dquant]; |
819 |
|
if (quant > 31) quant = 31; |
820 |
|
else if (quant < 1) quant = 1; |
821 |
|
} |
822 |
|
pMB->quant = quant; |
823 |
|
} |
824 |
|
|
825 |
|
/* initial skip decision */ |
826 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
827 |
|
|
828 |
|
if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
829 |
|
if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) |
830 |
|
if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
831 |
|
SkipMacroblockP(pMB, sad00); |
832 |
|
continue; |
833 |
} |
} |
834 |
} |
} |
835 |
|
|
836 |
if (MotionFlags & PMV_QUICKSTOP16) |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, |
837 |
goto PMVfast16_Terminate_without_Refine; |
y, current->motion_flags, pMB->quant, |
838 |
if (MotionFlags & PMV_EARLYSTOP16) |
&Data, pParam, pMBs, reference->mbs, |
839 |
goto PMVfast16_Terminate_with_Refine; |
current->global_flags & XVID_INTER4V, pMB); |
840 |
|
|
841 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
842 |
|
if (current->coding_type == P_VOP) { |
843 |
|
if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
844 |
|
&& ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) |
845 |
|
if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { |
846 |
|
SkipMacroblockP(pMB, sad00); |
847 |
|
continue; |
848 |
} |
} |
|
|
|
|
|
|
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
|
/* |
|
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
|
|
|
|
// (0,0) is always possible |
|
|
|
|
|
if (!MVzero(pmv[0])) |
|
|
CHECK_MV16_ZERO; |
|
|
|
|
|
// previous frame MV is always possible |
|
|
|
|
|
if (!MVzero(prevMB->mvs[0])) |
|
|
if (!MVequal(prevMB->mvs[0],pmv[0])) |
|
|
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
|
|
|
|
|
// left neighbour, if allowed |
|
|
|
|
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1],prevMB->mvs[0])) |
|
|
if (!MVequal(pmv[1],pmv[0])) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[1].x = EVEN(pmv[1].x); |
|
|
pmv[1].y = EVEN(pmv[1].y); |
|
849 |
} |
} |
850 |
|
|
851 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
/* finally, intra decision */ |
|
} |
|
852 |
|
|
853 |
// top neighbour, if allowed |
InterBias = MV16_INTER_BIAS; |
854 |
if (!MVzero(pmv[2])) |
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
855 |
if (!MVequal(pmv[2],prevMB->mvs[0])) |
if (y != 0) |
856 |
if (!MVequal(pmv[2],pmv[0])) |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
857 |
if (!MVequal(pmv[2],pmv[1])) |
if (x != 0) |
858 |
{ |
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
859 |
|
|
860 |
// top right neighbour, if allowed |
if (InterBias < pMB->sad16) { |
861 |
if (!MVzero(pmv[3])) |
const int32_t deviation = |
862 |
if (!MVequal(pmv[3],prevMB->mvs[0])) |
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
863 |
if (!MVequal(pmv[3],pmv[0])) |
pParam->edged_width); |
864 |
if (!MVequal(pmv[3],pmv[1])) |
|
865 |
if (!MVequal(pmv[3],pmv[2])) |
if (deviation < (pMB->sad16 - InterBias)) { |
866 |
{ |
if (++iIntra >= iLimit) { free(qimage); return 1; } |
867 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
pMB->mode = MODE_INTRA; |
868 |
{ pmv[3].x = EVEN(pmv[3].x); |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
869 |
pmv[3].y = EVEN(pmv[3].y); |
pMB->mvs[3] = zeroMV; |
870 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
871 |
|
pMB->qmvs[3] = zeroMV; |
872 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
873 |
|
pMB->sad8[3] = 0; |
874 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
|
875 |
} |
} |
876 |
} |
} |
|
|
|
|
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96)*/ ) |
|
|
iMinSAD -= MV16_00_BIAS; |
|
|
|
|
|
|
|
|
/* Step 6: If MinSAD <= thresa goto Step 10. |
|
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
|
|
|
|
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast16_Terminate_with_Refine; |
|
877 |
} |
} |
878 |
|
free(qimage); |
879 |
|
|
880 |
|
if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ |
881 |
/************ (Diamond Search) **************/ |
current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); |
|
/* |
|
|
Step 7: Perform Diamond search, with either the small or large diamond. |
|
|
If Found=2 only examine one Diamond pattern, and afterwards goto step 10 |
|
|
Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. |
|
|
If center then goto step 10. |
|
|
Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. |
|
|
Refine by using small diamond and goto step 10. |
|
|
*/ |
|
|
|
|
|
if (MotionFlags & PMV_USESQUARES16) |
|
|
MainSearchPtr = Square16_MainSearch; |
|
|
else |
|
|
if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
|
|
MainSearchPtr = AdvDiamond16_MainSearch; |
|
882 |
else |
else |
883 |
MainSearchPtr = Diamond16_MainSearch; |
current->GMC_MV = zeroMV; |
884 |
|
|
885 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
return 0; |
|
|
|
|
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
|
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
886 |
} |
} |
887 |
|
|
|
if (MotionFlags & PMV_EXTSEARCH16) |
|
|
{ |
|
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
|
|
|
|
if (!(MVequal(pmv[0],backupMV)) ) |
|
|
{ iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
888 |
|
|
889 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
|
{ iSAD = (*MainSearchPtr)(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); |
|
890 |
|
|
891 |
if (iSAD < iMinSAD) |
static __inline int |
892 |
|
make_mask(const VECTOR * const pmv, const int i) |
893 |
{ |
{ |
894 |
*currMV = newMV; |
int mask = 255, j; |
895 |
iMinSAD = iSAD; |
for (j = 0; j < i; j++) { |
896 |
} |
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
897 |
|
if (pmv[i].x == pmv[j].x) { |
898 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
899 |
|
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
900 |
|
} else |
901 |
|
if (pmv[i].y == pmv[j].y) { |
902 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
903 |
|
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
904 |
} |
} |
905 |
} |
} |
906 |
|
return mask; |
|
/* |
|
|
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
|
|
*/ |
|
|
|
|
|
PMVfast16_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); |
|
|
|
|
|
PMVfast16_Terminate_without_Refine: |
|
|
currPMV->x = currMV->x - pmv[0].x; |
|
|
currPMV->y = currMV->y - pmv[0].y; |
|
|
return iMinSAD; |
|
907 |
} |
} |
908 |
|
|
909 |
|
static __inline void |
910 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
911 |
|
const int iHcount, const MACROBLOCK * const prevMB) |
912 |
|
{ |
913 |
|
|
914 |
|
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
915 |
|
|
916 |
|
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
917 |
|
pmv[5].x = EVEN(pmv[3].x); |
918 |
|
pmv[5].y = EVEN(pmv[3].y); |
919 |
|
} else pmv[5].x = pmv[5].y = 0; |
920 |
|
|
921 |
|
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
922 |
|
else pmv[3].x = pmv[3].y = 0; |
923 |
|
|
924 |
|
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
925 |
|
else pmv[4].x = pmv[4].y = 0; |
926 |
|
|
927 |
int32_t Diamond8_MainSearch( |
// [1] median prediction |
928 |
const uint8_t * const pRef, |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
|
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_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; |
|
|
} |
|
|
|
|
|
int32_t Halfpel8_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) |
|
|
{ |
|
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
929 |
|
|
930 |
int32_t iSAD; |
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
|
VECTOR backupMV = *currMV; |
|
931 |
|
|
932 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
933 |
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
|
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); |
|
934 |
|
|
935 |
return iMinSAD; |
if ((x != iWcount-1) && (y != iHcount-1)) { |
936 |
|
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
937 |
|
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
938 |
|
} else pmv[6].x = pmv[6].y = 0; |
939 |
} |
} |
940 |
|
|
941 |
|
static void |
942 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
SearchP(const uint8_t * const pRef, |
|
|
|
|
int32_t PMVfastSearch8( |
|
|
const uint8_t * const pRef, |
|
943 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
944 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
945 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
946 |
|
const uint8_t * const pRefQ, |
947 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
948 |
const int x, const int y, |
const int x, |
949 |
const int start_x, const int start_y, |
const int y, |
950 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
951 |
const uint32_t iQuant, |
const uint32_t iQuant, |
952 |
const uint32_t iFcode, |
SearchData * const Data, |
953 |
const MBParam * const pParam, |
const MBParam * const pParam, |
954 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
955 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
956 |
VECTOR * const currMV, |
int inter4v, |
957 |
VECTOR * const currPMV) |
MACROBLOCK * const pMB) |
958 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
|
|
|
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
|
959 |
|
|
960 |
int32_t iDiamondSize; |
int i, iDirection = 255, mask, threshA; |
961 |
|
VECTOR pmv[7]; |
962 |
|
|
963 |
int32_t min_dx; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
964 |
int32_t max_dx; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
965 |
int32_t min_dy; |
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
966 |
int32_t max_dy; |
|
967 |
|
Data->predMV = pmv[0]; |
968 |
|
|
969 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
970 |
|
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
971 |
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
972 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
973 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
974 |
|
Data->RefQ = pRefQ; |
975 |
|
|
976 |
|
Data->lambda16 = lambda_vec16[iQuant]; |
977 |
|
Data->lambda8 = lambda_vec8[iQuant]; |
978 |
|
|
979 |
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
980 |
|
Data->min_dx = EVEN(Data->min_dx); |
981 |
|
Data->max_dx = EVEN(Data->max_dx); |
982 |
|
Data->min_dy = EVEN(Data->min_dy); |
983 |
|
Data->max_dy = EVEN(Data->max_dy); } |
984 |
|
|
985 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
986 |
|
|
987 |
|
for(i = 0; i < 5; i++) |
988 |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
989 |
|
|
990 |
|
if (pParam->m_quarterpel) { |
991 |
|
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
992 |
|
i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); |
993 |
|
} else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
994 |
|
|
995 |
|
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
996 |
|
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
997 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
998 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
999 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1000 |
|
|
1001 |
|
if ((x == 0) && (y == 0)) threshA = 512; |
1002 |
|
else { |
1003 |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
1004 |
|
if (threshA < 512) threshA = 512; |
1005 |
|
if (threshA > 1024) threshA = 1024; } |
1006 |
|
|
1007 |
VECTOR pmv[4]; |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1008 |
int32_t psad[4]; |
prevMBs + x + y * pParam->mb_width); |
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
|
VECTOR startMV; |
|
1009 |
|
|
1010 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
if (pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; |
1011 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
else |
1012 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
1013 |
|
else CheckCandidate = CheckCandidate16no4v; |
1014 |
|
|
1015 |
static int32_t threshA,threshB; |
/* main loop. checking all predictions */ |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
1016 |
|
|
1017 |
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
for (i = 1; i < 7; i++) { |
1018 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
1019 |
|
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1020 |
|
if (Data->iMinSAD[0] <= threshA) break; |
1021 |
|
} |
1022 |
|
|
1023 |
MainSearch8FuncPtr MainSearchPtr; |
if ((Data->iMinSAD[0] <= threshA) || |
1024 |
|
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1025 |
|
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
1026 |
|
inter4v = 0; |
1027 |
|
} else { |
1028 |
|
|
1029 |
/* Init variables */ |
MainSearchFunc * MainSearchPtr; |
1030 |
startMV.x = start_x; |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1031 |
startMV.y = start_y; |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1032 |
|
else MainSearchPtr = DiamondSearch; |
1033 |
|
|
1034 |
/* Get maximum range */ |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 8, iWidth, iHeight, iFcode); |
|
1035 |
|
|
1036 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
/* extended search, diamond starting in 0,0 and in prediction. |
1037 |
{ min_dx = EVEN(min_dx); |
note that this search is/might be done in halfpel positions, |
1038 |
max_dx = EVEN(max_dx); |
which makes it more different than the diamond above */ |
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
1039 |
|
|
1040 |
|
if (MotionFlags & PMV_EXTSEARCH16) { |
1041 |
|
int32_t bSAD; |
1042 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1043 |
|
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1044 |
|
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1045 |
|
if (!(MVequal(startMV, backupMV))) { |
1046 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1047 |
|
|
1048 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1049 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1050 |
|
if (bSAD < Data->iMinSAD[0]) { |
1051 |
|
Data->currentMV[0] = backupMV; |
1052 |
|
Data->iMinSAD[0] = bSAD; } |
1053 |
|
} |
1054 |
|
|
1055 |
if ((x==0) && (y==0) ) |
backupMV = Data->currentMV[0]; |
1056 |
{ |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
1057 |
threshA = 512/4; |
else startMV.x = startMV.y = 0; |
1058 |
threshB = 1024/4; |
if (!(MVequal(startMV, backupMV))) { |
1059 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1060 |
|
|
1061 |
|
(*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); |
1062 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1063 |
|
if (bSAD < Data->iMinSAD[0]) { |
1064 |
|
Data->currentMV[0] = backupMV; |
1065 |
|
Data->iMinSAD[0] = bSAD; } |
1066 |
|
} |
1067 |
} |
} |
|
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; |
|
1068 |
} |
} |
1069 |
|
|
1070 |
iFound=0; |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
1071 |
|
|
1072 |
/* Step 4: Calculate SAD around the Median prediction. |
for(i = 0; i < 5; i++) { |
1073 |
MinSAD=SAD |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1074 |
If Motion Vector equal to Previous frame motion vector |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1075 |
and MinSAD<PrevFrmSAD goto Step 10. |
} |
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1076 |
|
|
1077 |
|
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1078 |
|
|
1079 |
// Prepare for main loop |
CheckCandidate = CheckCandidate16_qpel; |
1080 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1081 |
|
pParam->width, pParam->height, Data->iFcode, 0); |
1082 |
|
|
1083 |
// if (MotionFlags & PMV_USESQUARES8) |
QuarterpelRefine(Data); |
1084 |
// MainSearchPtr = Square8_MainSearch; |
} |
|
// else |
|
1085 |
|
|
1086 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
if (inter4v) { |
1087 |
MainSearchPtr = AdvDiamond8_MainSearch; |
SearchData Data8; |
1088 |
else |
Data8.iFcode = Data->iFcode; |
1089 |
MainSearchPtr = Diamond8_MainSearch; |
Data8.lambda8 = Data->lambda8; |
1090 |
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
1091 |
|
Data8.RefQ = Data->RefQ; |
1092 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1093 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1094 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1095 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1096 |
|
} |
1097 |
|
|
1098 |
|
if (!(inter4v) || |
1099 |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1100 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1101 |
|
// INTER MODE |
1102 |
|
pMB->mode = MODE_INTER; |
1103 |
|
pMB->mvs[0] = pMB->mvs[1] |
1104 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1105 |
|
|
1106 |
*currMV = startMV; |
pMB->qmvs[0] = pMB->qmvs[1] |
1107 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1108 |
|
|
1109 |
iMinSAD = sad8( cur, |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1110 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
|
iEdgedWidth); |
|
|
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode, iQuant); |
|
1111 |
|
|
1112 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) |
if(pParam->m_quarterpel) { |
1113 |
&& ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
1114 |
{ |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
1115 |
if (MotionFlags & PMV_QUICKSTOP16) |
} |
1116 |
goto PMVfast8_Terminate_without_Refine; |
else { |
1117 |
if (MotionFlags & PMV_EARLYSTOP16) |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1118 |
goto PMVfast8_Terminate_with_Refine; |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1119 |
|
} |
1120 |
|
} else { |
1121 |
|
// INTER4V MODE; all other things are already set in Search8 |
1122 |
|
pMB->mode = MODE_INTER4V; |
1123 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1124 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1125 |
|
} |
1126 |
} |
} |
1127 |
|
|
1128 |
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
static void |
1129 |
vector of the median. |
Search8(const SearchData * const OldData, |
1130 |
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
const int x, const int y, |
1131 |
*/ |
const uint32_t MotionFlags, |
1132 |
|
const MBParam * const pParam, |
1133 |
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
MACROBLOCK * const pMB, |
1134 |
iFound=2; |
const MACROBLOCK * const pMBs, |
1135 |
|
const int block, |
1136 |
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
SearchData * const Data) |
1137 |
Otherwise select large Diamond Search. |
{ |
1138 |
*/ |
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1139 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1140 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1141 |
|
|
1142 |
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
if(pParam->m_quarterpel) { |
1143 |
iDiamondSize=1; // 1 halfpel! |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1144 |
else |
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
1145 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
1146 |
|
Data->currentQMV->y - Data->predQMV.y, |
1147 |
|
Data->iFcode); |
1148 |
|
CheckCandidate = CheckCandidate8Q; |
1149 |
|
} else { |
1150 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1151 |
|
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
1152 |
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
1153 |
|
Data->currentMV->y - Data->predMV.y, |
1154 |
|
Data->iFcode); |
1155 |
|
CheckCandidate = CheckCandidate8; |
1156 |
|
} |
1157 |
|
|
1158 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
|
iDiamondSize*=2; |
|
1159 |
|
|
1160 |
|
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1161 |
|
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1162 |
|
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1163 |
|
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1164 |
|
|
1165 |
/* |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
|
|
Also calculate (0,0) but do not subtract offset. |
|
|
Let MinSAD be the smallest SAD up to this point. |
|
|
If MV is (0,0) subtract offset. |
|
|
*/ |
|
1166 |
|
|
1167 |
// the median prediction might be even better than mv16 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1168 |
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1169 |
|
|
1170 |
if (!MVequal(pmv[0],startMV)) |
if (MotionFlags & PMV_EXTSEARCH8) { |
1171 |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1172 |
|
|
1173 |
// (0,0) if needed |
MainSearchFunc *MainSearchPtr; |
1174 |
if (!MVzero(pmv[0])) |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1175 |
if (!MVzero(startMV)) |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1176 |
CHECK_MV8_ZERO; |
else MainSearchPtr = DiamondSearch; |
1177 |
|
|
1178 |
// previous frame MV if needed |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
if (!MVzero(prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
|
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
|
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
|
1179 |
|
|
1180 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
if(*(Data->iMinSAD) < temp_sad) { |
1181 |
{ |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1182 |
if (MotionFlags & PMV_QUICKSTOP16) |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1183 |
goto PMVfast8_Terminate_without_Refine; |
} |
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
1184 |
} |
} |
1185 |
|
|
1186 |
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
1187 |
|
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1188 |
|
|
1189 |
// left neighbour, if allowed and needed |
HalfpelRefine(Data); // perform halfpel refine of current best vector |
|
if (!MVzero(pmv[1])) |
|
|
if (!MVequal(pmv[1],startMV)) |
|
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
|
|
if (!MVequal(pmv[1],pmv[0])) |
|
|
{ |
|
|
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); |
|
|
} |
|
1190 |
|
|
1191 |
// top neighbour, if allowed and needed |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1192 |
if (!MVzero(pmv[2])) |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1193 |
if (!MVequal(pmv[2],startMV)) |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1194 |
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
} |
|
if (!MVequal(pmv[2],pmv[0])) |
|
|
if (!MVequal(pmv[2],pmv[1])) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
1195 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1196 |
|
|
1197 |
// top right neighbour, if allowed and needed |
if(pParam->m_quarterpel) { |
1198 |
if (!MVzero(pmv[3])) |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1199 |
if (!MVequal(pmv[3],startMV)) |
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1200 |
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
CheckCandidate = CheckCandidate8_qpel; |
1201 |
if (!MVequal(pmv[3],pmv[0])) |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1202 |
if (!MVequal(pmv[3],pmv[1])) |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1203 |
if (!MVequal(pmv[3],pmv[2])) |
QuarterpelRefine(Data); |
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
1204 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1205 |
} |
} |
1206 |
} |
} |
1207 |
|
|
1208 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) |
if(pParam->m_quarterpel) { |
1209 |
iMinSAD -= MV8_00_BIAS; |
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
1210 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
1211 |
|
} |
1212 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
else { |
1213 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1214 |
*/ |
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
|
|
|
|
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP16) |
|
|
goto PMVfast8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto PMVfast8_Terminate_with_Refine; |
|
1215 |
} |
} |
1216 |
|
|
1217 |
/************ (Diamond Search) **************/ |
pMB->mvs[block] = *(Data->currentMV); |
1218 |
/* |
pMB->qmvs[block] = *(Data->currentQMV); |
|
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. |
|
|
*/ |
|
1219 |
|
|
1220 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1221 |
|
} |
1222 |
|
|
1223 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
/* B-frames code starts here */ |
|
iSAD = (*MainSearchPtr)(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); |
|
1224 |
|
|
1225 |
if (iSAD < iMinSAD) |
static __inline VECTOR |
1226 |
|
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1227 |
{ |
{ |
1228 |
*currMV = newMV; |
/* the stupidiest function ever */ |
1229 |
iMinSAD = iSAD; |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
1230 |
|
else return pMB->b_mvs[0]; |
1231 |
} |
} |
1232 |
|
|
1233 |
if (MotionFlags & PMV_EXTSEARCH8) |
static void __inline |
1234 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
1235 |
|
const uint32_t iWcount, |
1236 |
|
const MACROBLOCK * const pMB, |
1237 |
|
const uint32_t mode_curr) |
1238 |
{ |
{ |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
1239 |
|
|
1240 |
if (!(MVequal(pmv[0],backupMV)) ) |
// [0] is prediction |
1241 |
{ iSAD = (*MainSearchPtr)(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); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
1242 |
|
|
1243 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
{ iSAD = (*MainSearchPtr)(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); |
|
1244 |
|
|
1245 |
if (iSAD < iMinSAD) |
pmv[2] = ChoosePred(pMB, mode_curr); |
1246 |
{ |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1247 |
|
|
1248 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1249 |
By performing an optional local half-pixel search, we can refine this result even further. |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1250 |
*/ |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1251 |
|
} else pmv[3].x = pmv[3].y = 0; |
1252 |
|
|
1253 |
PMVfast8_Terminate_with_Refine: |
if (y != 0) { |
1254 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
1255 |
iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
1256 |
x, y, |
} else pmv[4].x = pmv[4].y = 0; |
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1257 |
|
|
1258 |
|
if (x != 0) { |
1259 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
1260 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1261 |
|
} else pmv[5].x = pmv[5].y = 0; |
1262 |
|
|
1263 |
PMVfast8_Terminate_without_Refine: |
if ((x != 0)&&(y != 0)) { |
1264 |
currPMV->x = currMV->x - pmv[0].x; |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1265 |
currPMV->y = currMV->y - pmv[0].y; |
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1266 |
|
} else pmv[6].x = pmv[6].y = 0; |
1267 |
|
|
1268 |
return iMinSAD; |
// more? |
1269 |
} |
} |
1270 |
|
|
1271 |
int32_t EPZSSearch16( |
|
1272 |
const uint8_t * const pRef, |
/* search backward or forward, for b-frames */ |
1273 |
|
static void |
1274 |
|
SearchBF( const uint8_t * const pRef, |
1275 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1276 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1277 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1278 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1279 |
const int x, const int y, |
const int x, const int y, |
1280 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
|
const uint32_t iQuant, |
|
1281 |
const uint32_t iFcode, |
const uint32_t iFcode, |
1282 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1283 |
const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, |
1284 |
const MACROBLOCK * const prevMBs, |
const VECTOR * const predMV, |
1285 |
VECTOR * const currMV, |
int32_t * const best_sad, |
1286 |
VECTOR * const currPMV) |
const int32_t mode_current, |
1287 |
|
SearchData * const Data) |
1288 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const uint32_t iHcount = pParam->mb_height; |
|
1289 |
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
1290 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1291 |
|
|
1292 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
int i, iDirection, mask; |
1293 |
|
VECTOR pmv[7]; |
1294 |
|
MainSearchFunc *MainSearchPtr; |
1295 |
|
*Data->iMinSAD = MV_MAX_ERROR; |
1296 |
|
Data->iFcode = iFcode; |
1297 |
|
|
1298 |
|
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1299 |
|
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
1300 |
|
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
1301 |
|
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
1302 |
|
|
1303 |
|
Data->predMV = *predMV; |
1304 |
|
|
1305 |
int32_t min_dx; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1306 |
int32_t max_dx; |
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
1307 |
|
|
1308 |
VECTOR newMV; |
pmv[0] = Data->predMV; |
1309 |
VECTOR backupMV; |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1310 |
|
|
1311 |
VECTOR pmv[4]; |
Data->currentMV->x = Data->currentMV->y = 0; |
|
int32_t psad[8]; |
|
1312 |
|
|
1313 |
static MACROBLOCK * oldMBs = NULL; |
CheckCandidate = CheckCandidate16no4v; |
|
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
|
|
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
|
|
MACROBLOCK * oldMB = NULL; |
|
1314 |
|
|
1315 |
static int32_t thresh2; |
// main loop. checking all predictions |
1316 |
int32_t bPredEq; |
for (i = 0; i < 8; i++) { |
1317 |
int32_t iMinSAD,iSAD=9999; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1318 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1319 |
|
} |
1320 |
|
|
1321 |
MainSearch16FuncPtr MainSearchPtr; |
if (MotionFlags & PMV_USESQUARES16) |
1322 |
|
MainSearchPtr = SquareSearch; |
1323 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1324 |
|
MainSearchPtr = AdvDiamondSearch; |
1325 |
|
else MainSearchPtr = DiamondSearch; |
1326 |
|
|
1327 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1328 |
|
|
1329 |
|
HalfpelRefine(Data); |
1330 |
|
|
1331 |
|
// three bits are needed to code backward mode. four for forward |
1332 |
|
// we treat the bits just like they were vector's |
1333 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16 * 2; |
1334 |
|
else *Data->iMinSAD += 3 * Data->lambda16 * 2; |
1335 |
|
|
1336 |
|
if (*Data->iMinSAD < *best_sad) { |
1337 |
|
*best_sad = *Data->iMinSAD; |
1338 |
|
pMB->mode = mode_current; |
1339 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1340 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1341 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; |
1342 |
|
else pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
1343 |
|
} |
1344 |
|
|
|
if (oldMBs == NULL) |
|
|
{ oldMBs = (MACROBLOCK*) calloc(iWcount*iHcount,sizeof(MACROBLOCK)); |
|
|
// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); |
|
1345 |
} |
} |
|
oldMB = oldMBs + x + y * iWcount; |
|
1346 |
|
|
1347 |
/* Get maximum range */ |
static int32_t |
1348 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
SearchDirect(const IMAGE * const f_Ref, |
1349 |
x, y, 16, iWidth, iHeight, iFcode); |
const uint8_t * const f_RefH, |
1350 |
|
const uint8_t * const f_RefV, |
1351 |
|
const uint8_t * const f_RefHV, |
1352 |
|
const IMAGE * const b_Ref, |
1353 |
|
const uint8_t * const b_RefH, |
1354 |
|
const uint8_t * const b_RefV, |
1355 |
|
const uint8_t * const b_RefHV, |
1356 |
|
const IMAGE * const pCur, |
1357 |
|
const int x, const int y, |
1358 |
|
const uint32_t MotionFlags, |
1359 |
|
const int32_t TRB, const int32_t TRD, |
1360 |
|
const MBParam * const pParam, |
1361 |
|
MACROBLOCK * const pMB, |
1362 |
|
const MACROBLOCK * const b_mb, |
1363 |
|
int32_t * const best_sad, |
1364 |
|
SearchData * const Data) |
1365 |
|
|
1366 |
|
{ |
1367 |
|
int32_t skip_sad; |
1368 |
|
int k; |
1369 |
|
|
1370 |
|
MainSearchFunc *MainSearchPtr; |
1371 |
|
|
1372 |
|
*Data->iMinSAD = 256*4096; |
1373 |
|
Data->referencemv = b_mb->mvs; |
1374 |
|
|
1375 |
|
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1376 |
|
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1377 |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
1378 |
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1379 |
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1380 |
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
1381 |
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
1382 |
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1383 |
|
|
1384 |
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
1385 |
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
1386 |
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
1387 |
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
1388 |
|
|
1389 |
|
for (k = 0; k < 4; k++) { |
1390 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1391 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1392 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1393 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1394 |
|
|
1395 |
|
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
1396 |
|
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
1397 |
|
|
1398 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1399 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1400 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1401 |
|
return 0; |
1402 |
|
} |
1403 |
|
if (b_mb->mode != MODE_INTER4V) { |
1404 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1405 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1406 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1407 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1408 |
|
break; |
1409 |
|
} |
1410 |
|
} |
1411 |
|
|
1412 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (b_mb->mode == MODE_INTER4V) |
1413 |
{ min_dx = EVEN(min_dx); |
CheckCandidate = CheckCandidateDirect; |
1414 |
max_dx = EVEN(max_dx); |
else CheckCandidate = CheckCandidateDirectno4v; |
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
1415 |
|
|
1416 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
(*CheckCandidate)(0, 0, 255, &k, Data); |
1417 |
|
|
1418 |
/* Step 4: Calculate SAD around the Median prediction. |
// skip decision |
1419 |
MinSAD=SAD |
if (*Data->iMinSAD - 2 * Data->lambda16 < (uint32_t)pMB->quant * SKIP_THRESH_B) { |
1420 |
If Motion Vector equal to Previous frame motion vector |
//possible skip - checking chroma. everything copied from MC |
1421 |
and MinSAD<PrevFrmSAD goto Step 10. |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1422 |
If SAD<=256 goto Step 10. |
int sum, dx, dy, b_dx, b_dy; |
|
*/ |
|
1423 |
|
|
1424 |
// Prepare for main loop |
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1425 |
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1426 |
|
|
1427 |
*currMV=pmv[0]; /* current best := median prediction */ |
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1428 |
if (!(MotionFlags & PMV_HALFPEL16)) |
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
{ |
|
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
|
} |
|
1429 |
|
|
1430 |
if (currMV->x > max_dx) |
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
1431 |
currMV->x=max_dx; |
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
|
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; |
|
1432 |
|
|
1433 |
/***************** This is predictor SET A: only median prediction ******************/ |
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
1434 |
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1435 |
|
|
1436 |
iMinSAD = sad16( cur, |
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1437 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1438 |
iEdgedWidth, MV_MAX_ERROR); |
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1439 |
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
Data->iEdgedWidth/2); |
1440 |
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1441 |
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1442 |
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1443 |
|
Data->iEdgedWidth/2); |
1444 |
|
|
1445 |
// thresh1 is fixed to 256 |
if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1446 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV, prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
pMB->mode = MODE_DIRECT_NONE_MV; |
1447 |
{ |
return *Data->iMinSAD; |
1448 |
if (MotionFlags & PMV_QUICKSTOP16) |
} |
|
goto EPZS16_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP16) |
|
|
goto EPZS16_Terminate_with_Refine; |
|
1449 |
} |
} |
1450 |
|
|
1451 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
skip_sad = *Data->iMinSAD; |
1452 |
|
|
1453 |
// previous frame MV |
// DIRECT MODE DELTA VECTOR SEARCH. |
1454 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
// This has to be made more effective, but at the moment I'm happy it's running at all |
1455 |
|
|
1456 |
// set threshhold based on Min of Prediction and SAD of collocated block |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1457 |
// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1458 |
|
else MainSearchPtr = DiamondSearch; |
1459 |
|
|
1460 |
if ((x==0) && (y==0) ) |
(*MainSearchPtr)(0, 0, Data, 255); |
|
{ |
|
|
thresh2 = 512; |
|
|
} |
|
|
else |
|
|
{ |
|
|
/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ |
|
1461 |
|
|
1462 |
thresh2 = MIN(psad[0],iSAD)*6/5 + 128; |
HalfpelRefine(Data); |
|
} |
|
1463 |
|
|
1464 |
// MV=(0,0) is often a good choice |
*Data->iMinSAD += 1 * Data->lambda16 * 2; // one bit is needed to code direct mode |
1465 |
|
*best_sad = *Data->iMinSAD; |
1466 |
|
|
1467 |
CHECK_MV16_ZERO; |
if (b_mb->mode == MODE_INTER4V) |
1468 |
|
pMB->mode = MODE_DIRECT; |
1469 |
|
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1470 |
|
|
1471 |
|
pMB->pmvs[3] = *Data->currentMV; |
1472 |
|
|
1473 |
// left neighbour, if allowed |
for (k = 0; k < 4; k++) { |
1474 |
if (x != 0) |
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1475 |
{ |
pMB->b_mvs[k].x = ((Data->currentMV->x == 0) |
1476 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
? Data->directmvB[k].x |
1477 |
{ pmv[1].x = EVEN(pmv[1].x); |
: pMB->mvs[k].x - Data->referencemv[k].x); |
1478 |
pmv[1].y = EVEN(pmv[1].y); |
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1479 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1480 |
|
? Data->directmvB[k].y |
1481 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
1482 |
|
if (b_mb->mode != MODE_INTER4V) { |
1483 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1484 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1485 |
|
break; |
1486 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1487 |
} |
} |
1488 |
|
return skip_sad; |
|
// top neighbour, if allowed |
|
|
if (y != 0) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
1489 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1490 |
|
|
1491 |
// top right neighbour, if allowed |
|
1492 |
if ((uint32_t)x != (iWcount-1)) |
static __inline void |
1493 |
|
SearchInterpolate(const uint8_t * const f_Ref, |
1494 |
|
const uint8_t * const f_RefH, |
1495 |
|
const uint8_t * const f_RefV, |
1496 |
|
const uint8_t * const f_RefHV, |
1497 |
|
const uint8_t * const b_Ref, |
1498 |
|
const uint8_t * const b_RefH, |
1499 |
|
const uint8_t * const b_RefV, |
1500 |
|
const uint8_t * const b_RefHV, |
1501 |
|
const IMAGE * const pCur, |
1502 |
|
const int x, const int y, |
1503 |
|
const uint32_t fcode, |
1504 |
|
const uint32_t bcode, |
1505 |
|
const uint32_t MotionFlags, |
1506 |
|
const MBParam * const pParam, |
1507 |
|
const VECTOR * const f_predMV, |
1508 |
|
const VECTOR * const b_predMV, |
1509 |
|
MACROBLOCK * const pMB, |
1510 |
|
int32_t * const best_sad, |
1511 |
|
SearchData * const fData) |
1512 |
|
|
1513 |
{ |
{ |
1514 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
1515 |
{ pmv[3].x = EVEN(pmv[3].x); |
const int32_t iEdgedWidth = pParam->edged_width; |
1516 |
pmv[3].y = EVEN(pmv[3].y); |
|
1517 |
} |
int iDirection, i, j; |
1518 |
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
SearchData bData; |
1519 |
|
|
1520 |
|
*(bData.iMinSAD = fData->iMinSAD) = 4096*256; |
1521 |
|
bData.Cur = fData->Cur; |
1522 |
|
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1523 |
|
bData.currentMV = fData->currentMV + 1; |
1524 |
|
bData.lambda16 = fData->lambda16; |
1525 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1526 |
|
|
1527 |
|
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
1528 |
|
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
1529 |
|
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
1530 |
|
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
1531 |
|
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
1532 |
|
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1533 |
|
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1534 |
|
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
1535 |
|
|
1536 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
1537 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
1538 |
|
|
1539 |
|
fData->currentMV[0] = fData->currentMV[3]; //forward search stored it's vector here. backward stored it in the place it's needed |
1540 |
|
get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); |
1541 |
|
get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); |
1542 |
|
|
1543 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1544 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
1545 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
1546 |
|
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1547 |
|
|
1548 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1549 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
1550 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
1551 |
|
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1552 |
|
|
1553 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1554 |
|
|
1555 |
|
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1556 |
|
|
1557 |
|
do { |
1558 |
|
iDirection = 255; |
1559 |
|
// forward MV moves |
1560 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1561 |
|
|
1562 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1563 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
1564 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
1565 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
1566 |
|
|
1567 |
|
// backward MV moves |
1568 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1569 |
|
fData->currentMV[2] = fData->currentMV[0]; |
1570 |
|
|
1571 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1572 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1573 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1574 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
1575 |
|
|
1576 |
|
} while (!(iDirection)); |
1577 |
|
|
1578 |
|
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
1579 |
|
*fData->iMinSAD += 2 * fData->lambda16 * 2; |
1580 |
|
|
1581 |
|
if (*fData->iMinSAD < *best_sad) { |
1582 |
|
*best_sad = *fData->iMinSAD; |
1583 |
|
pMB->mvs[0] = fData->currentMV[0]; |
1584 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
1585 |
|
pMB->mode = MODE_INTERPOLATE; |
1586 |
|
|
1587 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1588 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1589 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1590 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1591 |
} |
} |
1592 |
} |
} |
1593 |
|
|
|
/* Terminate if MinSAD <= T_2 |
|
|
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
|
|
*/ |
|
1594 |
|
|
1595 |
if ( (iMinSAD <= thresh2) |
void |
1596 |
|| ( MVequal(*currMV,prevMB->mvs[0]) && ((uint32_t)iMinSAD <= prevMB->sad16) ) ) |
MotionEstimationBVOP(MBParam * const pParam, |
1597 |
|
FRAMEINFO * const frame, |
1598 |
|
const int32_t time_bp, |
1599 |
|
const int32_t time_pp, |
1600 |
|
// forward (past) reference |
1601 |
|
const MACROBLOCK * const f_mbs, |
1602 |
|
const IMAGE * const f_ref, |
1603 |
|
const IMAGE * const f_refH, |
1604 |
|
const IMAGE * const f_refV, |
1605 |
|
const IMAGE * const f_refHV, |
1606 |
|
// backward (future) reference |
1607 |
|
const MACROBLOCK * const b_mbs, |
1608 |
|
const IMAGE * const b_ref, |
1609 |
|
const IMAGE * const b_refH, |
1610 |
|
const IMAGE * const b_refV, |
1611 |
|
const IMAGE * const b_refHV) |
1612 |
{ |
{ |
1613 |
if (MotionFlags & PMV_QUICKSTOP16) |
uint32_t i, j; |
1614 |
goto EPZS16_Terminate_without_Refine; |
int32_t best_sad, skip_sad; |
1615 |
if (MotionFlags & PMV_EARLYSTOP16) |
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1616 |
goto EPZS16_Terminate_with_Refine; |
static const VECTOR zeroMV={0,0}; |
1617 |
} |
|
1618 |
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1619 |
|
|
1620 |
|
const int32_t TRB = time_pp - time_bp; |
1621 |
|
const int32_t TRD = time_pp; |
1622 |
|
|
1623 |
|
// some pre-inintialized data for the rest of the search |
1624 |
|
|
1625 |
|
SearchData Data; |
1626 |
|
int32_t iMinSAD; |
1627 |
|
VECTOR currentMV[3]; |
1628 |
|
Data.iEdgedWidth = pParam->edged_width; |
1629 |
|
Data.currentMV = currentMV; |
1630 |
|
Data.iMinSAD = &iMinSAD; |
1631 |
|
Data.lambda16 = lambda_vec16[frame->quant]; |
1632 |
|
|
1633 |
/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ |
// note: i==horizontal, j==vertical |
1634 |
|
|
1635 |
backupMV = prevMB->mvs[0]; // collocated MV |
for (j = 0; j < pParam->mb_height; j++) { |
|
backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X |
|
|
backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y |
|
1636 |
|
|
1637 |
CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
1638 |
|
|
1639 |
// left neighbour |
for (i = 0; i < pParam->mb_width; i++) { |
1640 |
if (x != 0) |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1641 |
CHECK_MV16_CANDIDATE((prevMB-1)->mvs[0].x,(prevMB-1)->mvs[0].y); |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1642 |
|
|
1643 |
// top neighbour |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
1644 |
if (y != 0) |
if ( (b_mb->mode == MODE_NOT_CODED) ) { |
1645 |
CHECK_MV16_CANDIDATE((prevMB-iWcount)->mvs[0].x,(prevMB-iWcount)->mvs[0].y); |
pMB->mode = MODE_NOT_CODED; |
1646 |
|
continue; |
1647 |
|
} |
1648 |
|
|
1649 |
// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1650 |
|
pMB->quant = frame->quant; |
1651 |
|
/* direct search comes first, because it (1) checks for SKIP-mode |
1652 |
|
and (2) sets very good predictions for forward and backward search */ |
1653 |
|
|
1654 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1655 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1656 |
|
&frame->image, |
1657 |
|
i, j, |
1658 |
|
frame->motion_flags, |
1659 |
|
TRB, TRD, |
1660 |
|
pParam, |
1661 |
|
pMB, b_mb, |
1662 |
|
&best_sad, |
1663 |
|
&Data); |
1664 |
|
|
1665 |
if ((uint32_t)x != iWcount-1) |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
|
CHECK_MV16_CANDIDATE((prevMB+1)->mvs[0].x,(prevMB+1)->mvs[0].y); |
|
1666 |
|
|
1667 |
// bottom neighbour, dito |
// best_sad = 256*4096; //uncomment to disable Directsearch. |
1668 |
if ((uint32_t)y != iHcount-1) |
// To disable any other mode, just comment the function call |
|
CHECK_MV16_CANDIDATE((prevMB+iWcount)->mvs[0].x,(prevMB+iWcount)->mvs[0].y); |
|
1669 |
|
|
1670 |
/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ |
// forward search |
1671 |
if (iMinSAD <= thresh2) |
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1672 |
{ |
&frame->image, i, j, |
1673 |
if (MotionFlags & PMV_QUICKSTOP16) |
frame->motion_flags, |
1674 |
goto EPZS16_Terminate_without_Refine; |
frame->fcode, pParam, |
1675 |
if (MotionFlags & PMV_EARLYSTOP16) |
pMB, &f_predMV, &best_sad, |
1676 |
goto EPZS16_Terminate_with_Refine; |
MODE_FORWARD, &Data); |
|
} |
|
1677 |
|
|
1678 |
/************ (if Diamond Search) **************/ |
// backward search |
1679 |
|
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1680 |
|
&frame->image, i, j, |
1681 |
|
frame->motion_flags, |
1682 |
|
frame->bcode, pParam, |
1683 |
|
pMB, &b_predMV, &best_sad, |
1684 |
|
MODE_BACKWARD, &Data); |
1685 |
|
|
1686 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
// interpolate search comes last, because it uses data from forward and backward as prediction |
1687 |
|
|
1688 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1689 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1690 |
|
&frame->image, |
1691 |
|
i, j, |
1692 |
|
frame->fcode, frame->bcode, |
1693 |
|
frame->motion_flags, |
1694 |
|
pParam, |
1695 |
|
&f_predMV, &b_predMV, |
1696 |
|
pMB, &best_sad, |
1697 |
|
&Data); |
1698 |
|
|
1699 |
|
switch (pMB->mode) { |
1700 |
|
case MODE_FORWARD: |
1701 |
|
f_count++; |
1702 |
|
f_predMV = pMB->mvs[0]; |
1703 |
|
break; |
1704 |
|
case MODE_BACKWARD: |
1705 |
|
b_count++; |
1706 |
|
b_predMV = pMB->b_mvs[0]; |
1707 |
|
break; |
1708 |
|
case MODE_INTERPOLATE: |
1709 |
|
i_count++; |
1710 |
|
f_predMV = pMB->mvs[0]; |
1711 |
|
b_predMV = pMB->b_mvs[0]; |
1712 |
|
break; |
1713 |
|
case MODE_DIRECT: |
1714 |
|
case MODE_DIRECT_NO4V: |
1715 |
|
d_count++; |
1716 |
|
break; |
1717 |
|
default: |
1718 |
|
break; |
1719 |
|
} |
1720 |
|
} |
1721 |
|
} |
1722 |
|
|
1723 |
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
1724 |
x, y, |
// f_count,b_count,i_count,d_count,n_count); |
|
currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
2, iFcode, iQuant, 0); |
|
1725 |
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
1726 |
} |
} |
1727 |
|
|
1728 |
|
/* Hinted ME starts here */ |
1729 |
|
|
1730 |
if (MotionFlags & PMV_EXTSEARCH16) |
static void |
1731 |
|
Search8hinted(const SearchData * const OldData, |
1732 |
|
const int x, const int y, |
1733 |
|
const uint32_t MotionFlags, |
1734 |
|
const MBParam * const pParam, |
1735 |
|
MACROBLOCK * const pMB, |
1736 |
|
const MACROBLOCK * const pMBs, |
1737 |
|
const int block, |
1738 |
|
SearchData * const Data) |
1739 |
{ |
{ |
1740 |
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
int32_t temp_sad; |
1741 |
|
MainSearchFunc *MainSearchPtr; |
1742 |
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1743 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1744 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1745 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1746 |
|
|
1747 |
if (!(MVequal(pmv[0],backupMV)) ) |
if(pParam->m_quarterpel) { |
1748 |
{ |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1749 |
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
1750 |
x, y, |
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
1751 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
Data->currentQMV->y - Data->predQMV.y, |
1752 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
Data->iFcode); |
1753 |
|
CheckCandidate = CheckCandidate8Q; |
1754 |
|
} else { |
1755 |
|
if (block != 0) *(Data->iMinSAD) += Data->lambda8 * |
1756 |
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
1757 |
|
Data->currentMV->y - Data->predMV.y, |
1758 |
|
Data->iFcode); |
1759 |
|
CheckCandidate = CheckCandidate8; |
1760 |
} |
} |
1761 |
|
|
1762 |
if (iSAD < iMinSAD) |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1763 |
{ |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1764 |
*currMV = newMV; |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1765 |
iMinSAD = iSAD; |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
} |
|
1766 |
|
|
1767 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
|
{ |
|
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); |
|
1768 |
|
|
1769 |
if (iSAD < iMinSAD) |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1770 |
{ |
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
|
} |
|
|
} |
|
|
} |
|
1771 |
|
|
1772 |
/*************** Choose best MV found **************/ |
temp_sad = *(Data->iMinSAD); // store current MinSAD |
1773 |
|
|
1774 |
EPZS16_Terminate_with_Refine: |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1775 |
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1776 |
iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
else MainSearchPtr = DiamondSearch; |
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1777 |
|
|
1778 |
EPZS16_Terminate_without_Refine: |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1779 |
|
|
1780 |
*oldMB = *prevMB; |
if(*(Data->iMinSAD) < temp_sad) { |
1781 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1782 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1783 |
|
} |
1784 |
|
|
1785 |
|
if (MotionFlags & PMV_HALFPELREFINE8) { |
1786 |
|
temp_sad = *(Data->iMinSAD); // store current MinSAD |
1787 |
|
|
1788 |
|
HalfpelRefine(Data); // perform halfpel refine of current best vector |
1789 |
|
|
1790 |
|
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1791 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1792 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1793 |
|
} |
1794 |
|
} |
1795 |
|
|
1796 |
currPMV->x = currMV->x - pmv[0].x; |
if(pParam->m_quarterpel) { |
1797 |
currPMV->y = currMV->y - pmv[0].y; |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1798 |
return iMinSAD; |
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1799 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1800 |
|
pParam->width, pParam->height, Data->iFcode, 0); |
1801 |
|
CheckCandidate = CheckCandidate8_qpel; |
1802 |
|
QuarterpelRefine(Data); |
1803 |
} |
} |
1804 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
1805 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
1806 |
|
} else { |
1807 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1808 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1809 |
|
} |
1810 |
|
|
1811 |
|
pMB->mvs[block] = *(Data->currentMV); |
1812 |
|
pMB->qmvs[block] = *(Data->currentQMV); |
1813 |
|
|
1814 |
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1815 |
|
} |
1816 |
|
|
1817 |
int32_t EPZSSearch8( |
static void |
1818 |
const uint8_t * const pRef, |
SearchPhinted ( const uint8_t * const pRef, |
1819 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1820 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
1821 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
1822 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1823 |
const int x, const int y, |
const int x, |
1824 |
const int start_x, const int start_y, |
const int y, |
1825 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1826 |
const uint32_t iQuant, |
const uint32_t iQuant, |
|
const uint32_t iFcode, |
|
1827 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1828 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1829 |
const MACROBLOCK * const prevMBs, |
int inter4v, |
1830 |
VECTOR * const currMV, |
MACROBLOCK * const pMB, |
1831 |
VECTOR * const currPMV) |
SearchData * const Data) |
1832 |
{ |
{ |
|
/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ |
|
1833 |
|
|
|
const uint32_t iWcount = pParam->mb_width; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
1834 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1835 |
|
|
1836 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
int i, t; |
1837 |
|
MainSearchFunc * MainSearchPtr; |
|
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]; |
|
1838 |
|
|
1839 |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1840 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1841 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
1842 |
|
|
1843 |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1844 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
1845 |
|
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
1846 |
|
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
1847 |
|
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
1848 |
|
Data->lambda16 = lambda_vec16[iQuant]; |
1849 |
|
Data->lambda8 = lambda_vec8[iQuant]; |
1850 |
|
|
1851 |
int32_t bPredEq; |
if (!(MotionFlags & PMV_HALFPEL16)) { |
1852 |
int32_t iMinSAD,iSAD=9999; |
Data->min_dx = EVEN(Data->min_dx); |
1853 |
|
Data->max_dx = EVEN(Data->max_dx); |
1854 |
MainSearch8FuncPtr MainSearchPtr; |
Data->min_dy = EVEN(Data->min_dy); |
1855 |
|
Data->max_dy = EVEN(Data->max_dy); |
1856 |
/* Get maximum range */ |
} |
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
|
|
x, y, 8, iWidth, iHeight, iFcode); |
|
1857 |
|
|
1858 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
1859 |
|
|
1860 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
{ 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; */ |
|
1861 |
|
|
1862 |
bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); |
if(pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; |
1863 |
|
else |
1864 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
1865 |
|
else CheckCandidate = CheckCandidate16no4v; |
1866 |
|
|
1867 |
|
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1868 |
|
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
1869 |
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
1870 |
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
1871 |
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
1872 |
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
1873 |
|
|
1874 |
|
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
1875 |
|
|
1876 |
|
if (pMB->mode == MODE_INTER4V) |
1877 |
|
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
1878 |
|
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
1879 |
|
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
1880 |
|
if (!(make_mask(pMB->mvs, i))) |
1881 |
|
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
1882 |
|
} |
1883 |
|
|
1884 |
/* Step 4: Calculate SAD around the Median prediction. |
if (MotionFlags & PMV_USESQUARES16) |
1885 |
MinSAD=SAD |
MainSearchPtr = SquareSearch; |
1886 |
If Motion Vector equal to Previous frame motion vector |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1887 |
and MinSAD<PrevFrmSAD goto Step 10. |
MainSearchPtr = AdvDiamondSearch; |
1888 |
If SAD<=256 goto Step 10. |
else MainSearchPtr = DiamondSearch; |
1889 |
*/ |
|
1890 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1891 |
|
|
1892 |
|
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
1893 |
|
|
1894 |
|
for(i = 0; i < 5; i++) { |
1895 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1896 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1897 |
|
} |
1898 |
|
|
1899 |
|
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
1900 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1901 |
|
pParam->width, pParam->height, Data->iFcode, 0); |
1902 |
|
CheckCandidate = CheckCandidate16_qpel; |
1903 |
|
QuarterpelRefine(Data); |
1904 |
|
} |
1905 |
|
|
1906 |
|
if (inter4v) { |
1907 |
|
SearchData Data8; |
1908 |
|
Data8.iFcode = Data->iFcode; |
1909 |
|
Data8.lambda8 = Data->lambda8; |
1910 |
|
Data8.iEdgedWidth = Data->iEdgedWidth; |
1911 |
|
Data8.RefQ = Data->RefQ; |
1912 |
|
Search8hinted(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1913 |
|
Search8hinted(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1914 |
|
Search8hinted(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1915 |
|
Search8hinted(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1916 |
|
} |
1917 |
|
|
1918 |
|
if (!(inter4v) || |
1919 |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
1920 |
|
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1921 |
|
// INTER MODE |
1922 |
|
pMB->mode = MODE_INTER; |
1923 |
|
pMB->mvs[0] = pMB->mvs[1] |
1924 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1925 |
|
|
1926 |
// Prepare for main loop |
pMB->qmvs[0] = pMB->qmvs[1] |
1927 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1928 |
|
|
1929 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1930 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1931 |
|
|
1932 |
if (!(MotionFlags & PMV_HALFPEL8)) |
if(pParam->m_quarterpel) { |
1933 |
{ |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
1934 |
currMV->x = EVEN(currMV->x); |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
1935 |
currMV->y = EVEN(currMV->y); |
} |
1936 |
|
else { |
1937 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1938 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1939 |
|
} |
1940 |
|
} else { |
1941 |
|
// INTER4V MODE; all other things are already set in Search8hinted |
1942 |
|
pMB->mode = MODE_INTER4V; |
1943 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
1944 |
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
1945 |
} |
} |
1946 |
|
|
1947 |
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; |
|
1948 |
|
|
1949 |
/***************** This is predictor SET A: only median prediction ******************/ |
void |
1950 |
|
MotionEstimationHinted( MBParam * const pParam, |
1951 |
|
FRAMEINFO * const current, |
1952 |
|
FRAMEINFO * const reference, |
1953 |
|
const IMAGE * const pRefH, |
1954 |
|
const IMAGE * const pRefV, |
1955 |
|
const IMAGE * const pRefHV) |
1956 |
|
{ |
1957 |
|
MACROBLOCK *const pMBs = current->mbs; |
1958 |
|
const IMAGE *const pCurrent = ¤t->image; |
1959 |
|
const IMAGE *const pRef = &reference->image; |
1960 |
|
|
1961 |
|
uint32_t x, y; |
1962 |
|
uint8_t * qimage; |
1963 |
|
int32_t temp[5], quant = current->quant; |
1964 |
|
int32_t iMinSAD[5]; |
1965 |
|
VECTOR currentMV[5], currentQMV[5]; |
1966 |
|
SearchData Data; |
1967 |
|
Data.iEdgedWidth = pParam->edged_width; |
1968 |
|
Data.currentMV = currentMV; |
1969 |
|
Data.currentQMV = currentQMV; |
1970 |
|
Data.iMinSAD = iMinSAD; |
1971 |
|
Data.temp = temp; |
1972 |
|
Data.iFcode = current->fcode; |
1973 |
|
Data.rounding = pParam->m_rounding_type; |
1974 |
|
|
1975 |
iMinSAD = sad8( cur, |
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
1976 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
return; // allocate some mem for qpel interpolated blocks |
1977 |
iEdgedWidth); |
// somehow this is dirty since I think we shouldn't use malloc outside |
1978 |
iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode, iQuant); |
// encoder_create() - so please fix me! |
1979 |
|
|
1980 |
|
Data.RefQ = qimage; |
1981 |
|
|
1982 |
// thresh1 is fixed to 256 |
if (sadInit) (*sadInit) (); |
|
if (iMinSAD < 256/4 ) |
|
|
{ |
|
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
1983 |
|
|
1984 |
/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ |
for (y = 0; y < pParam->mb_height; y++) { |
1985 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1986 |
|
|
1987 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1988 |
|
|
1989 |
// MV=(0,0) is often a good choice |
//intra mode is copied from the first pass. At least for the time being |
1990 |
CHECK_MV8_ZERO; |
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
1991 |
|
|
|
// previous frame MV |
|
|
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
|
1992 |
|
|
1993 |
// left neighbour, if allowed |
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1994 |
if (psad[1] != MV_MAX_ERROR) |
pMB->dquant = NO_CHANGE; |
1995 |
{ |
pMB->quant = current->quant; } |
1996 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
else { |
1997 |
{ pmv[1].x = EVEN(pmv[1].x); |
if (pMB->dquant != NO_CHANGE) { |
1998 |
pmv[1].y = EVEN(pmv[1].y); |
quant += DQtab[pMB->dquant]; |
1999 |
|
if (quant > 31) quant = 31; |
2000 |
|
else if (quant < 1) quant = 1; |
2001 |
} |
} |
2002 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
pMB->quant = quant; |
2003 |
} |
} |
2004 |
|
|
2005 |
// top neighbour, if allowed |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
2006 |
if (psad[2] != MV_MAX_ERROR) |
y, current->motion_flags, pMB->quant, |
2007 |
{ |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
2008 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
&Data); |
|
{ pmv[2].x = EVEN(pmv[2].x); |
|
|
pmv[2].y = EVEN(pmv[2].y); |
|
|
} |
|
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
2009 |
|
|
|
// top right neighbour, if allowed |
|
|
if (psad[3] != MV_MAX_ERROR) |
|
|
{ |
|
|
if (!(MotionFlags & PMV_HALFPEL8 )) |
|
|
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
2010 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
2011 |
} |
} |
2012 |
|
free(qimage); |
2013 |
} |
} |
2014 |
|
|
2015 |
/* // this bias is zero anyway, at the moment! |
static __inline int |
2016 |
|
MEanalyzeMB ( const uint8_t * const pRef, |
2017 |
if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) |
const uint8_t * const pCur, |
2018 |
iMinSAD -= MV8_00_BIAS; |
const int x, |
2019 |
|
const int y, |
2020 |
*/ |
const MBParam * const pParam, |
2021 |
|
const MACROBLOCK * const pMBs, |
2022 |
/* Terminate if MinSAD <= T_2 |
MACROBLOCK * const pMB, |
2023 |
Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] |
SearchData * const Data) |
|
*/ |
|
|
|
|
|
if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ |
|
2024 |
{ |
{ |
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto EPZS8_Terminate_without_Refine; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto EPZS8_Terminate_with_Refine; |
|
|
} |
|
|
|
|
|
/************ (Diamond Search) **************/ |
|
2025 |
|
|
2026 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
int i = 255, mask; |
2027 |
|
VECTOR pmv[3]; |
2028 |
|
|
2029 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8)) |
*(Data->iMinSAD) = MV_MAX_ERROR; |
2030 |
iDiamondSize *= 2; |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2031 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2032 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
2033 |
|
|
2034 |
/* default: use best prediction as starting point for one call of EPZS_MainSearch */ |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
2035 |
|
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
2036 |
|
|
2037 |
/* // there is no EPZS^2 for inter4v at the moment |
CheckCandidate = CheckCandidate16no4vI; |
|
|
|
|
if (MotionFlags & PMV_USESQUARES8) |
|
|
MainSearchPtr = Square8_MainSearch; |
|
|
else |
|
|
*/ |
|
2038 |
|
|
2039 |
// if (MotionFlags & PMV_USESQUARES8) |
pmv[1].x = EVEN(pMB->mvs[0].x); |
2040 |
// MainSearchPtr = Square8_MainSearch; |
pmv[1].y = EVEN(pMB->mvs[0].y); |
2041 |
// else |
pmv[0].x = EVEN(Data->predMV.x); |
2042 |
|
pmv[0].y = EVEN(Data->predMV.y); |
2043 |
|
pmv[2].x = pmv[2].y = 0; |
2044 |
|
|
2045 |
if (MotionFlags & PMV_ADVANCEDDIAMOND8) |
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
2046 |
MainSearchPtr = AdvDiamond8_MainSearch; |
if (!(mask = make_mask(pmv, 1))) |
2047 |
else |
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
2048 |
MainSearchPtr = Diamond8_MainSearch; |
if (!(mask = make_mask(pmv, 2))) |
2049 |
|
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
2050 |
|
|
2051 |
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, |
|
|
iDiamondSize, iFcode, iQuant, 0); |
|
2052 |
|
|
2053 |
|
pMB->mvs[0] = *Data->currentMV; |
2054 |
|
|
2055 |
if (iSAD < iMinSAD) |
return *(Data->iMinSAD); |
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
2056 |
} |
} |
2057 |
|
|
2058 |
if (MotionFlags & PMV_EXTSEARCH8) |
#define INTRA_THRESH 1350 |
2059 |
{ |
#define INTER_THRESH 900 |
|
/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ |
|
2060 |
|
|
2061 |
if (!(MVequal(pmv[0],backupMV)) ) |
int |
2062 |
|
MEanalysis( const IMAGE * const pRef, |
2063 |
|
const IMAGE * const pCurrent, |
2064 |
|
MBParam * const pParam, |
2065 |
|
MACROBLOCK * const pMBs, |
2066 |
|
const uint32_t iFcode) |
2067 |
{ |
{ |
2068 |
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
uint32_t x, y, intra = 0; |
2069 |
x, y, |
int sSAD = 0; |
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
2070 |
|
|
2071 |
if (iSAD < iMinSAD) |
VECTOR currentMV; |
2072 |
{ |
int32_t iMinSAD; |
2073 |
*currMV = newMV; |
SearchData Data; |
2074 |
iMinSAD = iSAD; |
Data.iEdgedWidth = pParam->edged_width; |
2075 |
} |
Data.currentMV = ¤tMV; |
2076 |
} |
Data.iMinSAD = &iMinSAD; |
2077 |
|
Data.iFcode = iFcode; |
2078 |
|
|
2079 |
|
if (sadInit) (*sadInit) (); |
2080 |
|
|
2081 |
|
for (y = 1; y < pParam->mb_height-1; y++) { |
2082 |
|
for (x = 1; x < pParam->mb_width-1; x++) { |
2083 |
|
int sad, dev; |
2084 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
2085 |
|
|
2086 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
2087 |
{ |
pParam, pMBs, pMB, &Data); |
|
iSAD = (*MainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); |
|
2088 |
|
|
2089 |
if (iSAD < iMinSAD) |
if (sad > INTRA_THRESH) { |
2090 |
{ |
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
2091 |
*currMV = newMV; |
pParam->edged_width); |
2092 |
iMinSAD = iSAD; |
if (dev + INTRA_THRESH < sad) intra++; |
2093 |
|
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
2094 |
} |
} |
2095 |
|
sSAD += sad; |
2096 |
} |
} |
2097 |
} |
} |
2098 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2099 |
|
if (sSAD > INTER_THRESH ) return 1; //P frame |
2100 |
|
emms(); |
2101 |
|
return 0; // B frame |
2102 |
|
|
|
/*************** 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; |
|
2103 |
} |
} |
2104 |
|
|
2105 |
|
int |
2106 |
|
FindFcode( const MBParam * const pParam, |
2107 |
|
const FRAMEINFO * const current) |
2108 |
|
{ |
2109 |
|
uint32_t x, y; |
2110 |
|
int max = 0, min = 0, i; |
2111 |
|
|
2112 |
|
for (y = 0; y < pParam->mb_height; y++) { |
2113 |
|
for (x = 0; x < pParam->mb_width; x++) { |
2114 |
|
|
2115 |
|
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
2116 |
|
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
2117 |
|
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
2118 |
|
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
2119 |
|
|
2120 |
|
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
2121 |
|
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
2122 |
|
} |
2123 |
|
} |
2124 |
|
} |
2125 |
|
|
2126 |
/* *********************************************************** |
min = -min; |
2127 |
bvop motion estimation |
max += 1; |
2128 |
// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) |
if (min > max) max = min; |
2129 |
***************************************************************/ |
if (pParam->m_quarterpel) max *= 2; |
|
|
|
|
|
|
|
void MotionEstimationBVOP( |
|
|
MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
|
|
|
|
// forward (past) reference |
|
|
const MACROBLOCK * const f_mbs, |
|
|
const IMAGE * const f_ref, |
|
|
const IMAGE * const f_refH, |
|
|
const IMAGE * const f_refV, |
|
|
const IMAGE * const f_refHV, |
|
|
// backward (future) reference |
|
|
const MACROBLOCK * const b_mbs, |
|
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
|
{ |
|
|
const uint32_t mb_width = pParam->mb_width; |
|
|
const uint32_t mb_height = pParam->mb_height; |
|
|
const int32_t edged_width = pParam->edged_width; |
|
2130 |
|
|
2131 |
uint32_t i,j; |
for (i = 1; (max > 32 << (i - 1)); i++); |
2132 |
|
return i; |
2133 |
|
} |
2134 |
|
|
|
int32_t f_sad16; |
|
|
int32_t b_sad16; |
|
|
int32_t i_sad16; |
|
|
int32_t d_sad16; |
|
|
int32_t best_sad; |
|
2135 |
|
|
2136 |
VECTOR pmv_dontcare; |
/* Global Motion Estimation, (C) 2002 by sysKin */ |
2137 |
|
/* calculate global translation from local (halfpel) motion field */ |
2138 |
|
|
2139 |
// note: i==horizontal, j==vertical |
static VECTOR |
2140 |
for (j = 0; j < mb_height; j++) |
GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) |
2141 |
{ |
{ |
|
for (i = 0; i < mb_width; i++) |
|
|
{ |
|
|
MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; |
|
|
|
|
|
if (b_mb->mode == MODE_INTER |
|
|
&& b_mb->cbp == 0 |
|
|
&& b_mb->mvs[0].x == 0 |
|
|
&& b_mb->mvs[0].y == 0) |
|
|
{ |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->mvs[0].x = 0; |
|
|
mb->mvs[0].y = 0; |
|
|
mb->b_mvs[0].x = 0; |
|
|
mb->b_mvs[0].y = 0; |
|
|
continue; |
|
|
} |
|
|
|
|
2142 |
|
|
2143 |
// forward search |
int count, bestcount = 0; |
2144 |
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
int x, y; |
2145 |
&frame->image, |
VECTOR gmc; |
2146 |
i, j, |
int step, min_x, max_x, min_y, max_y; |
2147 |
frame->motion_flags, frame->quant, frame->fcode, |
uint32_t mx, my; |
|
pParam, |
|
|
f_mbs, f_mbs /* todo */, |
|
|
&mb->mvs[0], &pmv_dontcare); // ignore pmv |
|
2148 |
|
|
2149 |
// backward search |
min_x = min_y = -32<<iFcode; |
2150 |
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
max_x = max_y = 32<<iFcode; |
|
&frame->image, |
|
|
i, j, |
|
|
frame->motion_flags, frame->quant, frame->bcode, |
|
|
pParam, |
|
|
b_mbs, b_mbs, /* todo */ |
|
|
&mb->b_mvs[0], &pmv_dontcare); // ignore pmv |
|
2151 |
|
|
2152 |
// interpolate search (simple, but effective) |
for (step = 32; step >= 2; step /= 2) { |
2153 |
i_sad16 = sad16bi_c( |
bestcount = 0; |
2154 |
frame->image.y + i*16 + j*16*edged_width, |
for (y = min_y; y <= max_y; y += step) |
2155 |
get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
for (x = min_x ; x <= max_x; x += step) { |
2156 |
i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), |
count = 0; |
2157 |
get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
//for all macroblocks |
2158 |
i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), |
for (my = 1; my < pParam->mb_height-1; my++) |
2159 |
edged_width); |
for (mx = 1; mx < pParam->mb_width-1; mx++) { |
2160 |
|
const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; |
2161 |
// TODO: direct search |
VECTOR mv; |
|
// predictor + range of [-32,32] |
|
|
d_sad16 = 65535; |
|
2162 |
|
|
2163 |
|
if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) |
2164 |
|
continue; |
2165 |
|
|
2166 |
if (f_sad16 < b_sad16) |
mv = pMB->mvs[0]; |
2167 |
{ |
if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ |
2168 |
best_sad = f_sad16; |
count++; |
|
mb->mode = MODE_FORWARD; |
|
2169 |
} |
} |
2170 |
else |
if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } |
|
{ |
|
|
best_sad = b_sad16; |
|
|
mb->mode = MODE_BACKWARD; |
|
2171 |
} |
} |
2172 |
|
min_x = gmc.x - step; |
2173 |
|
max_x = gmc.x + step; |
2174 |
|
min_y = gmc.y - step; |
2175 |
|
max_y = gmc.y + step; |
2176 |
|
|
|
if (i_sad16 < best_sad) |
|
|
{ |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
2177 |
} |
} |
2178 |
|
if (pParam->m_quarterpel) { |
2179 |
if (d_sad16 < best_sad) |
gmc.x *= 2; |
2180 |
{ |
gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ |
|
best_sad = d_sad16; |
|
|
mb->mode = MODE_DIRECT; |
|
2181 |
} |
} |
2182 |
|
|
2183 |
} |
if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) |
2184 |
} |
gmc.x = gmc.y = 0; //no camara pan, no GMC |
2185 |
|
|
2186 |
|
return gmc; |
2187 |
} |
} |