1 |
/************************************************************************** |
/************************************************************************** |
2 |
* |
* |
3 |
* Modifications: |
* XVID MPEG-4 VIDEO CODEC |
4 |
|
* motion estimation |
5 |
* |
* |
6 |
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
* This program is an implementation of a part of one or more MPEG-4 |
7 |
* PMVFast_Refine to support multiple searches with different start points |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
8 |
* 07.01.2002 uv-block-based interpolation |
* to use this software module in hardware or software products are |
9 |
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
* advised that its use may infringe existing patents or copyrights, and |
10 |
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
* any such use would be at such party's own risk. The original |
11 |
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
* developer of this software module and his/her company, and subsequent |
12 |
* added new quality mode = 6 which performs halfpel refinement |
* editors and their companies, will have no liability for use of this |
13 |
* filesize difference between quality 5 and 6 is smaller than 1% |
* software or modifications or derivatives thereof. |
|
* (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> |
33 |
|
#include <stdlib.h> |
34 |
|
|
35 |
#include "../encoder.h" |
#include "../encoder.h" |
36 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.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" |
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 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
#define GET_REFERENCE(X, Y, REF) { \ |
56 |
/* nb = vop pixels * 2^(bpp-8) */ |
switch ( ((X&1)<<1) + (Y&1) ) \ |
57 |
#define MV16_00_BIAS (128+1) |
{ \ |
58 |
|
case 0 : REF = data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
59 |
|
case 1 : REF = data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
60 |
|
case 2 : REF = data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ |
61 |
|
default : REF = data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ |
62 |
|
} \ |
63 |
|
} |
64 |
|
|
65 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
#define iDiamondSize 2 |
|
#define 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 |
|
if (y == 0) yb = 1; |
82 |
|
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 |
|
|
|
// fast ((A)/2)*2 |
|
|
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
|
92 |
|
|
93 |
|
/* CHECK_CANDIATE FUNCTIONS START */ |
94 |
|
|
95 |
#define MIN(X, Y) ((X)<(Y)?(X):(Y)) |
static void |
96 |
#define MAX(X, Y) ((X)>(Y)?(X):(Y)) |
CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
97 |
#define ABS(X) (((X)>0)?(X):-(X)) |
{ |
98 |
#define SIGN(X) (((X)>0)?1:-1) |
int32_t * const sad = data->temp; |
99 |
|
int t; |
100 |
|
const uint8_t * Reference; |
101 |
|
|
102 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
103 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
104 |
|
|
105 |
int32_t PMVfastSearch8( |
switch ( ((x&1)<<1) + (y&1) ) { |
106 |
const uint8_t * const pRef, |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
107 |
const uint8_t * const pRefH, |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
108 |
const uint8_t * const pRefV, |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
109 |
const uint8_t * const pRefHV, |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
110 |
const IMAGE * const pCur, |
} |
|
const int x, const int y, |
|
|
const int start_x, int start_y, |
|
|
const uint32_t iQuality, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
111 |
|
|
112 |
int32_t PMVfastSearch16( |
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t iQuality, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV); |
|
113 |
|
|
114 |
|
t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
115 |
|
data->temp[0] += lambda_vec16[data->iQuant] * t; |
116 |
|
data->temp[1] += lambda_vec8[data->iQuant] * t; |
117 |
|
|
118 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
119 |
|
data->iMinSAD[0] = data->temp[0]; |
120 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
121 |
|
*dir = Direction; } |
122 |
|
|
123 |
/* |
if (data->temp[1] < data->iMinSAD[1]) { |
124 |
* diamond search stuff |
data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
125 |
* keep the the sequence in circular order (so optimization works) |
if (data->temp[2] < data->iMinSAD[2]) { |
126 |
*/ |
data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
127 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
128 |
|
data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
129 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
130 |
|
data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
131 |
|
|
132 |
typedef struct |
} |
133 |
|
|
134 |
|
static void |
135 |
|
CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
136 |
|
{ |
137 |
|
int32_t sad; |
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 |
|
switch ( ((x&1)<<1) + (y&1) ) |
144 |
{ |
{ |
145 |
int32_t dx; |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
146 |
int32_t dy; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
147 |
|
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
148 |
|
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
149 |
|
} |
150 |
|
|
151 |
|
sad = lambda_vec16[data->iQuant] * |
152 |
|
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
153 |
|
sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
154 |
|
|
155 |
|
if (sad < *(data->iMinSAD)) { |
156 |
|
*(data->iMinSAD) = sad; |
157 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
158 |
|
*dir = Direction; } |
159 |
} |
} |
|
DPOINT; |
|
160 |
|
|
161 |
|
static void |
162 |
|
CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
163 |
|
|
164 |
static const DPOINT diamond_small[4] = |
// CheckCandidate16 variant which expects x and y in quarter pixel resolution |
165 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
166 |
|
// around currentMV! |
167 |
{ |
{ |
168 |
{0, 1}, {1, 0}, {0, -1}, {-1, 0} |
int32_t * const sad = data->temp; |
169 |
}; |
int t; |
170 |
|
uint8_t * Reference = (uint8_t *) data->RefQ; |
171 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
172 |
|
VECTOR halfpelMV = *(data->currentMV); |
173 |
|
|
174 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
175 |
|
uint32_t rounding = data->rounding; |
176 |
|
|
177 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
178 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
179 |
|
|
180 |
static const DPOINT diamond_large[8] = |
switch( ((x&1)<<1) + (y&1) ) |
181 |
{ |
{ |
182 |
{0, 2}, {1, 1}, {2, 0}, {1, -1}, {0, -2}, {-1, -1}, {-2, 0}, {-1, 1} |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
183 |
}; |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); |
184 |
|
break; |
185 |
|
|
186 |
|
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
187 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
188 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
189 |
|
|
190 |
// mv.length table |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
191 |
static const uint32_t mvtab[33] = { |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
192 |
1, 2, 3, 4, 6, 7, 7, 7, |
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
193 |
9, 9, 9, 10, 10, 10, 10, 10, |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
194 |
10, 10, 10, 10, 10, 10, 10, 10, |
break; |
|
10, 11, 11, 11, 11, 11, 11, 12, 12 |
|
|
}; |
|
195 |
|
|
196 |
|
case 2: // x qpel, y halfpel - left or right during qpel refinement |
197 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
198 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
199 |
|
|
200 |
static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
201 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
202 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
203 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
204 |
|
break; |
205 |
|
|
206 |
|
default: // x and y in qpel resolution - the "corners" (top left/right and |
207 |
|
// bottom left/right) during qpel refinement |
208 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
209 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
210 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
211 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
212 |
|
|
213 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
214 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
215 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
216 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
217 |
|
break; |
218 |
|
} |
219 |
|
|
220 |
|
data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); |
221 |
|
|
222 |
|
t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
223 |
|
data->temp[0] += lambda_vec16[data->iQuant] * t; |
224 |
|
data->temp[1] += lambda_vec8[data->iQuant] * t; |
225 |
|
|
226 |
|
if (data->temp[0] < data->iMinSAD[0]) { |
227 |
|
data->iMinSAD[0] = data->temp[0]; |
228 |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
229 |
|
*dir = Direction; } |
230 |
|
|
231 |
|
if (data->temp[1] < data->iMinSAD[1]) { |
232 |
|
data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
233 |
|
if (data->temp[2] < data->iMinSAD[2]) { |
234 |
|
data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
235 |
|
if (data->temp[3] < data->iMinSAD[3]) { |
236 |
|
data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
237 |
|
if (data->temp[4] < data->iMinSAD[4]) { |
238 |
|
data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
239 |
|
} |
240 |
|
|
241 |
|
static void |
242 |
|
CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
243 |
|
|
244 |
|
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
245 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
246 |
|
// around currentMV! |
247 |
{ |
{ |
248 |
if (component == 0) |
int32_t sad; |
249 |
return 1; |
uint8_t * Reference = (uint8_t *) data->RefQ; |
250 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
251 |
|
VECTOR halfpelMV = *(data->currentMV); |
252 |
|
|
253 |
|
int32_t iEdgedWidth = data->iEdgedWidth; |
254 |
|
uint32_t rounding = data->rounding; |
255 |
|
|
256 |
if (component < 0) |
if (( x > data->max_dx) || ( x < data->min_dx) |
257 |
component = -component; |
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
258 |
|
|
259 |
if (iFcode == 1) |
switch( ((x&1)<<1) + (y&1) ) |
260 |
{ |
{ |
261 |
if (component > 32) |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
262 |
component = 32; |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); |
263 |
|
break; |
264 |
|
|
265 |
return mvtab[component] + 1; |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
266 |
} |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
267 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
268 |
|
|
269 |
|
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
270 |
|
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
271 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
272 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
273 |
|
break; |
274 |
|
|
275 |
component += (1 << (iFcode - 1)) - 1; |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
276 |
component >>= (iFcode - 1); |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
277 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
278 |
|
|
279 |
if (component > 32) |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
280 |
component = 32; |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); |
281 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); |
282 |
|
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); |
283 |
|
break; |
284 |
|
|
285 |
return mvtab[component] + 1 + iFcode - 1; |
default: // x and y in qpel resolution - the "corners" (top left/right and |
286 |
|
// bottom left/right) during qpel refinement |
287 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
288 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
289 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
290 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
291 |
|
|
292 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
293 |
|
interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); |
294 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); |
295 |
|
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); |
296 |
|
break; |
297 |
} |
} |
298 |
|
|
299 |
|
sad = lambda_vec16[data->iQuant] * |
300 |
|
d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
301 |
|
sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); |
302 |
|
|
303 |
|
if (sad < *(data->iMinSAD)) { |
304 |
|
*(data->iMinSAD) = sad; |
305 |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
306 |
|
*dir = Direction; |
307 |
|
} |
308 |
|
} |
309 |
|
|
310 |
static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
static void |
311 |
|
CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
312 |
{ |
{ |
313 |
return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int32_t sad; |
314 |
|
|
315 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
316 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
317 |
|
|
318 |
|
sad = lambda_vec16[data->iQuant] * |
319 |
|
d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
320 |
|
|
321 |
|
sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), |
322 |
|
data->iEdgedWidth, 256*4096); |
323 |
|
|
324 |
|
if (sad < *(data->iMinSAD)) { |
325 |
|
*(data->iMinSAD) = sad; |
326 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
327 |
|
*dir = Direction; } |
328 |
} |
} |
329 |
|
|
|
static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) |
|
330 |
|
|
331 |
|
static void |
332 |
|
CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) |
333 |
{ |
{ |
334 |
return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); |
int32_t sad; |
335 |
|
const int xb = data->currentMV[1].x; |
336 |
|
const int yb = data->currentMV[1].y; |
337 |
|
const uint8_t *ReferenceF, *ReferenceB; |
338 |
|
|
339 |
|
if (( xf > data->max_dx) || ( xf < data->min_dx) |
340 |
|
|| ( yf > data->max_dy) || (yf < data->min_dy)) return; |
341 |
|
|
342 |
|
switch ( ((xf&1)<<1) + (yf&1) ) { |
343 |
|
case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; |
344 |
|
case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
345 |
|
case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; |
346 |
|
default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; |
347 |
} |
} |
348 |
|
|
349 |
|
switch ( ((xb&1)<<1) + (yb&1) ) { |
350 |
|
case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; |
351 |
|
case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
352 |
|
case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; |
353 |
|
default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; |
354 |
|
} |
355 |
|
|
356 |
|
sad = lambda_vec16[data->iQuant] * |
357 |
|
( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + |
358 |
|
d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); |
359 |
|
|
360 |
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
361 |
|
|
362 |
/* calculate the min/max range (in halfpixels) |
if (sad < *(data->iMinSAD)) { |
363 |
relative to the _MACROBLOCK_ position |
*(data->iMinSAD) = sad; |
364 |
*/ |
data->currentMV->x = xf; data->currentMV->y = yf; |
365 |
|
*dir = Direction; } |
366 |
|
} |
367 |
|
|
368 |
static void __inline get_range( |
static void |
369 |
int32_t * const min_dx, int32_t * const max_dx, |
CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
int32_t * const min_dy, int32_t * const max_dy, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block_sz, // block dimension, 8 or 16 |
|
|
const uint32_t width, const uint32_t height, |
|
|
const uint32_t fcode) |
|
370 |
{ |
{ |
371 |
|
int32_t sad; |
372 |
|
int k; |
373 |
|
const uint8_t *ReferenceF; |
374 |
|
const uint8_t *ReferenceB; |
375 |
|
VECTOR mvs, b_mvs; |
376 |
|
|
377 |
const int search_range = 32 << (fcode - 1); |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
|
const int high = search_range - 1; |
|
|
const int low = -search_range; |
|
378 |
|
|
379 |
// convert full-pixel measurements to half pixel |
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
|
const int hp_width = 2 * width; |
|
|
const int hp_height = 2 * height; |
|
|
const int hp_edge = 2 * block_sz; |
|
|
const int hp_x = 2 * (x) * block_sz; // we need _right end_ of block, not x-coordinate |
|
|
const int hp_y = 2 * (y) * block_sz; // same for _bottom end_ |
|
380 |
|
|
381 |
*max_dx = MIN(high, hp_width - hp_x); |
for (k = 0; k < 4; k++) { |
382 |
*max_dy = MIN(high, hp_height - hp_y); |
mvs.x = data->directmvF[k].x + x; |
383 |
*min_dx = MAX(low, -(hp_edge + hp_x)); |
b_mvs.x = ((x == 0) ? |
384 |
*min_dy = MAX(low, -(hp_edge + hp_y)); |
data->directmvB[k].x |
385 |
|
: mvs.x - data->referencemv[k].x); |
386 |
|
|
387 |
|
mvs.y = data->directmvF[k].y + y; |
388 |
|
b_mvs.y = ((y == 0) ? |
389 |
|
data->directmvB[k].y |
390 |
|
: mvs.y - data->referencemv[k].y); |
391 |
|
|
392 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
393 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
394 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
395 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
396 |
|
|
397 |
|
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
398 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
399 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
400 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
401 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
402 |
} |
} |
403 |
|
|
404 |
|
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
405 |
|
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
406 |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
407 |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
408 |
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
409 |
|
} |
410 |
|
|
411 |
/* |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
412 |
* getref: calculate reference image pointer |
ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
413 |
* the decision to use interpolation h/v/hv or the normal image is |
ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
414 |
* based on dx & dy. |
data->iEdgedWidth); |
415 |
*/ |
if (sad > *(data->iMinSAD)) return; |
416 |
|
} |
417 |
|
|
418 |
static __inline const uint8_t * get_ref( |
if (sad < *(data->iMinSAD)) { |
419 |
const uint8_t * const refn, |
*(data->iMinSAD) = sad; |
420 |
const uint8_t * const refh, |
data->currentMV->x = x; data->currentMV->y = y; |
421 |
const uint8_t * const refv, |
*dir = Direction; } |
422 |
const uint8_t * const refhv, |
} |
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const int32_t dx, const int32_t dy, |
|
|
const uint32_t stride) |
|
|
{ |
|
423 |
|
|
424 |
switch ( ((dx&1)<<1) + (dy&1) ) // ((dx%2)?2:0)+((dy%2)?1:0) |
static void |
425 |
|
CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
426 |
{ |
{ |
427 |
case 0 : return refn + (x*block+dx/2) + (y*block+dy/2)*stride; |
int32_t sad; |
428 |
case 1 : return refv + (x*block+dx/2) + (y*block+(dy-1)/2)*stride; |
const uint8_t *ReferenceF; |
429 |
case 2 : return refh + (x*block+(dx-1)/2) + (y*block+dy/2)*stride; |
const uint8_t *ReferenceB; |
430 |
default : |
VECTOR mvs, b_mvs; |
431 |
case 3 : return refhv + (x*block+(dx-1)/2) + (y*block+(dy-1)/2)*stride; |
|
432 |
|
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
433 |
|
|
434 |
|
sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); |
435 |
|
|
436 |
|
mvs.x = data->directmvF[0].x + x; |
437 |
|
b_mvs.x = ((x == 0) ? |
438 |
|
data->directmvB[0].x |
439 |
|
: mvs.x - data->referencemv[0].x); |
440 |
|
|
441 |
|
mvs.y = data->directmvF[0].y + y; |
442 |
|
b_mvs.y = ((y == 0) ? |
443 |
|
data->directmvB[0].y |
444 |
|
: mvs.y - data->referencemv[0].y); |
445 |
|
|
446 |
|
if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) |
447 |
|
|| ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) |
448 |
|
|| ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) |
449 |
|
|| ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; |
450 |
|
|
451 |
|
switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { |
452 |
|
case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
453 |
|
case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
454 |
|
case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; |
455 |
|
default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; |
456 |
} |
} |
457 |
|
|
458 |
|
switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { |
459 |
|
case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
460 |
|
case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
461 |
|
case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; |
462 |
|
default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; |
463 |
} |
} |
464 |
|
|
465 |
|
sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
466 |
|
|
467 |
/* This is somehow a copy of get_ref, but with MV instead of X,Y */ |
if (sad < *(data->iMinSAD)) { |
468 |
|
*(data->iMinSAD) = sad; |
469 |
|
data->currentMV->x = x; data->currentMV->y = y; |
470 |
|
*dir = Direction; } |
471 |
|
} |
472 |
|
|
473 |
static __inline const uint8_t * get_ref_mv( |
static void |
474 |
const uint8_t * const refn, |
CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
|
const uint8_t * const refh, |
|
|
const uint8_t * const refv, |
|
|
const uint8_t * const refhv, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const uint32_t block, // block dimension, 8 or 16 |
|
|
const VECTOR* mv, // measured in half-pel! |
|
|
const uint32_t stride) |
|
475 |
{ |
{ |
476 |
|
int32_t sad; |
477 |
|
const uint8_t * Reference; |
478 |
|
|
479 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
480 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
481 |
|
|
482 |
switch ( (((mv->x)&1)<<1) + ((mv->y)&1) ) |
switch ( ((x&1)<<1) + (y&1) ) |
483 |
{ |
{ |
484 |
case 0 : return refn + (x*block+(mv->x)/2) + (y*block+(mv->y)/2)*stride; |
case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; |
485 |
case 1 : return refv + (x*block+(mv->x)/2) + (y*block+((mv->y)-1)/2)*stride; |
case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
486 |
case 2 : return refh + (x*block+((mv->x)-1)/2) + (y*block+(mv->y)/2)*stride; |
case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; |
487 |
default : |
default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; |
|
case 3 : return refhv + (x*block+((mv->x)-1)/2) + (y*block+((mv->y)-1)/2)*stride; |
|
488 |
} |
} |
489 |
|
|
490 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
491 |
|
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); |
492 |
|
|
493 |
|
if (sad < *(data->iMinSAD)) { |
494 |
|
*(data->iMinSAD) = sad; |
495 |
|
data->currentMV->x = x; data->currentMV->y = y; |
496 |
|
*dir = Direction; } |
497 |
} |
} |
498 |
|
|
499 |
#ifndef SEARCH16 |
static void |
500 |
#define SEARCH16 PMVfastSearch16 |
CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
501 |
#endif |
// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution |
502 |
|
// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) |
503 |
|
// around currentMV! |
504 |
|
|
505 |
#ifndef SEARCH8 |
{ |
506 |
#define SEARCH8 PMVfastSearch8 |
int32_t sad; |
507 |
#endif |
uint8_t *Reference = (uint8_t *) data->RefQ; |
508 |
|
const uint8_t *ref1, *ref2, *ref3, *ref4; |
509 |
|
VECTOR halfpelMV = *(data->currentMV); |
510 |
|
|
511 |
bool MotionEstimation( |
int32_t iEdgedWidth = data->iEdgedWidth; |
512 |
MACROBLOCK * const pMBs, |
uint32_t rounding = data->rounding; |
|
MBParam * const pParam, |
|
|
const IMAGE * const pRef, |
|
|
const IMAGE * const pRefH, |
|
|
const IMAGE * const pRefV, |
|
|
const IMAGE * const pRefHV, |
|
|
IMAGE * const pCurrent, |
|
|
const uint32_t iLimit) |
|
513 |
|
|
514 |
|
if (( x > data->max_dx) || ( x < data->min_dx) |
515 |
|
|| ( y > data->max_dy) || (y < data->min_dy)) return; |
516 |
|
|
517 |
|
switch( ((x&1)<<1) + (y&1) ) |
518 |
{ |
{ |
519 |
const uint32_t iWcount = pParam->mb_width; |
case 0: // pure halfpel position - shouldn't happen during a refinement step |
520 |
const uint32_t iHcount = pParam->mb_height; |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); |
521 |
|
break; |
522 |
|
|
523 |
uint32_t i, j, iIntra = 0; |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
524 |
|
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
525 |
|
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
526 |
|
|
527 |
VECTOR mv16; |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
528 |
VECTOR pmv16; |
break; |
529 |
|
|
530 |
int32_t sad8 = 0; |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
531 |
int32_t sad16; |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
532 |
int32_t deviation; |
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); |
533 |
|
|
534 |
// note: i==horizontal, j==vertical |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); |
535 |
for (i = 0; i < iHcount; i++) |
break; |
|
for (j = 0; j < iWcount; j++) |
|
|
{ |
|
|
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
|
536 |
|
|
537 |
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
default: // x and y in qpel resolution - the "corners" (top left/right and |
538 |
j, i, pParam->motion_flags, |
// bottom left/right) during qpel refinement |
539 |
pParam, pMBs, &mv16, &pmv16); |
GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); |
540 |
pMB->sad16=sad16; |
GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); |
541 |
|
GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); |
542 |
|
GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); |
543 |
|
|
544 |
|
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
545 |
|
break; |
546 |
|
} |
547 |
|
|
548 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
549 |
|
sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); |
550 |
|
|
551 |
/* decide: MODE_INTER or MODE_INTRA |
if (sad < *(data->iMinSAD)) { |
552 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
*(data->iMinSAD) = sad; |
553 |
*/ |
data->currentQMV->x = x; data->currentQMV->y = y; |
554 |
|
*dir = Direction; } |
555 |
|
} |
556 |
|
|
557 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
/* CHECK_CANDIATE FUNCTIONS END */ |
558 |
|
|
559 |
if (deviation < (sad16 - INTER_BIAS)) |
/* MAINSEARCH FUNCTIONS START */ |
560 |
|
|
561 |
|
static void |
562 |
|
AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
563 |
{ |
{ |
|
pMB->mode = MODE_INTRA; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
|
564 |
|
|
565 |
iIntra++; |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
if(iIntra >= iLimit) |
|
|
return 1; |
|
566 |
|
|
567 |
continue; |
int iDirection; |
568 |
|
|
569 |
|
do { |
570 |
|
iDirection = 0; |
571 |
|
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
572 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
573 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
574 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
575 |
|
|
576 |
|
/* now we're doing diagonal checks near our candidate */ |
577 |
|
|
578 |
|
if (iDirection) { //checking if anything found |
579 |
|
bDirection = iDirection; |
580 |
|
iDirection = 0; |
581 |
|
x = data->currentMV->x; y = data->currentMV->y; |
582 |
|
if (bDirection & 3) { //our candidate is left or right |
583 |
|
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
584 |
|
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
585 |
|
} else { // what remains here is up or down |
586 |
|
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
587 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
588 |
|
|
589 |
|
if (iDirection) { |
590 |
|
bDirection += iDirection; |
591 |
|
x = data->currentMV->x; y = data->currentMV->y; } |
592 |
|
} else { //about to quit, eh? not so fast.... |
593 |
|
switch (bDirection) { |
594 |
|
case 2: |
595 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
596 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
597 |
|
break; |
598 |
|
case 1: |
599 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
600 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
601 |
|
break; |
602 |
|
case 2 + 4: |
603 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
604 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
605 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
606 |
|
break; |
607 |
|
case 4: |
608 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
609 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
610 |
|
break; |
611 |
|
case 8: |
612 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
613 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
614 |
|
break; |
615 |
|
case 1 + 4: |
616 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
617 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
618 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
619 |
|
break; |
620 |
|
case 2 + 8: |
621 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
622 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
623 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
624 |
|
break; |
625 |
|
case 1 + 8: |
626 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
627 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
628 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
629 |
|
break; |
630 |
|
default: //1+2+4+8 == we didn't find anything at all |
631 |
|
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
632 |
|
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
633 |
|
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
634 |
|
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
635 |
|
break; |
636 |
|
} |
637 |
|
if (!iDirection) break; //ok, the end. really |
638 |
|
bDirection = iDirection; |
639 |
|
x = data->currentMV->x; y = data->currentMV->y; |
640 |
|
} |
641 |
|
} |
642 |
|
while (1); //forever |
643 |
} |
} |
644 |
|
|
645 |
if (pParam->global_flags & XVID_INTER4V) |
static void |
646 |
|
SquareSearch(int x, int y, const SearchData * const data, int bDirection) |
647 |
{ |
{ |
648 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
int iDirection; |
|
2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
|
|
pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
|
649 |
|
|
650 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
do { |
651 |
2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, |
iDirection = 0; |
652 |
pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); |
653 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); |
654 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); |
655 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); |
656 |
|
if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); |
657 |
|
if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); |
658 |
|
if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); |
659 |
|
if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); |
660 |
|
|
661 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
bDirection = iDirection; |
662 |
2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
x = data->currentMV->x; y = data->currentMV->y; |
663 |
pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
} while (iDirection); |
664 |
|
} |
665 |
|
|
666 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
static void |
667 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, |
DiamondSearch(int x, int y, const SearchData * const data, int bDirection) |
668 |
pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
{ |
669 |
|
|
670 |
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ |
|
} |
|
671 |
|
|
672 |
|
int iDirection; |
673 |
|
|
674 |
/* decide: MODE_INTER or MODE_INTER4V |
do { |
675 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
iDirection = 0; |
676 |
*/ |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
677 |
|
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
678 |
|
if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
679 |
|
if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
680 |
|
|
681 |
if (pMB->dquant == NO_CHANGE) { |
/* now we're doing diagonal checks near our candidate */ |
|
if (((pParam->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { |
|
682 |
|
|
683 |
sad8 = sad16; |
if (iDirection) { //checking if anything found |
684 |
pMB->mode = MODE_INTER; |
bDirection = iDirection; |
685 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
iDirection = 0; |
686 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
x = data->currentMV->x; y = data->currentMV->y; |
687 |
pMB->pmvs[0].x = pmv16.x; |
if (bDirection & 3) { //our candidate is left or right |
688 |
pMB->pmvs[0].y = pmv16.y; |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
689 |
} |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
690 |
else |
} else { // what remains here is up or down |
691 |
pMB->mode = MODE_INTER4V; |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
692 |
|
CHECK_CANDIDATE(x - iDiamondSize, y, 1); } |
693 |
|
|
694 |
|
bDirection += iDirection; |
695 |
|
x = data->currentMV->x; y = data->currentMV->y; |
696 |
} |
} |
|
else |
|
|
{ |
|
|
sad8 = sad16; |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
697 |
} |
} |
698 |
|
while (iDirection); |
699 |
} |
} |
700 |
|
|
701 |
return 0; |
/* MAINSEARCH FUNCTIONS END */ |
|
} |
|
702 |
|
|
703 |
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ |
704 |
|
|
705 |
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
static void |
706 |
|
HalfpelRefine(const SearchData * const data) |
707 |
|
{ |
708 |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
709 |
|
|
710 |
|
VECTOR backupMV = *(data->currentMV); |
711 |
|
int iDirection; //not needed |
712 |
|
|
713 |
#define CHECK_MV16_ZERO {\ |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
714 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
715 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
716 |
{ \ |
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ |
|
|
iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD <= iQuant * 96) \ |
|
|
iSAD -= MV16_00_BIAS; \ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ |
|
|
} |
|
717 |
|
|
718 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
719 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
720 |
|
|
721 |
#define CHECK_MV16_CANDIDATE(X,Y) { \ |
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
722 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
|
&& ((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); } } \ |
|
723 |
} |
} |
724 |
|
|
|
#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ |
|
|
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
|
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
|
{ \ |
|
|
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
|
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
|
|
} |
|
725 |
|
|
726 |
#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ |
static void |
727 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
QuarterpelRefine(const SearchData * const data) |
728 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
{ |
729 |
{ \ |
/* Perform quarter pixel refinement*/ |
730 |
iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ |
|
731 |
iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
VECTOR backupMV = *(data->currentQMV); |
732 |
if (iSAD < iMinSAD) \ |
int iDirection; //not needed |
733 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
|
734 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); |
735 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); |
736 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); |
737 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); |
738 |
|
|
739 |
|
CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); |
740 |
|
CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); |
741 |
|
|
742 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); |
743 |
|
CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); |
744 |
|
|
745 |
} |
} |
746 |
|
|
747 |
|
static __inline int |
748 |
|
SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
749 |
|
const int x, const int y, |
750 |
|
const uint32_t iEdgedWidth, const uint32_t iQuant) |
751 |
|
|
752 |
#define CHECK_MV8_ZERO {\ |
{ |
753 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ |
/* keep repeating checks for all b-frames before this P frame, |
754 |
iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
to make sure that SKIP is possible (todo) |
755 |
if (iSAD < iMinSAD) \ |
how: if skip is not possible set sad00 to a very high value */ |
756 |
{ iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ |
|
757 |
|
uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, |
758 |
|
reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); |
759 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
760 |
|
sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, |
761 |
|
reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
762 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
763 |
|
|
764 |
|
return 1; |
765 |
} |
} |
766 |
|
|
767 |
|
static __inline void |
768 |
|
SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
769 |
|
{ |
770 |
|
pMB->mode = MODE_NOT_CODED; |
771 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
772 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
773 |
|
|
774 |
#define CHECK_MV8_CANDIDATE(X,Y) { \ |
pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; |
775 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; |
776 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
|
777 |
{ \ |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
|
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
|
|
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
|
|
if (iSAD < iMinSAD) \ |
|
|
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ |
|
778 |
} |
} |
779 |
|
|
780 |
#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ |
bool |
781 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
MotionEstimation(MBParam * const pParam, |
782 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
FRAMEINFO * const current, |
783 |
{ \ |
FRAMEINFO * const reference, |
784 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
const IMAGE * const pRefH, |
785 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
const IMAGE * const pRefV, |
786 |
if (iSAD < iMinSAD) \ |
const IMAGE * const pRefHV, |
787 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ |
const uint32_t iLimit) |
788 |
|
{ |
789 |
|
MACROBLOCK *const pMBs = current->mbs; |
790 |
|
const IMAGE *const pCurrent = ¤t->image; |
791 |
|
const IMAGE *const pRef = &reference->image; |
792 |
|
|
793 |
|
const VECTOR zeroMV = { 0, 0 }; |
794 |
|
|
795 |
|
uint32_t x, y; |
796 |
|
uint32_t iIntra = 0; |
797 |
|
int32_t InterBias, quant = current->quant; |
798 |
|
uint8_t *qimage; |
799 |
|
|
800 |
|
// some pre-initialized thingies for SearchP |
801 |
|
int32_t temp[5]; |
802 |
|
VECTOR currentMV[5]; |
803 |
|
VECTOR currentQMV[5]; |
804 |
|
int32_t iMinSAD[5]; |
805 |
|
SearchData Data; |
806 |
|
Data.iEdgedWidth = pParam->edged_width; |
807 |
|
Data.currentMV = currentMV; |
808 |
|
Data.currentQMV = currentQMV; |
809 |
|
Data.iMinSAD = iMinSAD; |
810 |
|
Data.temp = temp; |
811 |
|
Data.iFcode = current->fcode; |
812 |
|
Data.rounding = pParam->m_rounding_type; |
813 |
|
|
814 |
|
if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) |
815 |
|
return 0; // allocate some mem for qpel interpolated blocks |
816 |
|
// somehow this is dirty since I think we shouldn't use malloc outside |
817 |
|
// encoder_create() - so please fix me! |
818 |
|
|
819 |
|
if (sadInit) (*sadInit) (); |
820 |
|
|
821 |
|
for (y = 0; y < pParam->mb_height; y++) { |
822 |
|
for (x = 0; x < pParam->mb_width; x++) { |
823 |
|
|
824 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
825 |
|
int32_t sad00 = pMB->sad16 |
826 |
|
= sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, |
827 |
|
pRef->y + (x + y * pParam->edged_width) * 16, |
828 |
|
pParam->edged_width, pMB->sad8 ); |
829 |
|
|
830 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
831 |
|
pMB->dquant = NO_CHANGE; |
832 |
|
pMB->quant = current->quant; } |
833 |
|
else |
834 |
|
if (pMB->dquant != NO_CHANGE) { |
835 |
|
quant += DQtab[pMB->dquant]; |
836 |
|
if (quant > 31) quant = 31; |
837 |
|
else if (quant < 1) quant = 1; |
838 |
|
pMB->quant = quant; |
839 |
} |
} |
840 |
|
|
841 |
#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ |
//initial skip decision |
842 |
if ( ((X) <= max_dx) && ((X) >= min_dx) \ |
|
843 |
&& ((Y) <= max_dy) && ((Y) >= min_dy) ) \ |
if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) |
844 |
{ \ |
&& (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { |
845 |
iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ |
if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { |
846 |
iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ |
SkipMacroblockP(pMB, sad00); |
847 |
if (iSAD < iMinSAD) \ |
continue; |
848 |
{ iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ |
sad00 = 256 * 4096; |
849 |
} |
} |
850 |
|
} else sad00 = 256*4096; // skip not allowed - for final skip decision |
851 |
|
|
852 |
/* too slow and not fully functional at the moment */ |
SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, |
853 |
/* |
y, current->motion_flags, pMB->quant, |
854 |
int32_t ZeroSearch16( |
&Data, pParam, pMBs, reference->mbs, |
855 |
const uint8_t * const pRef, |
current->global_flags & XVID_INTER4V, pMB); |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
MBParam * const pParam, |
|
|
MACROBLOCK * const pMBs, |
|
|
VECTOR * const currMV, |
|
|
VECTOR * const currPMV) |
|
|
{ |
|
|
const int32_t iEdgedWidth = pParam->edged_width; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
|
|
int32_t iSAD; |
|
|
int32_t pred_x,pred_y; |
|
856 |
|
|
857 |
get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); |
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
858 |
|
if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) |
859 |
|
if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) |
860 |
|
{ SkipMacroblockP(pMB, sad00); continue; } |
861 |
|
|
862 |
iSAD = sad16( cur, |
/* finally, intra decision */ |
|
get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
if (iSAD <= iQuant * 96) |
|
|
iSAD -= MV16_00_BIAS; |
|
863 |
|
|
864 |
currMV->x = 0; |
InterBias = MV16_INTER_BIAS; |
865 |
currMV->y = 0; |
if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work |
866 |
currPMV->x = -pred_x; |
if (y != 0) |
867 |
currPMV->y = -pred_y; |
if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; |
868 |
|
if (x != 0) |
869 |
|
if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; |
870 |
|
|
871 |
return iSAD; |
if (InterBias < pMB->sad16) { |
872 |
|
const int32_t deviation = |
873 |
|
dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
874 |
|
pParam->edged_width); |
875 |
|
|
876 |
|
if (deviation < (pMB->sad16 - InterBias)) { |
877 |
|
if (++iIntra >= iLimit) { free(qimage); return 1; } |
878 |
|
pMB->mode = MODE_INTRA; |
879 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = |
880 |
|
pMB->mvs[3] = zeroMV; |
881 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = |
882 |
|
pMB->qmvs[3] = zeroMV; |
883 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = |
884 |
|
pMB->sad8[3] = 0; |
885 |
} |
} |
|
*/ |
|
|
|
|
|
int32_t PMVfastSearch16_MainSearch( |
|
|
const uint8_t * const pRef, |
|
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const uint8_t * const cur, |
|
|
const int x, const int y, |
|
|
int32_t startx, int32_t starty, |
|
|
int32_t iMinSAD, |
|
|
VECTOR * const currMV, |
|
|
const VECTOR * const pmv, |
|
|
const int32_t min_dx, const int32_t max_dx, |
|
|
const int32_t min_dy, const int32_t max_dy, |
|
|
const int32_t iEdgedWidth, |
|
|
const int32_t iDiamondSize, |
|
|
const int32_t iFcode, |
|
|
const int32_t iQuant, |
|
|
int iFound) |
|
|
{ |
|
|
/* 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); |
|
886 |
} |
} |
887 |
else |
} |
888 |
|
} |
889 |
|
free(qimage); |
890 |
|
return 0; |
891 |
|
} |
892 |
|
|
893 |
|
|
894 |
|
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
895 |
|
|
896 |
|
static __inline int |
897 |
|
make_mask(const VECTOR * const pmv, const int i) |
898 |
{ |
{ |
899 |
currMV->x = startx; |
int mask = 255, j; |
900 |
currMV->y = starty; |
for (j = 0; j < i; j++) { |
901 |
|
if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already |
902 |
|
if (pmv[i].x == pmv[j].x) { |
903 |
|
if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } |
904 |
|
if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } |
905 |
|
} else |
906 |
|
if (pmv[i].y == pmv[j].y) { |
907 |
|
if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } |
908 |
|
if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } |
909 |
|
} |
910 |
} |
} |
911 |
return iMinSAD; |
return mask; |
912 |
} |
} |
913 |
|
|
914 |
int32_t PMVfastSearch16_Refine( |
static __inline void |
915 |
const uint8_t * const pRef, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, |
916 |
const uint8_t * const pRefH, |
const int iHcount, const MACROBLOCK * const prevMB) |
|
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) |
|
917 |
{ |
{ |
|
/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ |
|
918 |
|
|
919 |
int32_t iSAD; |
//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself |
|
VECTOR backupMV = *currMV; |
|
920 |
|
|
921 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour |
922 |
CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); |
pmv[5].x = EVEN(pmv[3].x); |
923 |
CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); |
pmv[5].y = EVEN(pmv[3].y); |
924 |
CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); |
} else pmv[5].x = pmv[5].y = 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); |
|
925 |
|
|
926 |
return iMinSAD; |
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour |
927 |
} |
else pmv[3].x = pmv[3].y = 0; |
928 |
|
|
929 |
#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) |
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour |
930 |
|
else pmv[4].x = pmv[4].y = 0; |
931 |
|
|
932 |
|
// [1] median prediction |
933 |
|
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
934 |
|
|
935 |
|
pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask |
936 |
|
|
937 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
938 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
939 |
|
|
940 |
|
if ((x != iWcount-1) && (y != iHcount-1)) { |
941 |
|
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame |
942 |
|
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
943 |
|
} else pmv[6].x = pmv[6].y = 0; |
944 |
|
} |
945 |
|
|
946 |
int32_t PMVfastSearch16( |
static void |
947 |
const uint8_t * const pRef, |
SearchP(const uint8_t * const pRef, |
948 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
949 |
const uint8_t * const pRefV, |
const uint8_t * const pRefV, |
950 |
const uint8_t * const pRefHV, |
const uint8_t * const pRefHV, |
951 |
|
const uint8_t * const pRefQ, |
952 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
953 |
const int x, const int y, |
const int x, |
954 |
|
const int y, |
955 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
956 |
MBParam * const pParam, |
const uint32_t iQuant, |
957 |
MACROBLOCK * const pMBs, |
SearchData * const Data, |
958 |
VECTOR * const currMV, |
const MBParam * const pParam, |
959 |
VECTOR * const currPMV) |
const MACROBLOCK * const pMBs, |
960 |
{ |
const MACROBLOCK * const prevMBs, |
961 |
const uint32_t iWcount = pParam->mb_width; |
int inter4v, |
962 |
const int32_t iFcode = pParam->fixed_code; |
MACROBLOCK * const pMB) |
963 |
const int32_t iQuant = pParam->quant; |
{ |
964 |
const int32_t iWidth = pParam->width; |
|
965 |
const int32_t iHeight = pParam->height; |
int i, iDirection = 255, mask, threshA; |
966 |
const int32_t iEdgedWidth = pParam->edged_width; |
VECTOR pmv[7]; |
967 |
|
|
968 |
const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; |
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
969 |
|
|
970 |
int32_t iDiamondSize; |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() |
971 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
972 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
973 |
|
|
974 |
|
Data->predMV = pmv[0]; |
975 |
|
|
976 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; |
977 |
|
Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; |
978 |
|
Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; |
979 |
|
Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; |
980 |
|
Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
981 |
|
Data->RefQ = pRefQ; |
982 |
|
|
983 |
|
Data->iQuant = iQuant; |
984 |
|
|
985 |
|
if (!(MotionFlags & PMV_HALFPEL16)) { |
986 |
|
Data->min_dx = EVEN(Data->min_dx); |
987 |
|
Data->max_dx = EVEN(Data->max_dx); |
988 |
|
Data->min_dy = EVEN(Data->min_dy); |
989 |
|
Data->max_dy = EVEN(Data->max_dy); } |
990 |
|
|
991 |
|
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
992 |
|
|
993 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
994 |
|
else CheckCandidate = CheckCandidate16no4v; |
995 |
|
|
996 |
|
for(i = 0; i < 5; i++) |
997 |
|
Data->currentMV[i].x = Data->currentMV[i].y = 0; |
998 |
|
|
999 |
|
i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); |
1000 |
|
Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; |
1001 |
|
Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; |
1002 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
1003 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
1004 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
1005 |
|
|
1006 |
|
if ((x == 0) && (y == 0)) threshA = 512; |
1007 |
|
else { |
1008 |
|
threshA = Data->temp[0]; // that's when we keep this SAD atm |
1009 |
|
if (threshA < 512) threshA = 512; |
1010 |
|
if (threshA > 1024) threshA = 1024; } |
1011 |
|
|
1012 |
int32_t min_dx; |
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
1013 |
int32_t max_dx; |
prevMBs + x + y * pParam->mb_width); |
|
int32_t min_dy; |
|
|
int32_t max_dy; |
|
1014 |
|
|
1015 |
int32_t iFound; |
if (inter4v) CheckCandidate = CheckCandidate16; |
1016 |
|
else CheckCandidate = CheckCandidate16no4v; |
1017 |
|
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; /* just for PMVFAST */ |
|
1018 |
|
|
1019 |
VECTOR pmv[4]; |
/* main loop. checking all predictions */ |
|
int32_t psad[4]; |
|
1020 |
|
|
1021 |
MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
for (i = 1; i < 7; i++) { |
1022 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
1023 |
|
(*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1024 |
|
if (Data->iMinSAD[0] <= threshA) break; |
1025 |
|
} |
1026 |
|
|
1027 |
static int32_t threshA,threshB; |
if ((Data->iMinSAD[0] <= threshA) || |
1028 |
int32_t bPredEq; |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1029 |
int32_t iMinSAD,iSAD; |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
1030 |
|
inter4v = 0; |
1031 |
|
} else { |
1032 |
|
|
1033 |
/* Get maximum range */ |
MainSearchFunc * MainSearchPtr; |
1034 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1035 |
x, y, 16, iWidth, iHeight, iFcode); |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1036 |
|
else MainSearchPtr = DiamondSearch; |
1037 |
|
|
1038 |
/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
1039 |
|
|
1040 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
/* extended search, diamond starting in 0,0 and in prediction. |
1041 |
{ min_dx = EVEN(min_dx); |
note that this search is/might be done in halfpel positions, |
1042 |
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 something like IF (dx>max_dx) THEN dx=max_dx; */ |
|
1043 |
|
|
1044 |
|
if (MotionFlags & PMV_EXTSEARCH16) { |
1045 |
|
int32_t bSAD; |
1046 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
1047 |
|
if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? |
1048 |
|
startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); |
1049 |
|
if (!(MVequal(startMV, backupMV))) { |
1050 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1051 |
|
|
1052 |
bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); |
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
1053 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1054 |
|
if (bSAD < Data->iMinSAD[0]) { |
1055 |
|
Data->currentMV[0] = backupMV; |
1056 |
|
Data->iMinSAD[0] = bSAD; } |
1057 |
|
} |
1058 |
|
|
1059 |
if ((x==0) && (y==0) ) |
backupMV = Data->currentMV[0]; |
1060 |
{ |
if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; |
1061 |
threshA = 512; |
else startMV.x = startMV.y = 0; |
1062 |
threshB = 1024; |
if (!(MVequal(startMV, backupMV))) { |
1063 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
1064 |
|
|
1065 |
|
CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); |
1066 |
|
(*MainSearchPtr)(startMV.x, startMV.y, Data, 255); |
1067 |
|
if (bSAD < Data->iMinSAD[0]) { |
1068 |
|
Data->currentMV[0] = backupMV; |
1069 |
|
Data->iMinSAD[0] = bSAD; } |
1070 |
|
} |
1071 |
} |
} |
|
else |
|
|
{ |
|
|
threshA = psad[0]; |
|
|
threshB = threshA+256; |
|
|
if (threshA< 512) threshA = 512; |
|
|
if (threshA>1024) threshA = 1024; |
|
|
if (threshB>1792) threshB = 1792; |
|
1072 |
} |
} |
1073 |
|
|
1074 |
iFound=0; |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
|
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
1075 |
|
|
1076 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) |
for(i = 0; i < 5; i++) { |
1077 |
iFound=2; |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
1078 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1079 |
|
} |
1080 |
|
|
1081 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { |
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
1082 |
|
|
1083 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
if(inter4v) |
1084 |
iDiamondSize=1; // halfpel! |
CheckCandidate = CheckCandidate16_qpel; |
1085 |
else |
else |
1086 |
iDiamondSize=2; // halfpel! |
CheckCandidate = CheckCandidate16no4v_qpel; |
1087 |
|
|
1088 |
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
QuarterpelRefine(Data); |
1089 |
iDiamondSize*=2; |
} |
1090 |
|
|
1091 |
/* Step 4: Calculate SAD around the Median prediction. |
if (inter4v) { |
1092 |
MinSAD=SAD |
SearchData Data8; |
1093 |
If Motion Vector equal to Previous frame motion vector |
Data8.iFcode = Data->iFcode; |
1094 |
and MinSAD<PrevFrmSAD goto Step 10. |
Data8.iQuant = Data->iQuant; |
1095 |
If SAD<=256 goto Step 10. |
Data8.iEdgedWidth = Data->iEdgedWidth; |
1096 |
*/ |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1097 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1098 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
1099 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1100 |
|
} |
1101 |
|
|
1102 |
|
if (!(inter4v) || |
1103 |
|
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1104 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1105 |
|
// INTER MODE |
1106 |
|
pMB->mode = MODE_INTER; |
1107 |
|
pMB->mvs[0] = pMB->mvs[1] |
1108 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1109 |
|
|
1110 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1111 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1112 |
|
|
1113 |
// Prepare for main loop |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1114 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1115 |
|
|
1116 |
*currMV=pmv[0]; /* current best := prediction */ |
if(pParam->m_quarterpel) { |
1117 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; |
1118 |
{ /* This should NOT be necessary! */ |
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; |
|
currMV->x = EVEN(currMV->x); |
|
|
currMV->y = EVEN(currMV->y); |
|
1119 |
} |
} |
1120 |
|
else { |
1121 |
if (currMV->x > max_dx) |
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1122 |
{ |
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
|
currMV->x=max_dx; |
|
1123 |
} |
} |
1124 |
if (currMV->x < min_dx) |
} else { |
1125 |
{ |
// INTER4V MODE; all other things are already set in Search8 |
1126 |
currMV->x=min_dx; |
pMB->mode = MODE_INTER4V; |
1127 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1128 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; |
1129 |
} |
} |
|
if (currMV->y > max_dy) |
|
|
{ |
|
|
currMV->y=max_dy; |
|
1130 |
} |
} |
1131 |
if (currMV->y < min_dy) |
|
1132 |
|
static void |
1133 |
|
Search8(const SearchData * const OldData, |
1134 |
|
const int x, const int y, |
1135 |
|
const uint32_t MotionFlags, |
1136 |
|
const MBParam * const pParam, |
1137 |
|
MACROBLOCK * const pMB, |
1138 |
|
const MACROBLOCK * const pMBs, |
1139 |
|
const int block, |
1140 |
|
SearchData * const Data) |
1141 |
{ |
{ |
1142 |
currMV->y=min_dy; |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1143 |
|
Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1144 |
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1145 |
|
Data->currentMV = OldData->currentMV + 1 + block; |
1146 |
|
Data->currentQMV = OldData->currentQMV + 1 + block; |
1147 |
|
|
1148 |
|
if (block != 0) { |
1149 |
|
if(pParam->m_quarterpel) { |
1150 |
|
*(Data->iMinSAD) += lambda_vec8[Data->iQuant] * |
1151 |
|
d_mv_bits( Data->currentQMV->x - Data->predQMV.x, |
1152 |
|
Data->currentQMV->y - Data->predQMV.y, |
1153 |
|
Data->iFcode); |
1154 |
|
} |
1155 |
|
else { |
1156 |
|
*(Data->iMinSAD) += lambda_vec8[Data->iQuant] * |
1157 |
|
d_mv_bits( Data->currentMV->x - Data->predMV.x, |
1158 |
|
Data->currentMV->y - Data->predMV.y, |
1159 |
|
Data->iFcode); |
1160 |
|
} |
1161 |
} |
} |
1162 |
|
|
|
iMinSAD = sad16( cur, |
|
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), |
|
|
iEdgedWidth, MV_MAX_ERROR); |
|
|
iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
1163 |
|
|
1164 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && (iMinSAD < pMB->sad16) ) ) |
if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { |
|
{ |
|
1165 |
|
|
1166 |
if (MotionFlags & PMV_QUICKSTOP16) |
Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1167 |
goto step10b; |
Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1168 |
if (MotionFlags & PMV_EARLYSTOP16) |
Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1169 |
goto step10; |
Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1170 |
} |
Data->RefQ = OldData->RefQ; |
1171 |
|
|
1172 |
/* |
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. ******** WHAT'S THIS 'OFFSET' ??? *********** |
|
|
*/ |
|
1173 |
|
|
1174 |
// (0,0) is always possible |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, |
1175 |
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1176 |
|
|
1177 |
CHECK_MV16_ZERO; |
CheckCandidate = CheckCandidate8; |
1178 |
|
|
1179 |
// previous frame MV is always possible |
if (MotionFlags & PMV_EXTSEARCH8) { |
1180 |
CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1181 |
|
|
1182 |
// left neighbour, if allowed |
MainSearchFunc *MainSearchPtr; |
1183 |
if (x != 0) |
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1184 |
{ |
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1185 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
else MainSearchPtr = DiamondSearch; |
1186 |
{ pmv[1].x = EVEN(pmv[1].x); |
|
1187 |
pmv[1].y = EVEN(pmv[1].y); |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1188 |
|
|
1189 |
|
if(*(Data->iMinSAD) < temp_sad) { |
1190 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1191 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1192 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1193 |
} |
} |
1194 |
|
|
1195 |
// top neighbour, if allowed |
if (MotionFlags & PMV_HALFPELREFINE8) { |
1196 |
if (y != 0) |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
1197 |
{ |
|
1198 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
HalfpelRefine(Data); // perform halfpel refine of current best vector |
1199 |
{ pmv[2].x = EVEN(pmv[2].x); |
|
1200 |
pmv[2].y = EVEN(pmv[2].y); |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
1201 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
1202 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
1203 |
|
} |
1204 |
} |
} |
|
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1205 |
|
|
1206 |
// top right neighbour, if allowed |
if(pParam->m_quarterpel) { |
1207 |
if (x != (iWcount-1)) |
if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && |
1208 |
{ |
(MotionFlags & PMV_QUARTERPELREFINE8)) { |
1209 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
|
1210 |
{ pmv[3].x = EVEN(pmv[3].x); |
CheckCandidate = CheckCandidate8_qpel; |
1211 |
pmv[3].y = EVEN(pmv[3].y); |
QuarterpelRefine(Data); |
1212 |
|
} |
1213 |
|
} |
1214 |
|
} |
1215 |
|
|
1216 |
|
if(pParam->m_quarterpel) { |
1217 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; |
1218 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; |
1219 |
} |
} |
1220 |
CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); |
else { |
1221 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
1222 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
1223 |
} |
} |
1224 |
|
|
1225 |
|
pMB->mvs[block] = *(Data->currentMV); |
1226 |
|
pMB->qmvs[block] = *(Data->currentQMV); |
1227 |
|
|
1228 |
|
pMB->sad8[block] = 4 * (*Data->iMinSAD); |
1229 |
} |
} |
1230 |
|
|
1231 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
/* B-frames code starts here */ |
|
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
|
|
*/ |
|
1232 |
|
|
1233 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[0]) && (iMinSAD < pMB->sad16) ) ) |
static __inline VECTOR |
1234 |
|
ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) |
1235 |
{ |
{ |
1236 |
if (MotionFlags & PMV_QUICKSTOP16) |
/* the stupidiest function ever */ |
1237 |
goto step10b; |
if (mode == MODE_FORWARD) return pMB->mvs[0]; |
1238 |
if (MotionFlags & PMV_EARLYSTOP16) |
else return pMB->b_mvs[0]; |
|
goto step10; |
|
1239 |
} |
} |
1240 |
|
|
1241 |
|
static void __inline |
1242 |
|
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
1243 |
|
const uint32_t iWcount, |
1244 |
|
const MACROBLOCK * const pMB, |
1245 |
|
const uint32_t mode_curr) |
1246 |
|
{ |
1247 |
|
|
1248 |
|
// [0] is prediction |
1249 |
|
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
1250 |
|
|
1251 |
/************ (Diamond Search) **************/ |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
|
/* |
|
|
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. |
|
|
*/ |
|
1252 |
|
|
1253 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
pmv[2] = ChoosePred(pMB, mode_curr); |
1254 |
|
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1255 |
|
|
1256 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
1257 |
iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1258 |
x, y, |
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
1259 |
currMV->x, currMV->y, iMinSAD, &newMV, |
} else pmv[3].x = pmv[3].y = 0; |
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1260 |
|
|
1261 |
if (iSAD < iMinSAD) |
if (y != 0) { |
1262 |
{ |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
1263 |
*currMV = newMV; |
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
1264 |
iMinSAD = iSAD; |
} else pmv[4].x = pmv[4].y = 0; |
1265 |
|
|
1266 |
|
if (x != 0) { |
1267 |
|
pmv[5] = ChoosePred(pMB-1, mode_curr); |
1268 |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1269 |
|
} else pmv[5].x = pmv[5].y = 0; |
1270 |
|
|
1271 |
|
if ((x != 0)&&(y != 0)) { |
1272 |
|
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
1273 |
|
pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
1274 |
|
} else pmv[6].x = pmv[6].y = 0; |
1275 |
|
|
1276 |
|
// more? |
1277 |
} |
} |
1278 |
|
|
1279 |
if (MotionFlags & PMV_EXTSEARCH16) |
|
1280 |
|
/* search backward or forward, for b-frames */ |
1281 |
|
static void |
1282 |
|
SearchBF( const uint8_t * const pRef, |
1283 |
|
const uint8_t * const pRefH, |
1284 |
|
const uint8_t * const pRefV, |
1285 |
|
const uint8_t * const pRefHV, |
1286 |
|
const IMAGE * const pCur, |
1287 |
|
const int x, const int y, |
1288 |
|
const uint32_t MotionFlags, |
1289 |
|
const uint32_t iFcode, |
1290 |
|
const MBParam * const pParam, |
1291 |
|
MACROBLOCK * const pMB, |
1292 |
|
const VECTOR * const predMV, |
1293 |
|
int32_t * const best_sad, |
1294 |
|
const int32_t mode_current, |
1295 |
|
SearchData * const Data) |
1296 |
{ |
{ |
|
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
|
1297 |
|
|
1298 |
if (!(MVequal(pmv[0],backupMV)) ) |
const int32_t iEdgedWidth = pParam->edged_width; |
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1299 |
|
|
1300 |
if (iSAD < iMinSAD) |
int i, iDirection, mask; |
1301 |
{ |
VECTOR pmv[7]; |
1302 |
*currMV = newMV; |
MainSearchFunc *MainSearchPtr; |
1303 |
iMinSAD = iSAD; |
*Data->iMinSAD = MV_MAX_ERROR; |
1304 |
|
Data->iFcode = iFcode; |
1305 |
|
|
1306 |
|
Data->Ref = pRef + (x + y * iEdgedWidth) * 16; |
1307 |
|
Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; |
1308 |
|
Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; |
1309 |
|
Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; |
1310 |
|
|
1311 |
|
Data->predMV = *predMV; |
1312 |
|
|
1313 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1314 |
|
pParam->width, pParam->height, iFcode, pParam->m_quarterpel); |
1315 |
|
|
1316 |
|
pmv[0] = Data->predMV; |
1317 |
|
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
1318 |
|
|
1319 |
|
Data->currentMV->x = Data->currentMV->y = 0; |
1320 |
|
|
1321 |
|
CheckCandidate = CheckCandidate16no4v; |
1322 |
|
|
1323 |
|
// main loop. checking all predictions |
1324 |
|
for (i = 0; i < 8; i++) { |
1325 |
|
if (!(mask = make_mask(pmv, i)) ) continue; |
1326 |
|
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1327 |
} |
} |
1328 |
|
|
1329 |
|
if (MotionFlags & PMV_USESQUARES16) |
1330 |
|
MainSearchPtr = SquareSearch; |
1331 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1332 |
|
MainSearchPtr = AdvDiamondSearch; |
1333 |
|
else MainSearchPtr = DiamondSearch; |
1334 |
|
|
1335 |
|
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1336 |
|
|
1337 |
|
HalfpelRefine(Data); |
1338 |
|
|
1339 |
|
// three bits are needed to code backward mode. four for forward |
1340 |
|
// we treat the bits just like they were vector's |
1341 |
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * lambda_vec16[Data->iQuant]; |
1342 |
|
else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; |
1343 |
|
|
1344 |
|
|
1345 |
|
if (*Data->iMinSAD < *best_sad) { |
1346 |
|
*best_sad = *Data->iMinSAD; |
1347 |
|
pMB->mode = mode_current; |
1348 |
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
1349 |
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
1350 |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
1351 |
|
else pMB->b_mvs[0] = *Data->currentMV; |
1352 |
} |
} |
1353 |
|
|
1354 |
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
} |
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1355 |
|
|
1356 |
if (iSAD < iMinSAD) |
static int32_t |
1357 |
{ |
SearchDirect(const IMAGE * const f_Ref, |
1358 |
*currMV = newMV; |
const uint8_t * const f_RefH, |
1359 |
iMinSAD = iSAD; |
const uint8_t * const f_RefV, |
1360 |
|
const uint8_t * const f_RefHV, |
1361 |
|
const IMAGE * const b_Ref, |
1362 |
|
const uint8_t * const b_RefH, |
1363 |
|
const uint8_t * const b_RefV, |
1364 |
|
const uint8_t * const b_RefHV, |
1365 |
|
const IMAGE * const pCur, |
1366 |
|
const int x, const int y, |
1367 |
|
const uint32_t MotionFlags, |
1368 |
|
const int32_t TRB, const int32_t TRD, |
1369 |
|
const MBParam * const pParam, |
1370 |
|
MACROBLOCK * const pMB, |
1371 |
|
const MACROBLOCK * const b_mb, |
1372 |
|
int32_t * const best_sad, |
1373 |
|
SearchData * const Data) |
1374 |
|
|
1375 |
|
{ |
1376 |
|
int32_t skip_sad; |
1377 |
|
int k; |
1378 |
|
|
1379 |
|
MainSearchFunc *MainSearchPtr; |
1380 |
|
|
1381 |
|
*Data->iMinSAD = 256*4096; |
1382 |
|
Data->referencemv = b_mb->mvs; |
1383 |
|
|
1384 |
|
Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1385 |
|
Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; |
1386 |
|
Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; |
1387 |
|
Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1388 |
|
Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; |
1389 |
|
Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; |
1390 |
|
Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; |
1391 |
|
Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; |
1392 |
|
|
1393 |
|
Data->max_dx = 2 * pParam->width - 2 * (x) * 16; |
1394 |
|
Data->max_dy = 2 * pParam->height - 2 * (y) * 16; |
1395 |
|
Data->min_dx = -(2 * 16 + 2 * (x) * 16); |
1396 |
|
Data->min_dy = -(2 * 16 + 2 * (y) * 16); |
1397 |
|
|
1398 |
|
for (k = 0; k < 4; k++) { |
1399 |
|
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
1400 |
|
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
1401 |
|
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
1402 |
|
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
1403 |
|
|
1404 |
|
if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) |
1405 |
|
|| ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { |
1406 |
|
|
1407 |
|
*best_sad = 256*4096; // in that case, we won't use direct mode |
1408 |
|
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
1409 |
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1410 |
|
return 0; |
1411 |
} |
} |
1412 |
|
if (b_mb->mode != MODE_INTER4V) { |
1413 |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
1414 |
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
1415 |
|
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
1416 |
|
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
1417 |
|
break; |
1418 |
} |
} |
1419 |
} |
} |
1420 |
|
|
1421 |
/* |
if (b_mb->mode == MODE_INTER4V) |
1422 |
Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
CheckCandidate = CheckCandidateDirect; |
1423 |
*/ |
else CheckCandidate = CheckCandidateDirectno4v; |
1424 |
|
|
1425 |
step10: |
(*CheckCandidate)(0, 0, 255, &k, Data); |
|
if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step |
|
|
iMinSAD = PMVfastSearch16_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
1426 |
|
|
1427 |
step10b: |
// skip decision |
1428 |
currPMV->x = currMV->x - pmv[0].x; |
if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { |
1429 |
currPMV->y = currMV->y - pmv[0].y; |
//checking chroma. everything copied from MC |
1430 |
return iMinSAD; |
//this is not full chroma compensation, only it's fullpel approximation. should work though |
1431 |
} |
int sum, dx, dy, b_dx, b_dy; |
1432 |
|
|
1433 |
|
sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; |
1434 |
|
dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1435 |
|
|
1436 |
|
sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; |
1437 |
|
dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1438 |
|
|
1439 |
|
sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; |
1440 |
|
b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1441 |
|
|
1442 |
|
sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; |
1443 |
|
b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); |
1444 |
|
|
1445 |
|
sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), |
1446 |
|
f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1447 |
|
b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1448 |
|
Data->iEdgedWidth/2); |
1449 |
|
sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), |
1450 |
|
f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, |
1451 |
|
b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, |
1452 |
|
Data->iEdgedWidth/2); |
1453 |
|
|
1454 |
int32_t PMVfastSearch8_MainSearch( |
if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { |
1455 |
const uint8_t * const pRef, |
pMB->mode = MODE_DIRECT_NONE_MV; |
1456 |
const uint8_t * const pRefH, |
return *Data->iMinSAD; |
|
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); |
|
1457 |
} |
} |
|
else |
|
|
{ |
|
|
currMV->x = startx; |
|
|
currMV->y = starty; |
|
|
} |
|
|
return iMinSAD; |
|
1458 |
} |
} |
1459 |
|
|
1460 |
int32_t PMVfastSearch8_Refine( |
skip_sad = *Data->iMinSAD; |
|
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) */ |
|
1461 |
|
|
1462 |
int32_t iSAD; |
// DIRECT MODE DELTA VECTOR SEARCH. |
1463 |
VECTOR backupMV = *currMV; |
// This has to be made more effective, but at the moment I'm happy it's running at all |
1464 |
|
|
1465 |
CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); |
if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; |
1466 |
CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1467 |
CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); |
else MainSearchPtr = DiamondSearch; |
|
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); |
|
1468 |
|
|
1469 |
return iMinSAD; |
(*MainSearchPtr)(0, 0, Data, 255); |
|
} |
|
1470 |
|
|
1471 |
|
HalfpelRefine(Data); |
1472 |
|
|
1473 |
#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) |
*Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's |
1474 |
|
*best_sad = *Data->iMinSAD; |
1475 |
|
|
1476 |
int32_t PMVfastSearch8( |
if (b_mb->mode == MODE_INTER4V) |
1477 |
const uint8_t * const pRef, |
pMB->mode = MODE_DIRECT; |
1478 |
const uint8_t * const pRefH, |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
1479 |
const uint8_t * const pRefV, |
|
1480 |
const uint8_t * const pRefHV, |
pMB->pmvs[3] = *Data->currentMV; |
1481 |
|
|
1482 |
|
for (k = 0; k < 4; k++) { |
1483 |
|
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
1484 |
|
pMB->b_mvs[k].x = ((Data->currentMV->x == 0) |
1485 |
|
? Data->directmvB[k].x |
1486 |
|
: pMB->mvs[k].x - Data->referencemv[k].x); |
1487 |
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
1488 |
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
1489 |
|
? Data->directmvB[k].y |
1490 |
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
1491 |
|
if (b_mb->mode != MODE_INTER4V) { |
1492 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
1493 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
1494 |
|
break; |
1495 |
|
} |
1496 |
|
} |
1497 |
|
return skip_sad; |
1498 |
|
} |
1499 |
|
|
1500 |
|
|
1501 |
|
static __inline void |
1502 |
|
SearchInterpolate(const uint8_t * const f_Ref, |
1503 |
|
const uint8_t * const f_RefH, |
1504 |
|
const uint8_t * const f_RefV, |
1505 |
|
const uint8_t * const f_RefHV, |
1506 |
|
const uint8_t * const b_Ref, |
1507 |
|
const uint8_t * const b_RefH, |
1508 |
|
const uint8_t * const b_RefV, |
1509 |
|
const uint8_t * const b_RefHV, |
1510 |
const IMAGE * const pCur, |
const IMAGE * const pCur, |
1511 |
const int x, const int y, |
const int x, const int y, |
1512 |
const int start_x, int start_y, |
const uint32_t fcode, |
1513 |
|
const uint32_t bcode, |
1514 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1515 |
MBParam * const pParam, |
const MBParam * const pParam, |
1516 |
MACROBLOCK * const pMBs, |
const VECTOR * const f_predMV, |
1517 |
VECTOR * const currMV, |
const VECTOR * const b_predMV, |
1518 |
VECTOR * const currPMV) |
MACROBLOCK * const pMB, |
1519 |
|
int32_t * const best_sad, |
1520 |
|
SearchData * const fData) |
1521 |
|
|
1522 |
{ |
{ |
|
const uint32_t iWcount = pParam->mb_width; |
|
1523 |
|
|
|
const int32_t iFcode = pParam->fixed_code; |
|
|
const int32_t iQuant = pParam->quant; |
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
1524 |
const int32_t iEdgedWidth = pParam->edged_width; |
const int32_t iEdgedWidth = pParam->edged_width; |
1525 |
|
|
1526 |
const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; |
int iDirection, i, j; |
1527 |
|
SearchData bData; |
|
int32_t iDiamondSize; |
|
1528 |
|
|
1529 |
int32_t min_dx; |
bData.iMinSAD = fData->iMinSAD; |
1530 |
int32_t max_dx; |
*bData.iMinSAD = 4096*256; |
1531 |
int32_t min_dy; |
bData.Cur = fData->Cur; |
1532 |
int32_t max_dy; |
fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; |
1533 |
|
bData.currentMV = fData->currentMV + 1; |
1534 |
|
bData.iQuant = fData->iQuant; |
1535 |
|
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
1536 |
|
|
1537 |
|
bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; |
1538 |
|
bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; |
1539 |
|
bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; |
1540 |
|
bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; |
1541 |
|
bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; |
1542 |
|
bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; |
1543 |
|
bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; |
1544 |
|
bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; |
1545 |
|
|
1546 |
|
bData.bpredMV = fData->predMV = *f_predMV; |
1547 |
|
fData->bpredMV = bData.predMV = *b_predMV; |
1548 |
|
|
1549 |
|
fData->currentMV[0] = pMB->mvs[0]; |
1550 |
|
fData->currentMV[1] = pMB->b_mvs[0]; |
1551 |
|
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); |
1552 |
|
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); |
1553 |
|
|
1554 |
|
if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; |
1555 |
|
if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; |
1556 |
|
if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; |
1557 |
|
if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; |
1558 |
|
|
1559 |
|
if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; |
1560 |
|
if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; |
1561 |
|
if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; |
1562 |
|
if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; |
1563 |
|
|
1564 |
|
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
1565 |
|
|
1566 |
|
//diamond. I wish we could use normal mainsearch functions (square, advdiamond) |
1567 |
|
|
1568 |
|
do { |
1569 |
|
iDirection = 255; |
1570 |
|
// forward MV moves |
1571 |
|
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
1572 |
|
|
1573 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
1574 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, fData); |
1575 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
1576 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
1577 |
|
|
1578 |
|
// backward MV moves |
1579 |
|
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
1580 |
|
fData->currentMV[2] = fData->currentMV[0]; |
1581 |
|
|
1582 |
|
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
1583 |
|
CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); |
1584 |
|
CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); |
1585 |
|
CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); |
1586 |
|
|
1587 |
|
} while (!(iDirection)); |
1588 |
|
|
1589 |
|
// two bits are needed to code interpolate mode. we treat the bits just like they were vector's |
1590 |
|
*fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; |
1591 |
|
if (*fData->iMinSAD < *best_sad) { |
1592 |
|
*best_sad = *fData->iMinSAD; |
1593 |
|
pMB->mvs[0] = fData->currentMV[0]; |
1594 |
|
pMB->b_mvs[0] = fData->currentMV[1]; |
1595 |
|
pMB->mode = MODE_INTERPOLATE; |
1596 |
|
|
1597 |
|
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
1598 |
|
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
1599 |
|
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
1600 |
|
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
1601 |
|
} |
1602 |
|
} |
1603 |
|
|
1604 |
|
|
1605 |
|
void |
1606 |
|
MotionEstimationBVOP(MBParam * const pParam, |
1607 |
|
FRAMEINFO * const frame, |
1608 |
|
const int32_t time_bp, |
1609 |
|
const int32_t time_pp, |
1610 |
|
// forward (past) reference |
1611 |
|
const MACROBLOCK * const f_mbs, |
1612 |
|
const IMAGE * const f_ref, |
1613 |
|
const IMAGE * const f_refH, |
1614 |
|
const IMAGE * const f_refV, |
1615 |
|
const IMAGE * const f_refHV, |
1616 |
|
// backward (future) reference |
1617 |
|
const MACROBLOCK * const b_mbs, |
1618 |
|
const IMAGE * const b_ref, |
1619 |
|
const IMAGE * const b_refH, |
1620 |
|
const IMAGE * const b_refV, |
1621 |
|
const IMAGE * const b_refHV) |
1622 |
|
{ |
1623 |
|
uint32_t i, j; |
1624 |
|
int32_t best_sad, skip_sad; |
1625 |
|
int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; |
1626 |
|
static const VECTOR zeroMV={0,0}; |
1627 |
|
|
1628 |
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
1629 |
|
|
1630 |
|
const int32_t TRB = time_pp - time_bp; |
1631 |
|
const int32_t TRD = time_pp; |
1632 |
|
|
1633 |
|
// some pre-inintialized data for the rest of the search |
1634 |
|
|
1635 |
|
SearchData Data; |
1636 |
|
int32_t iMinSAD; |
1637 |
|
VECTOR currentMV[3]; |
1638 |
|
Data.iEdgedWidth = pParam->edged_width; |
1639 |
|
Data.currentMV = currentMV; |
1640 |
|
Data.iMinSAD = &iMinSAD; |
1641 |
|
Data.iQuant = frame->quant; |
1642 |
|
|
1643 |
VECTOR pmv[4]; |
// note: i==horizontal, j==vertical |
|
int32_t psad[4]; |
|
|
VECTOR newMV; |
|
|
VECTOR backupMV; |
|
1644 |
|
|
1645 |
MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
for (j = 0; j < pParam->mb_height; j++) { |
1646 |
|
|
1647 |
static int32_t threshA,threshB; |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
int32_t iFound,bPredEq; |
|
|
int32_t iMinSAD,iSAD; |
|
1648 |
|
|
1649 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
for (i = 0; i < pParam->mb_width; i++) { |
1650 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
1651 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
1652 |
|
|
1653 |
/* Get maximum range */ |
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
1654 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
if (b_mb->mode == MODE_NOT_CODED) { |
1655 |
x, y, 8, iWidth, iHeight, iFcode); |
pMB->mode = MODE_NOT_CODED; |
1656 |
|
continue; |
1657 |
|
} |
1658 |
|
|
1659 |
/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ |
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
1660 |
|
/* direct search comes first, because it (1) checks for SKIP-mode |
1661 |
|
and (2) sets very good predictions for forward and backward search */ |
1662 |
|
|
1663 |
|
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
1664 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
1665 |
|
&frame->image, |
1666 |
|
i, j, |
1667 |
|
frame->motion_flags, |
1668 |
|
TRB, TRD, |
1669 |
|
pParam, |
1670 |
|
pMB, b_mb, |
1671 |
|
&best_sad, |
1672 |
|
&Data); |
1673 |
|
|
1674 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
1675 |
|
|
1676 |
|
// best_sad = 256*4096; //uncomment to disable Directsearch. |
1677 |
|
// To disable any other mode, just comment the function call |
1678 |
|
|
1679 |
|
// forward search |
1680 |
|
SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1681 |
|
&frame->image, i, j, |
1682 |
|
frame->motion_flags, |
1683 |
|
frame->fcode, pParam, |
1684 |
|
pMB, &f_predMV, &best_sad, |
1685 |
|
MODE_FORWARD, &Data); |
1686 |
|
|
1687 |
|
// backward search |
1688 |
|
SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1689 |
|
&frame->image, i, j, |
1690 |
|
frame->motion_flags, |
1691 |
|
frame->bcode, pParam, |
1692 |
|
pMB, &b_predMV, &best_sad, |
1693 |
|
MODE_BACKWARD, &Data); |
1694 |
|
|
1695 |
|
// interpolate search comes last, because it uses data from forward and backward as prediction |
1696 |
|
|
1697 |
|
SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
1698 |
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
1699 |
|
&frame->image, |
1700 |
|
i, j, |
1701 |
|
frame->fcode, frame->bcode, |
1702 |
|
frame->motion_flags, |
1703 |
|
pParam, |
1704 |
|
&f_predMV, &b_predMV, |
1705 |
|
pMB, &best_sad, |
1706 |
|
&Data); |
1707 |
|
|
1708 |
|
switch (pMB->mode) { |
1709 |
|
case MODE_FORWARD: |
1710 |
|
f_count++; |
1711 |
|
f_predMV = pMB->mvs[0]; |
1712 |
|
break; |
1713 |
|
case MODE_BACKWARD: |
1714 |
|
b_count++; |
1715 |
|
b_predMV = pMB->b_mvs[0]; |
1716 |
|
break; |
1717 |
|
case MODE_INTERPOLATE: |
1718 |
|
i_count++; |
1719 |
|
f_predMV = pMB->mvs[0]; |
1720 |
|
b_predMV = pMB->b_mvs[0]; |
1721 |
|
break; |
1722 |
|
case MODE_DIRECT: |
1723 |
|
case MODE_DIRECT_NO4V: |
1724 |
|
d_count++; |
1725 |
|
break; |
1726 |
|
default: |
1727 |
|
break; |
1728 |
|
} |
1729 |
|
} |
1730 |
|
} |
1731 |
|
|
1732 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) |
// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", |
1733 |
{ min_dx = EVEN(min_dx); |
// f_count,b_count,i_count,d_count,n_count); |
|
max_dx = EVEN(max_dx); |
|
|
min_dy = EVEN(min_dy); |
|
|
max_dy = EVEN(max_dy); |
|
|
} /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ |
|
1734 |
|
|
1735 |
|
} |
1736 |
|
|
1737 |
bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); |
/* Hinted ME starts here */ |
1738 |
|
|
1739 |
if ((x==0) && (y==0) ) |
static __inline void |
1740 |
|
Search8hinted( const SearchData * const OldData, |
1741 |
|
const int x, const int y, |
1742 |
|
const uint32_t MotionFlags, |
1743 |
|
const MBParam * const pParam, |
1744 |
|
MACROBLOCK * const pMB, |
1745 |
|
const MACROBLOCK * const pMBs, |
1746 |
|
const int block) |
1747 |
{ |
{ |
1748 |
threshA = 512/4; |
SearchData Data; |
1749 |
threshB = 1024/4; |
MainSearchFunc *MainSearchPtr; |
1750 |
|
|
1751 |
|
Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); |
1752 |
|
Data.iMinSAD = OldData->iMinSAD + 1 + block; |
1753 |
|
Data.currentMV = OldData->currentMV+1+block; |
1754 |
|
Data.iFcode = OldData->iFcode; |
1755 |
|
Data.iQuant = OldData->iQuant; |
1756 |
|
|
1757 |
|
Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1758 |
|
Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1759 |
|
Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1760 |
|
Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1761 |
|
Data.iEdgedWidth = pParam->edged_width; |
1762 |
|
Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); |
1763 |
|
|
1764 |
|
CheckCandidate = CheckCandidate8; |
1765 |
|
|
1766 |
|
if (block != 0) |
1767 |
|
*(Data.iMinSAD) += lambda_vec8[Data.iQuant] * |
1768 |
|
d_mv_bits( Data.currentMV->x - Data.predMV.x, |
1769 |
|
Data.currentMV->y - Data.predMV.y, |
1770 |
|
Data.iFcode); |
1771 |
|
|
1772 |
|
|
1773 |
|
get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, |
1774 |
|
pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); |
1775 |
|
|
1776 |
|
if (pMB->mode == MODE_INTER4V) { |
1777 |
|
int dummy; |
1778 |
|
CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } |
1779 |
|
|
1780 |
|
if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; |
1781 |
|
else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; |
1782 |
|
else MainSearchPtr = DiamondSearch; |
1783 |
|
|
1784 |
|
(*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); |
1785 |
|
|
1786 |
|
if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); |
1787 |
|
|
1788 |
|
pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; |
1789 |
|
pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; |
1790 |
|
pMB->mvs[block] = *(Data.currentMV); |
1791 |
|
pMB->sad8[block] = 4 * (*(Data.iMinSAD)); |
1792 |
} |
} |
1793 |
else |
|
1794 |
|
|
1795 |
|
static void |
1796 |
|
SearchPhinted ( const uint8_t * const pRef, |
1797 |
|
const uint8_t * const pRefH, |
1798 |
|
const uint8_t * const pRefV, |
1799 |
|
const uint8_t * const pRefHV, |
1800 |
|
const IMAGE * const pCur, |
1801 |
|
const int x, |
1802 |
|
const int y, |
1803 |
|
const uint32_t MotionFlags, |
1804 |
|
const uint32_t iQuant, |
1805 |
|
const MBParam * const pParam, |
1806 |
|
const MACROBLOCK * const pMBs, |
1807 |
|
int inter4v, |
1808 |
|
MACROBLOCK * const pMB, |
1809 |
|
SearchData * const Data) |
1810 |
{ |
{ |
|
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; |
|
|
} |
|
1811 |
|
|
1812 |
iFound=0; |
const int32_t iEdgedWidth = pParam->edged_width; |
1813 |
|
|
1814 |
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
int i, t; |
1815 |
vector of the median. |
MainSearchFunc * MainSearchPtr; |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
1816 |
|
|
1817 |
if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) |
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1818 |
iFound=2; |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1819 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
1820 |
|
|
1821 |
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; |
1822 |
Otherwise select large Diamond Search. |
Data->Ref = pRef + (x + iEdgedWidth*y)*16; |
1823 |
*/ |
Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; |
1824 |
|
Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; |
1825 |
|
Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; |
1826 |
|
Data->iQuant = iQuant; |
1827 |
|
|
1828 |
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
if (!(MotionFlags & PMV_HALFPEL16)) { |
1829 |
iDiamondSize=1; // 1 halfpel! |
Data->min_dx = EVEN(Data->min_dx); |
1830 |
else |
Data->max_dx = EVEN(Data->max_dx); |
1831 |
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
Data->min_dy = EVEN(Data->min_dy); |
1832 |
|
Data->max_dy = EVEN(Data->max_dy); |
1833 |
|
} |
1834 |
|
|
1835 |
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
|
iDiamondSize*=2; |
|
1836 |
|
|
1837 |
/* Step 4: Calculate SAD around the Median prediction. |
if (pMB->dquant != NO_CHANGE) inter4v = 0; |
|
MinSAD=SAD |
|
|
If Motion Vector equal to Previous frame motion vector |
|
|
and MinSAD<PrevFrmSAD goto Step 10. |
|
|
If SAD<=256 goto Step 10. |
|
|
*/ |
|
1838 |
|
|
1839 |
|
if (inter4v) |
1840 |
|
CheckCandidate = CheckCandidate16; |
1841 |
|
else CheckCandidate = CheckCandidate16no4v; |
1842 |
|
|
|
// Prepare for main loop |
|
1843 |
|
|
1844 |
currMV->x=start_x; /* start with mv16 */ |
pMB->mvs[0].x = EVEN(pMB->mvs[0].x); |
1845 |
currMV->y=start_y; |
pMB->mvs[0].y = EVEN(pMB->mvs[0].y); |
1846 |
|
if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed |
1847 |
|
if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; |
1848 |
|
if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; |
1849 |
|
if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; |
1850 |
|
|
1851 |
iMinSAD = sad8( cur, |
(*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); |
|
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
|
|
iEdgedWidth); |
|
|
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; |
|
1852 |
|
|
1853 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
if (pMB->mode == MODE_INTER4V) |
1854 |
{ |
for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search |
1855 |
if (MotionFlags & PMV_QUICKSTOP8) |
pMB->mvs[i].x = EVEN(pMB->mvs[i].x); |
1856 |
goto step10_8b; |
pMB->mvs[i].y = EVEN(pMB->mvs[i].y); |
1857 |
if (MotionFlags & PMV_EARLYSTOP8) |
if (!(make_mask(pMB->mvs, i))) |
1858 |
goto step10_8; |
(*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); |
1859 |
} |
} |
1860 |
|
|
1861 |
/* |
if (MotionFlags & PMV_USESQUARES16) |
1862 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
MainSearchPtr = SquareSearch; |
1863 |
Also calculate (0,0) but do not subtract offset. |
else if (MotionFlags & PMV_ADVANCEDDIAMOND16) |
1864 |
Let MinSAD be the smallest SAD up to this point. |
MainSearchPtr = AdvDiamondSearch; |
1865 |
If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** |
else MainSearchPtr = DiamondSearch; |
|
*/ |
|
1866 |
|
|
1867 |
// the prediction might be even better than mv16 |
(*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
|
1868 |
|
|
1869 |
// (0,0) is always possible |
if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); |
|
CHECK_MV8_ZERO; |
|
1870 |
|
|
1871 |
// previous frame MV is always possible |
if (inter4v) |
1872 |
CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); |
for(i = 0; i < 4; i++) |
1873 |
|
Search8hinted(Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); |
1874 |
|
|
1875 |
// left neighbour, if allowed |
if (!(inter4v) || |
1876 |
if (psad[1] != MV_MAX_ERROR) |
(Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + |
1877 |
{ |
Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { |
1878 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
// INTER MODE |
1879 |
{ pmv[1].x = EVEN(pmv[1].x); |
|
1880 |
pmv[1].y = EVEN(pmv[1].y); |
pMB->mode = MODE_INTER; |
1881 |
|
pMB->mvs[0] = pMB->mvs[1] |
1882 |
|
= pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1883 |
|
|
1884 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = |
1885 |
|
pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
1886 |
|
|
1887 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1888 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1889 |
|
} else { |
1890 |
|
// INTER4V MODE; all other things are already set in Search8hinted |
1891 |
|
pMB->mode = MODE_INTER4V; |
1892 |
|
pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] |
1893 |
|
+ Data->iMinSAD[4] + IMV16X16 * iQuant; |
1894 |
} |
} |
1895 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
|
1896 |
} |
} |
1897 |
|
|
1898 |
// top neighbour, if allowed |
void |
1899 |
if (psad[2] != MV_MAX_ERROR) |
MotionEstimationHinted( MBParam * const pParam, |
1900 |
|
FRAMEINFO * const current, |
1901 |
|
FRAMEINFO * const reference, |
1902 |
|
const IMAGE * const pRefH, |
1903 |
|
const IMAGE * const pRefV, |
1904 |
|
const IMAGE * const pRefHV) |
1905 |
{ |
{ |
1906 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
MACROBLOCK *const pMBs = current->mbs; |
1907 |
{ pmv[2].x = EVEN(pmv[2].x); |
const IMAGE *const pCurrent = ¤t->image; |
1908 |
pmv[2].y = EVEN(pmv[2].y); |
const IMAGE *const pRef = &reference->image; |
1909 |
|
|
1910 |
|
uint32_t x, y; |
1911 |
|
int32_t temp[5], quant = current->quant; |
1912 |
|
int32_t iMinSAD[5]; |
1913 |
|
VECTOR currentMV[5]; |
1914 |
|
SearchData Data; |
1915 |
|
Data.iEdgedWidth = pParam->edged_width; |
1916 |
|
Data.currentMV = currentMV; |
1917 |
|
Data.iMinSAD = iMinSAD; |
1918 |
|
Data.temp = temp; |
1919 |
|
Data.iFcode = current->fcode; |
1920 |
|
|
1921 |
|
if (sadInit) (*sadInit) (); |
1922 |
|
|
1923 |
|
for (y = 0; y < pParam->mb_height; y++) { |
1924 |
|
for (x = 0; x < pParam->mb_width; x++) { |
1925 |
|
|
1926 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
1927 |
|
|
1928 |
|
//intra mode is copied from the first pass. At least for the time being |
1929 |
|
if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; |
1930 |
|
|
1931 |
|
|
1932 |
|
if (!(current->global_flags & XVID_LUMIMASKING)) { |
1933 |
|
pMB->dquant = NO_CHANGE; |
1934 |
|
pMB->quant = current->quant; } |
1935 |
|
else |
1936 |
|
if (pMB->dquant != NO_CHANGE) { |
1937 |
|
quant += DQtab[pMB->dquant]; |
1938 |
|
if (quant > 31) quant = 31; |
1939 |
|
else if (quant < 1) quant = 1; |
1940 |
|
pMB->quant = quant; |
1941 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
|
1942 |
|
|
1943 |
// top right neighbour, if allowed |
SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1944 |
if (psad[3] != MV_MAX_ERROR) |
y, current->motion_flags, pMB->quant, |
1945 |
{ |
pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, |
1946 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
&Data); |
1947 |
{ pmv[3].x = EVEN(pmv[3].x); |
|
|
pmv[3].y = EVEN(pmv[3].y); |
|
1948 |
} |
} |
|
CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); |
|
1949 |
} |
} |
1950 |
} |
} |
1951 |
|
|
1952 |
/* Step 6: If MinSAD <= thresa goto Step 10. |
static __inline int |
1953 |
If Motion Vector equal to Previous frame motion vector and MinSAD<PrevFrmSAD goto Step 10. |
MEanalyzeMB ( const uint8_t * const pRef, |
1954 |
*/ |
const uint8_t * const pCur, |
1955 |
|
const int x, |
1956 |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,pMB->mvs[iSubBlock]) && (iMinSAD < pMB->sad8[iSubBlock]) ) ) |
const int y, |
1957 |
|
const MBParam * const pParam, |
1958 |
|
const MACROBLOCK * const pMBs, |
1959 |
|
MACROBLOCK * const pMB, |
1960 |
|
SearchData * const Data) |
1961 |
{ |
{ |
|
if (MotionFlags & PMV_QUICKSTOP8) |
|
|
goto step10_8b; |
|
|
if (MotionFlags & PMV_EARLYSTOP8) |
|
|
goto step10_8; |
|
|
} |
|
1962 |
|
|
1963 |
/************ (Diamond Search) **************/ |
int i, mask; |
1964 |
/* |
VECTOR pmv[3]; |
|
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. |
|
|
*/ |
|
1965 |
|
|
1966 |
backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ |
*(Data->iMinSAD) = MV_MAX_ERROR; |
1967 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
1968 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1969 |
|
pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); |
1970 |
|
|
1971 |
/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ |
Data->Cur = pCur + (x + y * pParam->edged_width) * 16; |
1972 |
iSAD = PMVfastSearch8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
Data->Ref = pRef + (x + y * pParam->edged_width) * 16; |
|
x, y, |
|
|
currMV->x, currMV->y, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
1973 |
|
|
1974 |
if (iSAD < iMinSAD) |
CheckCandidate = CheckCandidate16no4vI; |
1975 |
{ |
|
1976 |
*currMV = newMV; |
pmv[1].x = EVEN(pMB->mvs[0].x); |
1977 |
iMinSAD = iSAD; |
pmv[1].y = EVEN(pMB->mvs[0].y); |
1978 |
|
pmv[0].x = EVEN(Data->predMV.x); |
1979 |
|
pmv[0].y = EVEN(Data->predMV.y); |
1980 |
|
pmv[2].x = pmv[2].y = 0; |
1981 |
|
|
1982 |
|
CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); |
1983 |
|
if (!(mask = make_mask(pmv, 1))) |
1984 |
|
CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); |
1985 |
|
if (!(mask = make_mask(pmv, 2))) |
1986 |
|
CheckCandidate16no4vI(0, 0, mask, &i, Data); |
1987 |
|
|
1988 |
|
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
1989 |
|
|
1990 |
|
pMB->mvs[0] = pMB->mvs[1] |
1991 |
|
= pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() |
1992 |
|
|
1993 |
|
return *(Data->iMinSAD); |
1994 |
} |
} |
1995 |
|
|
1996 |
if (MotionFlags & PMV_EXTSEARCH8) |
#define INTRA_THRESH 1350 |
1997 |
|
#define INTER_THRESH 900 |
1998 |
|
|
1999 |
|
int |
2000 |
|
MEanalysis( const IMAGE * const pRef, |
2001 |
|
const IMAGE * const pCurrent, |
2002 |
|
MBParam * const pParam, |
2003 |
|
MACROBLOCK * const pMBs, |
2004 |
|
const uint32_t iFcode) |
2005 |
{ |
{ |
2006 |
/* extended: search (up to) two more times: orignal prediction and (0,0) */ |
uint32_t x, y, intra = 0; |
2007 |
|
int sSAD = 0; |
2008 |
|
|
2009 |
if (!(MVequal(pmv[0],backupMV)) ) |
VECTOR currentMV; |
2010 |
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
int32_t iMinSAD; |
2011 |
x, y, |
SearchData Data; |
2012 |
pmv[0].x, pmv[0].y, iMinSAD, &newMV, |
Data.iEdgedWidth = pParam->edged_width; |
2013 |
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
Data.currentMV = ¤tMV; |
2014 |
|
Data.iMinSAD = &iMinSAD; |
2015 |
|
Data.iFcode = iFcode; |
2016 |
|
Data.iQuant = 2; |
2017 |
|
|
2018 |
if (iSAD < iMinSAD) |
if (sadInit) (*sadInit) (); |
2019 |
{ |
|
2020 |
*currMV = newMV; |
for (y = 0; y < pParam->mb_height-1; y++) { |
2021 |
iMinSAD = iSAD; |
for (x = 0; x < pParam->mb_width; x++) { |
2022 |
|
int sad, dev; |
2023 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
2024 |
|
|
2025 |
|
sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, |
2026 |
|
pParam, pMBs, pMB, &Data); |
2027 |
|
|
2028 |
|
if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work |
2029 |
|
if (sad > INTRA_THRESH) { |
2030 |
|
dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, |
2031 |
|
pParam->edged_width); |
2032 |
|
if (dev + INTRA_THRESH < sad) intra++; |
2033 |
|
if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame |
2034 |
} |
} |
2035 |
|
sSAD += sad; |
2036 |
} |
} |
2037 |
|
|
|
if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) |
|
|
{ iSAD = PMVfastSearch16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, |
|
|
x, y, |
|
|
0, 0, iMinSAD, &newMV, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); |
|
|
|
|
|
if (iSAD < iMinSAD) |
|
|
{ |
|
|
*currMV = newMV; |
|
|
iMinSAD = iSAD; |
|
2038 |
} |
} |
2039 |
} |
} |
2040 |
|
sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); |
2041 |
|
if (sSAD > INTER_THRESH ) return 1; //P frame |
2042 |
|
emms(); |
2043 |
|
return 0; // B frame |
2044 |
|
|
2045 |
} |
} |
2046 |
|
|
2047 |
/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. |
int |
2048 |
By performing an optional local half-pixel search, we can refine this result even further. |
FindFcode( const MBParam * const pParam, |
2049 |
*/ |
const FRAMEINFO * const current) |
2050 |
|
{ |
2051 |
|
uint32_t x, y; |
2052 |
|
int max = 0, min = 0, i; |
2053 |
|
|
2054 |
|
for (y = 0; y < pParam->mb_height; y++) { |
2055 |
|
for (x = 0; x < pParam->mb_width; x++) { |
2056 |
|
|
2057 |
step10_8: |
MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; |
2058 |
if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step |
for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { |
2059 |
iMinSAD = PMVfastSearch8_Refine( pRef, pRefH, pRefV, pRefHV, cur, |
if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; |
2060 |
x, y, |
if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; |
|
currMV, iMinSAD, |
|
|
pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); |
|
2061 |
|
|
2062 |
step10_8b: |
if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; |
2063 |
|
if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; |
2064 |
|
} |
2065 |
|
} |
2066 |
|
} |
2067 |
|
|
2068 |
currPMV->x = currMV->x - pmv[0].x; |
min = -min; |
2069 |
currPMV->y = currMV->y - pmv[0].y; |
max += 1; |
2070 |
|
if (min > max) max = min; |
2071 |
|
|
2072 |
return iMinSAD; |
for (i = 1; (max > 32 << (i - 1)); i++); |
2073 |
|
return i; |
2074 |
} |
} |