?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
r1009: Move the dependencies to newer package names
[cinelerra_cv/mob.git] / quicktime / ffmpeg / libavcodec / motion_est_template.c
blobd8101ec3345a824d756efbfad8821869a567cd2c
1 /*
2 * Motion estimation
3 * Copyright (c) 2002-2004 Michael Niedermayer
4 *
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
9 *
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
14 *
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 *
19 */
20
21 /**
22 * @file motion_est_template.c
23 * Motion estimation template.
24 */
25
26 //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
27 #define LOAD_COMMON\
28 uint32_t attribute_unused * const score_map= c->score_map;\
29 const int attribute_unused xmin= c->xmin;\
30 const int attribute_unused ymin= c->ymin;\
31 const int attribute_unused xmax= c->xmax;\
32 const int attribute_unused ymax= c->ymax;\
33 uint8_t *mv_penalty= c->current_mv_penalty;\
34 const int pred_x= c->pred_x;\
35 const int pred_y= c->pred_y;\
36
37 #define CHECK_HALF_MV(dx, dy, x, y)\
38 {\
39 const int hx= 2*(x)+(dx);\
40 const int hy= 2*(y)+(dy);\
41 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
42 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
43 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
44 }
45
46 #if 0
47 static int hpel_motion_search)(MpegEncContext * s,
48 int *mx_ptr, int *my_ptr, int dmin,
49 uint8_t *ref_data[3],
50 int size)
51 {
52 const int xx = 16 * s->mb_x + 8*(n&1);
53 const int yy = 16 * s->mb_y + 8*(n>>1);
54 const int mx = *mx_ptr;
55 const int my = *my_ptr;
56 const int penalty_factor= c->sub_penalty_factor;
57
58 LOAD_COMMON
59
60 // INIT;
61 //FIXME factorize
62 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;
63
64 if(s->no_rounding /*FIXME b_type*/){
65 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];
66 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];
67 }else{
68 hpel_put=& s->dsp.put_pixels_tab[size];
69 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
70 }
71 cmpf= s->dsp.me_cmp[size];
72 chroma_cmpf= s->dsp.me_cmp[size+1];
73 cmp_sub= s->dsp.me_sub_cmp[size];
74 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
75
76 if(c->skip){ //FIXME somehow move up (benchmark)
77 *mx_ptr = 0;
78 *my_ptr = 0;
79 return dmin;
80 }
81
82 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
83 CMP_HPEL(dmin, 0, 0, mx, my, size);
84 if(mx || my)
85 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
86 }
87
88 if (mx > xmin && mx < xmax &&
89 my > ymin && my < ymax) {
90 int bx=2*mx, by=2*my;
91 int d= dmin;
92
93 CHECK_HALF_MV(1, 1, mx-1, my-1)
94 CHECK_HALF_MV(0, 1, mx , my-1)
95 CHECK_HALF_MV(1, 1, mx , my-1)
96 CHECK_HALF_MV(1, 0, mx-1, my )
97 CHECK_HALF_MV(1, 0, mx , my )
98 CHECK_HALF_MV(1, 1, mx-1, my )
99 CHECK_HALF_MV(0, 1, mx , my )
100 CHECK_HALF_MV(1, 1, mx , my )
101
102 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);
103
104 *mx_ptr = bx;
105 *my_ptr = by;
106 }else{
107 *mx_ptr =2*mx;
108 *my_ptr =2*my;
109 }
110
111 return dmin;
112 }
113
114 #else
115 static int hpel_motion_search(MpegEncContext * s,
116 int *mx_ptr, int *my_ptr, int dmin,
117 int src_index, int ref_index,
118 int size, int h)
119 {
120 MotionEstContext * const c= &s->me;
121 const int mx = *mx_ptr;
122 const int my = *my_ptr;
123 const int penalty_factor= c->sub_penalty_factor;
124 me_cmp_func cmp_sub, chroma_cmp_sub;
125 int bx=2*mx, by=2*my;
126
127 LOAD_COMMON
128 int flags= c->sub_flags;
129
130 //FIXME factorize
131
132 cmp_sub= s->dsp.me_sub_cmp[size];
133 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
134
135 if(c->skip){ //FIXME move out of hpel?
136 *mx_ptr = 0;
137 *my_ptr = 0;
138 return dmin;
139 }
140
141 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
142 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
143 if(mx || my || size>0)
144 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
145 }
146
147 if (mx > xmin && mx < xmax &&
148 my > ymin && my < ymax) {
149 int d= dmin;
150 const int index= (my<<ME_MAP_SHIFT) + mx;
151 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
152 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;
153 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]
154 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
155 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]
156 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
157 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
158 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;
159
160 #if 1
161 int key;
162 int map_generation= c->map_generation;
163 #ifndef NDEBUG
164 uint32_t *map= c->map;
165 #endif
166 key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
167 assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
168 key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
169 assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
170 key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;
171 assert(map[(index+1)&(ME_MAP_SIZE-1)] == key);
172 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
173 assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);
174 #endif
175 if(t<=b){
176 CHECK_HALF_MV(0, 1, mx ,my-1)
177 if(l<=r){
178 CHECK_HALF_MV(1, 1, mx-1, my-1)
179 if(t+r<=b+l){
180 CHECK_HALF_MV(1, 1, mx , my-1)
181 }else{
182 CHECK_HALF_MV(1, 1, mx-1, my )
183 }
184 CHECK_HALF_MV(1, 0, mx-1, my )
185 }else{
186 CHECK_HALF_MV(1, 1, mx , my-1)
187 if(t+l<=b+r){
188 CHECK_HALF_MV(1, 1, mx-1, my-1)
189 }else{
190 CHECK_HALF_MV(1, 1, mx , my )
191 }
192 CHECK_HALF_MV(1, 0, mx , my )
193 }
194 }else{
195 if(l<=r){
196 if(t+l<=b+r){
197 CHECK_HALF_MV(1, 1, mx-1, my-1)
198 }else{
199 CHECK_HALF_MV(1, 1, mx , my )
200 }
201 CHECK_HALF_MV(1, 0, mx-1, my)
202 CHECK_HALF_MV(1, 1, mx-1, my)
203 }else{
204 if(t+r<=b+l){
205 CHECK_HALF_MV(1, 1, mx , my-1)
206 }else{
207 CHECK_HALF_MV(1, 1, mx-1, my)
208 }
209 CHECK_HALF_MV(1, 0, mx , my)
210 CHECK_HALF_MV(1, 1, mx , my)
211 }
212 CHECK_HALF_MV(0, 1, mx , my)
213 }
214 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
215 }
216
217 *mx_ptr = bx;
218 *my_ptr = by;
219
220 return dmin;
221 }
222 #endif
223
224 static int no_sub_motion_search(MpegEncContext * s,
225 int *mx_ptr, int *my_ptr, int dmin,
226 int src_index, int ref_index,
227 int size, int h)
228 {
229 (*mx_ptr)<<=1;
230 (*my_ptr)<<=1;
231 return dmin;
232 }
233
234 int inline ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
235 int ref_index, int size, int h, int add_rate)
236 {
237 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
238 MotionEstContext * const c= &s->me;
239 const int penalty_factor= c->mb_penalty_factor;
240 const int flags= c->mb_flags;
241 const int qpel= flags & FLAG_QPEL;
242 const int mask= 1+2*qpel;
243 me_cmp_func cmp_sub, chroma_cmp_sub;
244 int d;
245
246 LOAD_COMMON
247
248 //FIXME factorize
249
250 cmp_sub= s->dsp.mb_cmp[size];
251 chroma_cmp_sub= s->dsp.mb_cmp[size+1];
252
253 // assert(!c->skip);
254 // assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);
255
256 d= cmp(s, mx>>(qpel+1), my>>(qpel+1), mx&mask, my&mask, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
257 //FIXME check cbp before adding penalty for (0,0) vector
258 if(add_rate && (mx || my || size>0))
259 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
260
261 return d;
262 }
263
264 #define CHECK_QUARTER_MV(dx, dy, x, y)\
265 {\
266 const int hx= 4*(x)+(dx);\
267 const int hy= 4*(y)+(dy);\
268 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
269 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
270 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
271 }
272
273 static int qpel_motion_search(MpegEncContext * s,
274 int *mx_ptr, int *my_ptr, int dmin,
275 int src_index, int ref_index,
276 int size, int h)
277 {
278 MotionEstContext * const c= &s->me;
279 const int mx = *mx_ptr;
280 const int my = *my_ptr;
281 const int penalty_factor= c->sub_penalty_factor;
282 const int map_generation= c->map_generation;
283 const int subpel_quality= c->avctx->me_subpel_quality;
284 uint32_t *map= c->map;
285 me_cmp_func cmpf, chroma_cmpf;
286 me_cmp_func cmp_sub, chroma_cmp_sub;
287
288 LOAD_COMMON
289 int flags= c->sub_flags;
290
291 cmpf= s->dsp.me_cmp[size];
292 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
293 //FIXME factorize
294
295 cmp_sub= s->dsp.me_sub_cmp[size];
296 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
297
298 if(c->skip){ //FIXME somehow move up (benchmark)
299 *mx_ptr = 0;
300 *my_ptr = 0;
301 return dmin;
302 }
303
304 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
305 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
306 if(mx || my || size>0)
307 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
308 }
309
310 if (mx > xmin && mx < xmax &&
311 my > ymin && my < ymax) {
312 int bx=4*mx, by=4*my;
313 int d= dmin;
314 int i, nx, ny;
315 const int index= (my<<ME_MAP_SHIFT) + mx;
316 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
317 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
318 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
319 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
320 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
321 int best[8];
322 int best_pos[8][2];
323
324 memset(best, 64, sizeof(int)*8);
325 #if 1
326 if(s->me.dia_size>=2){
327 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
328 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
329 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
330 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
331
332 for(ny= -3; ny <= 3; ny++){
333 for(nx= -3; nx <= 3; nx++){
334 //FIXME this could overflow (unlikely though)
335 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
336 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
337 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
338 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
339 int i;
340
341 if((nx&3)==0 && (ny&3)==0) continue;
342
343 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
344
345 // if(nx&1) score-=1024*c->penalty_factor;
346 // if(ny&1) score-=1024*c->penalty_factor;
347
348 for(i=0; i<8; i++){
349 if(score < best[i]){
350 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
351 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
352 best[i]= score;
353 best_pos[i][0]= nx + 4*mx;
354 best_pos[i][1]= ny + 4*my;
355 break;
356 }
357 }
358 }
359 }
360 }else{
361 int tl;
362 //FIXME this could overflow (unlikely though)
363 const int cx = 4*(r - l);
364 const int cx2= r + l - 2*c;
365 const int cy = 4*(b - t);
366 const int cy2= b + t - 2*c;
367 int cxy;
368
369 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
370 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
371 }else{
372 tl= cmp(s, mx-1, my-1, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);//FIXME wrong if chroma me is different
373 }
374
375 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
376
377 assert(16*cx2 + 4*cx + 32*c == 32*r);
378 assert(16*cx2 - 4*cx + 32*c == 32*l);
379 assert(16*cy2 + 4*cy + 32*c == 32*b);
380 assert(16*cy2 - 4*cy + 32*c == 32*t);
381 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
382
383 for(ny= -3; ny <= 3; ny++){
384 for(nx= -3; nx <= 3; nx++){
385 //FIXME this could overflow (unlikely though)
386 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
387 int i;
388
389 if((nx&3)==0 && (ny&3)==0) continue;
390
391 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
392 // if(nx&1) score-=32*c->penalty_factor;
393 // if(ny&1) score-=32*c->penalty_factor;
394
395 for(i=0; i<8; i++){
396 if(score < best[i]){
397 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
398 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
399 best[i]= score;
400 best_pos[i][0]= nx + 4*mx;
401 best_pos[i][1]= ny + 4*my;
402 break;
403 }
404 }
405 }
406 }
407 }
408 for(i=0; i<subpel_quality; i++){
409 nx= best_pos[i][0];
410 ny= best_pos[i][1];
411 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
412 }
413
414 #if 0
415 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
416 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
417 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
418 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
419 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
420 if(tl<br){
421
422 // nx= FFMAX(4*mx - bx, bx - 4*mx);
423 // ny= FFMAX(4*my - by, by - 4*my);
424
425 static int stats[7][7], count;
426 count++;
427 stats[4*mx - bx + 3][4*my - by + 3]++;
428 if(256*256*256*64 % count ==0){
429 for(i=0; i<49; i++){
430 if((i%7)==0) printf("\n");
431 printf("%6d ", stats[0][i]);
432 }
433 printf("\n");
434 }
435 }
436 #endif
437 #else
438
439 CHECK_QUARTER_MV(2, 2, mx-1, my-1)
440 CHECK_QUARTER_MV(0, 2, mx , my-1)
441 CHECK_QUARTER_MV(2, 2, mx , my-1)
442 CHECK_QUARTER_MV(2, 0, mx , my )
443 CHECK_QUARTER_MV(2, 2, mx , my )
444 CHECK_QUARTER_MV(0, 2, mx , my )
445 CHECK_QUARTER_MV(2, 2, mx-1, my )
446 CHECK_QUARTER_MV(2, 0, mx-1, my )
447
448 nx= bx;
449 ny= by;
450
451 for(i=0; i<8; i++){
452 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
453 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
454 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)
455 }
456 #endif
457 #if 0
458 //outer ring
459 CHECK_QUARTER_MV(1, 3, mx-1, my-1)
460 CHECK_QUARTER_MV(1, 2, mx-1, my-1)
461 CHECK_QUARTER_MV(1, 1, mx-1, my-1)
462 CHECK_QUARTER_MV(2, 1, mx-1, my-1)
463 CHECK_QUARTER_MV(3, 1, mx-1, my-1)
464 CHECK_QUARTER_MV(0, 1, mx , my-1)
465 CHECK_QUARTER_MV(1, 1, mx , my-1)
466 CHECK_QUARTER_MV(2, 1, mx , my-1)
467 CHECK_QUARTER_MV(3, 1, mx , my-1)
468 CHECK_QUARTER_MV(3, 2, mx , my-1)
469 CHECK_QUARTER_MV(3, 3, mx , my-1)
470 CHECK_QUARTER_MV(3, 0, mx , my )
471 CHECK_QUARTER_MV(3, 1, mx , my )
472 CHECK_QUARTER_MV(3, 2, mx , my )
473 CHECK_QUARTER_MV(3, 3, mx , my )
474 CHECK_QUARTER_MV(2, 3, mx , my )
475 CHECK_QUARTER_MV(1, 3, mx , my )
476 CHECK_QUARTER_MV(0, 3, mx , my )
477 CHECK_QUARTER_MV(3, 3, mx-1, my )
478 CHECK_QUARTER_MV(2, 3, mx-1, my )
479 CHECK_QUARTER_MV(1, 3, mx-1, my )
480 CHECK_QUARTER_MV(1, 2, mx-1, my )
481 CHECK_QUARTER_MV(1, 1, mx-1, my )
482 CHECK_QUARTER_MV(1, 0, mx-1, my )
483 #endif
484 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
485
486 *mx_ptr = bx;
487 *my_ptr = by;
488 }else{
489 *mx_ptr =4*mx;
490 *my_ptr =4*my;
491 }
492
493 return dmin;
494 }
495
496
497 #define CHECK_MV(x,y)\
498 {\
499 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
500 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
501 assert((x) >= xmin);\
502 assert((x) <= xmax);\
503 assert((y) >= ymin);\
504 assert((y) <= ymax);\
505 /*printf("check_mv %d %d\n", x, y);*/\
506 if(map[index]!=key){\
507 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
508 map[index]= key;\
509 score_map[index]= d;\
510 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
511 /*printf("score:%d\n", d);*/\
512 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
513 }\
514 }
515
516 #define CHECK_CLIPED_MV(ax,ay)\
517 {\
518 const int x= ax;\
519 const int y= ay;\
520 const int x2= FFMAX(xmin, FFMIN(x, xmax));\
521 const int y2= FFMAX(ymin, FFMIN(y, ymax));\
522 CHECK_MV(x2, y2)\
523 }
524
525 #define CHECK_MV_DIR(x,y,new_dir)\
526 {\
527 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
528 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
529 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
530 if(map[index]!=key){\
531 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
532 map[index]= key;\
533 score_map[index]= d;\
534 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
535 /*printf("score:%d\n", d);*/\
536 if(d<dmin){\
537 best[0]=x;\
538 best[1]=y;\
539 dmin=d;\
540 next_dir= new_dir;\
541 }\
542 }\
543 }
544
545 #define check(x,y,S,v)\
546 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
547 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
548 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
549 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
550
551 #define LOAD_COMMON2\
552 uint32_t *map= c->map;\
553 const int qpel= flags&FLAG_QPEL;\
554 const int shift= 1+qpel;\
555
556 static always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
557 int src_index, int ref_index, int const penalty_factor,
558 int size, int h, int flags)
559 {
560 MotionEstContext * const c= &s->me;
561 me_cmp_func cmpf, chroma_cmpf;
562 int next_dir=-1;
563 LOAD_COMMON
564 LOAD_COMMON2
565 int map_generation= c->map_generation;
566
567 cmpf= s->dsp.me_cmp[size];
568 chroma_cmpf= s->dsp.me_cmp[size+1];
569
570 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
571 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
572 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
573 if(map[index]!=key){ //this will be executed only very rarey
574 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
575 map[index]= key;
576 }
577 }
578
579 for(;;){
580 int d;
581 const int dir= next_dir;
582 const int x= best[0];
583 const int y= best[1];
584 next_dir=-1;
585
586 //printf("%d", dir);
587 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
588 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
589 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
590 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
591
592 if(next_dir==-1){
593 return dmin;
594 }
595 }
596 }
597
598 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
599 int src_index, int ref_index, int const penalty_factor,
600 int size, int h, int flags)
601 {
602 MotionEstContext * const c= &s->me;
603 me_cmp_func cmpf, chroma_cmpf;
604 int dia_size;
605 LOAD_COMMON
606 LOAD_COMMON2
607 int map_generation= c->map_generation;
608
609 cmpf= s->dsp.me_cmp[size];
610 chroma_cmpf= s->dsp.me_cmp[size+1];
611
612 for(dia_size=1; dia_size<=4; dia_size++){
613 int dir;
614 const int x= best[0];
615 const int y= best[1];
616
617 if(dia_size&(dia_size-1)) continue;
618
619 if( x + dia_size > xmax
620 || x - dia_size < xmin
621 || y + dia_size > ymax
622 || y - dia_size < ymin)
623 continue;
624
625 for(dir= 0; dir<dia_size; dir+=2){
626 int d;
627
628 CHECK_MV(x + dir , y + dia_size - dir);
629 CHECK_MV(x + dia_size - dir, y - dir );
630 CHECK_MV(x - dir , y - dia_size + dir);
631 CHECK_MV(x - dia_size + dir, y + dir );
632 }
633
634 if(x!=best[0] || y!=best[1])
635 dia_size=0;
636 #if 0
637 {
638 int dx, dy, i;
639 static int stats[8*8];
640 dx= ABS(x-best[0]);
641 dy= ABS(y-best[1]);
642 if(dy>dx){
643 dx^=dy; dy^=dx; dx^=dy;
644 }
645 stats[dy*8 + dx] ++;
646 if(256*256*256*64 % (stats[0]+1)==0){
647 for(i=0; i<64; i++){
648 if((i&7)==0) printf("\n");
649 printf("%8d ", stats[i]);
650 }
651 printf("\n");
652 }
653 }
654 #endif
655 }
656 return dmin;
657 }
658
659 #define SAB_CHECK_MV(ax,ay)\
660 {\
661 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
662 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
663 /*printf("sab check %d %d\n", ax, ay);*/\
664 if(map[index]!=key){\
665 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
666 map[index]= key;\
667 score_map[index]= d;\
668 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
669 /*printf("score: %d\n", d);*/\
670 if(d < minima[minima_count-1].height){\
671 int j=0;\
672 \
673 while(d >= minima[j].height) j++;\
674 \
675 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
676 \
677 minima[j].checked= 0;\
678 minima[j].height= d;\
679 minima[j].x= ax;\
680 minima[j].y= ay;\
681 \
682 i=-1;\
683 continue;\
684 }\
685 }\
686 }
687
688 #define MAX_SAB_SIZE ME_MAP_SIZE
689 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
690 int src_index, int ref_index, int const penalty_factor,
691 int size, int h, int flags)
692 {
693 MotionEstContext * const c= &s->me;
694 me_cmp_func cmpf, chroma_cmpf;
695 Minima minima[MAX_SAB_SIZE];
696 const int minima_count= ABS(c->dia_size);
697 int i, j;
698 LOAD_COMMON
699 LOAD_COMMON2
700 int map_generation= c->map_generation;
701
702 cmpf= s->dsp.me_cmp[size];
703 chroma_cmpf= s->dsp.me_cmp[size+1];
704
705 for(j=i=0; i<ME_MAP_SIZE; i++){
706 uint32_t key= map[i];
707
708 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
709
710 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
711
712 assert(j<MAX_SAB_SIZE); //max j = number of predictors
713
714 minima[j].height= score_map[i];
715 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
716 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
717 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
718 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
719 minima[j].checked=0;
720 if(minima[j].x || minima[j].y)
721 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
722
723 j++;
724 }
725
726 qsort(minima, j, sizeof(Minima), minima_cmp);
727
728 for(; j<minima_count; j++){
729 minima[j].height=256*256*256*64;
730 minima[j].checked=0;
731 minima[j].x= minima[j].y=0;
732 }
733
734 for(i=0; i<minima_count; i++){
735 const int x= minima[i].x;
736 const int y= minima[i].y;
737 int d;
738
739 if(minima[i].checked) continue;
740
741 if( x >= xmax || x <= xmin
742 || y >= ymax || y <= ymin)
743 continue;
744
745 SAB_CHECK_MV(x-1, y)
746 SAB_CHECK_MV(x+1, y)
747 SAB_CHECK_MV(x , y-1)
748 SAB_CHECK_MV(x , y+1)
749
750 minima[i].checked= 1;
751 }
752
753 best[0]= minima[0].x;
754 best[1]= minima[0].y;
755 dmin= minima[0].height;
756
757 if( best[0] < xmax && best[0] > xmin
758 && best[1] < ymax && best[1] > ymin){
759 int d;
760 //ensure that the refernece samples for hpel refinement are in the map
761 CHECK_MV(best[0]-1, best[1])
762 CHECK_MV(best[0]+1, best[1])
763 CHECK_MV(best[0], best[1]-1)
764 CHECK_MV(best[0], best[1]+1)
765 }
766 return dmin;
767 }
768
769 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
770 int src_index, int ref_index, int const penalty_factor,
771 int size, int h, int flags)
772 {
773 MotionEstContext * const c= &s->me;
774 me_cmp_func cmpf, chroma_cmpf;
775 int dia_size;
776 LOAD_COMMON
777 LOAD_COMMON2
778 int map_generation= c->map_generation;
779
780 cmpf= s->dsp.me_cmp[size];
781 chroma_cmpf= s->dsp.me_cmp[size+1];
782
783 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
784 int dir, start, end;
785 const int x= best[0];
786 const int y= best[1];
787
788 start= FFMAX(0, y + dia_size - ymax);
789 end = FFMIN(dia_size, xmax - x + 1);
790 for(dir= start; dir<end; dir++){
791 int d;
792
793 //check(x + dir,y + dia_size - dir,0, a0)
794 CHECK_MV(x + dir , y + dia_size - dir);
795 }
796
797 start= FFMAX(0, x + dia_size - xmax);
798 end = FFMIN(dia_size, y - ymin + 1);
799 for(dir= start; dir<end; dir++){
800 int d;
801
802 //check(x + dia_size - dir, y - dir,0, a1)
803 CHECK_MV(x + dia_size - dir, y - dir );
804 }
805
806 start= FFMAX(0, -y + dia_size + ymin );
807 end = FFMIN(dia_size, x - xmin + 1);
808 for(dir= start; dir<end; dir++){
809 int d;
810
811 //check(x - dir,y - dia_size + dir,0, a2)
812 CHECK_MV(x - dir , y - dia_size + dir);
813 }
814
815 start= FFMAX(0, -x + dia_size + xmin );
816 end = FFMIN(dia_size, ymax - y + 1);
817 for(dir= start; dir<end; dir++){
818 int d;
819
820 //check(x - dia_size + dir, y + dir,0, a3)
821 CHECK_MV(x - dia_size + dir, y + dir );
822 }
823
824 if(x!=best[0] || y!=best[1])
825 dia_size=0;
826 #if 0
827 {
828 int dx, dy, i;
829 static int stats[8*8];
830 dx= ABS(x-best[0]);
831 dy= ABS(y-best[1]);
832 stats[dy*8 + dx] ++;
833 if(256*256*256*64 % (stats[0]+1)==0){
834 for(i=0; i<64; i++){
835 if((i&7)==0) printf("\n");
836 printf("%6d ", stats[i]);
837 }
838 printf("\n");
839 }
840 }
841 #endif
842 }
843 return dmin;
844 }
845
846 static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
847 int src_index, int ref_index, int const penalty_factor,
848 int size, int h, int flags){
849 MotionEstContext * const c= &s->me;
850 if(c->dia_size==-1)
851 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
852 else if(c->dia_size<-1)
853 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
854 else if(c->dia_size<2)
855 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
856 else
857 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
858 }
859
860 static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
861 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
862 int ref_mv_scale, int flags, int size, int h)
863 {
864 MotionEstContext * const c= &s->me;
865 int best[2]={0, 0};
866 int d, dmin;
867 int map_generation;
868 int penalty_factor;
869 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
870 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
871 me_cmp_func cmpf, chroma_cmpf;
872
873 LOAD_COMMON
874 LOAD_COMMON2
875
876 if(c->pre_pass){
877 penalty_factor= c->pre_penalty_factor;
878 cmpf= s->dsp.me_pre_cmp[size];
879 chroma_cmpf= s->dsp.me_pre_cmp[size+1];
880 }else{
881 penalty_factor= c->penalty_factor;
882 cmpf= s->dsp.me_cmp[size];
883 chroma_cmpf= s->dsp.me_cmp[size+1];
884 }
885
886 map_generation= update_map_generation(c);
887
888 assert(cmpf);
889 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
890 map[0]= map_generation;
891 score_map[0]= dmin;
892
893 /* first line */
894 if (s->first_slice_line) {
895 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
896 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
897 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
898 }else{
899 if(dmin<h*h && ( P_LEFT[0] |P_LEFT[1]
900 |P_TOP[0] |P_TOP[1]
901 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
902 *mx_ptr= 0;
903 *my_ptr= 0;
904 c->skip=1;
905 return dmin;
906 }
907 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
908 if(dmin>h*h*2){
909 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
910 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
911 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
912 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
913 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
914 }
915 }
916 if(dmin>h*h*4){
917 if(c->pre_pass){
918 CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
919 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
920 if(!s->first_slice_line)
921 CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
922 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
923 }else{
924 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
925 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
926 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
927 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
928 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
929 }
930 }
931
932 if(c->avctx->last_predictor_count){
933 const int count= c->avctx->last_predictor_count;
934 const int xstart= FFMAX(0, s->mb_x - count);
935 const int ystart= FFMAX(0, s->mb_y - count);
936 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
937 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
938 int mb_y;
939
940 for(mb_y=ystart; mb_y<yend; mb_y++){
941 int mb_x;
942 for(mb_x=xstart; mb_x<xend; mb_x++){
943 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
944 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
945 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
946
947 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
948 CHECK_MV(mx,my)
949 }
950 }
951 }
952
953 //check(best[0],best[1],0, b0)
954 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
955
956 //check(best[0],best[1],0, b1)
957 *mx_ptr= best[0];
958 *my_ptr= best[1];
959
960 // printf("%d %d %d \n", best[0], best[1], dmin);
961 return dmin;
962 }
963
964 //this function is dedicated to the braindamaged gcc
965 inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
966 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
967 int ref_mv_scale, int size, int h)
968 {
969 MotionEstContext * const c= &s->me;
970 //FIXME convert other functions in the same way if faster
971 if(c->flags==0 && h==16 && size==0){
972 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
973 // case FLAG_QPEL:
974 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
975 }else{
976 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
977 }
978 }
979
980 static int epzs_motion_search4(MpegEncContext * s,
981 int *mx_ptr, int *my_ptr, int P[10][2],
982 int src_index, int ref_index, int16_t (*last_mv)[2],
983 int ref_mv_scale)
984 {
985 MotionEstContext * const c= &s->me;
986 int best[2]={0, 0};
987 int d, dmin;
988 int map_generation;
989 const int penalty_factor= c->penalty_factor;
990 const int size=1;
991 const int h=8;
992 const int ref_mv_stride= s->mb_stride;
993 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
994 me_cmp_func cmpf, chroma_cmpf;
995 LOAD_COMMON
996 int flags= c->flags;
997 LOAD_COMMON2
998
999 cmpf= s->dsp.me_cmp[size];
1000 chroma_cmpf= s->dsp.me_cmp[size+1];
1001
1002 map_generation= update_map_generation(c);
1003
1004 dmin = 1000000;
1005 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1006 /* first line */
1007 if (s->first_slice_line) {
1008 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1009 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1010 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1011 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1012 }else{
1013 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1014 //FIXME try some early stop
1015 if(dmin>64*2){
1016 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1017 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1018 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1019 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1020 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1021 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1022 }
1023 }
1024 if(dmin>64*4){
1025 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1026 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1027 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1028 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1029 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1030 }
1031
1032 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1033
1034 *mx_ptr= best[0];
1035 *my_ptr= best[1];
1036
1037 // printf("%d %d %d \n", best[0], best[1], dmin);
1038 return dmin;
1039 }
1040
1041 //try to merge with above FIXME (needs PSNR test)
1042 static int epzs_motion_search2(MpegEncContext * s,
1043 int *mx_ptr, int *my_ptr, int P[10][2],
1044 int src_index, int ref_index, int16_t (*last_mv)[2],
1045 int ref_mv_scale)
1046 {
1047 MotionEstContext * const c= &s->me;
1048 int best[2]={0, 0};
1049 int d, dmin;
1050 int map_generation;
1051 const int penalty_factor= c->penalty_factor;
1052 const int size=0; //FIXME pass as arg
1053 const int h=8;
1054 const int ref_mv_stride= s->mb_stride;
1055 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1056 me_cmp_func cmpf, chroma_cmpf;
1057 LOAD_COMMON
1058 int flags= c->flags;
1059 LOAD_COMMON2
1060
1061 cmpf= s->dsp.me_cmp[size];
1062 chroma_cmpf= s->dsp.me_cmp[size+1];
1063
1064 map_generation= update_map_generation(c);
1065
1066 dmin = 1000000;
1067 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1068 /* first line */
1069 if (s->first_slice_line) {
1070 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1071 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1072 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1073 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1074 }else{
1075 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1076 //FIXME try some early stop
1077 if(dmin>64*2){
1078 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1079 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1080 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1081 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1082 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1083 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1084 }
1085 }
1086 if(dmin>64*4){
1087 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1088 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1089 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1090 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1091 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1092 }
1093
1094 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1095
1096 *mx_ptr= best[0];
1097 *my_ptr= best[1];
1098
1099 // printf("%d %d %d \n", best[0], best[1], dmin);
1100 return dmin;
1101 }
mob repository