3 * Copyright (c) 2002-2004 Michael Niedermayer
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.
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.
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
22 * @file motion_est_template.c
23 * Motion estimation template.
26 //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
28 uint32_t * const score_map= c->score_map;\
29 const int xmin= c->xmin;\
30 const int ymin= c->ymin;\
31 const int xmax= c->xmax;\
32 const int 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;\
37 #define CHECK_HALF_MV(dx, dy, x, y)\
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)\
47 static int hpel_motion_search)(MpegEncContext * s,
48 int *mx_ptr, int *my_ptr, int dmin,
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;
62 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;
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];
68 hpel_put=& s->dsp.put_pixels_tab[size];
69 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
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];
76 if(c->skip){ //FIXME somehow move up (benchmark)
82 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
83 CMP_HPEL(dmin, 0, 0, mx, my, size);
85 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
88 if (mx > xmin && mx < xmax &&
89 my > ymin && my < ymax) {
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 )
102 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);
115 static int hpel_motion_search(MpegEncContext * s,
116 int *mx_ptr, int *my_ptr, int dmin,
117 int src_index, int ref_index,
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;
128 int flags= c->sub_flags;
132 cmp_sub= s->dsp.me_sub_cmp[size];
133 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
135 if(c->skip){ //FIXME move out of hpel?
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;
147 if (mx > xmin && mx < xmax &&
148 my > ymin && my < ymax) {
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;
162 int map_generation= c->map_generation;
164 uint32_t *map= c->map;
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);
176 CHECK_HALF_MV(0, 1, mx ,my-1)
178 CHECK_HALF_MV(1, 1, mx-1, my-1)
180 CHECK_HALF_MV(1, 1, mx , my-1)
182 CHECK_HALF_MV(1, 1, mx-1, my )
184 CHECK_HALF_MV(1, 0, mx-1, my )
186 CHECK_HALF_MV(1, 1, mx , my-1)
188 CHECK_HALF_MV(1, 1, mx-1, my-1)
190 CHECK_HALF_MV(1, 1, mx , my )
192 CHECK_HALF_MV(1, 0, mx , my )
197 CHECK_HALF_MV(1, 1, mx-1, my-1)
199 CHECK_HALF_MV(1, 1, mx , my )
201 CHECK_HALF_MV(1, 0, mx-1, my)
202 CHECK_HALF_MV(1, 1, mx-1, my)
205 CHECK_HALF_MV(1, 1, mx , my-1)
207 CHECK_HALF_MV(1, 1, mx-1, my)
209 CHECK_HALF_MV(1, 0, mx , my)
210 CHECK_HALF_MV(1, 1, mx , my)
212 CHECK_HALF_MV(0, 1, mx , my)
214 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
224 static int inline get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
227 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
228 MotionEstContext * const c= &s->me;
231 const int penalty_factor= c->mb_penalty_factor;
232 const int flags= c->mb_flags;
233 const int qpel= flags & FLAG_QPEL;
234 const int mask= 1+2*qpel;
235 me_cmp_func cmp_sub, chroma_cmp_sub;
242 cmp_sub= s->dsp.mb_cmp[size];
243 chroma_cmp_sub= s->dsp.mb_cmp[size+1];
246 assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);
248 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);
249 //FIXME check cbp before adding penalty for (0,0) vector
250 if(mx || my || size>0)
251 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
256 #define CHECK_QUARTER_MV(dx, dy, x, y)\
258 const int hx= 4*(x)+(dx);\
259 const int hy= 4*(y)+(dy);\
260 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
261 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
262 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
265 static int qpel_motion_search(MpegEncContext * s,
266 int *mx_ptr, int *my_ptr, int dmin,
267 int src_index, int ref_index,
270 MotionEstContext * const c= &s->me;
271 const int mx = *mx_ptr;
272 const int my = *my_ptr;
273 const int penalty_factor= c->sub_penalty_factor;
274 const int map_generation= c->map_generation;
275 const int subpel_quality= c->avctx->me_subpel_quality;
276 uint32_t *map= c->map;
277 me_cmp_func cmpf, chroma_cmpf;
278 me_cmp_func cmp_sub, chroma_cmp_sub;
281 int flags= c->sub_flags;
283 cmpf= s->dsp.me_cmp[size];
284 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
287 cmp_sub= s->dsp.me_sub_cmp[size];
288 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
290 if(c->skip){ //FIXME somehow move up (benchmark)
296 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
297 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
298 if(mx || my || size>0)
299 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
302 if (mx > xmin && mx < xmax &&
303 my > ymin && my < ymax) {
304 int bx=4*mx, by=4*my;
307 const int index= (my<<ME_MAP_SHIFT) + mx;
308 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
309 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
310 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
311 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
312 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
316 memset(best, 64, sizeof(int)*8);
318 if(s->me.dia_size>=2){
319 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
320 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
321 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
322 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
324 for(ny= -3; ny <= 3; ny++){
325 for(nx= -3; nx <= 3; nx++){
326 //FIXME this could overflow (unlikely though)
327 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
328 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
329 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
330 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
333 if((nx&3)==0 && (ny&3)==0) continue;
335 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
337 // if(nx&1) score-=1024*c->penalty_factor;
338 // if(ny&1) score-=1024*c->penalty_factor;
342 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
343 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
345 best_pos[i][0]= nx + 4*mx;
346 best_pos[i][1]= ny + 4*my;
354 //FIXME this could overflow (unlikely though)
355 const int cx = 4*(r - l);
356 const int cx2= r + l - 2*c;
357 const int cy = 4*(b - t);
358 const int cy2= b + t - 2*c;
361 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
362 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
364 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
367 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
369 assert(16*cx2 + 4*cx + 32*c == 32*r);
370 assert(16*cx2 - 4*cx + 32*c == 32*l);
371 assert(16*cy2 + 4*cy + 32*c == 32*b);
372 assert(16*cy2 - 4*cy + 32*c == 32*t);
373 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
375 for(ny= -3; ny <= 3; ny++){
376 for(nx= -3; nx <= 3; nx++){
377 //FIXME this could overflow (unlikely though)
378 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
381 if((nx&3)==0 && (ny&3)==0) continue;
383 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
384 // if(nx&1) score-=32*c->penalty_factor;
385 // if(ny&1) score-=32*c->penalty_factor;
389 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
390 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
392 best_pos[i][0]= nx + 4*mx;
393 best_pos[i][1]= ny + 4*my;
400 for(i=0; i<subpel_quality; i++){
403 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
407 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
408 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
409 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
410 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
411 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
414 // nx= FFMAX(4*mx - bx, bx - 4*mx);
415 // ny= FFMAX(4*my - by, by - 4*my);
417 static int stats[7][7], count;
419 stats[4*mx - bx + 3][4*my - by + 3]++;
420 if(256*256*256*64 % count ==0){
422 if((i%7)==0) printf("\n");
423 printf("%6d ", stats[0][i]);
431 CHECK_QUARTER_MV(2, 2, mx-1, my-1)
432 CHECK_QUARTER_MV(0, 2, mx , my-1)
433 CHECK_QUARTER_MV(2, 2, mx , my-1)
434 CHECK_QUARTER_MV(2, 0, mx , my )
435 CHECK_QUARTER_MV(2, 2, mx , my )
436 CHECK_QUARTER_MV(0, 2, mx , my )
437 CHECK_QUARTER_MV(2, 2, mx-1, my )
438 CHECK_QUARTER_MV(2, 0, mx-1, my )
444 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
445 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
446 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)
451 CHECK_QUARTER_MV(1, 3, mx-1, my-1)
452 CHECK_QUARTER_MV(1, 2, mx-1, my-1)
453 CHECK_QUARTER_MV(1, 1, mx-1, my-1)
454 CHECK_QUARTER_MV(2, 1, mx-1, my-1)
455 CHECK_QUARTER_MV(3, 1, mx-1, my-1)
456 CHECK_QUARTER_MV(0, 1, mx , my-1)
457 CHECK_QUARTER_MV(1, 1, mx , my-1)
458 CHECK_QUARTER_MV(2, 1, mx , my-1)
459 CHECK_QUARTER_MV(3, 1, mx , my-1)
460 CHECK_QUARTER_MV(3, 2, mx , my-1)
461 CHECK_QUARTER_MV(3, 3, mx , my-1)
462 CHECK_QUARTER_MV(3, 0, mx , my )
463 CHECK_QUARTER_MV(3, 1, mx , my )
464 CHECK_QUARTER_MV(3, 2, mx , my )
465 CHECK_QUARTER_MV(3, 3, mx , my )
466 CHECK_QUARTER_MV(2, 3, mx , my )
467 CHECK_QUARTER_MV(1, 3, mx , my )
468 CHECK_QUARTER_MV(0, 3, mx , my )
469 CHECK_QUARTER_MV(3, 3, mx-1, my )
470 CHECK_QUARTER_MV(2, 3, mx-1, my )
471 CHECK_QUARTER_MV(1, 3, mx-1, my )
472 CHECK_QUARTER_MV(1, 2, mx-1, my )
473 CHECK_QUARTER_MV(1, 1, mx-1, my )
474 CHECK_QUARTER_MV(1, 0, mx-1, my )
476 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
489 #define CHECK_MV(x,y)\
491 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
492 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
493 assert((x) >= xmin);\
494 assert((x) <= xmax);\
495 assert((y) >= ymin);\
496 assert((y) <= ymax);\
497 /*printf("check_mv %d %d\n", x, y);*/\
498 if(map[index]!=key){\
499 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
501 score_map[index]= d;\
502 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
503 /*printf("score:%d\n", d);*/\
504 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
508 #define CHECK_CLIPED_MV(ax,ay)\
512 const int x2= FFMAX(xmin, FFMIN(x, xmax));\
513 const int y2= FFMAX(ymin, FFMIN(y, ymax));\
517 #define CHECK_MV_DIR(x,y,new_dir)\
519 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
520 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
521 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
522 if(map[index]!=key){\
523 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
525 score_map[index]= d;\
526 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
527 /*printf("score:%d\n", d);*/\
537 #define check(x,y,S,v)\
538 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
539 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
540 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
541 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
543 #define LOAD_COMMON2\
544 uint32_t *map= c->map;\
545 const int qpel= flags&FLAG_QPEL;\
546 const int shift= 1+qpel;\
548 static always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
549 int src_index, int ref_index, int const penalty_factor,
550 int size, int h, int flags)
552 MotionEstContext * const c= &s->me;
553 me_cmp_func cmpf, chroma_cmpf;
557 int map_generation= c->map_generation;
559 cmpf= s->dsp.me_cmp[size];
560 chroma_cmpf= s->dsp.me_cmp[size+1];
562 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
563 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
564 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
565 if(map[index]!=key){ //this will be executed only very rarey
566 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
573 const int dir= next_dir;
574 const int x= best[0];
575 const int y= best[1];
579 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
580 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
581 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
582 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
590 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
591 int src_index, int ref_index, int const penalty_factor,
592 int size, int h, int flags)
594 MotionEstContext * const c= &s->me;
595 me_cmp_func cmpf, chroma_cmpf;
599 int map_generation= c->map_generation;
601 cmpf= s->dsp.me_cmp[size];
602 chroma_cmpf= s->dsp.me_cmp[size+1];
604 for(dia_size=1; dia_size<=4; dia_size++){
606 const int x= best[0];
607 const int y= best[1];
609 if(dia_size&(dia_size-1)) continue;
611 if( x + dia_size > xmax
612 || x - dia_size < xmin
613 || y + dia_size > ymax
614 || y - dia_size < ymin)
617 for(dir= 0; dir<dia_size; dir+=2){
620 CHECK_MV(x + dir , y + dia_size - dir);
621 CHECK_MV(x + dia_size - dir, y - dir );
622 CHECK_MV(x - dir , y - dia_size + dir);
623 CHECK_MV(x - dia_size + dir, y + dir );
626 if(x!=best[0] || y!=best[1])
631 static int stats[8*8];
635 dx^=dy; dy^=dx; dx^=dy;
638 if(256*256*256*64 % (stats[0]+1)==0){
640 if((i&7)==0) printf("\n");
641 printf("%8d ", stats[i]);
651 #define SAB_CHECK_MV(ax,ay)\
653 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
654 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
655 /*printf("sab check %d %d\n", ax, ay);*/\
656 if(map[index]!=key){\
657 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
659 score_map[index]= d;\
660 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
661 /*printf("score: %d\n", d);*/\
662 if(d < minima[minima_count-1].height){\
665 while(d >= minima[j].height) j++;\
667 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
669 minima[j].checked= 0;\
670 minima[j].height= d;\
680 #define MAX_SAB_SIZE 16
681 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
682 int src_index, int ref_index, int const penalty_factor,
683 int size, int h, int flags)
685 MotionEstContext * const c= &s->me;
686 me_cmp_func cmpf, chroma_cmpf;
687 Minima minima[MAX_SAB_SIZE];
688 const int minima_count= ABS(c->dia_size);
692 int map_generation= c->map_generation;
694 cmpf= s->dsp.me_cmp[size];
695 chroma_cmpf= s->dsp.me_cmp[size+1];
697 for(j=i=0; i<ME_MAP_SIZE; i++){
698 uint32_t key= map[i];
700 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
702 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
704 assert(j<MAX_SAB_SIZE); //max j = number of predictors
706 minima[j].height= score_map[i];
707 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
708 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
709 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
710 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
712 if(minima[j].x || minima[j].y)
713 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
718 qsort(minima, j, sizeof(Minima), minima_cmp);
720 for(; j<minima_count; j++){
721 minima[j].height=256*256*256*64;
723 minima[j].x= minima[j].y=0;
726 for(i=0; i<minima_count; i++){
727 const int x= minima[i].x;
728 const int y= minima[i].y;
731 if(minima[i].checked) continue;
733 if( x >= xmax || x <= xmin
734 || y >= ymax || y <= ymin)
739 SAB_CHECK_MV(x , y-1)
740 SAB_CHECK_MV(x , y+1)
742 minima[i].checked= 1;
745 best[0]= minima[0].x;
746 best[1]= minima[0].y;
747 dmin= minima[0].height;
749 if( best[0] < xmax && best[0] > xmin
750 && best[1] < ymax && best[1] > ymin){
752 //ensure that the refernece samples for hpel refinement are in the map
753 CHECK_MV(best[0]-1, best[1])
754 CHECK_MV(best[0]+1, best[1])
755 CHECK_MV(best[0], best[1]-1)
756 CHECK_MV(best[0], best[1]+1)
761 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
762 int src_index, int ref_index, int const penalty_factor,
763 int size, int h, int flags)
765 MotionEstContext * const c= &s->me;
766 me_cmp_func cmpf, chroma_cmpf;
770 int map_generation= c->map_generation;
772 cmpf= s->dsp.me_cmp[size];
773 chroma_cmpf= s->dsp.me_cmp[size+1];
775 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
777 const int x= best[0];
778 const int y= best[1];
780 start= FFMAX(0, y + dia_size - ymax);
781 end = FFMIN(dia_size, xmax - x + 1);
782 for(dir= start; dir<end; dir++){
785 //check(x + dir,y + dia_size - dir,0, a0)
786 CHECK_MV(x + dir , y + dia_size - dir);
789 start= FFMAX(0, x + dia_size - xmax);
790 end = FFMIN(dia_size, y - ymin + 1);
791 for(dir= start; dir<end; dir++){
794 //check(x + dia_size - dir, y - dir,0, a1)
795 CHECK_MV(x + dia_size - dir, y - dir );
798 start= FFMAX(0, -y + dia_size + ymin );
799 end = FFMIN(dia_size, x - xmin + 1);
800 for(dir= start; dir<end; dir++){
803 //check(x - dir,y - dia_size + dir,0, a2)
804 CHECK_MV(x - dir , y - dia_size + dir);
807 start= FFMAX(0, -x + dia_size + xmin );
808 end = FFMIN(dia_size, ymax - y + 1);
809 for(dir= start; dir<end; dir++){
812 //check(x - dia_size + dir, y + dir,0, a3)
813 CHECK_MV(x - dia_size + dir, y + dir );
816 if(x!=best[0] || y!=best[1])
821 static int stats[8*8];
825 if(256*256*256*64 % (stats[0]+1)==0){
827 if((i&7)==0) printf("\n");
828 printf("%6d ", stats[i]);
838 static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
839 int src_index, int ref_index, int const penalty_factor,
840 int size, int h, int flags){
841 MotionEstContext * const c= &s->me;
843 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
844 else if(c->dia_size<-1)
845 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
846 else if(c->dia_size<2)
847 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
849 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
852 static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
853 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
854 int ref_mv_scale, int flags)
856 MotionEstContext * const c= &s->me;
860 const int penalty_factor= c->penalty_factor;
863 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
864 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
865 me_cmp_func cmpf, chroma_cmpf;
870 cmpf= s->dsp.me_cmp[size];
871 chroma_cmpf= s->dsp.me_cmp[size+1];
873 map_generation= update_map_generation(c);
875 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
876 map[0]= map_generation;
880 if (s->first_slice_line) {
881 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
882 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
883 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
885 if(dmin<256 && ( P_LEFT[0] |P_LEFT[1]
887 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
893 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
895 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
896 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
897 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
898 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
899 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
904 CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
905 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
906 if(!s->first_slice_line)
907 CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
908 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
910 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
911 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
912 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
913 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
914 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
918 if(c->avctx->last_predictor_count){
919 const int count= c->avctx->last_predictor_count;
920 const int xstart= FFMAX(0, s->mb_x - count);
921 const int ystart= FFMAX(0, s->mb_y - count);
922 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
923 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
926 for(mb_y=ystart; mb_y<yend; mb_y++){
928 for(mb_x=xstart; mb_x<xend; mb_x++){
929 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
930 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
931 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
933 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
939 //check(best[0],best[1],0, b0)
940 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
942 //check(best[0],best[1],0, b1)
946 // printf("%d %d %d \n", best[0], best[1], dmin);
950 //this function is dedicated to the braindamaged gcc
951 static inline int epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
952 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
955 MotionEstContext * const c= &s->me;
956 //FIXME convert other functions in the same way if faster
959 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0);
961 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
963 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags);
967 static int epzs_motion_search4(MpegEncContext * s,
968 int *mx_ptr, int *my_ptr, int P[10][2],
969 int src_index, int ref_index, int16_t (*last_mv)[2],
972 MotionEstContext * const c= &s->me;
976 const int penalty_factor= c->penalty_factor;
979 const int ref_mv_stride= s->mb_stride;
980 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
981 me_cmp_func cmpf, chroma_cmpf;
986 cmpf= s->dsp.me_cmp[size];
987 chroma_cmpf= s->dsp.me_cmp[size+1];
989 map_generation= update_map_generation(c);
992 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
994 if (s->first_slice_line) {
995 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
996 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
997 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
998 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1000 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1001 //FIXME try some early stop
1003 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1004 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1005 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1006 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1007 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1008 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1012 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1013 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1014 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1015 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1016 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1019 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1024 // printf("%d %d %d \n", best[0], best[1], dmin);
1028 //try to merge with above FIXME (needs PSNR test)
1029 static int epzs_motion_search2(MpegEncContext * s,
1030 int *mx_ptr, int *my_ptr, int P[10][2],
1031 int src_index, int ref_index, int16_t (*last_mv)[2],
1034 MotionEstContext * const c= &s->me;
1038 const int penalty_factor= c->penalty_factor;
1039 const int size=0; //FIXME pass as arg
1041 const int ref_mv_stride= s->mb_stride;
1042 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1043 me_cmp_func cmpf, chroma_cmpf;
1045 int flags= c->flags;
1048 cmpf= s->dsp.me_cmp[size];
1049 chroma_cmpf= s->dsp.me_cmp[size+1];
1051 map_generation= update_map_generation(c);
1054 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1056 if (s->first_slice_line) {
1057 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1058 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1059 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1060 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1062 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1063 //FIXME try some early stop
1065 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1066 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1067 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1068 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1069 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1070 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1074 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1075 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1076 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1077 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1078 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1081 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1086 // printf("%d %d %d \n", best[0], best[1], dmin);