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 int inline ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
225 int ref_index, int size, int h, int add_rate)
227 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
228 MotionEstContext * const c= &s->me;
229 const int penalty_factor= c->mb_penalty_factor;
230 const int flags= c->mb_flags;
231 const int qpel= flags & FLAG_QPEL;
232 const int mask= 1+2*qpel;
233 me_cmp_func cmp_sub, chroma_cmp_sub;
240 cmp_sub= s->dsp.mb_cmp[size];
241 chroma_cmp_sub= s->dsp.mb_cmp[size+1];
244 // assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);
246 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);
247 //FIXME check cbp before adding penalty for (0,0) vector
248 if(add_rate && (mx || my || size>0))
249 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
254 #define CHECK_QUARTER_MV(dx, dy, x, y)\
256 const int hx= 4*(x)+(dx);\
257 const int hy= 4*(y)+(dy);\
258 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
259 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
260 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
263 static int qpel_motion_search(MpegEncContext * s,
264 int *mx_ptr, int *my_ptr, int dmin,
265 int src_index, int ref_index,
268 MotionEstContext * const c= &s->me;
269 const int mx = *mx_ptr;
270 const int my = *my_ptr;
271 const int penalty_factor= c->sub_penalty_factor;
272 const int map_generation= c->map_generation;
273 const int subpel_quality= c->avctx->me_subpel_quality;
274 uint32_t *map= c->map;
275 me_cmp_func cmpf, chroma_cmpf;
276 me_cmp_func cmp_sub, chroma_cmp_sub;
279 int flags= c->sub_flags;
281 cmpf= s->dsp.me_cmp[size];
282 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
285 cmp_sub= s->dsp.me_sub_cmp[size];
286 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
288 if(c->skip){ //FIXME somehow move up (benchmark)
294 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
295 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
296 if(mx || my || size>0)
297 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
300 if (mx > xmin && mx < xmax &&
301 my > ymin && my < ymax) {
302 int bx=4*mx, by=4*my;
305 const int index= (my<<ME_MAP_SHIFT) + mx;
306 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
307 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
308 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
309 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
310 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
314 memset(best, 64, sizeof(int)*8);
316 if(s->me.dia_size>=2){
317 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
318 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
319 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
320 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
322 for(ny= -3; ny <= 3; ny++){
323 for(nx= -3; nx <= 3; nx++){
324 //FIXME this could overflow (unlikely though)
325 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
326 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
327 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
328 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
331 if((nx&3)==0 && (ny&3)==0) continue;
333 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
335 // if(nx&1) score-=1024*c->penalty_factor;
336 // if(ny&1) score-=1024*c->penalty_factor;
340 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
341 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
343 best_pos[i][0]= nx + 4*mx;
344 best_pos[i][1]= ny + 4*my;
352 //FIXME this could overflow (unlikely though)
353 const int cx = 4*(r - l);
354 const int cx2= r + l - 2*c;
355 const int cy = 4*(b - t);
356 const int cy2= b + t - 2*c;
359 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
360 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
362 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
365 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
367 assert(16*cx2 + 4*cx + 32*c == 32*r);
368 assert(16*cx2 - 4*cx + 32*c == 32*l);
369 assert(16*cy2 + 4*cy + 32*c == 32*b);
370 assert(16*cy2 - 4*cy + 32*c == 32*t);
371 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
373 for(ny= -3; ny <= 3; ny++){
374 for(nx= -3; nx <= 3; nx++){
375 //FIXME this could overflow (unlikely though)
376 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
379 if((nx&3)==0 && (ny&3)==0) continue;
381 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
382 // if(nx&1) score-=32*c->penalty_factor;
383 // if(ny&1) score-=32*c->penalty_factor;
387 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
388 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
390 best_pos[i][0]= nx + 4*mx;
391 best_pos[i][1]= ny + 4*my;
398 for(i=0; i<subpel_quality; i++){
401 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
405 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
406 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
407 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
408 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
409 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
412 // nx= FFMAX(4*mx - bx, bx - 4*mx);
413 // ny= FFMAX(4*my - by, by - 4*my);
415 static int stats[7][7], count;
417 stats[4*mx - bx + 3][4*my - by + 3]++;
418 if(256*256*256*64 % count ==0){
420 if((i%7)==0) printf("\n");
421 printf("%6d ", stats[0][i]);
429 CHECK_QUARTER_MV(2, 2, mx-1, my-1)
430 CHECK_QUARTER_MV(0, 2, mx , my-1)
431 CHECK_QUARTER_MV(2, 2, mx , my-1)
432 CHECK_QUARTER_MV(2, 0, mx , my )
433 CHECK_QUARTER_MV(2, 2, mx , my )
434 CHECK_QUARTER_MV(0, 2, mx , my )
435 CHECK_QUARTER_MV(2, 2, mx-1, my )
436 CHECK_QUARTER_MV(2, 0, mx-1, my )
442 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
443 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
444 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)
449 CHECK_QUARTER_MV(1, 3, mx-1, my-1)
450 CHECK_QUARTER_MV(1, 2, mx-1, my-1)
451 CHECK_QUARTER_MV(1, 1, mx-1, my-1)
452 CHECK_QUARTER_MV(2, 1, mx-1, my-1)
453 CHECK_QUARTER_MV(3, 1, mx-1, my-1)
454 CHECK_QUARTER_MV(0, 1, mx , my-1)
455 CHECK_QUARTER_MV(1, 1, mx , my-1)
456 CHECK_QUARTER_MV(2, 1, mx , my-1)
457 CHECK_QUARTER_MV(3, 1, mx , my-1)
458 CHECK_QUARTER_MV(3, 2, mx , my-1)
459 CHECK_QUARTER_MV(3, 3, mx , my-1)
460 CHECK_QUARTER_MV(3, 0, mx , my )
461 CHECK_QUARTER_MV(3, 1, mx , my )
462 CHECK_QUARTER_MV(3, 2, mx , my )
463 CHECK_QUARTER_MV(3, 3, mx , my )
464 CHECK_QUARTER_MV(2, 3, mx , my )
465 CHECK_QUARTER_MV(1, 3, mx , my )
466 CHECK_QUARTER_MV(0, 3, mx , my )
467 CHECK_QUARTER_MV(3, 3, mx-1, my )
468 CHECK_QUARTER_MV(2, 3, mx-1, my )
469 CHECK_QUARTER_MV(1, 3, mx-1, my )
470 CHECK_QUARTER_MV(1, 2, mx-1, my )
471 CHECK_QUARTER_MV(1, 1, mx-1, my )
472 CHECK_QUARTER_MV(1, 0, mx-1, my )
474 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
487 #define CHECK_MV(x,y)\
489 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
490 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
491 assert((x) >= xmin);\
492 assert((x) <= xmax);\
493 assert((y) >= ymin);\
494 assert((y) <= ymax);\
495 /*printf("check_mv %d %d\n", x, y);*/\
496 if(map[index]!=key){\
497 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
499 score_map[index]= d;\
500 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
501 /*printf("score:%d\n", d);*/\
502 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
506 #define CHECK_CLIPED_MV(ax,ay)\
510 const int x2= FFMAX(xmin, FFMIN(x, xmax));\
511 const int y2= FFMAX(ymin, FFMIN(y, ymax));\
515 #define CHECK_MV_DIR(x,y,new_dir)\
517 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
518 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
519 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
520 if(map[index]!=key){\
521 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
523 score_map[index]= d;\
524 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
525 /*printf("score:%d\n", d);*/\
535 #define check(x,y,S,v)\
536 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
537 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
538 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
539 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
541 #define LOAD_COMMON2\
542 uint32_t *map= c->map;\
543 const int qpel= flags&FLAG_QPEL;\
544 const int shift= 1+qpel;\
546 static always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
547 int src_index, int ref_index, int const penalty_factor,
548 int size, int h, int flags)
550 MotionEstContext * const c= &s->me;
551 me_cmp_func cmpf, chroma_cmpf;
555 int map_generation= c->map_generation;
557 cmpf= s->dsp.me_cmp[size];
558 chroma_cmpf= s->dsp.me_cmp[size+1];
560 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
561 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
562 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
563 if(map[index]!=key){ //this will be executed only very rarey
564 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
571 const int dir= next_dir;
572 const int x= best[0];
573 const int y= best[1];
577 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
578 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
579 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
580 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
588 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
589 int src_index, int ref_index, int const penalty_factor,
590 int size, int h, int flags)
592 MotionEstContext * const c= &s->me;
593 me_cmp_func cmpf, chroma_cmpf;
597 int map_generation= c->map_generation;
599 cmpf= s->dsp.me_cmp[size];
600 chroma_cmpf= s->dsp.me_cmp[size+1];
602 for(dia_size=1; dia_size<=4; dia_size++){
604 const int x= best[0];
605 const int y= best[1];
607 if(dia_size&(dia_size-1)) continue;
609 if( x + dia_size > xmax
610 || x - dia_size < xmin
611 || y + dia_size > ymax
612 || y - dia_size < ymin)
615 for(dir= 0; dir<dia_size; dir+=2){
618 CHECK_MV(x + dir , y + dia_size - dir);
619 CHECK_MV(x + dia_size - dir, y - dir );
620 CHECK_MV(x - dir , y - dia_size + dir);
621 CHECK_MV(x - dia_size + dir, y + dir );
624 if(x!=best[0] || y!=best[1])
629 static int stats[8*8];
633 dx^=dy; dy^=dx; dx^=dy;
636 if(256*256*256*64 % (stats[0]+1)==0){
638 if((i&7)==0) printf("\n");
639 printf("%8d ", stats[i]);
649 #define SAB_CHECK_MV(ax,ay)\
651 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
652 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
653 /*printf("sab check %d %d\n", ax, ay);*/\
654 if(map[index]!=key){\
655 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
657 score_map[index]= d;\
658 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
659 /*printf("score: %d\n", d);*/\
660 if(d < minima[minima_count-1].height){\
663 while(d >= minima[j].height) j++;\
665 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
667 minima[j].checked= 0;\
668 minima[j].height= d;\
678 #define MAX_SAB_SIZE ME_MAP_SIZE
679 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
680 int src_index, int ref_index, int const penalty_factor,
681 int size, int h, int flags)
683 MotionEstContext * const c= &s->me;
684 me_cmp_func cmpf, chroma_cmpf;
685 Minima minima[MAX_SAB_SIZE];
686 const int minima_count= ABS(c->dia_size);
690 int map_generation= c->map_generation;
692 cmpf= s->dsp.me_cmp[size];
693 chroma_cmpf= s->dsp.me_cmp[size+1];
695 for(j=i=0; i<ME_MAP_SIZE; i++){
696 uint32_t key= map[i];
698 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
700 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
702 assert(j<MAX_SAB_SIZE); //max j = number of predictors
704 minima[j].height= score_map[i];
705 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
706 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
707 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
708 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
710 if(minima[j].x || minima[j].y)
711 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
716 qsort(minima, j, sizeof(Minima), minima_cmp);
718 for(; j<minima_count; j++){
719 minima[j].height=256*256*256*64;
721 minima[j].x= minima[j].y=0;
724 for(i=0; i<minima_count; i++){
725 const int x= minima[i].x;
726 const int y= minima[i].y;
729 if(minima[i].checked) continue;
731 if( x >= xmax || x <= xmin
732 || y >= ymax || y <= ymin)
737 SAB_CHECK_MV(x , y-1)
738 SAB_CHECK_MV(x , y+1)
740 minima[i].checked= 1;
743 best[0]= minima[0].x;
744 best[1]= minima[0].y;
745 dmin= minima[0].height;
747 if( best[0] < xmax && best[0] > xmin
748 && best[1] < ymax && best[1] > ymin){
750 //ensure that the refernece samples for hpel refinement are in the map
751 CHECK_MV(best[0]-1, best[1])
752 CHECK_MV(best[0]+1, best[1])
753 CHECK_MV(best[0], best[1]-1)
754 CHECK_MV(best[0], best[1]+1)
759 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
760 int src_index, int ref_index, int const penalty_factor,
761 int size, int h, int flags)
763 MotionEstContext * const c= &s->me;
764 me_cmp_func cmpf, chroma_cmpf;
768 int map_generation= c->map_generation;
770 cmpf= s->dsp.me_cmp[size];
771 chroma_cmpf= s->dsp.me_cmp[size+1];
773 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
775 const int x= best[0];
776 const int y= best[1];
778 start= FFMAX(0, y + dia_size - ymax);
779 end = FFMIN(dia_size, xmax - x + 1);
780 for(dir= start; dir<end; dir++){
783 //check(x + dir,y + dia_size - dir,0, a0)
784 CHECK_MV(x + dir , y + dia_size - dir);
787 start= FFMAX(0, x + dia_size - xmax);
788 end = FFMIN(dia_size, y - ymin + 1);
789 for(dir= start; dir<end; dir++){
792 //check(x + dia_size - dir, y - dir,0, a1)
793 CHECK_MV(x + dia_size - dir, y - dir );
796 start= FFMAX(0, -y + dia_size + ymin );
797 end = FFMIN(dia_size, x - xmin + 1);
798 for(dir= start; dir<end; dir++){
801 //check(x - dir,y - dia_size + dir,0, a2)
802 CHECK_MV(x - dir , y - dia_size + dir);
805 start= FFMAX(0, -x + dia_size + xmin );
806 end = FFMIN(dia_size, ymax - y + 1);
807 for(dir= start; dir<end; dir++){
810 //check(x - dia_size + dir, y + dir,0, a3)
811 CHECK_MV(x - dia_size + dir, y + dir );
814 if(x!=best[0] || y!=best[1])
819 static int stats[8*8];
823 if(256*256*256*64 % (stats[0]+1)==0){
825 if((i&7)==0) printf("\n");
826 printf("%6d ", stats[i]);
836 static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
837 int src_index, int ref_index, int const penalty_factor,
838 int size, int h, int flags){
839 MotionEstContext * const c= &s->me;
841 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
842 else if(c->dia_size<-1)
843 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
844 else if(c->dia_size<2)
845 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
847 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
850 static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
851 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
852 int ref_mv_scale, int flags, int size, int h)
854 MotionEstContext * const c= &s->me;
858 const int penalty_factor= c->penalty_factor;
859 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
860 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
861 me_cmp_func cmpf, chroma_cmpf;
866 cmpf= s->dsp.me_cmp[size];
867 chroma_cmpf= s->dsp.me_cmp[size+1];
869 map_generation= update_map_generation(c);
872 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
873 map[0]= map_generation;
877 if (s->first_slice_line) {
878 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
879 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
880 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
882 if(dmin<h*h && ( P_LEFT[0] |P_LEFT[1]
884 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
890 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
892 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
893 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
894 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
895 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
896 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
901 CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
902 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
903 if(!s->first_slice_line)
904 CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
905 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
907 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
908 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
909 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
910 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
911 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
915 if(c->avctx->last_predictor_count){
916 const int count= c->avctx->last_predictor_count;
917 const int xstart= FFMAX(0, s->mb_x - count);
918 const int ystart= FFMAX(0, s->mb_y - count);
919 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
920 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
923 for(mb_y=ystart; mb_y<yend; mb_y++){
925 for(mb_x=xstart; mb_x<xend; mb_x++){
926 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
927 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
928 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
930 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
936 //check(best[0],best[1],0, b0)
937 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
939 //check(best[0],best[1],0, b1)
943 // printf("%d %d %d \n", best[0], best[1], dmin);
947 //this function is dedicated to the braindamaged gcc
948 inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
949 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
950 int ref_mv_scale, int size, int h)
952 MotionEstContext * const c= &s->me;
953 //FIXME convert other functions in the same way if faster
954 if(c->flags==0 && h==16 && size==0){
955 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
957 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
959 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
963 static int epzs_motion_search4(MpegEncContext * s,
964 int *mx_ptr, int *my_ptr, int P[10][2],
965 int src_index, int ref_index, int16_t (*last_mv)[2],
968 MotionEstContext * const c= &s->me;
972 const int penalty_factor= c->penalty_factor;
975 const int ref_mv_stride= s->mb_stride;
976 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
977 me_cmp_func cmpf, chroma_cmpf;
982 cmpf= s->dsp.me_cmp[size];
983 chroma_cmpf= s->dsp.me_cmp[size+1];
985 map_generation= update_map_generation(c);
988 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
990 if (s->first_slice_line) {
991 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
992 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
993 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
994 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
996 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
997 //FIXME try some early stop
999 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1000 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1001 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1002 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1003 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1004 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1008 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1009 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1010 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1011 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1012 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1015 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1020 // printf("%d %d %d \n", best[0], best[1], dmin);
1024 //try to merge with above FIXME (needs PSNR test)
1025 static int epzs_motion_search2(MpegEncContext * s,
1026 int *mx_ptr, int *my_ptr, int P[10][2],
1027 int src_index, int ref_index, int16_t (*last_mv)[2],
1030 MotionEstContext * const c= &s->me;
1034 const int penalty_factor= c->penalty_factor;
1035 const int size=0; //FIXME pass as arg
1037 const int ref_mv_stride= s->mb_stride;
1038 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1039 me_cmp_func cmpf, chroma_cmpf;
1041 int flags= c->flags;
1044 cmpf= s->dsp.me_cmp[size];
1045 chroma_cmpf= s->dsp.me_cmp[size+1];
1047 map_generation= update_map_generation(c);
1050 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1052 if (s->first_slice_line) {
1053 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1054 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1055 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1056 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1058 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1059 //FIXME try some early stop
1061 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1062 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1063 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1064 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1065 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1066 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1070 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1071 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1072 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1073 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1074 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1077 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1082 // printf("%d %d %d \n", best[0], best[1], dmin);