3 * Copyright (c) 2002-2004 Michael Niedermayer
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * @file motion_est_template.c
25 * Motion estimation template.
28 //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
30 uint32_t attribute_unused * const score_map= c->score_map;\
31 const int attribute_unused xmin= c->xmin;\
32 const int attribute_unused ymin= c->ymin;\
33 const int attribute_unused xmax= c->xmax;\
34 const int attribute_unused ymax= c->ymax;\
35 uint8_t *mv_penalty= c->current_mv_penalty;\
36 const int pred_x= c->pred_x;\
37 const int pred_y= c->pred_y;\
39 #define CHECK_HALF_MV(dx, dy, x, y)\
41 const int hx= 2*(x)+(dx);\
42 const int hy= 2*(y)+(dy);\
43 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
44 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
45 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
49 static int hpel_motion_search)(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
54 const int xx = 16 * s->mb_x + 8*(n&1);
55 const int yy = 16 * s->mb_y + 8*(n>>1);
56 const int mx = *mx_ptr;
57 const int my = *my_ptr;
58 const int penalty_factor= c->sub_penalty_factor;
64 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;
66 if(s->no_rounding /*FIXME b_type*/){
67 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];
68 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];
70 hpel_put=& s->dsp.put_pixels_tab[size];
71 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
73 cmpf= s->dsp.me_cmp[size];
74 chroma_cmpf= s->dsp.me_cmp[size+1];
75 cmp_sub= s->dsp.me_sub_cmp[size];
76 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
78 if(c->skip){ //FIXME somehow move up (benchmark)
84 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
85 CMP_HPEL(dmin, 0, 0, mx, my, size);
87 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
90 if (mx > xmin && mx < xmax &&
91 my > ymin && my < ymax) {
95 CHECK_HALF_MV(1, 1, mx-1, my-1)
96 CHECK_HALF_MV(0, 1, mx , my-1)
97 CHECK_HALF_MV(1, 1, mx , my-1)
98 CHECK_HALF_MV(1, 0, mx-1, my )
99 CHECK_HALF_MV(1, 0, mx , my )
100 CHECK_HALF_MV(1, 1, mx-1, my )
101 CHECK_HALF_MV(0, 1, mx , my )
102 CHECK_HALF_MV(1, 1, mx , my )
104 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);
117 static int hpel_motion_search(MpegEncContext * s,
118 int *mx_ptr, int *my_ptr, int dmin,
119 int src_index, int ref_index,
122 MotionEstContext * const c= &s->me;
123 const int mx = *mx_ptr;
124 const int my = *my_ptr;
125 const int penalty_factor= c->sub_penalty_factor;
126 me_cmp_func cmp_sub, chroma_cmp_sub;
127 int bx=2*mx, by=2*my;
130 int flags= c->sub_flags;
134 cmp_sub= s->dsp.me_sub_cmp[size];
135 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
137 if(c->skip){ //FIXME move out of hpel?
143 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
144 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
145 if(mx || my || size>0)
146 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
149 if (mx > xmin && mx < xmax &&
150 my > ymin && my < ymax) {
152 const int index= (my<<ME_MAP_SHIFT) + mx;
153 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
154 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;
155 const int l= 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 r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]
158 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
159 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
160 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;
164 int map_generation= c->map_generation;
166 uint32_t *map= c->map;
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+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
171 assert(map[(index+(1<<ME_MAP_SHIFT))&(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 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
175 assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);
178 CHECK_HALF_MV(0, 1, mx ,my-1)
180 CHECK_HALF_MV(1, 1, mx-1, my-1)
182 CHECK_HALF_MV(1, 1, mx , my-1)
184 CHECK_HALF_MV(1, 1, mx-1, my )
186 CHECK_HALF_MV(1, 0, mx-1, my )
188 CHECK_HALF_MV(1, 1, mx , my-1)
190 CHECK_HALF_MV(1, 1, mx-1, my-1)
192 CHECK_HALF_MV(1, 1, mx , my )
194 CHECK_HALF_MV(1, 0, mx , my )
199 CHECK_HALF_MV(1, 1, mx-1, my-1)
201 CHECK_HALF_MV(1, 1, mx , my )
203 CHECK_HALF_MV(1, 0, mx-1, my)
204 CHECK_HALF_MV(1, 1, mx-1, my)
207 CHECK_HALF_MV(1, 1, mx , my-1)
209 CHECK_HALF_MV(1, 1, mx-1, my)
211 CHECK_HALF_MV(1, 0, mx , my)
212 CHECK_HALF_MV(1, 1, mx , my)
214 CHECK_HALF_MV(0, 1, mx , my)
216 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
226 static int no_sub_motion_search(MpegEncContext * s,
227 int *mx_ptr, int *my_ptr, int dmin,
228 int src_index, int ref_index,
236 inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
237 int ref_index, int size, int h, int add_rate)
239 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
240 MotionEstContext * const c= &s->me;
241 const int penalty_factor= c->mb_penalty_factor;
242 const int flags= c->mb_flags;
243 const int qpel= flags & FLAG_QPEL;
244 const int mask= 1+2*qpel;
245 me_cmp_func cmp_sub, chroma_cmp_sub;
252 cmp_sub= s->dsp.mb_cmp[size];
253 chroma_cmp_sub= s->dsp.mb_cmp[size+1];
256 // assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);
258 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);
259 //FIXME check cbp before adding penalty for (0,0) vector
260 if(add_rate && (mx || my || size>0))
261 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
266 #define CHECK_QUARTER_MV(dx, dy, x, y)\
268 const int hx= 4*(x)+(dx);\
269 const int hy= 4*(y)+(dy);\
270 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
271 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
272 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
275 static int qpel_motion_search(MpegEncContext * s,
276 int *mx_ptr, int *my_ptr, int dmin,
277 int src_index, int ref_index,
280 MotionEstContext * const c= &s->me;
281 const int mx = *mx_ptr;
282 const int my = *my_ptr;
283 const int penalty_factor= c->sub_penalty_factor;
284 const int map_generation= c->map_generation;
285 const int subpel_quality= c->avctx->me_subpel_quality;
286 uint32_t *map= c->map;
287 me_cmp_func cmpf, chroma_cmpf;
288 me_cmp_func cmp_sub, chroma_cmp_sub;
291 int flags= c->sub_flags;
293 cmpf= s->dsp.me_cmp[size];
294 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
297 cmp_sub= s->dsp.me_sub_cmp[size];
298 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
300 if(c->skip){ //FIXME somehow move up (benchmark)
306 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
307 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
308 if(mx || my || size>0)
309 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
312 if (mx > xmin && mx < xmax &&
313 my > ymin && my < ymax) {
314 int bx=4*mx, by=4*my;
317 const int index= (my<<ME_MAP_SHIFT) + mx;
318 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
319 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
320 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
321 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
322 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
326 memset(best, 64, sizeof(int)*8);
328 if(s->me.dia_size>=2){
329 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
330 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
331 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
332 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
334 for(ny= -3; ny <= 3; ny++){
335 for(nx= -3; nx <= 3; nx++){
336 //FIXME this could overflow (unlikely though)
337 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
338 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
339 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
340 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
343 if((nx&3)==0 && (ny&3)==0) continue;
345 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
347 // if(nx&1) score-=1024*c->penalty_factor;
348 // if(ny&1) score-=1024*c->penalty_factor;
352 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
353 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
355 best_pos[i][0]= nx + 4*mx;
356 best_pos[i][1]= ny + 4*my;
364 //FIXME this could overflow (unlikely though)
365 const int cx = 4*(r - l);
366 const int cx2= r + l - 2*c;
367 const int cy = 4*(b - t);
368 const int cy2= b + t - 2*c;
371 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
372 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
374 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
377 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
379 assert(16*cx2 + 4*cx + 32*c == 32*r);
380 assert(16*cx2 - 4*cx + 32*c == 32*l);
381 assert(16*cy2 + 4*cy + 32*c == 32*b);
382 assert(16*cy2 - 4*cy + 32*c == 32*t);
383 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
385 for(ny= -3; ny <= 3; ny++){
386 for(nx= -3; nx <= 3; nx++){
387 //FIXME this could overflow (unlikely though)
388 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
391 if((nx&3)==0 && (ny&3)==0) continue;
393 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
394 // if(nx&1) score-=32*c->penalty_factor;
395 // if(ny&1) score-=32*c->penalty_factor;
399 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
400 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
402 best_pos[i][0]= nx + 4*mx;
403 best_pos[i][1]= ny + 4*my;
410 for(i=0; i<subpel_quality; i++){
413 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
417 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
418 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
419 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
420 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
421 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
424 // nx= FFMAX(4*mx - bx, bx - 4*mx);
425 // ny= FFMAX(4*my - by, by - 4*my);
427 static int stats[7][7], count;
429 stats[4*mx - bx + 3][4*my - by + 3]++;
430 if(256*256*256*64 % count ==0){
432 if((i%7)==0) printf("\n");
433 printf("%6d ", stats[0][i]);
441 CHECK_QUARTER_MV(2, 2, mx-1, my-1)
442 CHECK_QUARTER_MV(0, 2, mx , my-1)
443 CHECK_QUARTER_MV(2, 2, mx , my-1)
444 CHECK_QUARTER_MV(2, 0, mx , my )
445 CHECK_QUARTER_MV(2, 2, mx , my )
446 CHECK_QUARTER_MV(0, 2, mx , my )
447 CHECK_QUARTER_MV(2, 2, mx-1, my )
448 CHECK_QUARTER_MV(2, 0, mx-1, my )
454 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
455 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
456 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)
461 CHECK_QUARTER_MV(1, 3, mx-1, my-1)
462 CHECK_QUARTER_MV(1, 2, mx-1, my-1)
463 CHECK_QUARTER_MV(1, 1, mx-1, my-1)
464 CHECK_QUARTER_MV(2, 1, mx-1, my-1)
465 CHECK_QUARTER_MV(3, 1, mx-1, my-1)
466 CHECK_QUARTER_MV(0, 1, mx , my-1)
467 CHECK_QUARTER_MV(1, 1, mx , my-1)
468 CHECK_QUARTER_MV(2, 1, mx , my-1)
469 CHECK_QUARTER_MV(3, 1, mx , my-1)
470 CHECK_QUARTER_MV(3, 2, mx , my-1)
471 CHECK_QUARTER_MV(3, 3, mx , my-1)
472 CHECK_QUARTER_MV(3, 0, mx , my )
473 CHECK_QUARTER_MV(3, 1, mx , my )
474 CHECK_QUARTER_MV(3, 2, mx , my )
475 CHECK_QUARTER_MV(3, 3, mx , my )
476 CHECK_QUARTER_MV(2, 3, mx , my )
477 CHECK_QUARTER_MV(1, 3, mx , my )
478 CHECK_QUARTER_MV(0, 3, mx , my )
479 CHECK_QUARTER_MV(3, 3, mx-1, my )
480 CHECK_QUARTER_MV(2, 3, mx-1, my )
481 CHECK_QUARTER_MV(1, 3, mx-1, my )
482 CHECK_QUARTER_MV(1, 2, mx-1, my )
483 CHECK_QUARTER_MV(1, 1, mx-1, my )
484 CHECK_QUARTER_MV(1, 0, mx-1, my )
486 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
499 #define CHECK_MV(x,y)\
501 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
502 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
503 assert((x) >= xmin);\
504 assert((x) <= xmax);\
505 assert((y) >= ymin);\
506 assert((y) <= ymax);\
507 /*printf("check_mv %d %d\n", x, y);*/\
508 if(map[index]!=key){\
509 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
511 score_map[index]= d;\
512 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
513 /*printf("score:%d\n", d);*/\
514 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
518 #define CHECK_CLIPPED_MV(ax,ay)\
522 const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\
523 const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\
527 #define CHECK_MV_DIR(x,y,new_dir)\
529 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
530 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
531 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
532 if(map[index]!=key){\
533 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
535 score_map[index]= d;\
536 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
537 /*printf("score:%d\n", d);*/\
547 #define check(x,y,S,v)\
548 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
549 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
550 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
551 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
553 #define LOAD_COMMON2\
554 uint32_t *map= c->map;\
555 const int qpel= flags&FLAG_QPEL;\
556 const int shift= 1+qpel;\
558 static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
559 int src_index, int ref_index, int const penalty_factor,
560 int size, int h, int flags)
562 MotionEstContext * const c= &s->me;
563 me_cmp_func cmpf, chroma_cmpf;
567 int map_generation= c->map_generation;
569 cmpf= s->dsp.me_cmp[size];
570 chroma_cmpf= s->dsp.me_cmp[size+1];
572 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
573 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
574 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
575 if(map[index]!=key){ //this will be executed only very rarey
576 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
583 const int dir= next_dir;
584 const int x= best[0];
585 const int y= best[1];
589 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
590 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
591 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
592 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
600 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
601 int src_index, int ref_index, int const penalty_factor,
602 int size, int h, int flags)
604 MotionEstContext * const c= &s->me;
605 me_cmp_func cmpf, chroma_cmpf;
609 int map_generation= c->map_generation;
611 cmpf= s->dsp.me_cmp[size];
612 chroma_cmpf= s->dsp.me_cmp[size+1];
614 for(dia_size=1; dia_size<=4; dia_size++){
616 const int x= best[0];
617 const int y= best[1];
619 if(dia_size&(dia_size-1)) continue;
621 if( x + dia_size > xmax
622 || x - dia_size < xmin
623 || y + dia_size > ymax
624 || y - dia_size < ymin)
627 for(dir= 0; dir<dia_size; dir+=2){
630 CHECK_MV(x + dir , y + dia_size - dir);
631 CHECK_MV(x + dia_size - dir, y - dir );
632 CHECK_MV(x - dir , y - dia_size + dir);
633 CHECK_MV(x - dia_size + dir, y + dir );
636 if(x!=best[0] || y!=best[1])
641 static int stats[8*8];
642 dx= FFABS(x-best[0]);
643 dy= FFABS(y-best[1]);
645 dx^=dy; dy^=dx; dx^=dy;
648 if(256*256*256*64 % (stats[0]+1)==0){
650 if((i&7)==0) printf("\n");
651 printf("%8d ", stats[i]);
661 static int hex_search(MpegEncContext * s, int *best, int dmin,
662 int src_index, int ref_index, int const penalty_factor,
663 int size, int h, int flags, int dia_size)
665 MotionEstContext * const c= &s->me;
666 me_cmp_func cmpf, chroma_cmpf;
669 int map_generation= c->map_generation;
671 const int dec= dia_size & (dia_size-1);
673 cmpf= s->dsp.me_cmp[size];
674 chroma_cmpf= s->dsp.me_cmp[size+1];
676 for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
681 CHECK_CLIPPED_MV(x -dia_size , y);
682 CHECK_CLIPPED_MV(x+ dia_size , y);
683 CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size);
684 CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size);
686 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size);
687 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size);
689 }while(best[0] != x || best[1] != y);
695 static int l2s_dia_search(MpegEncContext * s, int *best, int dmin,
696 int src_index, int ref_index, int const penalty_factor,
697 int size, int h, int flags)
699 MotionEstContext * const c= &s->me;
700 me_cmp_func cmpf, chroma_cmpf;
703 int map_generation= c->map_generation;
705 int dia_size= c->dia_size&0xFF;
706 const int dec= dia_size & (dia_size-1);
707 static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1},
708 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};
710 cmpf= s->dsp.me_cmp[size];
711 chroma_cmpf= s->dsp.me_cmp[size+1];
713 for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
718 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size);
720 }while(best[0] != x || best[1] != y);
725 CHECK_CLIPPED_MV(x+1, y);
726 CHECK_CLIPPED_MV(x, y+1);
727 CHECK_CLIPPED_MV(x-1, y);
728 CHECK_CLIPPED_MV(x, y-1);
733 static int umh_search(MpegEncContext * s, int *best, int dmin,
734 int src_index, int ref_index, int const penalty_factor,
735 int size, int h, int flags)
737 MotionEstContext * const c= &s->me;
738 me_cmp_func cmpf, chroma_cmpf;
741 int map_generation= c->map_generation;
742 int x,y,x2,y2, i, j, d;
743 const int dia_size= c->dia_size&0xFE;
744 static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2},
745 { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},
746 {-2, 3}, { 0, 4}, { 2, 3},
747 {-2,-3}, { 0,-4}, { 2,-3},};
749 cmpf= s->dsp.me_cmp[size];
750 chroma_cmpf= s->dsp.me_cmp[size+1];
754 for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){
757 for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){
763 for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){
764 for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){
769 //FIXME prevent the CLIP stuff
771 for(j=1; j<=dia_size/4; j++){
773 CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j);
777 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2);
780 #define SAB_CHECK_MV(ax,ay)\
782 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
783 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
784 /*printf("sab check %d %d\n", ax, ay);*/\
785 if(map[index]!=key){\
786 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
788 score_map[index]= d;\
789 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
790 /*printf("score: %d\n", d);*/\
791 if(d < minima[minima_count-1].height){\
794 while(d >= minima[j].height) j++;\
796 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
798 minima[j].checked= 0;\
799 minima[j].height= d;\
809 #define MAX_SAB_SIZE ME_MAP_SIZE
810 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
811 int src_index, int ref_index, int const penalty_factor,
812 int size, int h, int flags)
814 MotionEstContext * const c= &s->me;
815 me_cmp_func cmpf, chroma_cmpf;
816 Minima minima[MAX_SAB_SIZE];
817 const int minima_count= FFABS(c->dia_size);
821 int map_generation= c->map_generation;
823 cmpf= s->dsp.me_cmp[size];
824 chroma_cmpf= s->dsp.me_cmp[size+1];
826 /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can
827 become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map
829 for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){
830 uint32_t key= map[i];
832 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
834 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
836 minima[j].height= score_map[i];
837 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
838 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
839 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
840 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
842 // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space
843 if( minima[j].x > xmax || minima[j].x < xmin
844 || minima[j].y > ymax || minima[j].y < ymin)
848 if(minima[j].x || minima[j].y)
849 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
854 qsort(minima, j, sizeof(Minima), minima_cmp);
856 for(; j<minima_count; j++){
857 minima[j].height=256*256*256*64;
859 minima[j].x= minima[j].y=0;
862 for(i=0; i<minima_count; i++){
863 const int x= minima[i].x;
864 const int y= minima[i].y;
867 if(minima[i].checked) continue;
869 if( x >= xmax || x <= xmin
870 || y >= ymax || y <= ymin)
875 SAB_CHECK_MV(x , y-1)
876 SAB_CHECK_MV(x , y+1)
878 minima[i].checked= 1;
881 best[0]= minima[0].x;
882 best[1]= minima[0].y;
883 dmin= minima[0].height;
885 if( best[0] < xmax && best[0] > xmin
886 && best[1] < ymax && best[1] > ymin){
888 //ensure that the refernece samples for hpel refinement are in the map
889 CHECK_MV(best[0]-1, best[1])
890 CHECK_MV(best[0]+1, best[1])
891 CHECK_MV(best[0], best[1]-1)
892 CHECK_MV(best[0], best[1]+1)
897 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
898 int src_index, int ref_index, int const penalty_factor,
899 int size, int h, int flags)
901 MotionEstContext * const c= &s->me;
902 me_cmp_func cmpf, chroma_cmpf;
906 int map_generation= c->map_generation;
908 cmpf= s->dsp.me_cmp[size];
909 chroma_cmpf= s->dsp.me_cmp[size+1];
911 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
913 const int x= best[0];
914 const int y= best[1];
916 start= FFMAX(0, y + dia_size - ymax);
917 end = FFMIN(dia_size, xmax - x + 1);
918 for(dir= start; dir<end; dir++){
921 //check(x + dir,y + dia_size - dir,0, a0)
922 CHECK_MV(x + dir , y + dia_size - dir);
925 start= FFMAX(0, x + dia_size - xmax);
926 end = FFMIN(dia_size, y - ymin + 1);
927 for(dir= start; dir<end; dir++){
930 //check(x + dia_size - dir, y - dir,0, a1)
931 CHECK_MV(x + dia_size - dir, y - dir );
934 start= FFMAX(0, -y + dia_size + ymin );
935 end = FFMIN(dia_size, x - xmin + 1);
936 for(dir= start; dir<end; dir++){
939 //check(x - dir,y - dia_size + dir,0, a2)
940 CHECK_MV(x - dir , y - dia_size + dir);
943 start= FFMAX(0, -x + dia_size + xmin );
944 end = FFMIN(dia_size, ymax - y + 1);
945 for(dir= start; dir<end; dir++){
948 //check(x - dia_size + dir, y + dir,0, a3)
949 CHECK_MV(x - dia_size + dir, y + dir );
952 if(x!=best[0] || y!=best[1])
957 static int stats[8*8];
958 dx= FFABS(x-best[0]);
959 dy= FFABS(y-best[1]);
961 if(256*256*256*64 % (stats[0]+1)==0){
963 if((i&7)==0) printf("\n");
964 printf("%6d ", stats[i]);
974 static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
975 int src_index, int ref_index, int const penalty_factor,
976 int size, int h, int flags){
977 MotionEstContext * const c= &s->me;
979 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
980 else if(c->dia_size<-1)
981 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
982 else if(c->dia_size<2)
983 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
984 else if(c->dia_size>768)
985 return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
986 else if(c->dia_size>512)
987 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);
988 else if(c->dia_size>256)
989 return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
991 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
994 static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
995 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
996 int ref_mv_scale, int flags, int size, int h)
998 MotionEstContext * const c= &s->me;
1003 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
1004 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
1005 me_cmp_func cmpf, chroma_cmpf;
1011 penalty_factor= c->pre_penalty_factor;
1012 cmpf= s->dsp.me_pre_cmp[size];
1013 chroma_cmpf= s->dsp.me_pre_cmp[size+1];
1015 penalty_factor= c->penalty_factor;
1016 cmpf= s->dsp.me_cmp[size];
1017 chroma_cmpf= s->dsp.me_cmp[size+1];
1020 map_generation= update_map_generation(c);
1023 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
1024 map[0]= map_generation;
1028 if (s->first_slice_line) {
1029 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1030 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1031 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1033 if(dmin<((h*h*s->avctx->mv0_threshold)>>8)
1034 && ( P_LEFT[0] |P_LEFT[1]
1036 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
1042 CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift)
1043 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1)
1044 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1)
1045 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) )
1046 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) )
1047 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1048 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1049 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
1050 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
1051 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1055 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
1056 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
1057 if(!s->first_slice_line)
1058 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1059 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1061 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1062 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1063 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1064 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1065 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1069 if(c->avctx->last_predictor_count){
1070 const int count= c->avctx->last_predictor_count;
1071 const int xstart= FFMAX(0, s->mb_x - count);
1072 const int ystart= FFMAX(0, s->mb_y - count);
1073 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
1074 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
1077 for(mb_y=ystart; mb_y<yend; mb_y++){
1079 for(mb_x=xstart; mb_x<xend; mb_x++){
1080 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
1081 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
1082 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
1084 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
1090 //check(best[0],best[1],0, b0)
1091 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1093 //check(best[0],best[1],0, b1)
1097 // printf("%d %d %d \n", best[0], best[1], dmin);
1101 //this function is dedicated to the braindamaged gcc
1102 inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
1103 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
1104 int ref_mv_scale, int size, int h)
1106 MotionEstContext * const c= &s->me;
1107 //FIXME convert other functions in the same way if faster
1108 if(c->flags==0 && h==16 && size==0){
1109 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
1111 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
1113 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
1117 static int epzs_motion_search4(MpegEncContext * s,
1118 int *mx_ptr, int *my_ptr, int P[10][2],
1119 int src_index, int ref_index, int16_t (*last_mv)[2],
1122 MotionEstContext * const c= &s->me;
1126 const int penalty_factor= c->penalty_factor;
1129 const int ref_mv_stride= s->mb_stride;
1130 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1131 me_cmp_func cmpf, chroma_cmpf;
1133 int flags= c->flags;
1136 cmpf= s->dsp.me_cmp[size];
1137 chroma_cmpf= s->dsp.me_cmp[size+1];
1139 map_generation= update_map_generation(c);
1142 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1144 if (s->first_slice_line) {
1145 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1146 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1147 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1148 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1150 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1151 //FIXME try some early stop
1152 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1153 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1154 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1155 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1156 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1157 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1160 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1161 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1162 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1163 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1164 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1167 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1172 // printf("%d %d %d \n", best[0], best[1], dmin);
1176 //try to merge with above FIXME (needs PSNR test)
1177 static int epzs_motion_search2(MpegEncContext * s,
1178 int *mx_ptr, int *my_ptr, int P[10][2],
1179 int src_index, int ref_index, int16_t (*last_mv)[2],
1182 MotionEstContext * const c= &s->me;
1186 const int penalty_factor= c->penalty_factor;
1187 const int size=0; //FIXME pass as arg
1189 const int ref_mv_stride= s->mb_stride;
1190 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1191 me_cmp_func cmpf, chroma_cmpf;
1193 int flags= c->flags;
1196 cmpf= s->dsp.me_cmp[size];
1197 chroma_cmpf= s->dsp.me_cmp[size+1];
1199 map_generation= update_map_generation(c);
1202 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1204 if (s->first_slice_line) {
1205 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1206 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1207 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1208 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1210 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1211 //FIXME try some early stop
1212 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1213 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1214 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1215 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1216 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1217 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1220 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1221 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1222 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1223 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1224 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1227 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1232 // printf("%d %d %d \n", best[0], best[1], dmin);