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 * Motion estimation template.
27 #include "libavutil/qsort.h"
28 #include "mpegvideo.h"
30 //Let us hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
32 uint32_t av_unused * const score_map= c->score_map;\
33 const int av_unused xmin= c->xmin;\
34 const int av_unused ymin= c->ymin;\
35 const int av_unused xmax= c->xmax;\
36 const int av_unused ymax= c->ymax;\
37 uint8_t *mv_penalty= c->current_mv_penalty;\
38 const int pred_x= c->pred_x;\
39 const int pred_y= c->pred_y;\
41 #define CHECK_HALF_MV(dx, dy, x, y)\
43 const int hx= 2*(x)+(dx);\
44 const int hy= 2*(y)+(dy);\
45 d= cmp_hpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
46 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
47 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
50 static int hpel_motion_search(MpegEncContext * s,
51 int *mx_ptr, int *my_ptr, int dmin,
52 int src_index, int ref_index,
55 MotionEstContext * const c= &s->me;
56 const int mx = *mx_ptr;
57 const int my = *my_ptr;
58 const int penalty_factor= c->sub_penalty_factor;
59 me_cmp_func cmp_sub, chroma_cmp_sub;
63 int flags= c->sub_flags;
67 cmp_sub = s->mecc.me_sub_cmp[size];
68 chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1];
70 if(c->skip){ //FIXME move out of hpel?
76 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
77 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
78 if(mx || my || size>0)
79 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
82 if (mx > xmin && mx < xmax &&
83 my > ymin && my < ymax) {
85 const int index= (my<<ME_MAP_SHIFT) + mx;
86 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
87 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;
88 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]
89 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
90 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]
91 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
92 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
93 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;
95 #if defined(ASSERT_LEVEL) && ASSERT_LEVEL > 1
97 unsigned map_generation= c->map_generation;
98 key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
99 av_assert2(c->map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
100 key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
101 av_assert2(c->map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
102 key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;
103 av_assert2(c->map[(index+1)&(ME_MAP_SIZE-1)] == key);
104 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
105 av_assert2(c->map[(index-1)&(ME_MAP_SIZE-1)] == key);
108 CHECK_HALF_MV(0, 1, mx ,my-1)
110 CHECK_HALF_MV(1, 1, mx-1, my-1)
112 CHECK_HALF_MV(1, 1, mx , my-1)
114 CHECK_HALF_MV(1, 1, mx-1, my )
116 CHECK_HALF_MV(1, 0, mx-1, my )
118 CHECK_HALF_MV(1, 1, mx , my-1)
120 CHECK_HALF_MV(1, 1, mx-1, my-1)
122 CHECK_HALF_MV(1, 1, mx , my )
124 CHECK_HALF_MV(1, 0, mx , my )
129 CHECK_HALF_MV(1, 1, mx-1, my-1)
131 CHECK_HALF_MV(1, 1, mx , my )
133 CHECK_HALF_MV(1, 0, mx-1, my)
134 CHECK_HALF_MV(1, 1, mx-1, my)
137 CHECK_HALF_MV(1, 1, mx , my-1)
139 CHECK_HALF_MV(1, 1, mx-1, my)
141 CHECK_HALF_MV(1, 0, mx , my)
142 CHECK_HALF_MV(1, 1, mx , my)
144 CHECK_HALF_MV(0, 1, mx , my)
146 av_assert2(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
155 static int no_sub_motion_search(MpegEncContext * s,
156 int *mx_ptr, int *my_ptr, int dmin,
157 int src_index, int ref_index,
165 static inline int get_mb_score(MpegEncContext *s, int mx, int my,
166 int src_index, int ref_index, int size,
169 MotionEstContext * const c= &s->me;
170 const int penalty_factor= c->mb_penalty_factor;
171 const int flags= c->mb_flags;
172 const int qpel= flags & FLAG_QPEL;
173 const int mask= 1+2*qpel;
174 me_cmp_func cmp_sub, chroma_cmp_sub;
181 cmp_sub = s->mecc.mb_cmp[size];
182 chroma_cmp_sub = s->mecc.mb_cmp[size + 1];
184 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);
185 //FIXME check cbp before adding penalty for (0,0) vector
186 if(add_rate && (mx || my || size>0))
187 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
192 int ff_get_mb_score(MpegEncContext *s, int mx, int my, int src_index,
193 int ref_index, int size, int h, int add_rate)
195 return get_mb_score(s, mx, my, src_index, ref_index, size, h, add_rate);
198 #define CHECK_QUARTER_MV(dx, dy, x, y)\
200 const int hx= 4*(x)+(dx);\
201 const int hy= 4*(y)+(dy);\
202 d= cmp_qpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
203 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
204 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
207 static int qpel_motion_search(MpegEncContext * s,
208 int *mx_ptr, int *my_ptr, int dmin,
209 int src_index, int ref_index,
212 MotionEstContext * const c= &s->me;
213 const int mx = *mx_ptr;
214 const int my = *my_ptr;
215 const int penalty_factor= c->sub_penalty_factor;
216 const unsigned map_generation = c->map_generation;
217 const int subpel_quality= c->avctx->me_subpel_quality;
218 uint32_t *map= c->map;
219 me_cmp_func cmpf, chroma_cmpf;
220 me_cmp_func cmp_sub, chroma_cmp_sub;
223 int flags= c->sub_flags;
225 cmpf = s->mecc.me_cmp[size];
226 chroma_cmpf = s->mecc.me_cmp[size + 1]; // FIXME: factorize
229 cmp_sub = s->mecc.me_sub_cmp[size];
230 chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1];
232 if(c->skip){ //FIXME somehow move up (benchmark)
238 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
239 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
240 if(mx || my || size>0)
241 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
244 if (mx > xmin && mx < xmax &&
245 my > ymin && my < ymax) {
246 int bx=4*mx, by=4*my;
249 const int index= (my<<ME_MAP_SHIFT) + mx;
250 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
251 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
252 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
253 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
254 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
258 memset(best, 64, sizeof(int)*8);
259 if(s->me.dia_size>=2){
260 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
261 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
262 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
263 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
265 for(ny= -3; ny <= 3; ny++){
266 for(nx= -3; nx <= 3; nx++){
267 //FIXME this could overflow (unlikely though)
268 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
269 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
270 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
271 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
274 if((nx&3)==0 && (ny&3)==0) continue;
276 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
278 // if(nx&1) score-=1024*c->penalty_factor;
279 // if(ny&1) score-=1024*c->penalty_factor;
283 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
284 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
286 best_pos[i][0]= nx + 4*mx;
287 best_pos[i][1]= ny + 4*my;
295 //FIXME this could overflow (unlikely though)
296 const int cx = 4*(r - l);
297 const int cx2= r + l - 2*c;
298 const int cy = 4*(b - t);
299 const int cy2= b + t - 2*c;
302 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == ((my-1)<<ME_MAP_MV_BITS) + (mx-1) + map_generation){
303 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
305 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
308 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
310 av_assert2(16*cx2 + 4*cx + 32*c == 32*r);
311 av_assert2(16*cx2 - 4*cx + 32*c == 32*l);
312 av_assert2(16*cy2 + 4*cy + 32*c == 32*b);
313 av_assert2(16*cy2 - 4*cy + 32*c == 32*t);
314 av_assert2(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
316 for(ny= -3; ny <= 3; ny++){
317 for(nx= -3; nx <= 3; nx++){
318 //FIXME this could overflow (unlikely though)
319 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
322 if((nx&3)==0 && (ny&3)==0) continue;
324 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
325 // if(nx&1) score-=32*c->penalty_factor;
326 // if(ny&1) score-=32*c->penalty_factor;
330 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
331 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
333 best_pos[i][0]= nx + 4*mx;
334 best_pos[i][1]= ny + 4*my;
341 for(i=0; i<subpel_quality; i++){
344 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
347 av_assert2(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
360 #define CHECK_MV(x,y)\
362 const unsigned key = ((unsigned)(y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
363 const int index= (((unsigned)(y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
364 av_assert2((x) >= xmin);\
365 av_assert2((x) <= xmax);\
366 av_assert2((y) >= ymin);\
367 av_assert2((y) <= ymax);\
368 if(map[index]!=key){\
369 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
371 score_map[index]= d;\
372 d += (mv_penalty[((x)*(1<<shift))-pred_x] + mv_penalty[((y)*(1<<shift))-pred_y])*penalty_factor;\
373 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
377 #define CHECK_CLIPPED_MV(ax,ay)\
381 const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\
382 const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\
386 #define CHECK_MV_DIR(x,y,new_dir)\
388 const unsigned key = ((unsigned)(y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
389 const int index= (((unsigned)(y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
390 if(map[index]!=key){\
391 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
393 score_map[index]= d;\
394 d += (mv_penalty[(int)((unsigned)(x)<<shift)-pred_x] + mv_penalty[(int)((unsigned)(y)<<shift)-pred_y])*penalty_factor;\
404 #define check(x,y,S,v)\
405 if( (x)<(xmin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
406 if( (x)>(xmax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
407 if( (y)<(ymin<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
408 if( (y)>(ymax<<(S)) ) av_log(NULL, AV_LOG_ERROR, "%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
410 #define LOAD_COMMON2\
411 uint32_t *map= c->map;\
412 const int qpel= flags&FLAG_QPEL;\
413 const int shift= 1+qpel;\
415 static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
416 int src_index, int ref_index, const int penalty_factor,
417 int size, int h, int flags)
419 MotionEstContext * const c= &s->me;
420 me_cmp_func cmpf, chroma_cmpf;
424 unsigned map_generation = c->map_generation;
426 cmpf = s->mecc.me_cmp[size];
427 chroma_cmpf = s->mecc.me_cmp[size + 1];
429 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
430 const unsigned key = ((unsigned)best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
431 const int index= (((unsigned)best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
432 if (map[index] != key) { // this will be executed only very rarely
433 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
440 const int dir= next_dir;
441 const int x= best[0];
442 const int y= best[1];
445 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
446 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
447 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
448 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
456 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
457 int src_index, int ref_index, const int penalty_factor,
458 int size, int h, int flags)
460 MotionEstContext * const c= &s->me;
461 me_cmp_func cmpf, chroma_cmpf;
465 unsigned map_generation = c->map_generation;
467 cmpf = s->mecc.me_cmp[size];
468 chroma_cmpf = s->mecc.me_cmp[size + 1];
470 for(dia_size=1; dia_size<=4; dia_size++){
472 const int x= best[0];
473 const int y= best[1];
475 if(dia_size&(dia_size-1)) continue;
477 if( x + dia_size > xmax
478 || x - dia_size < xmin
479 || y + dia_size > ymax
480 || y - dia_size < ymin)
483 for(dir= 0; dir<dia_size; dir+=2){
486 CHECK_MV(x + dir , y + dia_size - dir);
487 CHECK_MV(x + dia_size - dir, y - dir );
488 CHECK_MV(x - dir , y - dia_size + dir);
489 CHECK_MV(x - dia_size + dir, y + dir );
492 if(x!=best[0] || y!=best[1])
498 static int hex_search(MpegEncContext * s, int *best, int dmin,
499 int src_index, int ref_index, const int penalty_factor,
500 int size, int h, int flags, int dia_size)
502 MotionEstContext * const c= &s->me;
503 me_cmp_func cmpf, chroma_cmpf;
506 unsigned map_generation = c->map_generation;
508 const int dec= dia_size & (dia_size-1);
510 cmpf = s->mecc.me_cmp[size];
511 chroma_cmpf = s->mecc.me_cmp[size + 1];
513 for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
518 CHECK_CLIPPED_MV(x -dia_size , y);
519 CHECK_CLIPPED_MV(x+ dia_size , y);
520 CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size);
521 CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size);
523 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size);
524 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size);
526 }while(best[0] != x || best[1] != y);
532 static int l2s_dia_search(MpegEncContext * s, int *best, int dmin,
533 int src_index, int ref_index, const int penalty_factor,
534 int size, int h, int flags)
536 MotionEstContext * const c= &s->me;
537 me_cmp_func cmpf, chroma_cmpf;
540 unsigned map_generation = c->map_generation;
542 int dia_size= c->dia_size&0xFF;
543 const int dec= dia_size & (dia_size-1);
544 static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1},
545 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};
547 cmpf = s->mecc.me_cmp[size];
548 chroma_cmpf = s->mecc.me_cmp[size + 1];
550 for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
555 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size);
557 }while(best[0] != x || best[1] != y);
562 CHECK_CLIPPED_MV(x+1, y);
563 CHECK_CLIPPED_MV(x, y+1);
564 CHECK_CLIPPED_MV(x-1, y);
565 CHECK_CLIPPED_MV(x, y-1);
570 static int umh_search(MpegEncContext * s, int *best, int dmin,
571 int src_index, int ref_index, const int penalty_factor,
572 int size, int h, int flags)
574 MotionEstContext * const c= &s->me;
575 me_cmp_func cmpf, chroma_cmpf;
578 unsigned map_generation = c->map_generation;
579 int x,y,x2,y2, i, j, d;
580 const int dia_size= c->dia_size&0xFE;
581 static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2},
582 { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},
583 {-2, 3}, { 0, 4}, { 2, 3},
584 {-2,-3}, { 0,-4}, { 2,-3},};
586 cmpf = s->mecc.me_cmp[size];
587 chroma_cmpf = s->mecc.me_cmp[size + 1];
591 for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){
594 for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){
600 for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){
601 for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){
606 //FIXME prevent the CLIP stuff
608 for(j=1; j<=dia_size/4; j++){
610 CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j);
614 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2);
617 static int full_search(MpegEncContext * s, int *best, int dmin,
618 int src_index, int ref_index, const int penalty_factor,
619 int size, int h, int flags)
621 MotionEstContext * const c= &s->me;
622 me_cmp_func cmpf, chroma_cmpf;
625 unsigned map_generation = c->map_generation;
627 const int dia_size= c->dia_size&0xFF;
629 cmpf = s->mecc.me_cmp[size];
630 chroma_cmpf = s->mecc.me_cmp[size + 1];
632 for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){
633 for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){
641 CHECK_CLIPPED_MV(x , y);
642 CHECK_CLIPPED_MV(x+1, y);
643 CHECK_CLIPPED_MV(x, y+1);
644 CHECK_CLIPPED_MV(x-1, y);
645 CHECK_CLIPPED_MV(x, y-1);
652 #define SAB_CHECK_MV(ax,ay)\
654 const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
655 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
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 if(d < minima[minima_count-1].height){\
664 while(d >= minima[j].height) j++;\
666 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
668 minima[j].checked= 0;\
669 minima[j].height= d;\
679 #define MAX_SAB_SIZE ME_MAP_SIZE
680 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
681 int src_index, int ref_index, const int penalty_factor,
682 int size, int h, int flags)
684 MotionEstContext * const c= &s->me;
685 me_cmp_func cmpf, chroma_cmpf;
686 Minima minima[MAX_SAB_SIZE];
687 const int minima_count= FFABS(c->dia_size);
691 unsigned map_generation = c->map_generation;
693 av_assert1(minima_count <= MAX_SAB_SIZE);
695 cmpf = s->mecc.me_cmp[size];
696 chroma_cmpf = s->mecc.me_cmp[size + 1];
698 /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can
699 become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map
701 for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){
702 uint32_t key= map[i];
704 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
706 if ((key & (-(1 << (2 * ME_MAP_MV_BITS)))) != map_generation)
709 minima[j].height= score_map[i];
710 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
711 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
712 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
713 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
715 // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space
716 if( minima[j].x > xmax || minima[j].x < xmin
717 || minima[j].y > ymax || minima[j].y < ymin)
721 if(minima[j].x || minima[j].y)
722 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
727 AV_QSORT(minima, j, Minima, minima_cmp);
729 for(; j<minima_count; j++){
730 minima[j].height=256*256*256*64;
732 minima[j].x= minima[j].y=0;
735 for(i=0; i<minima_count; i++){
736 const int x= minima[i].x;
737 const int y= minima[i].y;
740 if(minima[i].checked) continue;
742 if( x >= xmax || x <= xmin
743 || y >= ymax || y <= ymin)
748 SAB_CHECK_MV(x , y-1)
749 SAB_CHECK_MV(x , y+1)
751 minima[i].checked= 1;
754 best[0]= minima[0].x;
755 best[1]= minima[0].y;
756 dmin= minima[0].height;
758 if( best[0] < xmax && best[0] > xmin
759 && best[1] < ymax && best[1] > ymin){
761 // ensure that the reference samples for hpel refinement are in the map
762 CHECK_MV(best[0]-1, best[1])
763 CHECK_MV(best[0]+1, best[1])
764 CHECK_MV(best[0], best[1]-1)
765 CHECK_MV(best[0], best[1]+1)
770 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
771 int src_index, int ref_index, const int penalty_factor,
772 int size, int h, int flags)
774 MotionEstContext * const c= &s->me;
775 me_cmp_func cmpf, chroma_cmpf;
779 unsigned map_generation = c->map_generation;
781 cmpf = s->mecc.me_cmp[size];
782 chroma_cmpf = s->mecc.me_cmp[size + 1];
784 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
786 const int x= best[0];
787 const int y= best[1];
789 start= FFMAX(0, y + dia_size - ymax);
790 end = FFMIN(dia_size, xmax - x + 1);
791 for(dir= start; dir<end; dir++){
794 //check(x + dir,y + dia_size - dir,0, a0)
795 CHECK_MV(x + dir , y + dia_size - dir);
798 start= FFMAX(0, x + dia_size - xmax);
799 end = FFMIN(dia_size, y - ymin + 1);
800 for(dir= start; dir<end; dir++){
803 //check(x + dia_size - dir, y - dir,0, a1)
804 CHECK_MV(x + dia_size - dir, y - dir );
807 start= FFMAX(0, -y + dia_size + ymin );
808 end = FFMIN(dia_size, x - xmin + 1);
809 for(dir= start; dir<end; dir++){
812 //check(x - dir,y - dia_size + dir,0, a2)
813 CHECK_MV(x - dir , y - dia_size + dir);
816 start= FFMAX(0, -x + dia_size + xmin );
817 end = FFMIN(dia_size, ymax - y + 1);
818 for(dir= start; dir<end; dir++){
821 //check(x - dia_size + dir, y + dir,0, a3)
822 CHECK_MV(x - dia_size + dir, y + dir );
825 if(x!=best[0] || y!=best[1])
831 static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
832 int src_index, int ref_index, const int penalty_factor,
833 int size, int h, int flags){
834 MotionEstContext * const c= &s->me;
836 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
837 else if(c->dia_size<-1)
838 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
839 else if(c->dia_size<2)
840 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
841 else if(c->dia_size>1024)
842 return full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
843 else if(c->dia_size>768)
844 return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
845 else if(c->dia_size>512)
846 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);
847 else if(c->dia_size>256)
848 return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
850 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
854 @param P a list of candidate mvs to check before starting the
855 iterative search. If one of the candidates is close to the optimal mv, then
856 it takes fewer iterations. And it increases the chance that we find the
859 static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
860 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
861 int ref_mv_scale, int flags, int size, int h)
863 MotionEstContext * const c= &s->me;
864 int best[2]={0, 0}; /**< x and y coordinates of the best motion vector.
865 i.e. the difference between the position of the
866 block currently being encoded and the position of
867 the block chosen to predict it from. */
868 int d; ///< the score (cmp + penalty) of any given mv
869 int dmin; /**< the best value of d, i.e. the score
870 corresponding to the mv stored in best[]. */
871 unsigned map_generation;
873 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
874 const int ref_mv_xy = s->mb_x + s->mb_y * ref_mv_stride; // add to last_mv before passing FIXME
875 me_cmp_func cmpf, chroma_cmpf;
881 penalty_factor= c->pre_penalty_factor;
882 cmpf = s->mecc.me_pre_cmp[size];
883 chroma_cmpf = s->mecc.me_pre_cmp[size + 1];
885 penalty_factor= c->penalty_factor;
886 cmpf = s->mecc.me_cmp[size];
887 chroma_cmpf = s->mecc.me_cmp[size + 1];
890 map_generation= update_map_generation(c);
893 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
894 map[0]= map_generation;
897 //FIXME precalc first term below?
898 if ((s->pict_type == AV_PICTURE_TYPE_B && !(c->flags & FLAG_DIRECT)) ||
899 s->mpv_flags & FF_MPV_FLAG_MV0)
900 dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor;
903 if (s->first_slice_line) {
904 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
905 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
906 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
908 if(dmin<((h*h*s->avctx->mv0_threshold)>>8)
909 && ( P_LEFT[0] |P_LEFT[1]
911 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
917 CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift)
918 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1)
919 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1)
920 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) )
921 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) )
922 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
923 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
924 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
925 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
926 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
930 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
931 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
932 if(!s->first_slice_line)
933 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
934 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
936 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
937 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
938 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
939 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
940 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
944 if(c->avctx->last_predictor_count){
945 const int count= c->avctx->last_predictor_count;
946 const int xstart= FFMAX(0, s->mb_x - count);
947 const int ystart= FFMAX(0, s->mb_y - count);
948 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
949 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
952 for(mb_y=ystart; mb_y<yend; mb_y++){
954 for(mb_x=xstart; mb_x<xend; mb_x++){
955 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
956 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
957 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
959 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
965 //check(best[0],best[1],0, b0)
966 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
968 //check(best[0],best[1],0, b1)
975 //this function is dedicated to the brain damaged gcc
976 int ff_epzs_motion_search(MpegEncContext *s, int *mx_ptr, int *my_ptr,
977 int P[10][2], int src_index, int ref_index,
978 int16_t (*last_mv)[2], int ref_mv_scale,
981 MotionEstContext * const c= &s->me;
982 //FIXME convert other functions in the same way if faster
983 if(c->flags==0 && h==16 && size==0){
984 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
986 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
988 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
992 static int epzs_motion_search4(MpegEncContext * s,
993 int *mx_ptr, int *my_ptr, int P[10][2],
994 int src_index, int ref_index, int16_t (*last_mv)[2],
997 MotionEstContext * const c= &s->me;
1000 unsigned map_generation;
1001 const int penalty_factor= c->penalty_factor;
1004 const int ref_mv_stride= s->mb_stride;
1005 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1006 me_cmp_func cmpf, chroma_cmpf;
1008 int flags= c->flags;
1011 cmpf = s->mecc.me_cmp[size];
1012 chroma_cmpf = s->mecc.me_cmp[size + 1];
1014 map_generation= update_map_generation(c);
1019 if (s->first_slice_line) {
1020 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1021 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1022 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1023 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1025 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1026 //FIXME try some early stop
1027 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1028 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1029 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1030 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1031 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1032 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1035 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1036 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1037 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1038 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1039 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1042 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1050 //try to merge with above FIXME (needs PSNR test)
1051 static int epzs_motion_search2(MpegEncContext * s,
1052 int *mx_ptr, int *my_ptr, int P[10][2],
1053 int src_index, int ref_index, int16_t (*last_mv)[2],
1056 MotionEstContext * const c= &s->me;
1059 unsigned map_generation;
1060 const int penalty_factor= c->penalty_factor;
1061 const int size=0; //FIXME pass as arg
1063 const int ref_mv_stride= s->mb_stride;
1064 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
1065 me_cmp_func cmpf, chroma_cmpf;
1067 int flags= c->flags;
1070 cmpf = s->mecc.me_cmp[size];
1071 chroma_cmpf = s->mecc.me_cmp[size + 1];
1073 map_generation= update_map_generation(c);
1078 if (s->first_slice_line) {
1079 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1080 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1081 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1082 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1084 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1085 //FIXME try some early stop
1086 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1087 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1088 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1089 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1090 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1091 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1094 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1095 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
1096 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
1097 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1098 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1101 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);