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 const 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 * (1 << 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) * (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) * (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) * (1 << ME_MAP_MV_BITS) + (mx + 1) + map_generation;
103 av_assert2(c->map[(index+1)&(ME_MAP_SIZE-1)] == key);
104 key = (my) * (1 << 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 * (1 << 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)] ==
303 (my - 1) * (1 << ME_MAP_MV_BITS) + (mx - 1) + map_generation) {
304 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
306 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
309 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
311 av_assert2(16*cx2 + 4*cx + 32*c == 32*r);
312 av_assert2(16*cx2 - 4*cx + 32*c == 32*l);
313 av_assert2(16*cy2 + 4*cy + 32*c == 32*b);
314 av_assert2(16*cy2 - 4*cy + 32*c == 32*t);
315 av_assert2(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
317 for(ny= -3; ny <= 3; ny++){
318 for(nx= -3; nx <= 3; nx++){
319 //FIXME this could overflow (unlikely though)
320 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
323 if((nx&3)==0 && (ny&3)==0) continue;
325 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
326 // if(nx&1) score-=32*c->penalty_factor;
327 // if(ny&1) score-=32*c->penalty_factor;
331 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
332 memmove(best_pos[i + 1], best_pos[i], sizeof(best_pos[0]) * (7 - i));
334 best_pos[i][0]= nx + 4*mx;
335 best_pos[i][1]= ny + 4*my;
342 for(i=0; i<subpel_quality; i++){
345 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
348 av_assert2(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
361 #define CHECK_MV(x,y)\
363 const unsigned key = ((unsigned)(y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
364 const int index= (((unsigned)(y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
365 av_assert2((x) >= xmin);\
366 av_assert2((x) <= xmax);\
367 av_assert2((y) >= ymin);\
368 av_assert2((y) <= ymax);\
369 if(map[index]!=key){\
370 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
372 score_map[index]= d;\
373 d += (mv_penalty[((x)*(1<<shift))-pred_x] + mv_penalty[((y)*(1<<shift))-pred_y])*penalty_factor;\
374 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
378 #define CHECK_CLIPPED_MV(ax,ay)\
382 const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\
383 const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\
387 #define CHECK_MV_DIR(x,y,new_dir)\
389 const unsigned key = ((unsigned)(y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
390 const int index= (((unsigned)(y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
391 if(map[index]!=key){\
392 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
394 score_map[index]= d;\
395 d += (mv_penalty[(int)((unsigned)(x)<<shift)-pred_x] + mv_penalty[(int)((unsigned)(y)<<shift)-pred_y])*penalty_factor;\
405 #define check(x,y,S,v)\
406 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);\
407 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);\
408 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);\
409 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);\
411 #define LOAD_COMMON2\
412 uint32_t *map= c->map;\
413 const int qpel= flags&FLAG_QPEL;\
414 const int shift= 1+qpel;\
416 static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
417 int src_index, int ref_index, const int penalty_factor,
418 int size, int h, int flags)
420 MotionEstContext * const c= &s->me;
421 me_cmp_func cmpf, chroma_cmpf;
425 unsigned map_generation = c->map_generation;
427 cmpf = s->mecc.me_cmp[size];
428 chroma_cmpf = s->mecc.me_cmp[size + 1];
430 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
431 const unsigned key = ((unsigned)best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
432 const int index= (((unsigned)best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
433 if (map[index] != key) { // this will be executed only very rarely
434 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
441 const int dir= next_dir;
442 const int x= best[0];
443 const int y= best[1];
446 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
447 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
448 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
449 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
457 static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
458 int src_index, int ref_index, const int penalty_factor,
459 int size, int h, int flags)
461 MotionEstContext * const c= &s->me;
462 me_cmp_func cmpf, chroma_cmpf;
466 unsigned map_generation = c->map_generation;
468 cmpf = s->mecc.me_cmp[size];
469 chroma_cmpf = s->mecc.me_cmp[size + 1];
471 for(dia_size=1; dia_size<=4; dia_size++){
473 const int x= best[0];
474 const int y= best[1];
476 if(dia_size&(dia_size-1)) continue;
478 if( x + dia_size > xmax
479 || x - dia_size < xmin
480 || y + dia_size > ymax
481 || y - dia_size < ymin)
484 for(dir= 0; dir<dia_size; dir+=2){
487 CHECK_MV(x + dir , y + dia_size - dir);
488 CHECK_MV(x + dia_size - dir, y - dir );
489 CHECK_MV(x - dir , y - dia_size + dir);
490 CHECK_MV(x - dia_size + dir, y + dir );
493 if(x!=best[0] || y!=best[1])
499 static int hex_search(MpegEncContext * s, int *best, int dmin,
500 int src_index, int ref_index, const int penalty_factor,
501 int size, int h, int flags, int dia_size)
503 MotionEstContext * const c= &s->me;
504 me_cmp_func cmpf, chroma_cmpf;
507 unsigned map_generation = c->map_generation;
509 const int dec= dia_size & (dia_size-1);
511 cmpf = s->mecc.me_cmp[size];
512 chroma_cmpf = s->mecc.me_cmp[size + 1];
514 for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
519 CHECK_CLIPPED_MV(x -dia_size , y);
520 CHECK_CLIPPED_MV(x+ dia_size , y);
521 CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size);
522 CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size);
524 CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size);
525 CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size);
527 }while(best[0] != x || best[1] != y);
533 static int l2s_dia_search(MpegEncContext * s, int *best, int dmin,
534 int src_index, int ref_index, const int penalty_factor,
535 int size, int h, int flags)
537 MotionEstContext * const c= &s->me;
538 me_cmp_func cmpf, chroma_cmpf;
541 unsigned map_generation = c->map_generation;
543 int dia_size= c->dia_size&0xFF;
544 const int dec= dia_size & (dia_size-1);
545 static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1},
546 { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}};
548 cmpf = s->mecc.me_cmp[size];
549 chroma_cmpf = s->mecc.me_cmp[size + 1];
551 for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){
556 CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size);
558 }while(best[0] != x || best[1] != y);
563 CHECK_CLIPPED_MV(x+1, y);
564 CHECK_CLIPPED_MV(x, y+1);
565 CHECK_CLIPPED_MV(x-1, y);
566 CHECK_CLIPPED_MV(x, y-1);
571 static int umh_search(MpegEncContext * s, int *best, int dmin,
572 int src_index, int ref_index, const int penalty_factor,
573 int size, int h, int flags)
575 MotionEstContext * const c= &s->me;
576 me_cmp_func cmpf, chroma_cmpf;
579 unsigned map_generation = c->map_generation;
580 int x,y,x2,y2, i, j, d;
581 const int dia_size= c->dia_size&0xFE;
582 static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2},
583 { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2},
584 {-2, 3}, { 0, 4}, { 2, 3},
585 {-2,-3}, { 0,-4}, { 2,-3},};
587 cmpf = s->mecc.me_cmp[size];
588 chroma_cmpf = s->mecc.me_cmp[size + 1];
592 for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){
595 for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){
601 for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){
602 for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){
607 //FIXME prevent the CLIP stuff
609 for(j=1; j<=dia_size/4; j++){
611 CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j);
615 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2);
618 static int full_search(MpegEncContext * s, int *best, int dmin,
619 int src_index, int ref_index, const int penalty_factor,
620 int size, int h, int flags)
622 MotionEstContext * const c= &s->me;
623 me_cmp_func cmpf, chroma_cmpf;
626 unsigned map_generation = c->map_generation;
628 const int dia_size= c->dia_size&0xFF;
630 cmpf = s->mecc.me_cmp[size];
631 chroma_cmpf = s->mecc.me_cmp[size + 1];
633 for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){
634 for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){
642 CHECK_CLIPPED_MV(x , y);
643 CHECK_CLIPPED_MV(x+1, y);
644 CHECK_CLIPPED_MV(x, y+1);
645 CHECK_CLIPPED_MV(x-1, y);
646 CHECK_CLIPPED_MV(x, y-1);
653 #define SAB_CHECK_MV(ax,ay)\
655 const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
656 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
657 if(map[index]!=key){\
658 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
660 score_map[index]= d;\
661 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
662 if(d < minima[minima_count-1].height){\
665 while(d >= minima[j].height) j++;\
667 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
669 minima[j].checked= 0;\
670 minima[j].height= d;\
680 #define MAX_SAB_SIZE ME_MAP_SIZE
681 static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
682 int src_index, int ref_index, const int penalty_factor,
683 int size, int h, int flags)
685 MotionEstContext * const c= &s->me;
686 me_cmp_func cmpf, chroma_cmpf;
687 Minima minima[MAX_SAB_SIZE];
688 const int minima_count= FFABS(c->dia_size);
692 unsigned map_generation = c->map_generation;
694 av_assert1(minima_count <= MAX_SAB_SIZE);
696 cmpf = s->mecc.me_cmp[size];
697 chroma_cmpf = s->mecc.me_cmp[size + 1];
699 /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can
700 become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map
702 for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){
703 uint32_t key= map[i];
705 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
707 if ((key & (-(1 << (2 * ME_MAP_MV_BITS)))) != map_generation)
710 minima[j].height= score_map[i];
711 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
712 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
713 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
714 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
716 // all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space
717 if( minima[j].x > xmax || minima[j].x < xmin
718 || minima[j].y > ymax || minima[j].y < ymin)
722 if(minima[j].x || minima[j].y)
723 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
728 AV_QSORT(minima, j, Minima, minima_cmp);
730 for(; j<minima_count; j++){
731 minima[j].height=256*256*256*64;
733 minima[j].x= minima[j].y=0;
736 for(i=0; i<minima_count; i++){
737 const int x= minima[i].x;
738 const int y= minima[i].y;
741 if(minima[i].checked) continue;
743 if( x >= xmax || x <= xmin
744 || y >= ymax || y <= ymin)
749 SAB_CHECK_MV(x , y-1)
750 SAB_CHECK_MV(x , y+1)
752 minima[i].checked= 1;
755 best[0]= minima[0].x;
756 best[1]= minima[0].y;
757 dmin= minima[0].height;
759 if( best[0] < xmax && best[0] > xmin
760 && best[1] < ymax && best[1] > ymin){
762 // ensure that the reference samples for hpel refinement are in the map
763 CHECK_MV(best[0]-1, best[1])
764 CHECK_MV(best[0]+1, best[1])
765 CHECK_MV(best[0], best[1]-1)
766 CHECK_MV(best[0], best[1]+1)
771 static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
772 int src_index, int ref_index, const int penalty_factor,
773 int size, int h, int flags)
775 MotionEstContext * const c= &s->me;
776 me_cmp_func cmpf, chroma_cmpf;
780 unsigned map_generation = c->map_generation;
782 cmpf = s->mecc.me_cmp[size];
783 chroma_cmpf = s->mecc.me_cmp[size + 1];
785 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
787 const int x= best[0];
788 const int y= best[1];
790 start= FFMAX(0, y + dia_size - ymax);
791 end = FFMIN(dia_size, xmax - x + 1);
792 for(dir= start; dir<end; dir++){
795 //check(x + dir,y + dia_size - dir,0, a0)
796 CHECK_MV(x + dir , y + dia_size - dir);
799 start= FFMAX(0, x + dia_size - xmax);
800 end = FFMIN(dia_size, y - ymin + 1);
801 for(dir= start; dir<end; dir++){
804 //check(x + dia_size - dir, y - dir,0, a1)
805 CHECK_MV(x + dia_size - dir, y - dir );
808 start= FFMAX(0, -y + dia_size + ymin );
809 end = FFMIN(dia_size, x - xmin + 1);
810 for(dir= start; dir<end; dir++){
813 //check(x - dir,y - dia_size + dir,0, a2)
814 CHECK_MV(x - dir , y - dia_size + dir);
817 start= FFMAX(0, -x + dia_size + xmin );
818 end = FFMIN(dia_size, ymax - y + 1);
819 for(dir= start; dir<end; dir++){
822 //check(x - dia_size + dir, y + dir,0, a3)
823 CHECK_MV(x - dia_size + dir, y + dir );
826 if(x!=best[0] || y!=best[1])
832 static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
833 int src_index, int ref_index, const int penalty_factor,
834 int size, int h, int flags){
835 MotionEstContext * const c= &s->me;
837 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
838 else if(c->dia_size<-1)
839 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
840 else if(c->dia_size<2)
841 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
842 else if(c->dia_size>1024)
843 return full_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
844 else if(c->dia_size>768)
845 return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
846 else if(c->dia_size>512)
847 return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);
848 else if(c->dia_size>256)
849 return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
851 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
855 @param P a list of candidate mvs to check before starting the
856 iterative search. If one of the candidates is close to the optimal mv, then
857 it takes fewer iterations. And it increases the chance that we find the
860 static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
861 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
862 int ref_mv_scale, int flags, int size, int h)
864 MotionEstContext * const c= &s->me;
865 int best[2]={0, 0}; /**< x and y coordinates of the best motion vector.
866 i.e. the difference between the position of the
867 block currently being encoded and the position of
868 the block chosen to predict it from. */
869 int d; ///< the score (cmp + penalty) of any given mv
870 int dmin; /**< the best value of d, i.e. the score
871 corresponding to the mv stored in best[]. */
872 unsigned map_generation;
874 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
875 const int ref_mv_xy = s->mb_x + s->mb_y * ref_mv_stride; // add to last_mv before passing FIXME
876 me_cmp_func cmpf, chroma_cmpf;
882 penalty_factor= c->pre_penalty_factor;
883 cmpf = s->mecc.me_pre_cmp[size];
884 chroma_cmpf = s->mecc.me_pre_cmp[size + 1];
886 penalty_factor= c->penalty_factor;
887 cmpf = s->mecc.me_cmp[size];
888 chroma_cmpf = s->mecc.me_cmp[size + 1];
891 map_generation= update_map_generation(c);
894 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
895 map[0]= map_generation;
898 //FIXME precalc first term below?
899 if ((s->pict_type == AV_PICTURE_TYPE_B && !(c->flags & FLAG_DIRECT)) ||
900 s->mpv_flags & FF_MPV_FLAG_MV0)
901 dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor;
904 if (s->first_slice_line) {
905 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
906 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
907 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
909 if(dmin<((h*h*s->avctx->mv0_threshold)>>8)
910 && ( P_LEFT[0] |P_LEFT[1]
912 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
918 CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift)
919 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1)
920 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1)
921 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) )
922 CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) )
923 CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
924 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
925 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
926 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
927 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
931 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
932 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
933 if(!s->first_slice_line)
934 CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
935 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
937 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
938 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
939 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
940 CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
941 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
945 if(c->avctx->last_predictor_count){
946 const int count= c->avctx->last_predictor_count;
947 const int xstart= FFMAX(0, s->mb_x - count);
948 const int ystart= FFMAX(0, s->mb_y - count);
949 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
950 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
953 for(mb_y=ystart; mb_y<yend; mb_y++){
955 for(mb_x=xstart; mb_x<xend; mb_x++){
956 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
957 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
958 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
960 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
966 //check(best[0],best[1],0, b0)
967 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
969 //check(best[0],best[1],0, b1)
976 //this function is dedicated to the brain damaged gcc
977 int ff_epzs_motion_search(MpegEncContext *s, int *mx_ptr, int *my_ptr,
978 int P[10][2], int src_index, int ref_index,
979 int16_t (*last_mv)[2], int ref_mv_scale,
982 MotionEstContext * const c= &s->me;
983 //FIXME convert other functions in the same way if faster
984 if(c->flags==0 && h==16 && size==0){
985 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
987 // return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
989 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
993 static int epzs_motion_search2(MpegEncContext * s,
994 int *mx_ptr, int *my_ptr, int P[10][2],
995 int src_index, int ref_index, int16_t (*last_mv)[2],
996 int ref_mv_scale, const int size)
998 MotionEstContext * const c= &s->me;
1001 unsigned map_generation;
1002 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);