4x4 SSE compare function
[libav.git] / libavcodec / motion_est_template.c
CommitLineData
1457ab52
MN
1/*
2 * Motion estimation
8f2ab833 3 * Copyright (c) 2002-2004 Michael Niedermayer
1457ab52
MN
4 *
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
9 *
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
14 *
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 *
19 */
983e3246
MN
20
21/**
22 * @file motion_est_template.c
23 * Motion estimation template.
24 */
2750b827 25
1457ab52 26//lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...)
bb198e19 27#define LOAD_COMMON\
af4091f1
MN
28 uint32_t * const score_map= c->score_map;\
29 const int xmin= c->xmin;\
30 const int ymin= c->ymin;\
31 const int xmax= c->xmax;\
32 const int ymax= c->ymax;\
33 uint8_t *mv_penalty= c->current_mv_penalty;\
34 const int pred_x= c->pred_x;\
35 const int pred_y= c->pred_y;\
1457ab52 36
1457ab52
MN
37#define CHECK_HALF_MV(dx, dy, x, y)\
38{\
39 const int hx= 2*(x)+(dx);\
40 const int hy= 2*(y)+(dy);\
2750b827 41 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\
1457ab52
MN
42 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
43 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
44}
45
46#if 0
2750b827 47static int hpel_motion_search)(MpegEncContext * s,
1457ab52 48 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN
49 uint8_t *ref_data[3],
50 int size)
1457ab52 51{
1457ab52
MN
52 const int xx = 16 * s->mb_x + 8*(n&1);
53 const int yy = 16 * s->mb_y + 8*(n>>1);
54 const int mx = *mx_ptr;
55 const int my = *my_ptr;
af4091f1 56 const int penalty_factor= c->sub_penalty_factor;
1457ab52 57
bb198e19 58 LOAD_COMMON
1457ab52
MN
59
60 // INIT;
61 //FIXME factorize
62 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub;
63
64 if(s->no_rounding /*FIXME b_type*/){
65 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];
66 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];
67 }else{
68 hpel_put=& s->dsp.put_pixels_tab[size];
69 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
70 }
2750b827
MN
71 cmpf= s->dsp.me_cmp[size];
72 chroma_cmpf= s->dsp.me_cmp[size+1];
1457ab52
MN
73 cmp_sub= s->dsp.me_sub_cmp[size];
74 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
75
af4091f1 76 if(c->skip){ //FIXME somehow move up (benchmark)
1457ab52
MN
77 *mx_ptr = 0;
78 *my_ptr = 0;
79 return dmin;
80 }
81
1f202b0d 82 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
1457ab52
MN
83 CMP_HPEL(dmin, 0, 0, mx, my, size);
84 if(mx || my)
85 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
86 }
87
88 if (mx > xmin && mx < xmax &&
89 my > ymin && my < ymax) {
90 int bx=2*mx, by=2*my;
91 int d= dmin;
92
93 CHECK_HALF_MV(1, 1, mx-1, my-1)
94 CHECK_HALF_MV(0, 1, mx , my-1)
95 CHECK_HALF_MV(1, 1, mx , my-1)
96 CHECK_HALF_MV(1, 0, mx-1, my )
97 CHECK_HALF_MV(1, 0, mx , my )
98 CHECK_HALF_MV(1, 1, mx-1, my )
99 CHECK_HALF_MV(0, 1, mx , my )
100 CHECK_HALF_MV(1, 1, mx , my )
101
b07a5980 102 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2);
1457ab52
MN
103
104 *mx_ptr = bx;
105 *my_ptr = by;
106 }else{
107 *mx_ptr =2*mx;
108 *my_ptr =2*my;
109 }
110
111 return dmin;
112}
113
114#else
2750b827 115static int hpel_motion_search(MpegEncContext * s,
1457ab52 116 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN
117 int src_index, int ref_index,
118 int size, int h)
1457ab52 119{
af4091f1 120 MotionEstContext * const c= &s->me;
1457ab52
MN
121 const int mx = *mx_ptr;
122 const int my = *my_ptr;
af4091f1 123 const int penalty_factor= c->sub_penalty_factor;
1457ab52 124 me_cmp_func cmp_sub, chroma_cmp_sub;
67725183 125 int bx=2*mx, by=2*my;
1457ab52 126
bb198e19 127 LOAD_COMMON
af4091f1 128 int flags= c->sub_flags;
1457ab52
MN
129
130 //FIXME factorize
131
132 cmp_sub= s->dsp.me_sub_cmp[size];
133 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
134
af4091f1 135 if(c->skip){ //FIXME move out of hpel?
1457ab52
MN
136 *mx_ptr = 0;
137 *my_ptr = 0;
138 return dmin;
139 }
140
1f202b0d 141 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
2750b827 142 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
1d0eab1d 143 if(mx || my || size>0)
1457ab52
MN
144 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor;
145 }
146
147 if (mx > xmin && mx < xmax &&
148 my > ymin && my < ymax) {
1457ab52
MN
149 int d= dmin;
150 const int index= (my<<ME_MAP_SHIFT) + mx;
151 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
af4091f1 152 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*c->penalty_factor;
1457ab52 153 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]
af4091f1 154 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
1457ab52 155 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]
af4091f1 156 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*c->penalty_factor;
1457ab52 157 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
af4091f1 158 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->penalty_factor;
b07a5980 159
67725183 160#if 1
b07a5980 161 int key;
af4091f1 162 int map_generation= c->map_generation;
802f454e 163#ifndef NDEBUG
af4091f1 164 uint32_t *map= c->map;
802f454e 165#endif
b07a5980
MN
166 key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
167 assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
168 key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation;
169 assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key);
170 key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation;
171 assert(map[(index+1)&(ME_MAP_SIZE-1)] == key);
172 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
173 assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);
174#endif
1457ab52
MN
175 if(t<=b){
176 CHECK_HALF_MV(0, 1, mx ,my-1)
177 if(l<=r){
178 CHECK_HALF_MV(1, 1, mx-1, my-1)
179 if(t+r<=b+l){
180 CHECK_HALF_MV(1, 1, mx , my-1)
181 }else{
182 CHECK_HALF_MV(1, 1, mx-1, my )
183 }
184 CHECK_HALF_MV(1, 0, mx-1, my )
185 }else{
186 CHECK_HALF_MV(1, 1, mx , my-1)
187 if(t+l<=b+r){
188 CHECK_HALF_MV(1, 1, mx-1, my-1)
189 }else{
190 CHECK_HALF_MV(1, 1, mx , my )
191 }
192 CHECK_HALF_MV(1, 0, mx , my )
193 }
194 }else{
195 if(l<=r){
196 if(t+l<=b+r){
197 CHECK_HALF_MV(1, 1, mx-1, my-1)
198 }else{
199 CHECK_HALF_MV(1, 1, mx , my )
200 }
201 CHECK_HALF_MV(1, 0, mx-1, my)
202 CHECK_HALF_MV(1, 1, mx-1, my)
203 }else{
204 if(t+r<=b+l){
205 CHECK_HALF_MV(1, 1, mx , my-1)
206 }else{
207 CHECK_HALF_MV(1, 1, mx-1, my)
208 }
209 CHECK_HALF_MV(1, 0, mx , my)
210 CHECK_HALF_MV(1, 1, mx , my)
211 }
212 CHECK_HALF_MV(0, 1, mx , my)
213 }
214 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2);
1457ab52
MN
215 }
216
67725183
MN
217 *mx_ptr = bx;
218 *my_ptr = by;
219
1457ab52
MN
220 return dmin;
221}
222#endif
223
2750b827
MN
224static int inline get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
225 int ref_index)
67725183
MN
226{
227// const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp;
af4091f1 228 MotionEstContext * const c= &s->me;
67725183 229 const int size= 0;
bb198e19 230 const int h= 16;
af4091f1
MN
231 const int penalty_factor= c->mb_penalty_factor;
232 const int flags= c->mb_flags;
2750b827
MN
233 const int qpel= flags & FLAG_QPEL;
234 const int mask= 1+2*qpel;
67725183
MN
235 me_cmp_func cmp_sub, chroma_cmp_sub;
236 int d;
237
bb198e19 238 LOAD_COMMON
67725183
MN
239
240 //FIXME factorize
241
242 cmp_sub= s->dsp.mb_cmp[size];
243 chroma_cmp_sub= s->dsp.mb_cmp[size+1];
244
af4091f1 245 assert(!c->skip);
1f202b0d 246 assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp);
67725183 247
2750b827 248 d= cmp(s, mx>>(qpel+1), my>>(qpel+1), mx&mask, my&mask, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
67725183
MN
249 //FIXME check cbp before adding penalty for (0,0) vector
250 if(mx || my || size>0)
251 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor;
252
253 return d;
254}
255
1457ab52
MN
256#define CHECK_QUARTER_MV(dx, dy, x, y)\
257{\
258 const int hx= 4*(x)+(dx);\
259 const int hy= 4*(y)+(dy);\
2750b827 260 d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
1457ab52
MN
261 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\
262 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\
263}
264
2750b827 265static int qpel_motion_search(MpegEncContext * s,
1457ab52 266 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN
267 int src_index, int ref_index,
268 int size, int h)
1457ab52 269{
af4091f1 270 MotionEstContext * const c= &s->me;
1457ab52
MN
271 const int mx = *mx_ptr;
272 const int my = *my_ptr;
af4091f1
MN
273 const int penalty_factor= c->sub_penalty_factor;
274 const int map_generation= c->map_generation;
1f202b0d 275 const int subpel_quality= c->avctx->me_subpel_quality;
af4091f1 276 uint32_t *map= c->map;
2750b827 277 me_cmp_func cmpf, chroma_cmpf;
1457ab52
MN
278 me_cmp_func cmp_sub, chroma_cmp_sub;
279
bb198e19 280 LOAD_COMMON
af4091f1 281 int flags= c->sub_flags;
1457ab52 282
2750b827
MN
283 cmpf= s->dsp.me_cmp[size];
284 chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME
1457ab52
MN
285 //FIXME factorize
286
287 cmp_sub= s->dsp.me_sub_cmp[size];
288 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1];
289
af4091f1 290 if(c->skip){ //FIXME somehow move up (benchmark)
1457ab52
MN
291 *mx_ptr = 0;
292 *my_ptr = 0;
293 return dmin;
294 }
295
1f202b0d 296 if(c->avctx->me_cmp != c->avctx->me_sub_cmp){
2750b827 297 dmin= cmp(s, mx, my, 0, 0, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);
1d0eab1d 298 if(mx || my || size>0)
1457ab52
MN
299 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor;
300 }
301
302 if (mx > xmin && mx < xmax &&
303 my > ymin && my < ymax) {
304 int bx=4*mx, by=4*my;
305 int d= dmin;
306 int i, nx, ny;
307 const int index= (my<<ME_MAP_SHIFT) + mx;
308 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
309 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
310 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
311 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)];
312 const int c= score_map[(index )&(ME_MAP_SIZE-1)];
313 int best[8];
314 int best_pos[8][2];
315
316 memset(best, 64, sizeof(int)*8);
317#if 1
826f429a 318 if(s->me.dia_size>=2){
1457ab52
MN
319 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
320 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
321 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
322 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
323
324 for(ny= -3; ny <= 3; ny++){
325 for(nx= -3; nx <= 3; nx++){
0da6315a
MN
326 //FIXME this could overflow (unlikely though)
327 const int64_t t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t;
328 const int64_t c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c;
329 const int64_t b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b;
330 int score= (ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2 + 512)>>10;
1457ab52
MN
331 int i;
332
333 if((nx&3)==0 && (ny&3)==0) continue;
334
0da6315a 335 score += (mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
1457ab52 336
af4091f1
MN
337// if(nx&1) score-=1024*c->penalty_factor;
338// if(ny&1) score-=1024*c->penalty_factor;
1457ab52
MN
339
340 for(i=0; i<8; i++){
341 if(score < best[i]){
342 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
343 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
344 best[i]= score;
345 best_pos[i][0]= nx + 4*mx;
346 best_pos[i][1]= ny + 4*my;
347 break;
348 }
349 }
350 }
351 }
352 }else{
353 int tl;
0da6315a 354 //FIXME this could overflow (unlikely though)
1457ab52
MN
355 const int cx = 4*(r - l);
356 const int cx2= r + l - 2*c;
357 const int cy = 4*(b - t);
358 const int cy2= b + t - 2*c;
359 int cxy;
360
361 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME
362 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
363 }else{
2750b827 364 tl= cmp(s, mx-1, my-1, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);//FIXME wrong if chroma me is different
1457ab52
MN
365 }
366
367 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c;
368
369 assert(16*cx2 + 4*cx + 32*c == 32*r);
370 assert(16*cx2 - 4*cx + 32*c == 32*l);
371 assert(16*cy2 + 4*cy + 32*c == 32*b);
372 assert(16*cy2 - 4*cy + 32*c == 32*t);
373 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl);
374
375 for(ny= -3; ny <= 3; ny++){
376 for(nx= -3; nx <= 3; nx++){
0da6315a 377 //FIXME this could overflow (unlikely though)
1457ab52
MN
378 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor
379 int i;
380
381 if((nx&3)==0 && (ny&3)==0) continue;
382
383 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor;
af4091f1
MN
384// if(nx&1) score-=32*c->penalty_factor;
385 // if(ny&1) score-=32*c->penalty_factor;
1457ab52
MN
386
387 for(i=0; i<8; i++){
388 if(score < best[i]){
389 memmove(&best[i+1], &best[i], sizeof(int)*(7-i));
390 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i));
391 best[i]= score;
392 best_pos[i][0]= nx + 4*mx;
393 best_pos[i][1]= ny + 4*my;
394 break;
395 }
396 }
397 }
398 }
399 }
826f429a 400 for(i=0; i<subpel_quality; i++){
1457ab52
MN
401 nx= best_pos[i][0];
402 ny= best_pos[i][1];
403 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
404 }
826f429a 405
1457ab52 406#if 0
826f429a
MN
407 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
408 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
409 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
410 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
411// if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
412 if(tl<br){
413
414// nx= FFMAX(4*mx - bx, bx - 4*mx);
415// ny= FFMAX(4*my - by, by - 4*my);
1457ab52 416
826f429a
MN
417 static int stats[7][7], count;
418 count++;
419 stats[4*mx - bx + 3][4*my - by + 3]++;
420 if(256*256*256*64 % count ==0){
421 for(i=0; i<49; i++){
422 if((i%7)==0) printf("\n");
1457ab52
MN
423 printf("%6d ", stats[0][i]);
424 }
425 printf("\n");
426 }
826f429a 427 }
1457ab52
MN
428#endif
429#else
430
431 CHECK_QUARTER_MV(2, 2, mx-1, my-1)
432 CHECK_QUARTER_MV(0, 2, mx , my-1)
433 CHECK_QUARTER_MV(2, 2, mx , my-1)
434 CHECK_QUARTER_MV(2, 0, mx , my )
435 CHECK_QUARTER_MV(2, 2, mx , my )
436 CHECK_QUARTER_MV(0, 2, mx , my )
437 CHECK_QUARTER_MV(2, 2, mx-1, my )
438 CHECK_QUARTER_MV(2, 0, mx-1, my )
439
440 nx= bx;
441 ny= by;
442
443 for(i=0; i<8; i++){
444 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
445 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
446 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2)
447 }
448#endif
449#if 0
450 //outer ring
451 CHECK_QUARTER_MV(1, 3, mx-1, my-1)
452 CHECK_QUARTER_MV(1, 2, mx-1, my-1)
453 CHECK_QUARTER_MV(1, 1, mx-1, my-1)
454 CHECK_QUARTER_MV(2, 1, mx-1, my-1)
455 CHECK_QUARTER_MV(3, 1, mx-1, my-1)
456 CHECK_QUARTER_MV(0, 1, mx , my-1)
457 CHECK_QUARTER_MV(1, 1, mx , my-1)
458 CHECK_QUARTER_MV(2, 1, mx , my-1)
459 CHECK_QUARTER_MV(3, 1, mx , my-1)
460 CHECK_QUARTER_MV(3, 2, mx , my-1)
461 CHECK_QUARTER_MV(3, 3, mx , my-1)
462 CHECK_QUARTER_MV(3, 0, mx , my )
463 CHECK_QUARTER_MV(3, 1, mx , my )
464 CHECK_QUARTER_MV(3, 2, mx , my )
465 CHECK_QUARTER_MV(3, 3, mx , my )
466 CHECK_QUARTER_MV(2, 3, mx , my )
467 CHECK_QUARTER_MV(1, 3, mx , my )
468 CHECK_QUARTER_MV(0, 3, mx , my )
469 CHECK_QUARTER_MV(3, 3, mx-1, my )
470 CHECK_QUARTER_MV(2, 3, mx-1, my )
471 CHECK_QUARTER_MV(1, 3, mx-1, my )
472 CHECK_QUARTER_MV(1, 2, mx-1, my )
473 CHECK_QUARTER_MV(1, 1, mx-1, my )
474 CHECK_QUARTER_MV(1, 0, mx-1, my )
475#endif
476 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4);
477
478 *mx_ptr = bx;
479 *my_ptr = by;
480 }else{
481 *mx_ptr =4*mx;
482 *my_ptr =4*my;
483 }
484
485 return dmin;
486}
487
1457ab52
MN
488
489#define CHECK_MV(x,y)\
490{\
491 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
492 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
ae1dbde1
MN
493 assert((x) >= xmin);\
494 assert((x) <= xmax);\
495 assert((y) >= ymin);\
496 assert((y) <= ymax);\
b07a5980 497/*printf("check_mv %d %d\n", x, y);*/\
1457ab52 498 if(map[index]!=key){\
2750b827 499 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
1457ab52
MN
500 map[index]= key;\
501 score_map[index]= d;\
502 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
b07a5980 503/*printf("score:%d\n", d);*/\
1457ab52
MN
504 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
505 }\
506}
507
b07a5980
MN
508#define CHECK_CLIPED_MV(ax,ay)\
509{\
9c3d33d6
MN
510 const int x= ax;\
511 const int y= ay;\
512 const int x2= FFMAX(xmin, FFMIN(x, xmax));\
513 const int y2= FFMAX(ymin, FFMIN(y, ymax));\
514 CHECK_MV(x2, y2)\
b07a5980
MN
515}
516
1457ab52
MN
517#define CHECK_MV_DIR(x,y,new_dir)\
518{\
519 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
520 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
b07a5980 521/*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
1457ab52 522 if(map[index]!=key){\
2750b827 523 d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
1457ab52
MN
524 map[index]= key;\
525 score_map[index]= d;\
526 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
b07a5980 527/*printf("score:%d\n", d);*/\
1457ab52
MN
528 if(d<dmin){\
529 best[0]=x;\
530 best[1]=y;\
531 dmin=d;\
532 next_dir= new_dir;\
533 }\
534 }\
535}
536
537#define check(x,y,S,v)\
538if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\
539if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\
540if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\
541if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\
542
2750b827 543#define LOAD_COMMON2\
af4091f1 544 uint32_t *map= c->map;\
2750b827
MN
545 const int qpel= flags&FLAG_QPEL;\
546 const int shift= 1+qpel;\
1457ab52 547
2750b827
MN
548static always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
549 int src_index, int ref_index, int const penalty_factor,
550 int size, int h, int flags)
1457ab52 551{
af4091f1 552 MotionEstContext * const c= &s->me;
2750b827 553 me_cmp_func cmpf, chroma_cmpf;
1457ab52 554 int next_dir=-1;
bb198e19 555 LOAD_COMMON
2750b827 556 LOAD_COMMON2
af4091f1 557 int map_generation= c->map_generation;
1457ab52 558
2750b827
MN
559 cmpf= s->dsp.me_cmp[size];
560 chroma_cmpf= s->dsp.me_cmp[size+1];
1457ab52 561
b07a5980
MN
562 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */
563 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
564 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
565 if(map[index]!=key){ //this will be executed only very rarey
2750b827 566 score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
b07a5980
MN
567 map[index]= key;
568 }
569 }
570
1457ab52
MN
571 for(;;){
572 int d;
573 const int dir= next_dir;
574 const int x= best[0];
575 const int y= best[1];
576 next_dir=-1;
577
578//printf("%d", dir);
579 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
580 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
581 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
582 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
583
584 if(next_dir==-1){
585 return dmin;
586 }
587 }
588}
589
2750b827
MN
590static int funny_diamond_search(MpegEncContext * s, int *best, int dmin,
591 int src_index, int ref_index, int const penalty_factor,
592 int size, int h, int flags)
b07a5980 593{
af4091f1 594 MotionEstContext * const c= &s->me;
2750b827 595 me_cmp_func cmpf, chroma_cmpf;
b07a5980 596 int dia_size;
bb198e19 597 LOAD_COMMON
2750b827 598 LOAD_COMMON2
af4091f1 599 int map_generation= c->map_generation;
b07a5980 600
2750b827
MN
601 cmpf= s->dsp.me_cmp[size];
602 chroma_cmpf= s->dsp.me_cmp[size+1];
b07a5980
MN
603
604 for(dia_size=1; dia_size<=4; dia_size++){
605 int dir;
606 const int x= best[0];
607 const int y= best[1];
608
609 if(dia_size&(dia_size-1)) continue;
610
611 if( x + dia_size > xmax
612 || x - dia_size < xmin
613 || y + dia_size > ymax
614 || y - dia_size < ymin)
615 continue;
616
617 for(dir= 0; dir<dia_size; dir+=2){
618 int d;
619
620 CHECK_MV(x + dir , y + dia_size - dir);
621 CHECK_MV(x + dia_size - dir, y - dir );
622 CHECK_MV(x - dir , y - dia_size + dir);
623 CHECK_MV(x - dia_size + dir, y + dir );
624 }
625
626 if(x!=best[0] || y!=best[1])
627 dia_size=0;
628#if 0
629{
630int dx, dy, i;
631static int stats[8*8];
632dx= ABS(x-best[0]);
633dy= ABS(y-best[1]);
634if(dy>dx){
635 dx^=dy; dy^=dx; dx^=dy;
636}
637stats[dy*8 + dx] ++;
638if(256*256*256*64 % (stats[0]+1)==0){
639 for(i=0; i<64; i++){
640 if((i&7)==0) printf("\n");
641 printf("%8d ", stats[i]);
642 }
643 printf("\n");
644}
645}
646#endif
647 }
648 return dmin;
649}
650
651#define SAB_CHECK_MV(ax,ay)\
652{\
653 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
654 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
655/*printf("sab check %d %d\n", ax, ay);*/\
656 if(map[index]!=key){\
2750b827 657 d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
b07a5980
MN
658 map[index]= key;\
659 score_map[index]= d;\
660 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
661/*printf("score: %d\n", d);*/\
662 if(d < minima[minima_count-1].height){\
663 int j=0;\
664 \
665 while(d >= minima[j].height) j++;\
666\
667 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\
668\
669 minima[j].checked= 0;\
670 minima[j].height= d;\
671 minima[j].x= ax;\
672 minima[j].y= ay;\
673 \
674 i=-1;\
675 continue;\
676 }\
677 }\
678}
679
4994af2f 680#define MAX_SAB_SIZE ME_MAP_SIZE
2750b827
MN
681static int sab_diamond_search(MpegEncContext * s, int *best, int dmin,
682 int src_index, int ref_index, int const penalty_factor,
683 int size, int h, int flags)
b07a5980 684{
af4091f1 685 MotionEstContext * const c= &s->me;
2750b827 686 me_cmp_func cmpf, chroma_cmpf;
b07a5980 687 Minima minima[MAX_SAB_SIZE];
af4091f1 688 const int minima_count= ABS(c->dia_size);
b07a5980 689 int i, j;
bb198e19 690 LOAD_COMMON
2750b827 691 LOAD_COMMON2
af4091f1 692 int map_generation= c->map_generation;
b07a5980 693
2750b827
MN
694 cmpf= s->dsp.me_cmp[size];
695 chroma_cmpf= s->dsp.me_cmp[size+1];
b07a5980
MN
696
697 for(j=i=0; i<ME_MAP_SIZE; i++){
698 uint32_t key= map[i];
699
700 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
701
702 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
703
704 assert(j<MAX_SAB_SIZE); //max j = number of predictors
705
706 minima[j].height= score_map[i];
707 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
708 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
709 minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
710 minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
711 minima[j].checked=0;
712 if(minima[j].x || minima[j].y)
713 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
714
715 j++;
716 }
717
718 qsort(minima, j, sizeof(Minima), minima_cmp);
719
720 for(; j<minima_count; j++){
721 minima[j].height=256*256*256*64;
722 minima[j].checked=0;
723 minima[j].x= minima[j].y=0;
724 }
725
726 for(i=0; i<minima_count; i++){
727 const int x= minima[i].x;
728 const int y= minima[i].y;
729 int d;
730
731 if(minima[i].checked) continue;
732
733 if( x >= xmax || x <= xmin
734 || y >= ymax || y <= ymin)
735 continue;
736
737 SAB_CHECK_MV(x-1, y)
738 SAB_CHECK_MV(x+1, y)
739 SAB_CHECK_MV(x , y-1)
740 SAB_CHECK_MV(x , y+1)
741
742 minima[i].checked= 1;
743 }
744
745 best[0]= minima[0].x;
746 best[1]= minima[0].y;
747 dmin= minima[0].height;
748
749 if( best[0] < xmax && best[0] > xmin
750 && best[1] < ymax && best[1] > ymin){
751 int d;
752 //ensure that the refernece samples for hpel refinement are in the map
753 CHECK_MV(best[0]-1, best[1])
754 CHECK_MV(best[0]+1, best[1])
755 CHECK_MV(best[0], best[1]-1)
756 CHECK_MV(best[0], best[1]+1)
757 }
758 return dmin;
759}
760
2750b827
MN
761static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
762 int src_index, int ref_index, int const penalty_factor,
763 int size, int h, int flags)
1457ab52 764{
af4091f1 765 MotionEstContext * const c= &s->me;
2750b827 766 me_cmp_func cmpf, chroma_cmpf;
b07a5980 767 int dia_size;
bb198e19 768 LOAD_COMMON
2750b827 769 LOAD_COMMON2
af4091f1 770 int map_generation= c->map_generation;
1457ab52 771
2750b827
MN
772 cmpf= s->dsp.me_cmp[size];
773 chroma_cmpf= s->dsp.me_cmp[size+1];
1457ab52 774
af4091f1 775 for(dia_size=1; dia_size<=c->dia_size; dia_size++){
1457ab52
MN
776 int dir, start, end;
777 const int x= best[0];
778 const int y= best[1];
779
780 start= FFMAX(0, y + dia_size - ymax);
b07a5980 781 end = FFMIN(dia_size, xmax - x + 1);
1457ab52
MN
782 for(dir= start; dir<end; dir++){
783 int d;
784
785//check(x + dir,y + dia_size - dir,0, a0)
786 CHECK_MV(x + dir , y + dia_size - dir);
787 }
788
789 start= FFMAX(0, x + dia_size - xmax);
b07a5980 790 end = FFMIN(dia_size, y - ymin + 1);
1457ab52
MN
791 for(dir= start; dir<end; dir++){
792 int d;
793
794//check(x + dia_size - dir, y - dir,0, a1)
795 CHECK_MV(x + dia_size - dir, y - dir );
796 }
797
798 start= FFMAX(0, -y + dia_size + ymin );
b07a5980 799 end = FFMIN(dia_size, x - xmin + 1);
1457ab52
MN
800 for(dir= start; dir<end; dir++){
801 int d;
802
803//check(x - dir,y - dia_size + dir,0, a2)
804 CHECK_MV(x - dir , y - dia_size + dir);
805 }
806
807 start= FFMAX(0, -x + dia_size + xmin );
b07a5980 808 end = FFMIN(dia_size, ymax - y + 1);
1457ab52
MN
809 for(dir= start; dir<end; dir++){
810 int d;
811
812//check(x - dia_size + dir, y + dir,0, a3)
813 CHECK_MV(x - dia_size + dir, y + dir );
814 }
815
816 if(x!=best[0] || y!=best[1])
817 dia_size=0;
b07a5980
MN
818#if 0
819{
820int dx, dy, i;
821static int stats[8*8];
822dx= ABS(x-best[0]);
823dy= ABS(y-best[1]);
824stats[dy*8 + dx] ++;
825if(256*256*256*64 % (stats[0]+1)==0){
826 for(i=0; i<64; i++){
827 if((i&7)==0) printf("\n");
828 printf("%6d ", stats[i]);
829 }
830 printf("\n");
831}
832}
833#endif
1457ab52
MN
834 }
835 return dmin;
836}
837
2750b827
MN
838static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
839 int src_index, int ref_index, int const penalty_factor,
840 int size, int h, int flags){
af4091f1
MN
841 MotionEstContext * const c= &s->me;
842 if(c->dia_size==-1)
2750b827 843 return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
af4091f1 844 else if(c->dia_size<-1)
2750b827 845 return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
af4091f1 846 else if(c->dia_size<2)
2750b827
MN
847 return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
848 else
849 return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
850}
851
852static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
853 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
26efc54e 854 int ref_mv_scale, int flags, int size, int h)
1457ab52 855{
af4091f1 856 MotionEstContext * const c= &s->me;
1457ab52 857 int best[2]={0, 0};
2750b827 858 int d, dmin;
1457ab52 859 int map_generation;
af4091f1 860 const int penalty_factor= c->penalty_factor;
bb198e19
MN
861 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
862 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
2750b827
MN
863 me_cmp_func cmpf, chroma_cmpf;
864
bb198e19 865 LOAD_COMMON
2750b827 866 LOAD_COMMON2
1457ab52 867
2750b827
MN
868 cmpf= s->dsp.me_cmp[size];
869 chroma_cmpf= s->dsp.me_cmp[size+1];
1457ab52 870
af4091f1 871 map_generation= update_map_generation(c);
1457ab52 872
26efc54e 873 assert(cmpf);
2750b827 874 dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
1457ab52
MN
875 map[0]= map_generation;
876 score_map[0]= dmin;
877
878 /* first line */
9c3d33d6 879 if (s->first_slice_line) {
1457ab52 880 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
b07a5980
MN
881 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
882 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1457ab52 883 }else{
26efc54e 884 if(dmin<h*h && ( P_LEFT[0] |P_LEFT[1]
1457ab52 885 |P_TOP[0] |P_TOP[1]
b07a5980 886 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
1457ab52
MN
887 *mx_ptr= 0;
888 *my_ptr= 0;
af4091f1 889 c->skip=1;
1457ab52
MN
890 return dmin;
891 }
892 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
26efc54e 893 if(dmin>h*h*2){
b07a5980
MN
894 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
895 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1457ab52
MN
896 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
897 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
898 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
899 }
900 }
26efc54e 901 if(dmin>h*h*4){
af4091f1 902 if(c->pre_pass){
f931ff7b
MN
903 CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
904 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
9c3d33d6
MN
905 if(!s->first_slice_line)
906 CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
907 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
f931ff7b
MN
908 }else{
909 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
910 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
80ee9fc0 911 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
9c3d33d6
MN
912 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
913 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
f931ff7b 914 }
1457ab52 915 }
b07a5980 916
1f202b0d
MN
917 if(c->avctx->last_predictor_count){
918 const int count= c->avctx->last_predictor_count;
b07a5980
MN
919 const int xstart= FFMAX(0, s->mb_x - count);
920 const int ystart= FFMAX(0, s->mb_y - count);
921 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
922 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
923 int mb_y;
924
925 for(mb_y=ystart; mb_y<yend; mb_y++){
926 int mb_x;
927 for(mb_x=xstart; mb_x<xend; mb_x++){
928 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
929 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
930 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
931
932 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
933 CHECK_MV(mx,my)
1457ab52
MN
934 }
935 }
936 }
b07a5980 937
1457ab52 938//check(best[0],best[1],0, b0)
2750b827 939 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
1457ab52
MN
940
941//check(best[0],best[1],0, b1)
942 *mx_ptr= best[0];
943 *my_ptr= best[1];
944
945// printf("%d %d %d \n", best[0], best[1], dmin);
946 return dmin;
947}
948
2750b827 949//this function is dedicated to the braindamaged gcc
26efc54e 950inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
2750b827 951 int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
26efc54e 952 int ref_mv_scale, int size, int h)
2750b827 953{
af4091f1 954 MotionEstContext * const c= &s->me;
2750b827 955//FIXME convert other functions in the same way if faster
26efc54e
MN
956 if(c->flags==0 && h==16 && size==0){
957 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
2750b827
MN
958// case FLAG_QPEL:
959// return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
26efc54e
MN
960 }else{
961 return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
2750b827
MN
962 }
963}
964
965static int epzs_motion_search4(MpegEncContext * s,
966 int *mx_ptr, int *my_ptr, int P[10][2],
967 int src_index, int ref_index, int16_t (*last_mv)[2],
968 int ref_mv_scale)
1457ab52 969{
af4091f1 970 MotionEstContext * const c= &s->me;
1457ab52
MN
971 int best[2]={0, 0};
972 int d, dmin;
1457ab52 973 int map_generation;
af4091f1 974 const int penalty_factor= c->penalty_factor;
1457ab52 975 const int size=1;
bb198e19 976 const int h=8;
7bc9090a
MN
977 const int ref_mv_stride= s->mb_stride;
978 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
2750b827 979 me_cmp_func cmpf, chroma_cmpf;
bb198e19 980 LOAD_COMMON
af4091f1 981 int flags= c->flags;
2750b827 982 LOAD_COMMON2
1457ab52 983
2750b827
MN
984 cmpf= s->dsp.me_cmp[size];
985 chroma_cmpf= s->dsp.me_cmp[size+1];
1457ab52 986
af4091f1 987 map_generation= update_map_generation(c);
1457ab52
MN
988
989 dmin = 1000000;
990//printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
991 /* first line */
9c3d33d6 992 if (s->first_slice_line) {
1457ab52 993 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
b07a5980
MN
994 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
995 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1457ab52
MN
996 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
997 }else{
998 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
999 //FIXME try some early stop
1000 if(dmin>64*2){
1001 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1002 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1003 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1004 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
b07a5980
MN
1005 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1006 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1457ab52
MN
1007 }
1008 }
1009 if(dmin>64*4){
b07a5980
MN
1010 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1011 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
80ee9fc0 1012 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
9c3d33d6
MN
1013 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1014 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
1457ab52
MN
1015 }
1016
2750b827 1017 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
bb198e19
MN
1018
1019 *mx_ptr= best[0];
1020 *my_ptr= best[1];
1021
1022// printf("%d %d %d \n", best[0], best[1], dmin);
1023 return dmin;
1024}
1025
1026//try to merge with above FIXME (needs PSNR test)
2750b827
MN
1027static int epzs_motion_search2(MpegEncContext * s,
1028 int *mx_ptr, int *my_ptr, int P[10][2],
1029 int src_index, int ref_index, int16_t (*last_mv)[2],
1030 int ref_mv_scale)
bb198e19 1031{
af4091f1 1032 MotionEstContext * const c= &s->me;
bb198e19
MN
1033 int best[2]={0, 0};
1034 int d, dmin;
bb198e19 1035 int map_generation;
af4091f1 1036 const int penalty_factor= c->penalty_factor;
bb198e19
MN
1037 const int size=0; //FIXME pass as arg
1038 const int h=8;
1039 const int ref_mv_stride= s->mb_stride;
1040 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
2750b827 1041 me_cmp_func cmpf, chroma_cmpf;
bb198e19 1042 LOAD_COMMON
af4091f1 1043 int flags= c->flags;
2750b827 1044 LOAD_COMMON2
bb198e19 1045
2750b827
MN
1046 cmpf= s->dsp.me_cmp[size];
1047 chroma_cmpf= s->dsp.me_cmp[size+1];
bb198e19 1048
af4091f1 1049 map_generation= update_map_generation(c);
bb198e19
MN
1050
1051 dmin = 1000000;
1052//printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
1053 /* first line */
9c3d33d6 1054 if (s->first_slice_line) {
bb198e19
MN
1055 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1056 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1057 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1058 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1059 }else{
1060 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
1061 //FIXME try some early stop
1062 if(dmin>64*2){
1063 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
1064 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
1065 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
1066 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
1067 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
1068 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
1069 }
1070 }
1071 if(dmin>64*4){
1072 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
1073 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
80ee9fc0 1074 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
9c3d33d6
MN
1075 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
1076 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
bb198e19
MN
1077 }
1078
2750b827 1079 dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
b07a5980 1080
1457ab52
MN
1081 *mx_ptr= best[0];
1082 *my_ptr= best[1];
1083
1084// printf("%d %d %d \n", best[0], best[1], dmin);
1085 return dmin;
1086}