Mercurial > libavcodec.hg
annotate motion_est_template.c @ 1805:5dd5b2052f5b libavcodec
mpeg4 b frame + adaptive quantization assertion failure fix
author | michael |
---|---|
date | Sat, 14 Feb 2004 03:49:43 +0000 |
parents | e039d79185c2 |
children | a3c60fa850dc |
rev | line source |
---|---|
936 | 1 /* |
2 * Motion estimation | |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1708
diff
changeset
|
3 * Copyright (c) 2002-2004 Michael Niedermayer |
936 | 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 */ | |
1106 | 20 |
21 /** | |
22 * @file motion_est_template.c | |
23 * Motion estimation template. | |
24 */ | |
1708 | 25 //FIXME ref2_y next_pic? |
936 | 26 //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...) |
27 //Note, the last line is there to kill these ugly unused var warnings | |
1708 | 28 #define LOAD_COMMON\ |
936 | 29 uint32_t * const score_map= s->me.score_map;\ |
30 const int time_pp= s->pp_time;\ | |
31 const int time_pb= s->pb_time;\ | |
1708 | 32 const int xmin= s->me.xmin;\ |
33 const int ymin= s->me.ymin;\ | |
34 const int xmax= s->me.xmax;\ | |
35 const int ymax= s->me.ymax;\ | |
36 uint8_t * const src_y= src_data[0];\ | |
37 uint8_t * const src_u= src_data[1];\ | |
38 uint8_t * const src_v= src_data[2];\ | |
39 uint8_t * const ref_y= ref_data[0];\ | |
40 uint8_t * const ref_u= ref_data[1];\ | |
41 uint8_t * const ref_v= ref_data[2];\ | |
936 | 42 op_pixels_func (*hpel_put)[4];\ |
43 op_pixels_func (*hpel_avg)[4]= &s->dsp.avg_pixels_tab[size];\ | |
44 op_pixels_func (*chroma_hpel_put)[4];\ | |
45 qpel_mc_func (*qpel_put)[16];\ | |
46 qpel_mc_func (*qpel_avg)[16]= &s->dsp.avg_qpel_pixels_tab[size];\ | |
1266 | 47 const __attribute__((unused)) int unu= time_pp + time_pb + (size_t)src_u + (size_t)src_v + (size_t)ref_u + (size_t)ref_v\ |
1708 | 48 + (size_t)hpel_avg + (size_t)qpel_avg + (size_t)score_map\ |
49 + xmin + xmax + ymin + ymax;\ | |
936 | 50 if(s->no_rounding /*FIXME b_type*/){\ |
51 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];\ | |
52 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];\ | |
53 qpel_put= &s->dsp.put_no_rnd_qpel_pixels_tab[size];\ | |
54 }else{\ | |
55 hpel_put=& s->dsp.put_pixels_tab[size];\ | |
56 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];\ | |
57 qpel_put= &s->dsp.put_qpel_pixels_tab[size];\ | |
58 } | |
59 | |
60 | |
61 #ifdef CMP_HPEL | |
62 | |
63 #define CHECK_HALF_MV(dx, dy, x, y)\ | |
64 {\ | |
65 const int hx= 2*(x)+(dx);\ | |
66 const int hy= 2*(y)+(dy);\ | |
67 CMP_HPEL(d, dx, dy, x, y, size);\ | |
68 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ | |
69 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ | |
70 } | |
71 | |
72 #if 0 | |
73 static int RENAME(hpel_motion_search)(MpegEncContext * s, | |
74 int *mx_ptr, int *my_ptr, int dmin, | |
1708 | 75 int pred_x, int pred_y, uint8_t *ref_data[3], |
76 int size, uint8_t * const mv_penalty) | |
936 | 77 { |
78 const int xx = 16 * s->mb_x + 8*(n&1); | |
79 const int yy = 16 * s->mb_y + 8*(n>>1); | |
80 const int mx = *mx_ptr; | |
81 const int my = *my_ptr; | |
948 | 82 const int penalty_factor= s->me.sub_penalty_factor; |
936 | 83 |
1708 | 84 LOAD_COMMON |
936 | 85 |
86 // INIT; | |
87 //FIXME factorize | |
88 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub; | |
89 | |
90 if(s->no_rounding /*FIXME b_type*/){ | |
91 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size]; | |
92 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1]; | |
93 }else{ | |
94 hpel_put=& s->dsp.put_pixels_tab[size]; | |
95 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1]; | |
96 } | |
97 cmp= s->dsp.me_cmp[size]; | |
98 chroma_cmp= s->dsp.me_cmp[size+1]; | |
99 cmp_sub= s->dsp.me_sub_cmp[size]; | |
100 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
101 | |
102 if(s->me.skip){ //FIXME somehow move up (benchmark) | |
103 *mx_ptr = 0; | |
104 *my_ptr = 0; | |
105 return dmin; | |
106 } | |
107 | |
108 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
109 CMP_HPEL(dmin, 0, 0, mx, my, size); | |
110 if(mx || my) | |
111 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; | |
112 } | |
113 | |
114 if (mx > xmin && mx < xmax && | |
115 my > ymin && my < ymax) { | |
116 int bx=2*mx, by=2*my; | |
117 int d= dmin; | |
118 | |
119 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
120 CHECK_HALF_MV(0, 1, mx , my-1) | |
121 CHECK_HALF_MV(1, 1, mx , my-1) | |
122 CHECK_HALF_MV(1, 0, mx-1, my ) | |
123 CHECK_HALF_MV(1, 0, mx , my ) | |
124 CHECK_HALF_MV(1, 1, mx-1, my ) | |
125 CHECK_HALF_MV(0, 1, mx , my ) | |
126 CHECK_HALF_MV(1, 1, mx , my ) | |
127 | |
948 | 128 assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2); |
936 | 129 |
130 *mx_ptr = bx; | |
131 *my_ptr = by; | |
132 }else{ | |
133 *mx_ptr =2*mx; | |
134 *my_ptr =2*my; | |
135 } | |
136 | |
137 return dmin; | |
138 } | |
139 | |
140 #else | |
141 static int RENAME(hpel_motion_search)(MpegEncContext * s, | |
142 int *mx_ptr, int *my_ptr, int dmin, | |
1708 | 143 int pred_x, int pred_y, uint8_t *src_data[3], |
144 uint8_t *ref_data[3], int stride, int uvstride, | |
145 int size, int h, uint8_t * const mv_penalty) | |
936 | 146 { |
147 const int mx = *mx_ptr; | |
148 const int my = *my_ptr; | |
149 const int penalty_factor= s->me.sub_penalty_factor; | |
150 me_cmp_func cmp_sub, chroma_cmp_sub; | |
1013 | 151 int bx=2*mx, by=2*my; |
936 | 152 |
1708 | 153 LOAD_COMMON |
936 | 154 |
155 //FIXME factorize | |
156 | |
157 cmp_sub= s->dsp.me_sub_cmp[size]; | |
158 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
159 | |
160 if(s->me.skip){ //FIXME move out of hpel? | |
161 *mx_ptr = 0; | |
162 *my_ptr = 0; | |
163 return dmin; | |
164 } | |
165 | |
166 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
167 CMP_HPEL(dmin, 0, 0, mx, my, size); | |
1011 | 168 if(mx || my || size>0) |
936 | 169 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; |
170 } | |
171 | |
172 if (mx > xmin && mx < xmax && | |
173 my > ymin && my < ymax) { | |
174 int d= dmin; | |
175 const int index= (my<<ME_MAP_SHIFT) + mx; | |
176 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] | |
948 | 177 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*s->me.penalty_factor; |
936 | 178 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)] |
948 | 179 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*s->me.penalty_factor; |
936 | 180 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)] |
948 | 181 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*s->me.penalty_factor; |
936 | 182 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] |
948 | 183 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*s->me.penalty_factor; |
184 | |
1013 | 185 #if 1 |
948 | 186 int key; |
187 int map_generation= s->me.map_generation; | |
1419 | 188 #ifndef NDEBUG |
948 | 189 uint32_t *map= s->me.map; |
1419 | 190 #endif |
948 | 191 key= ((my-1)<<ME_MAP_MV_BITS) + (mx) + map_generation; |
192 assert(map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); | |
193 key= ((my+1)<<ME_MAP_MV_BITS) + (mx) + map_generation; | |
194 assert(map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] == key); | |
195 key= ((my)<<ME_MAP_MV_BITS) + (mx+1) + map_generation; | |
196 assert(map[(index+1)&(ME_MAP_SIZE-1)] == key); | |
197 key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation; | |
198 assert(map[(index-1)&(ME_MAP_SIZE-1)] == key); | |
199 #endif | |
936 | 200 if(t<=b){ |
201 CHECK_HALF_MV(0, 1, mx ,my-1) | |
202 if(l<=r){ | |
203 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
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-1, my ) | |
210 }else{ | |
211 CHECK_HALF_MV(1, 1, mx , my-1) | |
212 if(t+l<=b+r){ | |
213 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
214 }else{ | |
215 CHECK_HALF_MV(1, 1, mx , my ) | |
216 } | |
217 CHECK_HALF_MV(1, 0, mx , my ) | |
218 } | |
219 }else{ | |
220 if(l<=r){ | |
221 if(t+l<=b+r){ | |
222 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
223 }else{ | |
224 CHECK_HALF_MV(1, 1, mx , my ) | |
225 } | |
226 CHECK_HALF_MV(1, 0, mx-1, my) | |
227 CHECK_HALF_MV(1, 1, mx-1, my) | |
228 }else{ | |
229 if(t+r<=b+l){ | |
230 CHECK_HALF_MV(1, 1, mx , my-1) | |
231 }else{ | |
232 CHECK_HALF_MV(1, 1, mx-1, my) | |
233 } | |
234 CHECK_HALF_MV(1, 0, mx , my) | |
235 CHECK_HALF_MV(1, 1, mx , my) | |
236 } | |
237 CHECK_HALF_MV(0, 1, mx , my) | |
238 } | |
239 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2); | |
240 } | |
241 | |
1013 | 242 *mx_ptr = bx; |
243 *my_ptr = by; | |
244 | |
936 | 245 return dmin; |
246 } | |
247 #endif | |
248 | |
1708 | 249 static int RENAME(hpel_get_mb_score)(MpegEncContext * s, int mx, int my, int pred_x, int pred_y, uint8_t *src_data[3], |
250 uint8_t *ref_data[3], int stride, int uvstride, | |
1162 | 251 uint8_t * const mv_penalty) |
1013 | 252 { |
253 // const int check_luma= s->dsp.me_sub_cmp != s->dsp.mb_cmp; | |
254 const int size= 0; | |
1708 | 255 const int h= 16; |
1013 | 256 const int penalty_factor= s->me.mb_penalty_factor; |
257 me_cmp_func cmp_sub, chroma_cmp_sub; | |
258 int d; | |
259 | |
1708 | 260 LOAD_COMMON |
1013 | 261 |
262 //FIXME factorize | |
263 | |
264 cmp_sub= s->dsp.mb_cmp[size]; | |
265 chroma_cmp_sub= s->dsp.mb_cmp[size+1]; | |
266 | |
267 assert(!s->me.skip); | |
268 assert(s->avctx->me_sub_cmp != s->avctx->mb_cmp); | |
269 | |
270 CMP_HPEL(d, mx&1, my&1, mx>>1, my>>1, size); | |
271 //FIXME check cbp before adding penalty for (0,0) vector | |
272 if(mx || my || size>0) | |
273 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor; | |
274 | |
275 return d; | |
276 } | |
277 | |
936 | 278 #endif /* CMP_HPEL */ |
279 | |
1013 | 280 |
281 | |
936 | 282 #ifdef CMP_QPEL |
283 | |
284 #define CHECK_QUARTER_MV(dx, dy, x, y)\ | |
285 {\ | |
286 const int hx= 4*(x)+(dx);\ | |
287 const int hy= 4*(y)+(dy);\ | |
288 CMP_QPEL(d, dx, dy, x, y, size);\ | |
289 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ | |
290 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ | |
291 } | |
292 | |
293 static int RENAME(qpel_motion_search)(MpegEncContext * s, | |
294 int *mx_ptr, int *my_ptr, int dmin, | |
1708 | 295 int pred_x, int pred_y, uint8_t *src_data[3], |
296 uint8_t *ref_data[3], int stride, int uvstride, | |
297 int size, int h, uint8_t * const mv_penalty) | |
936 | 298 { |
299 const int mx = *mx_ptr; | |
300 const int my = *my_ptr; | |
301 const int penalty_factor= s->me.sub_penalty_factor; | |
302 const int map_generation= s->me.map_generation; | |
954 | 303 const int subpel_quality= s->avctx->me_subpel_quality; |
936 | 304 uint32_t *map= s->me.map; |
305 me_cmp_func cmp, chroma_cmp; | |
306 me_cmp_func cmp_sub, chroma_cmp_sub; | |
307 | |
1708 | 308 LOAD_COMMON |
936 | 309 |
310 cmp= s->dsp.me_cmp[size]; | |
311 chroma_cmp= s->dsp.me_cmp[size+1]; //factorize FIXME | |
312 //FIXME factorize | |
313 | |
314 cmp_sub= s->dsp.me_sub_cmp[size]; | |
315 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
316 | |
317 if(s->me.skip){ //FIXME somehow move up (benchmark) | |
318 *mx_ptr = 0; | |
319 *my_ptr = 0; | |
320 return dmin; | |
321 } | |
322 | |
323 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
324 CMP_QPEL(dmin, 0, 0, mx, my, size); | |
1011 | 325 if(mx || my || size>0) |
936 | 326 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor; |
327 } | |
328 | |
329 if (mx > xmin && mx < xmax && | |
330 my > ymin && my < ymax) { | |
331 int bx=4*mx, by=4*my; | |
332 int d= dmin; | |
333 int i, nx, ny; | |
334 const int index= (my<<ME_MAP_SHIFT) + mx; | |
335 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; | |
336 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
337 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
338 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; | |
339 const int c= score_map[(index )&(ME_MAP_SIZE-1)]; | |
340 int best[8]; | |
341 int best_pos[8][2]; | |
342 | |
343 memset(best, 64, sizeof(int)*8); | |
344 #if 1 | |
954 | 345 if(s->me.dia_size>=2){ |
936 | 346 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; |
347 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
348 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
349 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
350 | |
351 for(ny= -3; ny <= 3; ny++){ | |
352 for(nx= -3; nx <= 3; nx++){ | |
353 const int t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t; | |
354 const int c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c; | |
355 const int b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b; | |
356 int score= ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2; | |
357 int i; | |
358 | |
359 if((nx&3)==0 && (ny&3)==0) continue; | |
360 | |
361 score += 1024*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; | |
362 | |
363 // if(nx&1) score-=1024*s->me.penalty_factor; | |
364 // if(ny&1) score-=1024*s->me.penalty_factor; | |
365 | |
366 for(i=0; i<8; i++){ | |
367 if(score < best[i]){ | |
368 memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); | |
369 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); | |
370 best[i]= score; | |
371 best_pos[i][0]= nx + 4*mx; | |
372 best_pos[i][1]= ny + 4*my; | |
373 break; | |
374 } | |
375 } | |
376 } | |
377 } | |
378 }else{ | |
379 int tl; | |
380 const int cx = 4*(r - l); | |
381 const int cx2= r + l - 2*c; | |
382 const int cy = 4*(b - t); | |
383 const int cy2= b + t - 2*c; | |
384 int cxy; | |
385 | |
386 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME | |
387 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
388 }else{ | |
389 CMP(tl, mx-1, my-1, size); //FIXME wrong if chroma me is different | |
390 } | |
391 | |
392 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c; | |
393 | |
394 assert(16*cx2 + 4*cx + 32*c == 32*r); | |
395 assert(16*cx2 - 4*cx + 32*c == 32*l); | |
396 assert(16*cy2 + 4*cy + 32*c == 32*b); | |
397 assert(16*cy2 - 4*cy + 32*c == 32*t); | |
398 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl); | |
399 | |
400 for(ny= -3; ny <= 3; ny++){ | |
401 for(nx= -3; nx <= 3; nx++){ | |
402 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor | |
403 int i; | |
404 | |
405 if((nx&3)==0 && (ny&3)==0) continue; | |
406 | |
407 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; | |
408 // if(nx&1) score-=32*s->me.penalty_factor; | |
409 // if(ny&1) score-=32*s->me.penalty_factor; | |
410 | |
411 for(i=0; i<8; i++){ | |
412 if(score < best[i]){ | |
413 memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); | |
414 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); | |
415 best[i]= score; | |
416 best_pos[i][0]= nx + 4*mx; | |
417 best_pos[i][1]= ny + 4*my; | |
418 break; | |
419 } | |
420 } | |
421 } | |
422 } | |
423 } | |
954 | 424 for(i=0; i<subpel_quality; i++){ |
936 | 425 nx= best_pos[i][0]; |
426 ny= best_pos[i][1]; | |
427 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2) | |
428 } | |
954 | 429 |
936 | 430 #if 0 |
954 | 431 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; |
432 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
433 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
434 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
435 // if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){ | |
436 if(tl<br){ | |
437 | |
438 // nx= FFMAX(4*mx - bx, bx - 4*mx); | |
439 // ny= FFMAX(4*my - by, by - 4*my); | |
936 | 440 |
954 | 441 static int stats[7][7], count; |
442 count++; | |
443 stats[4*mx - bx + 3][4*my - by + 3]++; | |
444 if(256*256*256*64 % count ==0){ | |
445 for(i=0; i<49; i++){ | |
446 if((i%7)==0) printf("\n"); | |
936 | 447 printf("%6d ", stats[0][i]); |
448 } | |
449 printf("\n"); | |
450 } | |
954 | 451 } |
936 | 452 #endif |
453 #else | |
454 | |
455 CHECK_QUARTER_MV(2, 2, mx-1, my-1) | |
456 CHECK_QUARTER_MV(0, 2, mx , my-1) | |
457 CHECK_QUARTER_MV(2, 2, mx , my-1) | |
458 CHECK_QUARTER_MV(2, 0, mx , my ) | |
459 CHECK_QUARTER_MV(2, 2, mx , my ) | |
460 CHECK_QUARTER_MV(0, 2, mx , my ) | |
461 CHECK_QUARTER_MV(2, 2, mx-1, my ) | |
462 CHECK_QUARTER_MV(2, 0, mx-1, my ) | |
463 | |
464 nx= bx; | |
465 ny= by; | |
466 | |
467 for(i=0; i<8; i++){ | |
468 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1}; | |
469 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1}; | |
470 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2) | |
471 } | |
472 #endif | |
473 #if 0 | |
474 //outer ring | |
475 CHECK_QUARTER_MV(1, 3, mx-1, my-1) | |
476 CHECK_QUARTER_MV(1, 2, mx-1, my-1) | |
477 CHECK_QUARTER_MV(1, 1, mx-1, my-1) | |
478 CHECK_QUARTER_MV(2, 1, mx-1, my-1) | |
479 CHECK_QUARTER_MV(3, 1, mx-1, my-1) | |
480 CHECK_QUARTER_MV(0, 1, mx , my-1) | |
481 CHECK_QUARTER_MV(1, 1, mx , my-1) | |
482 CHECK_QUARTER_MV(2, 1, mx , my-1) | |
483 CHECK_QUARTER_MV(3, 1, mx , my-1) | |
484 CHECK_QUARTER_MV(3, 2, mx , my-1) | |
485 CHECK_QUARTER_MV(3, 3, mx , my-1) | |
486 CHECK_QUARTER_MV(3, 0, mx , my ) | |
487 CHECK_QUARTER_MV(3, 1, mx , my ) | |
488 CHECK_QUARTER_MV(3, 2, mx , my ) | |
489 CHECK_QUARTER_MV(3, 3, mx , my ) | |
490 CHECK_QUARTER_MV(2, 3, mx , my ) | |
491 CHECK_QUARTER_MV(1, 3, mx , my ) | |
492 CHECK_QUARTER_MV(0, 3, mx , my ) | |
493 CHECK_QUARTER_MV(3, 3, mx-1, my ) | |
494 CHECK_QUARTER_MV(2, 3, mx-1, my ) | |
495 CHECK_QUARTER_MV(1, 3, mx-1, my ) | |
496 CHECK_QUARTER_MV(1, 2, mx-1, my ) | |
497 CHECK_QUARTER_MV(1, 1, mx-1, my ) | |
498 CHECK_QUARTER_MV(1, 0, mx-1, my ) | |
499 #endif | |
500 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4); | |
501 | |
502 *mx_ptr = bx; | |
503 *my_ptr = by; | |
504 }else{ | |
505 *mx_ptr =4*mx; | |
506 *my_ptr =4*my; | |
507 } | |
508 | |
509 return dmin; | |
510 } | |
511 | |
1708 | 512 static int RENAME(qpel_get_mb_score)(MpegEncContext * s, int mx, int my, int pred_x, int pred_y, uint8_t *src_data[3], |
513 uint8_t *ref_data[3], int stride, int uvstride, | |
1162 | 514 uint8_t * const mv_penalty) |
1013 | 515 { |
516 const int size= 0; | |
1708 | 517 const int h= 16; |
1013 | 518 const int penalty_factor= s->me.mb_penalty_factor; |
519 me_cmp_func cmp_sub, chroma_cmp_sub; | |
520 int d; | |
521 | |
1708 | 522 LOAD_COMMON |
1013 | 523 |
524 //FIXME factorize | |
525 | |
526 cmp_sub= s->dsp.mb_cmp[size]; | |
527 chroma_cmp_sub= s->dsp.mb_cmp[size+1]; | |
528 | |
529 assert(!s->me.skip); | |
530 assert(s->avctx->me_sub_cmp != s->avctx->mb_cmp); | |
531 | |
532 CMP_QPEL(d, mx&3, my&3, mx>>2, my>>2, size); | |
533 //FIXME check cbp before adding penalty for (0,0) vector | |
534 if(mx || my || size>0) | |
535 d += (mv_penalty[mx - pred_x] + mv_penalty[my - pred_y])*penalty_factor; | |
536 | |
537 return d; | |
538 } | |
539 | |
540 | |
936 | 541 #endif /* CMP_QPEL */ |
542 | |
543 #define CHECK_MV(x,y)\ | |
544 {\ | |
545 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ | |
546 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ | |
948 | 547 /*printf("check_mv %d %d\n", x, y);*/\ |
936 | 548 if(map[index]!=key){\ |
549 CMP(d, x, y, size);\ | |
550 map[index]= key;\ | |
551 score_map[index]= d;\ | |
552 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ | |
948 | 553 /*printf("score:%d\n", d);*/\ |
936 | 554 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\ |
555 }\ | |
556 } | |
557 | |
948 | 558 #define CHECK_CLIPED_MV(ax,ay)\ |
559 {\ | |
1799 | 560 const int x= ax;\ |
561 const int y= ay;\ | |
562 const int x2= FFMAX(xmin, FFMIN(x, xmax));\ | |
563 const int y2= FFMAX(ymin, FFMIN(y, ymax));\ | |
564 CHECK_MV(x2, y2)\ | |
948 | 565 } |
566 | |
936 | 567 #define CHECK_MV_DIR(x,y,new_dir)\ |
568 {\ | |
569 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ | |
570 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ | |
948 | 571 /*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\ |
936 | 572 if(map[index]!=key){\ |
573 CMP(d, x, y, size);\ | |
574 map[index]= key;\ | |
575 score_map[index]= d;\ | |
576 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ | |
948 | 577 /*printf("score:%d\n", d);*/\ |
936 | 578 if(d<dmin){\ |
579 best[0]=x;\ | |
580 best[1]=y;\ | |
581 dmin=d;\ | |
582 next_dir= new_dir;\ | |
583 }\ | |
584 }\ | |
585 } | |
586 | |
587 #define check(x,y,S,v)\ | |
588 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\ | |
589 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\ | |
590 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\ | |
591 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\ | |
592 | |
593 | |
594 static inline int RENAME(small_diamond_search)(MpegEncContext * s, int *best, int dmin, | |
1708 | 595 uint8_t *src_data[3], |
596 uint8_t *ref_data[3], int stride, int uvstride, | |
936 | 597 int const pred_x, int const pred_y, int const penalty_factor, |
1708 | 598 int const shift, |
599 uint32_t *map, int map_generation, int size, int h, uint8_t * const mv_penalty | |
936 | 600 ) |
601 { | |
602 me_cmp_func cmp, chroma_cmp; | |
603 int next_dir=-1; | |
1708 | 604 LOAD_COMMON |
936 | 605 |
606 cmp= s->dsp.me_cmp[size]; | |
607 chroma_cmp= s->dsp.me_cmp[size+1]; | |
608 | |
948 | 609 { /* ensure that the best point is in the MAP as h/qpel refinement needs it */ |
610 const int key= (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation; | |
611 const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1); | |
612 if(map[index]!=key){ //this will be executed only very rarey | |
613 CMP(score_map[index], best[0], best[1], size); | |
614 map[index]= key; | |
615 } | |
616 } | |
617 | |
936 | 618 for(;;){ |
619 int d; | |
620 const int dir= next_dir; | |
621 const int x= best[0]; | |
622 const int y= best[1]; | |
623 next_dir=-1; | |
624 | |
625 //printf("%d", dir); | |
626 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0) | |
627 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1) | |
628 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2) | |
629 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3) | |
630 | |
631 if(next_dir==-1){ | |
632 return dmin; | |
633 } | |
634 } | |
635 } | |
636 | |
948 | 637 static inline int RENAME(funny_diamond_search)(MpegEncContext * s, int *best, int dmin, |
1708 | 638 uint8_t *src_data[3], |
639 uint8_t *ref_data[3], int stride, int uvstride, | |
948 | 640 int const pred_x, int const pred_y, int const penalty_factor, |
1708 | 641 int const shift, |
642 uint32_t *map, int map_generation, int size, int h, uint8_t * const mv_penalty | |
948 | 643 ) |
644 { | |
645 me_cmp_func cmp, chroma_cmp; | |
646 int dia_size; | |
1708 | 647 LOAD_COMMON |
948 | 648 |
649 cmp= s->dsp.me_cmp[size]; | |
650 chroma_cmp= s->dsp.me_cmp[size+1]; | |
651 | |
652 for(dia_size=1; dia_size<=4; dia_size++){ | |
653 int dir; | |
654 const int x= best[0]; | |
655 const int y= best[1]; | |
656 | |
657 if(dia_size&(dia_size-1)) continue; | |
658 | |
659 if( x + dia_size > xmax | |
660 || x - dia_size < xmin | |
661 || y + dia_size > ymax | |
662 || y - dia_size < ymin) | |
663 continue; | |
664 | |
665 for(dir= 0; dir<dia_size; dir+=2){ | |
666 int d; | |
667 | |
668 CHECK_MV(x + dir , y + dia_size - dir); | |
669 CHECK_MV(x + dia_size - dir, y - dir ); | |
670 CHECK_MV(x - dir , y - dia_size + dir); | |
671 CHECK_MV(x - dia_size + dir, y + dir ); | |
672 } | |
673 | |
674 if(x!=best[0] || y!=best[1]) | |
675 dia_size=0; | |
676 #if 0 | |
677 { | |
678 int dx, dy, i; | |
679 static int stats[8*8]; | |
680 dx= ABS(x-best[0]); | |
681 dy= ABS(y-best[1]); | |
682 if(dy>dx){ | |
683 dx^=dy; dy^=dx; dx^=dy; | |
684 } | |
685 stats[dy*8 + dx] ++; | |
686 if(256*256*256*64 % (stats[0]+1)==0){ | |
687 for(i=0; i<64; i++){ | |
688 if((i&7)==0) printf("\n"); | |
689 printf("%8d ", stats[i]); | |
690 } | |
691 printf("\n"); | |
692 } | |
693 } | |
694 #endif | |
695 } | |
696 return dmin; | |
697 } | |
698 | |
699 #define SAB_CHECK_MV(ax,ay)\ | |
700 {\ | |
701 const int key= ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\ | |
702 const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\ | |
703 /*printf("sab check %d %d\n", ax, ay);*/\ | |
704 if(map[index]!=key){\ | |
705 CMP(d, ax, ay, size);\ | |
706 map[index]= key;\ | |
707 score_map[index]= d;\ | |
708 d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\ | |
709 /*printf("score: %d\n", d);*/\ | |
710 if(d < minima[minima_count-1].height){\ | |
711 int j=0;\ | |
712 \ | |
713 while(d >= minima[j].height) j++;\ | |
714 \ | |
715 memmove(&minima [j+1], &minima [j], (minima_count - j - 1)*sizeof(Minima));\ | |
716 \ | |
717 minima[j].checked= 0;\ | |
718 minima[j].height= d;\ | |
719 minima[j].x= ax;\ | |
720 minima[j].y= ay;\ | |
721 \ | |
722 i=-1;\ | |
723 continue;\ | |
724 }\ | |
725 }\ | |
726 } | |
727 | |
728 #define MAX_SAB_SIZE 16 | |
729 static inline int RENAME(sab_diamond_search)(MpegEncContext * s, int *best, int dmin, | |
1708 | 730 uint8_t *src_data[3], |
731 uint8_t *ref_data[3], int stride, int uvstride, | |
948 | 732 int const pred_x, int const pred_y, int const penalty_factor, |
1708 | 733 int const shift, |
734 uint32_t *map, int map_generation, int size, int h, uint8_t * const mv_penalty | |
948 | 735 ) |
736 { | |
737 me_cmp_func cmp, chroma_cmp; | |
738 Minima minima[MAX_SAB_SIZE]; | |
954 | 739 const int minima_count= ABS(s->me.dia_size); |
948 | 740 int i, j; |
1708 | 741 LOAD_COMMON |
948 | 742 |
743 cmp= s->dsp.me_cmp[size]; | |
744 chroma_cmp= s->dsp.me_cmp[size+1]; | |
745 | |
746 for(j=i=0; i<ME_MAP_SIZE; i++){ | |
747 uint32_t key= map[i]; | |
748 | |
749 key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1)); | |
750 | |
751 if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue; | |
752 | |
753 assert(j<MAX_SAB_SIZE); //max j = number of predictors | |
754 | |
755 minima[j].height= score_map[i]; | |
756 minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS; | |
757 minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1); | |
758 minima[j].x-= (1<<(ME_MAP_MV_BITS-1)); | |
759 minima[j].y-= (1<<(ME_MAP_MV_BITS-1)); | |
760 minima[j].checked=0; | |
761 if(minima[j].x || minima[j].y) | |
762 minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor; | |
763 | |
764 j++; | |
765 } | |
766 | |
767 qsort(minima, j, sizeof(Minima), minima_cmp); | |
768 | |
769 for(; j<minima_count; j++){ | |
770 minima[j].height=256*256*256*64; | |
771 minima[j].checked=0; | |
772 minima[j].x= minima[j].y=0; | |
773 } | |
774 | |
775 for(i=0; i<minima_count; i++){ | |
776 const int x= minima[i].x; | |
777 const int y= minima[i].y; | |
778 int d; | |
779 | |
780 if(minima[i].checked) continue; | |
781 | |
782 if( x >= xmax || x <= xmin | |
783 || y >= ymax || y <= ymin) | |
784 continue; | |
785 | |
786 SAB_CHECK_MV(x-1, y) | |
787 SAB_CHECK_MV(x+1, y) | |
788 SAB_CHECK_MV(x , y-1) | |
789 SAB_CHECK_MV(x , y+1) | |
790 | |
791 minima[i].checked= 1; | |
792 } | |
793 | |
794 best[0]= minima[0].x; | |
795 best[1]= minima[0].y; | |
796 dmin= minima[0].height; | |
797 | |
798 if( best[0] < xmax && best[0] > xmin | |
799 && best[1] < ymax && best[1] > ymin){ | |
800 int d; | |
801 //ensure that the refernece samples for hpel refinement are in the map | |
802 CHECK_MV(best[0]-1, best[1]) | |
803 CHECK_MV(best[0]+1, best[1]) | |
804 CHECK_MV(best[0], best[1]-1) | |
805 CHECK_MV(best[0], best[1]+1) | |
806 } | |
807 return dmin; | |
808 } | |
809 | |
936 | 810 static inline int RENAME(var_diamond_search)(MpegEncContext * s, int *best, int dmin, |
1708 | 811 uint8_t *src_data[3], |
812 uint8_t *ref_data[3], int stride, int uvstride, | |
936 | 813 int const pred_x, int const pred_y, int const penalty_factor, |
1708 | 814 int const shift, |
815 uint32_t *map, int map_generation, int size, int h, uint8_t * const mv_penalty | |
936 | 816 ) |
817 { | |
818 me_cmp_func cmp, chroma_cmp; | |
948 | 819 int dia_size; |
1708 | 820 LOAD_COMMON |
936 | 821 |
822 cmp= s->dsp.me_cmp[size]; | |
823 chroma_cmp= s->dsp.me_cmp[size+1]; | |
824 | |
954 | 825 for(dia_size=1; dia_size<=s->me.dia_size; dia_size++){ |
936 | 826 int dir, start, end; |
827 const int x= best[0]; | |
828 const int y= best[1]; | |
829 | |
830 start= FFMAX(0, y + dia_size - ymax); | |
948 | 831 end = FFMIN(dia_size, xmax - x + 1); |
936 | 832 for(dir= start; dir<end; dir++){ |
833 int d; | |
834 | |
835 //check(x + dir,y + dia_size - dir,0, a0) | |
836 CHECK_MV(x + dir , y + dia_size - dir); | |
837 } | |
838 | |
839 start= FFMAX(0, x + dia_size - xmax); | |
948 | 840 end = FFMIN(dia_size, y - ymin + 1); |
936 | 841 for(dir= start; dir<end; dir++){ |
842 int d; | |
843 | |
844 //check(x + dia_size - dir, y - dir,0, a1) | |
845 CHECK_MV(x + dia_size - dir, y - dir ); | |
846 } | |
847 | |
848 start= FFMAX(0, -y + dia_size + ymin ); | |
948 | 849 end = FFMIN(dia_size, x - xmin + 1); |
936 | 850 for(dir= start; dir<end; dir++){ |
851 int d; | |
852 | |
853 //check(x - dir,y - dia_size + dir,0, a2) | |
854 CHECK_MV(x - dir , y - dia_size + dir); | |
855 } | |
856 | |
857 start= FFMAX(0, -x + dia_size + xmin ); | |
948 | 858 end = FFMIN(dia_size, ymax - y + 1); |
936 | 859 for(dir= start; dir<end; dir++){ |
860 int d; | |
861 | |
862 //check(x - dia_size + dir, y + dir,0, a3) | |
863 CHECK_MV(x - dia_size + dir, y + dir ); | |
864 } | |
865 | |
866 if(x!=best[0] || y!=best[1]) | |
867 dia_size=0; | |
948 | 868 #if 0 |
869 { | |
870 int dx, dy, i; | |
871 static int stats[8*8]; | |
872 dx= ABS(x-best[0]); | |
873 dy= ABS(y-best[1]); | |
874 stats[dy*8 + dx] ++; | |
875 if(256*256*256*64 % (stats[0]+1)==0){ | |
876 for(i=0; i<64; i++){ | |
877 if((i&7)==0) printf("\n"); | |
878 printf("%6d ", stats[i]); | |
879 } | |
880 printf("\n"); | |
881 } | |
882 } | |
883 #endif | |
936 | 884 } |
885 return dmin; | |
886 } | |
887 | |
1708 | 888 static int RENAME(epzs_motion_search)(MpegEncContext * s, |
936 | 889 int *mx_ptr, int *my_ptr, |
1708 | 890 int P[10][2], int pred_x, int pred_y, uint8_t *src_data[3], |
891 uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], | |
1162 | 892 int ref_mv_scale, uint8_t * const mv_penalty) |
936 | 893 { |
894 int best[2]={0, 0}; | |
895 int d, dmin; | |
896 const int shift= 1+s->quarter_sample; | |
897 uint32_t *map= s->me.map; | |
898 int map_generation; | |
899 const int penalty_factor= s->me.penalty_factor; | |
900 const int size=0; | |
1708 | 901 const int h=16; |
902 const int ref_mv_stride= s->mb_stride; //pass as arg FIXME | |
903 const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME | |
936 | 904 me_cmp_func cmp, chroma_cmp; |
1708 | 905 LOAD_COMMON |
936 | 906 |
907 cmp= s->dsp.me_cmp[size]; | |
908 chroma_cmp= s->dsp.me_cmp[size+1]; | |
909 | |
910 map_generation= update_map_generation(s); | |
911 | |
912 CMP(dmin, 0, 0, size); | |
913 map[0]= map_generation; | |
914 score_map[0]= dmin; | |
915 | |
916 /* first line */ | |
1799 | 917 if (s->first_slice_line) { |
936 | 918 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
948 | 919 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
920 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
936 | 921 }else{ |
922 if(dmin<256 && ( P_LEFT[0] |P_LEFT[1] | |
923 |P_TOP[0] |P_TOP[1] | |
948 | 924 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){ |
936 | 925 *mx_ptr= 0; |
926 *my_ptr= 0; | |
927 s->me.skip=1; | |
928 return dmin; | |
929 } | |
930 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) | |
931 if(dmin>256*2){ | |
948 | 932 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
933 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
936 | 934 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift) |
935 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift) | |
936 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) | |
937 } | |
938 } | |
939 if(dmin>256*4){ | |
952 | 940 if(s->me.pre_pass){ |
941 CHECK_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, | |
942 (last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16) | |
1799 | 943 if(!s->first_slice_line) |
944 CHECK_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, | |
945 (last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) | |
952 | 946 }else{ |
947 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, | |
948 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) | |
1800
e039d79185c2
simplify MV availability check / dont use below last row MV which is always zero
michael
parents:
1799
diff
changeset
|
949 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
1799 | 950 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
951 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) | |
952 | 952 } |
936 | 953 } |
948 | 954 |
955 if(s->avctx->last_predictor_count){ | |
956 const int count= s->avctx->last_predictor_count; | |
957 const int xstart= FFMAX(0, s->mb_x - count); | |
958 const int ystart= FFMAX(0, s->mb_y - count); | |
959 const int xend= FFMIN(s->mb_width , s->mb_x + count + 1); | |
960 const int yend= FFMIN(s->mb_height, s->mb_y + count + 1); | |
961 int mb_y; | |
936 | 962 |
948 | 963 for(mb_y=ystart; mb_y<yend; mb_y++){ |
964 int mb_x; | |
965 for(mb_x=xstart; mb_x<xend; mb_x++){ | |
966 const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride; | |
967 int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16; | |
968 int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16; | |
969 | |
970 if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue; | |
971 CHECK_MV(mx,my) | |
936 | 972 } |
973 } | |
974 } | |
948 | 975 |
936 | 976 //check(best[0],best[1],0, b0) |
954 | 977 if(s->me.dia_size==-1) |
1708 | 978 dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
979 pred_x, pred_y, penalty_factor, | |
980 shift, map, map_generation, size, h, mv_penalty); | |
954 | 981 else if(s->me.dia_size<-1) |
1708 | 982 dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
983 pred_x, pred_y, penalty_factor, | |
984 shift, map, map_generation, size, h, mv_penalty); | |
954 | 985 else if(s->me.dia_size<2) |
1708 | 986 dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
987 pred_x, pred_y, penalty_factor, | |
988 shift, map, map_generation, size, h, mv_penalty); | |
936 | 989 else |
1708 | 990 dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
991 pred_x, pred_y, penalty_factor, | |
992 shift, map, map_generation, size, h, mv_penalty); | |
936 | 993 |
994 //check(best[0],best[1],0, b1) | |
995 *mx_ptr= best[0]; | |
996 *my_ptr= best[1]; | |
997 | |
998 // printf("%d %d %d \n", best[0], best[1], dmin); | |
999 return dmin; | |
1000 } | |
1001 | |
1002 #ifndef CMP_DIRECT /* no 4mv search needed in direct mode */ | |
1708 | 1003 static int RENAME(epzs_motion_search4)(MpegEncContext * s, |
936 | 1004 int *mx_ptr, int *my_ptr, |
1005 int P[10][2], int pred_x, int pred_y, | |
1708 | 1006 uint8_t *src_data[3], |
1007 uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], | |
1162 | 1008 int ref_mv_scale, uint8_t * const mv_penalty) |
936 | 1009 { |
1010 int best[2]={0, 0}; | |
1011 int d, dmin; | |
1012 const int shift= 1+s->quarter_sample; | |
1013 uint32_t *map= s->me.map; | |
1014 int map_generation; | |
1015 const int penalty_factor= s->me.penalty_factor; | |
1016 const int size=1; | |
1708 | 1017 const int h=8; |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1018 const int ref_mv_stride= s->mb_stride; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1019 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; |
936 | 1020 me_cmp_func cmp, chroma_cmp; |
1708 | 1021 LOAD_COMMON |
936 | 1022 |
1023 cmp= s->dsp.me_cmp[size]; | |
1024 chroma_cmp= s->dsp.me_cmp[size+1]; | |
1025 | |
1026 map_generation= update_map_generation(s); | |
1027 | |
1028 dmin = 1000000; | |
1029 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax); | |
1030 /* first line */ | |
1799 | 1031 if (s->first_slice_line) { |
936 | 1032 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
948 | 1033 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
1034 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
936 | 1035 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) |
1036 }else{ | |
1037 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) | |
1038 //FIXME try some early stop | |
1039 if(dmin>64*2){ | |
1040 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) | |
1041 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) | |
1042 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) | |
1043 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) | |
948 | 1044 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, |
1045 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
936 | 1046 } |
1047 } | |
1048 if(dmin>64*4){ | |
948 | 1049 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, |
1050 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) | |
1800
e039d79185c2
simplify MV availability check / dont use below last row MV which is always zero
michael
parents:
1799
diff
changeset
|
1051 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
1799 | 1052 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
1053 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) | |
936 | 1054 } |
1055 | |
954 | 1056 if(s->me.dia_size==-1) |
1708 | 1057 dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
1058 pred_x, pred_y, penalty_factor, | |
1059 shift, map, map_generation, size, h, mv_penalty); | |
954 | 1060 else if(s->me.dia_size<-1) |
1708 | 1061 dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
1062 pred_x, pred_y, penalty_factor, | |
1063 shift, map, map_generation, size, h, mv_penalty); | |
954 | 1064 else if(s->me.dia_size<2) |
1708 | 1065 dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
1066 pred_x, pred_y, penalty_factor, | |
1067 shift, map, map_generation, size, h, mv_penalty); | |
936 | 1068 else |
1708 | 1069 dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, |
1070 pred_x, pred_y, penalty_factor, | |
1071 shift, map, map_generation, size, h, mv_penalty); | |
1072 | |
1073 | |
1074 *mx_ptr= best[0]; | |
1075 *my_ptr= best[1]; | |
1076 | |
1077 // printf("%d %d %d \n", best[0], best[1], dmin); | |
1078 return dmin; | |
1079 } | |
1080 | |
1081 //try to merge with above FIXME (needs PSNR test) | |
1082 static int RENAME(epzs_motion_search2)(MpegEncContext * s, | |
1083 int *mx_ptr, int *my_ptr, | |
1084 int P[10][2], int pred_x, int pred_y, | |
1085 uint8_t *src_data[3], | |
1086 uint8_t *ref_data[3], int stride, int uvstride, int16_t (*last_mv)[2], | |
1087 int ref_mv_scale, uint8_t * const mv_penalty) | |
1088 { | |
1089 int best[2]={0, 0}; | |
1090 int d, dmin; | |
1091 const int shift= 1+s->quarter_sample; | |
1092 uint32_t *map= s->me.map; | |
1093 int map_generation; | |
1094 const int penalty_factor= s->me.penalty_factor; | |
1095 const int size=0; //FIXME pass as arg | |
1096 const int h=8; | |
1097 const int ref_mv_stride= s->mb_stride; | |
1098 const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride; | |
1099 me_cmp_func cmp, chroma_cmp; | |
1100 LOAD_COMMON | |
1101 | |
1102 cmp= s->dsp.me_cmp[size]; | |
1103 chroma_cmp= s->dsp.me_cmp[size+1]; | |
1104 | |
1105 map_generation= update_map_generation(s); | |
1106 | |
1107 dmin = 1000000; | |
1108 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax); | |
1109 /* first line */ | |
1799 | 1110 if (s->first_slice_line) { |
1708 | 1111 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) |
1112 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, | |
1113 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
1114 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) | |
1115 }else{ | |
1116 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) | |
1117 //FIXME try some early stop | |
1118 if(dmin>64*2){ | |
1119 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) | |
1120 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) | |
1121 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) | |
1122 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) | |
1123 CHECK_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, | |
1124 (last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16) | |
1125 } | |
1126 } | |
1127 if(dmin>64*4){ | |
1128 CHECK_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, | |
1129 (last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16) | |
1800
e039d79185c2
simplify MV availability check / dont use below last row MV which is always zero
michael
parents:
1799
diff
changeset
|
1130 if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line |
1799 | 1131 CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, |
1132 (last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16) | |
1708 | 1133 } |
1134 | |
1135 if(s->me.dia_size==-1) | |
1136 dmin= RENAME(funny_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, | |
1137 pred_x, pred_y, penalty_factor, | |
1138 shift, map, map_generation, size, h, mv_penalty); | |
1139 else if(s->me.dia_size<-1) | |
1140 dmin= RENAME(sab_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, | |
1141 pred_x, pred_y, penalty_factor, | |
1142 shift, map, map_generation, size, h, mv_penalty); | |
1143 else if(s->me.dia_size<2) | |
1144 dmin= RENAME(small_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, | |
1145 pred_x, pred_y, penalty_factor, | |
1146 shift, map, map_generation, size, h, mv_penalty); | |
1147 else | |
1148 dmin= RENAME(var_diamond_search)(s, best, dmin, src_data, ref_data, stride, uvstride, | |
1149 pred_x, pred_y, penalty_factor, | |
1150 shift, map, map_generation, size, h, mv_penalty); | |
1151 | |
948 | 1152 |
936 | 1153 *mx_ptr= best[0]; |
1154 *my_ptr= best[1]; | |
1155 | |
1156 // printf("%d %d %d \n", best[0], best[1], dmin); | |
1157 return dmin; | |
1158 } | |
1159 #endif /* !CMP_DIRECT */ |