Mercurial > libavcodec.hg
comparison motion_est_template.c @ 936:caa77cd960c0 libavcodec
qpel encoding
4mv+b frames encoding finally fixed
chroma ME
5 comparission functions for ME
b frame encoding speedup
wmv2 codec (unfinished)
user specified diamond size for EPZS
author | michaelni |
---|---|
date | Fri, 27 Dec 2002 23:51:46 +0000 |
parents | |
children | 371bc36a9c5c |
comparison
equal
deleted
inserted
replaced
935:c9bbd35064b6 | 936:caa77cd960c0 |
---|---|
1 /* | |
2 * Motion estimation | |
3 * Copyright (c) 2002 Michael Niedermayer | |
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 */ | |
20 | |
21 //lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...) | |
22 //Note, the last line is there to kill these ugly unused var warnings | |
23 #define LOAD_COMMON(x, y)\ | |
24 uint32_t * const score_map= s->me.score_map;\ | |
25 const int stride= s->linesize;\ | |
26 const int uvstride= s->uvlinesize;\ | |
27 const int time_pp= s->pp_time;\ | |
28 const int time_pb= s->pb_time;\ | |
29 uint8_t * const src_y= s->new_picture.data[0] + ((y) * stride) + (x);\ | |
30 uint8_t * const src_u= s->new_picture.data[1] + (((y)>>1) * uvstride) + ((x)>>1);\ | |
31 uint8_t * const src_v= s->new_picture.data[2] + (((y)>>1) * uvstride) + ((x)>>1);\ | |
32 uint8_t * const ref_y= ref_picture->data[0] + ((y) * stride) + (x);\ | |
33 uint8_t * const ref_u= ref_picture->data[1] + (((y)>>1) * uvstride) + ((x)>>1);\ | |
34 uint8_t * const ref_v= ref_picture->data[2] + (((y)>>1) * uvstride) + ((x)>>1);\ | |
35 uint8_t * const ref2_y= s->next_picture.data[0] + ((y) * stride) + (x);\ | |
36 op_pixels_func (*hpel_put)[4];\ | |
37 op_pixels_func (*hpel_avg)[4]= &s->dsp.avg_pixels_tab[size];\ | |
38 op_pixels_func (*chroma_hpel_put)[4];\ | |
39 qpel_mc_func (*qpel_put)[16];\ | |
40 qpel_mc_func (*qpel_avg)[16]= &s->dsp.avg_qpel_pixels_tab[size];\ | |
41 const __attribute__((unused)) int unu= time_pp + time_pb + (int)src_u + (int)src_v + (int)ref_u + (int)ref_v\ | |
42 + (int)ref2_y + (int)hpel_avg + (int)qpel_avg;\ | |
43 if(s->no_rounding /*FIXME b_type*/){\ | |
44 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];\ | |
45 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];\ | |
46 qpel_put= &s->dsp.put_no_rnd_qpel_pixels_tab[size];\ | |
47 }else{\ | |
48 hpel_put=& s->dsp.put_pixels_tab[size];\ | |
49 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];\ | |
50 qpel_put= &s->dsp.put_qpel_pixels_tab[size];\ | |
51 } | |
52 | |
53 | |
54 #ifdef CMP_HPEL | |
55 | |
56 #define CHECK_HALF_MV(dx, dy, x, y)\ | |
57 {\ | |
58 const int hx= 2*(x)+(dx);\ | |
59 const int hy= 2*(y)+(dy);\ | |
60 CMP_HPEL(d, dx, dy, x, y, size);\ | |
61 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ | |
62 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ | |
63 } | |
64 | |
65 #if 0 | |
66 static int RENAME(hpel_motion_search)(MpegEncContext * s, | |
67 int *mx_ptr, int *my_ptr, int dmin, | |
68 int xmin, int ymin, int xmax, int ymax, | |
69 int pred_x, int pred_y, Picture *ref_picture, | |
70 int n, int size) | |
71 { | |
72 UINT8 *ptr; | |
73 | |
74 const int xx = 16 * s->mb_x + 8*(n&1); | |
75 const int yy = 16 * s->mb_y + 8*(n>>1); | |
76 const int mx = *mx_ptr; | |
77 const int my = *my_ptr; | |
78 | |
79 LOAD_COMMON(xx, yy); | |
80 | |
81 // INIT; | |
82 //FIXME factorize | |
83 me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub; | |
84 | |
85 if(s->no_rounding /*FIXME b_type*/){ | |
86 hpel_put= &s->dsp.put_no_rnd_pixels_tab[size]; | |
87 chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1]; | |
88 }else{ | |
89 hpel_put=& s->dsp.put_pixels_tab[size]; | |
90 chroma_hpel_put= &s->dsp.put_pixels_tab[size+1]; | |
91 } | |
92 cmp= s->dsp.me_cmp[size]; | |
93 chroma_cmp= s->dsp.me_cmp[size+1]; | |
94 cmp_sub= s->dsp.me_sub_cmp[size]; | |
95 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
96 | |
97 if(s->me.skip){ //FIXME somehow move up (benchmark) | |
98 *mx_ptr = 0; | |
99 *my_ptr = 0; | |
100 return dmin; | |
101 } | |
102 | |
103 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
104 CMP_HPEL(dmin, 0, 0, mx, my, size); | |
105 if(mx || my) | |
106 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; | |
107 } | |
108 | |
109 if (mx > xmin && mx < xmax && | |
110 my > ymin && my < ymax) { | |
111 int bx=2*mx, by=2*my; | |
112 int d= dmin; | |
113 | |
114 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
115 CHECK_HALF_MV(0, 1, mx , my-1) | |
116 CHECK_HALF_MV(1, 1, mx , my-1) | |
117 CHECK_HALF_MV(1, 0, mx-1, my ) | |
118 CHECK_HALF_MV(1, 0, mx , my ) | |
119 CHECK_HALF_MV(1, 1, mx-1, my ) | |
120 CHECK_HALF_MV(0, 1, mx , my ) | |
121 CHECK_HALF_MV(1, 1, mx , my ) | |
122 | |
123 assert(bx < xmin*2 || bx > xmax*2 || by < ymin*2 || by > ymax*2); | |
124 | |
125 *mx_ptr = bx; | |
126 *my_ptr = by; | |
127 }else{ | |
128 *mx_ptr =2*mx; | |
129 *my_ptr =2*my; | |
130 } | |
131 | |
132 return dmin; | |
133 } | |
134 | |
135 #else | |
136 static int RENAME(hpel_motion_search)(MpegEncContext * s, | |
137 int *mx_ptr, int *my_ptr, int dmin, | |
138 int xmin, int ymin, int xmax, int ymax, | |
139 int pred_x, int pred_y, Picture *ref_picture, | |
140 int n, int size, uint16_t * const mv_penalty) | |
141 { | |
142 const int xx = 16 * s->mb_x + 8*(n&1); | |
143 const int yy = 16 * s->mb_y + 8*(n>>1); | |
144 const int mx = *mx_ptr; | |
145 const int my = *my_ptr; | |
146 const int penalty_factor= s->me.sub_penalty_factor; | |
147 me_cmp_func cmp_sub, chroma_cmp_sub; | |
148 | |
149 LOAD_COMMON(xx, yy); | |
150 | |
151 //FIXME factorize | |
152 | |
153 cmp_sub= s->dsp.me_sub_cmp[size]; | |
154 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
155 | |
156 if(s->me.skip){ //FIXME move out of hpel? | |
157 *mx_ptr = 0; | |
158 *my_ptr = 0; | |
159 return dmin; | |
160 } | |
161 | |
162 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
163 CMP_HPEL(dmin, 0, 0, mx, my, size); | |
164 if(mx || my) | |
165 dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; | |
166 } | |
167 | |
168 if (mx > xmin && mx < xmax && | |
169 my > ymin && my < ymax) { | |
170 int bx=2*mx, by=2*my; | |
171 int d= dmin; | |
172 const int index= (my<<ME_MAP_SHIFT) + mx; | |
173 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] | |
174 + (mv_penalty[bx - pred_x] + mv_penalty[by-2 - pred_y])*penalty_factor; | |
175 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)] | |
176 + (mv_penalty[bx-2 - pred_x] + mv_penalty[by - pred_y])*penalty_factor; | |
177 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)] | |
178 + (mv_penalty[bx+2 - pred_x] + mv_penalty[by - pred_y])*penalty_factor; | |
179 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)] | |
180 + (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*penalty_factor; | |
181 | |
182 if(t<=b){ | |
183 CHECK_HALF_MV(0, 1, mx ,my-1) | |
184 if(l<=r){ | |
185 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
186 if(t+r<=b+l){ | |
187 CHECK_HALF_MV(1, 1, mx , my-1) | |
188 }else{ | |
189 CHECK_HALF_MV(1, 1, mx-1, my ) | |
190 } | |
191 CHECK_HALF_MV(1, 0, mx-1, my ) | |
192 }else{ | |
193 CHECK_HALF_MV(1, 1, mx , my-1) | |
194 if(t+l<=b+r){ | |
195 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
196 }else{ | |
197 CHECK_HALF_MV(1, 1, mx , my ) | |
198 } | |
199 CHECK_HALF_MV(1, 0, mx , my ) | |
200 } | |
201 }else{ | |
202 if(l<=r){ | |
203 if(t+l<=b+r){ | |
204 CHECK_HALF_MV(1, 1, mx-1, my-1) | |
205 }else{ | |
206 CHECK_HALF_MV(1, 1, mx , my ) | |
207 } | |
208 CHECK_HALF_MV(1, 0, mx-1, my) | |
209 CHECK_HALF_MV(1, 1, mx-1, my) | |
210 }else{ | |
211 if(t+r<=b+l){ | |
212 CHECK_HALF_MV(1, 1, mx , my-1) | |
213 }else{ | |
214 CHECK_HALF_MV(1, 1, mx-1, my) | |
215 } | |
216 CHECK_HALF_MV(1, 0, mx , my) | |
217 CHECK_HALF_MV(1, 1, mx , my) | |
218 } | |
219 CHECK_HALF_MV(0, 1, mx , my) | |
220 } | |
221 assert(bx >= xmin*2 && bx <= xmax*2 && by >= ymin*2 && by <= ymax*2); | |
222 | |
223 *mx_ptr = bx; | |
224 *my_ptr = by; | |
225 }else{ | |
226 *mx_ptr =2*mx; | |
227 *my_ptr =2*my; | |
228 } | |
229 | |
230 return dmin; | |
231 } | |
232 #endif | |
233 | |
234 #endif /* CMP_HPEL */ | |
235 | |
236 #ifdef CMP_QPEL | |
237 | |
238 #define CHECK_QUARTER_MV(dx, dy, x, y)\ | |
239 {\ | |
240 const int hx= 4*(x)+(dx);\ | |
241 const int hy= 4*(y)+(dy);\ | |
242 CMP_QPEL(d, dx, dy, x, y, size);\ | |
243 d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ | |
244 COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ | |
245 } | |
246 | |
247 static int RENAME(qpel_motion_search)(MpegEncContext * s, | |
248 int *mx_ptr, int *my_ptr, int dmin, | |
249 int xmin, int ymin, int xmax, int ymax, | |
250 int pred_x, int pred_y, Picture *ref_picture, | |
251 int n, int size, uint16_t * const mv_penalty) | |
252 { | |
253 const int xx = 16 * s->mb_x + 8*(n&1); | |
254 const int yy = 16 * s->mb_y + 8*(n>>1); | |
255 const int mx = *mx_ptr; | |
256 const int my = *my_ptr; | |
257 const int penalty_factor= s->me.sub_penalty_factor; | |
258 const int map_generation= s->me.map_generation; | |
259 uint32_t *map= s->me.map; | |
260 me_cmp_func cmp, chroma_cmp; | |
261 me_cmp_func cmp_sub, chroma_cmp_sub; | |
262 | |
263 LOAD_COMMON(xx, yy); | |
264 | |
265 cmp= s->dsp.me_cmp[size]; | |
266 chroma_cmp= s->dsp.me_cmp[size+1]; //factorize FIXME | |
267 //FIXME factorize | |
268 | |
269 cmp_sub= s->dsp.me_sub_cmp[size]; | |
270 chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; | |
271 | |
272 if(s->me.skip){ //FIXME somehow move up (benchmark) | |
273 *mx_ptr = 0; | |
274 *my_ptr = 0; | |
275 return dmin; | |
276 } | |
277 | |
278 if(s->avctx->me_cmp != s->avctx->me_sub_cmp){ | |
279 CMP_QPEL(dmin, 0, 0, mx, my, size); | |
280 if(mx || my) | |
281 dmin += (mv_penalty[4*mx - pred_x] + mv_penalty[4*my - pred_y])*penalty_factor; | |
282 } | |
283 | |
284 if (mx > xmin && mx < xmax && | |
285 my > ymin && my < ymax) { | |
286 int bx=4*mx, by=4*my; | |
287 int d= dmin; | |
288 int i, nx, ny; | |
289 const int index= (my<<ME_MAP_SHIFT) + mx; | |
290 const int t= score_map[(index-(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; | |
291 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
292 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
293 const int b= score_map[(index+(1<<ME_MAP_SHIFT) )&(ME_MAP_SIZE-1)]; | |
294 const int c= score_map[(index )&(ME_MAP_SIZE-1)]; | |
295 int best[8]; | |
296 int best_pos[8][2]; | |
297 | |
298 memset(best, 64, sizeof(int)*8); | |
299 #if 1 | |
300 if(s->avctx->dia_size>=2){ | |
301 const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
302 const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
303 const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
304 const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)]; | |
305 | |
306 for(ny= -3; ny <= 3; ny++){ | |
307 for(nx= -3; nx <= 3; nx++){ | |
308 const int t2= nx*nx*(tr + tl - 2*t) + 4*nx*(tr-tl) + 32*t; | |
309 const int c2= nx*nx*( r + l - 2*c) + 4*nx*( r- l) + 32*c; | |
310 const int b2= nx*nx*(br + bl - 2*b) + 4*nx*(br-bl) + 32*b; | |
311 int score= ny*ny*(b2 + t2 - 2*c2) + 4*ny*(b2 - t2) + 32*c2; | |
312 int i; | |
313 | |
314 if((nx&3)==0 && (ny&3)==0) continue; | |
315 | |
316 score += 1024*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; | |
317 | |
318 // if(nx&1) score-=1024*s->me.penalty_factor; | |
319 // if(ny&1) score-=1024*s->me.penalty_factor; | |
320 | |
321 for(i=0; i<8; i++){ | |
322 if(score < best[i]){ | |
323 memmove(&best[i+1], &best[i], sizeof(int)*(7-i)); | |
324 memmove(&best_pos[i+1][0], &best_pos[i][0], sizeof(int)*2*(7-i)); | |
325 best[i]= score; | |
326 best_pos[i][0]= nx + 4*mx; | |
327 best_pos[i][1]= ny + 4*my; | |
328 break; | |
329 } | |
330 } | |
331 } | |
332 } | |
333 }else{ | |
334 int tl; | |
335 const int cx = 4*(r - l); | |
336 const int cx2= r + l - 2*c; | |
337 const int cy = 4*(b - t); | |
338 const int cy2= b + t - 2*c; | |
339 int cxy; | |
340 | |
341 if(map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)] == (my<<ME_MAP_MV_BITS) + mx + map_generation && 0){ //FIXME | |
342 tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)]; | |
343 }else{ | |
344 CMP(tl, mx-1, my-1, size); //FIXME wrong if chroma me is different | |
345 } | |
346 | |
347 cxy= 2*tl + (cx + cy)/4 - (cx2 + cy2) - 2*c; | |
348 | |
349 assert(16*cx2 + 4*cx + 32*c == 32*r); | |
350 assert(16*cx2 - 4*cx + 32*c == 32*l); | |
351 assert(16*cy2 + 4*cy + 32*c == 32*b); | |
352 assert(16*cy2 - 4*cy + 32*c == 32*t); | |
353 assert(16*cxy + 16*cy2 + 16*cx2 - 4*cy - 4*cx + 32*c == 32*tl); | |
354 | |
355 for(ny= -3; ny <= 3; ny++){ | |
356 for(nx= -3; nx <= 3; nx++){ | |
357 int score= ny*nx*cxy + nx*nx*cx2 + ny*ny*cy2 + nx*cx + ny*cy + 32*c; //FIXME factor | |
358 int i; | |
359 | |
360 if((nx&3)==0 && (ny&3)==0) continue; | |
361 | |
362 score += 32*(mv_penalty[4*mx + nx - pred_x] + mv_penalty[4*my + ny - pred_y])*penalty_factor; | |
363 // if(nx&1) score-=32*s->me.penalty_factor; | |
364 // if(ny&1) score-=32*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 } | |
379 for(i=0; i<8; i++){ | |
380 nx= best_pos[i][0]; | |
381 ny= best_pos[i][1]; | |
382 CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2) | |
383 } | |
384 #if 0 | |
385 nx= FFMAX(4*mx - bx, bx - 4*mx); | |
386 ny= FFMAX(4*my - by, by - 4*my); | |
387 | |
388 static int stats[4][4]; | |
389 stats[nx][ny]++; | |
390 if(256*256*256*64 % (stats[0][0]+1) ==0){ | |
391 for(i=0; i<16; i++){ | |
392 if((i&3)==0) printf("\n"); | |
393 printf("%6d ", stats[0][i]); | |
394 } | |
395 printf("\n"); | |
396 } | |
397 #endif | |
398 #else | |
399 | |
400 CHECK_QUARTER_MV(2, 2, mx-1, my-1) | |
401 CHECK_QUARTER_MV(0, 2, mx , my-1) | |
402 CHECK_QUARTER_MV(2, 2, mx , my-1) | |
403 CHECK_QUARTER_MV(2, 0, mx , my ) | |
404 CHECK_QUARTER_MV(2, 2, mx , my ) | |
405 CHECK_QUARTER_MV(0, 2, mx , my ) | |
406 CHECK_QUARTER_MV(2, 2, mx-1, my ) | |
407 CHECK_QUARTER_MV(2, 0, mx-1, my ) | |
408 | |
409 nx= bx; | |
410 ny= by; | |
411 | |
412 for(i=0; i<8; i++){ | |
413 int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1}; | |
414 int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1}; | |
415 CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>2, (ny + oy[i])>>2) | |
416 } | |
417 #endif | |
418 #if 0 | |
419 //outer ring | |
420 CHECK_QUARTER_MV(1, 3, mx-1, my-1) | |
421 CHECK_QUARTER_MV(1, 2, mx-1, my-1) | |
422 CHECK_QUARTER_MV(1, 1, mx-1, my-1) | |
423 CHECK_QUARTER_MV(2, 1, mx-1, my-1) | |
424 CHECK_QUARTER_MV(3, 1, mx-1, my-1) | |
425 CHECK_QUARTER_MV(0, 1, mx , my-1) | |
426 CHECK_QUARTER_MV(1, 1, mx , my-1) | |
427 CHECK_QUARTER_MV(2, 1, mx , my-1) | |
428 CHECK_QUARTER_MV(3, 1, mx , my-1) | |
429 CHECK_QUARTER_MV(3, 2, mx , my-1) | |
430 CHECK_QUARTER_MV(3, 3, mx , my-1) | |
431 CHECK_QUARTER_MV(3, 0, mx , my ) | |
432 CHECK_QUARTER_MV(3, 1, mx , my ) | |
433 CHECK_QUARTER_MV(3, 2, mx , my ) | |
434 CHECK_QUARTER_MV(3, 3, mx , my ) | |
435 CHECK_QUARTER_MV(2, 3, mx , my ) | |
436 CHECK_QUARTER_MV(1, 3, mx , my ) | |
437 CHECK_QUARTER_MV(0, 3, mx , my ) | |
438 CHECK_QUARTER_MV(3, 3, mx-1, my ) | |
439 CHECK_QUARTER_MV(2, 3, mx-1, my ) | |
440 CHECK_QUARTER_MV(1, 3, mx-1, my ) | |
441 CHECK_QUARTER_MV(1, 2, mx-1, my ) | |
442 CHECK_QUARTER_MV(1, 1, mx-1, my ) | |
443 CHECK_QUARTER_MV(1, 0, mx-1, my ) | |
444 #endif | |
445 assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4); | |
446 | |
447 *mx_ptr = bx; | |
448 *my_ptr = by; | |
449 }else{ | |
450 *mx_ptr =4*mx; | |
451 *my_ptr =4*my; | |
452 } | |
453 | |
454 return dmin; | |
455 } | |
456 | |
457 #endif /* CMP_QPEL */ | |
458 | |
459 #define CHECK_MV(x,y)\ | |
460 {\ | |
461 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ | |
462 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ | |
463 if(map[index]!=key){\ | |
464 CMP(d, x, y, size);\ | |
465 map[index]= key;\ | |
466 score_map[index]= d;\ | |
467 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ | |
468 COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\ | |
469 }\ | |
470 } | |
471 | |
472 #define CHECK_MV_DIR(x,y,new_dir)\ | |
473 {\ | |
474 const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\ | |
475 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\ | |
476 if(map[index]!=key){\ | |
477 CMP(d, x, y, size);\ | |
478 map[index]= key;\ | |
479 score_map[index]= d;\ | |
480 d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\ | |
481 if(d<dmin){\ | |
482 best[0]=x;\ | |
483 best[1]=y;\ | |
484 dmin=d;\ | |
485 next_dir= new_dir;\ | |
486 }\ | |
487 }\ | |
488 } | |
489 | |
490 #define check(x,y,S,v)\ | |
491 if( (x)<(xmin<<(S)) ) printf("%d %d %d %d %d xmin" #v, xmin, (x), (y), s->mb_x, s->mb_y);\ | |
492 if( (x)>(xmax<<(S)) ) printf("%d %d %d %d %d xmax" #v, xmax, (x), (y), s->mb_x, s->mb_y);\ | |
493 if( (y)<(ymin<<(S)) ) printf("%d %d %d %d %d ymin" #v, ymin, (x), (y), s->mb_x, s->mb_y);\ | |
494 if( (y)>(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, s->mb_y);\ | |
495 | |
496 | |
497 static inline int RENAME(small_diamond_search)(MpegEncContext * s, int *best, int dmin, | |
498 Picture *ref_picture, | |
499 int const pred_x, int const pred_y, int const penalty_factor, | |
500 int const xmin, int const ymin, int const xmax, int const ymax, int const shift, | |
501 uint32_t *map, int map_generation, int size, uint16_t * const mv_penalty | |
502 ) | |
503 { | |
504 me_cmp_func cmp, chroma_cmp; | |
505 int next_dir=-1; | |
506 LOAD_COMMON(s->mb_x*16, s->mb_y*16); | |
507 | |
508 cmp= s->dsp.me_cmp[size]; | |
509 chroma_cmp= s->dsp.me_cmp[size+1]; | |
510 | |
511 for(;;){ | |
512 int d; | |
513 const int dir= next_dir; | |
514 const int x= best[0]; | |
515 const int y= best[1]; | |
516 next_dir=-1; | |
517 | |
518 //printf("%d", dir); | |
519 if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0) | |
520 if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1) | |
521 if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2) | |
522 if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3) | |
523 | |
524 if(next_dir==-1){ | |
525 return dmin; | |
526 } | |
527 } | |
528 } | |
529 | |
530 static inline int RENAME(var_diamond_search)(MpegEncContext * s, int *best, int dmin, | |
531 Picture *ref_picture, | |
532 int const pred_x, int const pred_y, int const penalty_factor, | |
533 int const xmin, int const ymin, int const xmax, int const ymax, int const shift, | |
534 uint32_t *map, int map_generation, int size, uint16_t * const mv_penalty | |
535 ) | |
536 { | |
537 me_cmp_func cmp, chroma_cmp; | |
538 int dia_size=1; | |
539 LOAD_COMMON(s->mb_x*16, s->mb_y*16); | |
540 | |
541 cmp= s->dsp.me_cmp[size]; | |
542 chroma_cmp= s->dsp.me_cmp[size+1]; | |
543 | |
544 for(dia_size=1; dia_size<=s->avctx->dia_size; dia_size++){ | |
545 int dir, start, end; | |
546 const int x= best[0]; | |
547 const int y= best[1]; | |
548 | |
549 start= FFMAX(0, y + dia_size - ymax); | |
550 end = FFMIN(dia_size, xmax - x); | |
551 for(dir= start; dir<end; dir++){ | |
552 int d; | |
553 | |
554 //check(x + dir,y + dia_size - dir,0, a0) | |
555 CHECK_MV(x + dir , y + dia_size - dir); | |
556 } | |
557 | |
558 start= FFMAX(0, x + dia_size - xmax); | |
559 end = FFMIN(dia_size, y - ymin); | |
560 for(dir= start; dir<end; dir++){ | |
561 int d; | |
562 | |
563 //check(x + dia_size - dir, y - dir,0, a1) | |
564 CHECK_MV(x + dia_size - dir, y - dir ); | |
565 } | |
566 | |
567 start= FFMAX(0, -y + dia_size + ymin ); | |
568 end = FFMIN(dia_size, x - xmin); | |
569 for(dir= start; dir<end; dir++){ | |
570 int d; | |
571 | |
572 //check(x - dir,y - dia_size + dir,0, a2) | |
573 CHECK_MV(x - dir , y - dia_size + dir); | |
574 } | |
575 | |
576 start= FFMAX(0, -x + dia_size + xmin ); | |
577 end = FFMIN(dia_size, ymax - y); | |
578 for(dir= start; dir<end; dir++){ | |
579 int d; | |
580 | |
581 //check(x - dia_size + dir, y + dir,0, a3) | |
582 CHECK_MV(x - dia_size + dir, y + dir ); | |
583 } | |
584 | |
585 if(x!=best[0] || y!=best[1]) | |
586 dia_size=0; | |
587 } | |
588 return dmin; | |
589 } | |
590 | |
591 static int RENAME(epzs_motion_search)(MpegEncContext * s, int block, | |
592 int *mx_ptr, int *my_ptr, | |
593 int P[10][2], int pred_x, int pred_y, | |
594 int xmin, int ymin, int xmax, int ymax, Picture *ref_picture, uint16_t * const mv_penalty) | |
595 { | |
596 int best[2]={0, 0}; | |
597 int d, dmin; | |
598 const int shift= 1+s->quarter_sample; | |
599 uint32_t *map= s->me.map; | |
600 int map_generation; | |
601 const int penalty_factor= s->me.penalty_factor; | |
602 const int size=0; | |
603 me_cmp_func cmp, chroma_cmp; | |
604 LOAD_COMMON(s->mb_x*16, s->mb_y*16); | |
605 | |
606 cmp= s->dsp.me_cmp[size]; | |
607 chroma_cmp= s->dsp.me_cmp[size+1]; | |
608 | |
609 map_generation= update_map_generation(s); | |
610 | |
611 CMP(dmin, 0, 0, size); | |
612 map[0]= map_generation; | |
613 score_map[0]= dmin; | |
614 | |
615 /* first line */ | |
616 if ((s->mb_y == 0 || s->first_slice_line)) { | |
617 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) | |
618 CHECK_MV(P_LAST[0]>>shift, P_LAST[1]>>shift) | |
619 }else{ | |
620 if(dmin<256 && ( P_LEFT[0] |P_LEFT[1] | |
621 |P_TOP[0] |P_TOP[1] | |
622 |P_TOPRIGHT[0]|P_TOPRIGHT[1])==0 && s->avctx->dia_size==0){ | |
623 *mx_ptr= 0; | |
624 *my_ptr= 0; | |
625 s->me.skip=1; | |
626 return dmin; | |
627 } | |
628 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) | |
629 if(dmin>256*2){ | |
630 CHECK_MV(P_LAST[0] >>shift, P_LAST[1] >>shift) | |
631 CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift) | |
632 CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift) | |
633 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) | |
634 } | |
635 } | |
636 if(dmin>256*4){ | |
637 CHECK_MV(P_LAST_RIGHT[0] >>shift, P_LAST_RIGHT[1] >>shift) | |
638 CHECK_MV(P_LAST_BOTTOM[0]>>shift, P_LAST_BOTTOM[1]>>shift) | |
639 } | |
640 #if 0 //doest only slow things down | |
641 if(dmin>512*3){ | |
642 int step; | |
643 dmin= score_map[0]; | |
644 best[0]= best[1]=0; | |
645 for(step=128; step>0; step>>=1){ | |
646 const int step2= step; | |
647 int y; | |
648 for(y=-step2+best[1]; y<=step2+best[1]; y+=step){ | |
649 int x; | |
650 if(y<ymin || y>ymax) continue; | |
651 | |
652 for(x=-step2+best[0]; x<=step2+best[0]; x+=step){ | |
653 if(x<xmin || x>xmax) continue; | |
654 if(x==best[0] && y==best[1]) continue; | |
655 CHECK_MV(x,y) | |
656 } | |
657 } | |
658 } | |
659 } | |
660 #endif | |
661 //check(best[0],best[1],0, b0) | |
662 if(s->avctx->dia_size<2) | |
663 dmin= RENAME(small_diamond_search)(s, best, dmin, ref_picture, | |
664 pred_x, pred_y, penalty_factor, xmin, ymin, xmax, ymax, | |
665 shift, map, map_generation, size, mv_penalty); | |
666 else | |
667 dmin= RENAME(var_diamond_search)(s, best, dmin, ref_picture, | |
668 pred_x, pred_y, penalty_factor, xmin, ymin, xmax, ymax, | |
669 shift, map, map_generation, size, mv_penalty); | |
670 | |
671 //check(best[0],best[1],0, b1) | |
672 *mx_ptr= best[0]; | |
673 *my_ptr= best[1]; | |
674 | |
675 // printf("%d %d %d \n", best[0], best[1], dmin); | |
676 return dmin; | |
677 } | |
678 | |
679 #ifndef CMP_DIRECT /* no 4mv search needed in direct mode */ | |
680 static int RENAME(epzs_motion_search4)(MpegEncContext * s, int block, | |
681 int *mx_ptr, int *my_ptr, | |
682 int P[10][2], int pred_x, int pred_y, | |
683 int xmin, int ymin, int xmax, int ymax, Picture *ref_picture, uint16_t * const mv_penalty) | |
684 { | |
685 int best[2]={0, 0}; | |
686 int d, dmin; | |
687 const int shift= 1+s->quarter_sample; | |
688 uint32_t *map= s->me.map; | |
689 int map_generation; | |
690 const int penalty_factor= s->me.penalty_factor; | |
691 const int size=1; | |
692 me_cmp_func cmp, chroma_cmp; | |
693 LOAD_COMMON((s->mb_x*2 + (block&1))*8, (s->mb_y*2 + (block>>1))*8); | |
694 | |
695 cmp= s->dsp.me_cmp[size]; | |
696 chroma_cmp= s->dsp.me_cmp[size+1]; | |
697 | |
698 map_generation= update_map_generation(s); | |
699 | |
700 dmin = 1000000; | |
701 //printf("%d %d %d %d //",xmin, ymin, xmax, ymax); | |
702 /* first line */ | |
703 if ((s->mb_y == 0 || s->first_slice_line) && block<2) { | |
704 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) | |
705 CHECK_MV(P_LAST[0]>>shift, P_LAST[1]>>shift) | |
706 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) | |
707 }else{ | |
708 CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift) | |
709 //FIXME try some early stop | |
710 if(dmin>64*2){ | |
711 CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift) | |
712 CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift) | |
713 CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift) | |
714 CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift) | |
715 CHECK_MV(P_LAST[0]>>shift, P_LAST[1]>>shift) | |
716 } | |
717 } | |
718 if(dmin>64*4){ | |
719 CHECK_MV(P_LAST_RIGHT[0]>>shift, P_LAST_RIGHT[1]>>shift) | |
720 CHECK_MV(P_LAST_BOTTOM[0]>>shift, P_LAST_BOTTOM[1]>>shift) | |
721 } | |
722 | |
723 if(s->avctx->dia_size<2) | |
724 dmin= RENAME(small_diamond_search)(s, best, dmin, ref_picture, | |
725 pred_x, pred_y, penalty_factor, xmin, ymin, xmax, ymax, | |
726 shift, map, map_generation, size, mv_penalty); | |
727 else | |
728 dmin= RENAME(var_diamond_search)(s, best, dmin, ref_picture, | |
729 pred_x, pred_y, penalty_factor, xmin, ymin, xmax, ymax, | |
730 shift, map, map_generation, size, mv_penalty); | |
731 *mx_ptr= best[0]; | |
732 *my_ptr= best[1]; | |
733 | |
734 // printf("%d %d %d \n", best[0], best[1], dmin); | |
735 return dmin; | |
736 } | |
737 #endif /* !CMP_DIRECT */ |