?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ac3dec: simplify zero-bit mantissa dithering by calculating it
[ffmpeg-lucabe.git] / libavcodec / motion_est.c
blob047fbd669797d111101d5a93b104a2b0cd6d4a68
1 /*
2 * Motion estimation
3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
5 *
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
7 *
8 * This file is part of FFmpeg.
9 *
10 * FFmpeg is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * FFmpeg is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with FFmpeg; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 */
24
25 /**
26 * @file libavcodec/motion_est.c
27 * Motion estimation.
28 */
29
30 #include <stdlib.h>
31 #include <stdio.h>
32 #include <limits.h>
33 #include "avcodec.h"
34 #include "dsputil.h"
35 #include "mathops.h"
36 #include "mpegvideo.h"
37
38 #undef NDEBUG
39 #include <assert.h>
40
41 #define SQ(a) ((a)*(a))
42
43 #define P_LEFT P[1]
44 #define P_TOP P[2]
45 #define P_TOPRIGHT P[3]
46 #define P_MEDIAN P[4]
47 #define P_MV1 P[9]
48
49 static inline int sad_hpel_motion_search(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
51 int src_index, int ref_index,
52 int size, int h);
53
54 static inline int update_map_generation(MotionEstContext *c)
55 {
56 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57 if(c->map_generation==0){
58 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
60 }
61 return c->map_generation;
62 }
63
64 /* shape adaptive search stuff */
65 typedef struct Minima{
66 int height;
67 int x, y;
68 int checked;
69 }Minima;
70
71 static int minima_cmp(const void *a, const void *b){
72 const Minima *da = (const Minima *) a;
73 const Minima *db = (const Minima *) b;
74
75 return da->height - db->height;
76 }
77
78 #define FLAG_QPEL 1 //must be 1
79 #define FLAG_CHROMA 2
80 #define FLAG_DIRECT 4
81
82 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
83 const int offset[3]= {
84 y*c-> stride + x,
85 ((y*c->uvstride + x)>>1),
86 ((y*c->uvstride + x)>>1),
87 };
88 int i;
89 for(i=0; i<3; i++){
90 c->src[0][i]= src [i] + offset[i];
91 c->ref[0][i]= ref [i] + offset[i];
92 }
93 if(ref_index){
94 for(i=0; i<3; i++){
95 c->ref[ref_index][i]= ref2[i] + offset[i];
96 }
97 }
98 }
99
100 static int get_flags(MotionEstContext *c, int direct, int chroma){
101 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102 + (direct ? FLAG_DIRECT : 0)
103 + (chroma ? FLAG_CHROMA : 0);
104 }
105
106 /*! \brief compares a block (either a full macroblock or a partition thereof)
107 against a proposed motion-compensated prediction of that block
108 */
109 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
110 const int size, const int h, int ref_index, int src_index,
111 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
112 MotionEstContext * const c= &s->me;
113 const int stride= c->stride;
114 const int uvstride= c->uvstride;
115 const int qpel= flags&FLAG_QPEL;
116 const int chroma= flags&FLAG_CHROMA;
117 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
118 const int hx= subx + (x<<(1+qpel));
119 const int hy= suby + (y<<(1+qpel));
120 uint8_t * const * const ref= c->ref[ref_index];
121 uint8_t * const * const src= c->src[src_index];
122 int d;
123 //FIXME check chroma 4mv, (no crashes ...)
124 if(flags&FLAG_DIRECT){
125 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
126 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
127 const int time_pp= s->pp_time;
128 const int time_pb= s->pb_time;
129 const int mask= 2*qpel+1;
130 if(s->mv_type==MV_TYPE_8X8){
131 int i;
132 for(i=0; i<4; i++){
133 int fx = c->direct_basis_mv[i][0] + hx;
134 int fy = c->direct_basis_mv[i][1] + hy;
135 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
136 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
137 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
138 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
139
140 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
141 if(qpel){
142 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
143 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
144 }else{
145 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
146 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
147 }
148 }
149 }else{
150 int fx = c->direct_basis_mv[0][0] + hx;
151 int fy = c->direct_basis_mv[0][1] + hy;
152 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
153 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
154 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
155 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
156
157 if(qpel){
158 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
159 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
160 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
161 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
162 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
163 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
164 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
165 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
166 }else{
167 assert((fx>>1) + 16*s->mb_x >= -16);
168 assert((fy>>1) + 16*s->mb_y >= -16);
169 assert((fx>>1) + 16*s->mb_x <= s->width);
170 assert((fy>>1) + 16*s->mb_y <= s->height);
171 assert((bx>>1) + 16*s->mb_x >= -16);
172 assert((by>>1) + 16*s->mb_y >= -16);
173 assert((bx>>1) + 16*s->mb_x <= s->width);
174 assert((by>>1) + 16*s->mb_y <= s->height);
175
176 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
177 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
178 }
179 }
180 d = cmp_func(s, c->temp, src[0], stride, 16);
181 }else
182 d= 256*256*256*32;
183 }else{
184 int uvdxy; /* no, it might not be used uninitialized */
185 if(dxy){
186 if(qpel){
187 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
188 if(chroma){
189 int cx= hx/2;
190 int cy= hy/2;
191 cx= (cx>>1)|(cx&1);
192 cy= (cy>>1)|(cy&1);
193 uvdxy= (cx&1) + 2*(cy&1);
194 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
195 }
196 }else{
197 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
198 if(chroma)
199 uvdxy= dxy | (x&1) | (2*(y&1));
200 }
201 d = cmp_func(s, c->temp, src[0], stride, h);
202 }else{
203 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
204 if(chroma)
205 uvdxy= (x&1) + 2*(y&1);
206 }
207 if(chroma){
208 uint8_t * const uvtemp= c->temp + 16*stride;
209 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
211 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
212 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
213 }
214 }
215 #if 0
216 if(full_pel){
217 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
218 score_map[index]= d;
219 }
220
221 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
222 #endif
223 return d;
224 }
225
226 #include "motion_est_template.c"
227
228 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
229 return 0;
230 }
231
232 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
233 }
234
235 int ff_init_me(MpegEncContext *s){
236 MotionEstContext * const c= &s->me;
237 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
238 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
239
240 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
241 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
242 return -1;
243 }
244 //special case of snow is needed because snow uses its own iterative ME code
245 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
246 av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
247 return -1;
248 }
249
250 c->avctx= s->avctx;
251
252 if(cache_size < 2*dia_size && !c->stride){
253 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
254 }
255
256 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
257 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
258 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
259 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
260
261 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
262 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
263 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
264
265 /*FIXME s->no_rounding b_type*/
266 if(s->flags&CODEC_FLAG_QPEL){
267 c->sub_motion_search= qpel_motion_search;
268 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
269 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
270 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
271 }else{
272 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
273 c->sub_motion_search= hpel_motion_search;
274 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
275 && c->avctx-> me_cmp == FF_CMP_SAD
276 && c->avctx-> mb_cmp == FF_CMP_SAD)
277 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
278 else
279 c->sub_motion_search= hpel_motion_search;
280 }
281 c->hpel_avg= s->dsp.avg_pixels_tab;
282 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
283 else c->hpel_put= s->dsp.put_pixels_tab;
284
285 if(s->linesize){
286 c->stride = s->linesize;
287 c->uvstride= s->uvlinesize;
288 }else{
289 c->stride = 16*s->mb_width + 32;
290 c->uvstride= 8*s->mb_width + 16;
291 }
292
293 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
294 * not have yet, and even if we had, the motion estimation code
295 * does not expect it. */
296 if(s->codec_id != CODEC_ID_SNOW){
297 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
298 s->dsp.me_cmp[2]= zero_cmp;
299 }
300 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
301 s->dsp.me_sub_cmp[2]= zero_cmp;
302 }
303 c->hpel_put[2][0]= c->hpel_put[2][1]=
304 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
305 }
306
307 if(s->codec_id == CODEC_ID_H261){
308 c->sub_motion_search= no_sub_motion_search;
309 }
310
311 return 0;
312 }
313
314 #if 0
315 static int pix_dev(uint8_t * pix, int line_size, int mean)
316 {
317 int s, i, j;
318
319 s = 0;
320 for (i = 0; i < 16; i++) {
321 for (j = 0; j < 16; j += 8) {
322 s += FFABS(pix[0]-mean);
323 s += FFABS(pix[1]-mean);
324 s += FFABS(pix[2]-mean);
325 s += FFABS(pix[3]-mean);
326 s += FFABS(pix[4]-mean);
327 s += FFABS(pix[5]-mean);
328 s += FFABS(pix[6]-mean);
329 s += FFABS(pix[7]-mean);
330 pix += 8;
331 }
332 pix += line_size - 16;
333 }
334 return s;
335 }
336 #endif
337
338 static inline void no_motion_search(MpegEncContext * s,
339 int *mx_ptr, int *my_ptr)
340 {
341 *mx_ptr = 16 * s->mb_x;
342 *my_ptr = 16 * s->mb_y;
343 }
344
345 #define Z_THRESHOLD 256
346
347 #define CHECK_SAD_HALF_MV(suffix, x, y) \
348 {\
349 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
350 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
351 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
352 }
353
354 static inline int sad_hpel_motion_search(MpegEncContext * s,
355 int *mx_ptr, int *my_ptr, int dmin,
356 int src_index, int ref_index,
357 int size, int h)
358 {
359 MotionEstContext * const c= &s->me;
360 const int penalty_factor= c->sub_penalty_factor;
361 int mx, my, dminh;
362 uint8_t *pix, *ptr;
363 int stride= c->stride;
364 const int flags= c->sub_flags;
365 LOAD_COMMON
366
367 assert(flags == 0);
368
369 if(c->skip){
370 // printf("S");
371 *mx_ptr = 0;
372 *my_ptr = 0;
373 return dmin;
374 }
375 // printf("N");
376
377 pix = c->src[src_index][0];
378
379 mx = *mx_ptr;
380 my = *my_ptr;
381 ptr = c->ref[ref_index][0] + (my * stride) + mx;
382
383 dminh = dmin;
384
385 if (mx > xmin && mx < xmax &&
386 my > ymin && my < ymax) {
387 int dx=0, dy=0;
388 int d, pen_x, pen_y;
389 const int index= (my<<ME_MAP_SHIFT) + mx;
390 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
391 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
392 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
393 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
394 mx<<=1;
395 my<<=1;
396
397
398 pen_x= pred_x + mx;
399 pen_y= pred_y + my;
400
401 ptr-= stride;
402 if(t<=b){
403 CHECK_SAD_HALF_MV(y2 , 0, -1)
404 if(l<=r){
405 CHECK_SAD_HALF_MV(xy2, -1, -1)
406 if(t+r<=b+l){
407 CHECK_SAD_HALF_MV(xy2, +1, -1)
408 ptr+= stride;
409 }else{
410 ptr+= stride;
411 CHECK_SAD_HALF_MV(xy2, -1, +1)
412 }
413 CHECK_SAD_HALF_MV(x2 , -1, 0)
414 }else{
415 CHECK_SAD_HALF_MV(xy2, +1, -1)
416 if(t+l<=b+r){
417 CHECK_SAD_HALF_MV(xy2, -1, -1)
418 ptr+= stride;
419 }else{
420 ptr+= stride;
421 CHECK_SAD_HALF_MV(xy2, +1, +1)
422 }
423 CHECK_SAD_HALF_MV(x2 , +1, 0)
424 }
425 }else{
426 if(l<=r){
427 if(t+l<=b+r){
428 CHECK_SAD_HALF_MV(xy2, -1, -1)
429 ptr+= stride;
430 }else{
431 ptr+= stride;
432 CHECK_SAD_HALF_MV(xy2, +1, +1)
433 }
434 CHECK_SAD_HALF_MV(x2 , -1, 0)
435 CHECK_SAD_HALF_MV(xy2, -1, +1)
436 }else{
437 if(t+r<=b+l){
438 CHECK_SAD_HALF_MV(xy2, +1, -1)
439 ptr+= stride;
440 }else{
441 ptr+= stride;
442 CHECK_SAD_HALF_MV(xy2, -1, +1)
443 }
444 CHECK_SAD_HALF_MV(x2 , +1, 0)
445 CHECK_SAD_HALF_MV(xy2, +1, +1)
446 }
447 CHECK_SAD_HALF_MV(y2 , 0, +1)
448 }
449 mx+=dx;
450 my+=dy;
451
452 }else{
453 mx<<=1;
454 my<<=1;
455 }
456
457 *mx_ptr = mx;
458 *my_ptr = my;
459 return dminh;
460 }
461
462 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
463 {
464 const int xy= s->mb_x + s->mb_y*s->mb_stride;
465
466 s->p_mv_table[xy][0] = mx;
467 s->p_mv_table[xy][1] = my;
468
469 /* has already been set to the 4 MV if 4MV is done */
470 if(mv4){
471 int mot_xy= s->block_index[0];
472
473 s->current_picture.motion_val[0][mot_xy ][0]= mx;
474 s->current_picture.motion_val[0][mot_xy ][1]= my;
475 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
476 s->current_picture.motion_val[0][mot_xy+1][1]= my;
477
478 mot_xy += s->b8_stride;
479 s->current_picture.motion_val[0][mot_xy ][0]= mx;
480 s->current_picture.motion_val[0][mot_xy ][1]= my;
481 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
482 s->current_picture.motion_val[0][mot_xy+1][1]= my;
483 }
484 }
485
486 /**
487 * get fullpel ME search limits.
488 */
489 static inline void get_limits(MpegEncContext *s, int x, int y)
490 {
491 MotionEstContext * const c= &s->me;
492 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
493 /*
494 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
495 else c->range= 16;
496 */
497 if (s->unrestricted_mv) {
498 c->xmin = - x - 16;
499 c->ymin = - y - 16;
500 c->xmax = - x + s->mb_width *16;
501 c->ymax = - y + s->mb_height*16;
502 } else if (s->out_format == FMT_H261){
503 // Search range of H261 is different from other codec standards
504 c->xmin = (x > 15) ? - 15 : 0;
505 c->ymin = (y > 15) ? - 15 : 0;
506 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
507 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
508 } else {
509 c->xmin = - x;
510 c->ymin = - y;
511 c->xmax = - x + s->mb_width *16 - 16;
512 c->ymax = - y + s->mb_height*16 - 16;
513 }
514 if(range){
515 c->xmin = FFMAX(c->xmin,-range);
516 c->xmax = FFMIN(c->xmax, range);
517 c->ymin = FFMAX(c->ymin,-range);
518 c->ymax = FFMIN(c->ymax, range);
519 }
520 }
521
522 static inline void init_mv4_ref(MotionEstContext *c){
523 const int stride= c->stride;
524
525 c->ref[1][0] = c->ref[0][0] + 8;
526 c->ref[2][0] = c->ref[0][0] + 8*stride;
527 c->ref[3][0] = c->ref[2][0] + 8;
528 c->src[1][0] = c->src[0][0] + 8;
529 c->src[2][0] = c->src[0][0] + 8*stride;
530 c->src[3][0] = c->src[2][0] + 8;
531 }
532
533 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
534 {
535 MotionEstContext * const c= &s->me;
536 const int size= 1;
537 const int h=8;
538 int block;
539 int P[10][2];
540 int dmin_sum=0, mx4_sum=0, my4_sum=0;
541 int same=1;
542 const int stride= c->stride;
543 uint8_t *mv_penalty= c->current_mv_penalty;
544
545 init_mv4_ref(c);
546
547 for(block=0; block<4; block++){
548 int mx4, my4;
549 int pred_x4, pred_y4;
550 int dmin4;
551 static const int off[4]= {2, 1, 1, -1};
552 const int mot_stride = s->b8_stride;
553 const int mot_xy = s->block_index[block];
554
555 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
556 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
557
558 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
559
560 /* special case for first line */
561 if (s->first_slice_line && block<2) {
562 c->pred_x= pred_x4= P_LEFT[0];
563 c->pred_y= pred_y4= P_LEFT[1];
564 } else {
565 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
566 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
567 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
568 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
569 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
570 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
571 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
572 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
573
574 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
575 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
576
577 c->pred_x= pred_x4 = P_MEDIAN[0];
578 c->pred_y= pred_y4 = P_MEDIAN[1];
579 }
580 P_MV1[0]= mx;
581 P_MV1[1]= my;
582
583 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
584
585 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
586
587 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
588 int dxy;
589 const int offset= ((block&1) + (block>>1)*stride)*8;
590 uint8_t *dest_y = c->scratchpad + offset;
591 if(s->quarter_sample){
592 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
593 dxy = ((my4 & 3) << 2) | (mx4 & 3);
594
595 if(s->no_rounding)
596 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
597 else
598 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
599 }else{
600 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
601 dxy = ((my4 & 1) << 1) | (mx4 & 1);
602
603 if(s->no_rounding)
604 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
605 else
606 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
607 }
608 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
609 }else
610 dmin_sum+= dmin4;
611
612 if(s->quarter_sample){
613 mx4_sum+= mx4/2;
614 my4_sum+= my4/2;
615 }else{
616 mx4_sum+= mx4;
617 my4_sum+= my4;
618 }
619
620 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
621 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
622
623 if(mx4 != mx || my4 != my) same=0;
624 }
625
626 if(same)
627 return INT_MAX;
628
629 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
630 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
631 }
632
633 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
634 int dxy;
635 int mx, my;
636 int offset;
637
638 mx= ff_h263_round_chroma(mx4_sum);
639 my= ff_h263_round_chroma(my4_sum);
640 dxy = ((my & 1) << 1) | (mx & 1);
641
642 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
643
644 if(s->no_rounding){
645 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
646 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
647 }else{
648 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
649 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
650 }
651
652 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
653 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
654 }
655
656 c->pred_x= mx;
657 c->pred_y= my;
658
659 switch(c->avctx->mb_cmp&0xFF){
660 /*case FF_CMP_SSE:
661 return dmin_sum+ 32*s->qscale*s->qscale;*/
662 case FF_CMP_RD:
663 return dmin_sum;
664 default:
665 return dmin_sum+ 11*c->mb_penalty_factor;
666 }
667 }
668
669 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
670 MotionEstContext * const c= &s->me;
671
672 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
673 c->src[1][0] = c->src[0][0] + s->linesize;
674 if(c->flags & FLAG_CHROMA){
675 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
676 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
677 c->src[1][1] = c->src[0][1] + s->uvlinesize;
678 c->src[1][2] = c->src[0][2] + s->uvlinesize;
679 }
680 }
681
682 static int interlaced_search(MpegEncContext *s, int ref_index,
683 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
684 {
685 MotionEstContext * const c= &s->me;
686 const int size=0;
687 const int h=8;
688 int block;
689 int P[10][2];
690 uint8_t * const mv_penalty= c->current_mv_penalty;
691 int same=1;
692 const int stride= 2*s->linesize;
693 int dmin_sum= 0;
694 const int mot_stride= s->mb_stride;
695 const int xy= s->mb_x + s->mb_y*mot_stride;
696
697 c->ymin>>=1;
698 c->ymax>>=1;
699 c->stride<<=1;
700 c->uvstride<<=1;
701 init_interlaced_ref(s, ref_index);
702
703 for(block=0; block<2; block++){
704 int field_select;
705 int best_dmin= INT_MAX;
706 int best_field= -1;
707
708 for(field_select=0; field_select<2; field_select++){
709 int dmin, mx_i, my_i;
710 int16_t (*mv_table)[2]= mv_tables[block][field_select];
711
712 if(user_field_select){
713 assert(field_select==0 || field_select==1);
714 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
715 if(field_select_tables[block][xy] != field_select)
716 continue;
717 }
718
719 P_LEFT[0] = mv_table[xy - 1][0];
720 P_LEFT[1] = mv_table[xy - 1][1];
721 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
722
723 c->pred_x= P_LEFT[0];
724 c->pred_y= P_LEFT[1];
725
726 if(!s->first_slice_line){
727 P_TOP[0] = mv_table[xy - mot_stride][0];
728 P_TOP[1] = mv_table[xy - mot_stride][1];
729 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
730 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
731 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
732 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
733 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
734 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
735
736 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
737 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
738 }
739 P_MV1[0]= mx; //FIXME not correct if block != field_select
740 P_MV1[1]= my / 2;
741
742 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
743
744 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
745
746 mv_table[xy][0]= mx_i;
747 mv_table[xy][1]= my_i;
748
749 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
750 int dxy;
751
752 //FIXME chroma ME
753 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
754 dxy = ((my_i & 1) << 1) | (mx_i & 1);
755
756 if(s->no_rounding){
757 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
758 }else{
759 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
760 }
761 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
762 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
763 }else
764 dmin+= c->mb_penalty_factor; //field_select bits
765
766 dmin += field_select != block; //slightly prefer same field
767
768 if(dmin < best_dmin){
769 best_dmin= dmin;
770 best_field= field_select;
771 }
772 }
773 {
774 int16_t (*mv_table)[2]= mv_tables[block][best_field];
775
776 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
777 if(mv_table[xy][1]&1) same=0;
778 if(mv_table[xy][1]*2 != my) same=0;
779 if(best_field != block) same=0;
780 }
781
782 field_select_tables[block][xy]= best_field;
783 dmin_sum += best_dmin;
784 }
785
786 c->ymin<<=1;
787 c->ymax<<=1;
788 c->stride>>=1;
789 c->uvstride>>=1;
790
791 if(same)
792 return INT_MAX;
793
794 switch(c->avctx->mb_cmp&0xFF){
795 /*case FF_CMP_SSE:
796 return dmin_sum+ 32*s->qscale*s->qscale;*/
797 case FF_CMP_RD:
798 return dmin_sum;
799 default:
800 return dmin_sum+ 11*c->mb_penalty_factor;
801 }
802 }
803
804 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
805 int ymax= s->me.ymax>>interlaced;
806 int ymin= s->me.ymin>>interlaced;
807
808 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
809 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
810 if(mv[1] < ymin) mv[1] = ymin;
811 if(mv[1] > ymax) mv[1] = ymax;
812 }
813
814 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
815 MotionEstContext * const c= &s->me;
816 Picture *p= s->current_picture_ptr;
817 int mb_xy= mb_x + mb_y*s->mb_stride;
818 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
819 int mb_type= s->current_picture.mb_type[mb_xy];
820 int flags= c->flags;
821 int shift= (flags&FLAG_QPEL) + 1;
822 int mask= (1<<shift)-1;
823 int x, y, i;
824 int d=0;
825 me_cmp_func cmpf= s->dsp.sse[0];
826 me_cmp_func chroma_cmpf= s->dsp.sse[1];
827
828 if(p_type && USES_LIST(mb_type, 1)){
829 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
830 return INT_MAX/2;
831 }
832 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
833
834 for(i=0; i<4; i++){
835 int xy= s->block_index[i];
836 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
837 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
838 }
839
840 if(IS_INTERLACED(mb_type)){
841 int xy2= xy + s->b8_stride;
842 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
843 c->stride<<=1;
844 c->uvstride<<=1;
845
846 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
847 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
848 return INT_MAX/2;
849 }
850
851 if(USES_LIST(mb_type, 0)){
852 int field_select0= p->ref_index[0][xy ];
853 int field_select1= p->ref_index[0][xy2];
854 assert(field_select0==0 ||field_select0==1);
855 assert(field_select1==0 ||field_select1==1);
856 init_interlaced_ref(s, 0);
857
858 if(p_type){
859 s->p_field_select_table[0][mb_xy]= field_select0;
860 s->p_field_select_table[1][mb_xy]= field_select1;
861 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
862 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
863 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
864 }else{
865 s->b_field_select_table[0][0][mb_xy]= field_select0;
866 s->b_field_select_table[0][1][mb_xy]= field_select1;
867 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
868 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
869 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
870 }
871
872 x= p->motion_val[0][xy ][0];
873 y= p->motion_val[0][xy ][1];
874 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
875 x= p->motion_val[0][xy2][0];
876 y= p->motion_val[0][xy2][1];
877 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
878 }
879 if(USES_LIST(mb_type, 1)){
880 int field_select0= p->ref_index[1][xy ];
881 int field_select1= p->ref_index[1][xy2];
882 assert(field_select0==0 ||field_select0==1);
883 assert(field_select1==0 ||field_select1==1);
884 init_interlaced_ref(s, 2);
885
886 s->b_field_select_table[1][0][mb_xy]= field_select0;
887 s->b_field_select_table[1][1][mb_xy]= field_select1;
888 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
889 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
890 if(USES_LIST(mb_type, 0)){
891 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
892 }else{
893 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
894 }
895
896 x= p->motion_val[1][xy ][0];
897 y= p->motion_val[1][xy ][1];
898 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
899 x= p->motion_val[1][xy2][0];
900 y= p->motion_val[1][xy2][1];
901 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
902 //FIXME bidir scores
903 }
904 c->stride>>=1;
905 c->uvstride>>=1;
906 }else if(IS_8X8(mb_type)){
907 if(!(s->flags & CODEC_FLAG_4MV)){
908 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
909 return INT_MAX/2;
910 }
911 cmpf= s->dsp.sse[1];
912 chroma_cmpf= s->dsp.sse[1];
913 init_mv4_ref(c);
914 for(i=0; i<4; i++){
915 xy= s->block_index[i];
916 x= p->motion_val[0][xy][0];
917 y= p->motion_val[0][xy][1];
918 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
919 }
920 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
921 }else{
922 if(USES_LIST(mb_type, 0)){
923 if(p_type){
924 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
925 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
926 }else if(USES_LIST(mb_type, 1)){
927 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
928 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
929 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
930 }else{
931 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
932 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
933 }
934 x= p->motion_val[0][xy][0];
935 y= p->motion_val[0][xy][1];
936 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
937 }else if(USES_LIST(mb_type, 1)){
938 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
939 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
940
941 x= p->motion_val[1][xy][0];
942 y= p->motion_val[1][xy][1];
943 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
944 }else
945 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
946 }
947 return d;
948 }
949
950 void ff_estimate_p_frame_motion(MpegEncContext * s,
951 int mb_x, int mb_y)
952 {
953 MotionEstContext * const c= &s->me;
954 uint8_t *pix, *ppix;
955 int sum, mx, my, dmin;
956 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
957 int vard; ///< sum of squared differences with the estimated motion vector
958 int P[10][2];
959 const int shift= 1+s->quarter_sample;
960 int mb_type=0;
961 Picture * const pic= &s->current_picture;
962
963 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
964
965 assert(s->quarter_sample==0 || s->quarter_sample==1);
966 assert(s->linesize == c->stride);
967 assert(s->uvlinesize == c->uvstride);
968
969 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
970 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
971 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
972 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
973
974 get_limits(s, 16*mb_x, 16*mb_y);
975 c->skip=0;
976
977 /* intra / predictive decision */
978 pix = c->src[0][0];
979 sum = s->dsp.pix_sum(pix, s->linesize);
980 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
981
982 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
983 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
984 c->mb_var_sum_temp += (varc+128)>>8;
985
986 if(c->avctx->me_threshold){
987 vard= check_input_motion(s, mb_x, mb_y, 1);
988
989 if((vard+128)>>8 < c->avctx->me_threshold){
990 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
991 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
992 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
993 c->mc_mb_var_sum_temp += (vard+128)>>8;
994 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
995 return;
996 }
997 if((vard+128)>>8 < c->avctx->mb_threshold)
998 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
999 }
1000
1001 switch(s->me_method) {
1002 case ME_ZERO:
1003 default:
1004 no_motion_search(s, &mx, &my);
1005 mx-= mb_x*16;
1006 my-= mb_y*16;
1007 dmin = 0;
1008 break;
1009 case ME_X1:
1010 case ME_EPZS:
1011 {
1012 const int mot_stride = s->b8_stride;
1013 const int mot_xy = s->block_index[0];
1014
1015 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1016 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1017
1018 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1019
1020 if(!s->first_slice_line) {
1021 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1022 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1023 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1024 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1025 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1026 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1027 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1028
1029 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1030 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1031
1032 if(s->out_format == FMT_H263){
1033 c->pred_x = P_MEDIAN[0];
1034 c->pred_y = P_MEDIAN[1];
1035 }else { /* mpeg1 at least */
1036 c->pred_x= P_LEFT[0];
1037 c->pred_y= P_LEFT[1];
1038 }
1039 }else{
1040 c->pred_x= P_LEFT[0];
1041 c->pred_y= P_LEFT[1];
1042 }
1043
1044 }
1045 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1046
1047 break;
1048 }
1049
1050 /* At this point (mx,my) are full-pell and the relative displacement */
1051 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1052
1053 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1054
1055 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1056 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1057 c->mc_mb_var_sum_temp += (vard+128)>>8;
1058
1059 #if 0
1060 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1061 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1062 #endif
1063 if(mb_type){
1064 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1065 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1066 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1067
1068 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1069 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1070 set_p_mv_tables(s, mx, my, 1);
1071 }else{
1072 mx <<=shift;
1073 my <<=shift;
1074 }
1075 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1076 h263_mv4_search(s, mx, my, shift);
1077
1078 set_p_mv_tables(s, mx, my, 0);
1079 }
1080 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1081 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1082 }
1083 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1084 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1085 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1086 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1087
1088 if (vard*2 + 200*256 > varc)
1089 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1090 if (varc*2 + 200*256 > vard || s->qscale > 24){
1091 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1092 mb_type|= CANDIDATE_MB_TYPE_INTER;
1093 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1094 if(s->flags&CODEC_FLAG_MV0)
1095 if(mx || my)
1096 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1097 }else{
1098 mx <<=shift;
1099 my <<=shift;
1100 }
1101 if((s->flags&CODEC_FLAG_4MV)
1102 && !c->skip && varc>50<<8 && vard>10<<8){
1103 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1104 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1105
1106 set_p_mv_tables(s, mx, my, 0);
1107 }else
1108 set_p_mv_tables(s, mx, my, 1);
1109 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1110 && !c->skip){ //FIXME varc/d checks
1111 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1112 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1113 }
1114 }else{
1115 int intra_score, i;
1116 mb_type= CANDIDATE_MB_TYPE_INTER;
1117
1118 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1119 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1120 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1121
1122 if((s->flags&CODEC_FLAG_4MV)
1123 && !c->skip && varc>50<<8 && vard>10<<8){
1124 int dmin4= h263_mv4_search(s, mx, my, shift);
1125 if(dmin4 < dmin){
1126 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1127 dmin=dmin4;
1128 }
1129 }
1130 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1131 && !c->skip){ //FIXME varc/d checks
1132 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1133 if(dmin_i < dmin){
1134 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1135 dmin= dmin_i;
1136 }
1137 }
1138
1139 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1140 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1141
1142 /* get intra luma score */
1143 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1144 intra_score= varc - 500;
1145 }else{
1146 int mean= (sum+128)>>8;
1147 mean*= 0x01010101;
1148
1149 for(i=0; i<16; i++){
1150 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1151 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1152 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1153 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1154 }
1155
1156 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1157 }
1158 #if 0 //FIXME
1159 /* get chroma score */
1160 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1161 for(i=1; i<3; i++){
1162 uint8_t *dest_c;
1163 int mean;
1164
1165 if(s->out_format == FMT_H263){
1166 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1167 }else{
1168 mean= (s->last_dc[i] + 4)>>3;
1169 }
1170 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1171
1172 mean*= 0x01010101;
1173 for(i=0; i<8; i++){
1174 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1175 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1176 }
1177
1178 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1179 }
1180 }
1181 #endif
1182 intra_score += c->mb_penalty_factor*16;
1183
1184 if(intra_score < dmin){
1185 mb_type= CANDIDATE_MB_TYPE_INTRA;
1186 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1187 }else
1188 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1189
1190 {
1191 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1192 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1193 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1194 }
1195 }
1196
1197 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1198 }
1199
1200 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1201 int mb_x, int mb_y)
1202 {
1203 MotionEstContext * const c= &s->me;
1204 int mx, my, dmin;
1205 int P[10][2];
1206 const int shift= 1+s->quarter_sample;
1207 const int xy= mb_x + mb_y*s->mb_stride;
1208 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1209
1210 assert(s->quarter_sample==0 || s->quarter_sample==1);
1211
1212 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1213 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1214
1215 get_limits(s, 16*mb_x, 16*mb_y);
1216 c->skip=0;
1217
1218 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1219 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1220
1221 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1222
1223 /* special case for first line */
1224 if (s->first_slice_line) {
1225 c->pred_x= P_LEFT[0];
1226 c->pred_y= P_LEFT[1];
1227 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1228 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1229 } else {
1230 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1231 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1232 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1233 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1234 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1235 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1236 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1237
1238 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1239 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1240
1241 c->pred_x = P_MEDIAN[0];
1242 c->pred_y = P_MEDIAN[1];
1243 }
1244
1245 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1246
1247 s->p_mv_table[xy][0] = mx<<shift;
1248 s->p_mv_table[xy][1] = my<<shift;
1249
1250 return dmin;
1251 }
1252
1253 static int ff_estimate_motion_b(MpegEncContext * s,
1254 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1255 {
1256 MotionEstContext * const c= &s->me;
1257 int mx, my, dmin;
1258 int P[10][2];
1259 const int shift= 1+s->quarter_sample;
1260 const int mot_stride = s->mb_stride;
1261 const int mot_xy = mb_y*mot_stride + mb_x;
1262 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1263 int mv_scale;
1264
1265 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1266 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1267 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1268 c->current_mv_penalty= mv_penalty;
1269
1270 get_limits(s, 16*mb_x, 16*mb_y);
1271
1272 switch(s->me_method) {
1273 case ME_ZERO:
1274 default:
1275 no_motion_search(s, &mx, &my);
1276 dmin = 0;
1277 mx-= mb_x*16;
1278 my-= mb_y*16;
1279 break;
1280 case ME_X1:
1281 case ME_EPZS:
1282 {
1283 P_LEFT[0] = mv_table[mot_xy - 1][0];
1284 P_LEFT[1] = mv_table[mot_xy - 1][1];
1285
1286 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1287
1288 /* special case for first line */
1289 if (!s->first_slice_line) {
1290 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1291 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1292 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1293 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1294 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1295 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1296 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1297
1298 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1299 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1300 }
1301 c->pred_x= P_LEFT[0];
1302 c->pred_y= P_LEFT[1];
1303 }
1304
1305 if(mv_table == s->b_forw_mv_table){
1306 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1307 }else{
1308 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1309 }
1310
1311 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1312
1313 break;
1314 }
1315
1316 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1317
1318 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1319 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1320
1321 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1322 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1323 mv_table[mot_xy][0]= mx;
1324 mv_table[mot_xy][1]= my;
1325
1326 return dmin;
1327 }
1328
1329 static inline int check_bidir_mv(MpegEncContext * s,
1330 int motion_fx, int motion_fy,
1331 int motion_bx, int motion_by,
1332 int pred_fx, int pred_fy,
1333 int pred_bx, int pred_by,
1334 int size, int h)
1335 {
1336 //FIXME optimize?
1337 //FIXME better f_code prediction (max mv & distance)
1338 //FIXME pointers
1339 MotionEstContext * const c= &s->me;
1340 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1341 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1342 int stride= c->stride;
1343 uint8_t *dest_y = c->scratchpad;
1344 uint8_t *ptr;
1345 int dxy;
1346 int src_x, src_y;
1347 int fbmin;
1348 uint8_t **src_data= c->src[0];
1349 uint8_t **ref_data= c->ref[0];
1350 uint8_t **ref2_data= c->ref[2];
1351
1352 if(s->quarter_sample){
1353 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1354 src_x = motion_fx >> 2;
1355 src_y = motion_fy >> 2;
1356
1357 ptr = ref_data[0] + (src_y * stride) + src_x;
1358 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1359
1360 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1361 src_x = motion_bx >> 2;
1362 src_y = motion_by >> 2;
1363
1364 ptr = ref2_data[0] + (src_y * stride) + src_x;
1365 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1366 }else{
1367 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1368 src_x = motion_fx >> 1;
1369 src_y = motion_fy >> 1;
1370
1371 ptr = ref_data[0] + (src_y * stride) + src_x;
1372 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1373
1374 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1375 src_x = motion_bx >> 1;
1376 src_y = motion_by >> 1;
1377
1378 ptr = ref2_data[0] + (src_y * stride) + src_x;
1379 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1380 }
1381
1382 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1383 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1384 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1385
1386 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1387 }
1388 //FIXME CHROMA !!!
1389
1390 return fbmin;
1391 }
1392
1393 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1394 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1395 {
1396 MotionEstContext * const c= &s->me;
1397 const int mot_stride = s->mb_stride;
1398 const int xy = mb_y *mot_stride + mb_x;
1399 int fbmin;
1400 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1401 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1402 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1403 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1404 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1405 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1406 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1407 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1408 const int flags= c->sub_flags;
1409 const int qpel= flags&FLAG_QPEL;
1410 const int shift= 1+qpel;
1411 const int xmin= c->xmin<<shift;
1412 const int ymin= c->ymin<<shift;
1413 const int xmax= c->xmax<<shift;
1414 const int ymax= c->ymax<<shift;
1415 uint8_t map[8][8][8][8];
1416
1417 memset(map,0,sizeof(map));
1418 #define BIDIR_MAP(fx,fy,bx,by) \
1419 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1420 BIDIR_MAP(0,0,0,0) = 1;
1421
1422 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1423 motion_bx, motion_by,
1424 pred_fx, pred_fy,
1425 pred_bx, pred_by,
1426 0, 16);
1427
1428 if(s->avctx->bidir_refine){
1429 int score, end;
1430 #define CHECK_BIDIR(fx,fy,bx,by)\
1431 if( !BIDIR_MAP(fx,fy,bx,by)\
1432 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1433 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1434 BIDIR_MAP(fx,fy,bx,by) = 1;\
1435 score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
1436 if(score < fbmin){\
1437 fbmin= score;\
1438 motion_fx+=fx;\
1439 motion_fy+=fy;\
1440 motion_bx+=bx;\
1441 motion_by+=by;\
1442 end=0;\
1443 }\
1444 }
1445 #define CHECK_BIDIR2(a,b,c,d)\
1446 CHECK_BIDIR(a,b,c,d)\
1447 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1448
1449 #define CHECK_BIDIRR(a,b,c,d)\
1450 CHECK_BIDIR2(a,b,c,d)\
1451 CHECK_BIDIR2(b,c,d,a)\
1452 CHECK_BIDIR2(c,d,a,b)\
1453 CHECK_BIDIR2(d,a,b,c)
1454
1455 do{
1456 end=1;
1457
1458 CHECK_BIDIRR( 0, 0, 0, 1)
1459 if(s->avctx->bidir_refine > 1){
1460 CHECK_BIDIRR( 0, 0, 1, 1)
1461 CHECK_BIDIR2( 0, 1, 0, 1)
1462 CHECK_BIDIR2( 1, 0, 1, 0)
1463 CHECK_BIDIRR( 0, 0,-1, 1)
1464 CHECK_BIDIR2( 0,-1, 0, 1)
1465 CHECK_BIDIR2(-1, 0, 1, 0)
1466 if(s->avctx->bidir_refine > 2){
1467 CHECK_BIDIRR( 0, 1, 1, 1)
1468 CHECK_BIDIRR( 0,-1, 1, 1)
1469 CHECK_BIDIRR( 0, 1,-1, 1)
1470 CHECK_BIDIRR( 0, 1, 1,-1)
1471 if(s->avctx->bidir_refine > 3){
1472 CHECK_BIDIR2( 1, 1, 1, 1)
1473 CHECK_BIDIRR( 1, 1, 1,-1)
1474 CHECK_BIDIR2( 1, 1,-1,-1)
1475 CHECK_BIDIR2( 1,-1,-1, 1)
1476 CHECK_BIDIR2( 1,-1, 1,-1)
1477 }
1478 }
1479 }
1480 }while(!end);
1481 }
1482
1483 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1484 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1485 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1486 s->b_bidir_back_mv_table[xy][1]= motion_by;
1487
1488 return fbmin;
1489 }
1490
1491 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1492 {
1493 MotionEstContext * const c= &s->me;
1494 int P[10][2];
1495 const int mot_stride = s->mb_stride;
1496 const int mot_xy = mb_y*mot_stride + mb_x;
1497 const int shift= 1+s->quarter_sample;
1498 int dmin, i;
1499 const int time_pp= s->pp_time;
1500 const int time_pb= s->pb_time;
1501 int mx, my, xmin, xmax, ymin, ymax;
1502 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1503
1504 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1505 ymin= xmin=(-32)>>shift;
1506 ymax= xmax= 31>>shift;
1507
1508 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1509 s->mv_type= MV_TYPE_8X8;
1510 }else{
1511 s->mv_type= MV_TYPE_16X16;
1512 }
1513
1514 for(i=0; i<4; i++){
1515 int index= s->block_index[i];
1516 int min, max;
1517
1518 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1519 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1520 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1521 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1522 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1523 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1524
1525 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1526 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1527 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1528 min+= 16*mb_x - 1;
1529 xmax= FFMIN(xmax, s->width - max);
1530 xmin= FFMAX(xmin, - 16 - min);
1531
1532 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1533 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1534 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1535 min+= 16*mb_y - 1;
1536 ymax= FFMIN(ymax, s->height - max);
1537 ymin= FFMAX(ymin, - 16 - min);
1538
1539 if(s->mv_type == MV_TYPE_16X16) break;
1540 }
1541
1542 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1543
1544 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1545 s->b_direct_mv_table[mot_xy][0]= 0;
1546 s->b_direct_mv_table[mot_xy][1]= 0;
1547
1548 return 256*256*256*64;
1549 }
1550
1551 c->xmin= xmin;
1552 c->ymin= ymin;
1553 c->xmax= xmax;
1554 c->ymax= ymax;
1555 c->flags |= FLAG_DIRECT;
1556 c->sub_flags |= FLAG_DIRECT;
1557 c->pred_x=0;
1558 c->pred_y=0;
1559
1560 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1561 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1562
1563 /* special case for first line */
1564 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1565 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1566 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1567 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1568 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1569
1570 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1571 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1572 }
1573
1574 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1575 if(c->sub_flags&FLAG_QPEL)
1576 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1577 else
1578 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1579
1580 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1581 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1582
1583 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1584
1585 mv_table[mot_xy][0]= mx;
1586 mv_table[mot_xy][1]= my;
1587 c->flags &= ~FLAG_DIRECT;
1588 c->sub_flags &= ~FLAG_DIRECT;
1589
1590 return dmin;
1591 }
1592
1593 void ff_estimate_b_frame_motion(MpegEncContext * s,
1594 int mb_x, int mb_y)
1595 {
1596 MotionEstContext * const c= &s->me;
1597 const int penalty_factor= c->mb_penalty_factor;
1598 int fmin, bmin, dmin, fbmin, bimin, fimin;
1599 int type=0;
1600 const int xy = mb_y*s->mb_stride + mb_x;
1601 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1602
1603 get_limits(s, 16*mb_x, 16*mb_y);
1604
1605 c->skip=0;
1606
1607 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1608 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1609
1610 score= ((unsigned)(score*score + 128*256))>>16;
1611 c->mc_mb_var_sum_temp += score;
1612 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1613 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1614
1615 return;
1616 }
1617
1618 if(c->avctx->me_threshold){
1619 int vard= check_input_motion(s, mb_x, mb_y, 0);
1620
1621 if((vard+128)>>8 < c->avctx->me_threshold){
1622 // pix = c->src[0][0];
1623 // sum = s->dsp.pix_sum(pix, s->linesize);
1624 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1625
1626 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1627 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1628 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1629 c->mb_var_sum_temp += (varc+128)>>8;*/
1630 c->mc_mb_var_sum_temp += (vard+128)>>8;
1631 /* if (vard <= 64<<8 || vard < varc) {
1632 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1633 }else{
1634 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1635 }*/
1636 return;
1637 }
1638 if((vard+128)>>8 < c->avctx->mb_threshold){
1639 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1640 if(type == CANDIDATE_MB_TYPE_DIRECT){
1641 direct_search(s, mb_x, mb_y);
1642 }
1643 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1644 c->skip=0;
1645 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1646 }
1647 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1648 c->skip=0;
1649 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1650 }
1651 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1652 c->skip=0;
1653 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1654 interlaced_search(s, 0,
1655 s->b_field_mv_table[0], s->b_field_select_table[0],
1656 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1657 }
1658 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1659 c->skip=0;
1660 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1661 interlaced_search(s, 2,
1662 s->b_field_mv_table[1], s->b_field_select_table[1],
1663 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1664 }
1665 return;
1666 }
1667 }
1668
1669 if (s->codec_id == CODEC_ID_MPEG4)
1670 dmin= direct_search(s, mb_x, mb_y);
1671 else
1672 dmin= INT_MAX;
1673 //FIXME penalty stuff for non mpeg4
1674 c->skip=0;
1675 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1676
1677 c->skip=0;
1678 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1679 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1680
1681 c->skip=0;
1682 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1683 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1684
1685 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1686 //FIXME mb type penalty
1687 c->skip=0;
1688 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1689 fimin= interlaced_search(s, 0,
1690 s->b_field_mv_table[0], s->b_field_select_table[0],
1691 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1692 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1693 bimin= interlaced_search(s, 2,
1694 s->b_field_mv_table[1], s->b_field_select_table[1],
1695 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1696 }else
1697 fimin= bimin= INT_MAX;
1698
1699 {
1700 int score= fmin;
1701 type = CANDIDATE_MB_TYPE_FORWARD;
1702
1703 if (dmin <= score){
1704 score = dmin;
1705 type = CANDIDATE_MB_TYPE_DIRECT;
1706 }
1707 if(bmin<score){
1708 score=bmin;
1709 type= CANDIDATE_MB_TYPE_BACKWARD;
1710 }
1711 if(fbmin<score){
1712 score=fbmin;
1713 type= CANDIDATE_MB_TYPE_BIDIR;
1714 }
1715 if(fimin<score){
1716 score=fimin;
1717 type= CANDIDATE_MB_TYPE_FORWARD_I;
1718 }
1719 if(bimin<score){
1720 score=bimin;
1721 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1722 }
1723
1724 score= ((unsigned)(score*score + 128*256))>>16;
1725 c->mc_mb_var_sum_temp += score;
1726 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1727 }
1728
1729 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1730 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1731 if(fimin < INT_MAX)
1732 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1733 if(bimin < INT_MAX)
1734 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1735 if(fimin < INT_MAX && bimin < INT_MAX){
1736 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1737 }
1738 //FIXME something smarter
1739 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1740 if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1741 type |= CANDIDATE_MB_TYPE_DIRECT0;
1742 #if 0
1743 if(s->out_format == FMT_MPEG1)
1744 type |= CANDIDATE_MB_TYPE_INTRA;
1745 #endif
1746 }
1747
1748 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1749 }
1750
1751 /* find best f_code for ME which do unlimited searches */
1752 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1753 {
1754 if(s->me_method>=ME_EPZS){
1755 int score[8];
1756 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1757 uint8_t * fcode_tab= s->fcode_tab;
1758 int best_fcode=-1;
1759 int best_score=-10000000;
1760
1761 if(s->msmpeg4_version)
1762 range= FFMIN(range, 16);
1763 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1764 range= FFMIN(range, 256);
1765
1766 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1767
1768 for(y=0; y<s->mb_height; y++){
1769 int x;
1770 int xy= y*s->mb_stride;
1771 for(x=0; x<s->mb_width; x++){
1772 if(s->mb_type[xy] & type){
1773 int mx= mv_table[xy][0];
1774 int my= mv_table[xy][1];
1775 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1776 fcode_tab[my + MAX_MV]);
1777 int j;
1778
1779 if(mx >= range || mx < -range ||
1780 my >= range || my < -range)
1781 continue;
1782
1783 for(j=0; j<fcode && j<8; j++){
1784 if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1785 score[j]-= 170;
1786 }
1787 }
1788 xy++;
1789 }
1790 }
1791
1792 for(i=1; i<8; i++){
1793 if(score[i] > best_score){
1794 best_score= score[i];
1795 best_fcode= i;
1796 }
1797 // printf("%d %d\n", i, score[i]);
1798 }
1799
1800 // printf("fcode: %d type: %d\n", i, s->pict_type);
1801 return best_fcode;
1802 /* for(i=0; i<=MAX_FCODE; i++){
1803 printf("%d ", mv_num[i]);
1804 }
1805 printf("\n");*/
1806 }else{
1807 return 1;
1808 }
1809 }
1810
1811 void ff_fix_long_p_mvs(MpegEncContext * s)
1812 {
1813 MotionEstContext * const c= &s->me;
1814 const int f_code= s->f_code;
1815 int y, range;
1816 assert(s->pict_type==FF_P_TYPE);
1817
1818 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1819
1820 assert(range <= 16 || !s->msmpeg4_version);
1821 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1822
1823 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1824
1825 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1826 if(s->flags&CODEC_FLAG_4MV){
1827 const int wrap= s->b8_stride;
1828
1829 /* clip / convert to intra 8x8 type MVs */
1830 for(y=0; y<s->mb_height; y++){
1831 int xy= y*2*wrap;
1832 int i= y*s->mb_stride;
1833 int x;
1834
1835 for(x=0; x<s->mb_width; x++){
1836 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1837 int block;
1838 for(block=0; block<4; block++){
1839 int off= (block& 1) + (block>>1)*wrap;
1840 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1841 int my= s->current_picture.motion_val[0][ xy + off ][1];
1842
1843 if( mx >=range || mx <-range
1844 || my >=range || my <-range){
1845 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1846 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1847 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1848 }
1849 }
1850 }
1851 xy+=2;
1852 i++;
1853 }
1854 }
1855 }
1856 }
1857
1858 /**
1859 *
1860 * @param truncate 1 for truncation, 0 for using intra
1861 */
1862 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1863 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1864 {
1865 MotionEstContext * const c= &s->me;
1866 int y, h_range, v_range;
1867
1868 // RAL: 8 in MPEG-1, 16 in MPEG-4
1869 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1870
1871 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1872
1873 h_range= range;
1874 v_range= field_select_table ? range>>1 : range;
1875
1876 /* clip / convert to intra 16x16 type MVs */
1877 for(y=0; y<s->mb_height; y++){
1878 int x;
1879 int xy= y*s->mb_stride;
1880 for(x=0; x<s->mb_width; x++){
1881 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1882 if(field_select_table==NULL || field_select_table[xy] == field_select){
1883 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1884 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1885
1886 if(truncate){
1887 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1888 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1889 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1890 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1891 }else{
1892 s->mb_type[xy] &= ~type;
1893 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1894 mv_table[xy][0]=
1895 mv_table[xy][1]= 0;
1896 }
1897 }
1898 }
1899 }
1900 xy++;
1901 }
1902 }
1903 }
lucabe's unfinished and experimental patches for ffmpeg