?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Minor simplification of rotate_block()
[ffmpeg-lucabe.git] / libavcodec / motion_est.c
blob54bc93da38a5daf0b7ce2c42ffc44fb31799c6ca
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 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 "mpegvideo.h"
36
37 #undef NDEBUG
38 #include <assert.h>
39
40 #define SQ(a) ((a)*(a))
41
42 #define P_LEFT P[1]
43 #define P_TOP P[2]
44 #define P_TOPRIGHT P[3]
45 #define P_MEDIAN P[4]
46 #define P_MV1 P[9]
47
48 static inline int sad_hpel_motion_search(MpegEncContext * s,
49 int *mx_ptr, int *my_ptr, int dmin,
50 int src_index, int ref_index,
51 int size, int h);
52
53 static inline int update_map_generation(MotionEstContext *c)
54 {
55 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
56 if(c->map_generation==0){
57 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
58 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59 }
60 return c->map_generation;
61 }
62
63 /* shape adaptive search stuff */
64 typedef struct Minima{
65 int height;
66 int x, y;
67 int checked;
68 }Minima;
69
70 static int minima_cmp(const void *a, const void *b){
71 const Minima *da = (const Minima *) a;
72 const Minima *db = (const Minima *) b;
73
74 return da->height - db->height;
75 }
76
77 #define FLAG_QPEL 1 //must be 1
78 #define FLAG_CHROMA 2
79 #define FLAG_DIRECT 4
80
81 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){
82 const int offset[3]= {
83 y*c-> stride + x,
84 ((y*c->uvstride + x)>>1),
85 ((y*c->uvstride + x)>>1),
86 };
87 int i;
88 for(i=0; i<3; i++){
89 c->src[0][i]= src [i] + offset[i];
90 c->ref[0][i]= ref [i] + offset[i];
91 }
92 if(ref_index){
93 for(i=0; i<3; i++){
94 c->ref[ref_index][i]= ref2[i] + offset[i];
95 }
96 }
97 }
98
99 static int get_flags(MotionEstContext *c, int direct, int chroma){
100 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
101 + (direct ? FLAG_DIRECT : 0)
102 + (chroma ? FLAG_CHROMA : 0);
103 }
104
105 /*! \brief compares a block (either a full macroblock or a partition thereof)
106 against a proposed motion-compensated prediction of that block
107 */
108 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
109 const int size, const int h, int ref_index, int src_index,
110 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
111 MotionEstContext * const c= &s->me;
112 const int stride= c->stride;
113 const int uvstride= c->uvstride;
114 const int qpel= flags&FLAG_QPEL;
115 const int chroma= flags&FLAG_CHROMA;
116 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
117 const int hx= subx + (x<<(1+qpel));
118 const int hy= suby + (y<<(1+qpel));
119 uint8_t * const * const ref= c->ref[ref_index];
120 uint8_t * const * const src= c->src[src_index];
121 int d;
122 //FIXME check chroma 4mv, (no crashes ...)
123 if(flags&FLAG_DIRECT){
124 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
125 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
126 const int time_pp= s->pp_time;
127 const int time_pb= s->pb_time;
128 const int mask= 2*qpel+1;
129 if(s->mv_type==MV_TYPE_8X8){
130 int i;
131 for(i=0; i<4; i++){
132 int fx = c->direct_basis_mv[i][0] + hx;
133 int fy = c->direct_basis_mv[i][1] + hy;
134 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));
135 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));
136 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
137 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
138
139 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
140 if(qpel){
141 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
142 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
143 }else{
144 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
145 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
146 }
147 }
148 }else{
149 int fx = c->direct_basis_mv[0][0] + hx;
150 int fy = c->direct_basis_mv[0][1] + hy;
151 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
152 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
153 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
154 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
155
156 if(qpel){
157 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
158 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
159 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
160 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
161 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
162 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
163 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
164 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
165 }else{
166 assert((fx>>1) + 16*s->mb_x >= -16);
167 assert((fy>>1) + 16*s->mb_y >= -16);
168 assert((fx>>1) + 16*s->mb_x <= s->width);
169 assert((fy>>1) + 16*s->mb_y <= s->height);
170 assert((bx>>1) + 16*s->mb_x >= -16);
171 assert((by>>1) + 16*s->mb_y >= -16);
172 assert((bx>>1) + 16*s->mb_x <= s->width);
173 assert((by>>1) + 16*s->mb_y <= s->height);
174
175 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
176 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
177 }
178 }
179 d = cmp_func(s, c->temp, src[0], stride, 16);
180 }else
181 d= 256*256*256*32;
182 }else{
183 int uvdxy; /* no, it might not be used uninitialized */
184 if(dxy){
185 if(qpel){
186 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
187 if(chroma){
188 int cx= hx/2;
189 int cy= hy/2;
190 cx= (cx>>1)|(cx&1);
191 cy= (cy>>1)|(cy&1);
192 uvdxy= (cx&1) + 2*(cy&1);
193 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
194 }
195 }else{
196 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
197 if(chroma)
198 uvdxy= dxy | (x&1) | (2*(y&1));
199 }
200 d = cmp_func(s, c->temp, src[0], stride, h);
201 }else{
202 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
203 if(chroma)
204 uvdxy= (x&1) + 2*(y&1);
205 }
206 if(chroma){
207 uint8_t * const uvtemp= c->temp + 16*stride;
208 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
209 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
211 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
212 }
213 }
214 #if 0
215 if(full_pel){
216 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
217 score_map[index]= d;
218 }
219
220 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
221 #endif
222 return d;
223 }
224
225 #include "motion_est_template.c"
226
227 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
228 return 0;
229 }
230
231 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
232 }
233
234 void ff_init_me(MpegEncContext *s){
235 MotionEstContext * const c= &s->me;
236 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
237 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
238 c->avctx= s->avctx;
239
240 if(cache_size < 2*dia_size && !c->stride){
241 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
242 }
243
244 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
245 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
246 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
247 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
248
249 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
250 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
251 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
252
253 /*FIXME s->no_rounding b_type*/
254 if(s->flags&CODEC_FLAG_QPEL){
255 c->sub_motion_search= qpel_motion_search;
256 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
257 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
258 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
259 }else{
260 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
261 c->sub_motion_search= hpel_motion_search;
262 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
263 && c->avctx-> me_cmp == FF_CMP_SAD
264 && c->avctx-> mb_cmp == FF_CMP_SAD)
265 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
266 else
267 c->sub_motion_search= hpel_motion_search;
268 }
269 c->hpel_avg= s->dsp.avg_pixels_tab;
270 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
271 else c->hpel_put= s->dsp.put_pixels_tab;
272
273 if(s->linesize){
274 c->stride = s->linesize;
275 c->uvstride= s->uvlinesize;
276 }else{
277 c->stride = 16*s->mb_width + 32;
278 c->uvstride= 8*s->mb_width + 16;
279 }
280
281 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
282 * not have yet, and even if we had, the motion estimation code
283 * does not expect it. */
284 if(s->codec_id != CODEC_ID_SNOW){
285 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
286 s->dsp.me_cmp[2]= zero_cmp;
287 }
288 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
289 s->dsp.me_sub_cmp[2]= zero_cmp;
290 }
291 c->hpel_put[2][0]= c->hpel_put[2][1]=
292 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
293 }
294
295 if(s->codec_id == CODEC_ID_H261){
296 c->sub_motion_search= no_sub_motion_search;
297 }
298
299 c->temp= c->scratchpad;
300 }
301
302 #if 0
303 static int pix_dev(uint8_t * pix, int line_size, int mean)
304 {
305 int s, i, j;
306
307 s = 0;
308 for (i = 0; i < 16; i++) {
309 for (j = 0; j < 16; j += 8) {
310 s += FFABS(pix[0]-mean);
311 s += FFABS(pix[1]-mean);
312 s += FFABS(pix[2]-mean);
313 s += FFABS(pix[3]-mean);
314 s += FFABS(pix[4]-mean);
315 s += FFABS(pix[5]-mean);
316 s += FFABS(pix[6]-mean);
317 s += FFABS(pix[7]-mean);
318 pix += 8;
319 }
320 pix += line_size - 16;
321 }
322 return s;
323 }
324 #endif
325
326 static inline void no_motion_search(MpegEncContext * s,
327 int *mx_ptr, int *my_ptr)
328 {
329 *mx_ptr = 16 * s->mb_x;
330 *my_ptr = 16 * s->mb_y;
331 }
332
333 #if 0 /* the use of these functions is inside #if 0 */
334 static int full_motion_search(MpegEncContext * s,
335 int *mx_ptr, int *my_ptr, int range,
336 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
337 {
338 int x1, y1, x2, y2, xx, yy, x, y;
339 int mx, my, dmin, d;
340 uint8_t *pix;
341
342 xx = 16 * s->mb_x;
343 yy = 16 * s->mb_y;
344 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
345 if (x1 < xmin)
346 x1 = xmin;
347 x2 = xx + range - 1;
348 if (x2 > xmax)
349 x2 = xmax;
350 y1 = yy - range + 1;
351 if (y1 < ymin)
352 y1 = ymin;
353 y2 = yy + range - 1;
354 if (y2 > ymax)
355 y2 = ymax;
356 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
357 dmin = 0x7fffffff;
358 mx = 0;
359 my = 0;
360 for (y = y1; y <= y2; y++) {
361 for (x = x1; x <= x2; x++) {
362 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
363 s->linesize, 16);
364 if (d < dmin ||
365 (d == dmin &&
366 (abs(x - xx) + abs(y - yy)) <
367 (abs(mx - xx) + abs(my - yy)))) {
368 dmin = d;
369 mx = x;
370 my = y;
371 }
372 }
373 }
374
375 *mx_ptr = mx;
376 *my_ptr = my;
377
378 #if 0
379 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
380 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
381 av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
382 }
383 #endif
384 return dmin;
385 }
386
387
388 static int log_motion_search(MpegEncContext * s,
389 int *mx_ptr, int *my_ptr, int range,
390 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
391 {
392 int x1, y1, x2, y2, xx, yy, x, y;
393 int mx, my, dmin, d;
394 uint8_t *pix;
395
396 xx = s->mb_x << 4;
397 yy = s->mb_y << 4;
398
399 /* Left limit */
400 x1 = xx - range;
401 if (x1 < xmin)
402 x1 = xmin;
403
404 /* Right limit */
405 x2 = xx + range;
406 if (x2 > xmax)
407 x2 = xmax;
408
409 /* Upper limit */
410 y1 = yy - range;
411 if (y1 < ymin)
412 y1 = ymin;
413
414 /* Lower limit */
415 y2 = yy + range;
416 if (y2 > ymax)
417 y2 = ymax;
418
419 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
420 dmin = 0x7fffffff;
421 mx = 0;
422 my = 0;
423
424 do {
425 for (y = y1; y <= y2; y += range) {
426 for (x = x1; x <= x2; x += range) {
427 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
428 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
429 dmin = d;
430 mx = x;
431 my = y;
432 }
433 }
434 }
435
436 range = range >> 1;
437
438 x1 = mx - range;
439 if (x1 < xmin)
440 x1 = xmin;
441
442 x2 = mx + range;
443 if (x2 > xmax)
444 x2 = xmax;
445
446 y1 = my - range;
447 if (y1 < ymin)
448 y1 = ymin;
449
450 y2 = my + range;
451 if (y2 > ymax)
452 y2 = ymax;
453
454 } while (range >= 1);
455
456 #ifdef DEBUG
457 av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
458 #endif
459 *mx_ptr = mx;
460 *my_ptr = my;
461 return dmin;
462 }
463
464 static int phods_motion_search(MpegEncContext * s,
465 int *mx_ptr, int *my_ptr, int range,
466 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
467 {
468 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
469 int mx, my, dminx, dminy;
470 uint8_t *pix;
471
472 xx = s->mb_x << 4;
473 yy = s->mb_y << 4;
474
475 /* Left limit */
476 x1 = xx - range;
477 if (x1 < xmin)
478 x1 = xmin;
479
480 /* Right limit */
481 x2 = xx + range;
482 if (x2 > xmax)
483 x2 = xmax;
484
485 /* Upper limit */
486 y1 = yy - range;
487 if (y1 < ymin)
488 y1 = ymin;
489
490 /* Lower limit */
491 y2 = yy + range;
492 if (y2 > ymax)
493 y2 = ymax;
494
495 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
496 mx = 0;
497 my = 0;
498
499 x = xx;
500 y = yy;
501 do {
502 dminx = 0x7fffffff;
503 dminy = 0x7fffffff;
504
505 lastx = x;
506 for (x = x1; x <= x2; x += range) {
507 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
508 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
509 dminx = d;
510 mx = x;
511 }
512 }
513
514 x = lastx;
515 for (y = y1; y <= y2; y += range) {
516 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
517 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
518 dminy = d;
519 my = y;
520 }
521 }
522
523 range = range >> 1;
524
525 x = mx;
526 y = my;
527 x1 = mx - range;
528 if (x1 < xmin)
529 x1 = xmin;
530
531 x2 = mx + range;
532 if (x2 > xmax)
533 x2 = xmax;
534
535 y1 = my - range;
536 if (y1 < ymin)
537 y1 = ymin;
538
539 y2 = my + range;
540 if (y2 > ymax)
541 y2 = ymax;
542
543 } while (range >= 1);
544
545 #ifdef DEBUG
546 av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
547 #endif
548
549 /* half pixel search */
550 *mx_ptr = mx;
551 *my_ptr = my;
552 return dminy;
553 }
554 #endif /* 0 */
555
556 #define Z_THRESHOLD 256
557
558 #define CHECK_SAD_HALF_MV(suffix, x, y) \
559 {\
560 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
561 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
562 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
563 }
564
565 static inline int sad_hpel_motion_search(MpegEncContext * s,
566 int *mx_ptr, int *my_ptr, int dmin,
567 int src_index, int ref_index,
568 int size, int h)
569 {
570 MotionEstContext * const c= &s->me;
571 const int penalty_factor= c->sub_penalty_factor;
572 int mx, my, dminh;
573 uint8_t *pix, *ptr;
574 int stride= c->stride;
575 const int flags= c->sub_flags;
576 LOAD_COMMON
577
578 assert(flags == 0);
579
580 if(c->skip){
581 // printf("S");
582 *mx_ptr = 0;
583 *my_ptr = 0;
584 return dmin;
585 }
586 // printf("N");
587
588 pix = c->src[src_index][0];
589
590 mx = *mx_ptr;
591 my = *my_ptr;
592 ptr = c->ref[ref_index][0] + (my * stride) + mx;
593
594 dminh = dmin;
595
596 if (mx > xmin && mx < xmax &&
597 my > ymin && my < ymax) {
598 int dx=0, dy=0;
599 int d, pen_x, pen_y;
600 const int index= (my<<ME_MAP_SHIFT) + mx;
601 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
602 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
603 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
604 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
605 mx<<=1;
606 my<<=1;
607
608
609 pen_x= pred_x + mx;
610 pen_y= pred_y + my;
611
612 ptr-= stride;
613 if(t<=b){
614 CHECK_SAD_HALF_MV(y2 , 0, -1)
615 if(l<=r){
616 CHECK_SAD_HALF_MV(xy2, -1, -1)
617 if(t+r<=b+l){
618 CHECK_SAD_HALF_MV(xy2, +1, -1)
619 ptr+= stride;
620 }else{
621 ptr+= stride;
622 CHECK_SAD_HALF_MV(xy2, -1, +1)
623 }
624 CHECK_SAD_HALF_MV(x2 , -1, 0)
625 }else{
626 CHECK_SAD_HALF_MV(xy2, +1, -1)
627 if(t+l<=b+r){
628 CHECK_SAD_HALF_MV(xy2, -1, -1)
629 ptr+= stride;
630 }else{
631 ptr+= stride;
632 CHECK_SAD_HALF_MV(xy2, +1, +1)
633 }
634 CHECK_SAD_HALF_MV(x2 , +1, 0)
635 }
636 }else{
637 if(l<=r){
638 if(t+l<=b+r){
639 CHECK_SAD_HALF_MV(xy2, -1, -1)
640 ptr+= stride;
641 }else{
642 ptr+= stride;
643 CHECK_SAD_HALF_MV(xy2, +1, +1)
644 }
645 CHECK_SAD_HALF_MV(x2 , -1, 0)
646 CHECK_SAD_HALF_MV(xy2, -1, +1)
647 }else{
648 if(t+r<=b+l){
649 CHECK_SAD_HALF_MV(xy2, +1, -1)
650 ptr+= stride;
651 }else{
652 ptr+= stride;
653 CHECK_SAD_HALF_MV(xy2, -1, +1)
654 }
655 CHECK_SAD_HALF_MV(x2 , +1, 0)
656 CHECK_SAD_HALF_MV(xy2, +1, +1)
657 }
658 CHECK_SAD_HALF_MV(y2 , 0, +1)
659 }
660 mx+=dx;
661 my+=dy;
662
663 }else{
664 mx<<=1;
665 my<<=1;
666 }
667
668 *mx_ptr = mx;
669 *my_ptr = my;
670 return dminh;
671 }
672
673 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
674 {
675 const int xy= s->mb_x + s->mb_y*s->mb_stride;
676
677 s->p_mv_table[xy][0] = mx;
678 s->p_mv_table[xy][1] = my;
679
680 /* has already been set to the 4 MV if 4MV is done */
681 if(mv4){
682 int mot_xy= s->block_index[0];
683
684 s->current_picture.motion_val[0][mot_xy ][0]= mx;
685 s->current_picture.motion_val[0][mot_xy ][1]= my;
686 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
687 s->current_picture.motion_val[0][mot_xy+1][1]= my;
688
689 mot_xy += s->b8_stride;
690 s->current_picture.motion_val[0][mot_xy ][0]= mx;
691 s->current_picture.motion_val[0][mot_xy ][1]= my;
692 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
693 s->current_picture.motion_val[0][mot_xy+1][1]= my;
694 }
695 }
696
697 /**
698 * get fullpel ME search limits.
699 */
700 static inline void get_limits(MpegEncContext *s, int x, int y)
701 {
702 MotionEstContext * const c= &s->me;
703 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
704 /*
705 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
706 else c->range= 16;
707 */
708 if (s->unrestricted_mv) {
709 c->xmin = - x - 16;
710 c->ymin = - y - 16;
711 c->xmax = - x + s->mb_width *16;
712 c->ymax = - y + s->mb_height*16;
713 } else if (s->out_format == FMT_H261){
714 // Search range of H261 is different from other codec standards
715 c->xmin = (x > 15) ? - 15 : 0;
716 c->ymin = (y > 15) ? - 15 : 0;
717 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
718 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
719 } else {
720 c->xmin = - x;
721 c->ymin = - y;
722 c->xmax = - x + s->mb_width *16 - 16;
723 c->ymax = - y + s->mb_height*16 - 16;
724 }
725 if(range){
726 c->xmin = FFMAX(c->xmin,-range);
727 c->xmax = FFMIN(c->xmax, range);
728 c->ymin = FFMAX(c->ymin,-range);
729 c->ymax = FFMIN(c->ymax, range);
730 }
731 }
732
733 static inline void init_mv4_ref(MotionEstContext *c){
734 const int stride= c->stride;
735
736 c->ref[1][0] = c->ref[0][0] + 8;
737 c->ref[2][0] = c->ref[0][0] + 8*stride;
738 c->ref[3][0] = c->ref[2][0] + 8;
739 c->src[1][0] = c->src[0][0] + 8;
740 c->src[2][0] = c->src[0][0] + 8*stride;
741 c->src[3][0] = c->src[2][0] + 8;
742 }
743
744 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
745 {
746 MotionEstContext * const c= &s->me;
747 const int size= 1;
748 const int h=8;
749 int block;
750 int P[10][2];
751 int dmin_sum=0, mx4_sum=0, my4_sum=0;
752 int same=1;
753 const int stride= c->stride;
754 uint8_t *mv_penalty= c->current_mv_penalty;
755
756 init_mv4_ref(c);
757
758 for(block=0; block<4; block++){
759 int mx4, my4;
760 int pred_x4, pred_y4;
761 int dmin4;
762 static const int off[4]= {2, 1, 1, -1};
763 const int mot_stride = s->b8_stride;
764 const int mot_xy = s->block_index[block];
765
766 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
767 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
768
769 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
770
771 /* special case for first line */
772 if (s->first_slice_line && block<2) {
773 c->pred_x= pred_x4= P_LEFT[0];
774 c->pred_y= pred_y4= P_LEFT[1];
775 } else {
776 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
777 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
778 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
779 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
780 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
781 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
782 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
783 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
784
785 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
786 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
787
788 c->pred_x= pred_x4 = P_MEDIAN[0];
789 c->pred_y= pred_y4 = P_MEDIAN[1];
790 }
791 P_MV1[0]= mx;
792 P_MV1[1]= my;
793
794 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
795
796 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
797
798 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
799 int dxy;
800 const int offset= ((block&1) + (block>>1)*stride)*8;
801 uint8_t *dest_y = c->scratchpad + offset;
802 if(s->quarter_sample){
803 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
804 dxy = ((my4 & 3) << 2) | (mx4 & 3);
805
806 if(s->no_rounding)
807 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
808 else
809 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
810 }else{
811 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
812 dxy = ((my4 & 1) << 1) | (mx4 & 1);
813
814 if(s->no_rounding)
815 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
816 else
817 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
818 }
819 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
820 }else
821 dmin_sum+= dmin4;
822
823 if(s->quarter_sample){
824 mx4_sum+= mx4/2;
825 my4_sum+= my4/2;
826 }else{
827 mx4_sum+= mx4;
828 my4_sum+= my4;
829 }
830
831 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
832 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
833
834 if(mx4 != mx || my4 != my) same=0;
835 }
836
837 if(same)
838 return INT_MAX;
839
840 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
841 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);
842 }
843
844 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
845 int dxy;
846 int mx, my;
847 int offset;
848
849 mx= ff_h263_round_chroma(mx4_sum);
850 my= ff_h263_round_chroma(my4_sum);
851 dxy = ((my & 1) << 1) | (mx & 1);
852
853 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
854
855 if(s->no_rounding){
856 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
857 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
858 }else{
859 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
860 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
861 }
862
863 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);
864 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);
865 }
866
867 c->pred_x= mx;
868 c->pred_y= my;
869
870 switch(c->avctx->mb_cmp&0xFF){
871 /*case FF_CMP_SSE:
872 return dmin_sum+ 32*s->qscale*s->qscale;*/
873 case FF_CMP_RD:
874 return dmin_sum;
875 default:
876 return dmin_sum+ 11*c->mb_penalty_factor;
877 }
878 }
879
880 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
881 MotionEstContext * const c= &s->me;
882
883 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
884 c->src[1][0] = c->src[0][0] + s->linesize;
885 if(c->flags & FLAG_CHROMA){
886 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
887 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
888 c->src[1][1] = c->src[0][1] + s->uvlinesize;
889 c->src[1][2] = c->src[0][2] + s->uvlinesize;
890 }
891 }
892
893 static int interlaced_search(MpegEncContext *s, int ref_index,
894 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
895 {
896 MotionEstContext * const c= &s->me;
897 const int size=0;
898 const int h=8;
899 int block;
900 int P[10][2];
901 uint8_t * const mv_penalty= c->current_mv_penalty;
902 int same=1;
903 const int stride= 2*s->linesize;
904 int dmin_sum= 0;
905 const int mot_stride= s->mb_stride;
906 const int xy= s->mb_x + s->mb_y*mot_stride;
907
908 c->ymin>>=1;
909 c->ymax>>=1;
910 c->stride<<=1;
911 c->uvstride<<=1;
912 init_interlaced_ref(s, ref_index);
913
914 for(block=0; block<2; block++){
915 int field_select;
916 int best_dmin= INT_MAX;
917 int best_field= -1;
918
919 for(field_select=0; field_select<2; field_select++){
920 int dmin, mx_i, my_i;
921 int16_t (*mv_table)[2]= mv_tables[block][field_select];
922
923 if(user_field_select){
924 assert(field_select==0 || field_select==1);
925 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
926 if(field_select_tables[block][xy] != field_select)
927 continue;
928 }
929
930 P_LEFT[0] = mv_table[xy - 1][0];
931 P_LEFT[1] = mv_table[xy - 1][1];
932 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
933
934 c->pred_x= P_LEFT[0];
935 c->pred_y= P_LEFT[1];
936
937 if(!s->first_slice_line){
938 P_TOP[0] = mv_table[xy - mot_stride][0];
939 P_TOP[1] = mv_table[xy - mot_stride][1];
940 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
941 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
942 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
943 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
944 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
945 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
946
947 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
948 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
949 }
950 P_MV1[0]= mx; //FIXME not correct if block != field_select
951 P_MV1[1]= my / 2;
952
953 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
954
955 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
956
957 mv_table[xy][0]= mx_i;
958 mv_table[xy][1]= my_i;
959
960 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
961 int dxy;
962
963 //FIXME chroma ME
964 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
965 dxy = ((my_i & 1) << 1) | (mx_i & 1);
966
967 if(s->no_rounding){
968 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
969 }else{
970 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
971 }
972 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
973 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
974 }else
975 dmin+= c->mb_penalty_factor; //field_select bits
976
977 dmin += field_select != block; //slightly prefer same field
978
979 if(dmin < best_dmin){
980 best_dmin= dmin;
981 best_field= field_select;
982 }
983 }
984 {
985 int16_t (*mv_table)[2]= mv_tables[block][best_field];
986
987 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
988 if(mv_table[xy][1]&1) same=0;
989 if(mv_table[xy][1]*2 != my) same=0;
990 if(best_field != block) same=0;
991 }
992
993 field_select_tables[block][xy]= best_field;
994 dmin_sum += best_dmin;
995 }
996
997 c->ymin<<=1;
998 c->ymax<<=1;
999 c->stride>>=1;
1000 c->uvstride>>=1;
1001
1002 if(same)
1003 return INT_MAX;
1004
1005 switch(c->avctx->mb_cmp&0xFF){
1006 /*case FF_CMP_SSE:
1007 return dmin_sum+ 32*s->qscale*s->qscale;*/
1008 case FF_CMP_RD:
1009 return dmin_sum;
1010 default:
1011 return dmin_sum+ 11*c->mb_penalty_factor;
1012 }
1013 }
1014
1015 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1016 int ymax= s->me.ymax>>interlaced;
1017 int ymin= s->me.ymin>>interlaced;
1018
1019 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1020 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1021 if(mv[1] < ymin) mv[1] = ymin;
1022 if(mv[1] > ymax) mv[1] = ymax;
1023 }
1024
1025 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1026 MotionEstContext * const c= &s->me;
1027 Picture *p= s->current_picture_ptr;
1028 int mb_xy= mb_x + mb_y*s->mb_stride;
1029 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1030 int mb_type= s->current_picture.mb_type[mb_xy];
1031 int flags= c->flags;
1032 int shift= (flags&FLAG_QPEL) + 1;
1033 int mask= (1<<shift)-1;
1034 int x, y, i;
1035 int d=0;
1036 me_cmp_func cmpf= s->dsp.sse[0];
1037 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1038
1039 if(p_type && USES_LIST(mb_type, 1)){
1040 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1041 return INT_MAX/2;
1042 }
1043 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1044
1045 for(i=0; i<4; i++){
1046 int xy= s->block_index[i];
1047 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1048 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1049 }
1050
1051 if(IS_INTERLACED(mb_type)){
1052 int xy2= xy + s->b8_stride;
1053 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1054 c->stride<<=1;
1055 c->uvstride<<=1;
1056
1057 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1058 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1059 return INT_MAX/2;
1060 }
1061
1062 if(USES_LIST(mb_type, 0)){
1063 int field_select0= p->ref_index[0][xy ];
1064 int field_select1= p->ref_index[0][xy2];
1065 assert(field_select0==0 ||field_select0==1);
1066 assert(field_select1==0 ||field_select1==1);
1067 init_interlaced_ref(s, 0);
1068
1069 if(p_type){
1070 s->p_field_select_table[0][mb_xy]= field_select0;
1071 s->p_field_select_table[1][mb_xy]= field_select1;
1072 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1073 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1074 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1075 }else{
1076 s->b_field_select_table[0][0][mb_xy]= field_select0;
1077 s->b_field_select_table[0][1][mb_xy]= field_select1;
1078 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1079 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1080 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1081 }
1082
1083 x= p->motion_val[0][xy ][0];
1084 y= p->motion_val[0][xy ][1];
1085 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1086 x= p->motion_val[0][xy2][0];
1087 y= p->motion_val[0][xy2][1];
1088 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1089 }
1090 if(USES_LIST(mb_type, 1)){
1091 int field_select0= p->ref_index[1][xy ];
1092 int field_select1= p->ref_index[1][xy2];
1093 assert(field_select0==0 ||field_select0==1);
1094 assert(field_select1==0 ||field_select1==1);
1095 init_interlaced_ref(s, 2);
1096
1097 s->b_field_select_table[1][0][mb_xy]= field_select0;
1098 s->b_field_select_table[1][1][mb_xy]= field_select1;
1099 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1100 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1101 if(USES_LIST(mb_type, 0)){
1102 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1103 }else{
1104 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1105 }
1106
1107 x= p->motion_val[1][xy ][0];
1108 y= p->motion_val[1][xy ][1];
1109 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1110 x= p->motion_val[1][xy2][0];
1111 y= p->motion_val[1][xy2][1];
1112 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1113 //FIXME bidir scores
1114 }
1115 c->stride>>=1;
1116 c->uvstride>>=1;
1117 }else if(IS_8X8(mb_type)){
1118 if(!(s->flags & CODEC_FLAG_4MV)){
1119 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1120 return INT_MAX/2;
1121 }
1122 cmpf= s->dsp.sse[1];
1123 chroma_cmpf= s->dsp.sse[1];
1124 init_mv4_ref(c);
1125 for(i=0; i<4; i++){
1126 xy= s->block_index[i];
1127 x= p->motion_val[0][xy][0];
1128 y= p->motion_val[0][xy][1];
1129 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1130 }
1131 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1132 }else{
1133 if(USES_LIST(mb_type, 0)){
1134 if(p_type){
1135 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1136 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1137 }else if(USES_LIST(mb_type, 1)){
1138 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1139 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1140 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1141 }else{
1142 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1143 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1144 }
1145 x= p->motion_val[0][xy][0];
1146 y= p->motion_val[0][xy][1];
1147 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1148 }else if(USES_LIST(mb_type, 1)){
1149 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1150 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1151
1152 x= p->motion_val[1][xy][0];
1153 y= p->motion_val[1][xy][1];
1154 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1155 }else
1156 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1157 }
1158 return d;
1159 }
1160
1161 void ff_estimate_p_frame_motion(MpegEncContext * s,
1162 int mb_x, int mb_y)
1163 {
1164 MotionEstContext * const c= &s->me;
1165 uint8_t *pix, *ppix;
1166 int sum, mx, my, dmin;
1167 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1168 int vard; ///< sum of squared differences with the estimated motion vector
1169 int P[10][2];
1170 const int shift= 1+s->quarter_sample;
1171 int mb_type=0;
1172 Picture * const pic= &s->current_picture;
1173
1174 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1175
1176 assert(s->quarter_sample==0 || s->quarter_sample==1);
1177 assert(s->linesize == c->stride);
1178 assert(s->uvlinesize == c->uvstride);
1179
1180 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1181 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1182 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1183 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1184
1185 get_limits(s, 16*mb_x, 16*mb_y);
1186 c->skip=0;
1187
1188 /* intra / predictive decision */
1189 pix = c->src[0][0];
1190 sum = s->dsp.pix_sum(pix, s->linesize);
1191 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1192
1193 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1194 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1195 c->mb_var_sum_temp += (varc+128)>>8;
1196
1197 if(c->avctx->me_threshold){
1198 vard= check_input_motion(s, mb_x, mb_y, 1);
1199
1200 if((vard+128)>>8 < c->avctx->me_threshold){
1201 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1202 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1203 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1204 c->mc_mb_var_sum_temp += (vard+128)>>8;
1205 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1206 return;
1207 }
1208 if((vard+128)>>8 < c->avctx->mb_threshold)
1209 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1210 }
1211
1212 switch(s->me_method) {
1213 case ME_ZERO:
1214 default:
1215 no_motion_search(s, &mx, &my);
1216 mx-= mb_x*16;
1217 my-= mb_y*16;
1218 dmin = 0;
1219 break;
1220 #if 0
1221 case ME_FULL:
1222 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1223 mx-= mb_x*16;
1224 my-= mb_y*16;
1225 break;
1226 case ME_LOG:
1227 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1228 mx-= mb_x*16;
1229 my-= mb_y*16;
1230 break;
1231 case ME_PHODS:
1232 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1233 mx-= mb_x*16;
1234 my-= mb_y*16;
1235 break;
1236 #endif
1237 case ME_X1:
1238 case ME_EPZS:
1239 {
1240 const int mot_stride = s->b8_stride;
1241 const int mot_xy = s->block_index[0];
1242
1243 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1244 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1245
1246 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1247
1248 if(!s->first_slice_line) {
1249 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1250 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1251 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1252 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1253 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1254 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1255 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1256
1257 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1258 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1259
1260 if(s->out_format == FMT_H263){
1261 c->pred_x = P_MEDIAN[0];
1262 c->pred_y = P_MEDIAN[1];
1263 }else { /* mpeg1 at least */
1264 c->pred_x= P_LEFT[0];
1265 c->pred_y= P_LEFT[1];
1266 }
1267 }else{
1268 c->pred_x= P_LEFT[0];
1269 c->pred_y= P_LEFT[1];
1270 }
1271
1272 }
1273 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1274
1275 break;
1276 }
1277
1278 /* At this point (mx,my) are full-pell and the relative displacement */
1279 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1280
1281 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1282
1283 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1284 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1285 c->mc_mb_var_sum_temp += (vard+128)>>8;
1286
1287 #if 0
1288 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1289 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1290 #endif
1291 if(mb_type){
1292 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1293 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1294 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1295
1296 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1297 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1298 set_p_mv_tables(s, mx, my, 1);
1299 }else{
1300 mx <<=shift;
1301 my <<=shift;
1302 }
1303 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1304 h263_mv4_search(s, mx, my, shift);
1305
1306 set_p_mv_tables(s, mx, my, 0);
1307 }
1308 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1309 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1310 }
1311 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1312 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1313 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1314 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1315
1316 if (vard*2 + 200*256 > varc)
1317 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1318 if (varc*2 + 200*256 > vard || s->qscale > 24){
1319 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1320 mb_type|= CANDIDATE_MB_TYPE_INTER;
1321 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1322 if(s->flags&CODEC_FLAG_MV0)
1323 if(mx || my)
1324 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1325 }else{
1326 mx <<=shift;
1327 my <<=shift;
1328 }
1329 if((s->flags&CODEC_FLAG_4MV)
1330 && !c->skip && varc>50<<8 && vard>10<<8){
1331 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1332 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1333
1334 set_p_mv_tables(s, mx, my, 0);
1335 }else
1336 set_p_mv_tables(s, mx, my, 1);
1337 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1338 && !c->skip){ //FIXME varc/d checks
1339 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1340 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1341 }
1342 }else{
1343 int intra_score, i;
1344 mb_type= CANDIDATE_MB_TYPE_INTER;
1345
1346 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1347 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1348 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1349
1350 if((s->flags&CODEC_FLAG_4MV)
1351 && !c->skip && varc>50<<8 && vard>10<<8){
1352 int dmin4= h263_mv4_search(s, mx, my, shift);
1353 if(dmin4 < dmin){
1354 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1355 dmin=dmin4;
1356 }
1357 }
1358 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1359 && !c->skip){ //FIXME varc/d checks
1360 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1361 if(dmin_i < dmin){
1362 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1363 dmin= dmin_i;
1364 }
1365 }
1366
1367 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1368 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1369
1370 /* get intra luma score */
1371 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1372 intra_score= varc - 500;
1373 }else{
1374 int mean= (sum+128)>>8;
1375 mean*= 0x01010101;
1376
1377 for(i=0; i<16; i++){
1378 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1379 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1380 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1381 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1382 }
1383
1384 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1385 }
1386 #if 0 //FIXME
1387 /* get chroma score */
1388 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1389 for(i=1; i<3; i++){
1390 uint8_t *dest_c;
1391 int mean;
1392
1393 if(s->out_format == FMT_H263){
1394 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1395 }else{
1396 mean= (s->last_dc[i] + 4)>>3;
1397 }
1398 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1399
1400 mean*= 0x01010101;
1401 for(i=0; i<8; i++){
1402 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1403 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1404 }
1405
1406 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1407 }
1408 }
1409 #endif
1410 intra_score += c->mb_penalty_factor*16;
1411
1412 if(intra_score < dmin){
1413 mb_type= CANDIDATE_MB_TYPE_INTRA;
1414 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1415 }else
1416 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1417
1418 {
1419 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1420 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1421 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1422 }
1423 }
1424
1425 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1426 }
1427
1428 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1429 int mb_x, int mb_y)
1430 {
1431 MotionEstContext * const c= &s->me;
1432 int mx, my, dmin;
1433 int P[10][2];
1434 const int shift= 1+s->quarter_sample;
1435 const int xy= mb_x + mb_y*s->mb_stride;
1436 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1437
1438 assert(s->quarter_sample==0 || s->quarter_sample==1);
1439
1440 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1441 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1442
1443 get_limits(s, 16*mb_x, 16*mb_y);
1444 c->skip=0;
1445
1446 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1447 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1448
1449 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1450
1451 /* special case for first line */
1452 if (s->first_slice_line) {
1453 c->pred_x= P_LEFT[0];
1454 c->pred_y= P_LEFT[1];
1455 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1456 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1457 } else {
1458 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1459 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1460 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1461 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1462 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1463 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1464 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1465
1466 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1467 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1468
1469 c->pred_x = P_MEDIAN[0];
1470 c->pred_y = P_MEDIAN[1];
1471 }
1472
1473 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1474
1475 s->p_mv_table[xy][0] = mx<<shift;
1476 s->p_mv_table[xy][1] = my<<shift;
1477
1478 return dmin;
1479 }
1480
1481 static int ff_estimate_motion_b(MpegEncContext * s,
1482 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1483 {
1484 MotionEstContext * const c= &s->me;
1485 int mx, my, dmin;
1486 int P[10][2];
1487 const int shift= 1+s->quarter_sample;
1488 const int mot_stride = s->mb_stride;
1489 const int mot_xy = mb_y*mot_stride + mb_x;
1490 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1491 int mv_scale;
1492
1493 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1494 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1495 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1496 c->current_mv_penalty= mv_penalty;
1497
1498 get_limits(s, 16*mb_x, 16*mb_y);
1499
1500 switch(s->me_method) {
1501 case ME_ZERO:
1502 default:
1503 no_motion_search(s, &mx, &my);
1504 dmin = 0;
1505 mx-= mb_x*16;
1506 my-= mb_y*16;
1507 break;
1508 #if 0
1509 case ME_FULL:
1510 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1511 mx-= mb_x*16;
1512 my-= mb_y*16;
1513 break;
1514 case ME_LOG:
1515 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1516 mx-= mb_x*16;
1517 my-= mb_y*16;
1518 break;
1519 case ME_PHODS:
1520 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1521 mx-= mb_x*16;
1522 my-= mb_y*16;
1523 break;
1524 #endif
1525 case ME_X1:
1526 case ME_EPZS:
1527 {
1528 P_LEFT[0] = mv_table[mot_xy - 1][0];
1529 P_LEFT[1] = mv_table[mot_xy - 1][1];
1530
1531 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1532
1533 /* special case for first line */
1534 if (!s->first_slice_line) {
1535 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1536 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1537 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1538 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1539 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1540 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1541 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1542
1543 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1544 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1545 }
1546 c->pred_x= P_LEFT[0];
1547 c->pred_y= P_LEFT[1];
1548 }
1549
1550 if(mv_table == s->b_forw_mv_table){
1551 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1552 }else{
1553 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1554 }
1555
1556 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1557
1558 break;
1559 }
1560
1561 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1562
1563 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1564 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1565
1566 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1567 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1568 mv_table[mot_xy][0]= mx;
1569 mv_table[mot_xy][1]= my;
1570
1571 return dmin;
1572 }
1573
1574 static inline int check_bidir_mv(MpegEncContext * s,
1575 int motion_fx, int motion_fy,
1576 int motion_bx, int motion_by,
1577 int pred_fx, int pred_fy,
1578 int pred_bx, int pred_by,
1579 int size, int h)
1580 {
1581 //FIXME optimize?
1582 //FIXME better f_code prediction (max mv & distance)
1583 //FIXME pointers
1584 MotionEstContext * const c= &s->me;
1585 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1586 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1587 int stride= c->stride;
1588 uint8_t *dest_y = c->scratchpad;
1589 uint8_t *ptr;
1590 int dxy;
1591 int src_x, src_y;
1592 int fbmin;
1593 uint8_t **src_data= c->src[0];
1594 uint8_t **ref_data= c->ref[0];
1595 uint8_t **ref2_data= c->ref[2];
1596
1597 if(s->quarter_sample){
1598 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1599 src_x = motion_fx >> 2;
1600 src_y = motion_fy >> 2;
1601
1602 ptr = ref_data[0] + (src_y * stride) + src_x;
1603 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1604
1605 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1606 src_x = motion_bx >> 2;
1607 src_y = motion_by >> 2;
1608
1609 ptr = ref2_data[0] + (src_y * stride) + src_x;
1610 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1611 }else{
1612 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1613 src_x = motion_fx >> 1;
1614 src_y = motion_fy >> 1;
1615
1616 ptr = ref_data[0] + (src_y * stride) + src_x;
1617 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1618
1619 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1620 src_x = motion_bx >> 1;
1621 src_y = motion_by >> 1;
1622
1623 ptr = ref2_data[0] + (src_y * stride) + src_x;
1624 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1625 }
1626
1627 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1628 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1629 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1630
1631 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1632 }
1633 //FIXME CHROMA !!!
1634
1635 return fbmin;
1636 }
1637
1638 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1639 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1640 {
1641 MotionEstContext * const c= &s->me;
1642 const int mot_stride = s->mb_stride;
1643 const int xy = mb_y *mot_stride + mb_x;
1644 int fbmin;
1645 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1646 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1647 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1648 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1649 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1650 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1651 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1652 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1653 const int flags= c->sub_flags;
1654 const int qpel= flags&FLAG_QPEL;
1655 const int shift= 1+qpel;
1656 const int xmin= c->xmin<<shift;
1657 const int ymin= c->ymin<<shift;
1658 const int xmax= c->xmax<<shift;
1659 const int ymax= c->ymax<<shift;
1660 uint8_t map[8][8][8][8];
1661
1662 memset(map,0,sizeof(map));
1663 #define BIDIR_MAP(fx,fy,bx,by) \
1664 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1665 BIDIR_MAP(0,0,0,0) = 1;
1666
1667 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1668 motion_bx, motion_by,
1669 pred_fx, pred_fy,
1670 pred_bx, pred_by,
1671 0, 16);
1672
1673 if(s->avctx->bidir_refine){
1674 int score, end;
1675 #define CHECK_BIDIR(fx,fy,bx,by)\
1676 if( !BIDIR_MAP(fx,fy,bx,by)\
1677 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1678 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1679 BIDIR_MAP(fx,fy,bx,by) = 1;\
1680 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);\
1681 if(score < fbmin){\
1682 fbmin= score;\
1683 motion_fx+=fx;\
1684 motion_fy+=fy;\
1685 motion_bx+=bx;\
1686 motion_by+=by;\
1687 end=0;\
1688 }\
1689 }
1690 #define CHECK_BIDIR2(a,b,c,d)\
1691 CHECK_BIDIR(a,b,c,d)\
1692 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1693
1694 #define CHECK_BIDIRR(a,b,c,d)\
1695 CHECK_BIDIR2(a,b,c,d)\
1696 CHECK_BIDIR2(b,c,d,a)\
1697 CHECK_BIDIR2(c,d,a,b)\
1698 CHECK_BIDIR2(d,a,b,c)
1699
1700 do{
1701 end=1;
1702
1703 CHECK_BIDIRR( 0, 0, 0, 1)
1704 if(s->avctx->bidir_refine > 1){
1705 CHECK_BIDIRR( 0, 0, 1, 1)
1706 CHECK_BIDIR2( 0, 1, 0, 1)
1707 CHECK_BIDIR2( 1, 0, 1, 0)
1708 CHECK_BIDIRR( 0, 0,-1, 1)
1709 CHECK_BIDIR2( 0,-1, 0, 1)
1710 CHECK_BIDIR2(-1, 0, 1, 0)
1711 if(s->avctx->bidir_refine > 2){
1712 CHECK_BIDIRR( 0, 1, 1, 1)
1713 CHECK_BIDIRR( 0,-1, 1, 1)
1714 CHECK_BIDIRR( 0, 1,-1, 1)
1715 CHECK_BIDIRR( 0, 1, 1,-1)
1716 if(s->avctx->bidir_refine > 3){
1717 CHECK_BIDIR2( 1, 1, 1, 1)
1718 CHECK_BIDIRR( 1, 1, 1,-1)
1719 CHECK_BIDIR2( 1, 1,-1,-1)
1720 CHECK_BIDIR2( 1,-1,-1, 1)
1721 CHECK_BIDIR2( 1,-1, 1,-1)
1722 }
1723 }
1724 }
1725 }while(!end);
1726 }
1727
1728 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1729 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1730 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1731 s->b_bidir_back_mv_table[xy][1]= motion_by;
1732
1733 return fbmin;
1734 }
1735
1736 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1737 {
1738 MotionEstContext * const c= &s->me;
1739 int P[10][2];
1740 const int mot_stride = s->mb_stride;
1741 const int mot_xy = mb_y*mot_stride + mb_x;
1742 const int shift= 1+s->quarter_sample;
1743 int dmin, i;
1744 const int time_pp= s->pp_time;
1745 const int time_pb= s->pb_time;
1746 int mx, my, xmin, xmax, ymin, ymax;
1747 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1748
1749 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1750 ymin= xmin=(-32)>>shift;
1751 ymax= xmax= 31>>shift;
1752
1753 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1754 s->mv_type= MV_TYPE_8X8;
1755 }else{
1756 s->mv_type= MV_TYPE_16X16;
1757 }
1758
1759 for(i=0; i<4; i++){
1760 int index= s->block_index[i];
1761 int min, max;
1762
1763 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1764 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1765 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1766 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1767 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1768 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1769
1770 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1771 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1772 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1773 min+= 16*mb_x - 1;
1774 xmax= FFMIN(xmax, s->width - max);
1775 xmin= FFMAX(xmin, - 16 - min);
1776
1777 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1778 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1779 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1780 min+= 16*mb_y - 1;
1781 ymax= FFMIN(ymax, s->height - max);
1782 ymin= FFMAX(ymin, - 16 - min);
1783
1784 if(s->mv_type == MV_TYPE_16X16) break;
1785 }
1786
1787 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1788
1789 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1790 s->b_direct_mv_table[mot_xy][0]= 0;
1791 s->b_direct_mv_table[mot_xy][1]= 0;
1792
1793 return 256*256*256*64;
1794 }
1795
1796 c->xmin= xmin;
1797 c->ymin= ymin;
1798 c->xmax= xmax;
1799 c->ymax= ymax;
1800 c->flags |= FLAG_DIRECT;
1801 c->sub_flags |= FLAG_DIRECT;
1802 c->pred_x=0;
1803 c->pred_y=0;
1804
1805 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1806 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1807
1808 /* special case for first line */
1809 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1810 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1811 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1812 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1813 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1814
1815 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1816 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1817 }
1818
1819 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1820 if(c->sub_flags&FLAG_QPEL)
1821 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1822 else
1823 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1824
1825 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1826 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1827
1828 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1829
1830 mv_table[mot_xy][0]= mx;
1831 mv_table[mot_xy][1]= my;
1832 c->flags &= ~FLAG_DIRECT;
1833 c->sub_flags &= ~FLAG_DIRECT;
1834
1835 return dmin;
1836 }
1837
1838 void ff_estimate_b_frame_motion(MpegEncContext * s,
1839 int mb_x, int mb_y)
1840 {
1841 MotionEstContext * const c= &s->me;
1842 const int penalty_factor= c->mb_penalty_factor;
1843 int fmin, bmin, dmin, fbmin, bimin, fimin;
1844 int type=0;
1845 const int xy = mb_y*s->mb_stride + mb_x;
1846 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1847
1848 get_limits(s, 16*mb_x, 16*mb_y);
1849
1850 c->skip=0;
1851
1852 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1853 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1854
1855 score= ((unsigned)(score*score + 128*256))>>16;
1856 c->mc_mb_var_sum_temp += score;
1857 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1858 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1859
1860 return;
1861 }
1862
1863 if(c->avctx->me_threshold){
1864 int vard= check_input_motion(s, mb_x, mb_y, 0);
1865
1866 if((vard+128)>>8 < c->avctx->me_threshold){
1867 // pix = c->src[0][0];
1868 // sum = s->dsp.pix_sum(pix, s->linesize);
1869 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1870
1871 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1872 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1873 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1874 c->mb_var_sum_temp += (varc+128)>>8;*/
1875 c->mc_mb_var_sum_temp += (vard+128)>>8;
1876 /* if (vard <= 64<<8 || vard < varc) {
1877 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1878 }else{
1879 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1880 }*/
1881 return;
1882 }
1883 if((vard+128)>>8 < c->avctx->mb_threshold){
1884 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1885 if(type == CANDIDATE_MB_TYPE_DIRECT){
1886 direct_search(s, mb_x, mb_y);
1887 }
1888 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1889 c->skip=0;
1890 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1891 }
1892 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1893 c->skip=0;
1894 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1895 }
1896 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1897 c->skip=0;
1898 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1899 interlaced_search(s, 0,
1900 s->b_field_mv_table[0], s->b_field_select_table[0],
1901 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1902 }
1903 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1904 c->skip=0;
1905 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1906 interlaced_search(s, 2,
1907 s->b_field_mv_table[1], s->b_field_select_table[1],
1908 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1909 }
1910 return;
1911 }
1912 }
1913
1914 if (s->codec_id == CODEC_ID_MPEG4)
1915 dmin= direct_search(s, mb_x, mb_y);
1916 else
1917 dmin= INT_MAX;
1918 //FIXME penalty stuff for non mpeg4
1919 c->skip=0;
1920 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1921
1922 c->skip=0;
1923 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1924 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1925
1926 c->skip=0;
1927 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1928 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1929
1930 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1931 //FIXME mb type penalty
1932 c->skip=0;
1933 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1934 fimin= interlaced_search(s, 0,
1935 s->b_field_mv_table[0], s->b_field_select_table[0],
1936 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1937 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1938 bimin= interlaced_search(s, 2,
1939 s->b_field_mv_table[1], s->b_field_select_table[1],
1940 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1941 }else
1942 fimin= bimin= INT_MAX;
1943
1944 {
1945 int score= fmin;
1946 type = CANDIDATE_MB_TYPE_FORWARD;
1947
1948 if (dmin <= score){
1949 score = dmin;
1950 type = CANDIDATE_MB_TYPE_DIRECT;
1951 }
1952 if(bmin<score){
1953 score=bmin;
1954 type= CANDIDATE_MB_TYPE_BACKWARD;
1955 }
1956 if(fbmin<score){
1957 score=fbmin;
1958 type= CANDIDATE_MB_TYPE_BIDIR;
1959 }
1960 if(fimin<score){
1961 score=fimin;
1962 type= CANDIDATE_MB_TYPE_FORWARD_I;
1963 }
1964 if(bimin<score){
1965 score=bimin;
1966 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1967 }
1968
1969 score= ((unsigned)(score*score + 128*256))>>16;
1970 c->mc_mb_var_sum_temp += score;
1971 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1972 }
1973
1974 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1975 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1976 if(fimin < INT_MAX)
1977 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1978 if(bimin < INT_MAX)
1979 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1980 if(fimin < INT_MAX && bimin < INT_MAX){
1981 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1982 }
1983 //FIXME something smarter
1984 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1985 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])
1986 type |= CANDIDATE_MB_TYPE_DIRECT0;
1987 #if 0
1988 if(s->out_format == FMT_MPEG1)
1989 type |= CANDIDATE_MB_TYPE_INTRA;
1990 #endif
1991 }
1992
1993 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1994 }
1995
1996 /* find best f_code for ME which do unlimited searches */
1997 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1998 {
1999 if(s->me_method>=ME_EPZS){
2000 int score[8];
2001 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
2002 uint8_t * fcode_tab= s->fcode_tab;
2003 int best_fcode=-1;
2004 int best_score=-10000000;
2005
2006 if(s->msmpeg4_version)
2007 range= FFMIN(range, 16);
2008 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2009 range= FFMIN(range, 256);
2010
2011 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
2012
2013 for(y=0; y<s->mb_height; y++){
2014 int x;
2015 int xy= y*s->mb_stride;
2016 for(x=0; x<s->mb_width; x++){
2017 if(s->mb_type[xy] & type){
2018 int mx= mv_table[xy][0];
2019 int my= mv_table[xy][1];
2020 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2021 fcode_tab[my + MAX_MV]);
2022 int j;
2023
2024 if(mx >= range || mx < -range ||
2025 my >= range || my < -range)
2026 continue;
2027
2028 for(j=0; j<fcode && j<8; j++){
2029 if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2030 score[j]-= 170;
2031 }
2032 }
2033 xy++;
2034 }
2035 }
2036
2037 for(i=1; i<8; i++){
2038 if(score[i] > best_score){
2039 best_score= score[i];
2040 best_fcode= i;
2041 }
2042 // printf("%d %d\n", i, score[i]);
2043 }
2044
2045 // printf("fcode: %d type: %d\n", i, s->pict_type);
2046 return best_fcode;
2047 /* for(i=0; i<=MAX_FCODE; i++){
2048 printf("%d ", mv_num[i]);
2049 }
2050 printf("\n");*/
2051 }else{
2052 return 1;
2053 }
2054 }
2055
2056 void ff_fix_long_p_mvs(MpegEncContext * s)
2057 {
2058 MotionEstContext * const c= &s->me;
2059 const int f_code= s->f_code;
2060 int y, range;
2061 assert(s->pict_type==FF_P_TYPE);
2062
2063 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2064
2065 assert(range <= 16 || !s->msmpeg4_version);
2066 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2067
2068 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2069
2070 //printf("%d no:%d %d//\n", clip, noclip, f_code);
2071 if(s->flags&CODEC_FLAG_4MV){
2072 const int wrap= s->b8_stride;
2073
2074 /* clip / convert to intra 8x8 type MVs */
2075 for(y=0; y<s->mb_height; y++){
2076 int xy= y*2*wrap;
2077 int i= y*s->mb_stride;
2078 int x;
2079
2080 for(x=0; x<s->mb_width; x++){
2081 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2082 int block;
2083 for(block=0; block<4; block++){
2084 int off= (block& 1) + (block>>1)*wrap;
2085 int mx= s->current_picture.motion_val[0][ xy + off ][0];
2086 int my= s->current_picture.motion_val[0][ xy + off ][1];
2087
2088 if( mx >=range || mx <-range
2089 || my >=range || my <-range){
2090 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2091 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2092 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2093 }
2094 }
2095 }
2096 xy+=2;
2097 i++;
2098 }
2099 }
2100 }
2101 }
2102
2103 /**
2104 *
2105 * @param truncate 1 for truncation, 0 for using intra
2106 */
2107 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2108 int16_t (*mv_table)[2], int f_code, int type, int truncate)
2109 {
2110 MotionEstContext * const c= &s->me;
2111 int y, h_range, v_range;
2112
2113 // RAL: 8 in MPEG-1, 16 in MPEG-4
2114 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2115
2116 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2117
2118 h_range= range;
2119 v_range= field_select_table ? range>>1 : range;
2120
2121 /* clip / convert to intra 16x16 type MVs */
2122 for(y=0; y<s->mb_height; y++){
2123 int x;
2124 int xy= y*s->mb_stride;
2125 for(x=0; x<s->mb_width; x++){
2126 if (s->mb_type[xy] & type){ // RAL: "type" test added...
2127 if(field_select_table==NULL || field_select_table[xy] == field_select){
2128 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2129 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2130
2131 if(truncate){
2132 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2133 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2134 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2135 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2136 }else{
2137 s->mb_type[xy] &= ~type;
2138 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2139 mv_table[xy][0]=
2140 mv_table[xy][1]= 0;
2141 }
2142 }
2143 }
2144 }
2145 xy++;
2146 }
2147 }
2148 }
lucabe's unfinished and experimental patches for ffmpeg