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