Libav
motion_est.c
Go to the documentation of this file.
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 Libav.
9  *
10  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24 
30 #include <stdlib.h>
31 #include <stdio.h>
32 #include <limits.h>
33 
34 #include "avcodec.h"
35 #include "mathops.h"
36 #include "mpegvideo.h"
37 
38 #undef NDEBUG
39 #include <assert.h>
40 
41 #define P_LEFT P[1]
42 #define P_TOP P[2]
43 #define P_TOPRIGHT P[3]
44 #define P_MEDIAN P[4]
45 #define P_MV1 P[9]
46 
48  int *mx_ptr, int *my_ptr, int dmin,
49  int src_index, int ref_index,
50  int size, int h);
51 
52 static inline unsigned update_map_generation(MotionEstContext *c)
53 {
54  c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
55  if(c->map_generation==0){
56  c->map_generation= 1<<(ME_MAP_MV_BITS*2);
57  memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
58  }
59  return c->map_generation;
60 }
61 
62 /* shape adaptive search stuff */
63 typedef struct Minima{
64  int height;
65  int x, y;
66  int checked;
67 }Minima;
68 
69 static int minima_cmp(const void *a, const void *b){
70  const Minima *da = (const Minima *) a;
71  const Minima *db = (const Minima *) b;
72 
73  return da->height - db->height;
74 }
75 
76 #define FLAG_QPEL 1 //must be 1
77 #define FLAG_CHROMA 2
78 #define FLAG_DIRECT 4
79 
80 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){
81  const int offset[3]= {
82  y*c-> stride + x,
83  ((y*c->uvstride + x)>>1),
84  ((y*c->uvstride + x)>>1),
85  };
86  int i;
87  for(i=0; i<3; i++){
88  c->src[0][i]= src [i] + offset[i];
89  c->ref[0][i]= ref [i] + offset[i];
90  }
91  if(ref_index){
92  for(i=0; i<3; i++){
93  c->ref[ref_index][i]= ref2[i] + offset[i];
94  }
95  }
96 }
97 
98 static int get_flags(MotionEstContext *c, int direct, int chroma){
99  return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
100  + (direct ? FLAG_DIRECT : 0)
101  + (chroma ? FLAG_CHROMA : 0);
102 }
103 
104 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
105  const int size, const int h, int ref_index, int src_index,
106  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
107  MotionEstContext * const c= &s->me;
108  const int stride= c->stride;
109  const int hx= subx + (x<<(1+qpel));
110  const int hy= suby + (y<<(1+qpel));
111  uint8_t * const * const ref= c->ref[ref_index];
112  uint8_t * const * const src= c->src[src_index];
113  int d;
114  //FIXME check chroma 4mv, (no crashes ...)
115  assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
116  if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
117  const int time_pp= s->pp_time;
118  const int time_pb= s->pb_time;
119  const int mask= 2*qpel+1;
120  if(s->mv_type==MV_TYPE_8X8){
121  int i;
122  for(i=0; i<4; i++){
123  int fx = c->direct_basis_mv[i][0] + hx;
124  int fy = c->direct_basis_mv[i][1] + hy;
125  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));
126  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));
127  int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
128  int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
129 
130  uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
131  if(qpel){
132  c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
133  c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
134  }else{
135  c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
136  c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
137  }
138  }
139  }else{
140  int fx = c->direct_basis_mv[0][0] + hx;
141  int fy = c->direct_basis_mv[0][1] + hy;
142  int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
143  int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
144  int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
145  int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
146 
147  if(qpel){
148  c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
149  c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
150  c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
151  c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
152  c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
153  c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
154  c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
155  c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
156  }else{
157  assert((fx>>1) + 16*s->mb_x >= -16);
158  assert((fy>>1) + 16*s->mb_y >= -16);
159  assert((fx>>1) + 16*s->mb_x <= s->width);
160  assert((fy>>1) + 16*s->mb_y <= s->height);
161  assert((bx>>1) + 16*s->mb_x >= -16);
162  assert((by>>1) + 16*s->mb_y >= -16);
163  assert((bx>>1) + 16*s->mb_x <= s->width);
164  assert((by>>1) + 16*s->mb_y <= s->height);
165 
166  c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
167  c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
168  }
169  }
170  d = cmp_func(s, c->temp, src[0], stride, 16);
171  }else
172  d= 256*256*256*32;
173  return d;
174 }
175 
176 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
177  const int size, const int h, int ref_index, int src_index,
178  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
179  MotionEstContext * const c= &s->me;
180  const int stride= c->stride;
181  const int uvstride= c->uvstride;
182  const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
183  const int hx= subx + (x<<(1+qpel));
184  const int hy= suby + (y<<(1+qpel));
185  uint8_t * const * const ref= c->ref[ref_index];
186  uint8_t * const * const src= c->src[src_index];
187  int d;
188  //FIXME check chroma 4mv, (no crashes ...)
189  int uvdxy; /* no, it might not be used uninitialized */
190  if(dxy){
191  if(qpel){
192  c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
193  if(chroma){
194  int cx= hx/2;
195  int cy= hy/2;
196  cx= (cx>>1)|(cx&1);
197  cy= (cy>>1)|(cy&1);
198  uvdxy= (cx&1) + 2*(cy&1);
199  //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
200  }
201  }else{
202  c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
203  if(chroma)
204  uvdxy= dxy | (x&1) | (2*(y&1));
205  }
206  d = cmp_func(s, c->temp, src[0], stride, h);
207  }else{
208  d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
209  if(chroma)
210  uvdxy= (x&1) + 2*(y&1);
211  }
212  if(chroma){
213  uint8_t * const uvtemp= c->temp + 16*stride;
214  c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
215  c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
216  d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
217  d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
218  }
219  return d;
220 }
221 
222 static int cmp_simple(MpegEncContext *s, const int x, const int y,
223  int ref_index, int src_index,
224  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
225  return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
226 }
227 
228 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
229  const int size, const int h, int ref_index, int src_index,
230  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
231  if(flags&FLAG_DIRECT){
232  return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
233  }else{
234  return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
235  }
236 }
237 
238 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
239  const int size, const int h, int ref_index, int src_index,
240  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
241  if(flags&FLAG_DIRECT){
242  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
243  }else{
244  return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA);
245  }
246 }
247 
251 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
252  const int size, const int h, int ref_index, int src_index,
253  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
256  && flags==0 && h==16 && size==0 && subx==0 && suby==0){
257  return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
258  }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
259  && subx==0 && suby==0){
260  return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
261  }else{
262  return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
263  }
264 }
265 
266 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
267  const int size, const int h, int ref_index, int src_index,
268  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
269  if(flags&FLAG_DIRECT){
270  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
271  }else{
272  return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
273  }
274 }
275 
276 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
277  const int size, const int h, int ref_index, int src_index,
278  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
279  if(flags&FLAG_DIRECT){
280  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
281  }else{
282  return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
283  }
284 }
285 
286 #include "motion_est_template.c"
287 
288 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
289  return 0;
290 }
291 
292 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
293 }
294 
296  MotionEstContext * const c= &s->me;
297  int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
298  int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
299 
301  av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
302  return -1;
303  }
304  if (s->me_method != ME_ZERO &&
305  s->me_method != ME_EPZS &&
306  s->me_method != ME_X1) {
307  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");
308  return -1;
309  }
310 
311  c->avctx= s->avctx;
312 
313  if(cache_size < 2*dia_size && !c->stride){
314  av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
315  }
316 
318  ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
320  ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
321 
322  c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
324  c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
325 
326 /*FIXME s->no_rounding b_type*/
327  if(s->flags&CODEC_FLAG_QPEL){
331  else c->qpel_put= s->dsp.put_qpel_pixels_tab;
332  }else{
335  else if( c->avctx->me_sub_cmp == FF_CMP_SAD
336  && c->avctx-> me_cmp == FF_CMP_SAD
337  && c->avctx-> mb_cmp == FF_CMP_SAD)
338  c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
339  else
341  }
342  c->hpel_avg = s->hdsp.avg_pixels_tab;
343  if (s->no_rounding)
345  else
346  c->hpel_put = s->hdsp.put_pixels_tab;
347 
348  if(s->linesize){
349  c->stride = s->linesize;
350  c->uvstride= s->uvlinesize;
351  }else{
352  c->stride = 16*s->mb_width + 32;
353  c->uvstride= 8*s->mb_width + 16;
354  }
355 
356  /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
357  * not have yet, and even if we had, the motion estimation code
358  * does not expect it. */
359  if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
360  s->dsp.me_cmp[2]= zero_cmp;
361  }
362  if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
363  s->dsp.me_sub_cmp[2]= zero_cmp;
364  }
365  c->hpel_put[2][0]= c->hpel_put[2][1]=
366  c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
367 
368  if(s->codec_id == AV_CODEC_ID_H261){
370  }
371 
372  return 0;
373 }
374 
375 #define CHECK_SAD_HALF_MV(suffix, x, y) \
376 {\
377  d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
378  d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
379  COPY3_IF_LT(dminh, d, dx, x, dy, y)\
380 }
381 
383  int *mx_ptr, int *my_ptr, int dmin,
384  int src_index, int ref_index,
385  int size, int h)
386 {
387  MotionEstContext * const c= &s->me;
388  const int penalty_factor= c->sub_penalty_factor;
389  int mx, my, dminh;
390  uint8_t *pix, *ptr;
391  int stride= c->stride;
392  const int flags= c->sub_flags;
394 
395  assert(flags == 0);
396 
397  if(c->skip){
398  *mx_ptr = 0;
399  *my_ptr = 0;
400  return dmin;
401  }
402 
403  pix = c->src[src_index][0];
404 
405  mx = *mx_ptr;
406  my = *my_ptr;
407  ptr = c->ref[ref_index][0] + (my * stride) + mx;
408 
409  dminh = dmin;
410 
411  if (mx > xmin && mx < xmax &&
412  my > ymin && my < ymax) {
413  int dx=0, dy=0;
414  int d, pen_x, pen_y;
415  const int index= (my<<ME_MAP_SHIFT) + mx;
416  const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
417  const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
418  const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
419  const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
420  mx<<=1;
421  my<<=1;
422 
423 
424  pen_x= pred_x + mx;
425  pen_y= pred_y + my;
426 
427  ptr-= stride;
428  if(t<=b){
429  CHECK_SAD_HALF_MV(y2 , 0, -1)
430  if(l<=r){
431  CHECK_SAD_HALF_MV(xy2, -1, -1)
432  if(t+r<=b+l){
433  CHECK_SAD_HALF_MV(xy2, +1, -1)
434  ptr+= stride;
435  }else{
436  ptr+= stride;
437  CHECK_SAD_HALF_MV(xy2, -1, +1)
438  }
439  CHECK_SAD_HALF_MV(x2 , -1, 0)
440  }else{
441  CHECK_SAD_HALF_MV(xy2, +1, -1)
442  if(t+l<=b+r){
443  CHECK_SAD_HALF_MV(xy2, -1, -1)
444  ptr+= stride;
445  }else{
446  ptr+= stride;
447  CHECK_SAD_HALF_MV(xy2, +1, +1)
448  }
449  CHECK_SAD_HALF_MV(x2 , +1, 0)
450  }
451  }else{
452  if(l<=r){
453  if(t+l<=b+r){
454  CHECK_SAD_HALF_MV(xy2, -1, -1)
455  ptr+= stride;
456  }else{
457  ptr+= stride;
458  CHECK_SAD_HALF_MV(xy2, +1, +1)
459  }
460  CHECK_SAD_HALF_MV(x2 , -1, 0)
461  CHECK_SAD_HALF_MV(xy2, -1, +1)
462  }else{
463  if(t+r<=b+l){
464  CHECK_SAD_HALF_MV(xy2, +1, -1)
465  ptr+= stride;
466  }else{
467  ptr+= stride;
468  CHECK_SAD_HALF_MV(xy2, -1, +1)
469  }
470  CHECK_SAD_HALF_MV(x2 , +1, 0)
471  CHECK_SAD_HALF_MV(xy2, +1, +1)
472  }
473  CHECK_SAD_HALF_MV(y2 , 0, +1)
474  }
475  mx+=dx;
476  my+=dy;
477 
478  }else{
479  mx<<=1;
480  my<<=1;
481  }
482 
483  *mx_ptr = mx;
484  *my_ptr = my;
485  return dminh;
486 }
487 
488 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
489 {
490  const int xy= s->mb_x + s->mb_y*s->mb_stride;
491 
492  s->p_mv_table[xy][0] = mx;
493  s->p_mv_table[xy][1] = my;
494 
495  /* has already been set to the 4 MV if 4MV is done */
496  if(mv4){
497  int mot_xy= s->block_index[0];
498 
499  s->current_picture.motion_val[0][mot_xy ][0] = mx;
500  s->current_picture.motion_val[0][mot_xy ][1] = my;
501  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
502  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
503 
504  mot_xy += s->b8_stride;
505  s->current_picture.motion_val[0][mot_xy ][0] = mx;
506  s->current_picture.motion_val[0][mot_xy ][1] = my;
507  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
508  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
509  }
510 }
511 
515 static inline void get_limits(MpegEncContext *s, int x, int y)
516 {
517  MotionEstContext * const c= &s->me;
518  int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
519 /*
520  if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
521  else c->range= 16;
522 */
523  if (s->unrestricted_mv) {
524  c->xmin = - x - 16;
525  c->ymin = - y - 16;
526  c->xmax = - x + s->mb_width *16;
527  c->ymax = - y + s->mb_height*16;
528  } else if (s->out_format == FMT_H261){
529  // Search range of H261 is different from other codec standards
530  c->xmin = (x > 15) ? - 15 : 0;
531  c->ymin = (y > 15) ? - 15 : 0;
532  c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
533  c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
534  } else {
535  c->xmin = - x;
536  c->ymin = - y;
537  c->xmax = - x + s->mb_width *16 - 16;
538  c->ymax = - y + s->mb_height*16 - 16;
539  }
540  if(range){
541  c->xmin = FFMAX(c->xmin,-range);
542  c->xmax = FFMIN(c->xmax, range);
543  c->ymin = FFMAX(c->ymin,-range);
544  c->ymax = FFMIN(c->ymax, range);
545  }
546 }
547 
548 static inline void init_mv4_ref(MotionEstContext *c){
549  const int stride= c->stride;
550 
551  c->ref[1][0] = c->ref[0][0] + 8;
552  c->ref[2][0] = c->ref[0][0] + 8*stride;
553  c->ref[3][0] = c->ref[2][0] + 8;
554  c->src[1][0] = c->src[0][0] + 8;
555  c->src[2][0] = c->src[0][0] + 8*stride;
556  c->src[3][0] = c->src[2][0] + 8;
557 }
558 
559 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
560 {
561  MotionEstContext * const c= &s->me;
562  const int size= 1;
563  const int h=8;
564  int block;
565  int P[10][2];
566  int dmin_sum=0, mx4_sum=0, my4_sum=0;
567  int same=1;
568  const int stride= c->stride;
570 
571  init_mv4_ref(c);
572 
573  for(block=0; block<4; block++){
574  int mx4, my4;
575  int pred_x4, pred_y4;
576  int dmin4;
577  static const int off[4]= {2, 1, 1, -1};
578  const int mot_stride = s->b8_stride;
579  const int mot_xy = s->block_index[block];
580 
581  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
582  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
583 
584  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
585 
586  /* special case for first line */
587  if (s->first_slice_line && block<2) {
588  c->pred_x= pred_x4= P_LEFT[0];
589  c->pred_y= pred_y4= P_LEFT[1];
590  } else {
591  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
592  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
593  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
594  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
595  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
596  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
597  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
598  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
599 
600  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
601  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
602 
603  c->pred_x= pred_x4 = P_MEDIAN[0];
604  c->pred_y= pred_y4 = P_MEDIAN[1];
605  }
606  P_MV1[0]= mx;
607  P_MV1[1]= my;
608 
609  dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
610 
611  dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
612 
613  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
614  int dxy;
615  const int offset= ((block&1) + (block>>1)*stride)*8;
616  uint8_t *dest_y = c->scratchpad + offset;
617  if(s->quarter_sample){
618  uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
619  dxy = ((my4 & 3) << 2) | (mx4 & 3);
620 
621  if(s->no_rounding)
622  s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
623  else
624  s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
625  }else{
626  uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
627  dxy = ((my4 & 1) << 1) | (mx4 & 1);
628 
629  if(s->no_rounding)
630  s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
631  else
632  s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
633  }
634  dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
635  }else
636  dmin_sum+= dmin4;
637 
638  if(s->quarter_sample){
639  mx4_sum+= mx4/2;
640  my4_sum+= my4/2;
641  }else{
642  mx4_sum+= mx4;
643  my4_sum+= my4;
644  }
645 
646  s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
647  s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
648 
649  if(mx4 != mx || my4 != my) same=0;
650  }
651 
652  if(same)
653  return INT_MAX;
654 
655  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
656  dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
657  }
658 
659  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
660  int dxy;
661  int mx, my;
662  int offset;
663 
664  mx= ff_h263_round_chroma(mx4_sum);
665  my= ff_h263_round_chroma(my4_sum);
666  dxy = ((my & 1) << 1) | (mx & 1);
667 
668  offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
669 
670  if(s->no_rounding){
671  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
672  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
673  }else{
674  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
675  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
676  }
677 
678  dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.f.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
679  dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.f.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
680  }
681 
682  c->pred_x= mx;
683  c->pred_y= my;
684 
685  switch(c->avctx->mb_cmp&0xFF){
686  /*case FF_CMP_SSE:
687  return dmin_sum+ 32*s->qscale*s->qscale;*/
688  case FF_CMP_RD:
689  return dmin_sum;
690  default:
691  return dmin_sum+ 11*c->mb_penalty_factor;
692  }
693 }
694 
695 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
696  MotionEstContext * const c= &s->me;
697 
698  c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
699  c->src[1][0] = c->src[0][0] + s->linesize;
700  if(c->flags & FLAG_CHROMA){
701  c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
702  c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
703  c->src[1][1] = c->src[0][1] + s->uvlinesize;
704  c->src[1][2] = c->src[0][2] + s->uvlinesize;
705  }
706 }
707 
708 static int interlaced_search(MpegEncContext *s, int ref_index,
709  int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
710 {
711  MotionEstContext * const c= &s->me;
712  const int size=0;
713  const int h=8;
714  int block;
715  int P[10][2];
717  int same=1;
718  const int stride= 2*s->linesize;
719  int dmin_sum= 0;
720  const int mot_stride= s->mb_stride;
721  const int xy= s->mb_x + s->mb_y*mot_stride;
722 
723  c->ymin>>=1;
724  c->ymax>>=1;
725  c->stride<<=1;
726  c->uvstride<<=1;
727  init_interlaced_ref(s, ref_index);
728 
729  for(block=0; block<2; block++){
730  int field_select;
731  int best_dmin= INT_MAX;
732  int best_field= -1;
733 
734  for(field_select=0; field_select<2; field_select++){
735  int dmin, mx_i, my_i;
736  int16_t (*mv_table)[2]= mv_tables[block][field_select];
737 
738  if(user_field_select){
739  assert(field_select==0 || field_select==1);
740  assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
741  if(field_select_tables[block][xy] != field_select)
742  continue;
743  }
744 
745  P_LEFT[0] = mv_table[xy - 1][0];
746  P_LEFT[1] = mv_table[xy - 1][1];
747  if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
748 
749  c->pred_x= P_LEFT[0];
750  c->pred_y= P_LEFT[1];
751 
752  if(!s->first_slice_line){
753  P_TOP[0] = mv_table[xy - mot_stride][0];
754  P_TOP[1] = mv_table[xy - mot_stride][1];
755  P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
756  P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
757  if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
758  if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
759  if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
760  if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
761 
762  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
763  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
764  }
765  P_MV1[0]= mx; //FIXME not correct if block != field_select
766  P_MV1[1]= my / 2;
767 
768  dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
769 
770  dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
771 
772  mv_table[xy][0]= mx_i;
773  mv_table[xy][1]= my_i;
774 
775  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
776  int dxy;
777 
778  //FIXME chroma ME
779  uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
780  dxy = ((my_i & 1) << 1) | (mx_i & 1);
781 
782  if(s->no_rounding){
783  s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
784  }else{
785  s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
786  }
787  dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
788  dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
789  }else
790  dmin+= c->mb_penalty_factor; //field_select bits
791 
792  dmin += field_select != block; //slightly prefer same field
793 
794  if(dmin < best_dmin){
795  best_dmin= dmin;
796  best_field= field_select;
797  }
798  }
799  {
800  int16_t (*mv_table)[2]= mv_tables[block][best_field];
801 
802  if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
803  if(mv_table[xy][1]&1) same=0;
804  if(mv_table[xy][1]*2 != my) same=0;
805  if(best_field != block) same=0;
806  }
807 
808  field_select_tables[block][xy]= best_field;
809  dmin_sum += best_dmin;
810  }
811 
812  c->ymin<<=1;
813  c->ymax<<=1;
814  c->stride>>=1;
815  c->uvstride>>=1;
816 
817  if(same)
818  return INT_MAX;
819 
820  switch(c->avctx->mb_cmp&0xFF){
821  /*case FF_CMP_SSE:
822  return dmin_sum+ 32*s->qscale*s->qscale;*/
823  case FF_CMP_RD:
824  return dmin_sum;
825  default:
826  return dmin_sum+ 11*c->mb_penalty_factor;
827  }
828 }
829 
830 static inline int get_penalty_factor(int lambda, int lambda2, int type){
831  switch(type&0xFF){
832  default:
833  case FF_CMP_SAD:
834  return lambda>>FF_LAMBDA_SHIFT;
835  case FF_CMP_DCT:
836  return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
837  case FF_CMP_SATD:
838  case FF_CMP_DCT264:
839  return (2*lambda)>>FF_LAMBDA_SHIFT;
840  case FF_CMP_RD:
841  case FF_CMP_PSNR:
842  case FF_CMP_SSE:
843  case FF_CMP_NSSE:
844  return lambda2>>FF_LAMBDA_SHIFT;
845  case FF_CMP_BIT:
846  return 1;
847  }
848 }
849 
851  int mb_x, int mb_y)
852 {
853  MotionEstContext * const c= &s->me;
854  uint8_t *pix, *ppix;
855  int sum, mx, my, dmin;
856  int varc;
857  int vard;
858  int P[10][2];
859  const int shift= 1+s->quarter_sample;
860  int mb_type=0;
861  Picture * const pic= &s->current_picture;
862 
863  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
864 
865  assert(s->quarter_sample==0 || s->quarter_sample==1);
866  assert(s->linesize == c->stride);
867  assert(s->uvlinesize == c->uvstride);
868 
873 
874  get_limits(s, 16*mb_x, 16*mb_y);
875  c->skip=0;
876 
877  /* intra / predictive decision */
878  pix = c->src[0][0];
879  sum = s->dsp.pix_sum(pix, s->linesize);
880  varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
881 
882  pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
883  pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
884  c->mb_var_sum_temp += (varc+128)>>8;
885 
886  switch(s->me_method) {
887  case ME_ZERO:
888  default:
889  mx = 0;
890  my = 0;
891  dmin = 0;
892  break;
893  case ME_X1:
894  case ME_EPZS:
895  {
896  const int mot_stride = s->b8_stride;
897  const int mot_xy = s->block_index[0];
898 
899  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
900  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
901 
902  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
903 
904  if(!s->first_slice_line) {
905  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
906  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
907  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
908  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
909  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
910  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
911  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
912 
913  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
914  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
915 
916  if(s->out_format == FMT_H263){
917  c->pred_x = P_MEDIAN[0];
918  c->pred_y = P_MEDIAN[1];
919  }else { /* mpeg1 at least */
920  c->pred_x= P_LEFT[0];
921  c->pred_y= P_LEFT[1];
922  }
923  }else{
924  c->pred_x= P_LEFT[0];
925  c->pred_y= P_LEFT[1];
926  }
927 
928  }
929  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
930 
931  break;
932  }
933 
934  /* At this point (mx,my) are full-pell and the relative displacement */
935  ppix = c->ref[0][0] + (my * s->linesize) + mx;
936 
937  vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
938 
939  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
940  c->mc_mb_var_sum_temp += (vard+128)>>8;
941 
942  if(mb_type){
943  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
944  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
945  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
946 
947  if(mb_type == CANDIDATE_MB_TYPE_INTER){
948  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
949  set_p_mv_tables(s, mx, my, 1);
950  }else{
951  mx <<=shift;
952  my <<=shift;
953  }
954  if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
955  h263_mv4_search(s, mx, my, shift);
956 
957  set_p_mv_tables(s, mx, my, 0);
958  }
959  if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
961  }
962  }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
963  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
964  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
965  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
966 
967  if (vard*2 + 200*256 > varc)
968  mb_type|= CANDIDATE_MB_TYPE_INTRA;
969  if (varc*2 + 200*256 > vard || s->qscale > 24){
970 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
971  mb_type|= CANDIDATE_MB_TYPE_INTER;
972  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
973  if(s->flags&CODEC_FLAG_MV0)
974  if(mx || my)
975  mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
976  }else{
977  mx <<=shift;
978  my <<=shift;
979  }
980  if((s->flags&CODEC_FLAG_4MV)
981  && !c->skip && varc>50<<8 && vard>10<<8){
982  if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
983  mb_type|=CANDIDATE_MB_TYPE_INTER4V;
984 
985  set_p_mv_tables(s, mx, my, 0);
986  }else
987  set_p_mv_tables(s, mx, my, 1);
989  && !c->skip){ //FIXME varc/d checks
990  if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
991  mb_type |= CANDIDATE_MB_TYPE_INTER_I;
992  }
993  }else{
994  int intra_score, i;
995  mb_type= CANDIDATE_MB_TYPE_INTER;
996 
997  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
998  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
999  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1000 
1001  if((s->flags&CODEC_FLAG_4MV)
1002  && !c->skip && varc>50<<8 && vard>10<<8){
1003  int dmin4= h263_mv4_search(s, mx, my, shift);
1004  if(dmin4 < dmin){
1005  mb_type= CANDIDATE_MB_TYPE_INTER4V;
1006  dmin=dmin4;
1007  }
1008  }
1010  && !c->skip){ //FIXME varc/d checks
1011  int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1012  if(dmin_i < dmin){
1013  mb_type = CANDIDATE_MB_TYPE_INTER_I;
1014  dmin= dmin_i;
1015  }
1016  }
1017 
1018  set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1019 
1020  /* get intra luma score */
1021  if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1022  intra_score= varc - 500;
1023  }else{
1024  unsigned mean = (sum+128)>>8;
1025  mean*= 0x01010101;
1026 
1027  for(i=0; i<16; i++){
1028  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1029  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1030  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1031  *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1032  }
1033 
1034  intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1035  }
1036  intra_score += c->mb_penalty_factor*16;
1037 
1038  if(intra_score < dmin){
1039  mb_type= CANDIDATE_MB_TYPE_INTRA;
1040  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1041  }else
1042  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1043 
1044  {
1045  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1046  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1047  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1048  }
1049  }
1050 
1051  s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1052 }
1053 
1055  int mb_x, int mb_y)
1056 {
1057  MotionEstContext * const c= &s->me;
1058  int mx, my, dmin;
1059  int P[10][2];
1060  const int shift= 1+s->quarter_sample;
1061  const int xy= mb_x + mb_y*s->mb_stride;
1062  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1063 
1064  assert(s->quarter_sample==0 || s->quarter_sample==1);
1065 
1068 
1069  get_limits(s, 16*mb_x, 16*mb_y);
1070  c->skip=0;
1071 
1072  P_LEFT[0] = s->p_mv_table[xy + 1][0];
1073  P_LEFT[1] = s->p_mv_table[xy + 1][1];
1074 
1075  if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1076 
1077  /* special case for first line */
1078  if (s->first_slice_line) {
1079  c->pred_x= P_LEFT[0];
1080  c->pred_y= P_LEFT[1];
1081  P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1082  P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1083  } else {
1084  P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1085  P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1086  P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1087  P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1088  if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1089  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1090  if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1091 
1092  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1093  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1094 
1095  c->pred_x = P_MEDIAN[0];
1096  c->pred_y = P_MEDIAN[1];
1097  }
1098 
1099  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1100 
1101  s->p_mv_table[xy][0] = mx<<shift;
1102  s->p_mv_table[xy][1] = my<<shift;
1103 
1104  return dmin;
1105 }
1106 
1107 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1108  int16_t (*mv_table)[2], int ref_index, int f_code)
1109 {
1110  MotionEstContext * const c= &s->me;
1111  int mx, my, dmin;
1112  int P[10][2];
1113  const int shift= 1+s->quarter_sample;
1114  const int mot_stride = s->mb_stride;
1115  const int mot_xy = mb_y*mot_stride + mb_x;
1116  uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1117  int mv_scale;
1118 
1123 
1124  get_limits(s, 16*mb_x, 16*mb_y);
1125 
1126  switch(s->me_method) {
1127  case ME_ZERO:
1128  default:
1129  mx = 0;
1130  my = 0;
1131  dmin = 0;
1132  break;
1133  case ME_X1:
1134  case ME_EPZS:
1135  P_LEFT[0] = mv_table[mot_xy - 1][0];
1136  P_LEFT[1] = mv_table[mot_xy - 1][1];
1137 
1138  if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1139 
1140  /* special case for first line */
1141  if (!s->first_slice_line) {
1142  P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1143  P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1144  P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1145  P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1146  if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1147  if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1148  if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1149 
1150  P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1151  P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1152  }
1153  c->pred_x = P_LEFT[0];
1154  c->pred_y = P_LEFT[1];
1155 
1156  if(mv_table == s->b_forw_mv_table){
1157  mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1158  }else{
1159  mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1160  }
1161 
1162  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1163 
1164  break;
1165  }
1166 
1167  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1168 
1169  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1170  dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1171 
1172 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1173  mv_table[mot_xy][0]= mx;
1174  mv_table[mot_xy][1]= my;
1175 
1176  return dmin;
1177 }
1178 
1179 static inline int check_bidir_mv(MpegEncContext * s,
1180  int motion_fx, int motion_fy,
1181  int motion_bx, int motion_by,
1182  int pred_fx, int pred_fy,
1183  int pred_bx, int pred_by,
1184  int size, int h)
1185 {
1186  //FIXME optimize?
1187  //FIXME better f_code prediction (max mv & distance)
1188  //FIXME pointers
1189  MotionEstContext * const c= &s->me;
1190  uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1191  uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1192  int stride= c->stride;
1193  uint8_t *dest_y = c->scratchpad;
1194  uint8_t *ptr;
1195  int dxy;
1196  int src_x, src_y;
1197  int fbmin;
1198  uint8_t **src_data= c->src[0];
1199  uint8_t **ref_data= c->ref[0];
1200  uint8_t **ref2_data= c->ref[2];
1201 
1202  if(s->quarter_sample){
1203  dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1204  src_x = motion_fx >> 2;
1205  src_y = motion_fy >> 2;
1206 
1207  ptr = ref_data[0] + (src_y * stride) + src_x;
1208  s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1209 
1210  dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1211  src_x = motion_bx >> 2;
1212  src_y = motion_by >> 2;
1213 
1214  ptr = ref2_data[0] + (src_y * stride) + src_x;
1215  s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1216  }else{
1217  dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1218  src_x = motion_fx >> 1;
1219  src_y = motion_fy >> 1;
1220 
1221  ptr = ref_data[0] + (src_y * stride) + src_x;
1222  s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1223 
1224  dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1225  src_x = motion_bx >> 1;
1226  src_y = motion_by >> 1;
1227 
1228  ptr = ref2_data[0] + (src_y * stride) + src_x;
1229  s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1230  }
1231 
1232  fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1233  +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1234  + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1235 
1236  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1237  }
1238  //FIXME CHROMA !!!
1239 
1240  return fbmin;
1241 }
1242 
1243 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1244 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1245 {
1246  MotionEstContext * const c= &s->me;
1247  const int mot_stride = s->mb_stride;
1248  const int xy = mb_y *mot_stride + mb_x;
1249  int fbmin;
1250  int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1251  int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1252  int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1253  int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1254  int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1255  int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1256  int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1257  int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1258  const int flags= c->sub_flags;
1259  const int qpel= flags&FLAG_QPEL;
1260  const int shift= 1+qpel;
1261  const int xmin= c->xmin<<shift;
1262  const int ymin= c->ymin<<shift;
1263  const int xmax= c->xmax<<shift;
1264  const int ymax= c->ymax<<shift;
1265 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1266 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1267  int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1268  uint8_t map[256] = { 0 };
1269 
1270  map[hashidx&255] = 1;
1271 
1272  fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1273  motion_bx, motion_by,
1274  pred_fx, pred_fy,
1275  pred_bx, pred_by,
1276  0, 16);
1277 
1278  if(s->avctx->bidir_refine){
1279  int end;
1280  static const uint8_t limittab[5]={0,8,32,64,80};
1281  const int limit= limittab[s->avctx->bidir_refine];
1282  static const int8_t vect[][4]={
1283 { 0, 0, 0, 1}, { 0, 0, 0,-1}, { 0, 0, 1, 0}, { 0, 0,-1, 0}, { 0, 1, 0, 0}, { 0,-1, 0, 0}, { 1, 0, 0, 0}, {-1, 0, 0, 0},
1284 
1285 { 0, 0, 1, 1}, { 0, 0,-1,-1}, { 0, 1, 1, 0}, { 0,-1,-1, 0}, { 1, 1, 0, 0}, {-1,-1, 0, 0}, { 1, 0, 0, 1}, {-1, 0, 0,-1},
1286 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1287 { 0, 0,-1, 1}, { 0, 0, 1,-1}, { 0,-1, 1, 0}, { 0, 1,-1, 0}, {-1, 1, 0, 0}, { 1,-1, 0, 0}, { 1, 0, 0,-1}, {-1, 0, 0, 1},
1288 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1289 
1290 { 0, 1, 1, 1}, { 0,-1,-1,-1}, { 1, 1, 1, 0}, {-1,-1,-1, 0}, { 1, 1, 0, 1}, {-1,-1, 0,-1}, { 1, 0, 1, 1}, {-1, 0,-1,-1},
1291 { 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1},
1292 { 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1},
1293 { 0, 1, 1,-1}, { 0,-1,-1, 1}, { 1, 1,-1, 0}, {-1,-1, 1, 0}, { 1,-1, 0, 1}, {-1, 1, 0,-1}, {-1, 0, 1, 1}, { 1, 0,-1,-1},
1294 
1295 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1296 { 1, 1, 1,-1}, {-1,-1,-1, 1}, { 1, 1,-1, 1}, {-1,-1, 1,-1}, { 1,-1, 1, 1}, {-1, 1,-1,-1}, {-1, 1, 1, 1}, { 1,-1,-1,-1},
1297 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1298  };
1299  static const uint8_t hash[]={
1300 HASH8( 0, 0, 0, 1), HASH8( 0, 0, 0,-1), HASH8( 0, 0, 1, 0), HASH8( 0, 0,-1, 0), HASH8( 0, 1, 0, 0), HASH8( 0,-1, 0, 0), HASH8( 1, 0, 0, 0), HASH8(-1, 0, 0, 0),
1301 
1302 HASH8( 0, 0, 1, 1), HASH8( 0, 0,-1,-1), HASH8( 0, 1, 1, 0), HASH8( 0,-1,-1, 0), HASH8( 1, 1, 0, 0), HASH8(-1,-1, 0, 0), HASH8( 1, 0, 0, 1), HASH8(-1, 0, 0,-1),
1303 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1304 HASH8( 0, 0,-1, 1), HASH8( 0, 0, 1,-1), HASH8( 0,-1, 1, 0), HASH8( 0, 1,-1, 0), HASH8(-1, 1, 0, 0), HASH8( 1,-1, 0, 0), HASH8( 1, 0, 0,-1), HASH8(-1, 0, 0, 1),
1305 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1306 
1307 HASH8( 0, 1, 1, 1), HASH8( 0,-1,-1,-1), HASH8( 1, 1, 1, 0), HASH8(-1,-1,-1, 0), HASH8( 1, 1, 0, 1), HASH8(-1,-1, 0,-1), HASH8( 1, 0, 1, 1), HASH8(-1, 0,-1,-1),
1308 HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1),
1309 HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1),
1310 HASH8( 0, 1, 1,-1), HASH8( 0,-1,-1, 1), HASH8( 1, 1,-1, 0), HASH8(-1,-1, 1, 0), HASH8( 1,-1, 0, 1), HASH8(-1, 1, 0,-1), HASH8(-1, 0, 1, 1), HASH8( 1, 0,-1,-1),
1311 
1312 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1313 HASH8( 1, 1, 1,-1), HASH8(-1,-1,-1, 1), HASH8( 1, 1,-1, 1), HASH8(-1,-1, 1,-1), HASH8( 1,-1, 1, 1), HASH8(-1, 1,-1,-1), HASH8(-1, 1, 1, 1), HASH8( 1,-1,-1,-1),
1314 HASH8( 1, 1,-1,-1), HASH8(-1,-1, 1, 1), HASH8( 1,-1,-1, 1), HASH8(-1, 1, 1,-1), HASH8( 1,-1, 1,-1), HASH8(-1, 1,-1, 1),
1315 };
1316 
1317 #define CHECK_BIDIR(fx,fy,bx,by)\
1318  if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1319  &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1320  &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1321  int score;\
1322  map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1323  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);\
1324  if(score < fbmin){\
1325  hashidx += HASH(fx,fy,bx,by);\
1326  fbmin= score;\
1327  motion_fx+=fx;\
1328  motion_fy+=fy;\
1329  motion_bx+=bx;\
1330  motion_by+=by;\
1331  end=0;\
1332  }\
1333  }
1334 #define CHECK_BIDIR2(a,b,c,d)\
1335 CHECK_BIDIR(a,b,c,d)\
1336 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1337 
1338  do{
1339  int i;
1340  int borderdist=0;
1341  end=1;
1342 
1343  CHECK_BIDIR2(0,0,0,1)
1344  CHECK_BIDIR2(0,0,1,0)
1345  CHECK_BIDIR2(0,1,0,0)
1346  CHECK_BIDIR2(1,0,0,0)
1347 
1348  for(i=8; i<limit; i++){
1349  int fx= motion_fx+vect[i][0];
1350  int fy= motion_fy+vect[i][1];
1351  int bx= motion_bx+vect[i][2];
1352  int by= motion_by+vect[i][3];
1353  if(borderdist<=0){
1354  int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1355  int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1356  if((a|b) < 0)
1357  map[(hashidx+hash[i])&255] = 1;
1358  }
1359  if(!map[(hashidx+hash[i])&255]){
1360  int score;
1361  map[(hashidx+hash[i])&255] = 1;
1362  score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1363  if(score < fbmin){
1364  hashidx += hash[i];
1365  fbmin= score;
1366  motion_fx=fx;
1367  motion_fy=fy;
1368  motion_bx=bx;
1369  motion_by=by;
1370  end=0;
1371  borderdist--;
1372  if(borderdist<=0){
1373  int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1374  int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1375  borderdist= FFMIN(a,b);
1376  }
1377  }
1378  }
1379  }
1380  }while(!end);
1381  }
1382 
1383  s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1384  s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1385  s->b_bidir_back_mv_table[xy][0]= motion_bx;
1386  s->b_bidir_back_mv_table[xy][1]= motion_by;
1387 
1388  return fbmin;
1389 }
1390 
1391 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1392 {
1393  MotionEstContext * const c= &s->me;
1394  int P[10][2];
1395  const int mot_stride = s->mb_stride;
1396  const int mot_xy = mb_y*mot_stride + mb_x;
1397  const int shift= 1+s->quarter_sample;
1398  int dmin, i;
1399  const int time_pp= s->pp_time;
1400  const int time_pb= s->pb_time;
1401  int mx, my, xmin, xmax, ymin, ymax;
1402  int16_t (*mv_table)[2]= s->b_direct_mv_table;
1403 
1404  c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1405  ymin= xmin=(-32)>>shift;
1406  ymax= xmax= 31>>shift;
1407 
1408  if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1409  s->mv_type= MV_TYPE_8X8;
1410  }else{
1411  s->mv_type= MV_TYPE_16X16;
1412  }
1413 
1414  for(i=0; i<4; i++){
1415  int index= s->block_index[i];
1416  int min, max;
1417 
1418  c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1419  c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1420  c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1421  c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1422 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1423 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1424 
1425  max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1426  min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1427  max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1428  min+= 16*mb_x - 1;
1429  xmax= FFMIN(xmax, s->width - max);
1430  xmin= FFMAX(xmin, - 16 - min);
1431 
1432  max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1433  min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1434  max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1435  min+= 16*mb_y - 1;
1436  ymax= FFMIN(ymax, s->height - max);
1437  ymin= FFMAX(ymin, - 16 - min);
1438 
1439  if(s->mv_type == MV_TYPE_16X16) break;
1440  }
1441 
1442  assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1443 
1444  if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1445  s->b_direct_mv_table[mot_xy][0]= 0;
1446  s->b_direct_mv_table[mot_xy][1]= 0;
1447 
1448  return 256*256*256*64;
1449  }
1450 
1451  c->xmin= xmin;
1452  c->ymin= ymin;
1453  c->xmax= xmax;
1454  c->ymax= ymax;
1455  c->flags |= FLAG_DIRECT;
1456  c->sub_flags |= FLAG_DIRECT;
1457  c->pred_x=0;
1458  c->pred_y=0;
1459 
1460  P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1461  P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1462 
1463  /* special case for first line */
1464  if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1465  P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1466  P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1467  P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1468  P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1469 
1470  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1471  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1472  }
1473 
1474  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1475  if(c->sub_flags&FLAG_QPEL)
1476  dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1477  else
1478  dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1479 
1480  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1481  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1482 
1483  get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1484 
1485  mv_table[mot_xy][0]= mx;
1486  mv_table[mot_xy][1]= my;
1487  c->flags &= ~FLAG_DIRECT;
1488  c->sub_flags &= ~FLAG_DIRECT;
1489 
1490  return dmin;
1491 }
1492 
1494  int mb_x, int mb_y)
1495 {
1496  MotionEstContext * const c= &s->me;
1497  const int penalty_factor= c->mb_penalty_factor;
1498  int fmin, bmin, dmin, fbmin, bimin, fimin;
1499  int type=0;
1500  const int xy = mb_y*s->mb_stride + mb_x;
1502  s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1503 
1504  get_limits(s, 16*mb_x, 16*mb_y);
1505 
1506  c->skip=0;
1507 
1508  if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1509  int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1510 
1511  score= ((unsigned)(score*score + 128*256))>>16;
1512  c->mc_mb_var_sum_temp += score;
1513  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1514  s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1515 
1516  return;
1517  }
1518 
1519  if (s->codec_id == AV_CODEC_ID_MPEG4)
1520  dmin= direct_search(s, mb_x, mb_y);
1521  else
1522  dmin= INT_MAX;
1523 //FIXME penalty stuff for non mpeg4
1524  c->skip=0;
1525  fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1526  3 * penalty_factor;
1527 
1528  c->skip=0;
1529  bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1530  2 * penalty_factor;
1531  av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1532 
1533  c->skip=0;
1534  fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1535  av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1536 
1538 //FIXME mb type penalty
1539  c->skip=0;
1541  fimin= interlaced_search(s, 0,
1543  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1545  bimin= interlaced_search(s, 2,
1547  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1548  }else
1549  fimin= bimin= INT_MAX;
1550 
1551  {
1552  int score= fmin;
1554 
1555  if (dmin <= score){
1556  score = dmin;
1557  type = CANDIDATE_MB_TYPE_DIRECT;
1558  }
1559  if(bmin<score){
1560  score=bmin;
1562  }
1563  if(fbmin<score){
1564  score=fbmin;
1566  }
1567  if(fimin<score){
1568  score=fimin;
1570  }
1571  if(bimin<score){
1572  score=bimin;
1574  }
1575 
1576  score= ((unsigned)(score*score + 128*256))>>16;
1577  c->mc_mb_var_sum_temp += score;
1578  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1579  }
1580 
1583  if(fimin < INT_MAX)
1585  if(bimin < INT_MAX)
1587  if(fimin < INT_MAX && bimin < INT_MAX){
1588  type |= CANDIDATE_MB_TYPE_BIDIR_I;
1589  }
1590  //FIXME something smarter
1591  if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1592  if(s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1593  type |= CANDIDATE_MB_TYPE_DIRECT0;
1594  }
1595 
1596  s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1597 }
1598 
1599 /* find best f_code for ME which do unlimited searches */
1600 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1601 {
1602  if(s->me_method>=ME_EPZS){
1603  int score[8];
1604  int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1605  uint8_t * fcode_tab= s->fcode_tab;
1606  int best_fcode=-1;
1607  int best_score=-10000000;
1608 
1609  if(s->msmpeg4_version)
1610  range= FFMIN(range, 16);
1612  range= FFMIN(range, 256);
1613 
1614  for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1615 
1616  for(y=0; y<s->mb_height; y++){
1617  int x;
1618  int xy= y*s->mb_stride;
1619  for(x=0; x<s->mb_width; x++){
1620  if(s->mb_type[xy] & type){
1621  int mx= mv_table[xy][0];
1622  int my= mv_table[xy][1];
1623  int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1624  fcode_tab[my + MAX_MV]);
1625  int j;
1626 
1627  if(mx >= range || mx < -range ||
1628  my >= range || my < -range)
1629  continue;
1630 
1631  for(j=0; j<fcode && j<8; j++){
1633  score[j]-= 170;
1634  }
1635  }
1636  xy++;
1637  }
1638  }
1639 
1640  for(i=1; i<8; i++){
1641  if(score[i] > best_score){
1642  best_score= score[i];
1643  best_fcode= i;
1644  }
1645  }
1646 
1647  return best_fcode;
1648  }else{
1649  return 1;
1650  }
1651 }
1652 
1654 {
1655  MotionEstContext * const c= &s->me;
1656  const int f_code= s->f_code;
1657  int y, range;
1658  assert(s->pict_type==AV_PICTURE_TYPE_P);
1659 
1660  range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1661 
1662  assert(range <= 16 || !s->msmpeg4_version);
1663  assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1664 
1665  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1666 
1667  if(s->flags&CODEC_FLAG_4MV){
1668  const int wrap= s->b8_stride;
1669 
1670  /* clip / convert to intra 8x8 type MVs */
1671  for(y=0; y<s->mb_height; y++){
1672  int xy= y*2*wrap;
1673  int i= y*s->mb_stride;
1674  int x;
1675 
1676  for(x=0; x<s->mb_width; x++){
1678  int block;
1679  for(block=0; block<4; block++){
1680  int off= (block& 1) + (block>>1)*wrap;
1681  int mx = s->current_picture.motion_val[0][ xy + off ][0];
1682  int my = s->current_picture.motion_val[0][ xy + off ][1];
1683 
1684  if( mx >=range || mx <-range
1685  || my >=range || my <-range){
1686  s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1689  }
1690  }
1691  }
1692  xy+=2;
1693  i++;
1694  }
1695  }
1696  }
1697 }
1698 
1703 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1704  int16_t (*mv_table)[2], int f_code, int type, int truncate)
1705 {
1706  MotionEstContext * const c= &s->me;
1707  int y, h_range, v_range;
1708 
1709  // RAL: 8 in MPEG-1, 16 in MPEG-4
1710  int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1711 
1712  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1713 
1714  h_range= range;
1715  v_range= field_select_table ? range>>1 : range;
1716 
1717  /* clip / convert to intra 16x16 type MVs */
1718  for(y=0; y<s->mb_height; y++){
1719  int x;
1720  int xy= y*s->mb_stride;
1721  for(x=0; x<s->mb_width; x++){
1722  if (s->mb_type[xy] & type){ // RAL: "type" test added...
1723  if(field_select_table==NULL || field_select_table[xy] == field_select){
1724  if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1725  || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1726 
1727  if(truncate){
1728  if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1729  else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1730  if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1731  else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1732  }else{
1733  s->mb_type[xy] &= ~type;
1734  s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1735  mv_table[xy][0]=
1736  mv_table[xy][1]= 0;
1737  }
1738  }
1739  }
1740  }
1741  xy++;
1742  }
1743  }
1744 }