FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
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 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
27  * Motion estimation.
28  */
29 
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 
47 static int sad_hpel_motion_search(MpegEncContext * s,
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  av_assert2(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  av_assert2((fx>>1) + 16*s->mb_x >= -16);
158  av_assert2((fy>>1) + 16*s->mb_y >= -16);
159  av_assert2((fx>>1) + 16*s->mb_x <= s->width);
160  av_assert2((fy>>1) + 16*s->mb_y <= s->height);
161  av_assert2((bx>>1) + 16*s->mb_x >= -16);
162  av_assert2((by>>1) + 16*s->mb_y >= -16);
163  av_assert2((bx>>1) + 16*s->mb_x <= s->width);
164  av_assert2((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 
248 /** @brief compares a block (either a full macroblock or a partition thereof)
249  against a proposed motion-compensated prediction of that block
250  */
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){
254  if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
255  && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
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 
295 int ff_init_me(MpegEncContext *s){
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 
300  if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
301  av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
302  return -1;
303  }
304  //special case of snow is needed because snow uses its own iterative ME code
305  if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
306  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");
307  return -1;
308  }
309 
310  c->avctx= s->avctx;
311 
312  if(cache_size < 2*dia_size && !c->stride){
313  av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
314  }
315 
316  ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
317  ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
318  ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
319  ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
320 
321  c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
322  c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
323  c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
324 
325 /*FIXME s->no_rounding b_type*/
326  if(s->flags&CODEC_FLAG_QPEL){
327  c->sub_motion_search= qpel_motion_search;
328  c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
329  if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
330  else c->qpel_put= s->dsp.put_qpel_pixels_tab;
331  }else{
332  if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
333  c->sub_motion_search= hpel_motion_search;
334  else if( c->avctx->me_sub_cmp == FF_CMP_SAD
335  && c->avctx-> me_cmp == FF_CMP_SAD
336  && c->avctx-> mb_cmp == FF_CMP_SAD)
337  c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
338  else
339  c->sub_motion_search= hpel_motion_search;
340  }
341  c->hpel_avg= s->dsp.avg_pixels_tab;
342  if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
343  else c->hpel_put= s->dsp.put_pixels_tab;
344 
345  if(s->linesize){
346  c->stride = s->linesize;
347  c->uvstride= s->uvlinesize;
348  }else{
349  c->stride = 16*s->mb_width + 32;
350  c->uvstride= 8*s->mb_width + 16;
351  }
352 
353  /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
354  * not have yet, and even if we had, the motion estimation code
355  * does not expect it. */
356  if(s->codec_id != AV_CODEC_ID_SNOW){
357  if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
358  s->dsp.me_cmp[2]= zero_cmp;
359  }
360  if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
361  s->dsp.me_sub_cmp[2]= zero_cmp;
362  }
363  c->hpel_put[2][0]= c->hpel_put[2][1]=
364  c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
365  }
366 
367  if(s->codec_id == AV_CODEC_ID_H261){
368  c->sub_motion_search= no_sub_motion_search;
369  }
370 
371  return 0;
372 }
373 
374 #define CHECK_SAD_HALF_MV(suffix, x, y) \
375 {\
376  d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
377  d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
378  COPY3_IF_LT(dminh, d, dx, x, dy, y)\
379 }
380 
381 static int sad_hpel_motion_search(MpegEncContext * s,
382  int *mx_ptr, int *my_ptr, int dmin,
383  int src_index, int ref_index,
384  int size, int h)
385 {
386  MotionEstContext * const c= &s->me;
387  const int penalty_factor= c->sub_penalty_factor;
388  int mx, my, dminh;
389  uint8_t *pix, *ptr;
390  int stride= c->stride;
391  const int flags= c->sub_flags;
392  LOAD_COMMON
393 
394  av_assert2(flags == 0);
395 
396  if(c->skip){
397  *mx_ptr = 0;
398  *my_ptr = 0;
399  return dmin;
400  }
401 
402  pix = c->src[src_index][0];
403 
404  mx = *mx_ptr;
405  my = *my_ptr;
406  ptr = c->ref[ref_index][0] + (my * stride) + mx;
407 
408  dminh = dmin;
409 
410  if (mx > xmin && mx < xmax &&
411  my > ymin && my < ymax) {
412  int dx=0, dy=0;
413  int d, pen_x, pen_y;
414  const int index= (my<<ME_MAP_SHIFT) + mx;
415  const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
416  const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
417  const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
418  const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
419  mx<<=1;
420  my<<=1;
421 
422 
423  pen_x= pred_x + mx;
424  pen_y= pred_y + my;
425 
426  ptr-= stride;
427  if(t<=b){
428  CHECK_SAD_HALF_MV(y2 , 0, -1)
429  if(l<=r){
430  CHECK_SAD_HALF_MV(xy2, -1, -1)
431  if(t+r<=b+l){
432  CHECK_SAD_HALF_MV(xy2, +1, -1)
433  ptr+= stride;
434  }else{
435  ptr+= stride;
436  CHECK_SAD_HALF_MV(xy2, -1, +1)
437  }
438  CHECK_SAD_HALF_MV(x2 , -1, 0)
439  }else{
440  CHECK_SAD_HALF_MV(xy2, +1, -1)
441  if(t+l<=b+r){
442  CHECK_SAD_HALF_MV(xy2, -1, -1)
443  ptr+= stride;
444  }else{
445  ptr+= stride;
446  CHECK_SAD_HALF_MV(xy2, +1, +1)
447  }
448  CHECK_SAD_HALF_MV(x2 , +1, 0)
449  }
450  }else{
451  if(l<=r){
452  if(t+l<=b+r){
453  CHECK_SAD_HALF_MV(xy2, -1, -1)
454  ptr+= stride;
455  }else{
456  ptr+= stride;
457  CHECK_SAD_HALF_MV(xy2, +1, +1)
458  }
459  CHECK_SAD_HALF_MV(x2 , -1, 0)
460  CHECK_SAD_HALF_MV(xy2, -1, +1)
461  }else{
462  if(t+r<=b+l){
463  CHECK_SAD_HALF_MV(xy2, +1, -1)
464  ptr+= stride;
465  }else{
466  ptr+= stride;
467  CHECK_SAD_HALF_MV(xy2, -1, +1)
468  }
469  CHECK_SAD_HALF_MV(x2 , +1, 0)
470  CHECK_SAD_HALF_MV(xy2, +1, +1)
471  }
472  CHECK_SAD_HALF_MV(y2 , 0, +1)
473  }
474  mx+=dx;
475  my+=dy;
476 
477  }else{
478  mx<<=1;
479  my<<=1;
480  }
481 
482  *mx_ptr = mx;
483  *my_ptr = my;
484  return dminh;
485 }
486 
487 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
488 {
489  const int xy= s->mb_x + s->mb_y*s->mb_stride;
490 
491  s->p_mv_table[xy][0] = mx;
492  s->p_mv_table[xy][1] = my;
493 
494  /* has already been set to the 4 MV if 4MV is done */
495  if(mv4){
496  int mot_xy= s->block_index[0];
497 
498  s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
499  s->current_picture.f.motion_val[0][mot_xy ][1] = my;
500  s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
501  s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
502 
503  mot_xy += s->b8_stride;
504  s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
505  s->current_picture.f.motion_val[0][mot_xy ][1] = my;
506  s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
507  s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
508  }
509 }
510 
511 /**
512  * get fullpel ME search limits.
513  */
514 static inline void get_limits(MpegEncContext *s, int x, int y)
515 {
516  MotionEstContext * const c= &s->me;
517  int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
518  int max_range = MAX_MV >> (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->width;
527  c->ymax = - y + s->height;
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 || range > max_range)
541  range = max_range;
542  if(range){
543  c->xmin = FFMAX(c->xmin,-range);
544  c->xmax = FFMIN(c->xmax, range);
545  c->ymin = FFMAX(c->ymin,-range);
546  c->ymax = FFMIN(c->ymax, range);
547  }
548 }
549 
550 static inline void init_mv4_ref(MotionEstContext *c){
551  const int stride= c->stride;
552 
553  c->ref[1][0] = c->ref[0][0] + 8;
554  c->ref[2][0] = c->ref[0][0] + 8*stride;
555  c->ref[3][0] = c->ref[2][0] + 8;
556  c->src[1][0] = c->src[0][0] + 8;
557  c->src[2][0] = c->src[0][0] + 8*stride;
558  c->src[3][0] = c->src[2][0] + 8;
559 }
560 
561 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
562 {
563  MotionEstContext * const c= &s->me;
564  const int size= 1;
565  const int h=8;
566  int block;
567  int P[10][2];
568  int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
569  int same=1;
570  const int stride= c->stride;
571  uint8_t *mv_penalty= c->current_mv_penalty;
572  int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
573 
574  init_mv4_ref(c);
575 
576  for(block=0; block<4; block++){
577  int mx4, my4;
578  int pred_x4, pred_y4;
579  int dmin4;
580  static const int off[4]= {2, 1, 1, -1};
581  const int mot_stride = s->b8_stride;
582  const int mot_xy = s->block_index[block];
583 
584  if(saftey_cliping){
585  c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
586  c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
587  }
588 
589  P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
590  P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
591 
592  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
593 
594  /* special case for first line */
595  if (s->first_slice_line && block<2) {
596  c->pred_x= pred_x4= P_LEFT[0];
597  c->pred_y= pred_y4= P_LEFT[1];
598  } else {
599  P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
600  P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
601  P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][0];
602  P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][1];
603  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
604  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
605  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
606  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
607 
608  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
609  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
610 
611  c->pred_x= pred_x4 = P_MEDIAN[0];
612  c->pred_y= pred_y4 = P_MEDIAN[1];
613  }
614  P_MV1[0]= mx;
615  P_MV1[1]= my;
616  if(saftey_cliping)
617  for(i=0; i<10; i++){
618  if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
619  if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
620  }
621 
622  dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
623 
624  dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
625 
626  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
627  int dxy;
628  const int offset= ((block&1) + (block>>1)*stride)*8;
629  uint8_t *dest_y = c->scratchpad + offset;
630  if(s->quarter_sample){
631  uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
632  dxy = ((my4 & 3) << 2) | (mx4 & 3);
633 
634  if(s->no_rounding)
635  s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
636  else
637  s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
638  }else{
639  uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
640  dxy = ((my4 & 1) << 1) | (mx4 & 1);
641 
642  if(s->no_rounding)
643  s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
644  else
645  s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
646  }
647  dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
648  }else
649  dmin_sum+= dmin4;
650 
651  if(s->quarter_sample){
652  mx4_sum+= mx4/2;
653  my4_sum+= my4/2;
654  }else{
655  mx4_sum+= mx4;
656  my4_sum+= my4;
657  }
658 
659  s->current_picture.f.motion_val[0][s->block_index[block]][0] = mx4;
660  s->current_picture.f.motion_val[0][s->block_index[block]][1] = my4;
661 
662  if(mx4 != mx || my4 != my) same=0;
663  }
664 
665  if(same)
666  return INT_MAX;
667 
668  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
669  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);
670  }
671 
672  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
673  int dxy;
674  int mx, my;
675  int offset;
676 
677  mx= ff_h263_round_chroma(mx4_sum);
678  my= ff_h263_round_chroma(my4_sum);
679  dxy = ((my & 1) << 1) | (mx & 1);
680 
681  offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
682 
683  if(s->no_rounding){
684  s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
685  s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
686  }else{
687  s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
688  s->dsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
689  }
690 
691  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);
692  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);
693  }
694 
695  c->pred_x= mx;
696  c->pred_y= my;
697 
698  switch(c->avctx->mb_cmp&0xFF){
699  /*case FF_CMP_SSE:
700  return dmin_sum+ 32*s->qscale*s->qscale;*/
701  case FF_CMP_RD:
702  return dmin_sum;
703  default:
704  return dmin_sum+ 11*c->mb_penalty_factor;
705  }
706 }
707 
708 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
709  MotionEstContext * const c= &s->me;
710 
711  c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
712  c->src[1][0] = c->src[0][0] + s->linesize;
713  if(c->flags & FLAG_CHROMA){
714  c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
715  c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
716  c->src[1][1] = c->src[0][1] + s->uvlinesize;
717  c->src[1][2] = c->src[0][2] + s->uvlinesize;
718  }
719 }
720 
721 static int interlaced_search(MpegEncContext *s, int ref_index,
722  int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
723 {
724  MotionEstContext * const c= &s->me;
725  const int size=0;
726  const int h=8;
727  int block;
728  int P[10][2];
729  uint8_t * const mv_penalty= c->current_mv_penalty;
730  int same=1;
731  const int stride= 2*s->linesize;
732  int dmin_sum= 0;
733  const int mot_stride= s->mb_stride;
734  const int xy= s->mb_x + s->mb_y*mot_stride;
735 
736  c->ymin>>=1;
737  c->ymax>>=1;
738  c->stride<<=1;
739  c->uvstride<<=1;
740  init_interlaced_ref(s, ref_index);
741 
742  for(block=0; block<2; block++){
743  int field_select;
744  int best_dmin= INT_MAX;
745  int best_field= -1;
746 
747  for(field_select=0; field_select<2; field_select++){
748  int dmin, mx_i, my_i;
749  int16_t (*mv_table)[2]= mv_tables[block][field_select];
750 
751  if(user_field_select){
752  av_assert1(field_select==0 || field_select==1);
753  av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
754  if(field_select_tables[block][xy] != field_select)
755  continue;
756  }
757 
758  P_LEFT[0] = mv_table[xy - 1][0];
759  P_LEFT[1] = mv_table[xy - 1][1];
760  if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
761 
762  c->pred_x= P_LEFT[0];
763  c->pred_y= P_LEFT[1];
764 
765  if(!s->first_slice_line){
766  P_TOP[0] = mv_table[xy - mot_stride][0];
767  P_TOP[1] = mv_table[xy - mot_stride][1];
768  P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
769  P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
770  if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
771  if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
772  if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
773  if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
774 
775  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
776  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
777  }
778  P_MV1[0]= mx; //FIXME not correct if block != field_select
779  P_MV1[1]= my / 2;
780 
781  dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
782 
783  dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
784 
785  mv_table[xy][0]= mx_i;
786  mv_table[xy][1]= my_i;
787 
788  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
789  int dxy;
790 
791  //FIXME chroma ME
792  uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
793  dxy = ((my_i & 1) << 1) | (mx_i & 1);
794 
795  if(s->no_rounding){
796  s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
797  }else{
798  s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
799  }
800  dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
801  dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
802  }else
803  dmin+= c->mb_penalty_factor; //field_select bits
804 
805  dmin += field_select != block; //slightly prefer same field
806 
807  if(dmin < best_dmin){
808  best_dmin= dmin;
809  best_field= field_select;
810  }
811  }
812  {
813  int16_t (*mv_table)[2]= mv_tables[block][best_field];
814 
815  if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
816  if(mv_table[xy][1]&1) same=0;
817  if(mv_table[xy][1]*2 != my) same=0;
818  if(best_field != block) same=0;
819  }
820 
821  field_select_tables[block][xy]= best_field;
822  dmin_sum += best_dmin;
823  }
824 
825  c->ymin<<=1;
826  c->ymax<<=1;
827  c->stride>>=1;
828  c->uvstride>>=1;
829 
830  if(same)
831  return INT_MAX;
832 
833  switch(c->avctx->mb_cmp&0xFF){
834  /*case FF_CMP_SSE:
835  return dmin_sum+ 32*s->qscale*s->qscale;*/
836  case FF_CMP_RD:
837  return dmin_sum;
838  default:
839  return dmin_sum+ 11*c->mb_penalty_factor;
840  }
841 }
842 
843 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
844  int ymax= s->me.ymax>>interlaced;
845  int ymin= s->me.ymin>>interlaced;
846 
847  if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
848  if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
849  if(mv[1] < ymin) mv[1] = ymin;
850  if(mv[1] > ymax) mv[1] = ymax;
851 }
852 
853 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
854  MotionEstContext * const c= &s->me;
855  Picture *p= s->current_picture_ptr;
856  int mb_xy= mb_x + mb_y*s->mb_stride;
857  int xy= 2*mb_x + 2*mb_y*s->b8_stride;
858  int mb_type= s->current_picture.f.mb_type[mb_xy];
859  int flags= c->flags;
860  int shift= (flags&FLAG_QPEL) + 1;
861  int mask= (1<<shift)-1;
862  int x, y, i;
863  int d=0;
864  me_cmp_func cmpf= s->dsp.sse[0];
865  me_cmp_func chroma_cmpf= s->dsp.sse[1];
866 
867  if(p_type && USES_LIST(mb_type, 1)){
868  av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
869  return INT_MAX/2;
870  }
871  av_assert0(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
872 
873  for(i=0; i<4; i++){
874  int xy= s->block_index[i];
875  clip_input_mv(s, p->f.motion_val[0][xy], !!IS_INTERLACED(mb_type));
876  clip_input_mv(s, p->f.motion_val[1][xy], !!IS_INTERLACED(mb_type));
877  }
878 
879  if(IS_INTERLACED(mb_type)){
880  int xy2= xy + s->b8_stride;
881  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
882  c->stride<<=1;
883  c->uvstride<<=1;
884 
885  if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
886  av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
887  return INT_MAX/2;
888  }
889 
890  if(USES_LIST(mb_type, 0)){
891  int field_select0= p->f.ref_index[0][4*mb_xy ];
892  int field_select1= p->f.ref_index[0][4*mb_xy+2];
893  av_assert0(field_select0==0 ||field_select0==1);
894  av_assert0(field_select1==0 ||field_select1==1);
895  init_interlaced_ref(s, 0);
896 
897  if(p_type){
898  s->p_field_select_table[0][mb_xy]= field_select0;
899  s->p_field_select_table[1][mb_xy]= field_select1;
900  *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
901  *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
902  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
903  }else{
904  s->b_field_select_table[0][0][mb_xy]= field_select0;
905  s->b_field_select_table[0][1][mb_xy]= field_select1;
906  *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
907  *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
908  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
909  }
910 
911  x = p->f.motion_val[0][xy ][0];
912  y = p->f.motion_val[0][xy ][1];
913  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
914  x = p->f.motion_val[0][xy2][0];
915  y = p->f.motion_val[0][xy2][1];
916  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
917  }
918  if(USES_LIST(mb_type, 1)){
919  int field_select0 = p->f.ref_index[1][4 * mb_xy ];
920  int field_select1 = p->f.ref_index[1][4 * mb_xy + 2];
921  av_assert0(field_select0==0 ||field_select0==1);
922  av_assert0(field_select1==0 ||field_select1==1);
923  init_interlaced_ref(s, 2);
924 
925  s->b_field_select_table[1][0][mb_xy]= field_select0;
926  s->b_field_select_table[1][1][mb_xy]= field_select1;
927  *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy ];
928  *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy2];
929  if(USES_LIST(mb_type, 0)){
930  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
931  }else{
932  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
933  }
934 
935  x = p->f.motion_val[1][xy ][0];
936  y = p->f.motion_val[1][xy ][1];
937  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
938  x = p->f.motion_val[1][xy2][0];
939  y = p->f.motion_val[1][xy2][1];
940  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
941  //FIXME bidir scores
942  }
943  c->stride>>=1;
944  c->uvstride>>=1;
945  }else if(IS_8X8(mb_type)){
946  if(!(s->flags & CODEC_FLAG_4MV)){
947  av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
948  return INT_MAX/2;
949  }
950  cmpf= s->dsp.sse[1];
951  chroma_cmpf= s->dsp.sse[1];
952  init_mv4_ref(c);
953  for(i=0; i<4; i++){
954  xy= s->block_index[i];
955  x= p->f.motion_val[0][xy][0];
956  y= p->f.motion_val[0][xy][1];
957  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
958  }
959  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
960  }else{
961  if(USES_LIST(mb_type, 0)){
962  if(p_type){
963  *(uint32_t*)s->p_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
964  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
965  }else if(USES_LIST(mb_type, 1)){
966  *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
967  *(uint32_t*)s->b_bidir_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
968  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
969  }else{
970  *(uint32_t*)s->b_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
971  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
972  }
973  x = p->f.motion_val[0][xy][0];
974  y = p->f.motion_val[0][xy][1];
975  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
976  }else if(USES_LIST(mb_type, 1)){
977  *(uint32_t*)s->b_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
978  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
979 
980  x = p->f.motion_val[1][xy][0];
981  y = p->f.motion_val[1][xy][1];
982  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
983  }else
984  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
985  }
986  return d;
987 }
988 
989 static inline int get_penalty_factor(int lambda, int lambda2, int type){
990  switch(type&0xFF){
991  default:
992  case FF_CMP_SAD:
993  return lambda>>FF_LAMBDA_SHIFT;
994  case FF_CMP_DCT:
995  return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
996  case FF_CMP_W53:
997  return (4*lambda)>>(FF_LAMBDA_SHIFT);
998  case FF_CMP_W97:
999  return (2*lambda)>>(FF_LAMBDA_SHIFT);
1000  case FF_CMP_SATD:
1001  case FF_CMP_DCT264:
1002  return (2*lambda)>>FF_LAMBDA_SHIFT;
1003  case FF_CMP_RD:
1004  case FF_CMP_PSNR:
1005  case FF_CMP_SSE:
1006  case FF_CMP_NSSE:
1007  return lambda2>>FF_LAMBDA_SHIFT;
1008  case FF_CMP_BIT:
1009  return 1;
1010  }
1011 }
1012 
1013 void ff_estimate_p_frame_motion(MpegEncContext * s,
1014  int mb_x, int mb_y)
1015 {
1016  MotionEstContext * const c= &s->me;
1017  uint8_t *pix, *ppix;
1018  int sum, mx, my, dmin;
1019  int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1020  int vard; ///< sum of squared differences with the estimated motion vector
1021  int P[10][2];
1022  const int shift= 1+s->quarter_sample;
1023  int mb_type=0;
1024  Picture * const pic= &s->current_picture;
1025 
1026  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1027 
1028  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1029  av_assert0(s->linesize == c->stride);
1030  av_assert0(s->uvlinesize == c->uvstride);
1031 
1032  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1033  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1034  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1035  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1036 
1037  get_limits(s, 16*mb_x, 16*mb_y);
1038  c->skip=0;
1039 
1040  /* intra / predictive decision */
1041  pix = c->src[0][0];
1042  sum = s->dsp.pix_sum(pix, s->linesize);
1043  varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
1044 
1045  pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1046  pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1047  c->mb_var_sum_temp += (varc+128)>>8;
1048 
1049  if(c->avctx->me_threshold){
1050  vard= check_input_motion(s, mb_x, mb_y, 1);
1051 
1052  if((vard+128)>>8 < c->avctx->me_threshold){
1053  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1054  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1055  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1056  c->mc_mb_var_sum_temp += (vard+128)>>8;
1057  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1058  return;
1059  }
1060  if((vard+128)>>8 < c->avctx->mb_threshold)
1061  mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1062  }
1063 
1064  switch(s->me_method) {
1065  case ME_ZERO:
1066  default:
1067  mx = 0;
1068  my = 0;
1069  dmin = 0;
1070  break;
1071  case ME_X1:
1072  case ME_EPZS:
1073  {
1074  const int mot_stride = s->b8_stride;
1075  const int mot_xy = s->block_index[0];
1076 
1077  P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
1078  P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
1079 
1080  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1081 
1082  if(!s->first_slice_line) {
1083  P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
1084  P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
1085  P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][0];
1086  P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][1];
1087  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1088  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1089  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1090 
1091  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1092  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1093 
1094  if(s->out_format == FMT_H263){
1095  c->pred_x = P_MEDIAN[0];
1096  c->pred_y = P_MEDIAN[1];
1097  }else { /* mpeg1 at least */
1098  c->pred_x= P_LEFT[0];
1099  c->pred_y= P_LEFT[1];
1100  }
1101  }else{
1102  c->pred_x= P_LEFT[0];
1103  c->pred_y= P_LEFT[1];
1104  }
1105 
1106  }
1107  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1108 
1109  break;
1110  }
1111 
1112  /* At this point (mx,my) are full-pell and the relative displacement */
1113  ppix = c->ref[0][0] + (my * s->linesize) + mx;
1114 
1115  vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1116 
1117  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1118  c->mc_mb_var_sum_temp += (vard+128)>>8;
1119 
1120  if(mb_type){
1121  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1122  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1123  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1124 
1125  if(mb_type == CANDIDATE_MB_TYPE_INTER){
1126  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1127  set_p_mv_tables(s, mx, my, 1);
1128  }else{
1129  mx <<=shift;
1130  my <<=shift;
1131  }
1132  if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1133  h263_mv4_search(s, mx, my, shift);
1134 
1135  set_p_mv_tables(s, mx, my, 0);
1136  }
1137  if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1138  interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1139  }
1140  }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1141  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1142  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1143  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1144 
1145  if (vard*2 + 200*256 > varc)
1146  mb_type|= CANDIDATE_MB_TYPE_INTRA;
1147  if (varc*2 + 200*256 > vard || s->qscale > 24){
1148 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1149  mb_type|= CANDIDATE_MB_TYPE_INTER;
1150  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1151  if(s->flags&CODEC_FLAG_MV0)
1152  if(mx || my)
1153  mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1154  }else{
1155  mx <<=shift;
1156  my <<=shift;
1157  }
1158  if((s->flags&CODEC_FLAG_4MV)
1159  && !c->skip && varc>50<<8 && vard>10<<8){
1160  if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1161  mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1162 
1163  set_p_mv_tables(s, mx, my, 0);
1164  }else
1165  set_p_mv_tables(s, mx, my, 1);
1166  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1167  && !c->skip){ //FIXME varc/d checks
1168  if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1169  mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1170  }
1171  }else{
1172  int intra_score, i;
1173  mb_type= CANDIDATE_MB_TYPE_INTER;
1174 
1175  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1176  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1177  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1178 
1179  if((s->flags&CODEC_FLAG_4MV)
1180  && !c->skip && varc>50<<8 && vard>10<<8){
1181  int dmin4= h263_mv4_search(s, mx, my, shift);
1182  if(dmin4 < dmin){
1183  mb_type= CANDIDATE_MB_TYPE_INTER4V;
1184  dmin=dmin4;
1185  }
1186  }
1187  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1188  && !c->skip){ //FIXME varc/d checks
1189  int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1190  if(dmin_i < dmin){
1191  mb_type = CANDIDATE_MB_TYPE_INTER_I;
1192  dmin= dmin_i;
1193  }
1194  }
1195 
1196  set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1197 
1198  /* get intra luma score */
1199  if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1200  intra_score= varc - 500;
1201  }else{
1202  unsigned mean = (sum+128)>>8;
1203  mean*= 0x01010101;
1204 
1205  for(i=0; i<16; i++){
1206  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1207  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1208  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1209  *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1210  }
1211 
1212  intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1213  }
1214  intra_score += c->mb_penalty_factor*16;
1215 
1216  if(intra_score < dmin){
1217  mb_type= CANDIDATE_MB_TYPE_INTRA;
1218  s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1219  }else
1220  s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1221 
1222  {
1223  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1224  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1225  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1226  }
1227  }
1228 
1229  s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1230 }
1231 
1232 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1233  int mb_x, int mb_y)
1234 {
1235  MotionEstContext * const c= &s->me;
1236  int mx, my, dmin;
1237  int P[10][2];
1238  const int shift= 1+s->quarter_sample;
1239  const int xy= mb_x + mb_y*s->mb_stride;
1240  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1241 
1242  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1243 
1244  c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1245  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1246 
1247  get_limits(s, 16*mb_x, 16*mb_y);
1248  c->skip=0;
1249 
1250  P_LEFT[0] = s->p_mv_table[xy + 1][0];
1251  P_LEFT[1] = s->p_mv_table[xy + 1][1];
1252 
1253  if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1254 
1255  /* special case for first line */
1256  if (s->first_slice_line) {
1257  c->pred_x= P_LEFT[0];
1258  c->pred_y= P_LEFT[1];
1259  P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1260  P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1261  } else {
1262  P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1263  P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1264  P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1265  P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1266  if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1267  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1268  if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1269 
1270  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1271  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1272 
1273  c->pred_x = P_MEDIAN[0];
1274  c->pred_y = P_MEDIAN[1];
1275  }
1276 
1277  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1278 
1279  s->p_mv_table[xy][0] = mx<<shift;
1280  s->p_mv_table[xy][1] = my<<shift;
1281 
1282  return dmin;
1283 }
1284 
1285 static int ff_estimate_motion_b(MpegEncContext * s,
1286  int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1287 {
1288  MotionEstContext * const c= &s->me;
1289  int mx, my, dmin;
1290  int P[10][2];
1291  const int shift= 1+s->quarter_sample;
1292  const int mot_stride = s->mb_stride;
1293  const int mot_xy = mb_y*mot_stride + mb_x;
1294  uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1295  int mv_scale;
1296 
1297  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1298  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1299  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1300  c->current_mv_penalty= mv_penalty;
1301 
1302  get_limits(s, 16*mb_x, 16*mb_y);
1303 
1304  switch(s->me_method) {
1305  case ME_ZERO:
1306  default:
1307  mx = 0;
1308  my = 0;
1309  dmin = 0;
1310  break;
1311  case ME_X1:
1312  case ME_EPZS:
1313  P_LEFT[0] = mv_table[mot_xy - 1][0];
1314  P_LEFT[1] = mv_table[mot_xy - 1][1];
1315 
1316  if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1317 
1318  /* special case for first line */
1319  if (!s->first_slice_line) {
1320  P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1321  P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1322  P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1323  P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1324  if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1325  if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1326  if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1327 
1328  P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1329  P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1330  }
1331  c->pred_x = P_LEFT[0];
1332  c->pred_y = P_LEFT[1];
1333 
1334  if(mv_table == s->b_forw_mv_table){
1335  mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1336  }else{
1337  mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1338  }
1339 
1340  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1341 
1342  break;
1343  }
1344 
1345  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1346 
1347  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1348  dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1349 
1350 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1351  mv_table[mot_xy][0]= mx;
1352  mv_table[mot_xy][1]= my;
1353 
1354  return dmin;
1355 }
1356 
1357 static inline int check_bidir_mv(MpegEncContext * s,
1358  int motion_fx, int motion_fy,
1359  int motion_bx, int motion_by,
1360  int pred_fx, int pred_fy,
1361  int pred_bx, int pred_by,
1362  int size, int h)
1363 {
1364  //FIXME optimize?
1365  //FIXME better f_code prediction (max mv & distance)
1366  //FIXME pointers
1367  MotionEstContext * const c= &s->me;
1368  uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1369  uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1370  int stride= c->stride;
1371  uint8_t *dest_y = c->scratchpad;
1372  uint8_t *ptr;
1373  int dxy;
1374  int src_x, src_y;
1375  int fbmin;
1376  uint8_t **src_data= c->src[0];
1377  uint8_t **ref_data= c->ref[0];
1378  uint8_t **ref2_data= c->ref[2];
1379 
1380  if(s->quarter_sample){
1381  dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1382  src_x = motion_fx >> 2;
1383  src_y = motion_fy >> 2;
1384 
1385  ptr = ref_data[0] + (src_y * stride) + src_x;
1386  s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1387 
1388  dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1389  src_x = motion_bx >> 2;
1390  src_y = motion_by >> 2;
1391 
1392  ptr = ref2_data[0] + (src_y * stride) + src_x;
1393  s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1394  }else{
1395  dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1396  src_x = motion_fx >> 1;
1397  src_y = motion_fy >> 1;
1398 
1399  ptr = ref_data[0] + (src_y * stride) + src_x;
1400  s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1401 
1402  dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1403  src_x = motion_bx >> 1;
1404  src_y = motion_by >> 1;
1405 
1406  ptr = ref2_data[0] + (src_y * stride) + src_x;
1407  s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1408  }
1409 
1410  fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1411  +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1412  + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1413 
1414  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1415  }
1416  //FIXME CHROMA !!!
1417 
1418  return fbmin;
1419 }
1420 
1421 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1422 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1423 {
1424  MotionEstContext * const c= &s->me;
1425  const int mot_stride = s->mb_stride;
1426  const int xy = mb_y *mot_stride + mb_x;
1427  int fbmin;
1428  int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1429  int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1430  int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1431  int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1432  int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1433  int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1434  int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1435  int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1436  const int flags= c->sub_flags;
1437  const int qpel= flags&FLAG_QPEL;
1438  const int shift= 1+qpel;
1439  const int xmin= c->xmin<<shift;
1440  const int ymin= c->ymin<<shift;
1441  const int xmax= c->xmax<<shift;
1442  const int ymax= c->ymax<<shift;
1443 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1444 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1445  int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1446  uint8_t map[256] = { 0 };
1447 
1448  map[hashidx&255] = 1;
1449 
1450  fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1451  motion_bx, motion_by,
1452  pred_fx, pred_fy,
1453  pred_bx, pred_by,
1454  0, 16);
1455 
1456  if(s->avctx->bidir_refine){
1457  int end;
1458  static const uint8_t limittab[5]={0,8,32,64,80};
1459  const int limit= limittab[s->avctx->bidir_refine];
1460  static const int8_t vect[][4]={
1461 { 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},
1462 
1463 { 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},
1464 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1465 { 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},
1466 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1467 
1468 { 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},
1469 { 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},
1470 { 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},
1471 { 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},
1472 
1473 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1474 { 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},
1475 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1476  };
1477  static const uint8_t hash[]={
1478 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),
1479 
1480 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),
1481 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1482 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),
1483 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1484 
1485 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),
1486 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),
1487 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),
1488 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),
1489 
1490 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1491 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),
1492 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),
1493 };
1494 
1495 #define CHECK_BIDIR(fx,fy,bx,by)\
1496  if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1497  &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1498  &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1499  int score;\
1500  map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1501  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);\
1502  if(score < fbmin){\
1503  hashidx += HASH(fx,fy,bx,by);\
1504  fbmin= score;\
1505  motion_fx+=fx;\
1506  motion_fy+=fy;\
1507  motion_bx+=bx;\
1508  motion_by+=by;\
1509  end=0;\
1510  }\
1511  }
1512 #define CHECK_BIDIR2(a,b,c,d)\
1513 CHECK_BIDIR(a,b,c,d)\
1514 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1515 
1516  do{
1517  int i;
1518  int borderdist=0;
1519  end=1;
1520 
1521  CHECK_BIDIR2(0,0,0,1)
1522  CHECK_BIDIR2(0,0,1,0)
1523  CHECK_BIDIR2(0,1,0,0)
1524  CHECK_BIDIR2(1,0,0,0)
1525 
1526  for(i=8; i<limit; i++){
1527  int fx= motion_fx+vect[i][0];
1528  int fy= motion_fy+vect[i][1];
1529  int bx= motion_bx+vect[i][2];
1530  int by= motion_by+vect[i][3];
1531  if(borderdist<=0){
1532  int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1533  int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1534  if((a|b) < 0)
1535  map[(hashidx+hash[i])&255] = 1;
1536  }
1537  if(!map[(hashidx+hash[i])&255]){
1538  int score;
1539  map[(hashidx+hash[i])&255] = 1;
1540  score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1541  if(score < fbmin){
1542  hashidx += hash[i];
1543  fbmin= score;
1544  motion_fx=fx;
1545  motion_fy=fy;
1546  motion_bx=bx;
1547  motion_by=by;
1548  end=0;
1549  borderdist--;
1550  if(borderdist<=0){
1551  int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1552  int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1553  borderdist= FFMIN(a,b);
1554  }
1555  }
1556  }
1557  }
1558  }while(!end);
1559  }
1560 
1561  s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1562  s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1563  s->b_bidir_back_mv_table[xy][0]= motion_bx;
1564  s->b_bidir_back_mv_table[xy][1]= motion_by;
1565 
1566  return fbmin;
1567 }
1568 
1569 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1570 {
1571  MotionEstContext * const c= &s->me;
1572  int P[10][2];
1573  const int mot_stride = s->mb_stride;
1574  const int mot_xy = mb_y*mot_stride + mb_x;
1575  const int shift= 1+s->quarter_sample;
1576  int dmin, i;
1577  const int time_pp= s->pp_time;
1578  const int time_pb= s->pb_time;
1579  int mx, my, xmin, xmax, ymin, ymax;
1580  int16_t (*mv_table)[2]= s->b_direct_mv_table;
1581 
1582  c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1583  ymin= xmin=(-32)>>shift;
1584  ymax= xmax= 31>>shift;
1585 
1586  if (IS_8X8(s->next_picture.f.mb_type[mot_xy])) {
1587  s->mv_type= MV_TYPE_8X8;
1588  }else{
1589  s->mv_type= MV_TYPE_16X16;
1590  }
1591 
1592  for(i=0; i<4; i++){
1593  int index= s->block_index[i];
1594  int min, max;
1595 
1596  c->co_located_mv[i][0] = s->next_picture.f.motion_val[0][index][0];
1597  c->co_located_mv[i][1] = s->next_picture.f.motion_val[0][index][1];
1598  c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1599  c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1600 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1601 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1602 
1603  max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1604  min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1605  max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1606  min+= 16*mb_x - 1;
1607  xmax= FFMIN(xmax, s->width - max);
1608  xmin= FFMAX(xmin, - 16 - min);
1609 
1610  max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1611  min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1612  max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1613  min+= 16*mb_y - 1;
1614  ymax= FFMIN(ymax, s->height - max);
1615  ymin= FFMAX(ymin, - 16 - min);
1616 
1617  if(s->mv_type == MV_TYPE_16X16) break;
1618  }
1619 
1620  av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1621 
1622  if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1623  s->b_direct_mv_table[mot_xy][0]= 0;
1624  s->b_direct_mv_table[mot_xy][1]= 0;
1625 
1626  return 256*256*256*64;
1627  }
1628 
1629  c->xmin= xmin;
1630  c->ymin= ymin;
1631  c->xmax= xmax;
1632  c->ymax= ymax;
1633  c->flags |= FLAG_DIRECT;
1634  c->sub_flags |= FLAG_DIRECT;
1635  c->pred_x=0;
1636  c->pred_y=0;
1637 
1638  P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1639  P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1640 
1641  /* special case for first line */
1642  if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1643  P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1644  P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1645  P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1646  P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1647 
1648  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1649  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1650  }
1651 
1652  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1653  if(c->sub_flags&FLAG_QPEL)
1654  dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1655  else
1656  dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1657 
1658  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1659  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1660 
1661  get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1662 
1663  mv_table[mot_xy][0]= mx;
1664  mv_table[mot_xy][1]= my;
1665  c->flags &= ~FLAG_DIRECT;
1666  c->sub_flags &= ~FLAG_DIRECT;
1667 
1668  return dmin;
1669 }
1670 
1671 void ff_estimate_b_frame_motion(MpegEncContext * s,
1672  int mb_x, int mb_y)
1673 {
1674  MotionEstContext * const c= &s->me;
1675  const int penalty_factor= c->mb_penalty_factor;
1676  int fmin, bmin, dmin, fbmin, bimin, fimin;
1677  int type=0;
1678  const int xy = mb_y*s->mb_stride + mb_x;
1679  init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1680  s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1681 
1682  get_limits(s, 16*mb_x, 16*mb_y);
1683 
1684  c->skip=0;
1685 
1686  if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.f.mbskip_table[xy]) {
1687  int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1688 
1689  score= ((unsigned)(score*score + 128*256))>>16;
1690  c->mc_mb_var_sum_temp += score;
1691  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1692  s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1693 
1694  return;
1695  }
1696 
1697  if(c->avctx->me_threshold){
1698  int vard= check_input_motion(s, mb_x, mb_y, 0);
1699 
1700  if((vard+128)>>8 < c->avctx->me_threshold){
1701 // pix = c->src[0][0];
1702 // sum = s->dsp.pix_sum(pix, s->linesize);
1703 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1704 
1705 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1706  s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1707 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1708  c->mb_var_sum_temp += (varc+128)>>8;*/
1709  c->mc_mb_var_sum_temp += (vard+128)>>8;
1710 /* if (vard <= 64<<8 || vard < varc) {
1711  c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1712  }else{
1713  c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1714  }*/
1715  return;
1716  }
1717  if((vard+128)>>8 < c->avctx->mb_threshold){
1718  type= s->mb_type[mb_y*s->mb_stride + mb_x];
1719  if(type == CANDIDATE_MB_TYPE_DIRECT){
1720  direct_search(s, mb_x, mb_y);
1721  }
1722  if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1723  c->skip=0;
1724  ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1725  }
1726  if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1727  c->skip=0;
1728  ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1729  }
1730  if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1731  c->skip=0;
1732  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1733  interlaced_search(s, 0,
1734  s->b_field_mv_table[0], s->b_field_select_table[0],
1735  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1736  }
1737  if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1738  c->skip=0;
1739  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1740  interlaced_search(s, 2,
1741  s->b_field_mv_table[1], s->b_field_select_table[1],
1742  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1743  }
1744  return;
1745  }
1746  }
1747 
1748  if (s->codec_id == AV_CODEC_ID_MPEG4)
1749  dmin= direct_search(s, mb_x, mb_y);
1750  else
1751  dmin= INT_MAX;
1752 //FIXME penalty stuff for non mpeg4
1753  c->skip=0;
1754  fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1755 
1756  c->skip=0;
1757  bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1758  av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1759 
1760  c->skip=0;
1761  fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1762  av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1763 
1764  if(s->flags & CODEC_FLAG_INTERLACED_ME){
1765 //FIXME mb type penalty
1766  c->skip=0;
1767  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1768  fimin= interlaced_search(s, 0,
1769  s->b_field_mv_table[0], s->b_field_select_table[0],
1770  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1771  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1772  bimin= interlaced_search(s, 2,
1773  s->b_field_mv_table[1], s->b_field_select_table[1],
1774  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1775  }else
1776  fimin= bimin= INT_MAX;
1777 
1778  {
1779  int score= fmin;
1780  type = CANDIDATE_MB_TYPE_FORWARD;
1781 
1782  if (dmin <= score){
1783  score = dmin;
1784  type = CANDIDATE_MB_TYPE_DIRECT;
1785  }
1786  if(bmin<score){
1787  score=bmin;
1788  type= CANDIDATE_MB_TYPE_BACKWARD;
1789  }
1790  if(fbmin<score){
1791  score=fbmin;
1792  type= CANDIDATE_MB_TYPE_BIDIR;
1793  }
1794  if(fimin<score){
1795  score=fimin;
1796  type= CANDIDATE_MB_TYPE_FORWARD_I;
1797  }
1798  if(bimin<score){
1799  score=bimin;
1800  type= CANDIDATE_MB_TYPE_BACKWARD_I;
1801  }
1802 
1803  score= ((unsigned)(score*score + 128*256))>>16;
1804  c->mc_mb_var_sum_temp += score;
1805  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1806  }
1807 
1808  if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1809  type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1810  if(fimin < INT_MAX)
1811  type |= CANDIDATE_MB_TYPE_FORWARD_I;
1812  if(bimin < INT_MAX)
1813  type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1814  if(fimin < INT_MAX && bimin < INT_MAX){
1815  type |= CANDIDATE_MB_TYPE_BIDIR_I;
1816  }
1817  //FIXME something smarter
1818  if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1819  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])
1820  type |= CANDIDATE_MB_TYPE_DIRECT0;
1821  }
1822 
1823  s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1824 }
1825 
1826 /* find best f_code for ME which do unlimited searches */
1827 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1828 {
1829  if(s->me_method>=ME_EPZS){
1830  int score[8];
1831  int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1832  uint8_t * fcode_tab= s->fcode_tab;
1833  int best_fcode=-1;
1834  int best_score=-10000000;
1835 
1836  if(s->msmpeg4_version)
1837  range= FFMIN(range, 16);
1838  else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1839  range= FFMIN(range, 256);
1840 
1841  for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1842 
1843  for(y=0; y<s->mb_height; y++){
1844  int x;
1845  int xy= y*s->mb_stride;
1846  for(x=0; x<s->mb_width; x++){
1847  if(s->mb_type[xy] & type){
1848  int mx= mv_table[xy][0];
1849  int my= mv_table[xy][1];
1850  int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1851  fcode_tab[my + MAX_MV]);
1852  int j;
1853 
1854  if(mx >= range || mx < -range ||
1855  my >= range || my < -range)
1856  continue;
1857 
1858  for(j=0; j<fcode && j<8; j++){
1859  if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1860  score[j]-= 170;
1861  }
1862  }
1863  xy++;
1864  }
1865  }
1866 
1867  for(i=1; i<8; i++){
1868  if(score[i] > best_score){
1869  best_score= score[i];
1870  best_fcode= i;
1871  }
1872  }
1873 
1874  return best_fcode;
1875  }else{
1876  return 1;
1877  }
1878 }
1879 
1880 void ff_fix_long_p_mvs(MpegEncContext * s)
1881 {
1882  MotionEstContext * const c= &s->me;
1883  const int f_code= s->f_code;
1884  int y, range;
1885  av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1886 
1887  range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1888 
1889  av_assert0(range <= 16 || !s->msmpeg4_version);
1890  av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1891 
1892  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1893 
1894  if(s->flags&CODEC_FLAG_4MV){
1895  const int wrap= s->b8_stride;
1896 
1897  /* clip / convert to intra 8x8 type MVs */
1898  for(y=0; y<s->mb_height; y++){
1899  int xy= y*2*wrap;
1900  int i= y*s->mb_stride;
1901  int x;
1902 
1903  for(x=0; x<s->mb_width; x++){
1904  if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1905  int block;
1906  for(block=0; block<4; block++){
1907  int off= (block& 1) + (block>>1)*wrap;
1908  int mx = s->current_picture.f.motion_val[0][ xy + off ][0];
1909  int my = s->current_picture.f.motion_val[0][ xy + off ][1];
1910 
1911  if( mx >=range || mx <-range
1912  || my >=range || my <-range){
1913  s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1914  s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1915  s->current_picture.f.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1916  }
1917  }
1918  }
1919  xy+=2;
1920  i++;
1921  }
1922  }
1923  }
1924 }
1925 
1926 /**
1927  *
1928  * @param truncate 1 for truncation, 0 for using intra
1929  */
1930 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1931  int16_t (*mv_table)[2], int f_code, int type, int truncate)
1932 {
1933  MotionEstContext * const c= &s->me;
1934  int y, h_range, v_range;
1935 
1936  // RAL: 8 in MPEG-1, 16 in MPEG-4
1937  int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1938 
1939  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1940 
1941  h_range= range;
1942  v_range= field_select_table ? range>>1 : range;
1943 
1944  /* clip / convert to intra 16x16 type MVs */
1945  for(y=0; y<s->mb_height; y++){
1946  int x;
1947  int xy= y*s->mb_stride;
1948  for(x=0; x<s->mb_width; x++){
1949  if (s->mb_type[xy] & type){ // RAL: "type" test added...
1950  if(field_select_table==NULL || field_select_table[xy] == field_select){
1951  if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1952  || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1953 
1954  if(truncate){
1955  if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1956  else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1957  if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1958  else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1959  }else{
1960  s->mb_type[xy] &= ~type;
1961  s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1962  mv_table[xy][0]=
1963  mv_table[xy][1]= 0;
1964  }
1965  }
1966  }
1967  }
1968  xy++;
1969  }
1970  }
1971 }
Generated on Sat May 25 2013 04:01:07 for FFmpeg by   doxygen 1.8.2