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 "dsputil.h"
36 #include "mathops.h"
37 #include "mpegvideo.h"
38 
39 #undef NDEBUG
40 #include <assert.h>
41 
42 #define P_LEFT P[1]
43 #define P_TOP P[2]
44 #define P_TOPRIGHT P[3]
45 #define P_MEDIAN P[4]
46 #define P_MV1 P[9]
47 
48 static int sad_hpel_motion_search(MpegEncContext * s,
49  int *mx_ptr, int *my_ptr, int dmin,
50  int src_index, int ref_index,
51  int size, int h);
52 
53 static inline unsigned update_map_generation(MotionEstContext *c)
54 {
55  c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
56  if(c->map_generation==0){
57  c->map_generation= 1<<(ME_MAP_MV_BITS*2);
58  memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59  }
60  return c->map_generation;
61 }
62 
63 /* shape adaptive search stuff */
64 typedef struct Minima{
65  int height;
66  int x, y;
67  int checked;
68 }Minima;
69 
70 static int minima_cmp(const void *a, const void *b){
71  const Minima *da = (const Minima *) a;
72  const Minima *db = (const Minima *) b;
73 
74  return da->height - db->height;
75 }
76 
77 #define FLAG_QPEL 1 //must be 1
78 #define FLAG_CHROMA 2
79 #define FLAG_DIRECT 4
80 
81 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
82  const int offset[3]= {
83  y*c-> stride + x,
84  ((y*c->uvstride + x)>>1),
85  ((y*c->uvstride + x)>>1),
86  };
87  int i;
88  for(i=0; i<3; i++){
89  c->src[0][i]= src [i] + offset[i];
90  c->ref[0][i]= ref [i] + offset[i];
91  }
92  if(ref_index){
93  for(i=0; i<3; i++){
94  c->ref[ref_index][i]= ref2[i] + offset[i];
95  }
96  }
97 }
98 
99 static int get_flags(MotionEstContext *c, int direct, int chroma){
100  return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
101  + (direct ? FLAG_DIRECT : 0)
102  + (chroma ? FLAG_CHROMA : 0);
103 }
104 
105 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
106  const int size, const int h, int ref_index, int src_index,
107  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
108  MotionEstContext * const c= &s->me;
109  const int stride= c->stride;
110  const int hx= subx + (x<<(1+qpel));
111  const int hy= suby + (y<<(1+qpel));
112  uint8_t * const * const ref= c->ref[ref_index];
113  uint8_t * const * const src= c->src[src_index];
114  int d;
115  //FIXME check chroma 4mv, (no crashes ...)
116  av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
117  if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
118  const int time_pp= s->pp_time;
119  const int time_pb= s->pb_time;
120  const int mask= 2*qpel+1;
121  if(s->mv_type==MV_TYPE_8X8){
122  int i;
123  for(i=0; i<4; i++){
124  int fx = c->direct_basis_mv[i][0] + hx;
125  int fy = c->direct_basis_mv[i][1] + hy;
126  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));
127  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));
128  int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
129  int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
130 
131  uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
132  if(qpel){
133  c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
134  c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
135  }else{
136  c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
137  c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
138  }
139  }
140  }else{
141  int fx = c->direct_basis_mv[0][0] + hx;
142  int fy = c->direct_basis_mv[0][1] + hy;
143  int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
144  int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
145  int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
146  int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
147 
148  if(qpel){
149  c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
150  c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
151  c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
152  c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
153  c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
154  c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
155  c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
156  c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
157  }else{
158  av_assert2((fx>>1) + 16*s->mb_x >= -16);
159  av_assert2((fy>>1) + 16*s->mb_y >= -16);
160  av_assert2((fx>>1) + 16*s->mb_x <= s->width);
161  av_assert2((fy>>1) + 16*s->mb_y <= s->height);
162  av_assert2((bx>>1) + 16*s->mb_x >= -16);
163  av_assert2((by>>1) + 16*s->mb_y >= -16);
164  av_assert2((bx>>1) + 16*s->mb_x <= s->width);
165  av_assert2((by>>1) + 16*s->mb_y <= s->height);
166 
167  c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
168  c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
169  }
170  }
171  d = cmp_func(s, c->temp, src[0], stride, 16);
172  }else
173  d= 256*256*256*32;
174  return d;
175 }
176 
177 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
178  const int size, const int h, int ref_index, int src_index,
179  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
180  MotionEstContext * const c= &s->me;
181  const int stride= c->stride;
182  const int uvstride= c->uvstride;
183  const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
184  const int hx= subx + (x<<(1+qpel));
185  const int hy= suby + (y<<(1+qpel));
186  uint8_t * const * const ref= c->ref[ref_index];
187  uint8_t * const * const src= c->src[src_index];
188  int d;
189  //FIXME check chroma 4mv, (no crashes ...)
190  int uvdxy; /* no, it might not be used uninitialized */
191  if(dxy){
192  if(qpel){
193  c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
194  if(chroma){
195  int cx= hx/2;
196  int cy= hy/2;
197  cx= (cx>>1)|(cx&1);
198  cy= (cy>>1)|(cy&1);
199  uvdxy= (cx&1) + 2*(cy&1);
200  //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
201  }
202  }else{
203  c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
204  if(chroma)
205  uvdxy= dxy | (x&1) | (2*(y&1));
206  }
207  d = cmp_func(s, c->temp, src[0], stride, h);
208  }else{
209  d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
210  if(chroma)
211  uvdxy= (x&1) + 2*(y&1);
212  }
213  if(chroma){
214  uint8_t * const uvtemp= c->temp + 16*stride;
215  c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
216  c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
217  d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
218  d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
219  }
220  return d;
221 }
222 
223 static int cmp_simple(MpegEncContext *s, const int x, const int y,
224  int ref_index, int src_index,
225  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
226  return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
227 }
228 
229 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
230  const int size, const int h, int ref_index, int src_index,
231  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
232  if(flags&FLAG_DIRECT){
233  return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
234  }else{
235  return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
236  }
237 }
238 
239 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
240  const int size, const int h, int ref_index, int src_index,
241  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
242  if(flags&FLAG_DIRECT){
243  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
244  }else{
245  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);
246  }
247 }
248 
249 /** @brief compares a block (either a full macroblock or a partition thereof)
250  against a proposed motion-compensated prediction of that block
251  */
252 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
253  const int size, const int h, int ref_index, int src_index,
254  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
255  if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
256  && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
257  && flags==0 && h==16 && size==0 && subx==0 && suby==0){
258  return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
259  }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
260  && subx==0 && suby==0){
261  return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
262  }else{
263  return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
264  }
265 }
266 
267 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
268  const int size, const int h, int ref_index, int src_index,
269  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
270  if(flags&FLAG_DIRECT){
271  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
272  }else{
273  return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
274  }
275 }
276 
277 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
278  const int size, const int h, int ref_index, int src_index,
279  me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
280  if(flags&FLAG_DIRECT){
281  return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
282  }else{
283  return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
284  }
285 }
286 
287 #include "motion_est_template.c"
288 
289 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
290  return 0;
291 }
292 
293 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
294 }
295 
296 int ff_init_me(MpegEncContext *s){
297  MotionEstContext * const c= &s->me;
298  int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
299  int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
300 
301  if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
302  av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
303  return -1;
304  }
305  //special case of snow is needed because snow uses its own iterative ME code
306  if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
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 
317  ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
318  ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
319  ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_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);
323  c->sub_flags= get_flags(c, 0, c->avctx->me_sub_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){
328  c->sub_motion_search= qpel_motion_search;
329  c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
330  if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
331  else c->qpel_put= s->dsp.put_qpel_pixels_tab;
332  }else{
333  if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
334  c->sub_motion_search= hpel_motion_search;
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
340  c->sub_motion_search= hpel_motion_search;
341  }
342  c->hpel_avg= s->dsp.avg_pixels_tab;
343  if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
344  else c->hpel_put= s->dsp.put_pixels_tab;
345 
346  if(s->linesize){
347  c->stride = s->linesize;
348  c->uvstride= s->uvlinesize;
349  }else{
350  c->stride = 16*s->mb_width + 32;
351  c->uvstride= 8*s->mb_width + 16;
352  }
353 
354  /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
355  * not have yet, and even if we had, the motion estimation code
356  * does not expect it. */
357  if(s->codec_id != AV_CODEC_ID_SNOW){
358  if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
359  s->dsp.me_cmp[2]= zero_cmp;
360  }
361  if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
362  s->dsp.me_sub_cmp[2]= zero_cmp;
363  }
364  c->hpel_put[2][0]= c->hpel_put[2][1]=
365  c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
366  }
367 
368  if(s->codec_id == AV_CODEC_ID_H261){
369  c->sub_motion_search= no_sub_motion_search;
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 
382 static int sad_hpel_motion_search(MpegEncContext * s,
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;
393  LOAD_COMMON
394 
395  av_assert2(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.f.motion_val[0][mot_xy ][0] = mx;
500  s->current_picture.f.motion_val[0][mot_xy ][1] = my;
501  s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
502  s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
503 
504  mot_xy += s->b8_stride;
505  s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
506  s->current_picture.f.motion_val[0][mot_xy ][1] = my;
507  s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
508  s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
509  }
510 }
511 
512 /**
513  * get fullpel ME search limits.
514  */
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->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){
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, i;
567  int same=1;
568  const int stride= c->stride;
569  uint8_t *mv_penalty= c->current_mv_penalty;
570  int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
571 
572  init_mv4_ref(c);
573 
574  for(block=0; block<4; block++){
575  int mx4, my4;
576  int pred_x4, pred_y4;
577  int dmin4;
578  static const int off[4]= {2, 1, 1, -1};
579  const int mot_stride = s->b8_stride;
580  const int mot_xy = s->block_index[block];
581 
582  if(saftey_cliping){
583  c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
584  c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
585  }
586 
587  P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
588  P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
589 
590  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
591 
592  /* special case for first line */
593  if (s->first_slice_line && block<2) {
594  c->pred_x= pred_x4= P_LEFT[0];
595  c->pred_y= pred_y4= P_LEFT[1];
596  } else {
597  P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
598  P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
599  P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][0];
600  P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][1];
601  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
602  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
603  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
604  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
605 
606  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
607  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
608 
609  c->pred_x= pred_x4 = P_MEDIAN[0];
610  c->pred_y= pred_y4 = P_MEDIAN[1];
611  }
612  P_MV1[0]= mx;
613  P_MV1[1]= my;
614  if(saftey_cliping)
615  for(i=0; i<10; i++){
616  if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
617  if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
618  }
619 
620  dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
621 
622  dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
623 
624  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
625  int dxy;
626  const int offset= ((block&1) + (block>>1)*stride)*8;
627  uint8_t *dest_y = c->scratchpad + offset;
628  if(s->quarter_sample){
629  uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
630  dxy = ((my4 & 3) << 2) | (mx4 & 3);
631 
632  if(s->no_rounding)
633  s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
634  else
635  s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
636  }else{
637  uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
638  dxy = ((my4 & 1) << 1) | (mx4 & 1);
639 
640  if(s->no_rounding)
641  s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
642  else
643  s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
644  }
645  dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
646  }else
647  dmin_sum+= dmin4;
648 
649  if(s->quarter_sample){
650  mx4_sum+= mx4/2;
651  my4_sum+= my4/2;
652  }else{
653  mx4_sum+= mx4;
654  my4_sum+= my4;
655  }
656 
657  s->current_picture.f.motion_val[0][s->block_index[block]][0] = mx4;
658  s->current_picture.f.motion_val[0][s->block_index[block]][1] = my4;
659 
660  if(mx4 != mx || my4 != my) same=0;
661  }
662 
663  if(same)
664  return INT_MAX;
665 
666  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
667  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);
668  }
669 
670  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
671  int dxy;
672  int mx, my;
673  int offset;
674 
675  mx= ff_h263_round_chroma(mx4_sum);
676  my= ff_h263_round_chroma(my4_sum);
677  dxy = ((my & 1) << 1) | (mx & 1);
678 
679  offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
680 
681  if(s->no_rounding){
682  s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
683  s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
684  }else{
685  s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
686  s->dsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
687  }
688 
689  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);
690  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);
691  }
692 
693  c->pred_x= mx;
694  c->pred_y= my;
695 
696  switch(c->avctx->mb_cmp&0xFF){
697  /*case FF_CMP_SSE:
698  return dmin_sum+ 32*s->qscale*s->qscale;*/
699  case FF_CMP_RD:
700  return dmin_sum;
701  default:
702  return dmin_sum+ 11*c->mb_penalty_factor;
703  }
704 }
705 
706 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
707  MotionEstContext * const c= &s->me;
708 
709  c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
710  c->src[1][0] = c->src[0][0] + s->linesize;
711  if(c->flags & FLAG_CHROMA){
712  c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
713  c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
714  c->src[1][1] = c->src[0][1] + s->uvlinesize;
715  c->src[1][2] = c->src[0][2] + s->uvlinesize;
716  }
717 }
718 
719 static int interlaced_search(MpegEncContext *s, int ref_index,
720  int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
721 {
722  MotionEstContext * const c= &s->me;
723  const int size=0;
724  const int h=8;
725  int block;
726  int P[10][2];
727  uint8_t * const mv_penalty= c->current_mv_penalty;
728  int same=1;
729  const int stride= 2*s->linesize;
730  int dmin_sum= 0;
731  const int mot_stride= s->mb_stride;
732  const int xy= s->mb_x + s->mb_y*mot_stride;
733 
734  c->ymin>>=1;
735  c->ymax>>=1;
736  c->stride<<=1;
737  c->uvstride<<=1;
738  init_interlaced_ref(s, ref_index);
739 
740  for(block=0; block<2; block++){
741  int field_select;
742  int best_dmin= INT_MAX;
743  int best_field= -1;
744 
745  for(field_select=0; field_select<2; field_select++){
746  int dmin, mx_i, my_i;
747  int16_t (*mv_table)[2]= mv_tables[block][field_select];
748 
749  if(user_field_select){
750  av_assert1(field_select==0 || field_select==1);
751  av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
752  if(field_select_tables[block][xy] != field_select)
753  continue;
754  }
755 
756  P_LEFT[0] = mv_table[xy - 1][0];
757  P_LEFT[1] = mv_table[xy - 1][1];
758  if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
759 
760  c->pred_x= P_LEFT[0];
761  c->pred_y= P_LEFT[1];
762 
763  if(!s->first_slice_line){
764  P_TOP[0] = mv_table[xy - mot_stride][0];
765  P_TOP[1] = mv_table[xy - mot_stride][1];
766  P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
767  P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
768  if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
769  if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
770  if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
771  if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
772 
773  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
774  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
775  }
776  P_MV1[0]= mx; //FIXME not correct if block != field_select
777  P_MV1[1]= my / 2;
778 
779  dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
780 
781  dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
782 
783  mv_table[xy][0]= mx_i;
784  mv_table[xy][1]= my_i;
785 
786  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
787  int dxy;
788 
789  //FIXME chroma ME
790  uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
791  dxy = ((my_i & 1) << 1) | (mx_i & 1);
792 
793  if(s->no_rounding){
794  s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
795  }else{
796  s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
797  }
798  dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
799  dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
800  }else
801  dmin+= c->mb_penalty_factor; //field_select bits
802 
803  dmin += field_select != block; //slightly prefer same field
804 
805  if(dmin < best_dmin){
806  best_dmin= dmin;
807  best_field= field_select;
808  }
809  }
810  {
811  int16_t (*mv_table)[2]= mv_tables[block][best_field];
812 
813  if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
814  if(mv_table[xy][1]&1) same=0;
815  if(mv_table[xy][1]*2 != my) same=0;
816  if(best_field != block) same=0;
817  }
818 
819  field_select_tables[block][xy]= best_field;
820  dmin_sum += best_dmin;
821  }
822 
823  c->ymin<<=1;
824  c->ymax<<=1;
825  c->stride>>=1;
826  c->uvstride>>=1;
827 
828  if(same)
829  return INT_MAX;
830 
831  switch(c->avctx->mb_cmp&0xFF){
832  /*case FF_CMP_SSE:
833  return dmin_sum+ 32*s->qscale*s->qscale;*/
834  case FF_CMP_RD:
835  return dmin_sum;
836  default:
837  return dmin_sum+ 11*c->mb_penalty_factor;
838  }
839 }
840 
841 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
842  int ymax= s->me.ymax>>interlaced;
843  int ymin= s->me.ymin>>interlaced;
844 
845  if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
846  if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
847  if(mv[1] < ymin) mv[1] = ymin;
848  if(mv[1] > ymax) mv[1] = ymax;
849 }
850 
851 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
852  MotionEstContext * const c= &s->me;
853  Picture *p= s->current_picture_ptr;
854  int mb_xy= mb_x + mb_y*s->mb_stride;
855  int xy= 2*mb_x + 2*mb_y*s->b8_stride;
856  int mb_type= s->current_picture.f.mb_type[mb_xy];
857  int flags= c->flags;
858  int shift= (flags&FLAG_QPEL) + 1;
859  int mask= (1<<shift)-1;
860  int x, y, i;
861  int d=0;
862  me_cmp_func cmpf= s->dsp.sse[0];
863  me_cmp_func chroma_cmpf= s->dsp.sse[1];
864 
865  if(p_type && USES_LIST(mb_type, 1)){
866  av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
867  return INT_MAX/2;
868  }
869  av_assert0(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
870 
871  for(i=0; i<4; i++){
872  int xy= s->block_index[i];
873  clip_input_mv(s, p->f.motion_val[0][xy], !!IS_INTERLACED(mb_type));
874  clip_input_mv(s, p->f.motion_val[1][xy], !!IS_INTERLACED(mb_type));
875  }
876 
877  if(IS_INTERLACED(mb_type)){
878  int xy2= xy + s->b8_stride;
879  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
880  c->stride<<=1;
881  c->uvstride<<=1;
882 
883  if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
884  av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
885  return INT_MAX/2;
886  }
887 
888  if(USES_LIST(mb_type, 0)){
889  int field_select0= p->f.ref_index[0][4*mb_xy ];
890  int field_select1= p->f.ref_index[0][4*mb_xy+2];
891  av_assert0(field_select0==0 ||field_select0==1);
892  av_assert0(field_select1==0 ||field_select1==1);
893  init_interlaced_ref(s, 0);
894 
895  if(p_type){
896  s->p_field_select_table[0][mb_xy]= field_select0;
897  s->p_field_select_table[1][mb_xy]= field_select1;
898  *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
899  *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
900  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
901  }else{
902  s->b_field_select_table[0][0][mb_xy]= field_select0;
903  s->b_field_select_table[0][1][mb_xy]= field_select1;
904  *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
905  *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
906  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
907  }
908 
909  x = p->f.motion_val[0][xy ][0];
910  y = p->f.motion_val[0][xy ][1];
911  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
912  x = p->f.motion_val[0][xy2][0];
913  y = p->f.motion_val[0][xy2][1];
914  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
915  }
916  if(USES_LIST(mb_type, 1)){
917  int field_select0 = p->f.ref_index[1][4 * mb_xy ];
918  int field_select1 = p->f.ref_index[1][4 * mb_xy + 2];
919  av_assert0(field_select0==0 ||field_select0==1);
920  av_assert0(field_select1==0 ||field_select1==1);
921  init_interlaced_ref(s, 2);
922 
923  s->b_field_select_table[1][0][mb_xy]= field_select0;
924  s->b_field_select_table[1][1][mb_xy]= field_select1;
925  *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy ];
926  *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy2];
927  if(USES_LIST(mb_type, 0)){
928  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
929  }else{
930  s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
931  }
932 
933  x = p->f.motion_val[1][xy ][0];
934  y = p->f.motion_val[1][xy ][1];
935  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
936  x = p->f.motion_val[1][xy2][0];
937  y = p->f.motion_val[1][xy2][1];
938  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
939  //FIXME bidir scores
940  }
941  c->stride>>=1;
942  c->uvstride>>=1;
943  }else if(IS_8X8(mb_type)){
944  if(!(s->flags & CODEC_FLAG_4MV)){
945  av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
946  return INT_MAX/2;
947  }
948  cmpf= s->dsp.sse[1];
949  chroma_cmpf= s->dsp.sse[1];
950  init_mv4_ref(c);
951  for(i=0; i<4; i++){
952  xy= s->block_index[i];
953  x= p->f.motion_val[0][xy][0];
954  y= p->f.motion_val[0][xy][1];
955  d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
956  }
957  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
958  }else{
959  if(USES_LIST(mb_type, 0)){
960  if(p_type){
961  *(uint32_t*)s->p_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
962  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
963  }else if(USES_LIST(mb_type, 1)){
964  *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
965  *(uint32_t*)s->b_bidir_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
966  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
967  }else{
968  *(uint32_t*)s->b_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
969  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
970  }
971  x = p->f.motion_val[0][xy][0];
972  y = p->f.motion_val[0][xy][1];
973  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
974  }else if(USES_LIST(mb_type, 1)){
975  *(uint32_t*)s->b_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
976  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
977 
978  x = p->f.motion_val[1][xy][0];
979  y = p->f.motion_val[1][xy][1];
980  d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
981  }else
982  s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
983  }
984  return d;
985 }
986 
987 void ff_estimate_p_frame_motion(MpegEncContext * s,
988  int mb_x, int mb_y)
989 {
990  MotionEstContext * const c= &s->me;
991  uint8_t *pix, *ppix;
992  int sum, mx, my, dmin;
993  int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
994  int vard; ///< sum of squared differences with the estimated motion vector
995  int P[10][2];
996  const int shift= 1+s->quarter_sample;
997  int mb_type=0;
998  Picture * const pic= &s->current_picture;
999 
1000  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1001 
1002  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1003  av_assert0(s->linesize == c->stride);
1004  av_assert0(s->uvlinesize == c->uvstride);
1005 
1006  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1007  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1008  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1009  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1010 
1011  get_limits(s, 16*mb_x, 16*mb_y);
1012  c->skip=0;
1013 
1014  /* intra / predictive decision */
1015  pix = c->src[0][0];
1016  sum = s->dsp.pix_sum(pix, s->linesize);
1017  varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
1018 
1019  pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1020  pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1021  c->mb_var_sum_temp += (varc+128)>>8;
1022 
1023  if(c->avctx->me_threshold){
1024  vard= check_input_motion(s, mb_x, mb_y, 1);
1025 
1026  if((vard+128)>>8 < c->avctx->me_threshold){
1027  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1028  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1029  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1030  c->mc_mb_var_sum_temp += (vard+128)>>8;
1031  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1032  return;
1033  }
1034  if((vard+128)>>8 < c->avctx->mb_threshold)
1035  mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1036  }
1037 
1038  switch(s->me_method) {
1039  case ME_ZERO:
1040  default:
1041  mx = 0;
1042  my = 0;
1043  dmin = 0;
1044  break;
1045  case ME_X1:
1046  case ME_EPZS:
1047  {
1048  const int mot_stride = s->b8_stride;
1049  const int mot_xy = s->block_index[0];
1050 
1051  P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
1052  P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
1053 
1054  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1055 
1056  if(!s->first_slice_line) {
1057  P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
1058  P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
1059  P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][0];
1060  P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][1];
1061  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1062  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1063  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1064 
1065  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1066  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1067 
1068  if(s->out_format == FMT_H263){
1069  c->pred_x = P_MEDIAN[0];
1070  c->pred_y = P_MEDIAN[1];
1071  }else { /* mpeg1 at least */
1072  c->pred_x= P_LEFT[0];
1073  c->pred_y= P_LEFT[1];
1074  }
1075  }else{
1076  c->pred_x= P_LEFT[0];
1077  c->pred_y= P_LEFT[1];
1078  }
1079 
1080  }
1081  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1082 
1083  break;
1084  }
1085 
1086  /* At this point (mx,my) are full-pell and the relative displacement */
1087  ppix = c->ref[0][0] + (my * s->linesize) + mx;
1088 
1089  vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1090 
1091  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1092 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1093  c->mc_mb_var_sum_temp += (vard+128)>>8;
1094 
1095  if(mb_type){
1096  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1097  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1098  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1099 
1100  if(mb_type == CANDIDATE_MB_TYPE_INTER){
1101  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1102  set_p_mv_tables(s, mx, my, 1);
1103  }else{
1104  mx <<=shift;
1105  my <<=shift;
1106  }
1107  if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1108  h263_mv4_search(s, mx, my, shift);
1109 
1110  set_p_mv_tables(s, mx, my, 0);
1111  }
1112  if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1113  interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1114  }
1115  }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1116  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1117  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1118  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1119 
1120  if (vard*2 + 200*256 > varc)
1121  mb_type|= CANDIDATE_MB_TYPE_INTRA;
1122  if (varc*2 + 200*256 > vard || s->qscale > 24){
1123 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1124  mb_type|= CANDIDATE_MB_TYPE_INTER;
1125  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1126  if(s->flags&CODEC_FLAG_MV0)
1127  if(mx || my)
1128  mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1129  }else{
1130  mx <<=shift;
1131  my <<=shift;
1132  }
1133  if((s->flags&CODEC_FLAG_4MV)
1134  && !c->skip && varc>50<<8 && vard>10<<8){
1135  if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1136  mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1137 
1138  set_p_mv_tables(s, mx, my, 0);
1139  }else
1140  set_p_mv_tables(s, mx, my, 1);
1141  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1142  && !c->skip){ //FIXME varc/d checks
1143  if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1144  mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1145  }
1146  }else{
1147  int intra_score, i;
1148  mb_type= CANDIDATE_MB_TYPE_INTER;
1149 
1150  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1151  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1152  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1153 
1154  if((s->flags&CODEC_FLAG_4MV)
1155  && !c->skip && varc>50<<8 && vard>10<<8){
1156  int dmin4= h263_mv4_search(s, mx, my, shift);
1157  if(dmin4 < dmin){
1158  mb_type= CANDIDATE_MB_TYPE_INTER4V;
1159  dmin=dmin4;
1160  }
1161  }
1162  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1163  && !c->skip){ //FIXME varc/d checks
1164  int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1165  if(dmin_i < dmin){
1166  mb_type = CANDIDATE_MB_TYPE_INTER_I;
1167  dmin= dmin_i;
1168  }
1169  }
1170 
1171 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1172  set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1173 
1174  /* get intra luma score */
1175  if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1176  intra_score= varc - 500;
1177  }else{
1178  unsigned mean = (sum+128)>>8;
1179  mean*= 0x01010101;
1180 
1181  for(i=0; i<16; i++){
1182  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1183  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1184  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1185  *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1186  }
1187 
1188  intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1189  }
1190  intra_score += c->mb_penalty_factor*16;
1191 
1192  if(intra_score < dmin){
1193  mb_type= CANDIDATE_MB_TYPE_INTRA;
1194  s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1195  }else
1196  s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1197 
1198  {
1199  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1200  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1201  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1202  }
1203  }
1204 
1205  s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1206 }
1207 
1208 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1209  int mb_x, int mb_y)
1210 {
1211  MotionEstContext * const c= &s->me;
1212  int mx, my, dmin;
1213  int P[10][2];
1214  const int shift= 1+s->quarter_sample;
1215  const int xy= mb_x + mb_y*s->mb_stride;
1216  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1217 
1218  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1219 
1220  c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1221  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1222 
1223  get_limits(s, 16*mb_x, 16*mb_y);
1224  c->skip=0;
1225 
1226  P_LEFT[0] = s->p_mv_table[xy + 1][0];
1227  P_LEFT[1] = s->p_mv_table[xy + 1][1];
1228 
1229  if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1230 
1231  /* special case for first line */
1232  if (s->first_slice_line) {
1233  c->pred_x= P_LEFT[0];
1234  c->pred_y= P_LEFT[1];
1235  P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1236  P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1237  } else {
1238  P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1239  P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1240  P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1241  P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1242  if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1243  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1244  if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1245 
1246  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1247  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1248 
1249  c->pred_x = P_MEDIAN[0];
1250  c->pred_y = P_MEDIAN[1];
1251  }
1252 
1253  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1254 
1255  s->p_mv_table[xy][0] = mx<<shift;
1256  s->p_mv_table[xy][1] = my<<shift;
1257 
1258  return dmin;
1259 }
1260 
1261 static int ff_estimate_motion_b(MpegEncContext * s,
1262  int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1263 {
1264  MotionEstContext * const c= &s->me;
1265  int mx, my, dmin;
1266  int P[10][2];
1267  const int shift= 1+s->quarter_sample;
1268  const int mot_stride = s->mb_stride;
1269  const int mot_xy = mb_y*mot_stride + mb_x;
1270  uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1271  int mv_scale;
1272 
1273  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1274  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1275  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1276  c->current_mv_penalty= mv_penalty;
1277 
1278  get_limits(s, 16*mb_x, 16*mb_y);
1279 
1280  switch(s->me_method) {
1281  case ME_ZERO:
1282  default:
1283  mx = 0;
1284  my = 0;
1285  dmin = 0;
1286  break;
1287  case ME_X1:
1288  case ME_EPZS:
1289  P_LEFT[0] = mv_table[mot_xy - 1][0];
1290  P_LEFT[1] = mv_table[mot_xy - 1][1];
1291 
1292  if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1293 
1294  /* special case for first line */
1295  if (!s->first_slice_line) {
1296  P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1297  P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1298  P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1299  P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1300  if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1301  if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1302  if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1303 
1304  P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1305  P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1306  }
1307  c->pred_x = P_LEFT[0];
1308  c->pred_y = P_LEFT[1];
1309 
1310  if(mv_table == s->b_forw_mv_table){
1311  mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1312  }else{
1313  mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1314  }
1315 
1316  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1317 
1318  break;
1319  }
1320 
1321  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1322 
1323  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1324  dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1325 
1326 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1327  mv_table[mot_xy][0]= mx;
1328  mv_table[mot_xy][1]= my;
1329 
1330  return dmin;
1331 }
1332 
1333 static inline int check_bidir_mv(MpegEncContext * s,
1334  int motion_fx, int motion_fy,
1335  int motion_bx, int motion_by,
1336  int pred_fx, int pred_fy,
1337  int pred_bx, int pred_by,
1338  int size, int h)
1339 {
1340  //FIXME optimize?
1341  //FIXME better f_code prediction (max mv & distance)
1342  //FIXME pointers
1343  MotionEstContext * const c= &s->me;
1344  uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1345  uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1346  int stride= c->stride;
1347  uint8_t *dest_y = c->scratchpad;
1348  uint8_t *ptr;
1349  int dxy;
1350  int src_x, src_y;
1351  int fbmin;
1352  uint8_t **src_data= c->src[0];
1353  uint8_t **ref_data= c->ref[0];
1354  uint8_t **ref2_data= c->ref[2];
1355 
1356  if(s->quarter_sample){
1357  dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1358  src_x = motion_fx >> 2;
1359  src_y = motion_fy >> 2;
1360 
1361  ptr = ref_data[0] + (src_y * stride) + src_x;
1362  s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1363 
1364  dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1365  src_x = motion_bx >> 2;
1366  src_y = motion_by >> 2;
1367 
1368  ptr = ref2_data[0] + (src_y * stride) + src_x;
1369  s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1370  }else{
1371  dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1372  src_x = motion_fx >> 1;
1373  src_y = motion_fy >> 1;
1374 
1375  ptr = ref_data[0] + (src_y * stride) + src_x;
1376  s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1377 
1378  dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1379  src_x = motion_bx >> 1;
1380  src_y = motion_by >> 1;
1381 
1382  ptr = ref2_data[0] + (src_y * stride) + src_x;
1383  s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1384  }
1385 
1386  fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1387  +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1388  + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1389 
1390  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1391  }
1392  //FIXME CHROMA !!!
1393 
1394  return fbmin;
1395 }
1396 
1397 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1398 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1399 {
1400  MotionEstContext * const c= &s->me;
1401  const int mot_stride = s->mb_stride;
1402  const int xy = mb_y *mot_stride + mb_x;
1403  int fbmin;
1404  int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1405  int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1406  int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1407  int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1408  int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1409  int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1410  int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1411  int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1412  const int flags= c->sub_flags;
1413  const int qpel= flags&FLAG_QPEL;
1414  const int shift= 1+qpel;
1415  const int xmin= c->xmin<<shift;
1416  const int ymin= c->ymin<<shift;
1417  const int xmax= c->xmax<<shift;
1418  const int ymax= c->ymax<<shift;
1419 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1420 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1421  int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1422  uint8_t map[256] = { 0 };
1423 
1424  map[hashidx&255] = 1;
1425 
1426  fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1427  motion_bx, motion_by,
1428  pred_fx, pred_fy,
1429  pred_bx, pred_by,
1430  0, 16);
1431 
1432  if(s->avctx->bidir_refine){
1433  int end;
1434  static const uint8_t limittab[5]={0,8,32,64,80};
1435  const int limit= limittab[s->avctx->bidir_refine];
1436  static const int8_t vect[][4]={
1437 { 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},
1438 
1439 { 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},
1440 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1441 { 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},
1442 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1443 
1444 { 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},
1445 { 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},
1446 { 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},
1447 { 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},
1448 
1449 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1450 { 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},
1451 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1452  };
1453  static const uint8_t hash[]={
1454 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),
1455 
1456 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),
1457 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1458 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),
1459 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1460 
1461 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),
1462 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),
1463 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),
1464 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),
1465 
1466 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1467 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),
1468 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),
1469 };
1470 
1471 #define CHECK_BIDIR(fx,fy,bx,by)\
1472  if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1473  &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1474  &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1475  int score;\
1476  map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1477  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);\
1478  if(score < fbmin){\
1479  hashidx += HASH(fx,fy,bx,by);\
1480  fbmin= score;\
1481  motion_fx+=fx;\
1482  motion_fy+=fy;\
1483  motion_bx+=bx;\
1484  motion_by+=by;\
1485  end=0;\
1486  }\
1487  }
1488 #define CHECK_BIDIR2(a,b,c,d)\
1489 CHECK_BIDIR(a,b,c,d)\
1490 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1491 
1492  do{
1493  int i;
1494  int borderdist=0;
1495  end=1;
1496 
1497  CHECK_BIDIR2(0,0,0,1)
1498  CHECK_BIDIR2(0,0,1,0)
1499  CHECK_BIDIR2(0,1,0,0)
1500  CHECK_BIDIR2(1,0,0,0)
1501 
1502  for(i=8; i<limit; i++){
1503  int fx= motion_fx+vect[i][0];
1504  int fy= motion_fy+vect[i][1];
1505  int bx= motion_bx+vect[i][2];
1506  int by= motion_by+vect[i][3];
1507  if(borderdist<=0){
1508  int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1509  int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1510  if((a|b) < 0)
1511  map[(hashidx+hash[i])&255] = 1;
1512  }
1513  if(!map[(hashidx+hash[i])&255]){
1514  int score;
1515  map[(hashidx+hash[i])&255] = 1;
1516  score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1517  if(score < fbmin){
1518  hashidx += hash[i];
1519  fbmin= score;
1520  motion_fx=fx;
1521  motion_fy=fy;
1522  motion_bx=bx;
1523  motion_by=by;
1524  end=0;
1525  borderdist--;
1526  if(borderdist<=0){
1527  int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1528  int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1529  borderdist= FFMIN(a,b);
1530  }
1531  }
1532  }
1533  }
1534  }while(!end);
1535  }
1536 
1537  s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1538  s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1539  s->b_bidir_back_mv_table[xy][0]= motion_bx;
1540  s->b_bidir_back_mv_table[xy][1]= motion_by;
1541 
1542  return fbmin;
1543 }
1544 
1545 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1546 {
1547  MotionEstContext * const c= &s->me;
1548  int P[10][2];
1549  const int mot_stride = s->mb_stride;
1550  const int mot_xy = mb_y*mot_stride + mb_x;
1551  const int shift= 1+s->quarter_sample;
1552  int dmin, i;
1553  const int time_pp= s->pp_time;
1554  const int time_pb= s->pb_time;
1555  int mx, my, xmin, xmax, ymin, ymax;
1556  int16_t (*mv_table)[2]= s->b_direct_mv_table;
1557 
1558  c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1559  ymin= xmin=(-32)>>shift;
1560  ymax= xmax= 31>>shift;
1561 
1562  if (IS_8X8(s->next_picture.f.mb_type[mot_xy])) {
1563  s->mv_type= MV_TYPE_8X8;
1564  }else{
1565  s->mv_type= MV_TYPE_16X16;
1566  }
1567 
1568  for(i=0; i<4; i++){
1569  int index= s->block_index[i];
1570  int min, max;
1571 
1572  c->co_located_mv[i][0] = s->next_picture.f.motion_val[0][index][0];
1573  c->co_located_mv[i][1] = s->next_picture.f.motion_val[0][index][1];
1574  c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1575  c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1576 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1577 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1578 
1579  max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1580  min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1581  max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1582  min+= 16*mb_x - 1;
1583  xmax= FFMIN(xmax, s->width - max);
1584  xmin= FFMAX(xmin, - 16 - min);
1585 
1586  max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1587  min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1588  max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1589  min+= 16*mb_y - 1;
1590  ymax= FFMIN(ymax, s->height - max);
1591  ymin= FFMAX(ymin, - 16 - min);
1592 
1593  if(s->mv_type == MV_TYPE_16X16) break;
1594  }
1595 
1596  av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1597 
1598  if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1599  s->b_direct_mv_table[mot_xy][0]= 0;
1600  s->b_direct_mv_table[mot_xy][1]= 0;
1601 
1602  return 256*256*256*64;
1603  }
1604 
1605  c->xmin= xmin;
1606  c->ymin= ymin;
1607  c->xmax= xmax;
1608  c->ymax= ymax;
1609  c->flags |= FLAG_DIRECT;
1610  c->sub_flags |= FLAG_DIRECT;
1611  c->pred_x=0;
1612  c->pred_y=0;
1613 
1614  P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1615  P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1616 
1617  /* special case for first line */
1618  if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1619  P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1620  P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1621  P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1622  P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1623 
1624  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1625  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1626  }
1627 
1628  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1629  if(c->sub_flags&FLAG_QPEL)
1630  dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1631  else
1632  dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1633 
1634  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1635  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1636 
1637  get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1638 
1639  mv_table[mot_xy][0]= mx;
1640  mv_table[mot_xy][1]= my;
1641  c->flags &= ~FLAG_DIRECT;
1642  c->sub_flags &= ~FLAG_DIRECT;
1643 
1644  return dmin;
1645 }
1646 
1647 void ff_estimate_b_frame_motion(MpegEncContext * s,
1648  int mb_x, int mb_y)
1649 {
1650  MotionEstContext * const c= &s->me;
1651  const int penalty_factor= c->mb_penalty_factor;
1652  int fmin, bmin, dmin, fbmin, bimin, fimin;
1653  int type=0;
1654  const int xy = mb_y*s->mb_stride + mb_x;
1655  init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1656  s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1657 
1658  get_limits(s, 16*mb_x, 16*mb_y);
1659 
1660  c->skip=0;
1661 
1662  if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.f.mbskip_table[xy]) {
1663  int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1664 
1665  score= ((unsigned)(score*score + 128*256))>>16;
1666  c->mc_mb_var_sum_temp += score;
1667  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1668  s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1669 
1670  return;
1671  }
1672 
1673  if(c->avctx->me_threshold){
1674  int vard= check_input_motion(s, mb_x, mb_y, 0);
1675 
1676  if((vard+128)>>8 < c->avctx->me_threshold){
1677 // pix = c->src[0][0];
1678 // sum = s->dsp.pix_sum(pix, s->linesize);
1679 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1680 
1681 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1682  s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1683 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1684  c->mb_var_sum_temp += (varc+128)>>8;*/
1685  c->mc_mb_var_sum_temp += (vard+128)>>8;
1686 /* if (vard <= 64<<8 || vard < varc) {
1687  c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1688  }else{
1689  c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1690  }*/
1691  return;
1692  }
1693  if((vard+128)>>8 < c->avctx->mb_threshold){
1694  type= s->mb_type[mb_y*s->mb_stride + mb_x];
1695  if(type == CANDIDATE_MB_TYPE_DIRECT){
1696  direct_search(s, mb_x, mb_y);
1697  }
1698  if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1699  c->skip=0;
1700  ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1701  }
1702  if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1703  c->skip=0;
1704  ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1705  }
1706  if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1707  c->skip=0;
1708  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1709  interlaced_search(s, 0,
1710  s->b_field_mv_table[0], s->b_field_select_table[0],
1711  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1712  }
1713  if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1714  c->skip=0;
1715  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1716  interlaced_search(s, 2,
1717  s->b_field_mv_table[1], s->b_field_select_table[1],
1718  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1719  }
1720  return;
1721  }
1722  }
1723 
1724  if (s->codec_id == AV_CODEC_ID_MPEG4)
1725  dmin= direct_search(s, mb_x, mb_y);
1726  else
1727  dmin= INT_MAX;
1728 //FIXME penalty stuff for non mpeg4
1729  c->skip=0;
1730  fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1731 
1732  c->skip=0;
1733  bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1734  av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1735 
1736  c->skip=0;
1737  fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1738  av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1739 
1740  if(s->flags & CODEC_FLAG_INTERLACED_ME){
1741 //FIXME mb type penalty
1742  c->skip=0;
1743  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1744  fimin= interlaced_search(s, 0,
1745  s->b_field_mv_table[0], s->b_field_select_table[0],
1746  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1747  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1748  bimin= interlaced_search(s, 2,
1749  s->b_field_mv_table[1], s->b_field_select_table[1],
1750  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1751  }else
1752  fimin= bimin= INT_MAX;
1753 
1754  {
1755  int score= fmin;
1756  type = CANDIDATE_MB_TYPE_FORWARD;
1757 
1758  if (dmin <= score){
1759  score = dmin;
1760  type = CANDIDATE_MB_TYPE_DIRECT;
1761  }
1762  if(bmin<score){
1763  score=bmin;
1764  type= CANDIDATE_MB_TYPE_BACKWARD;
1765  }
1766  if(fbmin<score){
1767  score=fbmin;
1768  type= CANDIDATE_MB_TYPE_BIDIR;
1769  }
1770  if(fimin<score){
1771  score=fimin;
1772  type= CANDIDATE_MB_TYPE_FORWARD_I;
1773  }
1774  if(bimin<score){
1775  score=bimin;
1776  type= CANDIDATE_MB_TYPE_BACKWARD_I;
1777  }
1778 
1779  score= ((unsigned)(score*score + 128*256))>>16;
1780  c->mc_mb_var_sum_temp += score;
1781  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1782  }
1783 
1784  if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1785  type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1786  if(fimin < INT_MAX)
1787  type |= CANDIDATE_MB_TYPE_FORWARD_I;
1788  if(bimin < INT_MAX)
1789  type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1790  if(fimin < INT_MAX && bimin < INT_MAX){
1791  type |= CANDIDATE_MB_TYPE_BIDIR_I;
1792  }
1793  //FIXME something smarter
1794  if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1795  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])
1796  type |= CANDIDATE_MB_TYPE_DIRECT0;
1797  }
1798 
1799  s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1800 }
1801 
1802 /* find best f_code for ME which do unlimited searches */
1803 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1804 {
1805  if(s->me_method>=ME_EPZS){
1806  int score[8];
1807  int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1808  uint8_t * fcode_tab= s->fcode_tab;
1809  int best_fcode=-1;
1810  int best_score=-10000000;
1811 
1812  if(s->msmpeg4_version)
1813  range= FFMIN(range, 16);
1814  else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1815  range= FFMIN(range, 256);
1816 
1817  for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1818 
1819  for(y=0; y<s->mb_height; y++){
1820  int x;
1821  int xy= y*s->mb_stride;
1822  for(x=0; x<s->mb_width; x++){
1823  if(s->mb_type[xy] & type){
1824  int mx= mv_table[xy][0];
1825  int my= mv_table[xy][1];
1826  int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1827  fcode_tab[my + MAX_MV]);
1828  int j;
1829 
1830  if(mx >= range || mx < -range ||
1831  my >= range || my < -range)
1832  continue;
1833 
1834  for(j=0; j<fcode && j<8; j++){
1835  if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1836  score[j]-= 170;
1837  }
1838  }
1839  xy++;
1840  }
1841  }
1842 
1843  for(i=1; i<8; i++){
1844  if(score[i] > best_score){
1845  best_score= score[i];
1846  best_fcode= i;
1847  }
1848  }
1849 
1850  return best_fcode;
1851  }else{
1852  return 1;
1853  }
1854 }
1855 
1856 void ff_fix_long_p_mvs(MpegEncContext * s)
1857 {
1858  MotionEstContext * const c= &s->me;
1859  const int f_code= s->f_code;
1860  int y, range;
1861  av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1862 
1863  range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1864 
1865  av_assert0(range <= 16 || !s->msmpeg4_version);
1866  av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1867 
1868  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1869 
1870  if(s->flags&CODEC_FLAG_4MV){
1871  const int wrap= s->b8_stride;
1872 
1873  /* clip / convert to intra 8x8 type MVs */
1874  for(y=0; y<s->mb_height; y++){
1875  int xy= y*2*wrap;
1876  int i= y*s->mb_stride;
1877  int x;
1878 
1879  for(x=0; x<s->mb_width; x++){
1880  if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1881  int block;
1882  for(block=0; block<4; block++){
1883  int off= (block& 1) + (block>>1)*wrap;
1884  int mx = s->current_picture.f.motion_val[0][ xy + off ][0];
1885  int my = s->current_picture.f.motion_val[0][ xy + off ][1];
1886 
1887  if( mx >=range || mx <-range
1888  || my >=range || my <-range){
1889  s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1890  s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1891  s->current_picture.f.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1892  }
1893  }
1894  }
1895  xy+=2;
1896  i++;
1897  }
1898  }
1899  }
1900 }
1901 
1902 /**
1903  *
1904  * @param truncate 1 for truncation, 0 for using intra
1905  */
1906 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1907  int16_t (*mv_table)[2], int f_code, int type, int truncate)
1908 {
1909  MotionEstContext * const c= &s->me;
1910  int y, h_range, v_range;
1911 
1912  // RAL: 8 in MPEG-1, 16 in MPEG-4
1913  int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1914 
1915  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1916 
1917  h_range= range;
1918  v_range= field_select_table ? range>>1 : range;
1919 
1920  /* clip / convert to intra 16x16 type MVs */
1921  for(y=0; y<s->mb_height; y++){
1922  int x;
1923  int xy= y*s->mb_stride;
1924  for(x=0; x<s->mb_width; x++){
1925  if (s->mb_type[xy] & type){ // RAL: "type" test added...
1926  if(field_select_table==NULL || field_select_table[xy] == field_select){
1927  if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1928  || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1929 
1930  if(truncate){
1931  if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1932  else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1933  if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1934  else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1935  }else{
1936  s->mb_type[xy] &= ~type;
1937  s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1938  mv_table[xy][0]=
1939  mv_table[xy][1]= 0;
1940  }
1941  }
1942  }
1943  }
1944  xy++;
1945  }
1946  }
1947 }
Generated on Sat May 25 2013 03:58:37 for FFmpeg by   doxygen 1.8.2