FFmpeg
 All Data Structures Namespaces 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->hdsp.avg_pixels_tab;
342  if (s->no_rounding)
343  c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
344  else
345  c->hpel_put = s->hdsp.put_pixels_tab;
346 
347  if(s->linesize){
348  c->stride = s->linesize;
349  c->uvstride= s->uvlinesize;
350  }else{
351  c->stride = 16*s->mb_width + 32;
352  c->uvstride= 8*s->mb_width + 16;
353  }
354 
355  /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
356  * not have yet, and even if we had, the motion estimation code
357  * does not expect it. */
358  if(s->codec_id != AV_CODEC_ID_SNOW){
359  if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
360  s->dsp.me_cmp[2]= zero_cmp;
361  }
362  if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
363  s->dsp.me_sub_cmp[2]= zero_cmp;
364  }
365  c->hpel_put[2][0]= c->hpel_put[2][1]=
366  c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
367  }
368 
369  if(s->codec_id == AV_CODEC_ID_H261){
370  c->sub_motion_search= no_sub_motion_search;
371  }
372 
373  return 0;
374 }
375 
376 #define CHECK_SAD_HALF_MV(suffix, x, y) \
377 {\
378  d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
379  d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
380  COPY3_IF_LT(dminh, d, dx, x, dy, y)\
381 }
382 
383 static int sad_hpel_motion_search(MpegEncContext * s,
384  int *mx_ptr, int *my_ptr, int dmin,
385  int src_index, int ref_index,
386  int size, int h)
387 {
388  MotionEstContext * const c= &s->me;
389  const int penalty_factor= c->sub_penalty_factor;
390  int mx, my, dminh;
391  uint8_t *pix, *ptr;
392  int stride= c->stride;
393  const int flags= c->sub_flags;
394  LOAD_COMMON
395 
396  av_assert2(flags == 0);
397 
398  if(c->skip){
399  *mx_ptr = 0;
400  *my_ptr = 0;
401  return dmin;
402  }
403 
404  pix = c->src[src_index][0];
405 
406  mx = *mx_ptr;
407  my = *my_ptr;
408  ptr = c->ref[ref_index][0] + (my * stride) + mx;
409 
410  dminh = dmin;
411 
412  if (mx > xmin && mx < xmax &&
413  my > ymin && my < ymax) {
414  int dx=0, dy=0;
415  int d, pen_x, pen_y;
416  const int index= (my<<ME_MAP_SHIFT) + mx;
417  const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
418  const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
419  const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
420  const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
421  mx<<=1;
422  my<<=1;
423 
424 
425  pen_x= pred_x + mx;
426  pen_y= pred_y + my;
427 
428  ptr-= stride;
429  if(t<=b){
430  CHECK_SAD_HALF_MV(y2 , 0, -1)
431  if(l<=r){
432  CHECK_SAD_HALF_MV(xy2, -1, -1)
433  if(t+r<=b+l){
434  CHECK_SAD_HALF_MV(xy2, +1, -1)
435  ptr+= stride;
436  }else{
437  ptr+= stride;
438  CHECK_SAD_HALF_MV(xy2, -1, +1)
439  }
440  CHECK_SAD_HALF_MV(x2 , -1, 0)
441  }else{
442  CHECK_SAD_HALF_MV(xy2, +1, -1)
443  if(t+l<=b+r){
444  CHECK_SAD_HALF_MV(xy2, -1, -1)
445  ptr+= stride;
446  }else{
447  ptr+= stride;
448  CHECK_SAD_HALF_MV(xy2, +1, +1)
449  }
450  CHECK_SAD_HALF_MV(x2 , +1, 0)
451  }
452  }else{
453  if(l<=r){
454  if(t+l<=b+r){
455  CHECK_SAD_HALF_MV(xy2, -1, -1)
456  ptr+= stride;
457  }else{
458  ptr+= stride;
459  CHECK_SAD_HALF_MV(xy2, +1, +1)
460  }
461  CHECK_SAD_HALF_MV(x2 , -1, 0)
462  CHECK_SAD_HALF_MV(xy2, -1, +1)
463  }else{
464  if(t+r<=b+l){
465  CHECK_SAD_HALF_MV(xy2, +1, -1)
466  ptr+= stride;
467  }else{
468  ptr+= stride;
469  CHECK_SAD_HALF_MV(xy2, -1, +1)
470  }
471  CHECK_SAD_HALF_MV(x2 , +1, 0)
472  CHECK_SAD_HALF_MV(xy2, +1, +1)
473  }
474  CHECK_SAD_HALF_MV(y2 , 0, +1)
475  }
476  mx+=dx;
477  my+=dy;
478 
479  }else{
480  mx<<=1;
481  my<<=1;
482  }
483 
484  *mx_ptr = mx;
485  *my_ptr = my;
486  return dminh;
487 }
488 
489 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
490 {
491  const int xy= s->mb_x + s->mb_y*s->mb_stride;
492 
493  s->p_mv_table[xy][0] = mx;
494  s->p_mv_table[xy][1] = my;
495 
496  /* has already been set to the 4 MV if 4MV is done */
497  if(mv4){
498  int mot_xy= s->block_index[0];
499 
500  s->current_picture.motion_val[0][mot_xy ][0] = mx;
501  s->current_picture.motion_val[0][mot_xy ][1] = my;
502  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
503  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
504 
505  mot_xy += s->b8_stride;
506  s->current_picture.motion_val[0][mot_xy ][0] = mx;
507  s->current_picture.motion_val[0][mot_xy ][1] = my;
508  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
509  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
510  }
511 }
512 
513 /**
514  * get fullpel ME search limits.
515  */
516 static inline void get_limits(MpegEncContext *s, int x, int y)
517 {
518  MotionEstContext * const c= &s->me;
519  int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
520  int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
521 /*
522  if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
523  else c->range= 16;
524 */
525  if (s->unrestricted_mv) {
526  c->xmin = - x - 16;
527  c->ymin = - y - 16;
528  c->xmax = - x + s->width;
529  c->ymax = - y + s->height;
530  } else if (s->out_format == FMT_H261){
531  // Search range of H261 is different from other codec standards
532  c->xmin = (x > 15) ? - 15 : 0;
533  c->ymin = (y > 15) ? - 15 : 0;
534  c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
535  c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
536  } else {
537  c->xmin = - x;
538  c->ymin = - y;
539  c->xmax = - x + s->mb_width *16 - 16;
540  c->ymax = - y + s->mb_height*16 - 16;
541  }
542  if(!range || range > max_range)
543  range = max_range;
544  if(range){
545  c->xmin = FFMAX(c->xmin,-range);
546  c->xmax = FFMIN(c->xmax, range);
547  c->ymin = FFMAX(c->ymin,-range);
548  c->ymax = FFMIN(c->ymax, range);
549  }
550 }
551 
552 static inline void init_mv4_ref(MotionEstContext *c){
553  const int stride= c->stride;
554 
555  c->ref[1][0] = c->ref[0][0] + 8;
556  c->ref[2][0] = c->ref[0][0] + 8*stride;
557  c->ref[3][0] = c->ref[2][0] + 8;
558  c->src[1][0] = c->src[0][0] + 8;
559  c->src[2][0] = c->src[0][0] + 8*stride;
560  c->src[3][0] = c->src[2][0] + 8;
561 }
562 
563 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
564 {
565  MotionEstContext * const c= &s->me;
566  const int size= 1;
567  const int h=8;
568  int block;
569  int P[10][2];
570  int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
571  int same=1;
572  const int stride= c->stride;
573  uint8_t *mv_penalty= c->current_mv_penalty;
574  int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
575 
576  init_mv4_ref(c);
577 
578  for(block=0; block<4; block++){
579  int mx4, my4;
580  int pred_x4, pred_y4;
581  int dmin4;
582  static const int off[4]= {2, 1, 1, -1};
583  const int mot_stride = s->b8_stride;
584  const int mot_xy = s->block_index[block];
585 
586  if(saftey_cliping){
587  c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
588  c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
589  }
590 
591  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
592  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
593 
594  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
595 
596  /* special case for first line */
597  if (s->first_slice_line && block<2) {
598  c->pred_x= pred_x4= P_LEFT[0];
599  c->pred_y= pred_y4= P_LEFT[1];
600  } else {
601  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
602  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
603  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
604  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
605  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
606  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
607  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
608  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
609 
610  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
611  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
612 
613  c->pred_x= pred_x4 = P_MEDIAN[0];
614  c->pred_y= pred_y4 = P_MEDIAN[1];
615  }
616  P_MV1[0]= mx;
617  P_MV1[1]= my;
618  if(saftey_cliping)
619  for(i=0; i<10; i++){
620  if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
621  if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
622  }
623 
624  dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
625 
626  dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
627 
628  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
629  int dxy;
630  const int offset= ((block&1) + (block>>1)*stride)*8;
631  uint8_t *dest_y = c->scratchpad + offset;
632  if(s->quarter_sample){
633  uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
634  dxy = ((my4 & 3) << 2) | (mx4 & 3);
635 
636  if(s->no_rounding)
637  s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
638  else
639  s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
640  }else{
641  uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
642  dxy = ((my4 & 1) << 1) | (mx4 & 1);
643 
644  if(s->no_rounding)
645  s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
646  else
647  s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
648  }
649  dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
650  }else
651  dmin_sum+= dmin4;
652 
653  if(s->quarter_sample){
654  mx4_sum+= mx4/2;
655  my4_sum+= my4/2;
656  }else{
657  mx4_sum+= mx4;
658  my4_sum+= my4;
659  }
660 
661  s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
662  s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
663 
664  if(mx4 != mx || my4 != my) same=0;
665  }
666 
667  if(same)
668  return INT_MAX;
669 
670  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
671  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);
672  }
673 
674  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
675  int dxy;
676  int mx, my;
677  int offset;
678 
679  mx= ff_h263_round_chroma(mx4_sum);
680  my= ff_h263_round_chroma(my4_sum);
681  dxy = ((my & 1) << 1) | (mx & 1);
682 
683  offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
684 
685  if(s->no_rounding){
686  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
687  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
688  }else{
689  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
690  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
691  }
692 
693  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);
694  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);
695  }
696 
697  c->pred_x= mx;
698  c->pred_y= my;
699 
700  switch(c->avctx->mb_cmp&0xFF){
701  /*case FF_CMP_SSE:
702  return dmin_sum+ 32*s->qscale*s->qscale;*/
703  case FF_CMP_RD:
704  return dmin_sum;
705  default:
706  return dmin_sum+ 11*c->mb_penalty_factor;
707  }
708 }
709 
710 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
711  MotionEstContext * const c= &s->me;
712 
713  c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
714  c->src[1][0] = c->src[0][0] + s->linesize;
715  if(c->flags & FLAG_CHROMA){
716  c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
717  c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
718  c->src[1][1] = c->src[0][1] + s->uvlinesize;
719  c->src[1][2] = c->src[0][2] + s->uvlinesize;
720  }
721 }
722 
723 static int interlaced_search(MpegEncContext *s, int ref_index,
724  int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
725 {
726  MotionEstContext * const c= &s->me;
727  const int size=0;
728  const int h=8;
729  int block;
730  int P[10][2];
731  uint8_t * const mv_penalty= c->current_mv_penalty;
732  int same=1;
733  const int stride= 2*s->linesize;
734  int dmin_sum= 0;
735  const int mot_stride= s->mb_stride;
736  const int xy= s->mb_x + s->mb_y*mot_stride;
737 
738  c->ymin>>=1;
739  c->ymax>>=1;
740  c->stride<<=1;
741  c->uvstride<<=1;
742  init_interlaced_ref(s, ref_index);
743 
744  for(block=0; block<2; block++){
745  int field_select;
746  int best_dmin= INT_MAX;
747  int best_field= -1;
748 
749  for(field_select=0; field_select<2; field_select++){
750  int dmin, mx_i, my_i;
751  int16_t (*mv_table)[2]= mv_tables[block][field_select];
752 
753  if(user_field_select){
754  av_assert1(field_select==0 || field_select==1);
755  av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
756  if(field_select_tables[block][xy] != field_select)
757  continue;
758  }
759 
760  P_LEFT[0] = mv_table[xy - 1][0];
761  P_LEFT[1] = mv_table[xy - 1][1];
762  if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
763 
764  c->pred_x= P_LEFT[0];
765  c->pred_y= P_LEFT[1];
766 
767  if(!s->first_slice_line){
768  P_TOP[0] = mv_table[xy - mot_stride][0];
769  P_TOP[1] = mv_table[xy - mot_stride][1];
770  P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
771  P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
772  if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
773  if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
774  if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
775  if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
776 
777  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
778  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
779  }
780  P_MV1[0]= mx; //FIXME not correct if block != field_select
781  P_MV1[1]= my / 2;
782 
783  dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
784 
785  dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
786 
787  mv_table[xy][0]= mx_i;
788  mv_table[xy][1]= my_i;
789 
790  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
791  int dxy;
792 
793  //FIXME chroma ME
794  uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
795  dxy = ((my_i & 1) << 1) | (mx_i & 1);
796 
797  if(s->no_rounding){
798  s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
799  }else{
800  s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
801  }
802  dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
803  dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
804  }else
805  dmin+= c->mb_penalty_factor; //field_select bits
806 
807  dmin += field_select != block; //slightly prefer same field
808 
809  if(dmin < best_dmin){
810  best_dmin= dmin;
811  best_field= field_select;
812  }
813  }
814  {
815  int16_t (*mv_table)[2]= mv_tables[block][best_field];
816 
817  if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
818  if(mv_table[xy][1]&1) same=0;
819  if(mv_table[xy][1]*2 != my) same=0;
820  if(best_field != block) same=0;
821  }
822 
823  field_select_tables[block][xy]= best_field;
824  dmin_sum += best_dmin;
825  }
826 
827  c->ymin<<=1;
828  c->ymax<<=1;
829  c->stride>>=1;
830  c->uvstride>>=1;
831 
832  if(same)
833  return INT_MAX;
834 
835  switch(c->avctx->mb_cmp&0xFF){
836  /*case FF_CMP_SSE:
837  return dmin_sum+ 32*s->qscale*s->qscale;*/
838  case FF_CMP_RD:
839  return dmin_sum;
840  default:
841  return dmin_sum+ 11*c->mb_penalty_factor;
842  }
843 }
844 
845 static inline int get_penalty_factor(int lambda, int lambda2, int type){
846  switch(type&0xFF){
847  default:
848  case FF_CMP_SAD:
849  return lambda>>FF_LAMBDA_SHIFT;
850  case FF_CMP_DCT:
851  return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
852  case FF_CMP_W53:
853  return (4*lambda)>>(FF_LAMBDA_SHIFT);
854  case FF_CMP_W97:
855  return (2*lambda)>>(FF_LAMBDA_SHIFT);
856  case FF_CMP_SATD:
857  case FF_CMP_DCT264:
858  return (2*lambda)>>FF_LAMBDA_SHIFT;
859  case FF_CMP_RD:
860  case FF_CMP_PSNR:
861  case FF_CMP_SSE:
862  case FF_CMP_NSSE:
863  return lambda2>>FF_LAMBDA_SHIFT;
864  case FF_CMP_BIT:
865  return 1;
866  }
867 }
868 
869 void ff_estimate_p_frame_motion(MpegEncContext * s,
870  int mb_x, int mb_y)
871 {
872  MotionEstContext * const c= &s->me;
873  uint8_t *pix, *ppix;
874  int sum, mx, my, dmin;
875  int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
876  int vard; ///< sum of squared differences with the estimated motion vector
877  int P[10][2];
878  const int shift= 1+s->quarter_sample;
879  int mb_type=0;
880  Picture * const pic= &s->current_picture;
881 
882  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
883 
884  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
885  av_assert0(s->linesize == c->stride);
886  av_assert0(s->uvlinesize == c->uvstride);
887 
888  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
889  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
890  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
891  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
892 
893  get_limits(s, 16*mb_x, 16*mb_y);
894  c->skip=0;
895 
896  /* intra / predictive decision */
897  pix = c->src[0][0];
898  sum = s->dsp.pix_sum(pix, s->linesize);
899  varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
900 
901  pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
902  pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
903  c->mb_var_sum_temp += (varc+128)>>8;
904 
905  switch(s->me_method) {
906  case ME_ZERO:
907  default:
908  mx = 0;
909  my = 0;
910  dmin = 0;
911  break;
912  case ME_X1:
913  case ME_EPZS:
914  {
915  const int mot_stride = s->b8_stride;
916  const int mot_xy = s->block_index[0];
917 
918  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
919  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
920 
921  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
922 
923  if(!s->first_slice_line) {
924  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
925  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
926  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
927  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
928  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
929  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
930  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
931 
932  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
933  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
934 
935  if(s->out_format == FMT_H263){
936  c->pred_x = P_MEDIAN[0];
937  c->pred_y = P_MEDIAN[1];
938  }else { /* mpeg1 at least */
939  c->pred_x= P_LEFT[0];
940  c->pred_y= P_LEFT[1];
941  }
942  }else{
943  c->pred_x= P_LEFT[0];
944  c->pred_y= P_LEFT[1];
945  }
946 
947  }
948  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
949 
950  break;
951  }
952 
953  /* At this point (mx,my) are full-pell and the relative displacement */
954  ppix = c->ref[0][0] + (my * s->linesize) + mx;
955 
956  vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
957 
958  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
959  c->mc_mb_var_sum_temp += (vard+128)>>8;
960 
961  if(mb_type){
962  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
963  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
964  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
965 
966  if(mb_type == CANDIDATE_MB_TYPE_INTER){
967  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
968  set_p_mv_tables(s, mx, my, 1);
969  }else{
970  mx <<=shift;
971  my <<=shift;
972  }
973  if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
974  h263_mv4_search(s, mx, my, shift);
975 
976  set_p_mv_tables(s, mx, my, 0);
977  }
978  if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
979  interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
980  }
981  }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
982  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
983  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
984  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
985 
986  if (vard*2 + 200*256 > varc)
987  mb_type|= CANDIDATE_MB_TYPE_INTRA;
988  if (varc*2 + 200*256 > vard || s->qscale > 24){
989 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
990  mb_type|= CANDIDATE_MB_TYPE_INTER;
991  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
992  if(s->flags&CODEC_FLAG_MV0)
993  if(mx || my)
994  mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
995  }else{
996  mx <<=shift;
997  my <<=shift;
998  }
999  if((s->flags&CODEC_FLAG_4MV)
1000  && !c->skip && varc>50<<8 && vard>10<<8){
1001  if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1002  mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1003 
1004  set_p_mv_tables(s, mx, my, 0);
1005  }else
1006  set_p_mv_tables(s, mx, my, 1);
1007  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1008  && !c->skip){ //FIXME varc/d checks
1009  if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1010  mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1011  }
1012  }else{
1013  int intra_score, i;
1014  mb_type= CANDIDATE_MB_TYPE_INTER;
1015 
1016  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1017  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1018  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1019 
1020  if((s->flags&CODEC_FLAG_4MV)
1021  && !c->skip && varc>50<<8 && vard>10<<8){
1022  int dmin4= h263_mv4_search(s, mx, my, shift);
1023  if(dmin4 < dmin){
1024  mb_type= CANDIDATE_MB_TYPE_INTER4V;
1025  dmin=dmin4;
1026  }
1027  }
1028  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1029  && !c->skip){ //FIXME varc/d checks
1030  int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1031  if(dmin_i < dmin){
1032  mb_type = CANDIDATE_MB_TYPE_INTER_I;
1033  dmin= dmin_i;
1034  }
1035  }
1036 
1037  set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1038 
1039  /* get intra luma score */
1040  if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1041  intra_score= varc - 500;
1042  }else{
1043  unsigned mean = (sum+128)>>8;
1044  mean*= 0x01010101;
1045 
1046  for(i=0; i<16; i++){
1047  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1048  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1049  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1050  *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1051  }
1052 
1053  intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1054  }
1055  intra_score += c->mb_penalty_factor*16;
1056 
1057  if(intra_score < dmin){
1058  mb_type= CANDIDATE_MB_TYPE_INTRA;
1059  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1060  }else
1061  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1062 
1063  {
1064  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1065  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1066  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1067  }
1068  }
1069 
1070  s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1071 }
1072 
1073 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1074  int mb_x, int mb_y)
1075 {
1076  MotionEstContext * const c= &s->me;
1077  int mx, my, dmin;
1078  int P[10][2];
1079  const int shift= 1+s->quarter_sample;
1080  const int xy= mb_x + mb_y*s->mb_stride;
1081  init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1082 
1083  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1084 
1085  c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1086  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1087 
1088  get_limits(s, 16*mb_x, 16*mb_y);
1089  c->skip=0;
1090 
1091  P_LEFT[0] = s->p_mv_table[xy + 1][0];
1092  P_LEFT[1] = s->p_mv_table[xy + 1][1];
1093 
1094  if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1095 
1096  /* special case for first line */
1097  if (s->first_slice_line) {
1098  c->pred_x= P_LEFT[0];
1099  c->pred_y= P_LEFT[1];
1100  P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1101  P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1102  } else {
1103  P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1104  P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1105  P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1106  P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1107  if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1108  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1109  if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1110 
1111  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1112  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1113 
1114  c->pred_x = P_MEDIAN[0];
1115  c->pred_y = P_MEDIAN[1];
1116  }
1117 
1118  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1119 
1120  s->p_mv_table[xy][0] = mx<<shift;
1121  s->p_mv_table[xy][1] = my<<shift;
1122 
1123  return dmin;
1124 }
1125 
1126 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1127  int16_t (*mv_table)[2], int ref_index, int f_code)
1128 {
1129  MotionEstContext * const c= &s->me;
1130  int mx, my, dmin;
1131  int P[10][2];
1132  const int shift= 1+s->quarter_sample;
1133  const int mot_stride = s->mb_stride;
1134  const int mot_xy = mb_y*mot_stride + mb_x;
1135  uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1136  int mv_scale;
1137 
1138  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1139  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1140  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1141  c->current_mv_penalty= mv_penalty;
1142 
1143  get_limits(s, 16*mb_x, 16*mb_y);
1144 
1145  switch(s->me_method) {
1146  case ME_ZERO:
1147  default:
1148  mx = 0;
1149  my = 0;
1150  dmin = 0;
1151  break;
1152  case ME_X1:
1153  case ME_EPZS:
1154  P_LEFT[0] = mv_table[mot_xy - 1][0];
1155  P_LEFT[1] = mv_table[mot_xy - 1][1];
1156 
1157  if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1158 
1159  /* special case for first line */
1160  if (!s->first_slice_line) {
1161  P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1162  P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1163  P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1164  P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1165  if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1166  if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1167  if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1168 
1169  P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1170  P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1171  }
1172  c->pred_x = P_LEFT[0];
1173  c->pred_y = P_LEFT[1];
1174 
1175  if(mv_table == s->b_forw_mv_table){
1176  mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1177  }else{
1178  mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1179  }
1180 
1181  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1182 
1183  break;
1184  }
1185 
1186  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1187 
1188  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1189  dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1190 
1191 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1192  mv_table[mot_xy][0]= mx;
1193  mv_table[mot_xy][1]= my;
1194 
1195  return dmin;
1196 }
1197 
1198 static inline int check_bidir_mv(MpegEncContext * s,
1199  int motion_fx, int motion_fy,
1200  int motion_bx, int motion_by,
1201  int pred_fx, int pred_fy,
1202  int pred_bx, int pred_by,
1203  int size, int h)
1204 {
1205  //FIXME optimize?
1206  //FIXME better f_code prediction (max mv & distance)
1207  //FIXME pointers
1208  MotionEstContext * const c= &s->me;
1209  uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1210  uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1211  int stride= c->stride;
1212  uint8_t *dest_y = c->scratchpad;
1213  uint8_t *ptr;
1214  int dxy;
1215  int src_x, src_y;
1216  int fbmin;
1217  uint8_t **src_data= c->src[0];
1218  uint8_t **ref_data= c->ref[0];
1219  uint8_t **ref2_data= c->ref[2];
1220 
1221  if(s->quarter_sample){
1222  dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1223  src_x = motion_fx >> 2;
1224  src_y = motion_fy >> 2;
1225 
1226  ptr = ref_data[0] + (src_y * stride) + src_x;
1227  s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1228 
1229  dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1230  src_x = motion_bx >> 2;
1231  src_y = motion_by >> 2;
1232 
1233  ptr = ref2_data[0] + (src_y * stride) + src_x;
1234  s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1235  }else{
1236  dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1237  src_x = motion_fx >> 1;
1238  src_y = motion_fy >> 1;
1239 
1240  ptr = ref_data[0] + (src_y * stride) + src_x;
1241  s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1242 
1243  dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1244  src_x = motion_bx >> 1;
1245  src_y = motion_by >> 1;
1246 
1247  ptr = ref2_data[0] + (src_y * stride) + src_x;
1248  s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1249  }
1250 
1251  fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1252  +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1253  + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1254 
1255  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1256  }
1257  //FIXME CHROMA !!!
1258 
1259  return fbmin;
1260 }
1261 
1262 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1263 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1264 {
1265  MotionEstContext * const c= &s->me;
1266  const int mot_stride = s->mb_stride;
1267  const int xy = mb_y *mot_stride + mb_x;
1268  int fbmin;
1269  int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1270  int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1271  int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1272  int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1273  int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1274  int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1275  int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1276  int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1277  const int flags= c->sub_flags;
1278  const int qpel= flags&FLAG_QPEL;
1279  const int shift= 1+qpel;
1280  const int xmin= c->xmin<<shift;
1281  const int ymin= c->ymin<<shift;
1282  const int xmax= c->xmax<<shift;
1283  const int ymax= c->ymax<<shift;
1284 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1285 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1286  int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1287  uint8_t map[256] = { 0 };
1288 
1289  map[hashidx&255] = 1;
1290 
1291  fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1292  motion_bx, motion_by,
1293  pred_fx, pred_fy,
1294  pred_bx, pred_by,
1295  0, 16);
1296 
1297  if(s->avctx->bidir_refine){
1298  int end;
1299  static const uint8_t limittab[5]={0,8,32,64,80};
1300  const int limit= limittab[s->avctx->bidir_refine];
1301  static const int8_t vect[][4]={
1302 { 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},
1303 
1304 { 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},
1305 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1306 { 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},
1307 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1308 
1309 { 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},
1310 { 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},
1311 { 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},
1312 { 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},
1313 
1314 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1315 { 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},
1316 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1317  };
1318  static const uint8_t hash[]={
1319 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),
1320 
1321 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),
1322 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1323 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),
1324 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1325 
1326 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),
1327 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),
1328 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),
1329 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),
1330 
1331 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1332 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),
1333 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),
1334 };
1335 
1336 #define CHECK_BIDIR(fx,fy,bx,by)\
1337  if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1338  &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1339  &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1340  int score;\
1341  map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1342  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);\
1343  if(score < fbmin){\
1344  hashidx += HASH(fx,fy,bx,by);\
1345  fbmin= score;\
1346  motion_fx+=fx;\
1347  motion_fy+=fy;\
1348  motion_bx+=bx;\
1349  motion_by+=by;\
1350  end=0;\
1351  }\
1352  }
1353 #define CHECK_BIDIR2(a,b,c,d)\
1354 CHECK_BIDIR(a,b,c,d)\
1355 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1356 
1357  do{
1358  int i;
1359  int borderdist=0;
1360  end=1;
1361 
1362  CHECK_BIDIR2(0,0,0,1)
1363  CHECK_BIDIR2(0,0,1,0)
1364  CHECK_BIDIR2(0,1,0,0)
1365  CHECK_BIDIR2(1,0,0,0)
1366 
1367  for(i=8; i<limit; i++){
1368  int fx= motion_fx+vect[i][0];
1369  int fy= motion_fy+vect[i][1];
1370  int bx= motion_bx+vect[i][2];
1371  int by= motion_by+vect[i][3];
1372  if(borderdist<=0){
1373  int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1374  int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1375  if((a|b) < 0)
1376  map[(hashidx+hash[i])&255] = 1;
1377  }
1378  if(!map[(hashidx+hash[i])&255]){
1379  int score;
1380  map[(hashidx+hash[i])&255] = 1;
1381  score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1382  if(score < fbmin){
1383  hashidx += hash[i];
1384  fbmin= score;
1385  motion_fx=fx;
1386  motion_fy=fy;
1387  motion_bx=bx;
1388  motion_by=by;
1389  end=0;
1390  borderdist--;
1391  if(borderdist<=0){
1392  int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1393  int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1394  borderdist= FFMIN(a,b);
1395  }
1396  }
1397  }
1398  }
1399  }while(!end);
1400  }
1401 
1402  s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1403  s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1404  s->b_bidir_back_mv_table[xy][0]= motion_bx;
1405  s->b_bidir_back_mv_table[xy][1]= motion_by;
1406 
1407  return fbmin;
1408 }
1409 
1410 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1411 {
1412  MotionEstContext * const c= &s->me;
1413  int P[10][2];
1414  const int mot_stride = s->mb_stride;
1415  const int mot_xy = mb_y*mot_stride + mb_x;
1416  const int shift= 1+s->quarter_sample;
1417  int dmin, i;
1418  const int time_pp= s->pp_time;
1419  const int time_pb= s->pb_time;
1420  int mx, my, xmin, xmax, ymin, ymax;
1421  int16_t (*mv_table)[2]= s->b_direct_mv_table;
1422 
1423  c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1424  ymin= xmin=(-32)>>shift;
1425  ymax= xmax= 31>>shift;
1426 
1427  if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1428  s->mv_type= MV_TYPE_8X8;
1429  }else{
1430  s->mv_type= MV_TYPE_16X16;
1431  }
1432 
1433  for(i=0; i<4; i++){
1434  int index= s->block_index[i];
1435  int min, max;
1436 
1437  c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1438  c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1439  c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1440  c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1441 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1442 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1443 
1444  max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1445  min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1446  max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1447  min+= 16*mb_x - 1;
1448  xmax= FFMIN(xmax, s->width - max);
1449  xmin= FFMAX(xmin, - 16 - min);
1450 
1451  max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1452  min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1453  max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1454  min+= 16*mb_y - 1;
1455  ymax= FFMIN(ymax, s->height - max);
1456  ymin= FFMAX(ymin, - 16 - min);
1457 
1458  if(s->mv_type == MV_TYPE_16X16) break;
1459  }
1460 
1461  av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1462 
1463  if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1464  s->b_direct_mv_table[mot_xy][0]= 0;
1465  s->b_direct_mv_table[mot_xy][1]= 0;
1466 
1467  return 256*256*256*64;
1468  }
1469 
1470  c->xmin= xmin;
1471  c->ymin= ymin;
1472  c->xmax= xmax;
1473  c->ymax= ymax;
1474  c->flags |= FLAG_DIRECT;
1475  c->sub_flags |= FLAG_DIRECT;
1476  c->pred_x=0;
1477  c->pred_y=0;
1478 
1479  P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1480  P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1481 
1482  /* special case for first line */
1483  if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1484  P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1485  P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1486  P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1487  P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1488 
1489  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1490  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1491  }
1492 
1493  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1494  if(c->sub_flags&FLAG_QPEL)
1495  dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1496  else
1497  dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1498 
1499  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1500  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1501 
1502  get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1503 
1504  mv_table[mot_xy][0]= mx;
1505  mv_table[mot_xy][1]= my;
1506  c->flags &= ~FLAG_DIRECT;
1507  c->sub_flags &= ~FLAG_DIRECT;
1508 
1509  return dmin;
1510 }
1511 
1512 void ff_estimate_b_frame_motion(MpegEncContext * s,
1513  int mb_x, int mb_y)
1514 {
1515  MotionEstContext * const c= &s->me;
1516  const int penalty_factor= c->mb_penalty_factor;
1517  int fmin, bmin, dmin, fbmin, bimin, fimin;
1518  int type=0;
1519  const int xy = mb_y*s->mb_stride + mb_x;
1520  init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1521  s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1522 
1523  get_limits(s, 16*mb_x, 16*mb_y);
1524 
1525  c->skip=0;
1526 
1527  if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1528  int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1529 
1530  score= ((unsigned)(score*score + 128*256))>>16;
1531  c->mc_mb_var_sum_temp += score;
1532  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1533  s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1534 
1535  return;
1536  }
1537 
1538  if (s->codec_id == AV_CODEC_ID_MPEG4)
1539  dmin= direct_search(s, mb_x, mb_y);
1540  else
1541  dmin= INT_MAX;
1542 //FIXME penalty stuff for non mpeg4
1543  c->skip=0;
1544  fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1545  3 * penalty_factor;
1546 
1547  c->skip=0;
1548  bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1549  2 * penalty_factor;
1550  av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1551 
1552  c->skip=0;
1553  fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1554  av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1555 
1556  if(s->flags & CODEC_FLAG_INTERLACED_ME){
1557 //FIXME mb type penalty
1558  c->skip=0;
1559  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1560  fimin= interlaced_search(s, 0,
1561  s->b_field_mv_table[0], s->b_field_select_table[0],
1562  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1563  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1564  bimin= interlaced_search(s, 2,
1565  s->b_field_mv_table[1], s->b_field_select_table[1],
1566  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1567  }else
1568  fimin= bimin= INT_MAX;
1569 
1570  {
1571  int score= fmin;
1572  type = CANDIDATE_MB_TYPE_FORWARD;
1573 
1574  if (dmin <= score){
1575  score = dmin;
1576  type = CANDIDATE_MB_TYPE_DIRECT;
1577  }
1578  if(bmin<score){
1579  score=bmin;
1580  type= CANDIDATE_MB_TYPE_BACKWARD;
1581  }
1582  if(fbmin<score){
1583  score=fbmin;
1584  type= CANDIDATE_MB_TYPE_BIDIR;
1585  }
1586  if(fimin<score){
1587  score=fimin;
1588  type= CANDIDATE_MB_TYPE_FORWARD_I;
1589  }
1590  if(bimin<score){
1591  score=bimin;
1592  type= CANDIDATE_MB_TYPE_BACKWARD_I;
1593  }
1594 
1595  score= ((unsigned)(score*score + 128*256))>>16;
1596  c->mc_mb_var_sum_temp += score;
1597  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1598  }
1599 
1600  if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1601  type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1602  if(fimin < INT_MAX)
1603  type |= CANDIDATE_MB_TYPE_FORWARD_I;
1604  if(bimin < INT_MAX)
1605  type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1606  if(fimin < INT_MAX && bimin < INT_MAX){
1607  type |= CANDIDATE_MB_TYPE_BIDIR_I;
1608  }
1609  //FIXME something smarter
1610  if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1611  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])
1612  type |= CANDIDATE_MB_TYPE_DIRECT0;
1613  }
1614 
1615  s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1616 }
1617 
1618 /* find best f_code for ME which do unlimited searches */
1619 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1620 {
1621  if(s->me_method>=ME_EPZS){
1622  int score[8];
1623  int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1624  uint8_t * fcode_tab= s->fcode_tab;
1625  int best_fcode=-1;
1626  int best_score=-10000000;
1627 
1628  if(s->msmpeg4_version)
1629  range= FFMIN(range, 16);
1630  else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1631  range= FFMIN(range, 256);
1632 
1633  for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1634 
1635  for(y=0; y<s->mb_height; y++){
1636  int x;
1637  int xy= y*s->mb_stride;
1638  for(x=0; x<s->mb_width; x++){
1639  if(s->mb_type[xy] & type){
1640  int mx= mv_table[xy][0];
1641  int my= mv_table[xy][1];
1642  int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1643  fcode_tab[my + MAX_MV]);
1644  int j;
1645 
1646  if(mx >= range || mx < -range ||
1647  my >= range || my < -range)
1648  continue;
1649 
1650  for(j=0; j<fcode && j<8; j++){
1651  if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1652  score[j]-= 170;
1653  }
1654  }
1655  xy++;
1656  }
1657  }
1658 
1659  for(i=1; i<8; i++){
1660  if(score[i] > best_score){
1661  best_score= score[i];
1662  best_fcode= i;
1663  }
1664  }
1665 
1666  return best_fcode;
1667  }else{
1668  return 1;
1669  }
1670 }
1671 
1672 void ff_fix_long_p_mvs(MpegEncContext * s)
1673 {
1674  MotionEstContext * const c= &s->me;
1675  const int f_code= s->f_code;
1676  int y, range;
1677  av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1678 
1679  range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1680 
1681  av_assert0(range <= 16 || !s->msmpeg4_version);
1682  av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1683 
1684  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1685 
1686  if(s->flags&CODEC_FLAG_4MV){
1687  const int wrap= s->b8_stride;
1688 
1689  /* clip / convert to intra 8x8 type MVs */
1690  for(y=0; y<s->mb_height; y++){
1691  int xy= y*2*wrap;
1692  int i= y*s->mb_stride;
1693  int x;
1694 
1695  for(x=0; x<s->mb_width; x++){
1696  if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1697  int block;
1698  for(block=0; block<4; block++){
1699  int off= (block& 1) + (block>>1)*wrap;
1700  int mx = s->current_picture.motion_val[0][ xy + off ][0];
1701  int my = s->current_picture.motion_val[0][ xy + off ][1];
1702 
1703  if( mx >=range || mx <-range
1704  || my >=range || my <-range){
1705  s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1706  s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1707  s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1708  }
1709  }
1710  }
1711  xy+=2;
1712  i++;
1713  }
1714  }
1715  }
1716 }
1717 
1718 /**
1719  *
1720  * @param truncate 1 for truncation, 0 for using intra
1721  */
1722 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1723  int16_t (*mv_table)[2], int f_code, int type, int truncate)
1724 {
1725  MotionEstContext * const c= &s->me;
1726  int y, h_range, v_range;
1727 
1728  // RAL: 8 in MPEG-1, 16 in MPEG-4
1729  int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1730 
1731  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1732 
1733  h_range= range;
1734  v_range= field_select_table ? range>>1 : range;
1735 
1736  /* clip / convert to intra 16x16 type MVs */
1737  for(y=0; y<s->mb_height; y++){
1738  int x;
1739  int xy= y*s->mb_stride;
1740  for(x=0; x<s->mb_width; x++){
1741  if (s->mb_type[xy] & type){ // RAL: "type" test added...
1742  if(field_select_table==NULL || field_select_table[xy] == field_select){
1743  if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1744  || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1745 
1746  if(truncate){
1747  if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1748  else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1749  if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1750  else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1751  }else{
1752  s->mb_type[xy] &= ~type;
1753  s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1754  mv_table[xy][0]=
1755  mv_table[xy][1]= 0;
1756  }
1757  }
1758  }
1759  }
1760  xy++;
1761  }
1762  }
1763 }
Generated on Wed Jul 10 2013 23:48:01 for FFmpeg by   doxygen 1.8.2