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 "mpegutils.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(MpegEncContext *s, uint8_t *a, uint8_t *b,
290  int stride, int h)
291 {
292  return 0;
293 }
294 
295 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
296 }
297 
298 int ff_init_me(MpegEncContext *s){
299  MotionEstContext * const c= &s->me;
300  int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
301  int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
302 
303  if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
304  av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
305  return -1;
306  }
307  //special case of snow is needed because snow uses its own iterative ME code
308  if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
309  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");
310  return -1;
311  }
312 
313  c->avctx= s->avctx;
314 
315  if(cache_size < 2*dia_size && !c->stride){
316  av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
317  }
318 
319  ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
320  ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
321  ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
322  ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
323 
324  c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
325  c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
326  c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
327 
328 /*FIXME s->no_rounding b_type*/
329  if(s->flags&CODEC_FLAG_QPEL){
330  c->sub_motion_search= qpel_motion_search;
331  c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
332  if (s->no_rounding)
333  c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
334  else
335  c->qpel_put = s->qdsp.put_qpel_pixels_tab;
336  }else{
337  if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
338  c->sub_motion_search= hpel_motion_search;
339  else if( c->avctx->me_sub_cmp == FF_CMP_SAD
340  && c->avctx-> me_cmp == FF_CMP_SAD
341  && c->avctx-> mb_cmp == FF_CMP_SAD)
342  c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
343  else
344  c->sub_motion_search= hpel_motion_search;
345  }
346  c->hpel_avg = s->hdsp.avg_pixels_tab;
347  if (s->no_rounding)
348  c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
349  else
350  c->hpel_put = s->hdsp.put_pixels_tab;
351 
352  if(s->linesize){
353  c->stride = s->linesize;
354  c->uvstride= s->uvlinesize;
355  }else{
356  c->stride = 16*s->mb_width + 32;
357  c->uvstride= 8*s->mb_width + 16;
358  }
359 
360  /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
361  * not have yet, and even if we had, the motion estimation code
362  * does not expect it. */
363  if(s->codec_id != AV_CODEC_ID_SNOW){
364  if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
365  s->dsp.me_cmp[2]= zero_cmp;
366  }
367  if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
368  s->dsp.me_sub_cmp[2]= zero_cmp;
369  }
370  c->hpel_put[2][0]= c->hpel_put[2][1]=
371  c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
372  }
373 
374  if(s->codec_id == AV_CODEC_ID_H261){
375  c->sub_motion_search= no_sub_motion_search;
376  }
377 
378  return 0;
379 }
380 
381 #define CHECK_SAD_HALF_MV(suffix, x, y) \
382 {\
383  d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
384  d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
385  COPY3_IF_LT(dminh, d, dx, x, dy, y)\
386 }
387 
388 static int sad_hpel_motion_search(MpegEncContext * s,
389  int *mx_ptr, int *my_ptr, int dmin,
390  int src_index, int ref_index,
391  int size, int h)
392 {
393  MotionEstContext * const c= &s->me;
394  const int penalty_factor= c->sub_penalty_factor;
395  int mx, my, dminh;
396  uint8_t *pix, *ptr;
397  int stride= c->stride;
398  LOAD_COMMON
399 
400  av_assert2(c->sub_flags == 0);
401 
402  if(c->skip){
403  *mx_ptr = 0;
404  *my_ptr = 0;
405  return dmin;
406  }
407 
408  pix = c->src[src_index][0];
409 
410  mx = *mx_ptr;
411  my = *my_ptr;
412  ptr = c->ref[ref_index][0] + (my * stride) + mx;
413 
414  dminh = dmin;
415 
416  if (mx > xmin && mx < xmax &&
417  my > ymin && my < ymax) {
418  int dx=0, dy=0;
419  int d, pen_x, pen_y;
420  const int index= (my<<ME_MAP_SHIFT) + mx;
421  const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
422  const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
423  const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
424  const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
425  mx<<=1;
426  my<<=1;
427 
428 
429  pen_x= pred_x + mx;
430  pen_y= pred_y + my;
431 
432  ptr-= stride;
433  if(t<=b){
434  CHECK_SAD_HALF_MV(y2 , 0, -1)
435  if(l<=r){
436  CHECK_SAD_HALF_MV(xy2, -1, -1)
437  if(t+r<=b+l){
438  CHECK_SAD_HALF_MV(xy2, +1, -1)
439  ptr+= stride;
440  }else{
441  ptr+= stride;
442  CHECK_SAD_HALF_MV(xy2, -1, +1)
443  }
444  CHECK_SAD_HALF_MV(x2 , -1, 0)
445  }else{
446  CHECK_SAD_HALF_MV(xy2, +1, -1)
447  if(t+l<=b+r){
448  CHECK_SAD_HALF_MV(xy2, -1, -1)
449  ptr+= stride;
450  }else{
451  ptr+= stride;
452  CHECK_SAD_HALF_MV(xy2, +1, +1)
453  }
454  CHECK_SAD_HALF_MV(x2 , +1, 0)
455  }
456  }else{
457  if(l<=r){
458  if(t+l<=b+r){
459  CHECK_SAD_HALF_MV(xy2, -1, -1)
460  ptr+= stride;
461  }else{
462  ptr+= stride;
463  CHECK_SAD_HALF_MV(xy2, +1, +1)
464  }
465  CHECK_SAD_HALF_MV(x2 , -1, 0)
466  CHECK_SAD_HALF_MV(xy2, -1, +1)
467  }else{
468  if(t+r<=b+l){
469  CHECK_SAD_HALF_MV(xy2, +1, -1)
470  ptr+= stride;
471  }else{
472  ptr+= stride;
473  CHECK_SAD_HALF_MV(xy2, -1, +1)
474  }
475  CHECK_SAD_HALF_MV(x2 , +1, 0)
476  CHECK_SAD_HALF_MV(xy2, +1, +1)
477  }
478  CHECK_SAD_HALF_MV(y2 , 0, +1)
479  }
480  mx+=dx;
481  my+=dy;
482 
483  }else{
484  mx<<=1;
485  my<<=1;
486  }
487 
488  *mx_ptr = mx;
489  *my_ptr = my;
490  return dminh;
491 }
492 
493 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
494 {
495  const int xy= s->mb_x + s->mb_y*s->mb_stride;
496 
497  s->p_mv_table[xy][0] = mx;
498  s->p_mv_table[xy][1] = my;
499 
500  /* has already been set to the 4 MV if 4MV is done */
501  if(mv4){
502  int mot_xy= s->block_index[0];
503 
504  s->current_picture.motion_val[0][mot_xy ][0] = mx;
505  s->current_picture.motion_val[0][mot_xy ][1] = my;
506  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
507  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
508 
509  mot_xy += s->b8_stride;
510  s->current_picture.motion_val[0][mot_xy ][0] = mx;
511  s->current_picture.motion_val[0][mot_xy ][1] = my;
512  s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
513  s->current_picture.motion_val[0][mot_xy + 1][1] = my;
514  }
515 }
516 
517 /**
518  * get fullpel ME search limits.
519  */
520 static inline void get_limits(MpegEncContext *s, int x, int y)
521 {
522  MotionEstContext * const c= &s->me;
523  int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
524  int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
525 /*
526  if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
527  else c->range= 16;
528 */
529  if (s->unrestricted_mv) {
530  c->xmin = - x - 16;
531  c->ymin = - y - 16;
532  c->xmax = - x + s->width;
533  c->ymax = - y + s->height;
534  } else if (s->out_format == FMT_H261){
535  // Search range of H261 is different from other codec standards
536  c->xmin = (x > 15) ? - 15 : 0;
537  c->ymin = (y > 15) ? - 15 : 0;
538  c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
539  c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
540  } else {
541  c->xmin = - x;
542  c->ymin = - y;
543  c->xmax = - x + s->mb_width *16 - 16;
544  c->ymax = - y + s->mb_height*16 - 16;
545  }
546  if(!range || range > max_range)
547  range = max_range;
548  if(range){
549  c->xmin = FFMAX(c->xmin,-range);
550  c->xmax = FFMIN(c->xmax, range);
551  c->ymin = FFMAX(c->ymin,-range);
552  c->ymax = FFMIN(c->ymax, range);
553  }
554 }
555 
556 static inline void init_mv4_ref(MotionEstContext *c){
557  const int stride= c->stride;
558 
559  c->ref[1][0] = c->ref[0][0] + 8;
560  c->ref[2][0] = c->ref[0][0] + 8*stride;
561  c->ref[3][0] = c->ref[2][0] + 8;
562  c->src[1][0] = c->src[0][0] + 8;
563  c->src[2][0] = c->src[0][0] + 8*stride;
564  c->src[3][0] = c->src[2][0] + 8;
565 }
566 
567 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
568 {
569  MotionEstContext * const c= &s->me;
570  const int size= 1;
571  const int h=8;
572  int block;
573  int P[10][2];
574  int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
575  int same=1;
576  const int stride= c->stride;
577  uint8_t *mv_penalty= c->current_mv_penalty;
578  int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
579 
580  init_mv4_ref(c);
581 
582  for(block=0; block<4; block++){
583  int mx4, my4;
584  int pred_x4, pred_y4;
585  int dmin4;
586  static const int off[4]= {2, 1, 1, -1};
587  const int mot_stride = s->b8_stride;
588  const int mot_xy = s->block_index[block];
589 
590  if(saftey_cliping){
591  c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
592  c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
593  }
594 
595  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
596  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
597 
598  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
599 
600  /* special case for first line */
601  if (s->first_slice_line && block<2) {
602  c->pred_x= pred_x4= P_LEFT[0];
603  c->pred_y= pred_y4= P_LEFT[1];
604  } else {
605  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
606  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
607  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
608  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
609  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
610  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
611  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
612  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
613 
614  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
615  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
616 
617  c->pred_x= pred_x4 = P_MEDIAN[0];
618  c->pred_y= pred_y4 = P_MEDIAN[1];
619  }
620  P_MV1[0]= mx;
621  P_MV1[1]= my;
622  if(saftey_cliping)
623  for(i=1; i<10; i++){
624  if (s->first_slice_line && block<2 && i>1 && i<9)
625  continue;
626  if (i>4 && i<9)
627  continue;
628  if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
629  if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
630  }
631 
632  dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
633 
634  dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
635 
636  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
637  int dxy;
638  const int offset= ((block&1) + (block>>1)*stride)*8;
639  uint8_t *dest_y = c->scratchpad + offset;
640  if(s->quarter_sample){
641  uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
642  dxy = ((my4 & 3) << 2) | (mx4 & 3);
643 
644  if(s->no_rounding)
645  s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
646  else
647  s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
648  }else{
649  uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
650  dxy = ((my4 & 1) << 1) | (mx4 & 1);
651 
652  if(s->no_rounding)
653  s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
654  else
655  s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
656  }
657  dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
658  }else
659  dmin_sum+= dmin4;
660 
661  if(s->quarter_sample){
662  mx4_sum+= mx4/2;
663  my4_sum+= my4/2;
664  }else{
665  mx4_sum+= mx4;
666  my4_sum+= my4;
667  }
668 
669  s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
670  s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
671 
672  if(mx4 != mx || my4 != my) same=0;
673  }
674 
675  if(same)
676  return INT_MAX;
677 
678  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
679  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);
680  }
681 
682  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
683  int dxy;
684  int mx, my;
685  int offset;
686 
687  mx= ff_h263_round_chroma(mx4_sum);
688  my= ff_h263_round_chroma(my4_sum);
689  dxy = ((my & 1) << 1) | (mx & 1);
690 
691  offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
692 
693  if(s->no_rounding){
694  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
695  s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
696  }else{
697  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
698  s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
699  }
700 
701  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);
702  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);
703  }
704 
705  c->pred_x= mx;
706  c->pred_y= my;
707 
708  switch(c->avctx->mb_cmp&0xFF){
709  /*case FF_CMP_SSE:
710  return dmin_sum+ 32*s->qscale*s->qscale;*/
711  case FF_CMP_RD:
712  return dmin_sum;
713  default:
714  return dmin_sum+ 11*c->mb_penalty_factor;
715  }
716 }
717 
718 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
719  MotionEstContext * const c= &s->me;
720 
721  c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
722  c->src[1][0] = c->src[0][0] + s->linesize;
723  if(c->flags & FLAG_CHROMA){
724  c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
725  c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
726  c->src[1][1] = c->src[0][1] + s->uvlinesize;
727  c->src[1][2] = c->src[0][2] + s->uvlinesize;
728  }
729 }
730 
731 static int interlaced_search(MpegEncContext *s, int ref_index,
732  int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
733 {
734  MotionEstContext * const c= &s->me;
735  const int size=0;
736  const int h=8;
737  int block;
738  int P[10][2];
739  uint8_t * const mv_penalty= c->current_mv_penalty;
740  int same=1;
741  const int stride= 2*s->linesize;
742  int dmin_sum= 0;
743  const int mot_stride= s->mb_stride;
744  const int xy= s->mb_x + s->mb_y*mot_stride;
745 
746  c->ymin>>=1;
747  c->ymax>>=1;
748  c->stride<<=1;
749  c->uvstride<<=1;
750  init_interlaced_ref(s, ref_index);
751 
752  for(block=0; block<2; block++){
753  int field_select;
754  int best_dmin= INT_MAX;
755  int best_field= -1;
756 
757  for(field_select=0; field_select<2; field_select++){
758  int dmin, mx_i, my_i;
759  int16_t (*mv_table)[2]= mv_tables[block][field_select];
760 
761  if(user_field_select){
762  av_assert1(field_select==0 || field_select==1);
763  av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
764  if(field_select_tables[block][xy] != field_select)
765  continue;
766  }
767 
768  P_LEFT[0] = mv_table[xy - 1][0];
769  P_LEFT[1] = mv_table[xy - 1][1];
770  if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
771 
772  c->pred_x= P_LEFT[0];
773  c->pred_y= P_LEFT[1];
774 
775  if(!s->first_slice_line){
776  P_TOP[0] = mv_table[xy - mot_stride][0];
777  P_TOP[1] = mv_table[xy - mot_stride][1];
778  P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
779  P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
780  if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
781  if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
782  if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
783  if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
784 
785  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
786  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
787  }
788  P_MV1[0]= mx; //FIXME not correct if block != field_select
789  P_MV1[1]= my / 2;
790 
791  dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
792 
793  dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
794 
795  mv_table[xy][0]= mx_i;
796  mv_table[xy][1]= my_i;
797 
798  if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
799  int dxy;
800 
801  //FIXME chroma ME
802  uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
803  dxy = ((my_i & 1) << 1) | (mx_i & 1);
804 
805  if(s->no_rounding){
806  s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
807  }else{
808  s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
809  }
810  dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
811  dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
812  }else
813  dmin+= c->mb_penalty_factor; //field_select bits
814 
815  dmin += field_select != block; //slightly prefer same field
816 
817  if(dmin < best_dmin){
818  best_dmin= dmin;
819  best_field= field_select;
820  }
821  }
822  {
823  int16_t (*mv_table)[2]= mv_tables[block][best_field];
824 
825  if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
826  if(mv_table[xy][1]&1) same=0;
827  if(mv_table[xy][1]*2 != my) same=0;
828  if(best_field != block) same=0;
829  }
830 
831  field_select_tables[block][xy]= best_field;
832  dmin_sum += best_dmin;
833  }
834 
835  c->ymin<<=1;
836  c->ymax<<=1;
837  c->stride>>=1;
838  c->uvstride>>=1;
839 
840  if(same)
841  return INT_MAX;
842 
843  switch(c->avctx->mb_cmp&0xFF){
844  /*case FF_CMP_SSE:
845  return dmin_sum+ 32*s->qscale*s->qscale;*/
846  case FF_CMP_RD:
847  return dmin_sum;
848  default:
849  return dmin_sum+ 11*c->mb_penalty_factor;
850  }
851 }
852 
853 static inline int get_penalty_factor(int lambda, int lambda2, int type){
854  switch(type&0xFF){
855  default:
856  case FF_CMP_SAD:
857  return lambda>>FF_LAMBDA_SHIFT;
858  case FF_CMP_DCT:
859  return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
860  case FF_CMP_W53:
861  return (4*lambda)>>(FF_LAMBDA_SHIFT);
862  case FF_CMP_W97:
863  return (2*lambda)>>(FF_LAMBDA_SHIFT);
864  case FF_CMP_SATD:
865  case FF_CMP_DCT264:
866  return (2*lambda)>>FF_LAMBDA_SHIFT;
867  case FF_CMP_RD:
868  case FF_CMP_PSNR:
869  case FF_CMP_SSE:
870  case FF_CMP_NSSE:
871  return lambda2>>FF_LAMBDA_SHIFT;
872  case FF_CMP_BIT:
873  return 1;
874  }
875 }
876 
877 void ff_estimate_p_frame_motion(MpegEncContext * s,
878  int mb_x, int mb_y)
879 {
880  MotionEstContext * const c= &s->me;
881  uint8_t *pix, *ppix;
882  int sum, mx, my, dmin;
883  int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
884  int vard; ///< sum of squared differences with the estimated motion vector
885  int P[10][2];
886  const int shift= 1+s->quarter_sample;
887  int mb_type=0;
888  Picture * const pic= &s->current_picture;
889 
890  init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
891 
892  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
893  av_assert0(s->linesize == c->stride);
894  av_assert0(s->uvlinesize == c->uvstride);
895 
896  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
897  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
898  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
899  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
900 
901  get_limits(s, 16*mb_x, 16*mb_y);
902  c->skip=0;
903 
904  /* intra / predictive decision */
905  pix = c->src[0][0];
906  sum = s->mpvencdsp.pix_sum(pix, s->linesize);
907  varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
908  (((unsigned) sum * sum) >> 8) + 500;
909 
910  pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
911  pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
912  c->mb_var_sum_temp += (varc+128)>>8;
913 
914  switch(s->me_method) {
915  case ME_ZERO:
916  default:
917  mx = 0;
918  my = 0;
919  dmin = 0;
920  break;
921  case ME_X1:
922  case ME_EPZS:
923  {
924  const int mot_stride = s->b8_stride;
925  const int mot_xy = s->block_index[0];
926 
927  P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
928  P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
929 
930  if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
931 
932  if(!s->first_slice_line) {
933  P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
934  P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
935  P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
936  P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
937  if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
938  if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
939  if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
940 
941  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
942  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
943 
944  if(s->out_format == FMT_H263){
945  c->pred_x = P_MEDIAN[0];
946  c->pred_y = P_MEDIAN[1];
947  }else { /* mpeg1 at least */
948  c->pred_x= P_LEFT[0];
949  c->pred_y= P_LEFT[1];
950  }
951  }else{
952  c->pred_x= P_LEFT[0];
953  c->pred_y= P_LEFT[1];
954  }
955 
956  }
957  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
958 
959  break;
960  }
961 
962  /* At this point (mx,my) are full-pell and the relative displacement */
963  ppix = c->ref[0][0] + (my * s->linesize) + mx;
964 
965  vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
966 
967  pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
968  c->mc_mb_var_sum_temp += (vard+128)>>8;
969 
970  if(mb_type){
971  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
972  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
973  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
974 
975  if(mb_type == CANDIDATE_MB_TYPE_INTER){
976  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
977  set_p_mv_tables(s, mx, my, 1);
978  }else{
979  mx <<=shift;
980  my <<=shift;
981  }
982  if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
983  h263_mv4_search(s, mx, my, shift);
984 
985  set_p_mv_tables(s, mx, my, 0);
986  }
987  if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
988  interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
989  }
990  }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
991  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
992  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
993  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
994 
995  if (vard*2 + 200*256 > varc)
996  mb_type|= CANDIDATE_MB_TYPE_INTRA;
997  if (varc*2 + 200*256 > vard || s->qscale > 24){
998 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
999  mb_type|= CANDIDATE_MB_TYPE_INTER;
1000  c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1001  if (s->mpv_flags & FF_MPV_FLAG_MV0)
1002  if(mx || my)
1003  mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1004  }else{
1005  mx <<=shift;
1006  my <<=shift;
1007  }
1008  if((s->flags&CODEC_FLAG_4MV)
1009  && !c->skip && varc>50<<8 && vard>10<<8){
1010  if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1011  mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1012 
1013  set_p_mv_tables(s, mx, my, 0);
1014  }else
1015  set_p_mv_tables(s, mx, my, 1);
1016  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1017  && !c->skip){ //FIXME varc/d checks
1018  if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1019  mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1020  }
1021  }else{
1022  int intra_score, i;
1023  mb_type= CANDIDATE_MB_TYPE_INTER;
1024 
1025  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1026  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1027  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1028 
1029  if((s->flags&CODEC_FLAG_4MV)
1030  && !c->skip && varc>50<<8 && vard>10<<8){
1031  int dmin4= h263_mv4_search(s, mx, my, shift);
1032  if(dmin4 < dmin){
1033  mb_type= CANDIDATE_MB_TYPE_INTER4V;
1034  dmin=dmin4;
1035  }
1036  }
1037  if((s->flags&CODEC_FLAG_INTERLACED_ME)
1038  && !c->skip){ //FIXME varc/d checks
1039  int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1040  if(dmin_i < dmin){
1041  mb_type = CANDIDATE_MB_TYPE_INTER_I;
1042  dmin= dmin_i;
1043  }
1044  }
1045 
1046  set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1047 
1048  /* get intra luma score */
1049  if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1050  intra_score= varc - 500;
1051  }else{
1052  unsigned mean = (sum+128)>>8;
1053  mean*= 0x01010101;
1054 
1055  for(i=0; i<16; i++){
1056  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1057  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1058  *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1059  *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1060  }
1061 
1062  intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1063  }
1064  intra_score += c->mb_penalty_factor*16;
1065 
1066  if(intra_score < dmin){
1067  mb_type= CANDIDATE_MB_TYPE_INTRA;
1068  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1069  }else
1070  s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1071 
1072  {
1073  int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1074  int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1075  c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1076  }
1077  }
1078 
1079  s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1080 }
1081 
1082 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1083  int mb_x, int mb_y)
1084 {
1085  MotionEstContext * const c= &s->me;
1086  int mx, my, dmin;
1087  int P[10][2];
1088  const int shift= 1+s->quarter_sample;
1089  const int xy= mb_x + mb_y*s->mb_stride;
1090  init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1091 
1092  av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1093 
1094  c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1095  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1096 
1097  get_limits(s, 16*mb_x, 16*mb_y);
1098  c->skip=0;
1099 
1100  P_LEFT[0] = s->p_mv_table[xy + 1][0];
1101  P_LEFT[1] = s->p_mv_table[xy + 1][1];
1102 
1103  if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1104 
1105  /* special case for first line */
1106  if (s->first_slice_line) {
1107  c->pred_x= P_LEFT[0];
1108  c->pred_y= P_LEFT[1];
1109  P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1110  P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1111  } else {
1112  P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1113  P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1114  P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1115  P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1116  if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1117  if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1118  if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1119 
1120  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1121  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1122 
1123  c->pred_x = P_MEDIAN[0];
1124  c->pred_y = P_MEDIAN[1];
1125  }
1126 
1127  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1128 
1129  s->p_mv_table[xy][0] = mx<<shift;
1130  s->p_mv_table[xy][1] = my<<shift;
1131 
1132  return dmin;
1133 }
1134 
1135 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1136  int16_t (*mv_table)[2], int ref_index, int f_code)
1137 {
1138  MotionEstContext * const c= &s->me;
1139  int mx, my, dmin;
1140  int P[10][2];
1141  const int shift= 1+s->quarter_sample;
1142  const int mot_stride = s->mb_stride;
1143  const int mot_xy = mb_y*mot_stride + mb_x;
1144  uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1145  int mv_scale;
1146 
1147  c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1148  c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1149  c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1150  c->current_mv_penalty= mv_penalty;
1151 
1152  get_limits(s, 16*mb_x, 16*mb_y);
1153 
1154  switch(s->me_method) {
1155  case ME_ZERO:
1156  default:
1157  mx = 0;
1158  my = 0;
1159  dmin = 0;
1160  break;
1161  case ME_X1:
1162  case ME_EPZS:
1163  P_LEFT[0] = mv_table[mot_xy - 1][0];
1164  P_LEFT[1] = mv_table[mot_xy - 1][1];
1165 
1166  if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1167 
1168  /* special case for first line */
1169  if (!s->first_slice_line) {
1170  P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1171  P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1172  P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1173  P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1174  if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1175  if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1176  if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1177 
1178  P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1179  P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1180  }
1181  c->pred_x = P_LEFT[0];
1182  c->pred_y = P_LEFT[1];
1183 
1184  if(mv_table == s->b_forw_mv_table){
1185  mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1186  }else{
1187  mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1188  }
1189 
1190  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1191 
1192  break;
1193  }
1194 
1195  dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1196 
1197  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1198  dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1199 
1200 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1201  mv_table[mot_xy][0]= mx;
1202  mv_table[mot_xy][1]= my;
1203 
1204  return dmin;
1205 }
1206 
1207 static inline int check_bidir_mv(MpegEncContext * s,
1208  int motion_fx, int motion_fy,
1209  int motion_bx, int motion_by,
1210  int pred_fx, int pred_fy,
1211  int pred_bx, int pred_by,
1212  int size, int h)
1213 {
1214  //FIXME optimize?
1215  //FIXME better f_code prediction (max mv & distance)
1216  //FIXME pointers
1217  MotionEstContext * const c= &s->me;
1218  uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1219  uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1220  int stride= c->stride;
1221  uint8_t *dest_y = c->scratchpad;
1222  uint8_t *ptr;
1223  int dxy;
1224  int src_x, src_y;
1225  int fbmin;
1226  uint8_t **src_data= c->src[0];
1227  uint8_t **ref_data= c->ref[0];
1228  uint8_t **ref2_data= c->ref[2];
1229 
1230  if(s->quarter_sample){
1231  dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1232  src_x = motion_fx >> 2;
1233  src_y = motion_fy >> 2;
1234 
1235  ptr = ref_data[0] + (src_y * stride) + src_x;
1236  s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1237 
1238  dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1239  src_x = motion_bx >> 2;
1240  src_y = motion_by >> 2;
1241 
1242  ptr = ref2_data[0] + (src_y * stride) + src_x;
1243  s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1244  }else{
1245  dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1246  src_x = motion_fx >> 1;
1247  src_y = motion_fy >> 1;
1248 
1249  ptr = ref_data[0] + (src_y * stride) + src_x;
1250  s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1251 
1252  dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1253  src_x = motion_bx >> 1;
1254  src_y = motion_by >> 1;
1255 
1256  ptr = ref2_data[0] + (src_y * stride) + src_x;
1257  s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1258  }
1259 
1260  fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1261  +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1262  + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1263 
1264  if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1265  }
1266  //FIXME CHROMA !!!
1267 
1268  return fbmin;
1269 }
1270 
1271 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1272 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1273 {
1274  MotionEstContext * const c= &s->me;
1275  const int mot_stride = s->mb_stride;
1276  const int xy = mb_y *mot_stride + mb_x;
1277  int fbmin;
1278  int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1279  int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1280  int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1281  int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1282  int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1283  int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1284  int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1285  int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1286  const int flags= c->sub_flags;
1287  const int qpel= flags&FLAG_QPEL;
1288  const int shift= 1+qpel;
1289  const int xmin= c->xmin<<shift;
1290  const int ymin= c->ymin<<shift;
1291  const int xmax= c->xmax<<shift;
1292  const int ymax= c->ymax<<shift;
1293 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1294 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1295  int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1296  uint8_t map[256] = { 0 };
1297 
1298  map[hashidx&255] = 1;
1299 
1300  fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1301  motion_bx, motion_by,
1302  pred_fx, pred_fy,
1303  pred_bx, pred_by,
1304  0, 16);
1305 
1306  if(s->avctx->bidir_refine){
1307  int end;
1308  static const uint8_t limittab[5]={0,8,32,64,80};
1309  const int limit= limittab[s->avctx->bidir_refine];
1310  static const int8_t vect[][4]={
1311 { 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},
1312 
1313 { 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},
1314 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1315 { 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},
1316 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1317 
1318 { 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},
1319 { 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},
1320 { 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},
1321 { 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},
1322 
1323 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1324 { 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},
1325 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1326  };
1327  static const uint8_t hash[]={
1328 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),
1329 
1330 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),
1331 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1332 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),
1333 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1334 
1335 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),
1336 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),
1337 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),
1338 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),
1339 
1340 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1341 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),
1342 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),
1343 };
1344 
1345 #define CHECK_BIDIR(fx,fy,bx,by)\
1346  if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1347  &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1348  &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1349  int score;\
1350  map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1351  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);\
1352  if(score < fbmin){\
1353  hashidx += HASH(fx,fy,bx,by);\
1354  fbmin= score;\
1355  motion_fx+=fx;\
1356  motion_fy+=fy;\
1357  motion_bx+=bx;\
1358  motion_by+=by;\
1359  end=0;\
1360  }\
1361  }
1362 #define CHECK_BIDIR2(a,b,c,d)\
1363 CHECK_BIDIR(a,b,c,d)\
1364 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1365 
1366  do{
1367  int i;
1368  int borderdist=0;
1369  end=1;
1370 
1371  CHECK_BIDIR2(0,0,0,1)
1372  CHECK_BIDIR2(0,0,1,0)
1373  CHECK_BIDIR2(0,1,0,0)
1374  CHECK_BIDIR2(1,0,0,0)
1375 
1376  for(i=8; i<limit; i++){
1377  int fx= motion_fx+vect[i][0];
1378  int fy= motion_fy+vect[i][1];
1379  int bx= motion_bx+vect[i][2];
1380  int by= motion_by+vect[i][3];
1381  if(borderdist<=0){
1382  int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1383  int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1384  if((a|b) < 0)
1385  map[(hashidx+hash[i])&255] = 1;
1386  }
1387  if(!map[(hashidx+hash[i])&255]){
1388  int score;
1389  map[(hashidx+hash[i])&255] = 1;
1390  score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1391  if(score < fbmin){
1392  hashidx += hash[i];
1393  fbmin= score;
1394  motion_fx=fx;
1395  motion_fy=fy;
1396  motion_bx=bx;
1397  motion_by=by;
1398  end=0;
1399  borderdist--;
1400  if(borderdist<=0){
1401  int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1402  int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1403  borderdist= FFMIN(a,b);
1404  }
1405  }
1406  }
1407  }
1408  }while(!end);
1409  }
1410 
1411  s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1412  s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1413  s->b_bidir_back_mv_table[xy][0]= motion_bx;
1414  s->b_bidir_back_mv_table[xy][1]= motion_by;
1415 
1416  return fbmin;
1417 }
1418 
1419 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1420 {
1421  MotionEstContext * const c= &s->me;
1422  int P[10][2];
1423  const int mot_stride = s->mb_stride;
1424  const int mot_xy = mb_y*mot_stride + mb_x;
1425  const int shift= 1+s->quarter_sample;
1426  int dmin, i;
1427  const int time_pp= s->pp_time;
1428  const int time_pb= s->pb_time;
1429  int mx, my, xmin, xmax, ymin, ymax;
1430  int16_t (*mv_table)[2]= s->b_direct_mv_table;
1431 
1432  c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1433  ymin= xmin=(-32)>>shift;
1434  ymax= xmax= 31>>shift;
1435 
1436  if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1437  s->mv_type= MV_TYPE_8X8;
1438  }else{
1439  s->mv_type= MV_TYPE_16X16;
1440  }
1441 
1442  for(i=0; i<4; i++){
1443  int index= s->block_index[i];
1444  int min, max;
1445 
1446  c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1447  c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1448  c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1449  c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1450 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1451 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1452 
1453  max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1454  min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1455  max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1456  min+= 16*mb_x - 1;
1457  xmax= FFMIN(xmax, s->width - max);
1458  xmin= FFMAX(xmin, - 16 - min);
1459 
1460  max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1461  min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1462  max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1463  min+= 16*mb_y - 1;
1464  ymax= FFMIN(ymax, s->height - max);
1465  ymin= FFMAX(ymin, - 16 - min);
1466 
1467  if(s->mv_type == MV_TYPE_16X16) break;
1468  }
1469 
1470  av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1471 
1472  if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1473  s->b_direct_mv_table[mot_xy][0]= 0;
1474  s->b_direct_mv_table[mot_xy][1]= 0;
1475 
1476  return 256*256*256*64;
1477  }
1478 
1479  c->xmin= xmin;
1480  c->ymin= ymin;
1481  c->xmax= xmax;
1482  c->ymax= ymax;
1483  c->flags |= FLAG_DIRECT;
1484  c->sub_flags |= FLAG_DIRECT;
1485  c->pred_x=0;
1486  c->pred_y=0;
1487 
1488  P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1489  P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1490 
1491  /* special case for first line */
1492  if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1493  P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1494  P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1495  P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1496  P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1497 
1498  P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1499  P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1500  }
1501 
1502  dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1503  if(c->sub_flags&FLAG_QPEL)
1504  dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1505  else
1506  dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1507 
1508  if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1509  dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1510 
1511  get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1512 
1513  mv_table[mot_xy][0]= mx;
1514  mv_table[mot_xy][1]= my;
1515  c->flags &= ~FLAG_DIRECT;
1516  c->sub_flags &= ~FLAG_DIRECT;
1517 
1518  return dmin;
1519 }
1520 
1521 void ff_estimate_b_frame_motion(MpegEncContext * s,
1522  int mb_x, int mb_y)
1523 {
1524  MotionEstContext * const c= &s->me;
1525  const int penalty_factor= c->mb_penalty_factor;
1526  int fmin, bmin, dmin, fbmin, bimin, fimin;
1527  int type=0;
1528  const int xy = mb_y*s->mb_stride + mb_x;
1529  init_ref(c, s->new_picture.f->data, s->last_picture.f->data,
1530  s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1531 
1532  get_limits(s, 16*mb_x, 16*mb_y);
1533 
1534  c->skip=0;
1535 
1536  if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1537  int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1538 
1539  score= ((unsigned)(score*score + 128*256))>>16;
1540  c->mc_mb_var_sum_temp += score;
1541  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1542  s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1543 
1544  return;
1545  }
1546 
1547  if (s->codec_id == AV_CODEC_ID_MPEG4)
1548  dmin= direct_search(s, mb_x, mb_y);
1549  else
1550  dmin= INT_MAX;
1551 //FIXME penalty stuff for non mpeg4
1552  c->skip=0;
1553  fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1554  3 * penalty_factor;
1555 
1556  c->skip=0;
1557  bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1558  2 * penalty_factor;
1559  av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1560 
1561  c->skip=0;
1562  fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1563  av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1564 
1565  if(s->flags & CODEC_FLAG_INTERLACED_ME){
1566 //FIXME mb type penalty
1567  c->skip=0;
1568  c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1569  fimin= interlaced_search(s, 0,
1570  s->b_field_mv_table[0], s->b_field_select_table[0],
1571  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1572  c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1573  bimin= interlaced_search(s, 2,
1574  s->b_field_mv_table[1], s->b_field_select_table[1],
1575  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1576  }else
1577  fimin= bimin= INT_MAX;
1578 
1579  {
1580  int score= fmin;
1581  type = CANDIDATE_MB_TYPE_FORWARD;
1582 
1583  if (dmin <= score){
1584  score = dmin;
1585  type = CANDIDATE_MB_TYPE_DIRECT;
1586  }
1587  if(bmin<score){
1588  score=bmin;
1589  type= CANDIDATE_MB_TYPE_BACKWARD;
1590  }
1591  if(fbmin<score){
1592  score=fbmin;
1593  type= CANDIDATE_MB_TYPE_BIDIR;
1594  }
1595  if(fimin<score){
1596  score=fimin;
1597  type= CANDIDATE_MB_TYPE_FORWARD_I;
1598  }
1599  if(bimin<score){
1600  score=bimin;
1601  type= CANDIDATE_MB_TYPE_BACKWARD_I;
1602  }
1603 
1604  score= ((unsigned)(score*score + 128*256))>>16;
1605  c->mc_mb_var_sum_temp += score;
1606  s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1607  }
1608 
1609  if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1610  type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1611  if(fimin < INT_MAX)
1612  type |= CANDIDATE_MB_TYPE_FORWARD_I;
1613  if(bimin < INT_MAX)
1614  type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1615  if(fimin < INT_MAX && bimin < INT_MAX){
1616  type |= CANDIDATE_MB_TYPE_BIDIR_I;
1617  }
1618  //FIXME something smarter
1619  if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1620  if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1621  s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1622  type |= CANDIDATE_MB_TYPE_DIRECT0;
1623  }
1624 
1625  s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1626 }
1627 
1628 /* find best f_code for ME which do unlimited searches */
1629 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1630 {
1631  if(s->me_method>=ME_EPZS){
1632  int score[8];
1633  int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1634  uint8_t * fcode_tab= s->fcode_tab;
1635  int best_fcode=-1;
1636  int best_score=-10000000;
1637 
1638  if(s->msmpeg4_version)
1639  range= FFMIN(range, 16);
1640  else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1641  range= FFMIN(range, 256);
1642 
1643  for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1644 
1645  for(y=0; y<s->mb_height; y++){
1646  int x;
1647  int xy= y*s->mb_stride;
1648  for(x=0; x<s->mb_width; x++){
1649  if(s->mb_type[xy] & type){
1650  int mx= mv_table[xy][0];
1651  int my= mv_table[xy][1];
1652  int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1653  fcode_tab[my + MAX_MV]);
1654  int j;
1655 
1656  if(mx >= range || mx < -range ||
1657  my >= range || my < -range)
1658  continue;
1659 
1660  for(j=0; j<fcode && j<8; j++){
1661  if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1662  score[j]-= 170;
1663  }
1664  }
1665  xy++;
1666  }
1667  }
1668 
1669  for(i=1; i<8; i++){
1670  if(score[i] > best_score){
1671  best_score= score[i];
1672  best_fcode= i;
1673  }
1674  }
1675 
1676  return best_fcode;
1677  }else{
1678  return 1;
1679  }
1680 }
1681 
1682 void ff_fix_long_p_mvs(MpegEncContext * s)
1683 {
1684  MotionEstContext * const c= &s->me;
1685  const int f_code= s->f_code;
1686  int y, range;
1687  av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1688 
1689  range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1690 
1691  av_assert0(range <= 16 || !s->msmpeg4_version);
1692  av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1693 
1694  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1695 
1696  if(s->flags&CODEC_FLAG_4MV){
1697  const int wrap= s->b8_stride;
1698 
1699  /* clip / convert to intra 8x8 type MVs */
1700  for(y=0; y<s->mb_height; y++){
1701  int xy= y*2*wrap;
1702  int i= y*s->mb_stride;
1703  int x;
1704 
1705  for(x=0; x<s->mb_width; x++){
1706  if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1707  int block;
1708  for(block=0; block<4; block++){
1709  int off= (block& 1) + (block>>1)*wrap;
1710  int mx = s->current_picture.motion_val[0][ xy + off ][0];
1711  int my = s->current_picture.motion_val[0][ xy + off ][1];
1712 
1713  if( mx >=range || mx <-range
1714  || my >=range || my <-range){
1715  s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1716  s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1717  s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1718  }
1719  }
1720  }
1721  xy+=2;
1722  i++;
1723  }
1724  }
1725  }
1726 }
1727 
1728 /**
1729  *
1730  * @param truncate 1 for truncation, 0 for using intra
1731  */
1732 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1733  int16_t (*mv_table)[2], int f_code, int type, int truncate)
1734 {
1735  MotionEstContext * const c= &s->me;
1736  int y, h_range, v_range;
1737 
1738  // RAL: 8 in MPEG-1, 16 in MPEG-4
1739  int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1740 
1741  if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1742 
1743  h_range= range;
1744  v_range= field_select_table ? range>>1 : range;
1745 
1746  /* clip / convert to intra 16x16 type MVs */
1747  for(y=0; y<s->mb_height; y++){
1748  int x;
1749  int xy= y*s->mb_stride;
1750  for(x=0; x<s->mb_width; x++){
1751  if (s->mb_type[xy] & type){ // RAL: "type" test added...
1752  if(field_select_table==NULL || field_select_table[xy] == field_select){
1753  if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1754  || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1755 
1756  if(truncate){
1757  if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1758  else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1759  if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1760  else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1761  }else{
1762  s->mb_type[xy] &= ~type;
1763  s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1764  mv_table[xy][0]=
1765  mv_table[xy][1]= 0;
1766  }
1767  }
1768  }
1769  }
1770  xy++;
1771  }
1772  }
1773 }
Generated on Sun Jul 20 2014 23:05:51 for FFmpeg by   doxygen 1.8.2