?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
lavfi: switch to AVFrame.
[FFMpeg-mirror/mplayer-patches.git] / libavcodec / motion_est.c
blob11285825a67969dc8399420aa9dd430fc2e4b963
1 /*
2 * Motion estimation
3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
5 *
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
7 *
8 * This file is part of Libav.
9 *
10 * Libav is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * Libav is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with Libav; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 */
24
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 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
116 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
117 const int time_pp= s->pp_time;
118 const int time_pb= s->pb_time;
119 const int mask= 2*qpel+1;
120 if(s->mv_type==MV_TYPE_8X8){
121 int i;
122 for(i=0; i<4; i++){
123 int fx = c->direct_basis_mv[i][0] + hx;
124 int fy = c->direct_basis_mv[i][1] + hy;
125 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
126 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
127 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
128 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
129
130 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
131 if(qpel){
132 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
133 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
134 }else{
135 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
136 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
137 }
138 }
139 }else{
140 int fx = c->direct_basis_mv[0][0] + hx;
141 int fy = c->direct_basis_mv[0][1] + hy;
142 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
143 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
144 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
145 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
146
147 if(qpel){
148 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
149 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
150 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
151 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
152 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
153 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
154 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
155 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
156 }else{
157 assert((fx>>1) + 16*s->mb_x >= -16);
158 assert((fy>>1) + 16*s->mb_y >= -16);
159 assert((fx>>1) + 16*s->mb_x <= s->width);
160 assert((fy>>1) + 16*s->mb_y <= s->height);
161 assert((bx>>1) + 16*s->mb_x >= -16);
162 assert((by>>1) + 16*s->mb_y >= -16);
163 assert((bx>>1) + 16*s->mb_x <= s->width);
164 assert((by>>1) + 16*s->mb_y <= s->height);
165
166 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
167 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
168 }
169 }
170 d = cmp_func(s, c->temp, src[0], stride, 16);
171 }else
172 d= 256*256*256*32;
173 return d;
174 }
175
176 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
177 const int size, const int h, int ref_index, int src_index,
178 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
179 MotionEstContext * const c= &s->me;
180 const int stride= c->stride;
181 const int uvstride= c->uvstride;
182 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
183 const int hx= subx + (x<<(1+qpel));
184 const int hy= suby + (y<<(1+qpel));
185 uint8_t * const * const ref= c->ref[ref_index];
186 uint8_t * const * const src= c->src[src_index];
187 int d;
188 //FIXME check chroma 4mv, (no crashes ...)
189 int uvdxy; /* no, it might not be used uninitialized */
190 if(dxy){
191 if(qpel){
192 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
193 if(chroma){
194 int cx= hx/2;
195 int cy= hy/2;
196 cx= (cx>>1)|(cx&1);
197 cy= (cy>>1)|(cy&1);
198 uvdxy= (cx&1) + 2*(cy&1);
199 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
200 }
201 }else{
202 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
203 if(chroma)
204 uvdxy= dxy | (x&1) | (2*(y&1));
205 }
206 d = cmp_func(s, c->temp, src[0], stride, h);
207 }else{
208 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
209 if(chroma)
210 uvdxy= (x&1) + 2*(y&1);
211 }
212 if(chroma){
213 uint8_t * const uvtemp= c->temp + 16*stride;
214 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
215 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
216 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
217 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
218 }
219 return d;
220 }
221
222 static int cmp_simple(MpegEncContext *s, const int x, const int y,
223 int ref_index, int src_index,
224 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
225 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
226 }
227
228 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
229 const int size, const int h, int ref_index, int src_index,
230 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
231 if(flags&FLAG_DIRECT){
232 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
233 }else{
234 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
235 }
236 }
237
238 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
239 const int size, const int h, int ref_index, int src_index,
240 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
241 if(flags&FLAG_DIRECT){
242 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
243 }else{
244 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA);
245 }
246 }
247
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) < -ME_MAP_SIZE){
301 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
302 return -1;
303 }
304 if (s->me_method != ME_ZERO &&
305 s->me_method != ME_EPZS &&
306 s->me_method != ME_X1) {
307 av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
308 return -1;
309 }
310
311 c->avctx= s->avctx;
312
313 if(cache_size < 2*dia_size && !c->stride){
314 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
315 }
316
317 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
318 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
319 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
320 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
321
322 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
323 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
324 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
325
326 /*FIXME s->no_rounding b_type*/
327 if(s->flags&CODEC_FLAG_QPEL){
328 c->sub_motion_search= qpel_motion_search;
329 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
330 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
331 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
332 }else{
333 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
334 c->sub_motion_search= hpel_motion_search;
335 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
336 && c->avctx-> me_cmp == FF_CMP_SAD
337 && c->avctx-> mb_cmp == FF_CMP_SAD)
338 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
339 else
340 c->sub_motion_search= hpel_motion_search;
341 }
342 c->hpel_avg= s->dsp.avg_pixels_tab;
343 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
344 else c->hpel_put= s->dsp.put_pixels_tab;
345
346 if(s->linesize){
347 c->stride = s->linesize;
348 c->uvstride= s->uvlinesize;
349 }else{
350 c->stride = 16*s->mb_width + 32;
351 c->uvstride= 8*s->mb_width + 16;
352 }
353
354 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
355 * not have yet, and even if we had, the motion estimation code
356 * does not expect it. */
357 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
358 s->dsp.me_cmp[2]= zero_cmp;
359 }
360 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
361 s->dsp.me_sub_cmp[2]= zero_cmp;
362 }
363 c->hpel_put[2][0]= c->hpel_put[2][1]=
364 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
365
366 if(s->codec_id == AV_CODEC_ID_H261){
367 c->sub_motion_search= no_sub_motion_search;
368 }
369
370 return 0;
371 }
372
373 #define CHECK_SAD_HALF_MV(suffix, x, y) \
374 {\
375 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
376 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
377 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
378 }
379
380 static int sad_hpel_motion_search(MpegEncContext * s,
381 int *mx_ptr, int *my_ptr, int dmin,
382 int src_index, int ref_index,
383 int size, int h)
384 {
385 MotionEstContext * const c= &s->me;
386 const int penalty_factor= c->sub_penalty_factor;
387 int mx, my, dminh;
388 uint8_t *pix, *ptr;
389 int stride= c->stride;
390 const int flags= c->sub_flags;
391 LOAD_COMMON
392
393 assert(flags == 0);
394
395 if(c->skip){
396 *mx_ptr = 0;
397 *my_ptr = 0;
398 return dmin;
399 }
400
401 pix = c->src[src_index][0];
402
403 mx = *mx_ptr;
404 my = *my_ptr;
405 ptr = c->ref[ref_index][0] + (my * stride) + mx;
406
407 dminh = dmin;
408
409 if (mx > xmin && mx < xmax &&
410 my > ymin && my < ymax) {
411 int dx=0, dy=0;
412 int d, pen_x, pen_y;
413 const int index= (my<<ME_MAP_SHIFT) + mx;
414 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
415 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
416 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
417 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
418 mx<<=1;
419 my<<=1;
420
421
422 pen_x= pred_x + mx;
423 pen_y= pred_y + my;
424
425 ptr-= stride;
426 if(t<=b){
427 CHECK_SAD_HALF_MV(y2 , 0, -1)
428 if(l<=r){
429 CHECK_SAD_HALF_MV(xy2, -1, -1)
430 if(t+r<=b+l){
431 CHECK_SAD_HALF_MV(xy2, +1, -1)
432 ptr+= stride;
433 }else{
434 ptr+= stride;
435 CHECK_SAD_HALF_MV(xy2, -1, +1)
436 }
437 CHECK_SAD_HALF_MV(x2 , -1, 0)
438 }else{
439 CHECK_SAD_HALF_MV(xy2, +1, -1)
440 if(t+l<=b+r){
441 CHECK_SAD_HALF_MV(xy2, -1, -1)
442 ptr+= stride;
443 }else{
444 ptr+= stride;
445 CHECK_SAD_HALF_MV(xy2, +1, +1)
446 }
447 CHECK_SAD_HALF_MV(x2 , +1, 0)
448 }
449 }else{
450 if(l<=r){
451 if(t+l<=b+r){
452 CHECK_SAD_HALF_MV(xy2, -1, -1)
453 ptr+= stride;
454 }else{
455 ptr+= stride;
456 CHECK_SAD_HALF_MV(xy2, +1, +1)
457 }
458 CHECK_SAD_HALF_MV(x2 , -1, 0)
459 CHECK_SAD_HALF_MV(xy2, -1, +1)
460 }else{
461 if(t+r<=b+l){
462 CHECK_SAD_HALF_MV(xy2, +1, -1)
463 ptr+= stride;
464 }else{
465 ptr+= stride;
466 CHECK_SAD_HALF_MV(xy2, -1, +1)
467 }
468 CHECK_SAD_HALF_MV(x2 , +1, 0)
469 CHECK_SAD_HALF_MV(xy2, +1, +1)
470 }
471 CHECK_SAD_HALF_MV(y2 , 0, +1)
472 }
473 mx+=dx;
474 my+=dy;
475
476 }else{
477 mx<<=1;
478 my<<=1;
479 }
480
481 *mx_ptr = mx;
482 *my_ptr = my;
483 return dminh;
484 }
485
486 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
487 {
488 const int xy= s->mb_x + s->mb_y*s->mb_stride;
489
490 s->p_mv_table[xy][0] = mx;
491 s->p_mv_table[xy][1] = my;
492
493 /* has already been set to the 4 MV if 4MV is done */
494 if(mv4){
495 int mot_xy= s->block_index[0];
496
497 s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
498 s->current_picture.f.motion_val[0][mot_xy ][1] = my;
499 s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
500 s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
501
502 mot_xy += s->b8_stride;
503 s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
504 s->current_picture.f.motion_val[0][mot_xy ][1] = my;
505 s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
506 s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
507 }
508 }
509
510 /**
511 * get fullpel ME search limits.
512 */
513 static inline void get_limits(MpegEncContext *s, int x, int y)
514 {
515 MotionEstContext * const c= &s->me;
516 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
517 /*
518 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
519 else c->range= 16;
520 */
521 if (s->unrestricted_mv) {
522 c->xmin = - x - 16;
523 c->ymin = - y - 16;
524 c->xmax = - x + s->mb_width *16;
525 c->ymax = - y + s->mb_height*16;
526 } else if (s->out_format == FMT_H261){
527 // Search range of H261 is different from other codec standards
528 c->xmin = (x > 15) ? - 15 : 0;
529 c->ymin = (y > 15) ? - 15 : 0;
530 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
531 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
532 } else {
533 c->xmin = - x;
534 c->ymin = - y;
535 c->xmax = - x + s->mb_width *16 - 16;
536 c->ymax = - y + s->mb_height*16 - 16;
537 }
538 if(range){
539 c->xmin = FFMAX(c->xmin,-range);
540 c->xmax = FFMIN(c->xmax, range);
541 c->ymin = FFMAX(c->ymin,-range);
542 c->ymax = FFMIN(c->ymax, range);
543 }
544 }
545
546 static inline void init_mv4_ref(MotionEstContext *c){
547 const int stride= c->stride;
548
549 c->ref[1][0] = c->ref[0][0] + 8;
550 c->ref[2][0] = c->ref[0][0] + 8*stride;
551 c->ref[3][0] = c->ref[2][0] + 8;
552 c->src[1][0] = c->src[0][0] + 8;
553 c->src[2][0] = c->src[0][0] + 8*stride;
554 c->src[3][0] = c->src[2][0] + 8;
555 }
556
557 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
558 {
559 MotionEstContext * const c= &s->me;
560 const int size= 1;
561 const int h=8;
562 int block;
563 int P[10][2];
564 int dmin_sum=0, mx4_sum=0, my4_sum=0;
565 int same=1;
566 const int stride= c->stride;
567 uint8_t *mv_penalty= c->current_mv_penalty;
568
569 init_mv4_ref(c);
570
571 for(block=0; block<4; block++){
572 int mx4, my4;
573 int pred_x4, pred_y4;
574 int dmin4;
575 static const int off[4]= {2, 1, 1, -1};
576 const int mot_stride = s->b8_stride;
577 const int mot_xy = s->block_index[block];
578
579 P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
580 P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
581
582 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
583
584 /* special case for first line */
585 if (s->first_slice_line && block<2) {
586 c->pred_x= pred_x4= P_LEFT[0];
587 c->pred_y= pred_y4= P_LEFT[1];
588 } else {
589 P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
590 P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
591 P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][0];
592 P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][1];
593 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
594 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
595 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
596 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
597
598 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
599 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
600
601 c->pred_x= pred_x4 = P_MEDIAN[0];
602 c->pred_y= pred_y4 = P_MEDIAN[1];
603 }
604 P_MV1[0]= mx;
605 P_MV1[1]= my;
606
607 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
608
609 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
610
611 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
612 int dxy;
613 const int offset= ((block&1) + (block>>1)*stride)*8;
614 uint8_t *dest_y = c->scratchpad + offset;
615 if(s->quarter_sample){
616 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
617 dxy = ((my4 & 3) << 2) | (mx4 & 3);
618
619 if(s->no_rounding)
620 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
621 else
622 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
623 }else{
624 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
625 dxy = ((my4 & 1) << 1) | (mx4 & 1);
626
627 if(s->no_rounding)
628 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
629 else
630 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
631 }
632 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
633 }else
634 dmin_sum+= dmin4;
635
636 if(s->quarter_sample){
637 mx4_sum+= mx4/2;
638 my4_sum+= my4/2;
639 }else{
640 mx4_sum+= mx4;
641 my4_sum+= my4;
642 }
643
644 s->current_picture.f.motion_val[0][s->block_index[block]][0] = mx4;
645 s->current_picture.f.motion_val[0][s->block_index[block]][1] = my4;
646
647 if(mx4 != mx || my4 != my) same=0;
648 }
649
650 if(same)
651 return INT_MAX;
652
653 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
654 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);
655 }
656
657 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
658 int dxy;
659 int mx, my;
660 int offset;
661
662 mx= ff_h263_round_chroma(mx4_sum);
663 my= ff_h263_round_chroma(my4_sum);
664 dxy = ((my & 1) << 1) | (mx & 1);
665
666 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
667
668 if(s->no_rounding){
669 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
670 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
671 }else{
672 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
673 s->dsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
674 }
675
676 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);
677 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);
678 }
679
680 c->pred_x= mx;
681 c->pred_y= my;
682
683 switch(c->avctx->mb_cmp&0xFF){
684 /*case FF_CMP_SSE:
685 return dmin_sum+ 32*s->qscale*s->qscale;*/
686 case FF_CMP_RD:
687 return dmin_sum;
688 default:
689 return dmin_sum+ 11*c->mb_penalty_factor;
690 }
691 }
692
693 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
694 MotionEstContext * const c= &s->me;
695
696 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
697 c->src[1][0] = c->src[0][0] + s->linesize;
698 if(c->flags & FLAG_CHROMA){
699 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
700 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
701 c->src[1][1] = c->src[0][1] + s->uvlinesize;
702 c->src[1][2] = c->src[0][2] + s->uvlinesize;
703 }
704 }
705
706 static int interlaced_search(MpegEncContext *s, int ref_index,
707 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
708 {
709 MotionEstContext * const c= &s->me;
710 const int size=0;
711 const int h=8;
712 int block;
713 int P[10][2];
714 uint8_t * const mv_penalty= c->current_mv_penalty;
715 int same=1;
716 const int stride= 2*s->linesize;
717 int dmin_sum= 0;
718 const int mot_stride= s->mb_stride;
719 const int xy= s->mb_x + s->mb_y*mot_stride;
720
721 c->ymin>>=1;
722 c->ymax>>=1;
723 c->stride<<=1;
724 c->uvstride<<=1;
725 init_interlaced_ref(s, ref_index);
726
727 for(block=0; block<2; block++){
728 int field_select;
729 int best_dmin= INT_MAX;
730 int best_field= -1;
731
732 for(field_select=0; field_select<2; field_select++){
733 int dmin, mx_i, my_i;
734 int16_t (*mv_table)[2]= mv_tables[block][field_select];
735
736 if(user_field_select){
737 assert(field_select==0 || field_select==1);
738 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
739 if(field_select_tables[block][xy] != field_select)
740 continue;
741 }
742
743 P_LEFT[0] = mv_table[xy - 1][0];
744 P_LEFT[1] = mv_table[xy - 1][1];
745 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
746
747 c->pred_x= P_LEFT[0];
748 c->pred_y= P_LEFT[1];
749
750 if(!s->first_slice_line){
751 P_TOP[0] = mv_table[xy - mot_stride][0];
752 P_TOP[1] = mv_table[xy - mot_stride][1];
753 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
754 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
755 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
756 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
757 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
758 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
759
760 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
761 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
762 }
763 P_MV1[0]= mx; //FIXME not correct if block != field_select
764 P_MV1[1]= my / 2;
765
766 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
767
768 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
769
770 mv_table[xy][0]= mx_i;
771 mv_table[xy][1]= my_i;
772
773 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
774 int dxy;
775
776 //FIXME chroma ME
777 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
778 dxy = ((my_i & 1) << 1) | (mx_i & 1);
779
780 if(s->no_rounding){
781 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
782 }else{
783 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
784 }
785 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
786 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
787 }else
788 dmin+= c->mb_penalty_factor; //field_select bits
789
790 dmin += field_select != block; //slightly prefer same field
791
792 if(dmin < best_dmin){
793 best_dmin= dmin;
794 best_field= field_select;
795 }
796 }
797 {
798 int16_t (*mv_table)[2]= mv_tables[block][best_field];
799
800 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
801 if(mv_table[xy][1]&1) same=0;
802 if(mv_table[xy][1]*2 != my) same=0;
803 if(best_field != block) same=0;
804 }
805
806 field_select_tables[block][xy]= best_field;
807 dmin_sum += best_dmin;
808 }
809
810 c->ymin<<=1;
811 c->ymax<<=1;
812 c->stride>>=1;
813 c->uvstride>>=1;
814
815 if(same)
816 return INT_MAX;
817
818 switch(c->avctx->mb_cmp&0xFF){
819 /*case FF_CMP_SSE:
820 return dmin_sum+ 32*s->qscale*s->qscale;*/
821 case FF_CMP_RD:
822 return dmin_sum;
823 default:
824 return dmin_sum+ 11*c->mb_penalty_factor;
825 }
826 }
827
828 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
829 int ymax= s->me.ymax>>interlaced;
830 int ymin= s->me.ymin>>interlaced;
831
832 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
833 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
834 if(mv[1] < ymin) mv[1] = ymin;
835 if(mv[1] > ymax) mv[1] = ymax;
836 }
837
838 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
839 MotionEstContext * const c= &s->me;
840 Picture *p= s->current_picture_ptr;
841 int mb_xy= mb_x + mb_y*s->mb_stride;
842 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
843 int mb_type= s->current_picture.f.mb_type[mb_xy];
844 int flags= c->flags;
845 int shift= (flags&FLAG_QPEL) + 1;
846 int mask= (1<<shift)-1;
847 int x, y, i;
848 int d=0;
849 me_cmp_func cmpf= s->dsp.sse[0];
850 me_cmp_func chroma_cmpf= s->dsp.sse[1];
851
852 if(p_type && USES_LIST(mb_type, 1)){
853 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
854 return INT_MAX/2;
855 }
856 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
857
858 for(i=0; i<4; i++){
859 int xy= s->block_index[i];
860 clip_input_mv(s, p->f.motion_val[0][xy], !!IS_INTERLACED(mb_type));
861 clip_input_mv(s, p->f.motion_val[1][xy], !!IS_INTERLACED(mb_type));
862 }
863
864 if(IS_INTERLACED(mb_type)){
865 int xy2= xy + s->b8_stride;
866 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
867 c->stride<<=1;
868 c->uvstride<<=1;
869
870 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
871 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
872 return INT_MAX/2;
873 }
874
875 if(USES_LIST(mb_type, 0)){
876 int field_select0= p->f.ref_index[0][4*mb_xy ];
877 int field_select1= p->f.ref_index[0][4*mb_xy+2];
878 assert(field_select0==0 ||field_select0==1);
879 assert(field_select1==0 ||field_select1==1);
880 init_interlaced_ref(s, 0);
881
882 if(p_type){
883 s->p_field_select_table[0][mb_xy]= field_select0;
884 s->p_field_select_table[1][mb_xy]= field_select1;
885 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
886 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
887 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
888 }else{
889 s->b_field_select_table[0][0][mb_xy]= field_select0;
890 s->b_field_select_table[0][1][mb_xy]= field_select1;
891 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
892 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
893 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
894 }
895
896 x = p->f.motion_val[0][xy ][0];
897 y = p->f.motion_val[0][xy ][1];
898 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
899 x = p->f.motion_val[0][xy2][0];
900 y = p->f.motion_val[0][xy2][1];
901 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
902 }
903 if(USES_LIST(mb_type, 1)){
904 int field_select0 = p->f.ref_index[1][4 * mb_xy ];
905 int field_select1 = p->f.ref_index[1][4 * mb_xy + 2];
906 assert(field_select0==0 ||field_select0==1);
907 assert(field_select1==0 ||field_select1==1);
908 init_interlaced_ref(s, 2);
909
910 s->b_field_select_table[1][0][mb_xy]= field_select0;
911 s->b_field_select_table[1][1][mb_xy]= field_select1;
912 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy ];
913 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy2];
914 if(USES_LIST(mb_type, 0)){
915 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
916 }else{
917 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
918 }
919
920 x = p->f.motion_val[1][xy ][0];
921 y = p->f.motion_val[1][xy ][1];
922 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
923 x = p->f.motion_val[1][xy2][0];
924 y = p->f.motion_val[1][xy2][1];
925 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
926 //FIXME bidir scores
927 }
928 c->stride>>=1;
929 c->uvstride>>=1;
930 }else if(IS_8X8(mb_type)){
931 if(!(s->flags & CODEC_FLAG_4MV)){
932 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
933 return INT_MAX/2;
934 }
935 cmpf= s->dsp.sse[1];
936 chroma_cmpf= s->dsp.sse[1];
937 init_mv4_ref(c);
938 for(i=0; i<4; i++){
939 xy= s->block_index[i];
940 x= p->f.motion_val[0][xy][0];
941 y= p->f.motion_val[0][xy][1];
942 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
943 }
944 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
945 }else{
946 if(USES_LIST(mb_type, 0)){
947 if(p_type){
948 *(uint32_t*)s->p_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
949 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
950 }else if(USES_LIST(mb_type, 1)){
951 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
952 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
953 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
954 }else{
955 *(uint32_t*)s->b_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
956 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
957 }
958 x = p->f.motion_val[0][xy][0];
959 y = p->f.motion_val[0][xy][1];
960 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
961 }else if(USES_LIST(mb_type, 1)){
962 *(uint32_t*)s->b_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
963 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
964
965 x = p->f.motion_val[1][xy][0];
966 y = p->f.motion_val[1][xy][1];
967 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
968 }else
969 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
970 }
971 return d;
972 }
973
974 static inline int get_penalty_factor(int lambda, int lambda2, int type){
975 switch(type&0xFF){
976 default:
977 case FF_CMP_SAD:
978 return lambda>>FF_LAMBDA_SHIFT;
979 case FF_CMP_DCT:
980 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
981 case FF_CMP_SATD:
982 case FF_CMP_DCT264:
983 return (2*lambda)>>FF_LAMBDA_SHIFT;
984 case FF_CMP_RD:
985 case FF_CMP_PSNR:
986 case FF_CMP_SSE:
987 case FF_CMP_NSSE:
988 return lambda2>>FF_LAMBDA_SHIFT;
989 case FF_CMP_BIT:
990 return 1;
991 }
992 }
993
994 void ff_estimate_p_frame_motion(MpegEncContext * s,
995 int mb_x, int mb_y)
996 {
997 MotionEstContext * const c= &s->me;
998 uint8_t *pix, *ppix;
999 int sum, mx, my, dmin;
1000 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1001 int vard; ///< sum of squared differences with the estimated motion vector
1002 int P[10][2];
1003 const int shift= 1+s->quarter_sample;
1004 int mb_type=0;
1005 Picture * const pic= &s->current_picture;
1006
1007 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1008
1009 assert(s->quarter_sample==0 || s->quarter_sample==1);
1010 assert(s->linesize == c->stride);
1011 assert(s->uvlinesize == c->uvstride);
1012
1013 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1014 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1015 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1016 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1017
1018 get_limits(s, 16*mb_x, 16*mb_y);
1019 c->skip=0;
1020
1021 /* intra / predictive decision */
1022 pix = c->src[0][0];
1023 sum = s->dsp.pix_sum(pix, s->linesize);
1024 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
1025
1026 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1027 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1028 c->mb_var_sum_temp += (varc+128)>>8;
1029
1030 if(c->avctx->me_threshold){
1031 vard= check_input_motion(s, mb_x, mb_y, 1);
1032
1033 if((vard+128)>>8 < c->avctx->me_threshold){
1034 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1035 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1036 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1037 c->mc_mb_var_sum_temp += (vard+128)>>8;
1038 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1039 return;
1040 }
1041 if((vard+128)>>8 < c->avctx->mb_threshold)
1042 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1043 }
1044
1045 switch(s->me_method) {
1046 case ME_ZERO:
1047 default:
1048 mx = 0;
1049 my = 0;
1050 dmin = 0;
1051 break;
1052 case ME_X1:
1053 case ME_EPZS:
1054 {
1055 const int mot_stride = s->b8_stride;
1056 const int mot_xy = s->block_index[0];
1057
1058 P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
1059 P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
1060
1061 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1062
1063 if(!s->first_slice_line) {
1064 P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
1065 P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
1066 P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][0];
1067 P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][1];
1068 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1069 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1070 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1071
1072 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1073 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1074
1075 if(s->out_format == FMT_H263){
1076 c->pred_x = P_MEDIAN[0];
1077 c->pred_y = P_MEDIAN[1];
1078 }else { /* mpeg1 at least */
1079 c->pred_x= P_LEFT[0];
1080 c->pred_y= P_LEFT[1];
1081 }
1082 }else{
1083 c->pred_x= P_LEFT[0];
1084 c->pred_y= P_LEFT[1];
1085 }
1086
1087 }
1088 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1089
1090 break;
1091 }
1092
1093 /* At this point (mx,my) are full-pell and the relative displacement */
1094 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1095
1096 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1097
1098 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1099 c->mc_mb_var_sum_temp += (vard+128)>>8;
1100
1101 if(mb_type){
1102 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1103 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1104 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1105
1106 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1107 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1108 set_p_mv_tables(s, mx, my, 1);
1109 }else{
1110 mx <<=shift;
1111 my <<=shift;
1112 }
1113 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1114 h263_mv4_search(s, mx, my, shift);
1115
1116 set_p_mv_tables(s, mx, my, 0);
1117 }
1118 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1119 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1120 }
1121 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1122 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1123 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1124 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1125
1126 if (vard*2 + 200*256 > varc)
1127 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1128 if (varc*2 + 200*256 > vard || s->qscale > 24){
1129 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1130 mb_type|= CANDIDATE_MB_TYPE_INTER;
1131 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1132 if(s->flags&CODEC_FLAG_MV0)
1133 if(mx || my)
1134 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1135 }else{
1136 mx <<=shift;
1137 my <<=shift;
1138 }
1139 if((s->flags&CODEC_FLAG_4MV)
1140 && !c->skip && varc>50<<8 && vard>10<<8){
1141 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1142 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1143
1144 set_p_mv_tables(s, mx, my, 0);
1145 }else
1146 set_p_mv_tables(s, mx, my, 1);
1147 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1148 && !c->skip){ //FIXME varc/d checks
1149 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1150 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1151 }
1152 }else{
1153 int intra_score, i;
1154 mb_type= CANDIDATE_MB_TYPE_INTER;
1155
1156 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1157 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1158 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1159
1160 if((s->flags&CODEC_FLAG_4MV)
1161 && !c->skip && varc>50<<8 && vard>10<<8){
1162 int dmin4= h263_mv4_search(s, mx, my, shift);
1163 if(dmin4 < dmin){
1164 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1165 dmin=dmin4;
1166 }
1167 }
1168 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1169 && !c->skip){ //FIXME varc/d checks
1170 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1171 if(dmin_i < dmin){
1172 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1173 dmin= dmin_i;
1174 }
1175 }
1176
1177 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1178
1179 /* get intra luma score */
1180 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1181 intra_score= varc - 500;
1182 }else{
1183 unsigned mean = (sum+128)>>8;
1184 mean*= 0x01010101;
1185
1186 for(i=0; i<16; i++){
1187 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1188 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1189 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1190 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1191 }
1192
1193 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1194 }
1195 intra_score += c->mb_penalty_factor*16;
1196
1197 if(intra_score < dmin){
1198 mb_type= CANDIDATE_MB_TYPE_INTRA;
1199 s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1200 }else
1201 s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1202
1203 {
1204 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1205 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1206 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1207 }
1208 }
1209
1210 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1211 }
1212
1213 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1214 int mb_x, int mb_y)
1215 {
1216 MotionEstContext * const c= &s->me;
1217 int mx, my, dmin;
1218 int P[10][2];
1219 const int shift= 1+s->quarter_sample;
1220 const int xy= mb_x + mb_y*s->mb_stride;
1221 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1222
1223 assert(s->quarter_sample==0 || s->quarter_sample==1);
1224
1225 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1226 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1227
1228 get_limits(s, 16*mb_x, 16*mb_y);
1229 c->skip=0;
1230
1231 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1232 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1233
1234 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1235
1236 /* special case for first line */
1237 if (s->first_slice_line) {
1238 c->pred_x= P_LEFT[0];
1239 c->pred_y= P_LEFT[1];
1240 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1241 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1242 } else {
1243 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1244 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1245 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1246 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1247 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1248 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1249 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1250
1251 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1252 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1253
1254 c->pred_x = P_MEDIAN[0];
1255 c->pred_y = P_MEDIAN[1];
1256 }
1257
1258 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1259
1260 s->p_mv_table[xy][0] = mx<<shift;
1261 s->p_mv_table[xy][1] = my<<shift;
1262
1263 return dmin;
1264 }
1265
1266 static int ff_estimate_motion_b(MpegEncContext * s,
1267 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1268 {
1269 MotionEstContext * const c= &s->me;
1270 int mx, my, dmin;
1271 int P[10][2];
1272 const int shift= 1+s->quarter_sample;
1273 const int mot_stride = s->mb_stride;
1274 const int mot_xy = mb_y*mot_stride + mb_x;
1275 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1276 int mv_scale;
1277
1278 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1279 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1280 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1281 c->current_mv_penalty= mv_penalty;
1282
1283 get_limits(s, 16*mb_x, 16*mb_y);
1284
1285 switch(s->me_method) {
1286 case ME_ZERO:
1287 default:
1288 mx = 0;
1289 my = 0;
1290 dmin = 0;
1291 break;
1292 case ME_X1:
1293 case ME_EPZS:
1294 P_LEFT[0] = mv_table[mot_xy - 1][0];
1295 P_LEFT[1] = mv_table[mot_xy - 1][1];
1296
1297 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1298
1299 /* special case for first line */
1300 if (!s->first_slice_line) {
1301 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1302 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1303 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1304 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1305 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1306 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1307 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1308
1309 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1310 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1311 }
1312 c->pred_x = P_LEFT[0];
1313 c->pred_y = P_LEFT[1];
1314
1315 if(mv_table == s->b_forw_mv_table){
1316 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1317 }else{
1318 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1319 }
1320
1321 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1322
1323 break;
1324 }
1325
1326 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1327
1328 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1329 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1330
1331 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1332 mv_table[mot_xy][0]= mx;
1333 mv_table[mot_xy][1]= my;
1334
1335 return dmin;
1336 }
1337
1338 static inline int check_bidir_mv(MpegEncContext * s,
1339 int motion_fx, int motion_fy,
1340 int motion_bx, int motion_by,
1341 int pred_fx, int pred_fy,
1342 int pred_bx, int pred_by,
1343 int size, int h)
1344 {
1345 //FIXME optimize?
1346 //FIXME better f_code prediction (max mv & distance)
1347 //FIXME pointers
1348 MotionEstContext * const c= &s->me;
1349 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1350 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1351 int stride= c->stride;
1352 uint8_t *dest_y = c->scratchpad;
1353 uint8_t *ptr;
1354 int dxy;
1355 int src_x, src_y;
1356 int fbmin;
1357 uint8_t **src_data= c->src[0];
1358 uint8_t **ref_data= c->ref[0];
1359 uint8_t **ref2_data= c->ref[2];
1360
1361 if(s->quarter_sample){
1362 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1363 src_x = motion_fx >> 2;
1364 src_y = motion_fy >> 2;
1365
1366 ptr = ref_data[0] + (src_y * stride) + src_x;
1367 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1368
1369 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1370 src_x = motion_bx >> 2;
1371 src_y = motion_by >> 2;
1372
1373 ptr = ref2_data[0] + (src_y * stride) + src_x;
1374 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1375 }else{
1376 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1377 src_x = motion_fx >> 1;
1378 src_y = motion_fy >> 1;
1379
1380 ptr = ref_data[0] + (src_y * stride) + src_x;
1381 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1382
1383 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1384 src_x = motion_bx >> 1;
1385 src_y = motion_by >> 1;
1386
1387 ptr = ref2_data[0] + (src_y * stride) + src_x;
1388 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1389 }
1390
1391 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1392 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1393 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1394
1395 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1396 }
1397 //FIXME CHROMA !!!
1398
1399 return fbmin;
1400 }
1401
1402 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1403 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1404 {
1405 MotionEstContext * const c= &s->me;
1406 const int mot_stride = s->mb_stride;
1407 const int xy = mb_y *mot_stride + mb_x;
1408 int fbmin;
1409 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1410 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1411 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1412 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1413 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1414 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1415 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1416 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1417 const int flags= c->sub_flags;
1418 const int qpel= flags&FLAG_QPEL;
1419 const int shift= 1+qpel;
1420 const int xmin= c->xmin<<shift;
1421 const int ymin= c->ymin<<shift;
1422 const int xmax= c->xmax<<shift;
1423 const int ymax= c->ymax<<shift;
1424 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1425 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1426 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1427 uint8_t map[256] = { 0 };
1428
1429 map[hashidx&255] = 1;
1430
1431 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1432 motion_bx, motion_by,
1433 pred_fx, pred_fy,
1434 pred_bx, pred_by,
1435 0, 16);
1436
1437 if(s->avctx->bidir_refine){
1438 int end;
1439 static const uint8_t limittab[5]={0,8,32,64,80};
1440 const int limit= limittab[s->avctx->bidir_refine];
1441 static const int8_t vect[][4]={
1442 { 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},
1443
1444 { 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},
1445 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1446 { 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},
1447 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1448
1449 { 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},
1450 { 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},
1451 { 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},
1452 { 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},
1453
1454 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1455 { 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},
1456 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1457 };
1458 static const uint8_t hash[]={
1459 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),
1460
1461 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),
1462 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1463 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),
1464 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1465
1466 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),
1467 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),
1468 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),
1469 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),
1470
1471 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1472 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),
1473 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),
1474 };
1475
1476 #define CHECK_BIDIR(fx,fy,bx,by)\
1477 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1478 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1479 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1480 int score;\
1481 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1482 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);\
1483 if(score < fbmin){\
1484 hashidx += HASH(fx,fy,bx,by);\
1485 fbmin= score;\
1486 motion_fx+=fx;\
1487 motion_fy+=fy;\
1488 motion_bx+=bx;\
1489 motion_by+=by;\
1490 end=0;\
1491 }\
1492 }
1493 #define CHECK_BIDIR2(a,b,c,d)\
1494 CHECK_BIDIR(a,b,c,d)\
1495 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1496
1497 do{
1498 int i;
1499 int borderdist=0;
1500 end=1;
1501
1502 CHECK_BIDIR2(0,0,0,1)
1503 CHECK_BIDIR2(0,0,1,0)
1504 CHECK_BIDIR2(0,1,0,0)
1505 CHECK_BIDIR2(1,0,0,0)
1506
1507 for(i=8; i<limit; i++){
1508 int fx= motion_fx+vect[i][0];
1509 int fy= motion_fy+vect[i][1];
1510 int bx= motion_bx+vect[i][2];
1511 int by= motion_by+vect[i][3];
1512 if(borderdist<=0){
1513 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1514 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1515 if((a|b) < 0)
1516 map[(hashidx+hash[i])&255] = 1;
1517 }
1518 if(!map[(hashidx+hash[i])&255]){
1519 int score;
1520 map[(hashidx+hash[i])&255] = 1;
1521 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1522 if(score < fbmin){
1523 hashidx += hash[i];
1524 fbmin= score;
1525 motion_fx=fx;
1526 motion_fy=fy;
1527 motion_bx=bx;
1528 motion_by=by;
1529 end=0;
1530 borderdist--;
1531 if(borderdist<=0){
1532 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1533 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1534 borderdist= FFMIN(a,b);
1535 }
1536 }
1537 }
1538 }
1539 }while(!end);
1540 }
1541
1542 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1543 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1544 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1545 s->b_bidir_back_mv_table[xy][1]= motion_by;
1546
1547 return fbmin;
1548 }
1549
1550 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1551 {
1552 MotionEstContext * const c= &s->me;
1553 int P[10][2];
1554 const int mot_stride = s->mb_stride;
1555 const int mot_xy = mb_y*mot_stride + mb_x;
1556 const int shift= 1+s->quarter_sample;
1557 int dmin, i;
1558 const int time_pp= s->pp_time;
1559 const int time_pb= s->pb_time;
1560 int mx, my, xmin, xmax, ymin, ymax;
1561 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1562
1563 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1564 ymin= xmin=(-32)>>shift;
1565 ymax= xmax= 31>>shift;
1566
1567 if (IS_8X8(s->next_picture.f.mb_type[mot_xy])) {
1568 s->mv_type= MV_TYPE_8X8;
1569 }else{
1570 s->mv_type= MV_TYPE_16X16;
1571 }
1572
1573 for(i=0; i<4; i++){
1574 int index= s->block_index[i];
1575 int min, max;
1576
1577 c->co_located_mv[i][0] = s->next_picture.f.motion_val[0][index][0];
1578 c->co_located_mv[i][1] = s->next_picture.f.motion_val[0][index][1];
1579 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1580 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1581 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1582 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1583
1584 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1585 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1586 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1587 min+= 16*mb_x - 1;
1588 xmax= FFMIN(xmax, s->width - max);
1589 xmin= FFMAX(xmin, - 16 - min);
1590
1591 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1592 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1593 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1594 min+= 16*mb_y - 1;
1595 ymax= FFMIN(ymax, s->height - max);
1596 ymin= FFMAX(ymin, - 16 - min);
1597
1598 if(s->mv_type == MV_TYPE_16X16) break;
1599 }
1600
1601 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1602
1603 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1604 s->b_direct_mv_table[mot_xy][0]= 0;
1605 s->b_direct_mv_table[mot_xy][1]= 0;
1606
1607 return 256*256*256*64;
1608 }
1609
1610 c->xmin= xmin;
1611 c->ymin= ymin;
1612 c->xmax= xmax;
1613 c->ymax= ymax;
1614 c->flags |= FLAG_DIRECT;
1615 c->sub_flags |= FLAG_DIRECT;
1616 c->pred_x=0;
1617 c->pred_y=0;
1618
1619 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1620 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1621
1622 /* special case for first line */
1623 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1624 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1625 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1626 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1627 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1628
1629 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1630 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1631 }
1632
1633 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1634 if(c->sub_flags&FLAG_QPEL)
1635 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1636 else
1637 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1638
1639 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1640 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1641
1642 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1643
1644 mv_table[mot_xy][0]= mx;
1645 mv_table[mot_xy][1]= my;
1646 c->flags &= ~FLAG_DIRECT;
1647 c->sub_flags &= ~FLAG_DIRECT;
1648
1649 return dmin;
1650 }
1651
1652 void ff_estimate_b_frame_motion(MpegEncContext * s,
1653 int mb_x, int mb_y)
1654 {
1655 MotionEstContext * const c= &s->me;
1656 const int penalty_factor= c->mb_penalty_factor;
1657 int fmin, bmin, dmin, fbmin, bimin, fimin;
1658 int type=0;
1659 const int xy = mb_y*s->mb_stride + mb_x;
1660 init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1661 s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1662
1663 get_limits(s, 16*mb_x, 16*mb_y);
1664
1665 c->skip=0;
1666
1667 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.f.mbskip_table[xy]) {
1668 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1669
1670 score= ((unsigned)(score*score + 128*256))>>16;
1671 c->mc_mb_var_sum_temp += score;
1672 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1673 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1674
1675 return;
1676 }
1677
1678 if(c->avctx->me_threshold){
1679 int vard= check_input_motion(s, mb_x, mb_y, 0);
1680
1681 if((vard+128)>>8 < c->avctx->me_threshold){
1682 // pix = c->src[0][0];
1683 // sum = s->dsp.pix_sum(pix, s->linesize);
1684 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1685
1686 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1687 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1688 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1689 c->mb_var_sum_temp += (varc+128)>>8;*/
1690 c->mc_mb_var_sum_temp += (vard+128)>>8;
1691 /* if (vard <= 64<<8 || vard < varc) {
1692 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1693 }else{
1694 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1695 }*/
1696 return;
1697 }
1698 if((vard+128)>>8 < c->avctx->mb_threshold){
1699 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1700 if(type == CANDIDATE_MB_TYPE_DIRECT){
1701 direct_search(s, mb_x, mb_y);
1702 }
1703 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1704 c->skip=0;
1705 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1706 }
1707 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1708 c->skip=0;
1709 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1710 }
1711 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1712 c->skip=0;
1713 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1714 interlaced_search(s, 0,
1715 s->b_field_mv_table[0], s->b_field_select_table[0],
1716 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1717 }
1718 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1719 c->skip=0;
1720 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1721 interlaced_search(s, 2,
1722 s->b_field_mv_table[1], s->b_field_select_table[1],
1723 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1724 }
1725 return;
1726 }
1727 }
1728
1729 if (s->codec_id == AV_CODEC_ID_MPEG4)
1730 dmin= direct_search(s, mb_x, mb_y);
1731 else
1732 dmin= INT_MAX;
1733 //FIXME penalty stuff for non mpeg4
1734 c->skip=0;
1735 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1736
1737 c->skip=0;
1738 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1739 av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1740
1741 c->skip=0;
1742 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1743 av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1744
1745 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1746 //FIXME mb type penalty
1747 c->skip=0;
1748 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1749 fimin= interlaced_search(s, 0,
1750 s->b_field_mv_table[0], s->b_field_select_table[0],
1751 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1752 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1753 bimin= interlaced_search(s, 2,
1754 s->b_field_mv_table[1], s->b_field_select_table[1],
1755 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1756 }else
1757 fimin= bimin= INT_MAX;
1758
1759 {
1760 int score= fmin;
1761 type = CANDIDATE_MB_TYPE_FORWARD;
1762
1763 if (dmin <= score){
1764 score = dmin;
1765 type = CANDIDATE_MB_TYPE_DIRECT;
1766 }
1767 if(bmin<score){
1768 score=bmin;
1769 type= CANDIDATE_MB_TYPE_BACKWARD;
1770 }
1771 if(fbmin<score){
1772 score=fbmin;
1773 type= CANDIDATE_MB_TYPE_BIDIR;
1774 }
1775 if(fimin<score){
1776 score=fimin;
1777 type= CANDIDATE_MB_TYPE_FORWARD_I;
1778 }
1779 if(bimin<score){
1780 score=bimin;
1781 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1782 }
1783
1784 score= ((unsigned)(score*score + 128*256))>>16;
1785 c->mc_mb_var_sum_temp += score;
1786 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1787 }
1788
1789 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1790 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1791 if(fimin < INT_MAX)
1792 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1793 if(bimin < INT_MAX)
1794 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1795 if(fimin < INT_MAX && bimin < INT_MAX){
1796 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1797 }
1798 //FIXME something smarter
1799 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1800 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])
1801 type |= CANDIDATE_MB_TYPE_DIRECT0;
1802 }
1803
1804 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1805 }
1806
1807 /* find best f_code for ME which do unlimited searches */
1808 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1809 {
1810 if(s->me_method>=ME_EPZS){
1811 int score[8];
1812 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1813 uint8_t * fcode_tab= s->fcode_tab;
1814 int best_fcode=-1;
1815 int best_score=-10000000;
1816
1817 if(s->msmpeg4_version)
1818 range= FFMIN(range, 16);
1819 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1820 range= FFMIN(range, 256);
1821
1822 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1823
1824 for(y=0; y<s->mb_height; y++){
1825 int x;
1826 int xy= y*s->mb_stride;
1827 for(x=0; x<s->mb_width; x++){
1828 if(s->mb_type[xy] & type){
1829 int mx= mv_table[xy][0];
1830 int my= mv_table[xy][1];
1831 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1832 fcode_tab[my + MAX_MV]);
1833 int j;
1834
1835 if(mx >= range || mx < -range ||
1836 my >= range || my < -range)
1837 continue;
1838
1839 for(j=0; j<fcode && j<8; j++){
1840 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1841 score[j]-= 170;
1842 }
1843 }
1844 xy++;
1845 }
1846 }
1847
1848 for(i=1; i<8; i++){
1849 if(score[i] > best_score){
1850 best_score= score[i];
1851 best_fcode= i;
1852 }
1853 }
1854
1855 return best_fcode;
1856 }else{
1857 return 1;
1858 }
1859 }
1860
1861 void ff_fix_long_p_mvs(MpegEncContext * s)
1862 {
1863 MotionEstContext * const c= &s->me;
1864 const int f_code= s->f_code;
1865 int y, range;
1866 assert(s->pict_type==AV_PICTURE_TYPE_P);
1867
1868 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1869
1870 assert(range <= 16 || !s->msmpeg4_version);
1871 assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1872
1873 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1874
1875 if(s->flags&CODEC_FLAG_4MV){
1876 const int wrap= s->b8_stride;
1877
1878 /* clip / convert to intra 8x8 type MVs */
1879 for(y=0; y<s->mb_height; y++){
1880 int xy= y*2*wrap;
1881 int i= y*s->mb_stride;
1882 int x;
1883
1884 for(x=0; x<s->mb_width; x++){
1885 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1886 int block;
1887 for(block=0; block<4; block++){
1888 int off= (block& 1) + (block>>1)*wrap;
1889 int mx = s->current_picture.f.motion_val[0][ xy + off ][0];
1890 int my = s->current_picture.f.motion_val[0][ xy + off ][1];
1891
1892 if( mx >=range || mx <-range
1893 || my >=range || my <-range){
1894 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1895 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1896 s->current_picture.f.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1897 }
1898 }
1899 }
1900 xy+=2;
1901 i++;
1902 }
1903 }
1904 }
1905 }
1906
1907 /**
1908 *
1909 * @param truncate 1 for truncation, 0 for using intra
1910 */
1911 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1912 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1913 {
1914 MotionEstContext * const c= &s->me;
1915 int y, h_range, v_range;
1916
1917 // RAL: 8 in MPEG-1, 16 in MPEG-4
1918 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1919
1920 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1921
1922 h_range= range;
1923 v_range= field_select_table ? range>>1 : range;
1924
1925 /* clip / convert to intra 16x16 type MVs */
1926 for(y=0; y<s->mb_height; y++){
1927 int x;
1928 int xy= y*s->mb_stride;
1929 for(x=0; x<s->mb_width; x++){
1930 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1931 if(field_select_table==NULL || field_select_table[xy] == field_select){
1932 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1933 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1934
1935 if(truncate){
1936 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1937 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1938 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1939 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1940 }else{
1941 s->mb_type[xy] &= ~type;
1942 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1943 mv_table[xy][0]=
1944 mv_table[xy][1]= 0;
1945 }
1946 }
1947 }
1948 }
1949 xy++;
1950 }
1951 }
1952 }
A couple of patches to make building with MPlayer easier