3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
8 * This file is part of FFmpeg.
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.
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.
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
37 #include "motion_est.h"
38 #include "mpegutils.h"
39 #include "mpegvideo.h"
43 #define P_TOPRIGHT P[3]
47 #define ME_MAP_SHIFT 3
48 #define ME_MAP_MV_BITS 11
50 static int sad_hpel_motion_search(MpegEncContext * s,
51 int *mx_ptr, int *my_ptr, int dmin,
52 int src_index, int ref_index,
55 static inline unsigned update_map_generation(MotionEstContext *c)
57 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
58 if(c->map_generation==0){
59 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
60 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
62 return c->map_generation;
65 /* shape adaptive search stuff */
66 typedef struct Minima{
72 static int minima_cmp(const void *a, const void *b){
73 const Minima *da = (const Minima *) a;
74 const Minima *db = (const Minima *) b;
76 return da->height - db->height;
79 #define FLAG_QPEL 1 //must be 1
83 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){
84 const int offset[3]= {
86 ((y*c->uvstride + x)>>1),
87 ((y*c->uvstride + x)>>1),
91 c->src[0][i]= src [i] + offset[i];
92 c->ref[0][i]= ref [i] + offset[i];
96 c->ref[ref_index][i]= ref2[i] + offset[i];
101 static int get_flags(MotionEstContext *c, int direct, int chroma){
102 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
103 + (direct ? FLAG_DIRECT : 0)
104 + (chroma ? FLAG_CHROMA : 0);
107 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
108 const int size, const int h, int ref_index, int src_index,
109 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
110 MotionEstContext * const c= &s->me;
111 const int stride= c->stride;
112 const int hx= subx + (x<<(1+qpel));
113 const int hy= suby + (y<<(1+qpel));
114 uint8_t * const * const ref= c->ref[ref_index];
115 uint8_t * const * const src= c->src[src_index];
117 //FIXME check chroma 4mv, (no crashes ...)
118 av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
119 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
120 const int time_pp= s->pp_time;
121 const int time_pb= s->pb_time;
122 const int mask= 2*qpel+1;
123 if(s->mv_type==MV_TYPE_8X8){
126 int fx = c->direct_basis_mv[i][0] + hx;
127 int fy = c->direct_basis_mv[i][1] + hy;
128 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));
129 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));
130 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
131 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
133 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
135 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
136 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
138 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
139 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
143 int fx = c->direct_basis_mv[0][0] + hx;
144 int fy = c->direct_basis_mv[0][1] + hy;
145 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
146 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
147 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
148 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
151 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
152 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
153 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
154 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
155 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
156 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
157 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
158 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
160 av_assert2((fx>>1) + 16*s->mb_x >= -16);
161 av_assert2((fy>>1) + 16*s->mb_y >= -16);
162 av_assert2((fx>>1) + 16*s->mb_x <= s->width);
163 av_assert2((fy>>1) + 16*s->mb_y <= s->height);
164 av_assert2((bx>>1) + 16*s->mb_x >= -16);
165 av_assert2((by>>1) + 16*s->mb_y >= -16);
166 av_assert2((bx>>1) + 16*s->mb_x <= s->width);
167 av_assert2((by>>1) + 16*s->mb_y <= s->height);
169 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
170 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
173 d = cmp_func(s, c->temp, src[0], stride, 16);
179 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
180 const int size, const int h, int ref_index, int src_index,
181 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
182 MotionEstContext * const c= &s->me;
183 const int stride= c->stride;
184 const int uvstride= c->uvstride;
185 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
186 const int hx= subx + (x<<(1+qpel));
187 const int hy= suby + (y<<(1+qpel));
188 uint8_t * const * const ref= c->ref[ref_index];
189 uint8_t * const * const src= c->src[src_index];
191 //FIXME check chroma 4mv, (no crashes ...)
192 int uvdxy; /* no, it might not be used uninitialized */
195 if (h << size == 16) {
196 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
197 } else if (size == 0 && h == 8) {
198 c->qpel_put[1][dxy](c->temp , ref[0] + x + y*stride , stride);
199 c->qpel_put[1][dxy](c->temp + 8, ref[0] + x + y*stride + 8, stride);
207 uvdxy= (cx&1) + 2*(cy&1);
208 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
211 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
213 uvdxy= dxy | (x&1) | (2*(y&1));
215 d = cmp_func(s, c->temp, src[0], stride, h);
217 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
219 uvdxy= (x&1) + 2*(y&1);
222 uint8_t * const uvtemp= c->temp + 16*stride;
223 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
224 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
225 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
226 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
231 static int cmp_simple(MpegEncContext *s, const int x, const int y,
232 int ref_index, int src_index,
233 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
234 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
237 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
238 const int size, const int h, int ref_index, int src_index,
239 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
240 if(flags&FLAG_DIRECT){
241 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
243 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
247 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
248 const int size, const int h, int ref_index, int src_index,
249 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
250 if(flags&FLAG_DIRECT){
251 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
253 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);
257 /** @brief compares a block (either a full macroblock or a partition thereof)
258 against a proposed motion-compensated prediction of that block
260 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
261 const int size, const int h, int ref_index, int src_index,
262 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
263 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
264 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
265 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
266 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
267 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
268 && subx==0 && suby==0){
269 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
271 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
275 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
276 const int size, const int h, int ref_index, int src_index,
277 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
278 if(flags&FLAG_DIRECT){
279 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
281 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
285 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
286 const int size, const int h, int ref_index, int src_index,
287 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
288 if(flags&FLAG_DIRECT){
289 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
291 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
295 #include "motion_est_template.c"
297 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
298 ptrdiff_t stride, int h)
303 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
306 int ff_init_me(MpegEncContext *s){
307 MotionEstContext * const c= &s->me;
308 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
309 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
311 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
312 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
315 //special case of snow is needed because snow uses its own iterative ME code
316 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
317 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");
323 if(s->codec_id == AV_CODEC_ID_H261)
324 c->avctx->me_sub_cmp = c->avctx->me_cmp;
326 if(cache_size < 2*dia_size && !c->stride){
327 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
330 ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
331 ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
332 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
333 ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
335 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
336 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
337 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
339 /*FIXME s->no_rounding b_type*/
340 if (s->avctx->flags & CODEC_FLAG_QPEL) {
341 c->sub_motion_search= qpel_motion_search;
342 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
344 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
346 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
348 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
349 c->sub_motion_search= hpel_motion_search;
350 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
351 && c->avctx-> me_cmp == FF_CMP_SAD
352 && c->avctx-> mb_cmp == FF_CMP_SAD)
353 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
355 c->sub_motion_search= hpel_motion_search;
357 c->hpel_avg = s->hdsp.avg_pixels_tab;
359 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
361 c->hpel_put = s->hdsp.put_pixels_tab;
364 c->stride = s->linesize;
365 c->uvstride= s->uvlinesize;
367 c->stride = 16*s->mb_width + 32;
368 c->uvstride= 8*s->mb_width + 16;
371 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
372 * not have yet, and even if we had, the motion estimation code
373 * does not expect it. */
374 if (s->codec_id != AV_CODEC_ID_SNOW) {
375 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
376 s->mecc.me_cmp[2] = zero_cmp;
377 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
378 s->mecc.me_sub_cmp[2] = zero_cmp;
379 c->hpel_put[2][0]= c->hpel_put[2][1]=
380 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
383 if(s->codec_id == AV_CODEC_ID_H261){
384 c->sub_motion_search= no_sub_motion_search;
390 #define CHECK_SAD_HALF_MV(suffix, x, y) \
392 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
393 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
394 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
397 static int sad_hpel_motion_search(MpegEncContext * s,
398 int *mx_ptr, int *my_ptr, int dmin,
399 int src_index, int ref_index,
402 MotionEstContext * const c= &s->me;
403 const int penalty_factor= c->sub_penalty_factor;
406 int stride= c->stride;
409 av_assert2(c->sub_flags == 0);
417 pix = c->src[src_index][0];
421 ptr = c->ref[ref_index][0] + (my * stride) + mx;
425 if (mx > xmin && mx < xmax &&
426 my > ymin && my < ymax) {
429 const int index= (my<<ME_MAP_SHIFT) + mx;
430 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
431 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
432 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
433 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
443 CHECK_SAD_HALF_MV(y2 , 0, -1)
445 CHECK_SAD_HALF_MV(xy2, -1, -1)
447 CHECK_SAD_HALF_MV(xy2, +1, -1)
451 CHECK_SAD_HALF_MV(xy2, -1, +1)
453 CHECK_SAD_HALF_MV(x2 , -1, 0)
455 CHECK_SAD_HALF_MV(xy2, +1, -1)
457 CHECK_SAD_HALF_MV(xy2, -1, -1)
461 CHECK_SAD_HALF_MV(xy2, +1, +1)
463 CHECK_SAD_HALF_MV(x2 , +1, 0)
468 CHECK_SAD_HALF_MV(xy2, -1, -1)
472 CHECK_SAD_HALF_MV(xy2, +1, +1)
474 CHECK_SAD_HALF_MV(x2 , -1, 0)
475 CHECK_SAD_HALF_MV(xy2, -1, +1)
478 CHECK_SAD_HALF_MV(xy2, +1, -1)
482 CHECK_SAD_HALF_MV(xy2, -1, +1)
484 CHECK_SAD_HALF_MV(x2 , +1, 0)
485 CHECK_SAD_HALF_MV(xy2, +1, +1)
487 CHECK_SAD_HALF_MV(y2 , 0, +1)
502 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
504 const int xy= s->mb_x + s->mb_y*s->mb_stride;
506 s->p_mv_table[xy][0] = mx;
507 s->p_mv_table[xy][1] = my;
509 /* has already been set to the 4 MV if 4MV is done */
511 int mot_xy= s->block_index[0];
513 s->current_picture.motion_val[0][mot_xy ][0] = mx;
514 s->current_picture.motion_val[0][mot_xy ][1] = my;
515 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
516 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
518 mot_xy += s->b8_stride;
519 s->current_picture.motion_val[0][mot_xy ][0] = mx;
520 s->current_picture.motion_val[0][mot_xy ][1] = my;
521 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
522 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
527 * get fullpel ME search limits.
529 static inline void get_limits(MpegEncContext *s, int x, int y)
531 MotionEstContext * const c= &s->me;
532 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
533 int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
535 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
538 if (s->unrestricted_mv) {
541 c->xmax = - x + s->width;
542 c->ymax = - y + s->height;
543 } else if (s->out_format == FMT_H261){
544 // Search range of H261 is different from other codec standards
545 c->xmin = (x > 15) ? - 15 : 0;
546 c->ymin = (y > 15) ? - 15 : 0;
547 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
548 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
552 c->xmax = - x + s->mb_width *16 - 16;
553 c->ymax = - y + s->mb_height*16 - 16;
555 if(!range || range > max_range)
558 c->xmin = FFMAX(c->xmin,-range);
559 c->xmax = FFMIN(c->xmax, range);
560 c->ymin = FFMAX(c->ymin,-range);
561 c->ymax = FFMIN(c->ymax, range);
565 static inline void init_mv4_ref(MotionEstContext *c){
566 const int stride= c->stride;
568 c->ref[1][0] = c->ref[0][0] + 8;
569 c->ref[2][0] = c->ref[0][0] + 8*stride;
570 c->ref[3][0] = c->ref[2][0] + 8;
571 c->src[1][0] = c->src[0][0] + 8;
572 c->src[2][0] = c->src[0][0] + 8*stride;
573 c->src[3][0] = c->src[2][0] + 8;
576 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
578 MotionEstContext * const c= &s->me;
583 int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
585 const int stride= c->stride;
586 uint8_t *mv_penalty= c->current_mv_penalty;
587 int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
591 for(block=0; block<4; block++){
593 int pred_x4, pred_y4;
595 static const int off[4]= {2, 1, 1, -1};
596 const int mot_stride = s->b8_stride;
597 const int mot_xy = s->block_index[block];
600 c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
601 c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
604 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
605 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
607 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
609 /* special case for first line */
610 if (s->first_slice_line && block<2) {
611 c->pred_x= pred_x4= P_LEFT[0];
612 c->pred_y= pred_y4= P_LEFT[1];
614 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
615 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
616 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
617 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
618 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
619 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
620 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
621 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
623 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
624 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
626 c->pred_x= pred_x4 = P_MEDIAN[0];
627 c->pred_y= pred_y4 = P_MEDIAN[1];
633 if (s->first_slice_line && block<2 && i>1 && i<9)
637 if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
638 if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
641 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
643 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
645 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
647 const int offset= ((block&1) + (block>>1)*stride)*8;
648 uint8_t *dest_y = c->scratchpad + offset;
649 if(s->quarter_sample){
650 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
651 dxy = ((my4 & 3) << 2) | (mx4 & 3);
654 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
656 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
658 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
659 dxy = ((my4 & 1) << 1) | (mx4 & 1);
662 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
664 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
666 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
670 if(s->quarter_sample){
678 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
679 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
681 if(mx4 != mx || my4 != my) same=0;
687 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
688 dmin_sum += s->mecc.mb_cmp[0](s,
689 s->new_picture.f->data[0] +
690 s->mb_x * 16 + s->mb_y * 16 * stride,
691 c->scratchpad, stride, 16);
694 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
699 mx= ff_h263_round_chroma(mx4_sum);
700 my= ff_h263_round_chroma(my4_sum);
701 dxy = ((my & 1) << 1) | (mx & 1);
703 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
706 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
707 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
709 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
710 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
713 dmin_sum += s->mecc.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);
714 dmin_sum += s->mecc.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);
720 switch(c->avctx->mb_cmp&0xFF){
722 return dmin_sum+ 32*s->qscale*s->qscale;*/
726 return dmin_sum+ 11*c->mb_penalty_factor;
730 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
731 MotionEstContext * const c= &s->me;
733 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
734 c->src[1][0] = c->src[0][0] + s->linesize;
735 if(c->flags & FLAG_CHROMA){
736 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
737 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
738 c->src[1][1] = c->src[0][1] + s->uvlinesize;
739 c->src[1][2] = c->src[0][2] + s->uvlinesize;
743 static int interlaced_search(MpegEncContext *s, int ref_index,
744 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
746 MotionEstContext * const c= &s->me;
751 uint8_t * const mv_penalty= c->current_mv_penalty;
753 const int stride= 2*s->linesize;
755 const int mot_stride= s->mb_stride;
756 const int xy= s->mb_x + s->mb_y*mot_stride;
762 init_interlaced_ref(s, ref_index);
764 for(block=0; block<2; block++){
766 int best_dmin= INT_MAX;
769 for(field_select=0; field_select<2; field_select++){
770 int dmin, mx_i, my_i;
771 int16_t (*mv_table)[2]= mv_tables[block][field_select];
773 if(user_field_select){
774 av_assert1(field_select==0 || field_select==1);
775 av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
776 if(field_select_tables[block][xy] != field_select)
780 P_LEFT[0] = mv_table[xy - 1][0];
781 P_LEFT[1] = mv_table[xy - 1][1];
782 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
784 c->pred_x= P_LEFT[0];
785 c->pred_y= P_LEFT[1];
787 if(!s->first_slice_line){
788 P_TOP[0] = mv_table[xy - mot_stride][0];
789 P_TOP[1] = mv_table[xy - mot_stride][1];
790 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
791 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
792 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
793 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
794 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
795 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
797 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
798 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
800 P_MV1[0]= mx; //FIXME not correct if block != field_select
803 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
805 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
807 mv_table[xy][0]= mx_i;
808 mv_table[xy][1]= my_i;
810 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
814 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
815 dxy = ((my_i & 1) << 1) | (mx_i & 1);
818 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
820 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
822 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
823 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
825 dmin+= c->mb_penalty_factor; //field_select bits
827 dmin += field_select != block; //slightly prefer same field
829 if(dmin < best_dmin){
831 best_field= field_select;
835 int16_t (*mv_table)[2]= mv_tables[block][best_field];
837 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
838 if(mv_table[xy][1]&1) same=0;
839 if(mv_table[xy][1]*2 != my) same=0;
840 if(best_field != block) same=0;
843 field_select_tables[block][xy]= best_field;
844 dmin_sum += best_dmin;
855 switch(c->avctx->mb_cmp&0xFF){
857 return dmin_sum+ 32*s->qscale*s->qscale;*/
861 return dmin_sum+ 11*c->mb_penalty_factor;
865 static inline int get_penalty_factor(int lambda, int lambda2, int type){
869 return lambda>>FF_LAMBDA_SHIFT;
871 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
873 return (4*lambda)>>(FF_LAMBDA_SHIFT);
875 return (2*lambda)>>(FF_LAMBDA_SHIFT);
878 return (2*lambda)>>FF_LAMBDA_SHIFT;
883 return lambda2>>FF_LAMBDA_SHIFT;
889 void ff_estimate_p_frame_motion(MpegEncContext * s,
892 MotionEstContext * const c= &s->me;
894 int sum, mx, my, dmin;
895 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
896 int vard; ///< sum of squared differences with the estimated motion vector
898 const int shift= 1+s->quarter_sample;
900 Picture * const pic= &s->current_picture;
902 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
904 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
905 av_assert0(s->linesize == c->stride);
906 av_assert0(s->uvlinesize == c->uvstride);
908 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
909 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
910 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
911 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
913 get_limits(s, 16*mb_x, 16*mb_y);
916 /* intra / predictive decision */
918 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
919 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
920 (((unsigned) sum * sum) >> 8) + 500;
922 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
923 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
924 c->mb_var_sum_temp += (varc+128)>>8;
926 switch(s->me_method) {
936 const int mot_stride = s->b8_stride;
937 const int mot_xy = s->block_index[0];
939 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
940 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
942 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
944 if(!s->first_slice_line) {
945 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
946 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
947 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
948 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
949 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
950 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
951 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
953 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
954 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
956 if(s->out_format == FMT_H263){
957 c->pred_x = P_MEDIAN[0];
958 c->pred_y = P_MEDIAN[1];
959 }else { /* mpeg1 at least */
960 c->pred_x= P_LEFT[0];
961 c->pred_y= P_LEFT[1];
964 c->pred_x= P_LEFT[0];
965 c->pred_y= P_LEFT[1];
969 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
974 /* At this point (mx,my) are full-pell and the relative displacement */
975 ppix = c->ref[0][0] + (my * s->linesize) + mx;
977 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
979 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
980 c->mc_mb_var_sum_temp += (vard+128)>>8;
982 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
983 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
984 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
985 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
987 if (vard*2 + 200*256 > varc)
988 mb_type|= CANDIDATE_MB_TYPE_INTRA;
989 if (varc*2 + 200*256 > vard || s->qscale > 24){
990 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
991 mb_type|= CANDIDATE_MB_TYPE_INTER;
992 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
993 if (s->mpv_flags & FF_MPV_FLAG_MV0)
995 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1000 if ((s->avctx->flags & CODEC_FLAG_4MV)
1001 && !c->skip && varc>50<<8 && vard>10<<8){
1002 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1003 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1005 set_p_mv_tables(s, mx, my, 0);
1007 set_p_mv_tables(s, mx, my, 1);
1008 if ((s->avctx->flags & CODEC_FLAG_INTERLACED_ME)
1009 && !c->skip){ //FIXME varc/d checks
1010 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1011 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1015 mb_type= CANDIDATE_MB_TYPE_INTER;
1017 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1018 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1019 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1021 if ((s->avctx->flags & CODEC_FLAG_4MV)
1022 && !c->skip && varc>50<<8 && vard>10<<8){
1023 int dmin4= h263_mv4_search(s, mx, my, shift);
1025 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1029 if ((s->avctx->flags & CODEC_FLAG_INTERLACED_ME)
1030 && !c->skip){ //FIXME varc/d checks
1031 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1033 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1038 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1040 /* get intra luma score */
1041 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1042 intra_score= varc - 500;
1044 unsigned mean = (sum+128)>>8;
1047 for(i=0; i<16; i++){
1048 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1049 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1050 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1051 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1054 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1056 intra_score += c->mb_penalty_factor*16;
1058 if(intra_score < dmin){
1059 mb_type= CANDIDATE_MB_TYPE_INTRA;
1060 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1062 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1065 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1066 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1067 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1071 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1074 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1077 MotionEstContext * const c= &s->me;
1080 const int shift= 1+s->quarter_sample;
1081 const int xy= mb_x + mb_y*s->mb_stride;
1082 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1084 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1086 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1087 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1089 get_limits(s, 16*mb_x, 16*mb_y);
1092 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1093 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1095 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1097 /* special case for first line */
1098 if (s->first_slice_line) {
1099 c->pred_x= P_LEFT[0];
1100 c->pred_y= P_LEFT[1];
1101 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1102 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1104 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1105 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1106 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1107 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1108 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1109 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1110 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1112 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1113 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1115 c->pred_x = P_MEDIAN[0];
1116 c->pred_y = P_MEDIAN[1];
1119 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1121 s->p_mv_table[xy][0] = mx<<shift;
1122 s->p_mv_table[xy][1] = my<<shift;
1127 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1128 int16_t (*mv_table)[2], int ref_index, int f_code)
1130 MotionEstContext * const c= &s->me;
1133 const int shift= 1+s->quarter_sample;
1134 const int mot_stride = s->mb_stride;
1135 const int mot_xy = mb_y*mot_stride + mb_x;
1136 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1139 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1140 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1141 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1142 c->current_mv_penalty= mv_penalty;
1144 get_limits(s, 16*mb_x, 16*mb_y);
1146 switch(s->me_method) {
1155 P_LEFT[0] = mv_table[mot_xy - 1][0];
1156 P_LEFT[1] = mv_table[mot_xy - 1][1];
1158 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1160 /* special case for first line */
1161 if (!s->first_slice_line) {
1162 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1163 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1164 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1165 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1166 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1167 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1168 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1170 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1171 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1173 c->pred_x = P_LEFT[0];
1174 c->pred_y = P_LEFT[1];
1176 if(mv_table == s->b_forw_mv_table){
1177 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1179 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1182 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1187 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1189 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1190 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1192 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1193 mv_table[mot_xy][0]= mx;
1194 mv_table[mot_xy][1]= my;
1199 static inline int check_bidir_mv(MpegEncContext * s,
1200 int motion_fx, int motion_fy,
1201 int motion_bx, int motion_by,
1202 int pred_fx, int pred_fy,
1203 int pred_bx, int pred_by,
1207 //FIXME better f_code prediction (max mv & distance)
1209 MotionEstContext * const c= &s->me;
1210 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1211 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1212 int stride= c->stride;
1213 uint8_t *dest_y = c->scratchpad;
1218 uint8_t **src_data= c->src[0];
1219 uint8_t **ref_data= c->ref[0];
1220 uint8_t **ref2_data= c->ref[2];
1222 if(s->quarter_sample){
1223 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1224 src_x = motion_fx >> 2;
1225 src_y = motion_fy >> 2;
1227 ptr = ref_data[0] + (src_y * stride) + src_x;
1228 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1230 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1231 src_x = motion_bx >> 2;
1232 src_y = motion_by >> 2;
1234 ptr = ref2_data[0] + (src_y * stride) + src_x;
1235 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1237 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1238 src_x = motion_fx >> 1;
1239 src_y = motion_fy >> 1;
1241 ptr = ref_data[0] + (src_y * stride) + src_x;
1242 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1244 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1245 src_x = motion_bx >> 1;
1246 src_y = motion_by >> 1;
1248 ptr = ref2_data[0] + (src_y * stride) + src_x;
1249 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1252 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1253 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1254 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1256 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1263 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1264 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1266 MotionEstContext * const c= &s->me;
1267 const int mot_stride = s->mb_stride;
1268 const int xy = mb_y *mot_stride + mb_x;
1270 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1271 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1272 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1273 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1274 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1275 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1276 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1277 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1278 const int flags= c->sub_flags;
1279 const int qpel= flags&FLAG_QPEL;
1280 const int shift= 1+qpel;
1281 const int xmin= c->xmin<<shift;
1282 const int ymin= c->ymin<<shift;
1283 const int xmax= c->xmax<<shift;
1284 const int ymax= c->ymax<<shift;
1285 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1286 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1287 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1288 uint8_t map[256] = { 0 };
1290 map[hashidx&255] = 1;
1292 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1293 motion_bx, motion_by,
1298 if(s->avctx->bidir_refine){
1300 static const uint8_t limittab[5]={0,8,32,64,80};
1301 const int limit= limittab[s->avctx->bidir_refine];
1302 static const int8_t vect[][4]={
1303 { 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},
1305 { 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},
1306 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1307 { 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},
1308 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1310 { 0, 1, 1, 1}, { 0,-1,-1,-1}, { 1, 1, 1, 0}, {-1,-1,-1, 0}, { 1, 1, 0, 1}, {-1,-1, 0,-1}, { 1, 0, 1, 1}, {-1, 0,-1,-1},
1311 { 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1},
1312 { 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1},
1313 { 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},
1315 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1316 { 1, 1, 1,-1}, {-1,-1,-1, 1}, { 1, 1,-1, 1}, {-1,-1, 1,-1}, { 1,-1, 1, 1}, {-1, 1,-1,-1}, {-1, 1, 1, 1}, { 1,-1,-1,-1},
1317 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1319 static const uint8_t hash[]={
1320 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),
1322 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),
1323 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1324 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),
1325 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1327 HASH8( 0, 1, 1, 1), HASH8( 0,-1,-1,-1), HASH8( 1, 1, 1, 0), HASH8(-1,-1,-1, 0), HASH8( 1, 1, 0, 1), HASH8(-1,-1, 0,-1), HASH8( 1, 0, 1, 1), HASH8(-1, 0,-1,-1),
1328 HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1),
1329 HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1),
1330 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),
1332 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1333 HASH8( 1, 1, 1,-1), HASH8(-1,-1,-1, 1), HASH8( 1, 1,-1, 1), HASH8(-1,-1, 1,-1), HASH8( 1,-1, 1, 1), HASH8(-1, 1,-1,-1), HASH8(-1, 1, 1, 1), HASH8( 1,-1,-1,-1),
1334 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),
1337 #define CHECK_BIDIR(fx,fy,bx,by)\
1338 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1339 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1340 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1342 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1343 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);\
1345 hashidx += HASH(fx,fy,bx,by);\
1354 #define CHECK_BIDIR2(a,b,c,d)\
1355 CHECK_BIDIR(a,b,c,d)\
1356 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1363 CHECK_BIDIR2(0,0,0,1)
1364 CHECK_BIDIR2(0,0,1,0)
1365 CHECK_BIDIR2(0,1,0,0)
1366 CHECK_BIDIR2(1,0,0,0)
1368 for(i=8; i<limit; i++){
1369 int fx= motion_fx+vect[i][0];
1370 int fy= motion_fy+vect[i][1];
1371 int bx= motion_bx+vect[i][2];
1372 int by= motion_by+vect[i][3];
1374 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1375 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1377 map[(hashidx+hash[i])&255] = 1;
1379 if(!map[(hashidx+hash[i])&255]){
1381 map[(hashidx+hash[i])&255] = 1;
1382 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1393 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1394 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1395 borderdist= FFMIN(a,b);
1403 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1404 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1405 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1406 s->b_bidir_back_mv_table[xy][1]= motion_by;
1411 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1413 MotionEstContext * const c= &s->me;
1415 const int mot_stride = s->mb_stride;
1416 const int mot_xy = mb_y*mot_stride + mb_x;
1417 const int shift= 1+s->quarter_sample;
1419 const int time_pp= s->pp_time;
1420 const int time_pb= s->pb_time;
1421 int mx, my, xmin, xmax, ymin, ymax;
1422 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1424 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1425 ymin= xmin=(-32)>>shift;
1426 ymax= xmax= 31>>shift;
1428 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1429 s->mv_type= MV_TYPE_8X8;
1431 s->mv_type= MV_TYPE_16X16;
1435 int index= s->block_index[i];
1438 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1439 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1440 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1441 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1442 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1443 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1445 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1446 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1447 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1449 xmax= FFMIN(xmax, s->width - max);
1450 xmin= FFMAX(xmin, - 16 - min);
1452 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1453 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1454 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1456 ymax= FFMIN(ymax, s->height - max);
1457 ymin= FFMAX(ymin, - 16 - min);
1459 if(s->mv_type == MV_TYPE_16X16) break;
1462 av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1464 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1465 s->b_direct_mv_table[mot_xy][0]= 0;
1466 s->b_direct_mv_table[mot_xy][1]= 0;
1468 return 256*256*256*64;
1475 c->flags |= FLAG_DIRECT;
1476 c->sub_flags |= FLAG_DIRECT;
1480 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1481 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1483 /* special case for first line */
1484 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1485 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1486 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1487 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1488 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1490 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1491 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1494 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1495 if(c->sub_flags&FLAG_QPEL)
1496 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1498 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1500 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1501 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1503 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1505 mv_table[mot_xy][0]= mx;
1506 mv_table[mot_xy][1]= my;
1507 c->flags &= ~FLAG_DIRECT;
1508 c->sub_flags &= ~FLAG_DIRECT;
1513 void ff_estimate_b_frame_motion(MpegEncContext * s,
1516 MotionEstContext * const c= &s->me;
1517 const int penalty_factor= c->mb_penalty_factor;
1518 int fmin, bmin, dmin, fbmin, bimin, fimin;
1520 const int xy = mb_y*s->mb_stride + mb_x;
1521 init_ref(c, s->new_picture.f->data, s->last_picture.f->data,
1522 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1524 get_limits(s, 16*mb_x, 16*mb_y);
1528 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1529 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1531 score= ((unsigned)(score*score + 128*256))>>16;
1532 c->mc_mb_var_sum_temp += score;
1533 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1534 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1539 if (s->codec_id == AV_CODEC_ID_MPEG4)
1540 dmin= direct_search(s, mb_x, mb_y);
1543 //FIXME penalty stuff for non mpeg4
1545 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1549 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1551 ff_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1554 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1555 ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1557 if (s->avctx->flags & CODEC_FLAG_INTERLACED_ME) {
1558 //FIXME mb type penalty
1560 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1561 fimin= interlaced_search(s, 0,
1562 s->b_field_mv_table[0], s->b_field_select_table[0],
1563 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1564 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1565 bimin= interlaced_search(s, 2,
1566 s->b_field_mv_table[1], s->b_field_select_table[1],
1567 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1569 fimin= bimin= INT_MAX;
1573 type = CANDIDATE_MB_TYPE_FORWARD;
1577 type = CANDIDATE_MB_TYPE_DIRECT;
1581 type= CANDIDATE_MB_TYPE_BACKWARD;
1585 type= CANDIDATE_MB_TYPE_BIDIR;
1589 type= CANDIDATE_MB_TYPE_FORWARD_I;
1593 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1596 score= ((unsigned)(score*score + 128*256))>>16;
1597 c->mc_mb_var_sum_temp += score;
1598 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1601 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1602 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1604 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1606 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1607 if(fimin < INT_MAX && bimin < INT_MAX){
1608 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1610 //FIXME something smarter
1611 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1612 if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1613 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1614 type |= CANDIDATE_MB_TYPE_DIRECT0;
1617 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1620 /* find best f_code for ME which do unlimited searches */
1621 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1623 if(s->me_method>=ME_EPZS){
1625 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1626 uint8_t * fcode_tab= s->fcode_tab;
1628 int best_score=-10000000;
1630 if(s->msmpeg4_version)
1631 range= FFMIN(range, 16);
1632 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1633 range= FFMIN(range, 256);
1635 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1637 for(y=0; y<s->mb_height; y++){
1639 int xy= y*s->mb_stride;
1640 for(x=0; x<s->mb_width; x++){
1641 if(s->mb_type[xy] & type){
1642 int mx= mv_table[xy][0];
1643 int my= mv_table[xy][1];
1644 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1645 fcode_tab[my + MAX_MV]);
1648 if(mx >= range || mx < -range ||
1649 my >= range || my < -range)
1652 for(j=0; j<fcode && j<8; j++){
1653 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1662 if(score[i] > best_score){
1663 best_score= score[i];
1674 void ff_fix_long_p_mvs(MpegEncContext * s)
1676 MotionEstContext * const c= &s->me;
1677 const int f_code= s->f_code;
1679 av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1681 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1683 av_assert0(range <= 16 || !s->msmpeg4_version);
1684 av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1686 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1688 if (s->avctx->flags & CODEC_FLAG_4MV) {
1689 const int wrap= s->b8_stride;
1691 /* clip / convert to intra 8x8 type MVs */
1692 for(y=0; y<s->mb_height; y++){
1694 int i= y*s->mb_stride;
1697 for(x=0; x<s->mb_width; x++){
1698 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1700 for(block=0; block<4; block++){
1701 int off= (block& 1) + (block>>1)*wrap;
1702 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1703 int my = s->current_picture.motion_val[0][ xy + off ][1];
1705 if( mx >=range || mx <-range
1706 || my >=range || my <-range){
1707 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1708 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1709 s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1722 * @param truncate 1 for truncation, 0 for using intra
1724 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1725 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1727 MotionEstContext * const c= &s->me;
1728 int y, h_range, v_range;
1730 // RAL: 8 in MPEG-1, 16 in MPEG-4
1731 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1733 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1736 v_range= field_select_table ? range>>1 : range;
1738 /* clip / convert to intra 16x16 type MVs */
1739 for(y=0; y<s->mb_height; y++){
1741 int xy= y*s->mb_stride;
1742 for(x=0; x<s->mb_width; x++){
1743 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1744 if (!field_select_table || field_select_table[xy] == field_select) {
1745 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1746 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1749 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1750 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1751 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1752 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1754 s->mb_type[xy] &= ~type;
1755 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;