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 Libav.
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.
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.
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
36 #include "mpegvideo.h"
43 #define P_TOPRIGHT P[3]
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,
52 static inline unsigned update_map_generation(MotionEstContext *c)
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);
59 return c->map_generation;
62 /* shape adaptive search stuff */
63 typedef struct Minima{
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;
73 return da->height - db->height;
76 #define FLAG_QPEL 1 //must be 1
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]= {
83 ((y*c->uvstride + x)>>1),
84 ((y*c->uvstride + x)>>1),
88 c->src[0][i]= src [i] + offset[i];
89 c->ref[0][i]= ref [i] + offset[i];
93 c->ref[ref_index][i]= ref2[i] + offset[i];
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);
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];
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){
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));
130 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
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);
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);
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));
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);
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);
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);
170 d = cmp_func(s, c->temp, src[0], stride, 16);
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];
188 //FIXME check chroma 4mv, (no crashes ...)
189 int uvdxy; /* no, it might not be used uninitialized */
192 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
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
202 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
204 uvdxy= dxy | (x&1) | (2*(y&1));
206 d = cmp_func(s, c->temp, src[0], stride, h);
208 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
210 uvdxy= (x&1) + 2*(y&1);
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);
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);
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);
234 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
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);
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);
248 /** @brief compares a block (either a full macroblock or a partition thereof)
249 against a proposed motion-compensated prediction of that block
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);
262 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
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);
272 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
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);
282 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
286 #include "motion_est_template.c"
288 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
292 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
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);
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");
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");
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");
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);
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);
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;
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
340 c->sub_motion_search= hpel_motion_search;
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;
347 c->stride = s->linesize;
348 c->uvstride= s->uvlinesize;
350 c->stride = 16*s->mb_width + 32;
351 c->uvstride= 8*s->mb_width + 16;
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;
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;
363 c->hpel_put[2][0]= c->hpel_put[2][1]=
364 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
366 if(s->codec_id == AV_CODEC_ID_H261){
367 c->sub_motion_search= no_sub_motion_search;
373 #define CHECK_SAD_HALF_MV(suffix, x, y) \
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)\
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,
385 MotionEstContext * const c= &s->me;
386 const int penalty_factor= c->sub_penalty_factor;
389 int stride= c->stride;
390 const int flags= c->sub_flags;
401 pix = c->src[src_index][0];
405 ptr = c->ref[ref_index][0] + (my * stride) + mx;
409 if (mx > xmin && mx < xmax &&
410 my > ymin && my < ymax) {
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)];
427 CHECK_SAD_HALF_MV(y2 , 0, -1)
429 CHECK_SAD_HALF_MV(xy2, -1, -1)
431 CHECK_SAD_HALF_MV(xy2, +1, -1)
435 CHECK_SAD_HALF_MV(xy2, -1, +1)
437 CHECK_SAD_HALF_MV(x2 , -1, 0)
439 CHECK_SAD_HALF_MV(xy2, +1, -1)
441 CHECK_SAD_HALF_MV(xy2, -1, -1)
445 CHECK_SAD_HALF_MV(xy2, +1, +1)
447 CHECK_SAD_HALF_MV(x2 , +1, 0)
452 CHECK_SAD_HALF_MV(xy2, -1, -1)
456 CHECK_SAD_HALF_MV(xy2, +1, +1)
458 CHECK_SAD_HALF_MV(x2 , -1, 0)
459 CHECK_SAD_HALF_MV(xy2, -1, +1)
462 CHECK_SAD_HALF_MV(xy2, +1, -1)
466 CHECK_SAD_HALF_MV(xy2, -1, +1)
468 CHECK_SAD_HALF_MV(x2 , +1, 0)
469 CHECK_SAD_HALF_MV(xy2, +1, +1)
471 CHECK_SAD_HALF_MV(y2 , 0, +1)
486 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
488 const int xy= s->mb_x + s->mb_y*s->mb_stride;
490 s->p_mv_table[xy][0] = mx;
491 s->p_mv_table[xy][1] = my;
493 /* has already been set to the 4 MV if 4MV is done */
495 int mot_xy= s->block_index[0];
497 s->current_picture.motion_val[0][mot_xy ][0] = mx;
498 s->current_picture.motion_val[0][mot_xy ][1] = my;
499 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
500 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
502 mot_xy += s->b8_stride;
503 s->current_picture.motion_val[0][mot_xy ][0] = mx;
504 s->current_picture.motion_val[0][mot_xy ][1] = my;
505 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
506 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
511 * get fullpel ME search limits.
513 static inline void get_limits(MpegEncContext *s, int x, int y)
515 MotionEstContext * const c= &s->me;
516 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
518 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
521 if (s->unrestricted_mv) {
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;
535 c->xmax = - x + s->mb_width *16 - 16;
536 c->ymax = - y + s->mb_height*16 - 16;
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);
546 static inline void init_mv4_ref(MotionEstContext *c){
547 const int stride= c->stride;
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;
557 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
559 MotionEstContext * const c= &s->me;
564 int dmin_sum=0, mx4_sum=0, my4_sum=0;
566 const int stride= c->stride;
567 uint8_t *mv_penalty= c->current_mv_penalty;
571 for(block=0; block<4; block++){
573 int pred_x4, pred_y4;
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];
579 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
580 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
582 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
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];
589 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
590 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
591 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
592 P_TOPRIGHT[1] = s->current_picture.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);
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]);
601 c->pred_x= pred_x4 = P_MEDIAN[0];
602 c->pred_y= pred_y4 = P_MEDIAN[1];
607 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
609 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
611 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
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);
620 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
622 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
624 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
625 dxy = ((my4 & 1) << 1) | (mx4 & 1);
628 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
630 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
632 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
636 if(s->quarter_sample){
644 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
645 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
647 if(mx4 != mx || my4 != my) same=0;
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);
657 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
662 mx= ff_h263_round_chroma(mx4_sum);
663 my= ff_h263_round_chroma(my4_sum);
664 dxy = ((my & 1) << 1) | (mx & 1);
666 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
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);
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);
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);
683 switch(c->avctx->mb_cmp&0xFF){
685 return dmin_sum+ 32*s->qscale*s->qscale;*/
689 return dmin_sum+ 11*c->mb_penalty_factor;
693 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
694 MotionEstContext * const c= &s->me;
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;
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)
709 MotionEstContext * const c= &s->me;
714 uint8_t * const mv_penalty= c->current_mv_penalty;
716 const int stride= 2*s->linesize;
718 const int mot_stride= s->mb_stride;
719 const int xy= s->mb_x + s->mb_y*mot_stride;
725 init_interlaced_ref(s, ref_index);
727 for(block=0; block<2; block++){
729 int best_dmin= INT_MAX;
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];
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)
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);
747 c->pred_x= P_LEFT[0];
748 c->pred_y= P_LEFT[1];
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);
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]);
763 P_MV1[0]= mx; //FIXME not correct if block != field_select
766 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
768 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
770 mv_table[xy][0]= mx_i;
771 mv_table[xy][1]= my_i;
773 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
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);
781 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
783 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
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;
788 dmin+= c->mb_penalty_factor; //field_select bits
790 dmin += field_select != block; //slightly prefer same field
792 if(dmin < best_dmin){
794 best_field= field_select;
798 int16_t (*mv_table)[2]= mv_tables[block][best_field];
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;
806 field_select_tables[block][xy]= best_field;
807 dmin_sum += best_dmin;
818 switch(c->avctx->mb_cmp&0xFF){
820 return dmin_sum+ 32*s->qscale*s->qscale;*/
824 return dmin_sum+ 11*c->mb_penalty_factor;
828 static inline int get_penalty_factor(int lambda, int lambda2, int type){
832 return lambda>>FF_LAMBDA_SHIFT;
834 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
837 return (2*lambda)>>FF_LAMBDA_SHIFT;
842 return lambda2>>FF_LAMBDA_SHIFT;
848 void ff_estimate_p_frame_motion(MpegEncContext * s,
851 MotionEstContext * const c= &s->me;
853 int sum, mx, my, dmin;
854 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
855 int vard; ///< sum of squared differences with the estimated motion vector
857 const int shift= 1+s->quarter_sample;
859 Picture * const pic= &s->current_picture;
861 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
863 assert(s->quarter_sample==0 || s->quarter_sample==1);
864 assert(s->linesize == c->stride);
865 assert(s->uvlinesize == c->uvstride);
867 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
868 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
869 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
870 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
872 get_limits(s, 16*mb_x, 16*mb_y);
875 /* intra / predictive decision */
877 sum = s->dsp.pix_sum(pix, s->linesize);
878 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
880 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
881 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
882 c->mb_var_sum_temp += (varc+128)>>8;
884 switch(s->me_method) {
894 const int mot_stride = s->b8_stride;
895 const int mot_xy = s->block_index[0];
897 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
898 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
900 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
902 if(!s->first_slice_line) {
903 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
904 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
905 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
906 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
907 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
908 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
909 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
911 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
912 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
914 if(s->out_format == FMT_H263){
915 c->pred_x = P_MEDIAN[0];
916 c->pred_y = P_MEDIAN[1];
917 }else { /* mpeg1 at least */
918 c->pred_x= P_LEFT[0];
919 c->pred_y= P_LEFT[1];
922 c->pred_x= P_LEFT[0];
923 c->pred_y= P_LEFT[1];
927 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
932 /* At this point (mx,my) are full-pell and the relative displacement */
933 ppix = c->ref[0][0] + (my * s->linesize) + mx;
935 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
937 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
938 c->mc_mb_var_sum_temp += (vard+128)>>8;
941 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
942 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
943 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
945 if(mb_type == CANDIDATE_MB_TYPE_INTER){
946 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
947 set_p_mv_tables(s, mx, my, 1);
952 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
953 h263_mv4_search(s, mx, my, shift);
955 set_p_mv_tables(s, mx, my, 0);
957 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
958 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
960 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
961 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
962 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
963 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
965 if (vard*2 + 200*256 > varc)
966 mb_type|= CANDIDATE_MB_TYPE_INTRA;
967 if (varc*2 + 200*256 > vard || s->qscale > 24){
968 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
969 mb_type|= CANDIDATE_MB_TYPE_INTER;
970 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
971 if(s->flags&CODEC_FLAG_MV0)
973 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
978 if((s->flags&CODEC_FLAG_4MV)
979 && !c->skip && varc>50<<8 && vard>10<<8){
980 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
981 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
983 set_p_mv_tables(s, mx, my, 0);
985 set_p_mv_tables(s, mx, my, 1);
986 if((s->flags&CODEC_FLAG_INTERLACED_ME)
987 && !c->skip){ //FIXME varc/d checks
988 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
989 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
993 mb_type= CANDIDATE_MB_TYPE_INTER;
995 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
996 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
997 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
999 if((s->flags&CODEC_FLAG_4MV)
1000 && !c->skip && varc>50<<8 && vard>10<<8){
1001 int dmin4= h263_mv4_search(s, mx, my, shift);
1003 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1007 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1008 && !c->skip){ //FIXME varc/d checks
1009 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1011 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1016 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1018 /* get intra luma score */
1019 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1020 intra_score= varc - 500;
1022 unsigned mean = (sum+128)>>8;
1025 for(i=0; i<16; i++){
1026 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1027 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1028 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1029 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1032 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1034 intra_score += c->mb_penalty_factor*16;
1036 if(intra_score < dmin){
1037 mb_type= CANDIDATE_MB_TYPE_INTRA;
1038 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1040 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1043 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1044 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1045 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1049 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1052 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1055 MotionEstContext * const c= &s->me;
1058 const int shift= 1+s->quarter_sample;
1059 const int xy= mb_x + mb_y*s->mb_stride;
1060 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1062 assert(s->quarter_sample==0 || s->quarter_sample==1);
1064 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1065 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1067 get_limits(s, 16*mb_x, 16*mb_y);
1070 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1071 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1073 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1075 /* special case for first line */
1076 if (s->first_slice_line) {
1077 c->pred_x= P_LEFT[0];
1078 c->pred_y= P_LEFT[1];
1079 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1080 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1082 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1083 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1084 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1085 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1086 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1087 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1088 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1090 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1091 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1093 c->pred_x = P_MEDIAN[0];
1094 c->pred_y = P_MEDIAN[1];
1097 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1099 s->p_mv_table[xy][0] = mx<<shift;
1100 s->p_mv_table[xy][1] = my<<shift;
1105 static int ff_estimate_motion_b(MpegEncContext * s,
1106 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1108 MotionEstContext * const c= &s->me;
1111 const int shift= 1+s->quarter_sample;
1112 const int mot_stride = s->mb_stride;
1113 const int mot_xy = mb_y*mot_stride + mb_x;
1114 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1117 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1118 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1119 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1120 c->current_mv_penalty= mv_penalty;
1122 get_limits(s, 16*mb_x, 16*mb_y);
1124 switch(s->me_method) {
1133 P_LEFT[0] = mv_table[mot_xy - 1][0];
1134 P_LEFT[1] = mv_table[mot_xy - 1][1];
1136 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1138 /* special case for first line */
1139 if (!s->first_slice_line) {
1140 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1141 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1142 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1143 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1144 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1145 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1146 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1148 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1149 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1151 c->pred_x = P_LEFT[0];
1152 c->pred_y = P_LEFT[1];
1154 if(mv_table == s->b_forw_mv_table){
1155 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1157 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1160 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1165 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1167 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1168 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1170 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1171 mv_table[mot_xy][0]= mx;
1172 mv_table[mot_xy][1]= my;
1177 static inline int check_bidir_mv(MpegEncContext * s,
1178 int motion_fx, int motion_fy,
1179 int motion_bx, int motion_by,
1180 int pred_fx, int pred_fy,
1181 int pred_bx, int pred_by,
1185 //FIXME better f_code prediction (max mv & distance)
1187 MotionEstContext * const c= &s->me;
1188 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1189 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1190 int stride= c->stride;
1191 uint8_t *dest_y = c->scratchpad;
1196 uint8_t **src_data= c->src[0];
1197 uint8_t **ref_data= c->ref[0];
1198 uint8_t **ref2_data= c->ref[2];
1200 if(s->quarter_sample){
1201 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1202 src_x = motion_fx >> 2;
1203 src_y = motion_fy >> 2;
1205 ptr = ref_data[0] + (src_y * stride) + src_x;
1206 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1208 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1209 src_x = motion_bx >> 2;
1210 src_y = motion_by >> 2;
1212 ptr = ref2_data[0] + (src_y * stride) + src_x;
1213 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1215 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1216 src_x = motion_fx >> 1;
1217 src_y = motion_fy >> 1;
1219 ptr = ref_data[0] + (src_y * stride) + src_x;
1220 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1222 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1223 src_x = motion_bx >> 1;
1224 src_y = motion_by >> 1;
1226 ptr = ref2_data[0] + (src_y * stride) + src_x;
1227 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1230 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1231 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1232 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1234 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1241 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1242 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1244 MotionEstContext * const c= &s->me;
1245 const int mot_stride = s->mb_stride;
1246 const int xy = mb_y *mot_stride + mb_x;
1248 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1249 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1250 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1251 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1252 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1253 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1254 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1255 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1256 const int flags= c->sub_flags;
1257 const int qpel= flags&FLAG_QPEL;
1258 const int shift= 1+qpel;
1259 const int xmin= c->xmin<<shift;
1260 const int ymin= c->ymin<<shift;
1261 const int xmax= c->xmax<<shift;
1262 const int ymax= c->ymax<<shift;
1263 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1264 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1265 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1266 uint8_t map[256] = { 0 };
1268 map[hashidx&255] = 1;
1270 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1271 motion_bx, motion_by,
1276 if(s->avctx->bidir_refine){
1278 static const uint8_t limittab[5]={0,8,32,64,80};
1279 const int limit= limittab[s->avctx->bidir_refine];
1280 static const int8_t vect[][4]={
1281 { 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},
1283 { 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},
1284 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1285 { 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},
1286 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1288 { 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},
1289 { 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},
1290 { 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},
1291 { 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},
1293 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1294 { 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},
1295 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1297 static const uint8_t hash[]={
1298 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),
1300 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),
1301 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1302 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),
1303 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1305 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),
1306 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),
1307 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),
1308 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),
1310 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1311 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),
1312 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),
1315 #define CHECK_BIDIR(fx,fy,bx,by)\
1316 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1317 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1318 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1320 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1321 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);\
1323 hashidx += HASH(fx,fy,bx,by);\
1332 #define CHECK_BIDIR2(a,b,c,d)\
1333 CHECK_BIDIR(a,b,c,d)\
1334 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1341 CHECK_BIDIR2(0,0,0,1)
1342 CHECK_BIDIR2(0,0,1,0)
1343 CHECK_BIDIR2(0,1,0,0)
1344 CHECK_BIDIR2(1,0,0,0)
1346 for(i=8; i<limit; i++){
1347 int fx= motion_fx+vect[i][0];
1348 int fy= motion_fy+vect[i][1];
1349 int bx= motion_bx+vect[i][2];
1350 int by= motion_by+vect[i][3];
1352 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1353 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1355 map[(hashidx+hash[i])&255] = 1;
1357 if(!map[(hashidx+hash[i])&255]){
1359 map[(hashidx+hash[i])&255] = 1;
1360 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1371 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1372 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1373 borderdist= FFMIN(a,b);
1381 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1382 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1383 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1384 s->b_bidir_back_mv_table[xy][1]= motion_by;
1389 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1391 MotionEstContext * const c= &s->me;
1393 const int mot_stride = s->mb_stride;
1394 const int mot_xy = mb_y*mot_stride + mb_x;
1395 const int shift= 1+s->quarter_sample;
1397 const int time_pp= s->pp_time;
1398 const int time_pb= s->pb_time;
1399 int mx, my, xmin, xmax, ymin, ymax;
1400 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1402 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1403 ymin= xmin=(-32)>>shift;
1404 ymax= xmax= 31>>shift;
1406 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1407 s->mv_type= MV_TYPE_8X8;
1409 s->mv_type= MV_TYPE_16X16;
1413 int index= s->block_index[i];
1416 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1417 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1418 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1419 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1420 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1421 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1423 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1424 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1425 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1427 xmax= FFMIN(xmax, s->width - max);
1428 xmin= FFMAX(xmin, - 16 - min);
1430 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1431 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1432 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1434 ymax= FFMIN(ymax, s->height - max);
1435 ymin= FFMAX(ymin, - 16 - min);
1437 if(s->mv_type == MV_TYPE_16X16) break;
1440 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1442 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1443 s->b_direct_mv_table[mot_xy][0]= 0;
1444 s->b_direct_mv_table[mot_xy][1]= 0;
1446 return 256*256*256*64;
1453 c->flags |= FLAG_DIRECT;
1454 c->sub_flags |= FLAG_DIRECT;
1458 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1459 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1461 /* special case for first line */
1462 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1463 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1464 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1465 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1466 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1468 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1469 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1472 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1473 if(c->sub_flags&FLAG_QPEL)
1474 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1476 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1478 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1479 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1481 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1483 mv_table[mot_xy][0]= mx;
1484 mv_table[mot_xy][1]= my;
1485 c->flags &= ~FLAG_DIRECT;
1486 c->sub_flags &= ~FLAG_DIRECT;
1491 void ff_estimate_b_frame_motion(MpegEncContext * s,
1494 MotionEstContext * const c= &s->me;
1495 const int penalty_factor= c->mb_penalty_factor;
1496 int fmin, bmin, dmin, fbmin, bimin, fimin;
1498 const int xy = mb_y*s->mb_stride + mb_x;
1499 init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1500 s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1502 get_limits(s, 16*mb_x, 16*mb_y);
1506 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1507 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1509 score= ((unsigned)(score*score + 128*256))>>16;
1510 c->mc_mb_var_sum_temp += score;
1511 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1512 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1517 if (s->codec_id == AV_CODEC_ID_MPEG4)
1518 dmin= direct_search(s, mb_x, mb_y);
1521 //FIXME penalty stuff for non mpeg4
1523 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1526 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1527 av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1530 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1531 av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1533 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1534 //FIXME mb type penalty
1536 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1537 fimin= interlaced_search(s, 0,
1538 s->b_field_mv_table[0], s->b_field_select_table[0],
1539 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1540 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1541 bimin= interlaced_search(s, 2,
1542 s->b_field_mv_table[1], s->b_field_select_table[1],
1543 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1545 fimin= bimin= INT_MAX;
1549 type = CANDIDATE_MB_TYPE_FORWARD;
1553 type = CANDIDATE_MB_TYPE_DIRECT;
1557 type= CANDIDATE_MB_TYPE_BACKWARD;
1561 type= CANDIDATE_MB_TYPE_BIDIR;
1565 type= CANDIDATE_MB_TYPE_FORWARD_I;
1569 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1572 score= ((unsigned)(score*score + 128*256))>>16;
1573 c->mc_mb_var_sum_temp += score;
1574 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1577 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1578 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1580 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1582 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1583 if(fimin < INT_MAX && bimin < INT_MAX){
1584 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1586 //FIXME something smarter
1587 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1588 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])
1589 type |= CANDIDATE_MB_TYPE_DIRECT0;
1592 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1595 /* find best f_code for ME which do unlimited searches */
1596 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1598 if(s->me_method>=ME_EPZS){
1600 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1601 uint8_t * fcode_tab= s->fcode_tab;
1603 int best_score=-10000000;
1605 if(s->msmpeg4_version)
1606 range= FFMIN(range, 16);
1607 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1608 range= FFMIN(range, 256);
1610 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1612 for(y=0; y<s->mb_height; y++){
1614 int xy= y*s->mb_stride;
1615 for(x=0; x<s->mb_width; x++){
1616 if(s->mb_type[xy] & type){
1617 int mx= mv_table[xy][0];
1618 int my= mv_table[xy][1];
1619 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1620 fcode_tab[my + MAX_MV]);
1623 if(mx >= range || mx < -range ||
1624 my >= range || my < -range)
1627 for(j=0; j<fcode && j<8; j++){
1628 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1637 if(score[i] > best_score){
1638 best_score= score[i];
1649 void ff_fix_long_p_mvs(MpegEncContext * s)
1651 MotionEstContext * const c= &s->me;
1652 const int f_code= s->f_code;
1654 assert(s->pict_type==AV_PICTURE_TYPE_P);
1656 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1658 assert(range <= 16 || !s->msmpeg4_version);
1659 assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1661 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1663 if(s->flags&CODEC_FLAG_4MV){
1664 const int wrap= s->b8_stride;
1666 /* clip / convert to intra 8x8 type MVs */
1667 for(y=0; y<s->mb_height; y++){
1669 int i= y*s->mb_stride;
1672 for(x=0; x<s->mb_width; x++){
1673 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1675 for(block=0; block<4; block++){
1676 int off= (block& 1) + (block>>1)*wrap;
1677 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1678 int my = s->current_picture.motion_val[0][ xy + off ][1];
1680 if( mx >=range || mx <-range
1681 || my >=range || my <-range){
1682 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1683 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1684 s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1697 * @param truncate 1 for truncation, 0 for using intra
1699 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1700 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1702 MotionEstContext * const c= &s->me;
1703 int y, h_range, v_range;
1705 // RAL: 8 in MPEG-1, 16 in MPEG-4
1706 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1708 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1711 v_range= field_select_table ? range>>1 : range;
1713 /* clip / convert to intra 16x16 type MVs */
1714 for(y=0; y<s->mb_height; y++){
1716 int xy= y*s->mb_stride;
1717 for(x=0; x<s->mb_width; x++){
1718 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1719 if(field_select_table==NULL || field_select_table[xy] == field_select){
1720 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1721 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1724 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1725 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1726 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1727 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1729 s->mb_type[xy] &= ~type;
1730 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;