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
36 #include "mpegutils.h"
37 #include "mpegvideo.h"
41 #define P_TOPRIGHT P[3]
45 #define ME_MAP_SHIFT 3
46 #define ME_MAP_MV_BITS 11
48 static int sad_hpel_motion_search(MpegEncContext * s,
49 int *mx_ptr, int *my_ptr, int dmin,
50 int src_index, int ref_index,
53 static inline unsigned update_map_generation(MotionEstContext *c)
55 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
56 if(c->map_generation==0){
57 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
58 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
60 return c->map_generation;
63 /* shape adaptive search stuff */
64 typedef struct Minima{
70 static int minima_cmp(const void *a, const void *b){
71 const Minima *da = (const Minima *) a;
72 const Minima *db = (const Minima *) b;
74 return da->height - db->height;
77 #define FLAG_QPEL 1 //must be 1
81 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
82 const int offset[3]= {
84 ((y*c->uvstride + x)>>1),
85 ((y*c->uvstride + x)>>1),
89 c->src[0][i]= src [i] + offset[i];
90 c->ref[0][i]= ref [i] + offset[i];
94 c->ref[ref_index][i]= ref2[i] + offset[i];
99 static int get_flags(MotionEstContext *c, int direct, int chroma){
100 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
101 + (direct ? FLAG_DIRECT : 0)
102 + (chroma ? FLAG_CHROMA : 0);
105 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
106 const int size, const int h, int ref_index, int src_index,
107 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
108 MotionEstContext * const c= &s->me;
109 const int stride= c->stride;
110 const int hx= subx + (x<<(1+qpel));
111 const int hy= suby + (y<<(1+qpel));
112 uint8_t * const * const ref= c->ref[ref_index];
113 uint8_t * const * const src= c->src[src_index];
115 //FIXME check chroma 4mv, (no crashes ...)
116 av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
117 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
118 const int time_pp= s->pp_time;
119 const int time_pb= s->pb_time;
120 const int mask= 2*qpel+1;
121 if(s->mv_type==MV_TYPE_8X8){
124 int fx = c->direct_basis_mv[i][0] + hx;
125 int fy = c->direct_basis_mv[i][1] + hy;
126 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
127 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
128 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
129 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
131 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
133 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
134 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
136 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
137 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
141 int fx = c->direct_basis_mv[0][0] + hx;
142 int fy = c->direct_basis_mv[0][1] + hy;
143 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
144 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
145 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
146 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
149 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
150 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
151 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
152 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
153 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
154 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
155 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
156 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
158 av_assert2((fx>>1) + 16*s->mb_x >= -16);
159 av_assert2((fy>>1) + 16*s->mb_y >= -16);
160 av_assert2((fx>>1) + 16*s->mb_x <= s->width);
161 av_assert2((fy>>1) + 16*s->mb_y <= s->height);
162 av_assert2((bx>>1) + 16*s->mb_x >= -16);
163 av_assert2((by>>1) + 16*s->mb_y >= -16);
164 av_assert2((bx>>1) + 16*s->mb_x <= s->width);
165 av_assert2((by>>1) + 16*s->mb_y <= s->height);
167 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
168 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
171 d = cmp_func(s, c->temp, src[0], stride, 16);
177 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
178 const int size, const int h, int ref_index, int src_index,
179 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
180 MotionEstContext * const c= &s->me;
181 const int stride= c->stride;
182 const int uvstride= c->uvstride;
183 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
184 const int hx= subx + (x<<(1+qpel));
185 const int hy= suby + (y<<(1+qpel));
186 uint8_t * const * const ref= c->ref[ref_index];
187 uint8_t * const * const src= c->src[src_index];
189 //FIXME check chroma 4mv, (no crashes ...)
190 int uvdxy; /* no, it might not be used uninitialized */
193 if (h << size == 16) {
194 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
195 } else if (size == 0 && h == 8) {
196 c->qpel_put[1][dxy](c->temp , ref[0] + x + y*stride , stride);
197 c->qpel_put[1][dxy](c->temp + 8, ref[0] + x + y*stride + 8, stride);
205 uvdxy= (cx&1) + 2*(cy&1);
206 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
209 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
211 uvdxy= dxy | (x&1) | (2*(y&1));
213 d = cmp_func(s, c->temp, src[0], stride, h);
215 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
217 uvdxy= (x&1) + 2*(y&1);
220 uint8_t * const uvtemp= c->temp + 16*stride;
221 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
222 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
223 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
224 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
229 static int cmp_simple(MpegEncContext *s, const int x, const int y,
230 int ref_index, int src_index,
231 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
232 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
235 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
236 const int size, const int h, int ref_index, int src_index,
237 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
238 if(flags&FLAG_DIRECT){
239 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
241 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
245 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
246 const int size, const int h, int ref_index, int src_index,
247 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
248 if(flags&FLAG_DIRECT){
249 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
251 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);
255 /** @brief compares a block (either a full macroblock or a partition thereof)
256 against a proposed motion-compensated prediction of that block
258 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
259 const int size, const int h, int ref_index, int src_index,
260 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
261 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
262 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
263 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
264 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
265 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
266 && subx==0 && suby==0){
267 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
269 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
273 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
274 const int size, const int h, int ref_index, int src_index,
275 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
276 if(flags&FLAG_DIRECT){
277 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
279 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
283 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
284 const int size, const int h, int ref_index, int src_index,
285 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
286 if(flags&FLAG_DIRECT){
287 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
289 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
293 #include "motion_est_template.c"
295 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
296 ptrdiff_t stride, int h)
301 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
304 int ff_init_me(MpegEncContext *s){
305 MotionEstContext * const c= &s->me;
306 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
307 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
309 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
310 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
313 //special case of snow is needed because snow uses its own iterative ME code
314 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
315 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");
321 if(s->codec_id == AV_CODEC_ID_H261)
322 c->avctx->me_sub_cmp = c->avctx->me_cmp;
324 if(cache_size < 2*dia_size && !c->stride){
325 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
328 ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
329 ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
330 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
331 ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
333 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
334 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
335 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
337 /*FIXME s->no_rounding b_type*/
338 if(s->flags&CODEC_FLAG_QPEL){
339 c->sub_motion_search= qpel_motion_search;
340 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
342 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
344 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
346 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
347 c->sub_motion_search= hpel_motion_search;
348 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
349 && c->avctx-> me_cmp == FF_CMP_SAD
350 && c->avctx-> mb_cmp == FF_CMP_SAD)
351 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
353 c->sub_motion_search= hpel_motion_search;
355 c->hpel_avg = s->hdsp.avg_pixels_tab;
357 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
359 c->hpel_put = s->hdsp.put_pixels_tab;
362 c->stride = s->linesize;
363 c->uvstride= s->uvlinesize;
365 c->stride = 16*s->mb_width + 32;
366 c->uvstride= 8*s->mb_width + 16;
369 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
370 * not have yet, and even if we had, the motion estimation code
371 * does not expect it. */
372 if (s->codec_id != AV_CODEC_ID_SNOW) {
373 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
374 s->mecc.me_cmp[2] = zero_cmp;
375 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
376 s->mecc.me_sub_cmp[2] = zero_cmp;
377 c->hpel_put[2][0]= c->hpel_put[2][1]=
378 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
381 if(s->codec_id == AV_CODEC_ID_H261){
382 c->sub_motion_search= no_sub_motion_search;
388 #define CHECK_SAD_HALF_MV(suffix, x, y) \
390 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
391 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
392 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
395 static int sad_hpel_motion_search(MpegEncContext * s,
396 int *mx_ptr, int *my_ptr, int dmin,
397 int src_index, int ref_index,
400 MotionEstContext * const c= &s->me;
401 const int penalty_factor= c->sub_penalty_factor;
404 int stride= c->stride;
407 av_assert2(c->sub_flags == 0);
415 pix = c->src[src_index][0];
419 ptr = c->ref[ref_index][0] + (my * stride) + mx;
423 if (mx > xmin && mx < xmax &&
424 my > ymin && my < ymax) {
427 const int index= (my<<ME_MAP_SHIFT) + mx;
428 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
429 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
430 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
431 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
441 CHECK_SAD_HALF_MV(y2 , 0, -1)
443 CHECK_SAD_HALF_MV(xy2, -1, -1)
445 CHECK_SAD_HALF_MV(xy2, +1, -1)
449 CHECK_SAD_HALF_MV(xy2, -1, +1)
451 CHECK_SAD_HALF_MV(x2 , -1, 0)
453 CHECK_SAD_HALF_MV(xy2, +1, -1)
455 CHECK_SAD_HALF_MV(xy2, -1, -1)
459 CHECK_SAD_HALF_MV(xy2, +1, +1)
461 CHECK_SAD_HALF_MV(x2 , +1, 0)
466 CHECK_SAD_HALF_MV(xy2, -1, -1)
470 CHECK_SAD_HALF_MV(xy2, +1, +1)
472 CHECK_SAD_HALF_MV(x2 , -1, 0)
473 CHECK_SAD_HALF_MV(xy2, -1, +1)
476 CHECK_SAD_HALF_MV(xy2, +1, -1)
480 CHECK_SAD_HALF_MV(xy2, -1, +1)
482 CHECK_SAD_HALF_MV(x2 , +1, 0)
483 CHECK_SAD_HALF_MV(xy2, +1, +1)
485 CHECK_SAD_HALF_MV(y2 , 0, +1)
500 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
502 const int xy= s->mb_x + s->mb_y*s->mb_stride;
504 s->p_mv_table[xy][0] = mx;
505 s->p_mv_table[xy][1] = my;
507 /* has already been set to the 4 MV if 4MV is done */
509 int mot_xy= s->block_index[0];
511 s->current_picture.motion_val[0][mot_xy ][0] = mx;
512 s->current_picture.motion_val[0][mot_xy ][1] = my;
513 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
514 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
516 mot_xy += s->b8_stride;
517 s->current_picture.motion_val[0][mot_xy ][0] = mx;
518 s->current_picture.motion_val[0][mot_xy ][1] = my;
519 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
520 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
525 * get fullpel ME search limits.
527 static inline void get_limits(MpegEncContext *s, int x, int y)
529 MotionEstContext * const c= &s->me;
530 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
531 int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
533 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
536 if (s->unrestricted_mv) {
539 c->xmax = - x + s->width;
540 c->ymax = - y + s->height;
541 } else if (s->out_format == FMT_H261){
542 // Search range of H261 is different from other codec standards
543 c->xmin = (x > 15) ? - 15 : 0;
544 c->ymin = (y > 15) ? - 15 : 0;
545 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
546 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
550 c->xmax = - x + s->mb_width *16 - 16;
551 c->ymax = - y + s->mb_height*16 - 16;
553 if(!range || range > max_range)
556 c->xmin = FFMAX(c->xmin,-range);
557 c->xmax = FFMIN(c->xmax, range);
558 c->ymin = FFMAX(c->ymin,-range);
559 c->ymax = FFMIN(c->ymax, range);
563 static inline void init_mv4_ref(MotionEstContext *c){
564 const int stride= c->stride;
566 c->ref[1][0] = c->ref[0][0] + 8;
567 c->ref[2][0] = c->ref[0][0] + 8*stride;
568 c->ref[3][0] = c->ref[2][0] + 8;
569 c->src[1][0] = c->src[0][0] + 8;
570 c->src[2][0] = c->src[0][0] + 8*stride;
571 c->src[3][0] = c->src[2][0] + 8;
574 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
576 MotionEstContext * const c= &s->me;
581 int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
583 const int stride= c->stride;
584 uint8_t *mv_penalty= c->current_mv_penalty;
585 int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15);
589 for(block=0; block<4; block++){
591 int pred_x4, pred_y4;
593 static const int off[4]= {2, 1, 1, -1};
594 const int mot_stride = s->b8_stride;
595 const int mot_xy = s->block_index[block];
598 c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
599 c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
602 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
603 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
605 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
607 /* special case for first line */
608 if (s->first_slice_line && block<2) {
609 c->pred_x= pred_x4= P_LEFT[0];
610 c->pred_y= pred_y4= P_LEFT[1];
612 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
613 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
614 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
615 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
616 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
617 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
618 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
619 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
621 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
622 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
624 c->pred_x= pred_x4 = P_MEDIAN[0];
625 c->pred_y= pred_y4 = P_MEDIAN[1];
631 if (s->first_slice_line && block<2 && i>1 && i<9)
635 if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
636 if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
639 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
641 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
643 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
645 const int offset= ((block&1) + (block>>1)*stride)*8;
646 uint8_t *dest_y = c->scratchpad + offset;
647 if(s->quarter_sample){
648 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
649 dxy = ((my4 & 3) << 2) | (mx4 & 3);
652 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
654 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
656 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
657 dxy = ((my4 & 1) << 1) | (mx4 & 1);
660 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
662 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
664 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
668 if(s->quarter_sample){
676 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
677 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
679 if(mx4 != mx || my4 != my) same=0;
685 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
686 dmin_sum += s->mecc.mb_cmp[0](s,
687 s->new_picture.f->data[0] +
688 s->mb_x * 16 + s->mb_y * 16 * stride,
689 c->scratchpad, stride, 16);
692 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
697 mx= ff_h263_round_chroma(mx4_sum);
698 my= ff_h263_round_chroma(my4_sum);
699 dxy = ((my & 1) << 1) | (mx & 1);
701 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
704 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
705 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
707 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
708 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
711 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);
712 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);
718 switch(c->avctx->mb_cmp&0xFF){
720 return dmin_sum+ 32*s->qscale*s->qscale;*/
724 return dmin_sum+ 11*c->mb_penalty_factor;
728 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
729 MotionEstContext * const c= &s->me;
731 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
732 c->src[1][0] = c->src[0][0] + s->linesize;
733 if(c->flags & FLAG_CHROMA){
734 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
735 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
736 c->src[1][1] = c->src[0][1] + s->uvlinesize;
737 c->src[1][2] = c->src[0][2] + s->uvlinesize;
741 static int interlaced_search(MpegEncContext *s, int ref_index,
742 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
744 MotionEstContext * const c= &s->me;
749 uint8_t * const mv_penalty= c->current_mv_penalty;
751 const int stride= 2*s->linesize;
753 const int mot_stride= s->mb_stride;
754 const int xy= s->mb_x + s->mb_y*mot_stride;
760 init_interlaced_ref(s, ref_index);
762 for(block=0; block<2; block++){
764 int best_dmin= INT_MAX;
767 for(field_select=0; field_select<2; field_select++){
768 int dmin, mx_i, my_i;
769 int16_t (*mv_table)[2]= mv_tables[block][field_select];
771 if(user_field_select){
772 av_assert1(field_select==0 || field_select==1);
773 av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
774 if(field_select_tables[block][xy] != field_select)
778 P_LEFT[0] = mv_table[xy - 1][0];
779 P_LEFT[1] = mv_table[xy - 1][1];
780 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
782 c->pred_x= P_LEFT[0];
783 c->pred_y= P_LEFT[1];
785 if(!s->first_slice_line){
786 P_TOP[0] = mv_table[xy - mot_stride][0];
787 P_TOP[1] = mv_table[xy - mot_stride][1];
788 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
789 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
790 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
791 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
792 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
793 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
795 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
796 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
798 P_MV1[0]= mx; //FIXME not correct if block != field_select
801 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
803 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
805 mv_table[xy][0]= mx_i;
806 mv_table[xy][1]= my_i;
808 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
812 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
813 dxy = ((my_i & 1) << 1) | (mx_i & 1);
816 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
818 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
820 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
821 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
823 dmin+= c->mb_penalty_factor; //field_select bits
825 dmin += field_select != block; //slightly prefer same field
827 if(dmin < best_dmin){
829 best_field= field_select;
833 int16_t (*mv_table)[2]= mv_tables[block][best_field];
835 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
836 if(mv_table[xy][1]&1) same=0;
837 if(mv_table[xy][1]*2 != my) same=0;
838 if(best_field != block) same=0;
841 field_select_tables[block][xy]= best_field;
842 dmin_sum += best_dmin;
853 switch(c->avctx->mb_cmp&0xFF){
855 return dmin_sum+ 32*s->qscale*s->qscale;*/
859 return dmin_sum+ 11*c->mb_penalty_factor;
863 static inline int get_penalty_factor(int lambda, int lambda2, int type){
867 return lambda>>FF_LAMBDA_SHIFT;
869 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
871 return (4*lambda)>>(FF_LAMBDA_SHIFT);
873 return (2*lambda)>>(FF_LAMBDA_SHIFT);
876 return (2*lambda)>>FF_LAMBDA_SHIFT;
881 return lambda2>>FF_LAMBDA_SHIFT;
887 void ff_estimate_p_frame_motion(MpegEncContext * s,
890 MotionEstContext * const c= &s->me;
892 int sum, mx, my, dmin;
893 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
894 int vard; ///< sum of squared differences with the estimated motion vector
896 const int shift= 1+s->quarter_sample;
898 Picture * const pic= &s->current_picture;
900 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
902 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
903 av_assert0(s->linesize == c->stride);
904 av_assert0(s->uvlinesize == c->uvstride);
906 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
907 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
908 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
909 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
911 get_limits(s, 16*mb_x, 16*mb_y);
914 /* intra / predictive decision */
916 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
917 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
918 (((unsigned) sum * sum) >> 8) + 500;
920 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
921 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
922 c->mb_var_sum_temp += (varc+128)>>8;
924 switch(s->me_method) {
934 const int mot_stride = s->b8_stride;
935 const int mot_xy = s->block_index[0];
937 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
938 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
940 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
942 if(!s->first_slice_line) {
943 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
944 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
945 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
946 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
947 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
948 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
949 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
951 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
952 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
954 if(s->out_format == FMT_H263){
955 c->pred_x = P_MEDIAN[0];
956 c->pred_y = P_MEDIAN[1];
957 }else { /* mpeg1 at least */
958 c->pred_x= P_LEFT[0];
959 c->pred_y= P_LEFT[1];
962 c->pred_x= P_LEFT[0];
963 c->pred_y= P_LEFT[1];
967 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
972 /* At this point (mx,my) are full-pell and the relative displacement */
973 ppix = c->ref[0][0] + (my * s->linesize) + mx;
975 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
977 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
978 c->mc_mb_var_sum_temp += (vard+128)>>8;
980 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
981 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
982 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
983 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
985 if (vard*2 + 200*256 > varc)
986 mb_type|= CANDIDATE_MB_TYPE_INTRA;
987 if (varc*2 + 200*256 > vard || s->qscale > 24){
988 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
989 mb_type|= CANDIDATE_MB_TYPE_INTER;
990 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
991 if (s->mpv_flags & FF_MPV_FLAG_MV0)
993 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
998 if((s->flags&CODEC_FLAG_4MV)
999 && !c->skip && varc>50<<8 && vard>10<<8){
1000 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1001 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1003 set_p_mv_tables(s, mx, my, 0);
1005 set_p_mv_tables(s, mx, my, 1);
1006 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1007 && !c->skip){ //FIXME varc/d checks
1008 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1009 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1013 mb_type= CANDIDATE_MB_TYPE_INTER;
1015 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1016 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1017 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1019 if((s->flags&CODEC_FLAG_4MV)
1020 && !c->skip && varc>50<<8 && vard>10<<8){
1021 int dmin4= h263_mv4_search(s, mx, my, shift);
1023 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1027 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1028 && !c->skip){ //FIXME varc/d checks
1029 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1031 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1036 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1038 /* get intra luma score */
1039 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1040 intra_score= varc - 500;
1042 unsigned mean = (sum+128)>>8;
1045 for(i=0; i<16; i++){
1046 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1047 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1048 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1049 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1052 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1054 intra_score += c->mb_penalty_factor*16;
1056 if(intra_score < dmin){
1057 mb_type= CANDIDATE_MB_TYPE_INTRA;
1058 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1060 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1063 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1064 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1065 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1069 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1072 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1075 MotionEstContext * const c= &s->me;
1078 const int shift= 1+s->quarter_sample;
1079 const int xy= mb_x + mb_y*s->mb_stride;
1080 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1082 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1084 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1085 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1087 get_limits(s, 16*mb_x, 16*mb_y);
1090 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1091 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1093 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1095 /* special case for first line */
1096 if (s->first_slice_line) {
1097 c->pred_x= P_LEFT[0];
1098 c->pred_y= P_LEFT[1];
1099 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1100 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1102 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1103 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1104 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1105 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1106 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1107 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1108 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1110 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1111 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1113 c->pred_x = P_MEDIAN[0];
1114 c->pred_y = P_MEDIAN[1];
1117 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1119 s->p_mv_table[xy][0] = mx<<shift;
1120 s->p_mv_table[xy][1] = my<<shift;
1125 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1126 int16_t (*mv_table)[2], int ref_index, int f_code)
1128 MotionEstContext * const c= &s->me;
1131 const int shift= 1+s->quarter_sample;
1132 const int mot_stride = s->mb_stride;
1133 const int mot_xy = mb_y*mot_stride + mb_x;
1134 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1137 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1138 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1139 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1140 c->current_mv_penalty= mv_penalty;
1142 get_limits(s, 16*mb_x, 16*mb_y);
1144 switch(s->me_method) {
1153 P_LEFT[0] = mv_table[mot_xy - 1][0];
1154 P_LEFT[1] = mv_table[mot_xy - 1][1];
1156 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1158 /* special case for first line */
1159 if (!s->first_slice_line) {
1160 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1161 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1162 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1163 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1164 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1165 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1166 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1168 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1169 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1171 c->pred_x = P_LEFT[0];
1172 c->pred_y = P_LEFT[1];
1174 if(mv_table == s->b_forw_mv_table){
1175 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1177 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1180 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1185 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1187 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1188 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1190 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1191 mv_table[mot_xy][0]= mx;
1192 mv_table[mot_xy][1]= my;
1197 static inline int check_bidir_mv(MpegEncContext * s,
1198 int motion_fx, int motion_fy,
1199 int motion_bx, int motion_by,
1200 int pred_fx, int pred_fy,
1201 int pred_bx, int pred_by,
1205 //FIXME better f_code prediction (max mv & distance)
1207 MotionEstContext * const c= &s->me;
1208 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1209 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1210 int stride= c->stride;
1211 uint8_t *dest_y = c->scratchpad;
1216 uint8_t **src_data= c->src[0];
1217 uint8_t **ref_data= c->ref[0];
1218 uint8_t **ref2_data= c->ref[2];
1220 if(s->quarter_sample){
1221 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1222 src_x = motion_fx >> 2;
1223 src_y = motion_fy >> 2;
1225 ptr = ref_data[0] + (src_y * stride) + src_x;
1226 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1228 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1229 src_x = motion_bx >> 2;
1230 src_y = motion_by >> 2;
1232 ptr = ref2_data[0] + (src_y * stride) + src_x;
1233 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1235 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1236 src_x = motion_fx >> 1;
1237 src_y = motion_fy >> 1;
1239 ptr = ref_data[0] + (src_y * stride) + src_x;
1240 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1242 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1243 src_x = motion_bx >> 1;
1244 src_y = motion_by >> 1;
1246 ptr = ref2_data[0] + (src_y * stride) + src_x;
1247 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1250 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1251 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1252 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1254 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1261 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1262 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1264 MotionEstContext * const c= &s->me;
1265 const int mot_stride = s->mb_stride;
1266 const int xy = mb_y *mot_stride + mb_x;
1268 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1269 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1270 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1271 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1272 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1273 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1274 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1275 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1276 const int flags= c->sub_flags;
1277 const int qpel= flags&FLAG_QPEL;
1278 const int shift= 1+qpel;
1279 const int xmin= c->xmin<<shift;
1280 const int ymin= c->ymin<<shift;
1281 const int xmax= c->xmax<<shift;
1282 const int ymax= c->ymax<<shift;
1283 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1284 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1285 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1286 uint8_t map[256] = { 0 };
1288 map[hashidx&255] = 1;
1290 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1291 motion_bx, motion_by,
1296 if(s->avctx->bidir_refine){
1298 static const uint8_t limittab[5]={0,8,32,64,80};
1299 const int limit= limittab[s->avctx->bidir_refine];
1300 static const int8_t vect[][4]={
1301 { 0, 0, 0, 1}, { 0, 0, 0,-1}, { 0, 0, 1, 0}, { 0, 0,-1, 0}, { 0, 1, 0, 0}, { 0,-1, 0, 0}, { 1, 0, 0, 0}, {-1, 0, 0, 0},
1303 { 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},
1304 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 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},
1308 { 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},
1309 { 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1},
1310 { 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1},
1311 { 0, 1, 1,-1}, { 0,-1,-1, 1}, { 1, 1,-1, 0}, {-1,-1, 1, 0}, { 1,-1, 0, 1}, {-1, 1, 0,-1}, {-1, 0, 1, 1}, { 1, 0,-1,-1},
1313 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1314 { 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},
1315 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1317 static const uint8_t hash[]={
1318 HASH8( 0, 0, 0, 1), HASH8( 0, 0, 0,-1), HASH8( 0, 0, 1, 0), HASH8( 0, 0,-1, 0), HASH8( 0, 1, 0, 0), HASH8( 0,-1, 0, 0), HASH8( 1, 0, 0, 0), HASH8(-1, 0, 0, 0),
1320 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),
1321 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 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),
1325 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),
1326 HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1),
1327 HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1),
1328 HASH8( 0, 1, 1,-1), HASH8( 0,-1,-1, 1), HASH8( 1, 1,-1, 0), HASH8(-1,-1, 1, 0), HASH8( 1,-1, 0, 1), HASH8(-1, 1, 0,-1), HASH8(-1, 0, 1, 1), HASH8( 1, 0,-1,-1),
1330 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1331 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),
1332 HASH8( 1, 1,-1,-1), HASH8(-1,-1, 1, 1), HASH8( 1,-1,-1, 1), HASH8(-1, 1, 1,-1), HASH8( 1,-1, 1,-1), HASH8(-1, 1,-1, 1),
1335 #define CHECK_BIDIR(fx,fy,bx,by)\
1336 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1337 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1338 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1340 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1341 score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
1343 hashidx += HASH(fx,fy,bx,by);\
1352 #define CHECK_BIDIR2(a,b,c,d)\
1353 CHECK_BIDIR(a,b,c,d)\
1354 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1361 CHECK_BIDIR2(0,0,0,1)
1362 CHECK_BIDIR2(0,0,1,0)
1363 CHECK_BIDIR2(0,1,0,0)
1364 CHECK_BIDIR2(1,0,0,0)
1366 for(i=8; i<limit; i++){
1367 int fx= motion_fx+vect[i][0];
1368 int fy= motion_fy+vect[i][1];
1369 int bx= motion_bx+vect[i][2];
1370 int by= motion_by+vect[i][3];
1372 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1373 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1375 map[(hashidx+hash[i])&255] = 1;
1377 if(!map[(hashidx+hash[i])&255]){
1379 map[(hashidx+hash[i])&255] = 1;
1380 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1391 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1392 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1393 borderdist= FFMIN(a,b);
1401 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1402 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1403 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1404 s->b_bidir_back_mv_table[xy][1]= motion_by;
1409 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1411 MotionEstContext * const c= &s->me;
1413 const int mot_stride = s->mb_stride;
1414 const int mot_xy = mb_y*mot_stride + mb_x;
1415 const int shift= 1+s->quarter_sample;
1417 const int time_pp= s->pp_time;
1418 const int time_pb= s->pb_time;
1419 int mx, my, xmin, xmax, ymin, ymax;
1420 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1422 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1423 ymin= xmin=(-32)>>shift;
1424 ymax= xmax= 31>>shift;
1426 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1427 s->mv_type= MV_TYPE_8X8;
1429 s->mv_type= MV_TYPE_16X16;
1433 int index= s->block_index[i];
1436 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1437 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1438 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1439 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1440 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1441 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1443 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1444 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1445 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1447 xmax= FFMIN(xmax, s->width - max);
1448 xmin= FFMAX(xmin, - 16 - min);
1450 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1451 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1452 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1454 ymax= FFMIN(ymax, s->height - max);
1455 ymin= FFMAX(ymin, - 16 - min);
1457 if(s->mv_type == MV_TYPE_16X16) break;
1460 av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1462 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1463 s->b_direct_mv_table[mot_xy][0]= 0;
1464 s->b_direct_mv_table[mot_xy][1]= 0;
1466 return 256*256*256*64;
1473 c->flags |= FLAG_DIRECT;
1474 c->sub_flags |= FLAG_DIRECT;
1478 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1479 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1481 /* special case for first line */
1482 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1483 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1484 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1485 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1486 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1488 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1489 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1492 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1493 if(c->sub_flags&FLAG_QPEL)
1494 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1496 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1498 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1499 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1501 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1503 mv_table[mot_xy][0]= mx;
1504 mv_table[mot_xy][1]= my;
1505 c->flags &= ~FLAG_DIRECT;
1506 c->sub_flags &= ~FLAG_DIRECT;
1511 void ff_estimate_b_frame_motion(MpegEncContext * s,
1514 MotionEstContext * const c= &s->me;
1515 const int penalty_factor= c->mb_penalty_factor;
1516 int fmin, bmin, dmin, fbmin, bimin, fimin;
1518 const int xy = mb_y*s->mb_stride + mb_x;
1519 init_ref(c, s->new_picture.f->data, s->last_picture.f->data,
1520 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1522 get_limits(s, 16*mb_x, 16*mb_y);
1526 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1527 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1529 score= ((unsigned)(score*score + 128*256))>>16;
1530 c->mc_mb_var_sum_temp += score;
1531 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1532 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1537 if (s->codec_id == AV_CODEC_ID_MPEG4)
1538 dmin= direct_search(s, mb_x, mb_y);
1541 //FIXME penalty stuff for non mpeg4
1543 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1547 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1549 av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1552 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1553 av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1555 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1556 //FIXME mb type penalty
1558 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1559 fimin= interlaced_search(s, 0,
1560 s->b_field_mv_table[0], s->b_field_select_table[0],
1561 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1562 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1563 bimin= interlaced_search(s, 2,
1564 s->b_field_mv_table[1], s->b_field_select_table[1],
1565 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1567 fimin= bimin= INT_MAX;
1571 type = CANDIDATE_MB_TYPE_FORWARD;
1575 type = CANDIDATE_MB_TYPE_DIRECT;
1579 type= CANDIDATE_MB_TYPE_BACKWARD;
1583 type= CANDIDATE_MB_TYPE_BIDIR;
1587 type= CANDIDATE_MB_TYPE_FORWARD_I;
1591 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1594 score= ((unsigned)(score*score + 128*256))>>16;
1595 c->mc_mb_var_sum_temp += score;
1596 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1599 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1600 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1602 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1604 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1605 if(fimin < INT_MAX && bimin < INT_MAX){
1606 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1608 //FIXME something smarter
1609 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1610 if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1611 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1612 type |= CANDIDATE_MB_TYPE_DIRECT0;
1615 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1618 /* find best f_code for ME which do unlimited searches */
1619 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1621 if(s->me_method>=ME_EPZS){
1623 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1624 uint8_t * fcode_tab= s->fcode_tab;
1626 int best_score=-10000000;
1628 if(s->msmpeg4_version)
1629 range= FFMIN(range, 16);
1630 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1631 range= FFMIN(range, 256);
1633 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1635 for(y=0; y<s->mb_height; y++){
1637 int xy= y*s->mb_stride;
1638 for(x=0; x<s->mb_width; x++){
1639 if(s->mb_type[xy] & type){
1640 int mx= mv_table[xy][0];
1641 int my= mv_table[xy][1];
1642 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1643 fcode_tab[my + MAX_MV]);
1646 if(mx >= range || mx < -range ||
1647 my >= range || my < -range)
1650 for(j=0; j<fcode && j<8; j++){
1651 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1660 if(score[i] > best_score){
1661 best_score= score[i];
1672 void ff_fix_long_p_mvs(MpegEncContext * s)
1674 MotionEstContext * const c= &s->me;
1675 const int f_code= s->f_code;
1677 av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1679 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1681 av_assert0(range <= 16 || !s->msmpeg4_version);
1682 av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1684 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1686 if(s->flags&CODEC_FLAG_4MV){
1687 const int wrap= s->b8_stride;
1689 /* clip / convert to intra 8x8 type MVs */
1690 for(y=0; y<s->mb_height; y++){
1692 int i= y*s->mb_stride;
1695 for(x=0; x<s->mb_width; x++){
1696 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1698 for(block=0; block<4; block++){
1699 int off= (block& 1) + (block>>1)*wrap;
1700 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1701 int my = s->current_picture.motion_val[0][ xy + off ][1];
1703 if( mx >=range || mx <-range
1704 || my >=range || my <-range){
1705 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1706 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1707 s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1720 * @param truncate 1 for truncation, 0 for using intra
1722 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1723 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1725 MotionEstContext * const c= &s->me;
1726 int y, h_range, v_range;
1728 // RAL: 8 in MPEG-1, 16 in MPEG-4
1729 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1731 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1734 v_range= field_select_table ? range>>1 : range;
1736 /* clip / convert to intra 16x16 type MVs */
1737 for(y=0; y<s->mb_height; y++){
1739 int xy= y*s->mb_stride;
1740 for(x=0; x<s->mb_width; x++){
1741 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1742 if (!field_select_table || field_select_table[xy] == field_select) {
1743 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1744 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1747 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1748 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1749 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1750 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1752 s->mb_type[xy] &= ~type;
1753 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;