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 "motion_est.h"
37 #include "mpegutils.h"
38 #include "mpegvideoenc.h"
42 #define P_TOPRIGHT P[3]
46 #define ME_MAP_SHIFT 3
47 #define ME_MAP_MV_BITS 11
49 static int sad_hpel_motion_search(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
51 int src_index, int ref_index,
54 static inline unsigned update_map_generation(MotionEstContext *c)
56 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57 if(c->map_generation==0){
58 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
61 return c->map_generation;
64 /* shape adaptive search stuff */
65 typedef struct Minima{
71 static int minima_cmp(const void *a, const void *b){
72 const Minima *da = (const Minima *) a;
73 const Minima *db = (const Minima *) b;
75 return da->height - db->height;
78 #define FLAG_QPEL 1 //must be 1
82 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){
83 const int offset[3]= {
85 ((y*c->uvstride + x)>>1),
86 ((y*c->uvstride + x)>>1),
90 c->src[0][i]= src [i] + offset[i];
91 c->ref[0][i]= ref [i] + offset[i];
95 c->ref[ref_index][i]= ref2[i] + offset[i];
100 static int get_flags(MotionEstContext *c, int direct, int chroma){
101 return ((c->avctx->flags&AV_CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102 + (direct ? FLAG_DIRECT : 0)
103 + (chroma ? FLAG_CHROMA : 0);
106 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
107 const int size, const int h, int ref_index, int src_index,
108 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
109 MotionEstContext * const c= &s->me;
110 const int stride= c->stride;
111 const int hx = subx + x * (1 << (1 + qpel));
112 const int hy = suby + y * (1 << (1 + qpel));
113 uint8_t * const * const ref= c->ref[ref_index];
114 uint8_t * const * const src= c->src[src_index];
116 //FIXME check chroma 4mv, (no crashes ...)
117 av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
118 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
119 const int time_pp= s->pp_time;
120 const int time_pb= s->pb_time;
121 const int mask= 2*qpel+1;
122 if(s->mv_type==MV_TYPE_8X8){
125 int fx = c->direct_basis_mv[i][0] + hx;
126 int fy = c->direct_basis_mv[i][1] + hy;
127 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));
128 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));
129 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
130 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
132 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
134 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
135 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
137 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
138 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
142 int fx = c->direct_basis_mv[0][0] + hx;
143 int fy = c->direct_basis_mv[0][1] + hy;
144 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
145 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
146 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
147 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
150 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
151 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
152 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
153 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
154 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
155 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
156 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
157 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
159 av_assert2((fx>>1) + 16*s->mb_x >= -16);
160 av_assert2((fy>>1) + 16*s->mb_y >= -16);
161 av_assert2((fx>>1) + 16*s->mb_x <= s->width);
162 av_assert2((fy>>1) + 16*s->mb_y <= s->height);
163 av_assert2((bx>>1) + 16*s->mb_x >= -16);
164 av_assert2((by>>1) + 16*s->mb_y >= -16);
165 av_assert2((bx>>1) + 16*s->mb_x <= s->width);
166 av_assert2((by>>1) + 16*s->mb_y <= s->height);
168 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
169 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
172 d = cmp_func(s, c->temp, src[0], stride, 16);
178 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
179 const int size, const int h, int ref_index, int src_index,
180 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
181 MotionEstContext * const c= &s->me;
182 const int stride= c->stride;
183 const int uvstride= c->uvstride;
184 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
185 const int hx= subx + x*(1<<(1+qpel));
186 const int hy= suby + y*(1<<(1+qpel));
187 uint8_t * const * const ref= c->ref[ref_index];
188 uint8_t * const * const src= c->src[src_index];
190 //FIXME check chroma 4mv, (no crashes ...)
191 int uvdxy; /* no, it might not be used uninitialized */
194 if (h << size == 16) {
195 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
196 } else if (size == 0 && h == 8) {
197 c->qpel_put[1][dxy](c->temp , ref[0] + x + y*stride , stride);
198 c->qpel_put[1][dxy](c->temp + 8, ref[0] + x + y*stride + 8, stride);
206 uvdxy= (cx&1) + 2*(cy&1);
207 // FIXME x/y wrong, but MPEG-4 qpel is sick anyway, we should drop as much of it as possible in favor for H.264
210 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
212 uvdxy= dxy | (x&1) | (2*(y&1));
214 d = cmp_func(s, c->temp, src[0], stride, h);
216 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
218 uvdxy= (x&1) + 2*(y&1);
221 uint8_t * const uvtemp= c->temp + 16*stride;
222 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
223 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
224 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
225 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
230 static int cmp_simple(MpegEncContext *s, const int x, const int y,
231 int ref_index, int src_index,
232 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
233 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
236 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
237 const int size, const int h, int ref_index, int src_index,
238 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
239 if(flags&FLAG_DIRECT){
240 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
242 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
246 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
247 const int size, const int h, int ref_index, int src_index,
248 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
249 if(flags&FLAG_DIRECT){
250 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
252 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);
256 /** @brief compares a block (either a full macroblock or a partition thereof)
257 against a proposed motion-compensated prediction of that block
259 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
260 const int size, const int h, int ref_index, int src_index,
261 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
262 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
263 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
264 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
265 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
266 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
267 && subx==0 && suby==0){
268 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
270 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
274 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
275 const int size, const int h, int ref_index, int src_index,
276 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
277 if(flags&FLAG_DIRECT){
278 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
280 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
284 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
285 const int size, const int h, int ref_index, int src_index,
286 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
287 if(flags&FLAG_DIRECT){
288 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
290 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
294 #include "motion_est_template.c"
296 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
297 ptrdiff_t stride, int h)
302 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
305 int ff_init_me(MpegEncContext *s){
306 MotionEstContext * const c= &s->me;
307 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
308 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
310 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
311 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
317 if(s->codec_id == AV_CODEC_ID_H261)
318 c->avctx->me_sub_cmp = c->avctx->me_cmp;
320 if(cache_size < 2*dia_size && !c->stride){
321 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
324 ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
325 ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
326 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
327 ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
329 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
330 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
331 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
333 /*FIXME s->no_rounding b_type*/
334 if (s->avctx->flags & AV_CODEC_FLAG_QPEL) {
335 c->sub_motion_search= qpel_motion_search;
336 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
338 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
340 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
342 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
343 c->sub_motion_search= hpel_motion_search;
344 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
345 && c->avctx-> me_cmp == FF_CMP_SAD
346 && c->avctx-> mb_cmp == FF_CMP_SAD)
347 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
349 c->sub_motion_search= hpel_motion_search;
351 c->hpel_avg = s->hdsp.avg_pixels_tab;
353 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
355 c->hpel_put = s->hdsp.put_pixels_tab;
358 c->stride = s->linesize;
359 c->uvstride= s->uvlinesize;
361 c->stride = 16*s->mb_width + 32;
362 c->uvstride= 8*s->mb_width + 16;
365 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
366 * not have yet, and even if we had, the motion estimation code
367 * does not expect it. */
368 if (s->codec_id != AV_CODEC_ID_SNOW) {
369 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
370 s->mecc.me_cmp[2] = zero_cmp;
371 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
372 s->mecc.me_sub_cmp[2] = zero_cmp;
373 c->hpel_put[2][0]= c->hpel_put[2][1]=
374 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
377 if(s->codec_id == AV_CODEC_ID_H261){
378 c->sub_motion_search= no_sub_motion_search;
384 #define CHECK_SAD_HALF_MV(suffix, x, y) \
386 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
387 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
388 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
391 static int sad_hpel_motion_search(MpegEncContext * s,
392 int *mx_ptr, int *my_ptr, int dmin,
393 int src_index, int ref_index,
396 MotionEstContext * const c= &s->me;
397 const int penalty_factor= c->sub_penalty_factor;
400 int stride= c->stride;
403 av_assert2(c->sub_flags == 0);
411 pix = c->src[src_index][0];
415 ptr = c->ref[ref_index][0] + (my * stride) + mx;
419 if (mx > xmin && mx < xmax &&
420 my > ymin && my < ymax) {
423 const int index= my*(1<<ME_MAP_SHIFT) + mx;
424 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
425 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
426 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
427 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
437 CHECK_SAD_HALF_MV(y2 , 0, -1)
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)
449 CHECK_SAD_HALF_MV(xy2, +1, -1)
451 CHECK_SAD_HALF_MV(xy2, -1, -1)
455 CHECK_SAD_HALF_MV(xy2, +1, +1)
457 CHECK_SAD_HALF_MV(x2 , +1, 0)
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)
472 CHECK_SAD_HALF_MV(xy2, +1, -1)
476 CHECK_SAD_HALF_MV(xy2, -1, +1)
478 CHECK_SAD_HALF_MV(x2 , +1, 0)
479 CHECK_SAD_HALF_MV(xy2, +1, +1)
481 CHECK_SAD_HALF_MV(y2 , 0, +1)
496 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
498 const int xy= s->mb_x + s->mb_y*s->mb_stride;
500 s->p_mv_table[xy][0] = mx;
501 s->p_mv_table[xy][1] = my;
503 /* has already been set to the 4 MV if 4MV is done */
505 int mot_xy= s->block_index[0];
507 s->current_picture.motion_val[0][mot_xy ][0] = mx;
508 s->current_picture.motion_val[0][mot_xy ][1] = my;
509 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
510 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
512 mot_xy += s->b8_stride;
513 s->current_picture.motion_val[0][mot_xy ][0] = mx;
514 s->current_picture.motion_val[0][mot_xy ][1] = my;
515 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
516 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
521 * get fullpel ME search limits.
523 static inline void get_limits(MpegEncContext *s, int x, int y)
525 MotionEstContext * const c= &s->me;
526 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
527 int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
529 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
532 if (s->unrestricted_mv) {
535 c->xmax = - x + s->width;
536 c->ymax = - y + s->height;
537 } else if (s->out_format == FMT_H261){
538 // Search range of H.261 is different from other codec standards
539 c->xmin = (x > 15) ? - 15 : 0;
540 c->ymin = (y > 15) ? - 15 : 0;
541 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
542 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
546 c->xmax = - x + s->mb_width *16 - 16;
547 c->ymax = - y + s->mb_height*16 - 16;
549 if(!range || range > max_range)
552 c->xmin = FFMAX(c->xmin,-range);
553 c->xmax = FFMIN(c->xmax, range);
554 c->ymin = FFMAX(c->ymin,-range);
555 c->ymax = FFMIN(c->ymax, range);
559 static inline void init_mv4_ref(MotionEstContext *c){
560 const int stride= c->stride;
562 c->ref[1][0] = c->ref[0][0] + 8;
563 c->ref[2][0] = c->ref[0][0] + 8*stride;
564 c->ref[3][0] = c->ref[2][0] + 8;
565 c->src[1][0] = c->src[0][0] + 8;
566 c->src[2][0] = c->src[0][0] + 8*stride;
567 c->src[3][0] = c->src[2][0] + 8;
570 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
572 MotionEstContext * const c= &s->me;
577 int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
579 const int stride= c->stride;
580 const uint8_t *mv_penalty = c->current_mv_penalty;
581 int safety_clipping= s->unrestricted_mv && (s->width&15) && (s->height&15);
585 for(block=0; block<4; block++){
587 int pred_x4, pred_y4;
589 static const int off[4]= {2, 1, 1, -1};
590 const int mot_stride = s->b8_stride;
591 const int mot_xy = s->block_index[block];
594 c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
595 c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
598 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
599 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
601 if (P_LEFT[0] > c->xmax * (1 << shift)) P_LEFT[0] = c->xmax * (1 << shift);
603 /* special case for first line */
604 if (s->first_slice_line && block<2) {
605 c->pred_x= pred_x4= P_LEFT[0];
606 c->pred_y= pred_y4= P_LEFT[1];
608 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
609 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
610 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
611 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
612 if (P_TOP[1] > c->ymax * (1 << shift)) P_TOP[1] = c->ymax * (1 << shift);
613 if (P_TOPRIGHT[0] < c->xmin * (1 << shift)) P_TOPRIGHT[0] = c->xmin * (1 << shift);
614 if (P_TOPRIGHT[0] > c->xmax * (1 << shift)) P_TOPRIGHT[0] = c->xmax * (1 << shift);
615 if (P_TOPRIGHT[1] > c->ymax * (1 << shift)) P_TOPRIGHT[1] = c->ymax * (1 << shift);
617 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
618 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
620 c->pred_x= pred_x4 = P_MEDIAN[0];
621 c->pred_y= pred_y4 = P_MEDIAN[1];
627 if (s->first_slice_line && block<2 && i>1 && i<9)
631 if (P[i][0] > c->xmax * (1 << shift)) P[i][0] = c->xmax * (1 << shift);
632 if (P[i][1] > c->ymax * (1 << shift)) P[i][1] = c->ymax * (1 <<shift );
635 dmin4 = epzs_motion_search2(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift, 1);
637 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
639 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
641 const int offset= ((block&1) + (block>>1)*stride)*8;
642 uint8_t *dest_y = c->scratchpad + offset;
643 if(s->quarter_sample){
644 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
645 dxy = ((my4 & 3) << 2) | (mx4 & 3);
648 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
650 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
652 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
653 dxy = ((my4 & 1) << 1) | (mx4 & 1);
656 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
658 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
660 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
664 if(s->quarter_sample){
672 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
673 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
675 if(mx4 != mx || my4 != my) same=0;
681 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
682 dmin_sum += s->mecc.mb_cmp[0](s,
683 s->new_picture->data[0] +
684 s->mb_x * 16 + s->mb_y * 16 * stride,
685 c->scratchpad, stride, 16);
688 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
693 mx= ff_h263_round_chroma(mx4_sum);
694 my= ff_h263_round_chroma(my4_sum);
695 dxy = ((my & 1) << 1) | (mx & 1);
697 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
700 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
701 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
703 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
704 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
707 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture->data[1] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad, s->uvlinesize, 8);
708 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture->data[2] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad + 8, s->uvlinesize, 8);
714 switch(c->avctx->mb_cmp&0xFF){
716 return dmin_sum+ 32*s->qscale*s->qscale;*/
720 return dmin_sum+ 11*c->mb_penalty_factor;
724 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
725 MotionEstContext * const c= &s->me;
727 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
728 c->src[1][0] = c->src[0][0] + s->linesize;
729 if(c->flags & FLAG_CHROMA){
730 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
731 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
732 c->src[1][1] = c->src[0][1] + s->uvlinesize;
733 c->src[1][2] = c->src[0][2] + s->uvlinesize;
737 static int interlaced_search(MpegEncContext *s, int ref_index,
738 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
740 MotionEstContext * const c= &s->me;
745 const uint8_t * const mv_penalty = c->current_mv_penalty;
747 const int stride= 2*s->linesize;
749 const int mot_stride= s->mb_stride;
750 const int xy= s->mb_x + s->mb_y*mot_stride;
756 init_interlaced_ref(s, ref_index);
758 for(block=0; block<2; block++){
760 int best_dmin= INT_MAX;
763 for(field_select=0; field_select<2; field_select++){
764 int dmin, mx_i, my_i;
765 int16_t (*mv_table)[2]= mv_tables[block][field_select];
767 if(user_field_select){
768 av_assert1(field_select==0 || field_select==1);
769 av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
770 if(field_select_tables[block][xy] != field_select)
774 P_LEFT[0] = mv_table[xy - 1][0];
775 P_LEFT[1] = mv_table[xy - 1][1];
776 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
778 c->pred_x= P_LEFT[0];
779 c->pred_y= P_LEFT[1];
781 if(!s->first_slice_line){
782 P_TOP[0] = mv_table[xy - mot_stride][0];
783 P_TOP[1] = mv_table[xy - mot_stride][1];
784 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
785 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
786 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
787 if (P_TOPRIGHT[0] < c->xmin * (1 << 1)) P_TOPRIGHT[0] = c->xmin * (1 << 1);
788 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
789 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
791 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
792 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
794 P_MV1[0]= mx; //FIXME not correct if block != field_select
797 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1, 0);
799 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
801 mv_table[xy][0]= mx_i;
802 mv_table[xy][1]= my_i;
804 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
808 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
809 dxy = ((my_i & 1) << 1) | (mx_i & 1);
812 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
814 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
816 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
817 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
819 dmin+= c->mb_penalty_factor; //field_select bits
821 dmin += field_select != block; //slightly prefer same field
823 if(dmin < best_dmin){
825 best_field= field_select;
829 int16_t (*mv_table)[2]= mv_tables[block][best_field];
831 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
832 if(mv_table[xy][1]&1) same=0;
833 if(mv_table[xy][1]*2 != my) same=0;
834 if(best_field != block) same=0;
837 field_select_tables[block][xy]= best_field;
838 dmin_sum += best_dmin;
849 switch(c->avctx->mb_cmp&0xFF){
851 return dmin_sum+ 32*s->qscale*s->qscale;*/
855 return dmin_sum+ 11*c->mb_penalty_factor;
859 static inline int get_penalty_factor(int lambda, int lambda2, int type){
863 return lambda>>FF_LAMBDA_SHIFT;
865 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
867 return (4*lambda)>>(FF_LAMBDA_SHIFT);
869 return (2*lambda)>>(FF_LAMBDA_SHIFT);
872 return (2*lambda)>>FF_LAMBDA_SHIFT;
877 return lambda2>>FF_LAMBDA_SHIFT;
879 case FF_CMP_MEDIAN_SAD:
884 void ff_estimate_p_frame_motion(MpegEncContext * s,
887 MotionEstContext * const c= &s->me;
889 int sum, mx = 0, my = 0, dmin = 0;
890 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
891 int vard; ///< sum of squared differences with the estimated motion vector
893 const int shift= 1+s->quarter_sample;
895 Picture * const pic= &s->current_picture;
897 init_ref(c, s->new_picture->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
899 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
900 av_assert0(s->linesize == c->stride);
901 av_assert0(s->uvlinesize == c->uvstride);
903 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
904 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
905 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
906 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
908 get_limits(s, 16*mb_x, 16*mb_y);
911 /* intra / predictive decision */
913 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
914 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
915 (((unsigned) sum * sum) >> 8) + 500;
917 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
918 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
919 c->mb_var_sum_temp += (varc+128)>>8;
921 if (s->motion_est != FF_ME_ZERO) {
922 const int mot_stride = s->b8_stride;
923 const int mot_xy = s->block_index[0];
925 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
926 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
928 if (P_LEFT[0] > (c->xmax << shift))
929 P_LEFT[0] = c->xmax << shift;
931 if (!s->first_slice_line) {
932 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
933 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
934 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
935 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
936 if (P_TOP[1] > (c->ymax << shift))
937 P_TOP[1] = c->ymax << shift;
938 if (P_TOPRIGHT[0] < (c->xmin * (1 << shift)))
939 P_TOPRIGHT[0] = c->xmin * (1 << shift);
940 if (P_TOPRIGHT[1] > (c->ymax * (1 << shift)))
941 P_TOPRIGHT[1] = c->ymax * (1 << shift);
943 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
944 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
946 if (s->out_format == FMT_H263) {
947 c->pred_x = P_MEDIAN[0];
948 c->pred_y = P_MEDIAN[1];
949 } else { /* MPEG-1 at least */
950 c->pred_x = P_LEFT[0];
951 c->pred_y = P_LEFT[1];
954 c->pred_x = P_LEFT[0];
955 c->pred_y = P_LEFT[1];
957 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
960 /* At this point (mx,my) are full-pell and the relative displacement */
961 ppix = c->ref[0][0] + (my * s->linesize) + mx;
963 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
965 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
966 c->mc_mb_var_sum_temp += (vard+128)>>8;
968 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
969 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
970 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
971 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
973 if (vard*2 + 200*256 > varc && !s->intra_penalty)
974 mb_type|= CANDIDATE_MB_TYPE_INTRA;
975 if (varc*2 + 200*256 > vard || s->qscale > 24){
976 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
977 mb_type|= CANDIDATE_MB_TYPE_INTER;
978 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
979 if (s->mpv_flags & FF_MPV_FLAG_MV0)
981 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
986 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
987 && !c->skip && varc>50<<8 && vard>10<<8){
988 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
989 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
991 set_p_mv_tables(s, mx, my, 0);
993 set_p_mv_tables(s, mx, my, 1);
994 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
995 && !c->skip){ //FIXME varc/d checks
996 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
997 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1001 mb_type= CANDIDATE_MB_TYPE_INTER;
1003 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1004 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1005 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1007 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
1008 && !c->skip && varc>50<<8 && vard>10<<8){
1009 int dmin4= h263_mv4_search(s, mx, my, shift);
1011 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1015 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
1016 && !c->skip){ //FIXME varc/d checks
1017 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1019 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1024 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1026 /* get intra luma score */
1027 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1028 intra_score= varc - 500;
1030 unsigned mean = (sum+128)>>8;
1033 for(i=0; i<16; i++){
1034 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1035 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1036 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1037 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1040 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1042 intra_score += c->mb_penalty_factor*16 + s->intra_penalty;
1044 if(intra_score < dmin){
1045 mb_type= CANDIDATE_MB_TYPE_INTRA;
1046 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1048 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1051 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1052 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1053 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1057 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1060 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1063 MotionEstContext * const c= &s->me;
1066 const int shift= 1+s->quarter_sample;
1067 const int xy= mb_x + mb_y*s->mb_stride;
1068 init_ref(c, s->new_picture->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1070 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1072 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1073 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
1075 get_limits(s, 16*mb_x, 16*mb_y);
1078 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1079 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1081 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1083 /* special case for first line */
1084 if (s->first_slice_line) {
1085 c->pred_x= P_LEFT[0];
1086 c->pred_y= P_LEFT[1];
1087 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1088 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1090 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1091 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1092 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1093 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1094 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1095 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1096 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1098 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1099 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1101 c->pred_x = P_MEDIAN[0];
1102 c->pred_y = P_MEDIAN[1];
1105 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1107 s->p_mv_table[xy][0] = mx<<shift;
1108 s->p_mv_table[xy][1] = my<<shift;
1113 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1114 int16_t (*mv_table)[2], int ref_index, int f_code)
1116 MotionEstContext * const c= &s->me;
1117 int mx = 0, my = 0, dmin = 0;
1119 const int shift= 1+s->quarter_sample;
1120 const int mot_stride = s->mb_stride;
1121 const int mot_xy = mb_y*mot_stride + mb_x;
1122 const uint8_t * const mv_penalty = c->mv_penalty[f_code] + MAX_DMV;
1125 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1126 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1127 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1128 c->current_mv_penalty= mv_penalty;
1130 get_limits(s, 16*mb_x, 16*mb_y);
1132 if (s->motion_est != FF_ME_ZERO) {
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 * (1 << shift)) P_TOPRIGHT[0] = c->xmin * (1 << 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) * (1 << 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);
1163 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1165 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1166 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1168 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1169 mv_table[mot_xy][0]= mx;
1170 mv_table[mot_xy][1]= my;
1175 static inline int check_bidir_mv(MpegEncContext * s,
1176 int motion_fx, int motion_fy,
1177 int motion_bx, int motion_by,
1178 int pred_fx, int pred_fy,
1179 int pred_bx, int pred_by,
1183 //FIXME better f_code prediction (max mv & distance)
1185 MotionEstContext * const c= &s->me;
1186 const uint8_t * const mv_penalty_f = c->mv_penalty[s->f_code] + MAX_DMV; // f_code of the prev frame
1187 const uint8_t * const mv_penalty_b = c->mv_penalty[s->b_code] + MAX_DMV; // f_code of the prev frame
1188 int stride= c->stride;
1189 uint8_t *dest_y = c->scratchpad;
1194 uint8_t **src_data= c->src[0];
1195 uint8_t **ref_data= c->ref[0];
1196 uint8_t **ref2_data= c->ref[2];
1198 if(s->quarter_sample){
1199 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1200 src_x = motion_fx >> 2;
1201 src_y = motion_fy >> 2;
1203 ptr = ref_data[0] + (src_y * stride) + src_x;
1204 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1206 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1207 src_x = motion_bx >> 2;
1208 src_y = motion_by >> 2;
1210 ptr = ref2_data[0] + (src_y * stride) + src_x;
1211 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1213 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1214 src_x = motion_fx >> 1;
1215 src_y = motion_fy >> 1;
1217 ptr = ref_data[0] + (src_y * stride) + src_x;
1218 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1220 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1221 src_x = motion_bx >> 1;
1222 src_y = motion_by >> 1;
1224 ptr = ref2_data[0] + (src_y * stride) + src_x;
1225 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1228 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1229 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1230 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1232 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1239 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1240 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1242 MotionEstContext * const c= &s->me;
1243 const int mot_stride = s->mb_stride;
1244 const int xy = mb_y *mot_stride + mb_x;
1246 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1247 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1248 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1249 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1250 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1251 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1252 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1253 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1254 const int flags= c->sub_flags;
1255 const int qpel= flags&FLAG_QPEL;
1256 const int shift= 1+qpel;
1257 const int xmin= c->xmin * (1 << shift);
1258 const int ymin= c->ymin * (1 << shift);
1259 const int xmax= c->xmax<<shift;
1260 const int ymax= c->ymax<<shift;
1261 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1262 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1263 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1264 uint8_t map[256] = { 0 };
1266 map[hashidx&255] = 1;
1268 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1269 motion_bx, motion_by,
1274 if(s->avctx->bidir_refine){
1276 static const uint8_t limittab[5]={0,8,32,64,80};
1277 const int limit= limittab[s->avctx->bidir_refine];
1278 static const int8_t vect[][4]={
1279 { 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},
1281 { 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},
1282 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 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},
1286 { 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},
1287 { 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},
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},
1291 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1292 { 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},
1293 { 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 static const uint8_t hash[]={
1296 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),
1298 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),
1299 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 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),
1303 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),
1304 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),
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),
1308 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1309 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),
1310 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),
1313 #define CHECK_BIDIR(fx,fy,bx,by)\
1314 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1315 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1316 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1318 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1319 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);\
1321 hashidx += HASH(fx,fy,bx,by);\
1330 #define CHECK_BIDIR2(a,b,c,d)\
1331 CHECK_BIDIR(a,b,c,d)\
1332 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1339 CHECK_BIDIR2(0,0,0,1)
1340 CHECK_BIDIR2(0,0,1,0)
1341 CHECK_BIDIR2(0,1,0,0)
1342 CHECK_BIDIR2(1,0,0,0)
1344 for(i=8; i<limit; i++){
1345 int fx= motion_fx+vect[i][0];
1346 int fy= motion_fy+vect[i][1];
1347 int bx= motion_bx+vect[i][2];
1348 int by= motion_by+vect[i][3];
1350 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1351 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1353 map[(hashidx+hash[i])&255] = 1;
1355 if(!map[(hashidx+hash[i])&255]){
1357 map[(hashidx+hash[i])&255] = 1;
1358 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1369 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1370 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1371 borderdist= FFMIN(a,b);
1379 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1380 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1381 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1382 s->b_bidir_back_mv_table[xy][1]= motion_by;
1387 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1389 MotionEstContext * const c= &s->me;
1391 const int mot_stride = s->mb_stride;
1392 const int mot_xy = mb_y*mot_stride + mb_x;
1393 const int shift= 1+s->quarter_sample;
1395 const int time_pp= s->pp_time;
1396 const int time_pb= s->pb_time;
1397 int mx, my, xmin, xmax, ymin, ymax;
1398 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1400 c->current_mv_penalty= c->mv_penalty[1] + MAX_DMV;
1401 ymin= xmin=(-32)>>shift;
1402 ymax= xmax= 31>>shift;
1404 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1405 s->mv_type= MV_TYPE_8X8;
1407 s->mv_type= MV_TYPE_16X16;
1411 int index= s->block_index[i];
1414 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1415 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1416 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1417 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1418 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1419 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1421 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1422 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1423 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1425 xmax= FFMIN(xmax, s->width - max);
1426 xmin= FFMAX(xmin, - 16 - min);
1428 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1429 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1430 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1432 ymax= FFMIN(ymax, s->height - max);
1433 ymin= FFMAX(ymin, - 16 - min);
1435 if(s->mv_type == MV_TYPE_16X16) break;
1438 av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1440 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1441 s->b_direct_mv_table[mot_xy][0]= 0;
1442 s->b_direct_mv_table[mot_xy][1]= 0;
1444 return 256*256*256*64;
1451 c->flags |= FLAG_DIRECT;
1452 c->sub_flags |= FLAG_DIRECT;
1456 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin * (1 << shift), xmax << shift);
1457 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin * (1 << shift), ymax << shift);
1459 /* special case for first line */
1460 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1461 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin * (1 << shift), xmax << shift);
1462 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin * (1 << shift), ymax << shift);
1463 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1][0], xmin * (1 << shift), xmax << shift);
1464 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1][1], ymin * (1 << shift), ymax << shift);
1466 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1467 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1470 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1471 if(c->sub_flags&FLAG_QPEL)
1472 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1474 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1476 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1477 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1479 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1481 mv_table[mot_xy][0]= mx;
1482 mv_table[mot_xy][1]= my;
1483 c->flags &= ~FLAG_DIRECT;
1484 c->sub_flags &= ~FLAG_DIRECT;
1489 void ff_estimate_b_frame_motion(MpegEncContext * s,
1492 MotionEstContext * const c= &s->me;
1493 const int penalty_factor= c->mb_penalty_factor;
1494 int fmin, bmin, dmin, fbmin, bimin, fimin;
1496 const int xy = mb_y*s->mb_stride + mb_x;
1497 init_ref(c, s->new_picture->data, s->last_picture.f->data,
1498 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1500 get_limits(s, 16*mb_x, 16*mb_y);
1504 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1505 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1507 score= ((unsigned)(score*score + 128*256))>>16;
1508 c->mc_mb_var_sum_temp += score;
1509 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1510 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1515 if (s->codec_id == AV_CODEC_ID_MPEG4)
1516 dmin= direct_search(s, mb_x, mb_y);
1519 // FIXME penalty stuff for non-MPEG-4
1521 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1525 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1527 ff_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 ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1533 if (s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME) {
1534 //FIXME mb type penalty
1536 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
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_DMV;
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 &&
1589 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1590 type |= CANDIDATE_MB_TYPE_DIRECT0;
1593 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1596 /* find best f_code for ME which do unlimited searches */
1597 int ff_get_best_fcode(MpegEncContext * s, const int16_t (*mv_table)[2], int type)
1599 if (s->motion_est != FF_ME_ZERO) {
1601 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1602 const uint8_t * fcode_tab = s->fcode_tab;
1604 int best_score=-10000000;
1606 if(s->msmpeg4_version)
1607 range= FFMIN(range, 16);
1608 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1609 range= FFMIN(range, 256);
1611 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1613 for(y=0; y<s->mb_height; y++){
1615 int xy= y*s->mb_stride;
1616 for(x=0; x<s->mb_width; x++, xy++){
1617 if(s->mb_type[xy] & type){
1618 int mx= mv_table[xy][0];
1619 int my= mv_table[xy][1];
1620 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1621 fcode_tab[my + MAX_MV]);
1624 if (mx >= range || mx < -range ||
1625 my >= range || my < -range)
1628 for(j=0; j<fcode && j<8; j++){
1629 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, int type)
1651 MotionEstContext * const c= &s->me;
1652 const int f_code= s->f_code;
1654 av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1656 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1658 av_assert0(range <= 16 || !s->msmpeg4_version);
1659 av_assert0(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->avctx->flags & AV_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] |= type;
1684 s->current_picture.mb_type[i] = type;
1696 * @param truncate 1 for truncation, 0 for using intra
1698 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1699 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1701 MotionEstContext * const c= &s->me;
1702 int y, h_range, v_range;
1704 // RAL: 8 in MPEG-1, 16 in MPEG-4
1705 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1707 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1710 v_range= field_select_table ? range>>1 : range;
1712 /* clip / convert to intra 16x16 type MVs */
1713 for(y=0; y<s->mb_height; y++){
1715 int xy= y*s->mb_stride;
1716 for(x=0; x<s->mb_width; x++){
1717 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1718 if (!field_select_table || field_select_table[xy] == field_select) {
1719 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1720 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1723 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1724 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1725 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1726 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1728 s->mb_type[xy] &= ~type;
1729 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;