3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
8 * This file is part of Libav.
10 * Libav is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
15 * Libav is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with Libav; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
33 #include "libavutil/intmath.h"
37 #include "mpegvideo.h"
44 #define P_TOPRIGHT P[3]
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 assert(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 assert((fx>>1) + 16*s->mb_x >= -16);
159 assert((fy>>1) + 16*s->mb_y >= -16);
160 assert((fx>>1) + 16*s->mb_x <= s->width);
161 assert((fy>>1) + 16*s->mb_y <= s->height);
162 assert((bx>>1) + 16*s->mb_x >= -16);
163 assert((by>>1) + 16*s->mb_y >= -16);
164 assert((bx>>1) + 16*s->mb_x <= s->width);
165 assert((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 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
199 uvdxy= (cx&1) + 2*(cy&1);
200 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
203 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
205 uvdxy= dxy | (x&1) | (2*(y&1));
207 d = cmp_func(s, c->temp, src[0], stride, h);
209 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
211 uvdxy= (x&1) + 2*(y&1);
214 uint8_t * const uvtemp= c->temp + 16*stride;
215 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
216 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
217 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
218 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
223 static int cmp_simple(MpegEncContext *s, const int x, const int y,
224 int ref_index, int src_index,
225 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
226 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
229 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
230 const int size, const int h, int ref_index, int src_index,
231 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
232 if(flags&FLAG_DIRECT){
233 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
235 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
239 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
240 const int size, const int h, int ref_index, int src_index,
241 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
242 if(flags&FLAG_DIRECT){
243 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
245 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);
249 /** @brief compares a block (either a full macroblock or a partition thereof)
250 against a proposed motion-compensated prediction of that block
252 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
253 const int size, const int h, int ref_index, int src_index,
254 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
255 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
256 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
257 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
258 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
259 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
260 && subx==0 && suby==0){
261 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
263 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
267 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
268 const int size, const int h, int ref_index, int src_index,
269 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
270 if(flags&FLAG_DIRECT){
271 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
273 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
277 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
278 const int size, const int h, int ref_index, int src_index,
279 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
280 if(flags&FLAG_DIRECT){
281 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
283 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
287 #include "motion_est_template.c"
289 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
293 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
296 int ff_init_me(MpegEncContext *s){
297 MotionEstContext * const c= &s->me;
298 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
299 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
301 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
302 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
305 //special case of snow is needed because snow uses its own iterative ME code
306 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){
307 av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
313 if(cache_size < 2*dia_size && !c->stride){
314 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
317 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
318 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
319 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
320 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
322 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
323 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
324 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
326 /*FIXME s->no_rounding b_type*/
327 if(s->flags&CODEC_FLAG_QPEL){
328 c->sub_motion_search= qpel_motion_search;
329 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
330 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
331 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
333 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
334 c->sub_motion_search= hpel_motion_search;
335 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
336 && c->avctx-> me_cmp == FF_CMP_SAD
337 && c->avctx-> mb_cmp == FF_CMP_SAD)
338 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
340 c->sub_motion_search= hpel_motion_search;
342 c->hpel_avg= s->dsp.avg_pixels_tab;
343 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
344 else c->hpel_put= s->dsp.put_pixels_tab;
347 c->stride = s->linesize;
348 c->uvstride= s->uvlinesize;
350 c->stride = 16*s->mb_width + 32;
351 c->uvstride= 8*s->mb_width + 16;
354 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
355 * not have yet, and even if we had, the motion estimation code
356 * does not expect it. */
357 if(s->codec_id != AV_CODEC_ID_SNOW){
358 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
359 s->dsp.me_cmp[2]= zero_cmp;
361 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
362 s->dsp.me_sub_cmp[2]= zero_cmp;
364 c->hpel_put[2][0]= c->hpel_put[2][1]=
365 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
368 if(s->codec_id == AV_CODEC_ID_H261){
369 c->sub_motion_search= no_sub_motion_search;
375 #define CHECK_SAD_HALF_MV(suffix, x, y) \
377 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
378 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
379 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
382 static int sad_hpel_motion_search(MpegEncContext * s,
383 int *mx_ptr, int *my_ptr, int dmin,
384 int src_index, int ref_index,
387 MotionEstContext * const c= &s->me;
388 const int penalty_factor= c->sub_penalty_factor;
391 int stride= c->stride;
392 const int flags= c->sub_flags;
405 pix = c->src[src_index][0];
409 ptr = c->ref[ref_index][0] + (my * stride) + mx;
413 if (mx > xmin && mx < xmax &&
414 my > ymin && my < ymax) {
417 const int index= (my<<ME_MAP_SHIFT) + mx;
418 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
419 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
420 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
421 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
431 CHECK_SAD_HALF_MV(y2 , 0, -1)
433 CHECK_SAD_HALF_MV(xy2, -1, -1)
435 CHECK_SAD_HALF_MV(xy2, +1, -1)
439 CHECK_SAD_HALF_MV(xy2, -1, +1)
441 CHECK_SAD_HALF_MV(x2 , -1, 0)
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)
456 CHECK_SAD_HALF_MV(xy2, -1, -1)
460 CHECK_SAD_HALF_MV(xy2, +1, +1)
462 CHECK_SAD_HALF_MV(x2 , -1, 0)
463 CHECK_SAD_HALF_MV(xy2, -1, +1)
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)
475 CHECK_SAD_HALF_MV(y2 , 0, +1)
490 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
492 const int xy= s->mb_x + s->mb_y*s->mb_stride;
494 s->p_mv_table[xy][0] = mx;
495 s->p_mv_table[xy][1] = my;
497 /* has already been set to the 4 MV if 4MV is done */
499 int mot_xy= s->block_index[0];
501 s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
502 s->current_picture.f.motion_val[0][mot_xy ][1] = my;
503 s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
504 s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
506 mot_xy += s->b8_stride;
507 s->current_picture.f.motion_val[0][mot_xy ][0] = mx;
508 s->current_picture.f.motion_val[0][mot_xy ][1] = my;
509 s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
510 s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
515 * get fullpel ME search limits.
517 static inline void get_limits(MpegEncContext *s, int x, int y)
519 MotionEstContext * const c= &s->me;
520 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
522 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
525 if (s->unrestricted_mv) {
528 c->xmax = - x + s->mb_width *16;
529 c->ymax = - y + s->mb_height*16;
530 } else if (s->out_format == FMT_H261){
531 // Search range of H261 is different from other codec standards
532 c->xmin = (x > 15) ? - 15 : 0;
533 c->ymin = (y > 15) ? - 15 : 0;
534 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
535 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
539 c->xmax = - x + s->mb_width *16 - 16;
540 c->ymax = - y + s->mb_height*16 - 16;
543 c->xmin = FFMAX(c->xmin,-range);
544 c->xmax = FFMIN(c->xmax, range);
545 c->ymin = FFMAX(c->ymin,-range);
546 c->ymax = FFMIN(c->ymax, range);
550 static inline void init_mv4_ref(MotionEstContext *c){
551 const int stride= c->stride;
553 c->ref[1][0] = c->ref[0][0] + 8;
554 c->ref[2][0] = c->ref[0][0] + 8*stride;
555 c->ref[3][0] = c->ref[2][0] + 8;
556 c->src[1][0] = c->src[0][0] + 8;
557 c->src[2][0] = c->src[0][0] + 8*stride;
558 c->src[3][0] = c->src[2][0] + 8;
561 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
563 MotionEstContext * const c= &s->me;
568 int dmin_sum=0, mx4_sum=0, my4_sum=0;
570 const int stride= c->stride;
571 uint8_t *mv_penalty= c->current_mv_penalty;
575 for(block=0; block<4; block++){
577 int pred_x4, pred_y4;
579 static const int off[4]= {2, 1, 1, -1};
580 const int mot_stride = s->b8_stride;
581 const int mot_xy = s->block_index[block];
583 P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
584 P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
586 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
588 /* special case for first line */
589 if (s->first_slice_line && block<2) {
590 c->pred_x= pred_x4= P_LEFT[0];
591 c->pred_y= pred_y4= P_LEFT[1];
593 P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
594 P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
595 P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][0];
596 P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][1];
597 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
598 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
599 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
600 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
602 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
603 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
605 c->pred_x= pred_x4 = P_MEDIAN[0];
606 c->pred_y= pred_y4 = P_MEDIAN[1];
611 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
613 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
615 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
617 const int offset= ((block&1) + (block>>1)*stride)*8;
618 uint8_t *dest_y = c->scratchpad + offset;
619 if(s->quarter_sample){
620 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
621 dxy = ((my4 & 3) << 2) | (mx4 & 3);
624 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
626 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
628 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
629 dxy = ((my4 & 1) << 1) | (mx4 & 1);
632 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
634 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
636 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
640 if(s->quarter_sample){
648 s->current_picture.f.motion_val[0][s->block_index[block]][0] = mx4;
649 s->current_picture.f.motion_val[0][s->block_index[block]][1] = my4;
651 if(mx4 != mx || my4 != my) same=0;
657 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
658 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
661 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
666 mx= ff_h263_round_chroma(mx4_sum);
667 my= ff_h263_round_chroma(my4_sum);
668 dxy = ((my & 1) << 1) | (mx & 1);
670 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
673 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
674 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
676 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
677 s->dsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
680 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.f.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
681 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.f.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
687 switch(c->avctx->mb_cmp&0xFF){
689 return dmin_sum+ 32*s->qscale*s->qscale;*/
693 return dmin_sum+ 11*c->mb_penalty_factor;
697 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
698 MotionEstContext * const c= &s->me;
700 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
701 c->src[1][0] = c->src[0][0] + s->linesize;
702 if(c->flags & FLAG_CHROMA){
703 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
704 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
705 c->src[1][1] = c->src[0][1] + s->uvlinesize;
706 c->src[1][2] = c->src[0][2] + s->uvlinesize;
710 static int interlaced_search(MpegEncContext *s, int ref_index,
711 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
713 MotionEstContext * const c= &s->me;
718 uint8_t * const mv_penalty= c->current_mv_penalty;
720 const int stride= 2*s->linesize;
722 const int mot_stride= s->mb_stride;
723 const int xy= s->mb_x + s->mb_y*mot_stride;
729 init_interlaced_ref(s, ref_index);
731 for(block=0; block<2; block++){
733 int best_dmin= INT_MAX;
736 for(field_select=0; field_select<2; field_select++){
737 int dmin, mx_i, my_i;
738 int16_t (*mv_table)[2]= mv_tables[block][field_select];
740 if(user_field_select){
741 assert(field_select==0 || field_select==1);
742 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
743 if(field_select_tables[block][xy] != field_select)
747 P_LEFT[0] = mv_table[xy - 1][0];
748 P_LEFT[1] = mv_table[xy - 1][1];
749 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
751 c->pred_x= P_LEFT[0];
752 c->pred_y= P_LEFT[1];
754 if(!s->first_slice_line){
755 P_TOP[0] = mv_table[xy - mot_stride][0];
756 P_TOP[1] = mv_table[xy - mot_stride][1];
757 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
758 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
759 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
760 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
761 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
762 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
764 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
765 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
767 P_MV1[0]= mx; //FIXME not correct if block != field_select
770 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
772 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
774 mv_table[xy][0]= mx_i;
775 mv_table[xy][1]= my_i;
777 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
781 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
782 dxy = ((my_i & 1) << 1) | (mx_i & 1);
785 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
787 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
789 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
790 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
792 dmin+= c->mb_penalty_factor; //field_select bits
794 dmin += field_select != block; //slightly prefer same field
796 if(dmin < best_dmin){
798 best_field= field_select;
802 int16_t (*mv_table)[2]= mv_tables[block][best_field];
804 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
805 if(mv_table[xy][1]&1) same=0;
806 if(mv_table[xy][1]*2 != my) same=0;
807 if(best_field != block) same=0;
810 field_select_tables[block][xy]= best_field;
811 dmin_sum += best_dmin;
822 switch(c->avctx->mb_cmp&0xFF){
824 return dmin_sum+ 32*s->qscale*s->qscale;*/
828 return dmin_sum+ 11*c->mb_penalty_factor;
832 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
833 int ymax= s->me.ymax>>interlaced;
834 int ymin= s->me.ymin>>interlaced;
836 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
837 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
838 if(mv[1] < ymin) mv[1] = ymin;
839 if(mv[1] > ymax) mv[1] = ymax;
842 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
843 MotionEstContext * const c= &s->me;
844 Picture *p= s->current_picture_ptr;
845 int mb_xy= mb_x + mb_y*s->mb_stride;
846 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
847 int mb_type= s->current_picture.f.mb_type[mb_xy];
849 int shift= (flags&FLAG_QPEL) + 1;
850 int mask= (1<<shift)-1;
853 me_cmp_func cmpf= s->dsp.sse[0];
854 me_cmp_func chroma_cmpf= s->dsp.sse[1];
856 if(p_type && USES_LIST(mb_type, 1)){
857 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
860 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
863 int xy= s->block_index[i];
864 clip_input_mv(s, p->f.motion_val[0][xy], !!IS_INTERLACED(mb_type));
865 clip_input_mv(s, p->f.motion_val[1][xy], !!IS_INTERLACED(mb_type));
868 if(IS_INTERLACED(mb_type)){
869 int xy2= xy + s->b8_stride;
870 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
874 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
875 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
879 if(USES_LIST(mb_type, 0)){
880 int field_select0= p->f.ref_index[0][4*mb_xy ];
881 int field_select1= p->f.ref_index[0][4*mb_xy+2];
882 assert(field_select0==0 ||field_select0==1);
883 assert(field_select1==0 ||field_select1==1);
884 init_interlaced_ref(s, 0);
887 s->p_field_select_table[0][mb_xy]= field_select0;
888 s->p_field_select_table[1][mb_xy]= field_select1;
889 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
890 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
891 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
893 s->b_field_select_table[0][0][mb_xy]= field_select0;
894 s->b_field_select_table[0][1][mb_xy]= field_select1;
895 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
896 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
897 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
900 x = p->f.motion_val[0][xy ][0];
901 y = p->f.motion_val[0][xy ][1];
902 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
903 x = p->f.motion_val[0][xy2][0];
904 y = p->f.motion_val[0][xy2][1];
905 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
907 if(USES_LIST(mb_type, 1)){
908 int field_select0 = p->f.ref_index[1][4 * mb_xy ];
909 int field_select1 = p->f.ref_index[1][4 * mb_xy + 2];
910 assert(field_select0==0 ||field_select0==1);
911 assert(field_select1==0 ||field_select1==1);
912 init_interlaced_ref(s, 2);
914 s->b_field_select_table[1][0][mb_xy]= field_select0;
915 s->b_field_select_table[1][1][mb_xy]= field_select1;
916 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy ];
917 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy2];
918 if(USES_LIST(mb_type, 0)){
919 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
921 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
924 x = p->f.motion_val[1][xy ][0];
925 y = p->f.motion_val[1][xy ][1];
926 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
927 x = p->f.motion_val[1][xy2][0];
928 y = p->f.motion_val[1][xy2][1];
929 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
934 }else if(IS_8X8(mb_type)){
935 if(!(s->flags & CODEC_FLAG_4MV)){
936 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
940 chroma_cmpf= s->dsp.sse[1];
943 xy= s->block_index[i];
944 x= p->f.motion_val[0][xy][0];
945 y= p->f.motion_val[0][xy][1];
946 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
948 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
950 if(USES_LIST(mb_type, 0)){
952 *(uint32_t*)s->p_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
953 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
954 }else if(USES_LIST(mb_type, 1)){
955 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
956 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
957 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
959 *(uint32_t*)s->b_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
960 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
962 x = p->f.motion_val[0][xy][0];
963 y = p->f.motion_val[0][xy][1];
964 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
965 }else if(USES_LIST(mb_type, 1)){
966 *(uint32_t*)s->b_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
967 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
969 x = p->f.motion_val[1][xy][0];
970 y = p->f.motion_val[1][xy][1];
971 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
973 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
978 void ff_estimate_p_frame_motion(MpegEncContext * s,
981 MotionEstContext * const c= &s->me;
983 int sum, mx, my, dmin;
984 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
985 int vard; ///< sum of squared differences with the estimated motion vector
987 const int shift= 1+s->quarter_sample;
989 Picture * const pic= &s->current_picture;
991 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
993 assert(s->quarter_sample==0 || s->quarter_sample==1);
994 assert(s->linesize == c->stride);
995 assert(s->uvlinesize == c->uvstride);
997 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
998 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
999 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1000 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1002 get_limits(s, 16*mb_x, 16*mb_y);
1005 /* intra / predictive decision */
1007 sum = s->dsp.pix_sum(pix, s->linesize);
1008 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
1010 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1011 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1012 c->mb_var_sum_temp += (varc+128)>>8;
1014 if(c->avctx->me_threshold){
1015 vard= check_input_motion(s, mb_x, mb_y, 1);
1017 if((vard+128)>>8 < c->avctx->me_threshold){
1018 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1019 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1020 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1021 c->mc_mb_var_sum_temp += (vard+128)>>8;
1022 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1025 if((vard+128)>>8 < c->avctx->mb_threshold)
1026 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1029 switch(s->me_method) {
1039 const int mot_stride = s->b8_stride;
1040 const int mot_xy = s->block_index[0];
1042 P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
1043 P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
1045 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1047 if(!s->first_slice_line) {
1048 P_TOP[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][0];
1049 P_TOP[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride ][1];
1050 P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][0];
1051 P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][1];
1052 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1053 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1054 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1056 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1057 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1059 if(s->out_format == FMT_H263){
1060 c->pred_x = P_MEDIAN[0];
1061 c->pred_y = P_MEDIAN[1];
1062 }else { /* mpeg1 at least */
1063 c->pred_x= P_LEFT[0];
1064 c->pred_y= P_LEFT[1];
1067 c->pred_x= P_LEFT[0];
1068 c->pred_y= P_LEFT[1];
1072 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1077 /* At this point (mx,my) are full-pell and the relative displacement */
1078 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1080 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1082 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1083 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1084 c->mc_mb_var_sum_temp += (vard+128)>>8;
1087 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1088 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1089 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1091 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1092 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1093 set_p_mv_tables(s, mx, my, 1);
1098 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1099 h263_mv4_search(s, mx, my, shift);
1101 set_p_mv_tables(s, mx, my, 0);
1103 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1104 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1106 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1107 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1108 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1109 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1111 if (vard*2 + 200*256 > varc)
1112 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1113 if (varc*2 + 200*256 > vard || s->qscale > 24){
1114 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1115 mb_type|= CANDIDATE_MB_TYPE_INTER;
1116 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1117 if(s->flags&CODEC_FLAG_MV0)
1119 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1124 if((s->flags&CODEC_FLAG_4MV)
1125 && !c->skip && varc>50<<8 && vard>10<<8){
1126 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1127 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1129 set_p_mv_tables(s, mx, my, 0);
1131 set_p_mv_tables(s, mx, my, 1);
1132 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1133 && !c->skip){ //FIXME varc/d checks
1134 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1135 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1139 mb_type= CANDIDATE_MB_TYPE_INTER;
1141 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1142 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1143 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1145 if((s->flags&CODEC_FLAG_4MV)
1146 && !c->skip && varc>50<<8 && vard>10<<8){
1147 int dmin4= h263_mv4_search(s, mx, my, shift);
1149 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1153 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1154 && !c->skip){ //FIXME varc/d checks
1155 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1157 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1162 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1163 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1165 /* get intra luma score */
1166 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1167 intra_score= varc - 500;
1169 unsigned mean = (sum+128)>>8;
1172 for(i=0; i<16; i++){
1173 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1174 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1175 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1176 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1179 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1181 intra_score += c->mb_penalty_factor*16;
1183 if(intra_score < dmin){
1184 mb_type= CANDIDATE_MB_TYPE_INTRA;
1185 s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1187 s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1190 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1191 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1192 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1196 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1199 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1202 MotionEstContext * const c= &s->me;
1205 const int shift= 1+s->quarter_sample;
1206 const int xy= mb_x + mb_y*s->mb_stride;
1207 init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
1209 assert(s->quarter_sample==0 || s->quarter_sample==1);
1211 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1212 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1214 get_limits(s, 16*mb_x, 16*mb_y);
1217 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1218 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1220 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1222 /* special case for first line */
1223 if (s->first_slice_line) {
1224 c->pred_x= P_LEFT[0];
1225 c->pred_y= P_LEFT[1];
1226 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1227 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1229 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1230 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1231 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1232 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1233 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1234 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1235 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1237 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1238 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1240 c->pred_x = P_MEDIAN[0];
1241 c->pred_y = P_MEDIAN[1];
1244 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1246 s->p_mv_table[xy][0] = mx<<shift;
1247 s->p_mv_table[xy][1] = my<<shift;
1252 static int ff_estimate_motion_b(MpegEncContext * s,
1253 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1255 MotionEstContext * const c= &s->me;
1258 const int shift= 1+s->quarter_sample;
1259 const int mot_stride = s->mb_stride;
1260 const int mot_xy = mb_y*mot_stride + mb_x;
1261 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1264 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1265 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1266 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1267 c->current_mv_penalty= mv_penalty;
1269 get_limits(s, 16*mb_x, 16*mb_y);
1271 switch(s->me_method) {
1280 P_LEFT[0] = mv_table[mot_xy - 1][0];
1281 P_LEFT[1] = mv_table[mot_xy - 1][1];
1283 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1285 /* special case for first line */
1286 if (!s->first_slice_line) {
1287 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1288 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1289 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1290 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1291 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1292 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1293 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1295 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1296 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1298 c->pred_x = P_LEFT[0];
1299 c->pred_y = P_LEFT[1];
1301 if(mv_table == s->b_forw_mv_table){
1302 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1304 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1307 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1312 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1314 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1315 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1317 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1318 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1319 mv_table[mot_xy][0]= mx;
1320 mv_table[mot_xy][1]= my;
1325 static inline int check_bidir_mv(MpegEncContext * s,
1326 int motion_fx, int motion_fy,
1327 int motion_bx, int motion_by,
1328 int pred_fx, int pred_fy,
1329 int pred_bx, int pred_by,
1333 //FIXME better f_code prediction (max mv & distance)
1335 MotionEstContext * const c= &s->me;
1336 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1337 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1338 int stride= c->stride;
1339 uint8_t *dest_y = c->scratchpad;
1344 uint8_t **src_data= c->src[0];
1345 uint8_t **ref_data= c->ref[0];
1346 uint8_t **ref2_data= c->ref[2];
1348 if(s->quarter_sample){
1349 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1350 src_x = motion_fx >> 2;
1351 src_y = motion_fy >> 2;
1353 ptr = ref_data[0] + (src_y * stride) + src_x;
1354 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1356 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1357 src_x = motion_bx >> 2;
1358 src_y = motion_by >> 2;
1360 ptr = ref2_data[0] + (src_y * stride) + src_x;
1361 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1363 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1364 src_x = motion_fx >> 1;
1365 src_y = motion_fy >> 1;
1367 ptr = ref_data[0] + (src_y * stride) + src_x;
1368 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1370 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1371 src_x = motion_bx >> 1;
1372 src_y = motion_by >> 1;
1374 ptr = ref2_data[0] + (src_y * stride) + src_x;
1375 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1378 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1379 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1380 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1382 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1389 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1390 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1392 MotionEstContext * const c= &s->me;
1393 const int mot_stride = s->mb_stride;
1394 const int xy = mb_y *mot_stride + mb_x;
1396 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1397 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1398 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1399 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1400 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1401 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1402 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1403 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1404 const int flags= c->sub_flags;
1405 const int qpel= flags&FLAG_QPEL;
1406 const int shift= 1+qpel;
1407 const int xmin= c->xmin<<shift;
1408 const int ymin= c->ymin<<shift;
1409 const int xmax= c->xmax<<shift;
1410 const int ymax= c->ymax<<shift;
1411 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1412 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1413 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1414 uint8_t map[256] = { 0 };
1416 map[hashidx&255] = 1;
1418 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1419 motion_bx, motion_by,
1424 if(s->avctx->bidir_refine){
1426 static const uint8_t limittab[5]={0,8,32,64,80};
1427 const int limit= limittab[s->avctx->bidir_refine];
1428 static const int8_t vect[][4]={
1429 { 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},
1431 { 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},
1432 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1433 { 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},
1434 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1436 { 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},
1437 { 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},
1438 { 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},
1439 { 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},
1441 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1442 { 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},
1443 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1445 static const uint8_t hash[]={
1446 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),
1448 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),
1449 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1450 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),
1451 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1453 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),
1454 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),
1455 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),
1456 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),
1458 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1459 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),
1460 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),
1463 #define CHECK_BIDIR(fx,fy,bx,by)\
1464 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1465 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1466 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1468 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1469 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);\
1471 hashidx += HASH(fx,fy,bx,by);\
1480 #define CHECK_BIDIR2(a,b,c,d)\
1481 CHECK_BIDIR(a,b,c,d)\
1482 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1489 CHECK_BIDIR2(0,0,0,1)
1490 CHECK_BIDIR2(0,0,1,0)
1491 CHECK_BIDIR2(0,1,0,0)
1492 CHECK_BIDIR2(1,0,0,0)
1494 for(i=8; i<limit; i++){
1495 int fx= motion_fx+vect[i][0];
1496 int fy= motion_fy+vect[i][1];
1497 int bx= motion_bx+vect[i][2];
1498 int by= motion_by+vect[i][3];
1500 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1501 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1503 map[(hashidx+hash[i])&255] = 1;
1505 if(!map[(hashidx+hash[i])&255]){
1507 map[(hashidx+hash[i])&255] = 1;
1508 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1519 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1520 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1521 borderdist= FFMIN(a,b);
1529 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1530 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1531 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1532 s->b_bidir_back_mv_table[xy][1]= motion_by;
1537 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1539 MotionEstContext * const c= &s->me;
1541 const int mot_stride = s->mb_stride;
1542 const int mot_xy = mb_y*mot_stride + mb_x;
1543 const int shift= 1+s->quarter_sample;
1545 const int time_pp= s->pp_time;
1546 const int time_pb= s->pb_time;
1547 int mx, my, xmin, xmax, ymin, ymax;
1548 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1550 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1551 ymin= xmin=(-32)>>shift;
1552 ymax= xmax= 31>>shift;
1554 if (IS_8X8(s->next_picture.f.mb_type[mot_xy])) {
1555 s->mv_type= MV_TYPE_8X8;
1557 s->mv_type= MV_TYPE_16X16;
1561 int index= s->block_index[i];
1564 c->co_located_mv[i][0] = s->next_picture.f.motion_val[0][index][0];
1565 c->co_located_mv[i][1] = s->next_picture.f.motion_val[0][index][1];
1566 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1567 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1568 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1569 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1571 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1572 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1573 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1575 xmax= FFMIN(xmax, s->width - max);
1576 xmin= FFMAX(xmin, - 16 - min);
1578 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1579 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1580 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1582 ymax= FFMIN(ymax, s->height - max);
1583 ymin= FFMAX(ymin, - 16 - min);
1585 if(s->mv_type == MV_TYPE_16X16) break;
1588 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1590 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1591 s->b_direct_mv_table[mot_xy][0]= 0;
1592 s->b_direct_mv_table[mot_xy][1]= 0;
1594 return 256*256*256*64;
1601 c->flags |= FLAG_DIRECT;
1602 c->sub_flags |= FLAG_DIRECT;
1606 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1607 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1609 /* special case for first line */
1610 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1611 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1612 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1613 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1614 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1616 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1617 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1620 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1621 if(c->sub_flags&FLAG_QPEL)
1622 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1624 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1626 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1627 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1629 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1631 mv_table[mot_xy][0]= mx;
1632 mv_table[mot_xy][1]= my;
1633 c->flags &= ~FLAG_DIRECT;
1634 c->sub_flags &= ~FLAG_DIRECT;
1639 void ff_estimate_b_frame_motion(MpegEncContext * s,
1642 MotionEstContext * const c= &s->me;
1643 const int penalty_factor= c->mb_penalty_factor;
1644 int fmin, bmin, dmin, fbmin, bimin, fimin;
1646 const int xy = mb_y*s->mb_stride + mb_x;
1647 init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
1648 s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
1650 get_limits(s, 16*mb_x, 16*mb_y);
1654 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.f.mbskip_table[xy]) {
1655 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1657 score= ((unsigned)(score*score + 128*256))>>16;
1658 c->mc_mb_var_sum_temp += score;
1659 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1660 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1665 if(c->avctx->me_threshold){
1666 int vard= check_input_motion(s, mb_x, mb_y, 0);
1668 if((vard+128)>>8 < c->avctx->me_threshold){
1669 // pix = c->src[0][0];
1670 // sum = s->dsp.pix_sum(pix, s->linesize);
1671 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1673 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1674 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1675 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1676 c->mb_var_sum_temp += (varc+128)>>8;*/
1677 c->mc_mb_var_sum_temp += (vard+128)>>8;
1678 /* if (vard <= 64<<8 || vard < varc) {
1679 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1681 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1685 if((vard+128)>>8 < c->avctx->mb_threshold){
1686 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1687 if(type == CANDIDATE_MB_TYPE_DIRECT){
1688 direct_search(s, mb_x, mb_y);
1690 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1692 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1694 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1696 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1698 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1700 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1701 interlaced_search(s, 0,
1702 s->b_field_mv_table[0], s->b_field_select_table[0],
1703 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1705 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1707 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1708 interlaced_search(s, 2,
1709 s->b_field_mv_table[1], s->b_field_select_table[1],
1710 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1716 if (s->codec_id == AV_CODEC_ID_MPEG4)
1717 dmin= direct_search(s, mb_x, mb_y);
1720 //FIXME penalty stuff for non mpeg4
1722 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1725 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1726 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1729 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1730 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1732 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1733 //FIXME mb type penalty
1735 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1736 fimin= interlaced_search(s, 0,
1737 s->b_field_mv_table[0], s->b_field_select_table[0],
1738 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1739 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1740 bimin= interlaced_search(s, 2,
1741 s->b_field_mv_table[1], s->b_field_select_table[1],
1742 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1744 fimin= bimin= INT_MAX;
1748 type = CANDIDATE_MB_TYPE_FORWARD;
1752 type = CANDIDATE_MB_TYPE_DIRECT;
1756 type= CANDIDATE_MB_TYPE_BACKWARD;
1760 type= CANDIDATE_MB_TYPE_BIDIR;
1764 type= CANDIDATE_MB_TYPE_FORWARD_I;
1768 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1771 score= ((unsigned)(score*score + 128*256))>>16;
1772 c->mc_mb_var_sum_temp += score;
1773 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1776 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1777 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1779 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1781 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1782 if(fimin < INT_MAX && bimin < INT_MAX){
1783 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1785 //FIXME something smarter
1786 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1787 if(s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1788 type |= CANDIDATE_MB_TYPE_DIRECT0;
1791 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1794 /* find best f_code for ME which do unlimited searches */
1795 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1797 if(s->me_method>=ME_EPZS){
1799 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1800 uint8_t * fcode_tab= s->fcode_tab;
1802 int best_score=-10000000;
1804 if(s->msmpeg4_version)
1805 range= FFMIN(range, 16);
1806 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1807 range= FFMIN(range, 256);
1809 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1811 for(y=0; y<s->mb_height; y++){
1813 int xy= y*s->mb_stride;
1814 for(x=0; x<s->mb_width; x++){
1815 if(s->mb_type[xy] & type){
1816 int mx= mv_table[xy][0];
1817 int my= mv_table[xy][1];
1818 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1819 fcode_tab[my + MAX_MV]);
1822 if(mx >= range || mx < -range ||
1823 my >= range || my < -range)
1826 for(j=0; j<fcode && j<8; j++){
1827 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1836 if(score[i] > best_score){
1837 best_score= score[i];
1840 // printf("%d %d\n", i, score[i]);
1843 // printf("fcode: %d type: %d\n", i, s->pict_type);
1845 /* for(i=0; i<=MAX_FCODE; i++){
1846 printf("%d ", mv_num[i]);
1854 void ff_fix_long_p_mvs(MpegEncContext * s)
1856 MotionEstContext * const c= &s->me;
1857 const int f_code= s->f_code;
1859 assert(s->pict_type==AV_PICTURE_TYPE_P);
1861 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1863 assert(range <= 16 || !s->msmpeg4_version);
1864 assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1866 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1868 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1869 if(s->flags&CODEC_FLAG_4MV){
1870 const int wrap= s->b8_stride;
1872 /* clip / convert to intra 8x8 type MVs */
1873 for(y=0; y<s->mb_height; y++){
1875 int i= y*s->mb_stride;
1878 for(x=0; x<s->mb_width; x++){
1879 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1881 for(block=0; block<4; block++){
1882 int off= (block& 1) + (block>>1)*wrap;
1883 int mx = s->current_picture.f.motion_val[0][ xy + off ][0];
1884 int my = s->current_picture.f.motion_val[0][ xy + off ][1];
1886 if( mx >=range || mx <-range
1887 || my >=range || my <-range){
1888 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1889 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1890 s->current_picture.f.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1903 * @param truncate 1 for truncation, 0 for using intra
1905 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1906 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1908 MotionEstContext * const c= &s->me;
1909 int y, h_range, v_range;
1911 // RAL: 8 in MPEG-1, 16 in MPEG-4
1912 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1914 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1917 v_range= field_select_table ? range>>1 : range;
1919 /* clip / convert to intra 16x16 type MVs */
1920 for(y=0; y<s->mb_height; y++){
1922 int xy= y*s->mb_stride;
1923 for(x=0; x<s->mb_width; x++){
1924 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1925 if(field_select_table==NULL || field_select_table[xy] == field_select){
1926 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1927 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1930 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1931 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1932 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1933 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1935 s->mb_type[xy] &= ~type;
1936 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;