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
35 #include "mpegvideo.h"
40 #define SQ(a) ((a)*(a))
44 #define P_TOPRIGHT P[3]
48 static inline 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 int 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 /*! \brief compares a block (either a full macroblock or a partition thereof)
106 against a proposed motion-compensated prediction of that block
108 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
109 const int size, const int h, int ref_index, int src_index,
110 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
111 MotionEstContext * const c= &s->me;
112 const int stride= c->stride;
113 const int uvstride= c->uvstride;
114 const int qpel= flags&FLAG_QPEL;
115 const int chroma= flags&FLAG_CHROMA;
116 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
117 const int hx= subx + (x<<(1+qpel));
118 const int hy= suby + (y<<(1+qpel));
119 uint8_t * const * const ref= c->ref[ref_index];
120 uint8_t * const * const src= c->src[src_index];
122 //FIXME check chroma 4mv, (no crashes ...)
123 if(flags&FLAG_DIRECT){
124 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
125 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
126 const int time_pp= s->pp_time;
127 const int time_pb= s->pb_time;
128 const int mask= 2*qpel+1;
129 if(s->mv_type==MV_TYPE_8X8){
132 int fx = c->direct_basis_mv[i][0] + hx;
133 int fy = c->direct_basis_mv[i][1] + hy;
134 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));
135 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));
136 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
137 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
139 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
141 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
142 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
144 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
145 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
149 int fx = c->direct_basis_mv[0][0] + hx;
150 int fy = c->direct_basis_mv[0][1] + hy;
151 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
152 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
153 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
154 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
157 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
158 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
159 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
160 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
161 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
162 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
163 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
164 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
166 assert((fx>>1) + 16*s->mb_x >= -16);
167 assert((fy>>1) + 16*s->mb_y >= -16);
168 assert((fx>>1) + 16*s->mb_x <= s->width);
169 assert((fy>>1) + 16*s->mb_y <= s->height);
170 assert((bx>>1) + 16*s->mb_x >= -16);
171 assert((by>>1) + 16*s->mb_y >= -16);
172 assert((bx>>1) + 16*s->mb_x <= s->width);
173 assert((by>>1) + 16*s->mb_y <= s->height);
175 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
176 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
179 d = cmp_func(s, c->temp, src[0], stride, 16);
183 int uvdxy; /* no, it might not be used uninitialized */
186 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
192 uvdxy= (cx&1) + 2*(cy&1);
193 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
196 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
198 uvdxy= dxy | (x&1) | (2*(y&1));
200 d = cmp_func(s, c->temp, src[0], stride, h);
202 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
204 uvdxy= (x&1) + 2*(y&1);
207 uint8_t * const uvtemp= c->temp + 16*stride;
208 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
209 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
211 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
216 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
220 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
225 #include "motion_est_template.c"
227 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
231 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
234 int ff_init_me(MpegEncContext *s){
235 MotionEstContext * const c= &s->me;
236 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
237 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
239 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
240 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
243 //special case of snow is needed because snow uses its own iterative ME code
244 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
245 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");
251 if(cache_size < 2*dia_size && !c->stride){
252 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
255 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
256 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
257 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
258 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
260 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
261 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
262 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
264 /*FIXME s->no_rounding b_type*/
265 if(s->flags&CODEC_FLAG_QPEL){
266 c->sub_motion_search= qpel_motion_search;
267 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
268 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
269 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
271 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
272 c->sub_motion_search= hpel_motion_search;
273 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
274 && c->avctx-> me_cmp == FF_CMP_SAD
275 && c->avctx-> mb_cmp == FF_CMP_SAD)
276 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
278 c->sub_motion_search= hpel_motion_search;
280 c->hpel_avg= s->dsp.avg_pixels_tab;
281 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
282 else c->hpel_put= s->dsp.put_pixels_tab;
285 c->stride = s->linesize;
286 c->uvstride= s->uvlinesize;
288 c->stride = 16*s->mb_width + 32;
289 c->uvstride= 8*s->mb_width + 16;
292 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
293 * not have yet, and even if we had, the motion estimation code
294 * does not expect it. */
295 if(s->codec_id != CODEC_ID_SNOW){
296 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
297 s->dsp.me_cmp[2]= zero_cmp;
299 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
300 s->dsp.me_sub_cmp[2]= zero_cmp;
302 c->hpel_put[2][0]= c->hpel_put[2][1]=
303 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
306 if(s->codec_id == CODEC_ID_H261){
307 c->sub_motion_search= no_sub_motion_search;
314 static int pix_dev(uint8_t * pix, int line_size, int mean)
319 for (i = 0; i < 16; i++) {
320 for (j = 0; j < 16; j += 8) {
321 s += FFABS(pix[0]-mean);
322 s += FFABS(pix[1]-mean);
323 s += FFABS(pix[2]-mean);
324 s += FFABS(pix[3]-mean);
325 s += FFABS(pix[4]-mean);
326 s += FFABS(pix[5]-mean);
327 s += FFABS(pix[6]-mean);
328 s += FFABS(pix[7]-mean);
331 pix += line_size - 16;
337 static inline void no_motion_search(MpegEncContext * s,
338 int *mx_ptr, int *my_ptr)
340 *mx_ptr = 16 * s->mb_x;
341 *my_ptr = 16 * s->mb_y;
344 #define Z_THRESHOLD 256
346 #define CHECK_SAD_HALF_MV(suffix, x, y) \
348 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
349 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
350 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
353 static inline int sad_hpel_motion_search(MpegEncContext * s,
354 int *mx_ptr, int *my_ptr, int dmin,
355 int src_index, int ref_index,
358 MotionEstContext * const c= &s->me;
359 const int penalty_factor= c->sub_penalty_factor;
362 int stride= c->stride;
363 const int flags= c->sub_flags;
376 pix = c->src[src_index][0];
380 ptr = c->ref[ref_index][0] + (my * stride) + mx;
384 if (mx > xmin && mx < xmax &&
385 my > ymin && my < ymax) {
388 const int index= (my<<ME_MAP_SHIFT) + mx;
389 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
390 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
391 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
392 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
402 CHECK_SAD_HALF_MV(y2 , 0, -1)
404 CHECK_SAD_HALF_MV(xy2, -1, -1)
406 CHECK_SAD_HALF_MV(xy2, +1, -1)
410 CHECK_SAD_HALF_MV(xy2, -1, +1)
412 CHECK_SAD_HALF_MV(x2 , -1, 0)
414 CHECK_SAD_HALF_MV(xy2, +1, -1)
416 CHECK_SAD_HALF_MV(xy2, -1, -1)
420 CHECK_SAD_HALF_MV(xy2, +1, +1)
422 CHECK_SAD_HALF_MV(x2 , +1, 0)
427 CHECK_SAD_HALF_MV(xy2, -1, -1)
431 CHECK_SAD_HALF_MV(xy2, +1, +1)
433 CHECK_SAD_HALF_MV(x2 , -1, 0)
434 CHECK_SAD_HALF_MV(xy2, -1, +1)
437 CHECK_SAD_HALF_MV(xy2, +1, -1)
441 CHECK_SAD_HALF_MV(xy2, -1, +1)
443 CHECK_SAD_HALF_MV(x2 , +1, 0)
444 CHECK_SAD_HALF_MV(xy2, +1, +1)
446 CHECK_SAD_HALF_MV(y2 , 0, +1)
461 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
463 const int xy= s->mb_x + s->mb_y*s->mb_stride;
465 s->p_mv_table[xy][0] = mx;
466 s->p_mv_table[xy][1] = my;
468 /* has already been set to the 4 MV if 4MV is done */
470 int mot_xy= s->block_index[0];
472 s->current_picture.motion_val[0][mot_xy ][0]= mx;
473 s->current_picture.motion_val[0][mot_xy ][1]= my;
474 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
475 s->current_picture.motion_val[0][mot_xy+1][1]= my;
477 mot_xy += s->b8_stride;
478 s->current_picture.motion_val[0][mot_xy ][0]= mx;
479 s->current_picture.motion_val[0][mot_xy ][1]= my;
480 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
481 s->current_picture.motion_val[0][mot_xy+1][1]= my;
486 * get fullpel ME search limits.
488 static inline void get_limits(MpegEncContext *s, int x, int y)
490 MotionEstContext * const c= &s->me;
491 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
493 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
496 if (s->unrestricted_mv) {
499 c->xmax = - x + s->mb_width *16;
500 c->ymax = - y + s->mb_height*16;
501 } else if (s->out_format == FMT_H261){
502 // Search range of H261 is different from other codec standards
503 c->xmin = (x > 15) ? - 15 : 0;
504 c->ymin = (y > 15) ? - 15 : 0;
505 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
506 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
510 c->xmax = - x + s->mb_width *16 - 16;
511 c->ymax = - y + s->mb_height*16 - 16;
514 c->xmin = FFMAX(c->xmin,-range);
515 c->xmax = FFMIN(c->xmax, range);
516 c->ymin = FFMAX(c->ymin,-range);
517 c->ymax = FFMIN(c->ymax, range);
521 static inline void init_mv4_ref(MotionEstContext *c){
522 const int stride= c->stride;
524 c->ref[1][0] = c->ref[0][0] + 8;
525 c->ref[2][0] = c->ref[0][0] + 8*stride;
526 c->ref[3][0] = c->ref[2][0] + 8;
527 c->src[1][0] = c->src[0][0] + 8;
528 c->src[2][0] = c->src[0][0] + 8*stride;
529 c->src[3][0] = c->src[2][0] + 8;
532 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
534 MotionEstContext * const c= &s->me;
539 int dmin_sum=0, mx4_sum=0, my4_sum=0;
541 const int stride= c->stride;
542 uint8_t *mv_penalty= c->current_mv_penalty;
546 for(block=0; block<4; block++){
548 int pred_x4, pred_y4;
550 static const int off[4]= {2, 1, 1, -1};
551 const int mot_stride = s->b8_stride;
552 const int mot_xy = s->block_index[block];
554 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
555 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
557 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
559 /* special case for first line */
560 if (s->first_slice_line && block<2) {
561 c->pred_x= pred_x4= P_LEFT[0];
562 c->pred_y= pred_y4= P_LEFT[1];
564 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
565 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
566 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
567 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
568 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
569 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
570 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
571 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
573 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
574 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
576 c->pred_x= pred_x4 = P_MEDIAN[0];
577 c->pred_y= pred_y4 = P_MEDIAN[1];
582 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
584 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
586 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
588 const int offset= ((block&1) + (block>>1)*stride)*8;
589 uint8_t *dest_y = c->scratchpad + offset;
590 if(s->quarter_sample){
591 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
592 dxy = ((my4 & 3) << 2) | (mx4 & 3);
595 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
597 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
599 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
600 dxy = ((my4 & 1) << 1) | (mx4 & 1);
603 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
605 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
607 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
611 if(s->quarter_sample){
619 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
620 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
622 if(mx4 != mx || my4 != my) same=0;
628 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
629 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
632 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
637 mx= ff_h263_round_chroma(mx4_sum);
638 my= ff_h263_round_chroma(my4_sum);
639 dxy = ((my & 1) << 1) | (mx & 1);
641 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
644 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
645 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
647 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
648 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
651 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
652 dmin_sum += s->dsp.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);
658 switch(c->avctx->mb_cmp&0xFF){
660 return dmin_sum+ 32*s->qscale*s->qscale;*/
664 return dmin_sum+ 11*c->mb_penalty_factor;
668 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
669 MotionEstContext * const c= &s->me;
671 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
672 c->src[1][0] = c->src[0][0] + s->linesize;
673 if(c->flags & FLAG_CHROMA){
674 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
675 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
676 c->src[1][1] = c->src[0][1] + s->uvlinesize;
677 c->src[1][2] = c->src[0][2] + s->uvlinesize;
681 static int interlaced_search(MpegEncContext *s, int ref_index,
682 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
684 MotionEstContext * const c= &s->me;
689 uint8_t * const mv_penalty= c->current_mv_penalty;
691 const int stride= 2*s->linesize;
693 const int mot_stride= s->mb_stride;
694 const int xy= s->mb_x + s->mb_y*mot_stride;
700 init_interlaced_ref(s, ref_index);
702 for(block=0; block<2; block++){
704 int best_dmin= INT_MAX;
707 for(field_select=0; field_select<2; field_select++){
708 int dmin, mx_i, my_i;
709 int16_t (*mv_table)[2]= mv_tables[block][field_select];
711 if(user_field_select){
712 assert(field_select==0 || field_select==1);
713 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
714 if(field_select_tables[block][xy] != field_select)
718 P_LEFT[0] = mv_table[xy - 1][0];
719 P_LEFT[1] = mv_table[xy - 1][1];
720 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
722 c->pred_x= P_LEFT[0];
723 c->pred_y= P_LEFT[1];
725 if(!s->first_slice_line){
726 P_TOP[0] = mv_table[xy - mot_stride][0];
727 P_TOP[1] = mv_table[xy - mot_stride][1];
728 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
729 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
730 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
731 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
732 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
733 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
735 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
736 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
738 P_MV1[0]= mx; //FIXME not correct if block != field_select
741 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
743 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
745 mv_table[xy][0]= mx_i;
746 mv_table[xy][1]= my_i;
748 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
752 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
753 dxy = ((my_i & 1) << 1) | (mx_i & 1);
756 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
758 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
760 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
761 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
763 dmin+= c->mb_penalty_factor; //field_select bits
765 dmin += field_select != block; //slightly prefer same field
767 if(dmin < best_dmin){
769 best_field= field_select;
773 int16_t (*mv_table)[2]= mv_tables[block][best_field];
775 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
776 if(mv_table[xy][1]&1) same=0;
777 if(mv_table[xy][1]*2 != my) same=0;
778 if(best_field != block) same=0;
781 field_select_tables[block][xy]= best_field;
782 dmin_sum += best_dmin;
793 switch(c->avctx->mb_cmp&0xFF){
795 return dmin_sum+ 32*s->qscale*s->qscale;*/
799 return dmin_sum+ 11*c->mb_penalty_factor;
803 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
804 int ymax= s->me.ymax>>interlaced;
805 int ymin= s->me.ymin>>interlaced;
807 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
808 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
809 if(mv[1] < ymin) mv[1] = ymin;
810 if(mv[1] > ymax) mv[1] = ymax;
813 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
814 MotionEstContext * const c= &s->me;
815 Picture *p= s->current_picture_ptr;
816 int mb_xy= mb_x + mb_y*s->mb_stride;
817 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
818 int mb_type= s->current_picture.mb_type[mb_xy];
820 int shift= (flags&FLAG_QPEL) + 1;
821 int mask= (1<<shift)-1;
824 me_cmp_func cmpf= s->dsp.sse[0];
825 me_cmp_func chroma_cmpf= s->dsp.sse[1];
827 if(p_type && USES_LIST(mb_type, 1)){
828 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
831 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
834 int xy= s->block_index[i];
835 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
836 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
839 if(IS_INTERLACED(mb_type)){
840 int xy2= xy + s->b8_stride;
841 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
845 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
846 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
850 if(USES_LIST(mb_type, 0)){
851 int field_select0= p->ref_index[0][xy ];
852 int field_select1= p->ref_index[0][xy2];
853 assert(field_select0==0 ||field_select0==1);
854 assert(field_select1==0 ||field_select1==1);
855 init_interlaced_ref(s, 0);
858 s->p_field_select_table[0][mb_xy]= field_select0;
859 s->p_field_select_table[1][mb_xy]= field_select1;
860 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
861 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
862 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
864 s->b_field_select_table[0][0][mb_xy]= field_select0;
865 s->b_field_select_table[0][1][mb_xy]= field_select1;
866 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
867 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
868 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
871 x= p->motion_val[0][xy ][0];
872 y= p->motion_val[0][xy ][1];
873 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
874 x= p->motion_val[0][xy2][0];
875 y= p->motion_val[0][xy2][1];
876 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
878 if(USES_LIST(mb_type, 1)){
879 int field_select0= p->ref_index[1][xy ];
880 int field_select1= p->ref_index[1][xy2];
881 assert(field_select0==0 ||field_select0==1);
882 assert(field_select1==0 ||field_select1==1);
883 init_interlaced_ref(s, 2);
885 s->b_field_select_table[1][0][mb_xy]= field_select0;
886 s->b_field_select_table[1][1][mb_xy]= field_select1;
887 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
888 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
889 if(USES_LIST(mb_type, 0)){
890 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
892 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
895 x= p->motion_val[1][xy ][0];
896 y= p->motion_val[1][xy ][1];
897 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
898 x= p->motion_val[1][xy2][0];
899 y= p->motion_val[1][xy2][1];
900 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
905 }else if(IS_8X8(mb_type)){
906 if(!(s->flags & CODEC_FLAG_4MV)){
907 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
911 chroma_cmpf= s->dsp.sse[1];
914 xy= s->block_index[i];
915 x= p->motion_val[0][xy][0];
916 y= p->motion_val[0][xy][1];
917 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
919 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
921 if(USES_LIST(mb_type, 0)){
923 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
924 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
925 }else if(USES_LIST(mb_type, 1)){
926 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
927 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
928 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
930 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
931 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
933 x= p->motion_val[0][xy][0];
934 y= p->motion_val[0][xy][1];
935 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
936 }else if(USES_LIST(mb_type, 1)){
937 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
938 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
940 x= p->motion_val[1][xy][0];
941 y= p->motion_val[1][xy][1];
942 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
944 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
949 void ff_estimate_p_frame_motion(MpegEncContext * s,
952 MotionEstContext * const c= &s->me;
954 int sum, mx, my, dmin;
955 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
956 int vard; ///< sum of squared differences with the estimated motion vector
958 const int shift= 1+s->quarter_sample;
960 Picture * const pic= &s->current_picture;
962 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
964 assert(s->quarter_sample==0 || s->quarter_sample==1);
965 assert(s->linesize == c->stride);
966 assert(s->uvlinesize == c->uvstride);
968 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
969 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
970 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
971 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
973 get_limits(s, 16*mb_x, 16*mb_y);
976 /* intra / predictive decision */
978 sum = s->dsp.pix_sum(pix, s->linesize);
979 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
981 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
982 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
983 c->mb_var_sum_temp += (varc+128)>>8;
985 if(c->avctx->me_threshold){
986 vard= check_input_motion(s, mb_x, mb_y, 1);
988 if((vard+128)>>8 < c->avctx->me_threshold){
989 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
990 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
991 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
992 c->mc_mb_var_sum_temp += (vard+128)>>8;
993 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
996 if((vard+128)>>8 < c->avctx->mb_threshold)
997 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1000 switch(s->me_method) {
1003 no_motion_search(s, &mx, &my);
1011 const int mot_stride = s->b8_stride;
1012 const int mot_xy = s->block_index[0];
1014 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1015 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1017 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1019 if(!s->first_slice_line) {
1020 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1021 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1022 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1023 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1024 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1025 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1026 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1028 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1029 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1031 if(s->out_format == FMT_H263){
1032 c->pred_x = P_MEDIAN[0];
1033 c->pred_y = P_MEDIAN[1];
1034 }else { /* mpeg1 at least */
1035 c->pred_x= P_LEFT[0];
1036 c->pred_y= P_LEFT[1];
1039 c->pred_x= P_LEFT[0];
1040 c->pred_y= P_LEFT[1];
1044 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1049 /* At this point (mx,my) are full-pell and the relative displacement */
1050 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1052 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1054 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1055 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1056 c->mc_mb_var_sum_temp += (vard+128)>>8;
1059 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1060 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1063 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1064 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1065 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1067 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1068 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1069 set_p_mv_tables(s, mx, my, 1);
1074 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1075 h263_mv4_search(s, mx, my, shift);
1077 set_p_mv_tables(s, mx, my, 0);
1079 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1080 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1082 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1083 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1084 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1085 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1087 if (vard*2 + 200*256 > varc)
1088 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1089 if (varc*2 + 200*256 > vard || s->qscale > 24){
1090 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1091 mb_type|= CANDIDATE_MB_TYPE_INTER;
1092 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1093 if(s->flags&CODEC_FLAG_MV0)
1095 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1100 if((s->flags&CODEC_FLAG_4MV)
1101 && !c->skip && varc>50<<8 && vard>10<<8){
1102 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1103 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1105 set_p_mv_tables(s, mx, my, 0);
1107 set_p_mv_tables(s, mx, my, 1);
1108 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1109 && !c->skip){ //FIXME varc/d checks
1110 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1111 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1115 mb_type= CANDIDATE_MB_TYPE_INTER;
1117 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1118 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1119 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1121 if((s->flags&CODEC_FLAG_4MV)
1122 && !c->skip && varc>50<<8 && vard>10<<8){
1123 int dmin4= h263_mv4_search(s, mx, my, shift);
1125 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1129 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1130 && !c->skip){ //FIXME varc/d checks
1131 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1133 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1138 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1139 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1141 /* get intra luma score */
1142 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1143 intra_score= varc - 500;
1145 int mean= (sum+128)>>8;
1148 for(i=0; i<16; i++){
1149 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1150 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1151 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1152 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1155 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1158 /* get chroma score */
1159 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1164 if(s->out_format == FMT_H263){
1165 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1167 mean= (s->last_dc[i] + 4)>>3;
1169 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1173 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1174 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1177 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1181 intra_score += c->mb_penalty_factor*16;
1183 if(intra_score < dmin){
1184 mb_type= CANDIDATE_MB_TYPE_INTRA;
1185 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1187 s->current_picture.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.data, s->last_picture.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) {
1274 no_motion_search(s, &mx, &my);
1282 P_LEFT[0] = mv_table[mot_xy - 1][0];
1283 P_LEFT[1] = mv_table[mot_xy - 1][1];
1285 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1287 /* special case for first line */
1288 if (!s->first_slice_line) {
1289 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1290 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1291 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1292 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1293 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1294 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1295 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1297 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1298 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1300 c->pred_x= P_LEFT[0];
1301 c->pred_y= P_LEFT[1];
1304 if(mv_table == s->b_forw_mv_table){
1305 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1307 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1310 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1315 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1317 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1318 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1320 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1321 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1322 mv_table[mot_xy][0]= mx;
1323 mv_table[mot_xy][1]= my;
1328 static inline int check_bidir_mv(MpegEncContext * s,
1329 int motion_fx, int motion_fy,
1330 int motion_bx, int motion_by,
1331 int pred_fx, int pred_fy,
1332 int pred_bx, int pred_by,
1336 //FIXME better f_code prediction (max mv & distance)
1338 MotionEstContext * const c= &s->me;
1339 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1340 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1341 int stride= c->stride;
1342 uint8_t *dest_y = c->scratchpad;
1347 uint8_t **src_data= c->src[0];
1348 uint8_t **ref_data= c->ref[0];
1349 uint8_t **ref2_data= c->ref[2];
1351 if(s->quarter_sample){
1352 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1353 src_x = motion_fx >> 2;
1354 src_y = motion_fy >> 2;
1356 ptr = ref_data[0] + (src_y * stride) + src_x;
1357 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1359 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1360 src_x = motion_bx >> 2;
1361 src_y = motion_by >> 2;
1363 ptr = ref2_data[0] + (src_y * stride) + src_x;
1364 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1366 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1367 src_x = motion_fx >> 1;
1368 src_y = motion_fy >> 1;
1370 ptr = ref_data[0] + (src_y * stride) + src_x;
1371 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1373 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1374 src_x = motion_bx >> 1;
1375 src_y = motion_by >> 1;
1377 ptr = ref2_data[0] + (src_y * stride) + src_x;
1378 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1381 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1382 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1383 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1385 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1392 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1393 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1395 MotionEstContext * const c= &s->me;
1396 const int mot_stride = s->mb_stride;
1397 const int xy = mb_y *mot_stride + mb_x;
1399 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1400 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1401 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1402 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1403 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1404 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1405 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1406 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1407 const int flags= c->sub_flags;
1408 const int qpel= flags&FLAG_QPEL;
1409 const int shift= 1+qpel;
1410 const int xmin= c->xmin<<shift;
1411 const int ymin= c->ymin<<shift;
1412 const int xmax= c->xmax<<shift;
1413 const int ymax= c->ymax<<shift;
1414 uint8_t map[8][8][8][8];
1416 memset(map,0,sizeof(map));
1417 #define BIDIR_MAP(fx,fy,bx,by) \
1418 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1419 BIDIR_MAP(0,0,0,0) = 1;
1421 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1422 motion_bx, motion_by,
1427 if(s->avctx->bidir_refine){
1429 #define CHECK_BIDIR(fx,fy,bx,by)\
1430 if( !BIDIR_MAP(fx,fy,bx,by)\
1431 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1432 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1433 BIDIR_MAP(fx,fy,bx,by) = 1;\
1434 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);\
1444 #define CHECK_BIDIR2(a,b,c,d)\
1445 CHECK_BIDIR(a,b,c,d)\
1446 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1448 #define CHECK_BIDIRR(a,b,c,d)\
1449 CHECK_BIDIR2(a,b,c,d)\
1450 CHECK_BIDIR2(b,c,d,a)\
1451 CHECK_BIDIR2(c,d,a,b)\
1452 CHECK_BIDIR2(d,a,b,c)
1457 CHECK_BIDIRR( 0, 0, 0, 1)
1458 if(s->avctx->bidir_refine > 1){
1459 CHECK_BIDIRR( 0, 0, 1, 1)
1460 CHECK_BIDIR2( 0, 1, 0, 1)
1461 CHECK_BIDIR2( 1, 0, 1, 0)
1462 CHECK_BIDIRR( 0, 0,-1, 1)
1463 CHECK_BIDIR2( 0,-1, 0, 1)
1464 CHECK_BIDIR2(-1, 0, 1, 0)
1465 if(s->avctx->bidir_refine > 2){
1466 CHECK_BIDIRR( 0, 1, 1, 1)
1467 CHECK_BIDIRR( 0,-1, 1, 1)
1468 CHECK_BIDIRR( 0, 1,-1, 1)
1469 CHECK_BIDIRR( 0, 1, 1,-1)
1470 if(s->avctx->bidir_refine > 3){
1471 CHECK_BIDIR2( 1, 1, 1, 1)
1472 CHECK_BIDIRR( 1, 1, 1,-1)
1473 CHECK_BIDIR2( 1, 1,-1,-1)
1474 CHECK_BIDIR2( 1,-1,-1, 1)
1475 CHECK_BIDIR2( 1,-1, 1,-1)
1482 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1483 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1484 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1485 s->b_bidir_back_mv_table[xy][1]= motion_by;
1490 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1492 MotionEstContext * const c= &s->me;
1494 const int mot_stride = s->mb_stride;
1495 const int mot_xy = mb_y*mot_stride + mb_x;
1496 const int shift= 1+s->quarter_sample;
1498 const int time_pp= s->pp_time;
1499 const int time_pb= s->pb_time;
1500 int mx, my, xmin, xmax, ymin, ymax;
1501 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1503 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1504 ymin= xmin=(-32)>>shift;
1505 ymax= xmax= 31>>shift;
1507 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1508 s->mv_type= MV_TYPE_8X8;
1510 s->mv_type= MV_TYPE_16X16;
1514 int index= s->block_index[i];
1517 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1518 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1519 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1520 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1521 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1522 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1524 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1525 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1526 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1528 xmax= FFMIN(xmax, s->width - max);
1529 xmin= FFMAX(xmin, - 16 - min);
1531 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1532 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1533 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1535 ymax= FFMIN(ymax, s->height - max);
1536 ymin= FFMAX(ymin, - 16 - min);
1538 if(s->mv_type == MV_TYPE_16X16) break;
1541 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1543 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1544 s->b_direct_mv_table[mot_xy][0]= 0;
1545 s->b_direct_mv_table[mot_xy][1]= 0;
1547 return 256*256*256*64;
1554 c->flags |= FLAG_DIRECT;
1555 c->sub_flags |= FLAG_DIRECT;
1559 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1560 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1562 /* special case for first line */
1563 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1564 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1565 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1566 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1567 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1569 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1570 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1573 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1574 if(c->sub_flags&FLAG_QPEL)
1575 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1577 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1579 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1580 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1582 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1584 mv_table[mot_xy][0]= mx;
1585 mv_table[mot_xy][1]= my;
1586 c->flags &= ~FLAG_DIRECT;
1587 c->sub_flags &= ~FLAG_DIRECT;
1592 void ff_estimate_b_frame_motion(MpegEncContext * s,
1595 MotionEstContext * const c= &s->me;
1596 const int penalty_factor= c->mb_penalty_factor;
1597 int fmin, bmin, dmin, fbmin, bimin, fimin;
1599 const int xy = mb_y*s->mb_stride + mb_x;
1600 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1602 get_limits(s, 16*mb_x, 16*mb_y);
1606 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1607 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1609 score= ((unsigned)(score*score + 128*256))>>16;
1610 c->mc_mb_var_sum_temp += score;
1611 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1612 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1617 if(c->avctx->me_threshold){
1618 int vard= check_input_motion(s, mb_x, mb_y, 0);
1620 if((vard+128)>>8 < c->avctx->me_threshold){
1621 // pix = c->src[0][0];
1622 // sum = s->dsp.pix_sum(pix, s->linesize);
1623 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1625 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1626 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1627 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1628 c->mb_var_sum_temp += (varc+128)>>8;*/
1629 c->mc_mb_var_sum_temp += (vard+128)>>8;
1630 /* if (vard <= 64<<8 || vard < varc) {
1631 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1633 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1637 if((vard+128)>>8 < c->avctx->mb_threshold){
1638 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1639 if(type == CANDIDATE_MB_TYPE_DIRECT){
1640 direct_search(s, mb_x, mb_y);
1642 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1644 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1646 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1648 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1650 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1652 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1653 interlaced_search(s, 0,
1654 s->b_field_mv_table[0], s->b_field_select_table[0],
1655 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1657 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1659 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1660 interlaced_search(s, 2,
1661 s->b_field_mv_table[1], s->b_field_select_table[1],
1662 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1668 if (s->codec_id == CODEC_ID_MPEG4)
1669 dmin= direct_search(s, mb_x, mb_y);
1672 //FIXME penalty stuff for non mpeg4
1674 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1677 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1678 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1681 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1682 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1684 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1685 //FIXME mb type penalty
1687 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1688 fimin= interlaced_search(s, 0,
1689 s->b_field_mv_table[0], s->b_field_select_table[0],
1690 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1691 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1692 bimin= interlaced_search(s, 2,
1693 s->b_field_mv_table[1], s->b_field_select_table[1],
1694 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1696 fimin= bimin= INT_MAX;
1700 type = CANDIDATE_MB_TYPE_FORWARD;
1704 type = CANDIDATE_MB_TYPE_DIRECT;
1708 type= CANDIDATE_MB_TYPE_BACKWARD;
1712 type= CANDIDATE_MB_TYPE_BIDIR;
1716 type= CANDIDATE_MB_TYPE_FORWARD_I;
1720 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1723 score= ((unsigned)(score*score + 128*256))>>16;
1724 c->mc_mb_var_sum_temp += score;
1725 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1728 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1729 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1731 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1733 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1734 if(fimin < INT_MAX && bimin < INT_MAX){
1735 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1737 //FIXME something smarter
1738 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1739 if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1740 type |= CANDIDATE_MB_TYPE_DIRECT0;
1742 if(s->out_format == FMT_MPEG1)
1743 type |= CANDIDATE_MB_TYPE_INTRA;
1747 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1750 /* find best f_code for ME which do unlimited searches */
1751 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1753 if(s->me_method>=ME_EPZS){
1755 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1756 uint8_t * fcode_tab= s->fcode_tab;
1758 int best_score=-10000000;
1760 if(s->msmpeg4_version)
1761 range= FFMIN(range, 16);
1762 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1763 range= FFMIN(range, 256);
1765 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1767 for(y=0; y<s->mb_height; y++){
1769 int xy= y*s->mb_stride;
1770 for(x=0; x<s->mb_width; x++){
1771 if(s->mb_type[xy] & type){
1772 int mx= mv_table[xy][0];
1773 int my= mv_table[xy][1];
1774 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1775 fcode_tab[my + MAX_MV]);
1778 if(mx >= range || mx < -range ||
1779 my >= range || my < -range)
1782 for(j=0; j<fcode && j<8; j++){
1783 if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1792 if(score[i] > best_score){
1793 best_score= score[i];
1796 // printf("%d %d\n", i, score[i]);
1799 // printf("fcode: %d type: %d\n", i, s->pict_type);
1801 /* for(i=0; i<=MAX_FCODE; i++){
1802 printf("%d ", mv_num[i]);
1810 void ff_fix_long_p_mvs(MpegEncContext * s)
1812 MotionEstContext * const c= &s->me;
1813 const int f_code= s->f_code;
1815 assert(s->pict_type==FF_P_TYPE);
1817 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1819 assert(range <= 16 || !s->msmpeg4_version);
1820 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1822 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1824 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1825 if(s->flags&CODEC_FLAG_4MV){
1826 const int wrap= s->b8_stride;
1828 /* clip / convert to intra 8x8 type MVs */
1829 for(y=0; y<s->mb_height; y++){
1831 int i= y*s->mb_stride;
1834 for(x=0; x<s->mb_width; x++){
1835 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1837 for(block=0; block<4; block++){
1838 int off= (block& 1) + (block>>1)*wrap;
1839 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1840 int my= s->current_picture.motion_val[0][ xy + off ][1];
1842 if( mx >=range || mx <-range
1843 || my >=range || my <-range){
1844 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1845 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1846 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1859 * @param truncate 1 for truncation, 0 for using intra
1861 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1862 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1864 MotionEstContext * const c= &s->me;
1865 int y, h_range, v_range;
1867 // RAL: 8 in MPEG-1, 16 in MPEG-4
1868 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1870 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1873 v_range= field_select_table ? range>>1 : range;
1875 /* clip / convert to intra 16x16 type MVs */
1876 for(y=0; y<s->mb_height; y++){
1878 int xy= y*s->mb_stride;
1879 for(x=0; x<s->mb_width; x++){
1880 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1881 if(field_select_table==NULL || field_select_table[xy] == field_select){
1882 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1883 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1886 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1887 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1888 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1889 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1891 s->mb_type[xy] &= ~type;
1892 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;