3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
36 #include "mpegvideo.h"
41 #define SQ(a) ((a)*(a))
45 #define P_TOPRIGHT P[3]
49 static inline int sad_hpel_motion_search(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
51 int src_index, int ref_index,
54 static inline int update_map_generation(MotionEstContext *c)
56 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57 if(c->map_generation==0){
58 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
61 return c->map_generation;
64 /* shape adaptive search stuff */
65 typedef struct Minima{
71 static int minima_cmp(const void *a, const void *b){
72 const Minima *da = (const Minima *) a;
73 const Minima *db = (const Minima *) b;
75 return da->height - db->height;
78 #define FLAG_QPEL 1 //must be 1
82 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
83 const int offset[3]= {
85 ((y*c->uvstride + x)>>1),
86 ((y*c->uvstride + x)>>1),
90 c->src[0][i]= src [i] + offset[i];
91 c->ref[0][i]= ref [i] + offset[i];
95 c->ref[ref_index][i]= ref2[i] + offset[i];
100 static int get_flags(MotionEstContext *c, int direct, int chroma){
101 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102 + (direct ? FLAG_DIRECT : 0)
103 + (chroma ? FLAG_CHROMA : 0);
106 static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
107 const int size, const int h, int ref_index, int src_index,
108 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
109 MotionEstContext * const c= &s->me;
110 const int stride= c->stride;
111 const int uvstride= c->uvstride;
112 const int qpel= flags&FLAG_QPEL;
113 const int chroma= flags&FLAG_CHROMA;
114 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
115 const int hx= subx + (x<<(1+qpel));
116 const int hy= suby + (y<<(1+qpel));
117 uint8_t * const * const ref= c->ref[ref_index];
118 uint8_t * const * const src= c->src[src_index];
120 //FIXME check chroma 4mv, (no crashes ...)
121 if(flags&FLAG_DIRECT){
122 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
123 const int time_pp= s->pp_time;
124 const int time_pb= s->pb_time;
125 const int mask= 2*qpel+1;
126 if(s->mv_type==MV_TYPE_8X8){
129 int fx = c->direct_basis_mv[i][0] + hx;
130 int fy = c->direct_basis_mv[i][1] + hy;
131 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));
132 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));
133 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
134 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
136 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
138 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
139 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
141 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
142 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
146 int fx = c->direct_basis_mv[0][0] + hx;
147 int fy = c->direct_basis_mv[0][1] + hy;
148 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
149 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
150 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
151 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
154 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
155 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
156 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
157 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
158 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
159 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
160 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
161 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
163 assert((fx>>1) + 16*s->mb_x >= -16);
164 assert((fy>>1) + 16*s->mb_y >= -16);
165 assert((fx>>1) + 16*s->mb_x <= s->width);
166 assert((fy>>1) + 16*s->mb_y <= s->height);
167 assert((bx>>1) + 16*s->mb_x >= -16);
168 assert((by>>1) + 16*s->mb_y >= -16);
169 assert((bx>>1) + 16*s->mb_x <= s->width);
170 assert((by>>1) + 16*s->mb_y <= s->height);
172 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
173 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
176 d = cmp_func(s, c->temp, src[0], stride, 16);
180 int uvdxy; /* no, it might not be used uninitialized */
183 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
189 uvdxy= (cx&1) + 2*(cy&1);
190 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
193 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
195 uvdxy= dxy | (x&1) | (2*(y&1));
197 d = cmp_func(s, c->temp, src[0], stride, h);
199 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
201 uvdxy= (x&1) + 2*(y&1);
204 uint8_t * const uvtemp= c->temp + 16*stride;
205 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
206 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
207 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
208 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
213 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
217 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
222 #include "motion_est_template.c"
224 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
228 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
231 void ff_init_me(MpegEncContext *s){
232 MotionEstContext * const c= &s->me;
235 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
236 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
237 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
238 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
240 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
241 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
242 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
244 /*FIXME s->no_rounding b_type*/
245 if(s->flags&CODEC_FLAG_QPEL){
246 c->sub_motion_search= qpel_motion_search;
247 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
248 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
249 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
251 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
252 c->sub_motion_search= hpel_motion_search;
253 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
254 && c->avctx-> me_cmp == FF_CMP_SAD
255 && c->avctx-> mb_cmp == FF_CMP_SAD)
256 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
258 c->sub_motion_search= hpel_motion_search;
260 c->hpel_avg= s->dsp.avg_pixels_tab;
261 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
262 else c->hpel_put= s->dsp.put_pixels_tab;
265 c->stride = s->linesize;
266 c->uvstride= s->uvlinesize;
268 c->stride = 16*s->mb_width + 32;
269 c->uvstride= 8*s->mb_width + 16;
272 // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
273 if(s->codec_id != CODEC_ID_SNOW){
274 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
275 s->dsp.me_cmp[2]= zero_cmp;
277 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
278 s->dsp.me_sub_cmp[2]= zero_cmp;
280 c->hpel_put[2][0]= c->hpel_put[2][1]=
281 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
284 if(s->codec_id == CODEC_ID_H261){
285 c->sub_motion_search= no_sub_motion_search;
288 c->temp= c->scratchpad;
292 static int pix_dev(uint8_t * pix, int line_size, int mean)
297 for (i = 0; i < 16; i++) {
298 for (j = 0; j < 16; j += 8) {
299 s += ABS(pix[0]-mean);
300 s += ABS(pix[1]-mean);
301 s += ABS(pix[2]-mean);
302 s += ABS(pix[3]-mean);
303 s += ABS(pix[4]-mean);
304 s += ABS(pix[5]-mean);
305 s += ABS(pix[6]-mean);
306 s += ABS(pix[7]-mean);
309 pix += line_size - 16;
315 static inline void no_motion_search(MpegEncContext * s,
316 int *mx_ptr, int *my_ptr)
318 *mx_ptr = 16 * s->mb_x;
319 *my_ptr = 16 * s->mb_y;
322 #if 0 /* the use of these functions is inside #if 0 */
323 static int full_motion_search(MpegEncContext * s,
324 int *mx_ptr, int *my_ptr, int range,
325 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
327 int x1, y1, x2, y2, xx, yy, x, y;
333 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
345 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
349 for (y = y1; y <= y2; y++) {
350 for (x = x1; x <= x2; x++) {
351 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
355 (abs(x - xx) + abs(y - yy)) <
356 (abs(mx - xx) + abs(my - yy)))) {
368 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
369 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
370 av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
377 static int log_motion_search(MpegEncContext * s,
378 int *mx_ptr, int *my_ptr, int range,
379 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
381 int x1, y1, x2, y2, xx, yy, x, y;
408 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
414 for (y = y1; y <= y2; y += range) {
415 for (x = x1; x <= x2; x += range) {
416 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
417 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
443 } while (range >= 1);
446 av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
453 static int phods_motion_search(MpegEncContext * s,
454 int *mx_ptr, int *my_ptr, int range,
455 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
457 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
458 int mx, my, dminx, dminy;
484 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
495 for (x = x1; x <= x2; x += range) {
496 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
497 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
504 for (y = y1; y <= y2; y += range) {
505 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
506 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
532 } while (range >= 1);
535 av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
538 /* half pixel search */
545 #define Z_THRESHOLD 256
547 #define CHECK_SAD_HALF_MV(suffix, x, y) \
549 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
550 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
551 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
554 static inline int sad_hpel_motion_search(MpegEncContext * s,
555 int *mx_ptr, int *my_ptr, int dmin,
556 int src_index, int ref_index,
559 MotionEstContext * const c= &s->me;
560 const int penalty_factor= c->sub_penalty_factor;
563 int stride= c->stride;
564 const int flags= c->sub_flags;
577 pix = c->src[src_index][0];
581 ptr = c->ref[ref_index][0] + (my * stride) + mx;
585 if (mx > xmin && mx < xmax &&
586 my > ymin && my < ymax) {
589 const int index= (my<<ME_MAP_SHIFT) + mx;
590 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
591 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
592 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
593 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
603 CHECK_SAD_HALF_MV(y2 , 0, -1)
605 CHECK_SAD_HALF_MV(xy2, -1, -1)
607 CHECK_SAD_HALF_MV(xy2, +1, -1)
611 CHECK_SAD_HALF_MV(xy2, -1, +1)
613 CHECK_SAD_HALF_MV(x2 , -1, 0)
615 CHECK_SAD_HALF_MV(xy2, +1, -1)
617 CHECK_SAD_HALF_MV(xy2, -1, -1)
621 CHECK_SAD_HALF_MV(xy2, +1, +1)
623 CHECK_SAD_HALF_MV(x2 , +1, 0)
628 CHECK_SAD_HALF_MV(xy2, -1, -1)
632 CHECK_SAD_HALF_MV(xy2, +1, +1)
634 CHECK_SAD_HALF_MV(x2 , -1, 0)
635 CHECK_SAD_HALF_MV(xy2, -1, +1)
638 CHECK_SAD_HALF_MV(xy2, +1, -1)
642 CHECK_SAD_HALF_MV(xy2, -1, +1)
644 CHECK_SAD_HALF_MV(x2 , +1, 0)
645 CHECK_SAD_HALF_MV(xy2, +1, +1)
647 CHECK_SAD_HALF_MV(y2 , 0, +1)
662 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
664 const int xy= s->mb_x + s->mb_y*s->mb_stride;
666 s->p_mv_table[xy][0] = mx;
667 s->p_mv_table[xy][1] = my;
669 /* has already been set to the 4 MV if 4MV is done */
671 int mot_xy= s->block_index[0];
673 s->current_picture.motion_val[0][mot_xy ][0]= mx;
674 s->current_picture.motion_val[0][mot_xy ][1]= my;
675 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
676 s->current_picture.motion_val[0][mot_xy+1][1]= my;
678 mot_xy += s->b8_stride;
679 s->current_picture.motion_val[0][mot_xy ][0]= mx;
680 s->current_picture.motion_val[0][mot_xy ][1]= my;
681 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
682 s->current_picture.motion_val[0][mot_xy+1][1]= my;
687 * get fullpel ME search limits.
689 static inline void get_limits(MpegEncContext *s, int x, int y)
691 MotionEstContext * const c= &s->me;
693 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
696 if (s->unrestricted_mv) {
699 c->xmax = - x + s->mb_width *16;
700 c->ymax = - y + s->mb_height*16;
701 } else if (s->out_format == FMT_H261){
702 // Search range of H261 is different from other codec standards
703 c->xmin = (x > 15) ? - 15 : 0;
704 c->ymin = (y > 15) ? - 15 : 0;
705 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
706 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
710 c->xmax = - x + s->mb_width *16 - 16;
711 c->ymax = - y + s->mb_height*16 - 16;
715 static inline void init_mv4_ref(MotionEstContext *c){
716 const int stride= c->stride;
718 c->ref[1][0] = c->ref[0][0] + 8;
719 c->ref[2][0] = c->ref[0][0] + 8*stride;
720 c->ref[3][0] = c->ref[2][0] + 8;
721 c->src[1][0] = c->src[0][0] + 8;
722 c->src[2][0] = c->src[0][0] + 8*stride;
723 c->src[3][0] = c->src[2][0] + 8;
726 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
728 MotionEstContext * const c= &s->me;
733 int dmin_sum=0, mx4_sum=0, my4_sum=0;
735 const int stride= c->stride;
736 uint8_t *mv_penalty= c->current_mv_penalty;
740 for(block=0; block<4; block++){
742 int pred_x4, pred_y4;
744 static const int off[4]= {2, 1, 1, -1};
745 const int mot_stride = s->b8_stride;
746 const int mot_xy = s->block_index[block];
748 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
749 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
751 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
753 /* special case for first line */
754 if (s->first_slice_line && block<2) {
755 c->pred_x= pred_x4= P_LEFT[0];
756 c->pred_y= pred_y4= P_LEFT[1];
758 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
759 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
760 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
761 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
762 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
763 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
764 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
765 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
767 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
768 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
770 c->pred_x= pred_x4 = P_MEDIAN[0];
771 c->pred_y= pred_y4 = P_MEDIAN[1];
776 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
778 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
780 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
782 const int offset= ((block&1) + (block>>1)*stride)*8;
783 uint8_t *dest_y = c->scratchpad + offset;
784 if(s->quarter_sample){
785 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
786 dxy = ((my4 & 3) << 2) | (mx4 & 3);
789 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
791 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
793 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
794 dxy = ((my4 & 1) << 1) | (mx4 & 1);
797 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
799 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
801 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
805 if(s->quarter_sample){
813 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
814 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
816 if(mx4 != mx || my4 != my) same=0;
822 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
823 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);
826 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
831 mx= ff_h263_round_chroma(mx4_sum);
832 my= ff_h263_round_chroma(my4_sum);
833 dxy = ((my & 1) << 1) | (mx & 1);
835 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
838 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
839 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
841 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
842 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
845 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);
846 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);
852 switch(c->avctx->mb_cmp&0xFF){
854 return dmin_sum+ 32*s->qscale*s->qscale;*/
858 return dmin_sum+ 11*c->mb_penalty_factor;
862 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
863 MotionEstContext * const c= &s->me;
865 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
866 c->src[1][0] = c->src[0][0] + s->linesize;
867 if(c->flags & FLAG_CHROMA){
868 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
869 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
870 c->src[1][1] = c->src[0][1] + s->uvlinesize;
871 c->src[1][2] = c->src[0][2] + s->uvlinesize;
875 static int interlaced_search(MpegEncContext *s, int ref_index,
876 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
878 MotionEstContext * const c= &s->me;
883 uint8_t * const mv_penalty= c->current_mv_penalty;
885 const int stride= 2*s->linesize;
887 const int mot_stride= s->mb_stride;
888 const int xy= s->mb_x + s->mb_y*mot_stride;
894 init_interlaced_ref(s, ref_index);
896 for(block=0; block<2; block++){
898 int best_dmin= INT_MAX;
901 for(field_select=0; field_select<2; field_select++){
902 int dmin, mx_i, my_i;
903 int16_t (*mv_table)[2]= mv_tables[block][field_select];
905 if(user_field_select){
906 assert(field_select==0 || field_select==1);
907 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
908 if(field_select_tables[block][xy] != field_select)
912 P_LEFT[0] = mv_table[xy - 1][0];
913 P_LEFT[1] = mv_table[xy - 1][1];
914 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
916 c->pred_x= P_LEFT[0];
917 c->pred_y= P_LEFT[1];
919 if(!s->first_slice_line){
920 P_TOP[0] = mv_table[xy - mot_stride][0];
921 P_TOP[1] = mv_table[xy - mot_stride][1];
922 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
923 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
924 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
925 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
926 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
927 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
929 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
930 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
932 P_MV1[0]= mx; //FIXME not correct if block != field_select
935 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
937 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
939 mv_table[xy][0]= mx_i;
940 mv_table[xy][1]= my_i;
942 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
946 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
947 dxy = ((my_i & 1) << 1) | (mx_i & 1);
950 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
952 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
954 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
955 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
957 dmin+= c->mb_penalty_factor; //field_select bits
959 dmin += field_select != block; //slightly prefer same field
961 if(dmin < best_dmin){
963 best_field= field_select;
967 int16_t (*mv_table)[2]= mv_tables[block][best_field];
969 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
970 if(mv_table[xy][1]&1) same=0;
971 if(mv_table[xy][1]*2 != my) same=0;
972 if(best_field != block) same=0;
975 field_select_tables[block][xy]= best_field;
976 dmin_sum += best_dmin;
987 switch(c->avctx->mb_cmp&0xFF){
989 return dmin_sum+ 32*s->qscale*s->qscale;*/
993 return dmin_sum+ 11*c->mb_penalty_factor;
997 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
998 int ymax= s->me.ymax>>interlaced;
999 int ymin= s->me.ymin>>interlaced;
1001 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1002 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1003 if(mv[1] < ymin) mv[1] = ymin;
1004 if(mv[1] > ymax) mv[1] = ymax;
1007 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1008 MotionEstContext * const c= &s->me;
1009 Picture *p= s->current_picture_ptr;
1010 int mb_xy= mb_x + mb_y*s->mb_stride;
1011 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1012 int mb_type= s->current_picture.mb_type[mb_xy];
1013 int flags= c->flags;
1014 int shift= (flags&FLAG_QPEL) + 1;
1015 int mask= (1<<shift)-1;
1018 me_cmp_func cmpf= s->dsp.sse[0];
1019 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1021 if(p_type && USES_LIST(mb_type, 1)){
1022 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1025 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1028 int xy= s->block_index[i];
1029 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1030 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1033 if(IS_INTERLACED(mb_type)){
1034 int xy2= xy + s->b8_stride;
1035 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1039 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1040 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1044 if(USES_LIST(mb_type, 0)){
1045 int field_select0= p->ref_index[0][xy ];
1046 int field_select1= p->ref_index[0][xy2];
1047 assert(field_select0==0 ||field_select0==1);
1048 assert(field_select1==0 ||field_select1==1);
1049 init_interlaced_ref(s, 0);
1052 s->p_field_select_table[0][mb_xy]= field_select0;
1053 s->p_field_select_table[1][mb_xy]= field_select1;
1054 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1055 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1056 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1058 s->b_field_select_table[0][0][mb_xy]= field_select0;
1059 s->b_field_select_table[0][1][mb_xy]= field_select1;
1060 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1061 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1062 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1065 x= p->motion_val[0][xy ][0];
1066 y= p->motion_val[0][xy ][1];
1067 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1068 x= p->motion_val[0][xy2][0];
1069 y= p->motion_val[0][xy2][1];
1070 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1072 if(USES_LIST(mb_type, 1)){
1073 int field_select0= p->ref_index[1][xy ];
1074 int field_select1= p->ref_index[1][xy2];
1075 assert(field_select0==0 ||field_select0==1);
1076 assert(field_select1==0 ||field_select1==1);
1077 init_interlaced_ref(s, 2);
1079 s->b_field_select_table[1][0][mb_xy]= field_select0;
1080 s->b_field_select_table[1][1][mb_xy]= field_select1;
1081 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1082 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1083 if(USES_LIST(mb_type, 0)){
1084 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1086 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1089 x= p->motion_val[1][xy ][0];
1090 y= p->motion_val[1][xy ][1];
1091 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1092 x= p->motion_val[1][xy2][0];
1093 y= p->motion_val[1][xy2][1];
1094 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1095 //FIXME bidir scores
1099 }else if(IS_8X8(mb_type)){
1100 if(!(s->flags & CODEC_FLAG_4MV)){
1101 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1104 cmpf= s->dsp.sse[1];
1105 chroma_cmpf= s->dsp.sse[1];
1108 xy= s->block_index[i];
1109 x= p->motion_val[0][xy][0];
1110 y= p->motion_val[0][xy][1];
1111 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1113 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1115 if(USES_LIST(mb_type, 0)){
1117 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1118 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1119 }else if(USES_LIST(mb_type, 1)){
1120 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1121 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1122 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1124 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1125 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1127 x= p->motion_val[0][xy][0];
1128 y= p->motion_val[0][xy][1];
1129 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1130 }else if(USES_LIST(mb_type, 1)){
1131 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1132 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1134 x= p->motion_val[1][xy][0];
1135 y= p->motion_val[1][xy][1];
1136 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1138 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1143 void ff_estimate_p_frame_motion(MpegEncContext * s,
1146 MotionEstContext * const c= &s->me;
1147 uint8_t *pix, *ppix;
1148 int sum, varc, vard, mx, my, dmin;
1150 const int shift= 1+s->quarter_sample;
1152 Picture * const pic= &s->current_picture;
1154 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1156 assert(s->quarter_sample==0 || s->quarter_sample==1);
1157 assert(s->linesize == c->stride);
1158 assert(s->uvlinesize == c->uvstride);
1160 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1161 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1162 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1163 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1165 get_limits(s, 16*mb_x, 16*mb_y);
1168 /* intra / predictive decision */
1170 sum = s->dsp.pix_sum(pix, s->linesize);
1171 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1173 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1174 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1175 c->mb_var_sum_temp += (varc+128)>>8;
1177 if(c->avctx->me_threshold){
1178 vard= check_input_motion(s, mb_x, mb_y, 1);
1180 if((vard+128)>>8 < c->avctx->me_threshold){
1181 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1182 c->mc_mb_var_sum_temp += (vard+128)>>8;
1183 if (vard <= 64<<8 || vard < varc) { //FIXME
1184 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1186 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1190 if((vard+128)>>8 < c->avctx->mb_threshold)
1191 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1194 switch(s->me_method) {
1197 no_motion_search(s, &mx, &my);
1204 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1209 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1214 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1222 const int mot_stride = s->b8_stride;
1223 const int mot_xy = s->block_index[0];
1225 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1226 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1228 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1230 if(!s->first_slice_line) {
1231 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1232 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1233 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1234 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1235 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1236 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1237 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1239 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1240 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1242 if(s->out_format == FMT_H263){
1243 c->pred_x = P_MEDIAN[0];
1244 c->pred_y = P_MEDIAN[1];
1245 }else { /* mpeg1 at least */
1246 c->pred_x= P_LEFT[0];
1247 c->pred_y= P_LEFT[1];
1250 c->pred_x= P_LEFT[0];
1251 c->pred_y= P_LEFT[1];
1255 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1260 /* At this point (mx,my) are full-pell and the relative displacement */
1261 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1263 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1265 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1266 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1267 c->mc_mb_var_sum_temp += (vard+128)>>8;
1270 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1271 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1274 if (vard <= 64<<8 || vard < varc)
1275 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1277 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1279 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1280 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1281 set_p_mv_tables(s, mx, my, 1);
1286 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1287 h263_mv4_search(s, mx, my, shift);
1289 set_p_mv_tables(s, mx, my, 0);
1291 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1292 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1294 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1295 if (vard <= 64<<8 || vard < varc)
1296 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1298 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1300 if (vard*2 + 200*256 > varc)
1301 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1302 if (varc*2 + 200*256 > vard){
1303 mb_type|= CANDIDATE_MB_TYPE_INTER;
1304 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1305 if(s->flags&CODEC_FLAG_MV0)
1307 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1312 if((s->flags&CODEC_FLAG_4MV)
1313 && !c->skip && varc>50<<8 && vard>10<<8){
1314 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1315 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1317 set_p_mv_tables(s, mx, my, 0);
1319 set_p_mv_tables(s, mx, my, 1);
1320 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1321 && !c->skip){ //FIXME varc/d checks
1322 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1323 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1327 mb_type= CANDIDATE_MB_TYPE_INTER;
1329 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1330 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1331 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1333 if((s->flags&CODEC_FLAG_4MV)
1334 && !c->skip && varc>50<<8 && vard>10<<8){
1335 int dmin4= h263_mv4_search(s, mx, my, shift);
1337 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1341 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1342 && !c->skip){ //FIXME varc/d checks
1343 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1345 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1350 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1351 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1353 /* get intra luma score */
1354 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1355 intra_score= varc - 500;
1357 int mean= (sum+128)>>8;
1360 for(i=0; i<16; i++){
1361 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1362 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1363 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1364 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1367 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1370 /* get chroma score */
1371 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1376 if(s->out_format == FMT_H263){
1377 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1379 mean= (s->last_dc[i] + 4)>>3;
1381 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1385 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1386 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1389 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1393 intra_score += c->mb_penalty_factor*16;
1395 if(intra_score < dmin){
1396 mb_type= CANDIDATE_MB_TYPE_INTRA;
1397 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1399 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1401 if (vard <= 64<<8 || vard < varc) { //FIXME
1402 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1404 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1408 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1411 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1414 MotionEstContext * const c= &s->me;
1417 const int shift= 1+s->quarter_sample;
1418 const int xy= mb_x + mb_y*s->mb_stride;
1419 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1421 assert(s->quarter_sample==0 || s->quarter_sample==1);
1423 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1424 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1426 get_limits(s, 16*mb_x, 16*mb_y);
1429 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1430 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1432 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1434 /* special case for first line */
1435 if (s->first_slice_line) {
1436 c->pred_x= P_LEFT[0];
1437 c->pred_y= P_LEFT[1];
1438 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1439 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1441 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1442 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1443 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1444 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1445 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1446 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1447 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1449 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1450 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1452 c->pred_x = P_MEDIAN[0];
1453 c->pred_y = P_MEDIAN[1];
1456 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1458 s->p_mv_table[xy][0] = mx<<shift;
1459 s->p_mv_table[xy][1] = my<<shift;
1464 static int ff_estimate_motion_b(MpegEncContext * s,
1465 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1467 MotionEstContext * const c= &s->me;
1470 const int shift= 1+s->quarter_sample;
1471 const int mot_stride = s->mb_stride;
1472 const int mot_xy = mb_y*mot_stride + mb_x;
1473 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1476 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1477 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1478 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1479 c->current_mv_penalty= mv_penalty;
1481 get_limits(s, 16*mb_x, 16*mb_y);
1483 switch(s->me_method) {
1486 no_motion_search(s, &mx, &my);
1493 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1498 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1503 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1511 P_LEFT[0] = mv_table[mot_xy - 1][0];
1512 P_LEFT[1] = mv_table[mot_xy - 1][1];
1514 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1516 /* special case for first line */
1517 if (!s->first_slice_line) {
1518 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1519 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1520 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1521 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1522 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1523 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1524 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1526 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1527 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1529 c->pred_x= P_LEFT[0];
1530 c->pred_y= P_LEFT[1];
1533 if(mv_table == s->b_forw_mv_table){
1534 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1536 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1539 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1544 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1546 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1547 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1549 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1550 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1551 mv_table[mot_xy][0]= mx;
1552 mv_table[mot_xy][1]= my;
1557 static inline int check_bidir_mv(MpegEncContext * s,
1558 int motion_fx, int motion_fy,
1559 int motion_bx, int motion_by,
1560 int pred_fx, int pred_fy,
1561 int pred_bx, int pred_by,
1565 //FIXME better f_code prediction (max mv & distance)
1567 MotionEstContext * const c= &s->me;
1568 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1569 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1570 int stride= c->stride;
1571 uint8_t *dest_y = c->scratchpad;
1576 uint8_t **src_data= c->src[0];
1577 uint8_t **ref_data= c->ref[0];
1578 uint8_t **ref2_data= c->ref[2];
1580 if(s->quarter_sample){
1581 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1582 src_x = motion_fx >> 2;
1583 src_y = motion_fy >> 2;
1585 ptr = ref_data[0] + (src_y * stride) + src_x;
1586 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1588 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1589 src_x = motion_bx >> 2;
1590 src_y = motion_by >> 2;
1592 ptr = ref2_data[0] + (src_y * stride) + src_x;
1593 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1595 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1596 src_x = motion_fx >> 1;
1597 src_y = motion_fy >> 1;
1599 ptr = ref_data[0] + (src_y * stride) + src_x;
1600 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1602 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1603 src_x = motion_bx >> 1;
1604 src_y = motion_by >> 1;
1606 ptr = ref2_data[0] + (src_y * stride) + src_x;
1607 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1610 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1611 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1612 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1614 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1621 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1622 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1624 MotionEstContext * const c= &s->me;
1625 const int mot_stride = s->mb_stride;
1626 const int xy = mb_y *mot_stride + mb_x;
1628 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1629 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1630 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1631 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1632 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1633 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1634 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1635 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1636 const int flags= c->sub_flags;
1637 const int qpel= flags&FLAG_QPEL;
1638 const int shift= 1+qpel;
1639 const int xmin= c->xmin<<shift;
1640 const int ymin= c->ymin<<shift;
1641 const int xmax= c->xmax<<shift;
1642 const int ymax= c->ymax<<shift;
1643 uint8_t map[8][8][8][8];
1645 memset(map,0,sizeof(map));
1646 #define BIDIR_MAP(fx,fy,bx,by) \
1647 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1648 BIDIR_MAP(0,0,0,0) = 1;
1650 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1651 motion_bx, motion_by,
1656 if(s->avctx->bidir_refine){
1658 #define CHECK_BIDIR(fx,fy,bx,by)\
1659 if( !BIDIR_MAP(fx,fy,bx,by)\
1660 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1661 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1662 BIDIR_MAP(fx,fy,bx,by) = 1;\
1663 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);\
1673 #define CHECK_BIDIR2(a,b,c,d)\
1674 CHECK_BIDIR(a,b,c,d)\
1675 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1677 #define CHECK_BIDIRR(a,b,c,d)\
1678 CHECK_BIDIR2(a,b,c,d)\
1679 CHECK_BIDIR2(b,c,d,a)\
1680 CHECK_BIDIR2(c,d,a,b)\
1681 CHECK_BIDIR2(d,a,b,c)
1686 CHECK_BIDIRR( 0, 0, 0, 1)
1687 if(s->avctx->bidir_refine > 1){
1688 CHECK_BIDIRR( 0, 0, 1, 1)
1689 CHECK_BIDIR2( 0, 1, 0, 1)
1690 CHECK_BIDIR2( 1, 0, 1, 0)
1691 CHECK_BIDIRR( 0, 0,-1, 1)
1692 CHECK_BIDIR2( 0,-1, 0, 1)
1693 CHECK_BIDIR2(-1, 0, 1, 0)
1694 if(s->avctx->bidir_refine > 2){
1695 CHECK_BIDIRR( 0, 1, 1, 1)
1696 CHECK_BIDIRR( 0,-1, 1, 1)
1697 CHECK_BIDIRR( 0, 1,-1, 1)
1698 CHECK_BIDIRR( 0, 1, 1,-1)
1699 if(s->avctx->bidir_refine > 3){
1700 CHECK_BIDIR2( 1, 1, 1, 1)
1701 CHECK_BIDIRR( 1, 1, 1,-1)
1702 CHECK_BIDIR2( 1, 1,-1,-1)
1703 CHECK_BIDIR2( 1,-1,-1, 1)
1704 CHECK_BIDIR2( 1,-1, 1,-1)
1711 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1712 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1713 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1714 s->b_bidir_back_mv_table[xy][1]= motion_by;
1719 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1721 MotionEstContext * const c= &s->me;
1723 const int mot_stride = s->mb_stride;
1724 const int mot_xy = mb_y*mot_stride + mb_x;
1725 const int shift= 1+s->quarter_sample;
1727 const int time_pp= s->pp_time;
1728 const int time_pb= s->pb_time;
1729 int mx, my, xmin, xmax, ymin, ymax;
1730 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1732 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1733 ymin= xmin=(-32)>>shift;
1734 ymax= xmax= 31>>shift;
1736 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1737 s->mv_type= MV_TYPE_8X8;
1739 s->mv_type= MV_TYPE_16X16;
1743 int index= s->block_index[i];
1746 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1747 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1748 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1749 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1750 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1751 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1753 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1754 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1755 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1757 xmax= FFMIN(xmax, s->width - max);
1758 xmin= FFMAX(xmin, - 16 - min);
1760 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1761 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1762 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1764 ymax= FFMIN(ymax, s->height - max);
1765 ymin= FFMAX(ymin, - 16 - min);
1767 if(s->mv_type == MV_TYPE_16X16) break;
1770 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1772 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1773 s->b_direct_mv_table[mot_xy][0]= 0;
1774 s->b_direct_mv_table[mot_xy][1]= 0;
1776 return 256*256*256*64;
1783 c->flags |= FLAG_DIRECT;
1784 c->sub_flags |= FLAG_DIRECT;
1788 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1789 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1791 /* special case for first line */
1792 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
1793 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1794 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1795 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1796 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1798 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1799 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1802 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1803 if(c->sub_flags&FLAG_QPEL)
1804 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1806 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1808 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1809 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1811 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1813 s->b_direct_mv_table[mot_xy][0]= mx;
1814 s->b_direct_mv_table[mot_xy][1]= my;
1815 c->flags &= ~FLAG_DIRECT;
1816 c->sub_flags &= ~FLAG_DIRECT;
1821 void ff_estimate_b_frame_motion(MpegEncContext * s,
1824 MotionEstContext * const c= &s->me;
1825 const int penalty_factor= c->mb_penalty_factor;
1826 int fmin, bmin, dmin, fbmin, bimin, fimin;
1828 const int xy = mb_y*s->mb_stride + mb_x;
1829 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1831 get_limits(s, 16*mb_x, 16*mb_y);
1834 if(c->avctx->me_threshold){
1835 int vard= check_input_motion(s, mb_x, mb_y, 0);
1837 if((vard+128)>>8 < c->avctx->me_threshold){
1838 // pix = c->src[0][0];
1839 // sum = s->dsp.pix_sum(pix, s->linesize);
1840 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1842 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1843 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1844 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1845 c->mb_var_sum_temp += (varc+128)>>8;*/
1846 c->mc_mb_var_sum_temp += (vard+128)>>8;
1847 /* if (vard <= 64<<8 || vard < varc) {
1848 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1850 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1854 if((vard+128)>>8 < c->avctx->mb_threshold){
1855 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1856 if(type == CANDIDATE_MB_TYPE_DIRECT){
1857 direct_search(s, mb_x, mb_y);
1859 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1861 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1863 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1865 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1867 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1869 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1870 interlaced_search(s, 0,
1871 s->b_field_mv_table[0], s->b_field_select_table[0],
1872 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1874 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1876 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1877 interlaced_search(s, 2,
1878 s->b_field_mv_table[1], s->b_field_select_table[1],
1879 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1885 if (s->codec_id == CODEC_ID_MPEG4)
1886 dmin= direct_search(s, mb_x, mb_y);
1889 //FIXME penalty stuff for non mpeg4
1891 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1894 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1895 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1898 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1899 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1901 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1902 //FIXME mb type penalty
1904 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1905 fimin= interlaced_search(s, 0,
1906 s->b_field_mv_table[0], s->b_field_select_table[0],
1907 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1908 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1909 bimin= interlaced_search(s, 2,
1910 s->b_field_mv_table[1], s->b_field_select_table[1],
1911 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1913 fimin= bimin= INT_MAX;
1917 type = CANDIDATE_MB_TYPE_FORWARD;
1921 type = CANDIDATE_MB_TYPE_DIRECT;
1925 type= CANDIDATE_MB_TYPE_BACKWARD;
1929 type= CANDIDATE_MB_TYPE_BIDIR;
1933 type= CANDIDATE_MB_TYPE_FORWARD_I;
1937 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1940 score= ((unsigned)(score*score + 128*256))>>16;
1941 c->mc_mb_var_sum_temp += score;
1942 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1945 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1946 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1948 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1950 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1951 if(fimin < INT_MAX && bimin < INT_MAX){
1952 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1954 //FIXME something smarter
1955 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1957 if(s->out_format == FMT_MPEG1)
1958 type |= CANDIDATE_MB_TYPE_INTRA;
1962 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1965 /* find best f_code for ME which do unlimited searches */
1966 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1968 if(s->me_method>=ME_EPZS){
1970 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1971 uint8_t * fcode_tab= s->fcode_tab;
1973 int best_score=-10000000;
1975 if(s->msmpeg4_version)
1976 range= FFMIN(range, 16);
1977 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1978 range= FFMIN(range, 256);
1980 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1982 for(y=0; y<s->mb_height; y++){
1984 int xy= y*s->mb_stride;
1985 for(x=0; x<s->mb_width; x++){
1986 if(s->mb_type[xy] & type){
1987 int mx= mv_table[xy][0];
1988 int my= mv_table[xy][1];
1989 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1990 fcode_tab[my + MAX_MV]);
1993 if(mx >= range || mx < -range ||
1994 my >= range || my < -range)
1997 for(j=0; j<fcode && j<8; j++){
1998 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2007 if(score[i] > best_score){
2008 best_score= score[i];
2011 // printf("%d %d\n", i, score[i]);
2014 // printf("fcode: %d type: %d\n", i, s->pict_type);
2016 /* for(i=0; i<=MAX_FCODE; i++){
2017 printf("%d ", mv_num[i]);
2025 void ff_fix_long_p_mvs(MpegEncContext * s)
2027 MotionEstContext * const c= &s->me;
2028 const int f_code= s->f_code;
2030 assert(s->pict_type==P_TYPE);
2032 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2034 assert(range <= 16 || !s->msmpeg4_version);
2035 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2037 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2039 //printf("%d no:%d %d//\n", clip, noclip, f_code);
2040 if(s->flags&CODEC_FLAG_4MV){
2041 const int wrap= s->b8_stride;
2043 /* clip / convert to intra 8x8 type MVs */
2044 for(y=0; y<s->mb_height; y++){
2046 int i= y*s->mb_stride;
2049 for(x=0; x<s->mb_width; x++){
2050 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2052 for(block=0; block<4; block++){
2053 int off= (block& 1) + (block>>1)*wrap;
2054 int mx= s->current_picture.motion_val[0][ xy + off ][0];
2055 int my= s->current_picture.motion_val[0][ xy + off ][1];
2057 if( mx >=range || mx <-range
2058 || my >=range || my <-range){
2059 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2060 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2061 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2074 * @param truncate 1 for truncation, 0 for using intra
2076 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2077 int16_t (*mv_table)[2], int f_code, int type, int truncate)
2079 MotionEstContext * const c= &s->me;
2080 int y, h_range, v_range;
2082 // RAL: 8 in MPEG-1, 16 in MPEG-4
2083 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2085 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2088 v_range= field_select_table ? range>>1 : range;
2090 /* clip / convert to intra 16x16 type MVs */
2091 for(y=0; y<s->mb_height; y++){
2093 int xy= y*s->mb_stride;
2094 for(x=0; x<s->mb_width; x++){
2095 if (s->mb_type[xy] & type){ // RAL: "type" test added...
2096 if(field_select_table==NULL || field_select_table[xy] == field_select){
2097 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2098 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2101 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2102 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2103 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2104 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2106 s->mb_type[xy] &= ~type;
2107 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;