3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2 of the License, or (at your option) any later version.
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
34 #include "mpegvideo.h"
39 #define SQ(a) ((a)*(a))
43 #define P_TOPRIGHT P[3]
47 static inline int sad_hpel_motion_search(MpegEncContext * s,
48 int *mx_ptr, int *my_ptr, int dmin,
49 int src_index, int ref_index,
52 static inline int update_map_generation(MpegEncContext * s)
54 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
55 if(s->me.map_generation==0){
56 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
57 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59 return s->me.map_generation;
62 /* shape adaptive search stuff */
63 typedef struct Minima{
69 static int minima_cmp(const void *a, const void *b){
70 const Minima *da = (const Minima *) a;
71 const Minima *db = (const Minima *) b;
73 return da->height - db->height;
76 #define FLAG_QPEL 1 //must be 1
80 static inline void init_ref(MpegEncContext *s, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
81 MotionEstContext * const c= &s->me;
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(MpegEncContext *s, int direct, int chroma){
100 return ((s->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
101 + (direct ? FLAG_DIRECT : 0)
102 + (chroma ? FLAG_CHROMA : 0);
105 static inline void init_mc(MpegEncContext *s, int size, int flags){
106 MotionEstContext * const c= &s->me;
108 /*FIXME s->no_rounding b_type*/
109 if(flags & FLAG_CHROMA){
110 if(s->no_rounding) c->chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1];
111 else c->chroma_hpel_put= &s->dsp.put_pixels_tab[size+1];
114 if(flags & FLAG_QPEL){
115 c->qpel_avg= &s->dsp.avg_qpel_pixels_tab[size];
116 if(s->no_rounding) c->qpel_put= &s->dsp.put_no_rnd_qpel_pixels_tab[size];
117 else c->qpel_put= &s->dsp.put_qpel_pixels_tab[size];
119 c->hpel_avg= &s->dsp.avg_pixels_tab[size];
120 if(s->no_rounding) c->hpel_put= &s->dsp.put_no_rnd_pixels_tab[size];
121 else c->hpel_put= &s->dsp.put_pixels_tab[size];
123 c->temp= c->scratchpad;
126 static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
127 const int size, const int h, int ref_index, int src_index,
128 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
129 MotionEstContext * const c= &s->me;
130 const int stride= c->stride;
131 const int uvstride= c->uvstride;
132 const int qpel= flags&FLAG_QPEL;
133 const int chroma= flags&FLAG_CHROMA;
134 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
135 const int hx= subx + (x<<(1+qpel));
136 const int hy= suby + (y<<(1+qpel));
137 uint8_t * const * const ref= c->ref[ref_index];
138 uint8_t * const * const src= c->src[src_index];
140 //FIXME check chroma 4mv, (no crashes ...)
141 if(flags&FLAG_DIRECT){
142 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
143 const int time_pp= s->pp_time;
144 const int time_pb= s->pb_time;
145 const int mask= 2*qpel+1;
146 if(s->mv_type==MV_TYPE_8X8){
149 int fx = c->direct_basis_mv[i][0] + hx;
150 int fy = c->direct_basis_mv[i][1] + hy;
151 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));
152 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));
153 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
154 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
156 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
158 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
159 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
161 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
162 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
166 int fx = c->direct_basis_mv[0][0] + hx;
167 int fy = c->direct_basis_mv[0][1] + hy;
168 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
169 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
170 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
171 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
174 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
175 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
176 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
177 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
178 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
179 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
180 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
181 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
183 assert((fx>>1) + 16*s->mb_x >= -16);
184 assert((fy>>1) + 16*s->mb_y >= -16);
185 assert((fx>>1) + 16*s->mb_x <= s->width);
186 assert((fy>>1) + 16*s->mb_y <= s->height);
187 assert((bx>>1) + 16*s->mb_x >= -16);
188 assert((by>>1) + 16*s->mb_y >= -16);
189 assert((bx>>1) + 16*s->mb_x <= s->width);
190 assert((by>>1) + 16*s->mb_y <= s->height);
192 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
193 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
196 d = cmp_func(s, c->temp, src[0], stride, 16);
203 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
209 uvdxy= (cx&1) + 2*(cy&1);
210 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
213 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
215 uvdxy= dxy | (x&1) | (2*(y&1));
217 d = cmp_func(s, c->temp, src[0], stride, h);
219 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
221 uvdxy= (x&1) + 2*(y&1);
224 uint8_t * const uvtemp= c->temp + 16*stride;
225 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
226 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
227 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
228 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
233 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
237 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
242 #include "motion_est_template.c"
244 static inline int get_penalty_factor(MpegEncContext *s, int type){
254 return s->qscale*s->qscale*2;
259 return (s->qscale*s->qscale*185 + 64)>>7;
263 void ff_init_me(MpegEncContext *s){
264 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
265 ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
266 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
267 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, s->avctx->mb_cmp);
269 s->me.flags = get_flags(s, 0, s->avctx->me_cmp &FF_CMP_CHROMA);
270 s->me.sub_flags= get_flags(s, 0, s->avctx->me_sub_cmp&FF_CMP_CHROMA);
271 s->me.mb_flags = get_flags(s, 0, s->avctx->mb_cmp &FF_CMP_CHROMA);
273 if(s->flags&CODEC_FLAG_QPEL){
274 s->me.sub_motion_search= qpel_motion_search;
276 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
277 s->me.sub_motion_search= hpel_motion_search;
278 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
279 && s->avctx-> me_cmp == FF_CMP_SAD
280 && s->avctx-> mb_cmp == FF_CMP_SAD)
281 s->me.sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
283 s->me.sub_motion_search= hpel_motion_search;
286 s->me.stride = s->linesize;
287 s->me.uvstride= s->uvlinesize;
289 s->me.stride = 16*s->mb_width + 32;
290 s->me.uvstride= 8*s->mb_width + 16;
295 static int pix_dev(uint8_t * pix, int line_size, int mean)
300 for (i = 0; i < 16; i++) {
301 for (j = 0; j < 16; j += 8) {
302 s += ABS(pix[0]-mean);
303 s += ABS(pix[1]-mean);
304 s += ABS(pix[2]-mean);
305 s += ABS(pix[3]-mean);
306 s += ABS(pix[4]-mean);
307 s += ABS(pix[5]-mean);
308 s += ABS(pix[6]-mean);
309 s += ABS(pix[7]-mean);
312 pix += line_size - 16;
318 static inline void no_motion_search(MpegEncContext * s,
319 int *mx_ptr, int *my_ptr)
321 *mx_ptr = 16 * s->mb_x;
322 *my_ptr = 16 * s->mb_y;
325 static int full_motion_search(MpegEncContext * s,
326 int *mx_ptr, int *my_ptr, int range,
327 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
329 int x1, y1, x2, y2, xx, yy, x, y;
335 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
347 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
351 for (y = y1; y <= y2; y++) {
352 for (x = x1; x <= x2; x++) {
353 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
357 (abs(x - xx) + abs(y - yy)) <
358 (abs(mx - xx) + abs(my - yy)))) {
370 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
371 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
372 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
379 static int log_motion_search(MpegEncContext * s,
380 int *mx_ptr, int *my_ptr, int range,
381 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
383 int x1, y1, x2, y2, xx, yy, x, y;
410 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
416 for (y = y1; y <= y2; y += range) {
417 for (x = x1; x <= x2; x += range) {
418 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
419 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
445 } while (range >= 1);
448 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
455 static int phods_motion_search(MpegEncContext * s,
456 int *mx_ptr, int *my_ptr, int range,
457 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
459 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
460 int mx, my, dminx, dminy;
486 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
497 for (x = x1; x <= x2; x += range) {
498 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
499 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
506 for (y = y1; y <= y2; y += range) {
507 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
508 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
534 } while (range >= 1);
537 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
540 /* half pixel search */
547 #define Z_THRESHOLD 256
549 #define CHECK_SAD_HALF_MV(suffix, x, y) \
551 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
552 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
553 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
556 static inline int sad_hpel_motion_search(MpegEncContext * s,
557 int *mx_ptr, int *my_ptr, int dmin,
558 int src_index, int ref_index,
561 const int penalty_factor= s->me.sub_penalty_factor;
564 int stride= s->me.stride;
565 const int flags= s->me.sub_flags;
578 pix = s->me.src[src_index][0];
582 ptr = s->me.ref[ref_index][0] + (my * stride) + mx;
586 if (mx > xmin && mx < xmax &&
587 my > ymin && my < ymax) {
590 const int index= (my<<ME_MAP_SHIFT) + mx;
591 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
592 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
593 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
594 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
604 CHECK_SAD_HALF_MV(y2 , 0, -1)
606 CHECK_SAD_HALF_MV(xy2, -1, -1)
608 CHECK_SAD_HALF_MV(xy2, +1, -1)
612 CHECK_SAD_HALF_MV(xy2, -1, +1)
614 CHECK_SAD_HALF_MV(x2 , -1, 0)
616 CHECK_SAD_HALF_MV(xy2, +1, -1)
618 CHECK_SAD_HALF_MV(xy2, -1, -1)
622 CHECK_SAD_HALF_MV(xy2, +1, +1)
624 CHECK_SAD_HALF_MV(x2 , +1, 0)
629 CHECK_SAD_HALF_MV(xy2, -1, -1)
633 CHECK_SAD_HALF_MV(xy2, +1, +1)
635 CHECK_SAD_HALF_MV(x2 , -1, 0)
636 CHECK_SAD_HALF_MV(xy2, -1, +1)
639 CHECK_SAD_HALF_MV(xy2, +1, -1)
643 CHECK_SAD_HALF_MV(xy2, -1, +1)
645 CHECK_SAD_HALF_MV(x2 , +1, 0)
646 CHECK_SAD_HALF_MV(xy2, +1, +1)
648 CHECK_SAD_HALF_MV(y2 , 0, +1)
663 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
665 const int xy= s->mb_x + s->mb_y*s->mb_stride;
667 s->p_mv_table[xy][0] = mx;
668 s->p_mv_table[xy][1] = my;
670 /* has allready been set to the 4 MV if 4MV is done */
672 int mot_xy= s->block_index[0];
674 s->current_picture.motion_val[0][mot_xy ][0]= mx;
675 s->current_picture.motion_val[0][mot_xy ][1]= my;
676 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
677 s->current_picture.motion_val[0][mot_xy+1][1]= my;
679 mot_xy += s->b8_stride;
680 s->current_picture.motion_val[0][mot_xy ][0]= mx;
681 s->current_picture.motion_val[0][mot_xy ][1]= my;
682 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
683 s->current_picture.motion_val[0][mot_xy+1][1]= my;
688 * get fullpel ME search limits.
690 static inline void get_limits(MpegEncContext *s, int x, int y)
693 if(s->avctx->me_range) s->me.range= s->avctx->me_range >> 1;
694 else s->me.range= 16;
696 if (s->unrestricted_mv) {
697 s->me.xmin = - x - 16;
698 s->me.ymin = - y - 16;
699 s->me.xmax = - x + s->mb_width *16;
700 s->me.ymax = - y + s->mb_height*16;
704 s->me.xmax = - x + s->mb_width *16 - 16;
705 s->me.ymax = - y + s->mb_height*16 - 16;
709 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
711 MotionEstContext * const c= &s->me;
716 int dmin_sum=0, mx4_sum=0, my4_sum=0;
718 const int stride= s->linesize;
719 const int uvstride= s->uvlinesize;
720 uint8_t *mv_penalty= s->me.current_mv_penalty;
722 c->ref[1][0] = c->ref[0][0] + 8;
723 c->ref[2][0] = c->ref[0][0] + 8*stride;
724 c->ref[3][0] = c->ref[2][0] + 8;
725 c->src[1][0] = c->src[0][0] + 8;
726 c->src[2][0] = c->src[0][0] + 8*stride;
727 c->src[3][0] = c->src[2][0] + 8;
728 init_mc(s, size, c->flags);
730 for(block=0; block<4; block++){
732 int pred_x4, pred_y4;
734 static const int off[4]= {2, 1, 1, -1};
735 const int mot_stride = s->b8_stride;
736 const int mot_xy = s->block_index[block];
738 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
739 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
741 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
743 /* special case for first line */
744 if (s->first_slice_line && block<2) {
745 s->me.pred_x= pred_x4= P_LEFT[0];
746 s->me.pred_y= pred_y4= P_LEFT[1];
748 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
749 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
750 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
751 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
752 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
753 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
754 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
755 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
757 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
758 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
760 s->me.pred_x= pred_x4 = P_MEDIAN[0];
761 s->me.pred_y= pred_y4 = P_MEDIAN[1];
766 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
768 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
770 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
772 const int offset= ((block&1) + (block>>1)*stride)*8;
773 uint8_t *dest_y = s->me.scratchpad + offset;
774 if(s->quarter_sample){
775 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
776 dxy = ((my4 & 3) << 2) | (mx4 & 3);
779 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
781 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
783 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
784 dxy = ((my4 & 1) << 1) | (mx4 & 1);
787 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
789 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
791 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
795 if(s->quarter_sample){
803 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
804 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
806 if(mx4 != mx || my4 != my) same=0;
812 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
813 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, s->me.scratchpad, stride, 16);
816 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
821 mx= ff_h263_round_chroma(mx4_sum);
822 my= ff_h263_round_chroma(my4_sum);
823 dxy = ((my & 1) << 1) | (mx & 1);
825 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
828 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
829 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
831 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
832 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
835 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad , s->uvlinesize, 8);
836 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad+8, s->uvlinesize, 8);
842 switch(s->avctx->mb_cmp&0xFF){
844 return dmin_sum+ 32*s->qscale*s->qscale;*/
848 return dmin_sum+ 11*s->me.mb_penalty_factor;
852 static int interlaced_search(MpegEncContext *s, int ref_index,
853 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my)
855 MotionEstContext * const c= &s->me;
860 uint8_t * const mv_penalty= c->current_mv_penalty;
862 const int stride= 2*s->linesize;
863 const int uvstride= 2*s->uvlinesize;
865 const int mot_stride= s->mb_stride;
866 const int xy= s->mb_x + s->mb_y*mot_stride;
872 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
873 c->src[1][0] = c->src[0][0] + s->linesize;
874 if(c->flags & FLAG_CHROMA){
875 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
876 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
877 c->src[1][1] = c->src[0][1] + s->uvlinesize;
878 c->src[1][2] = c->src[0][2] + s->uvlinesize;
880 init_mc(s, size, c->flags);
882 for(block=0; block<2; block++){
884 int best_dmin= INT_MAX;
887 for(field_select=0; field_select<2; field_select++){
888 int dmin, mx_i, my_i;
889 int16_t (*mv_table)[2]= mv_tables[block][field_select];
891 P_LEFT[0] = mv_table[xy - 1][0];
892 P_LEFT[1] = mv_table[xy - 1][1];
893 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
895 s->me.pred_x= P_LEFT[0];
896 s->me.pred_y= P_LEFT[1];
898 if(!s->first_slice_line){
899 P_TOP[0] = mv_table[xy - mot_stride][0];
900 P_TOP[1] = mv_table[xy - mot_stride][1];
901 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
902 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
903 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
904 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
905 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
906 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
908 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
909 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
911 P_MV1[0]= mx; //FIXME not correct if block != field_select
914 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
916 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
918 mv_table[xy][0]= mx_i;
919 mv_table[xy][1]= my_i;
921 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
925 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
926 dxy = ((my_i & 1) << 1) | (mx_i & 1);
929 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
931 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
933 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
934 dmin+= (mv_penalty[mx_i-s->me.pred_x] + mv_penalty[my_i-s->me.pred_y] + 1)*c->mb_penalty_factor;
936 dmin+= c->mb_penalty_factor; //field_select bits
938 dmin += field_select != block; //slightly prefer same field
940 if(dmin < best_dmin){
942 best_field= field_select;
946 int16_t (*mv_table)[2]= mv_tables[block][best_field];
948 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
949 if(mv_table[xy][1]&1) same=0;
950 if(mv_table[xy][1]*2 != my) same=0;
951 if(best_field != block) same=0;
954 field_select_tables[block][xy]= best_field;
955 dmin_sum += best_dmin;
966 switch(s->avctx->mb_cmp&0xFF){
968 return dmin_sum+ 32*s->qscale*s->qscale;*/
972 return dmin_sum+ 11*c->mb_penalty_factor;
976 void ff_estimate_p_frame_motion(MpegEncContext * s,
979 MotionEstContext * const c= &s->me;
981 int sum, varc, vard, mx, my, dmin, xx, yy;
983 const int shift= 1+s->quarter_sample;
985 Picture * const pic= &s->current_picture;
987 init_ref(s, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
988 init_mc(s, 0, c->flags);
990 assert(s->quarter_sample==0 || s->quarter_sample==1);
991 assert(s->linesize == s->me.stride);
992 assert(s->uvlinesize == s->me.uvstride);
994 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
995 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
996 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
997 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
999 get_limits(s, 16*mb_x, 16*mb_y);
1002 switch(s->me_method) {
1005 no_motion_search(s, &mx, &my);
1012 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1017 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1022 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1030 const int mot_stride = s->b8_stride;
1031 const int mot_xy = s->block_index[0];
1033 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1034 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1036 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1038 if(!s->first_slice_line) {
1039 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1040 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1041 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1042 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1043 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
1044 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1045 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1047 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1048 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1050 if(s->out_format == FMT_H263){
1051 c->pred_x = P_MEDIAN[0];
1052 c->pred_y = P_MEDIAN[1];
1053 }else { /* mpeg1 at least */
1054 c->pred_x= P_LEFT[0];
1055 c->pred_y= P_LEFT[1];
1058 c->pred_x= P_LEFT[0];
1059 c->pred_y= P_LEFT[1];
1063 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift);
1068 /* intra / predictive decision */
1073 /* At this point (mx,my) are full-pell and the relative displacement */
1074 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1076 sum = s->dsp.pix_sum(pix, s->linesize);
1078 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1079 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
1081 //printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
1082 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1083 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1084 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1085 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1086 s->mb_var_sum_temp += varc;
1087 s->mc_mb_var_sum_temp += vard;
1088 //printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
1091 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1092 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1094 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1095 if (vard <= 64 || vard < varc)
1096 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1098 s->scene_change_score+= s->qscale;
1100 if (vard*2 + 200 > varc)
1101 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1102 if (varc*2 + 200 > vard){
1103 mb_type|= CANDIDATE_MB_TYPE_INTER;
1104 s->me.sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1105 if(s->flags&CODEC_FLAG_MV0)
1107 mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
1112 if((s->flags&CODEC_FLAG_4MV)
1113 && !s->me.skip && varc>50 && vard>10){
1114 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1115 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1117 set_p_mv_tables(s, mx, my, 0);
1119 set_p_mv_tables(s, mx, my, 1);
1120 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1121 && !s->me.skip){ //FIXME varc/d checks
1122 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my) < INT_MAX)
1123 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1127 mb_type= CANDIDATE_MB_TYPE_INTER;
1129 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1130 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1131 dmin= get_mb_score(s, mx, my, 0, 0);
1133 if((s->flags&CODEC_FLAG_4MV)
1134 && !s->me.skip && varc>50 && vard>10){
1135 int dmin4= h263_mv4_search(s, mx, my, shift);
1137 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1141 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1142 && !s->me.skip){ //FIXME varc/d checks
1143 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my);
1145 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1150 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1151 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1153 /* get intra luma score */
1154 if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1155 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
1157 int mean= (sum+128)>>8;
1160 for(i=0; i<16; i++){
1161 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean;
1162 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean;
1163 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean;
1164 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean;
1167 intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize, 16);
1170 /* get chroma score */
1171 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1176 if(s->out_format == FMT_H263){
1177 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1179 mean= (s->last_dc[i] + 4)>>3;
1181 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1185 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean;
1186 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean;
1189 intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize);
1193 intra_score += s->me.mb_penalty_factor*16;
1195 if(intra_score < dmin){
1196 mb_type= CANDIDATE_MB_TYPE_INTRA;
1197 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1199 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1201 if (vard <= 64 || vard < varc) { //FIXME
1202 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1204 s->scene_change_score+= s->qscale;
1208 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1211 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1214 MotionEstContext * const c= &s->me;
1217 const int shift= 1+s->quarter_sample;
1218 const int xy= mb_x + mb_y*s->mb_stride;
1219 init_ref(s, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1220 init_mc(s, 0, c->flags);
1222 assert(s->quarter_sample==0 || s->quarter_sample==1);
1224 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1225 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1227 get_limits(s, 16*mb_x, 16*mb_y);
1230 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1231 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1233 if(P_LEFT[0] < (s->me.xmin<<shift)) P_LEFT[0] = (s->me.xmin<<shift);
1235 /* special case for first line */
1236 if (s->first_slice_line) {
1237 c->pred_x= P_LEFT[0];
1238 c->pred_y= P_LEFT[1];
1239 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1240 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1242 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1243 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1244 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1245 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1246 if(P_TOP[1] < (s->me.ymin<<shift)) P_TOP[1] = (s->me.ymin<<shift);
1247 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
1248 if(P_TOPRIGHT[1] < (s->me.ymin<<shift)) P_TOPRIGHT[1]= (s->me.ymin<<shift);
1250 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1251 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1253 c->pred_x = P_MEDIAN[0];
1254 c->pred_y = P_MEDIAN[1];
1257 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift);
1259 s->p_mv_table[xy][0] = mx<<shift;
1260 s->p_mv_table[xy][1] = my<<shift;
1265 static int ff_estimate_motion_b(MpegEncContext * s,
1266 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1270 const int shift= 1+s->quarter_sample;
1271 const int mot_stride = s->mb_stride;
1272 const int mot_xy = mb_y*mot_stride + mb_x;
1273 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1276 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1277 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1278 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1279 s->me.current_mv_penalty= mv_penalty;
1281 get_limits(s, 16*mb_x, 16*mb_y);
1283 switch(s->me_method) {
1286 no_motion_search(s, &mx, &my);
1293 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1298 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1303 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1311 P_LEFT[0] = mv_table[mot_xy - 1][0];
1312 P_LEFT[1] = mv_table[mot_xy - 1][1];
1314 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1316 /* special case for first line */
1317 if (!s->first_slice_line) {
1318 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1319 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1320 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1321 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1322 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1]= (s->me.ymax<<shift);
1323 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1324 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1326 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1327 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1329 s->me.pred_x= P_LEFT[0];
1330 s->me.pred_y= P_LEFT[1];
1333 if(mv_table == s->b_forw_mv_table){
1334 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1336 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1339 dmin = epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale);
1344 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1346 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1347 dmin= get_mb_score(s, mx, my, 0, ref_index);
1349 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1350 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1351 mv_table[mot_xy][0]= mx;
1352 mv_table[mot_xy][1]= my;
1357 static inline int check_bidir_mv(MpegEncContext * s,
1358 int motion_fx, int motion_fy,
1359 int motion_bx, int motion_by,
1360 int pred_fx, int pred_fy,
1361 int pred_bx, int pred_by,
1365 //FIXME better f_code prediction (max mv & distance)
1367 MotionEstContext * const c= &s->me;
1368 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1369 int stride= s->me.stride;
1370 int uvstride= s->me.uvstride;
1371 uint8_t *dest_y = s->me.scratchpad;
1376 uint8_t **src_data= c->src[0];
1377 uint8_t **ref_data= c->ref[0];
1378 uint8_t **ref2_data= c->ref[2];
1380 if(s->quarter_sample){
1381 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1382 src_x = motion_fx >> 2;
1383 src_y = motion_fy >> 2;
1385 ptr = ref_data[0] + (src_y * stride) + src_x;
1386 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1388 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1389 src_x = motion_bx >> 2;
1390 src_y = motion_by >> 2;
1392 ptr = ref2_data[0] + (src_y * stride) + src_x;
1393 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1395 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1396 src_x = motion_fx >> 1;
1397 src_y = motion_fy >> 1;
1399 ptr = ref_data[0] + (src_y * stride) + src_x;
1400 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1402 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1403 src_x = motion_bx >> 1;
1404 src_y = motion_by >> 1;
1406 ptr = ref2_data[0] + (src_y * stride) + src_x;
1407 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1410 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1411 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1412 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1414 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1421 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1422 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1424 const int mot_stride = s->mb_stride;
1425 const int xy = mb_y *mot_stride + mb_x;
1427 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1428 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1429 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1430 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1431 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1432 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1433 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1434 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1436 //FIXME do refinement and add flag
1438 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1439 motion_bx, motion_by,
1447 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1450 const int mot_stride = s->mb_stride;
1451 const int mot_xy = mb_y*mot_stride + mb_x;
1452 const int shift= 1+s->quarter_sample;
1454 const int time_pp= s->pp_time;
1455 const int time_pb= s->pb_time;
1456 int mx, my, xmin, xmax, ymin, ymax;
1457 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1459 s->me.current_mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1460 ymin= xmin=(-32)>>shift;
1461 ymax= xmax= 31>>shift;
1463 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1464 s->mv_type= MV_TYPE_8X8;
1466 s->mv_type= MV_TYPE_16X16;
1470 int index= s->block_index[i];
1473 s->me.co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1474 s->me.co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1475 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1476 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1477 // s->me.direct_basis_mv[1][i][0]= s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1478 // s->me.direct_basis_mv[1][i][1]= s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1480 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1481 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1482 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1484 xmax= FFMIN(xmax, s->width - max);
1485 xmin= FFMAX(xmin, - 16 - min);
1487 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1488 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1489 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1491 ymax= FFMIN(ymax, s->height - max);
1492 ymin= FFMAX(ymin, - 16 - min);
1494 if(s->mv_type == MV_TYPE_16X16) break;
1497 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1499 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1500 s->b_direct_mv_table[mot_xy][0]= 0;
1501 s->b_direct_mv_table[mot_xy][1]= 0;
1503 return 256*256*256*64;
1510 s->me.flags |= FLAG_DIRECT;
1511 s->me.sub_flags |= FLAG_DIRECT;
1515 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1516 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1518 /* special case for first line */
1519 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
1520 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1521 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1522 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1523 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1525 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1526 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1529 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift));
1530 if(s->me.sub_flags&FLAG_QPEL)
1531 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1533 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1535 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1536 dmin= get_mb_score(s, mx, my, 0, 0);
1538 get_limits(s, 16*mb_x, 16*mb_y); //restore s->me.?min/max, maybe not needed
1540 s->b_direct_mv_table[mot_xy][0]= mx;
1541 s->b_direct_mv_table[mot_xy][1]= my;
1542 s->me.flags &= ~FLAG_DIRECT;
1543 s->me.sub_flags &= ~FLAG_DIRECT;
1548 void ff_estimate_b_frame_motion(MpegEncContext * s,
1551 const int penalty_factor= s->me.mb_penalty_factor;
1552 int fmin, bmin, dmin, fbmin, bimin, fimin;
1554 init_ref(s, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1555 init_mc(s, 0, s->me.flags);
1558 if (s->codec_id == CODEC_ID_MPEG4)
1559 dmin= direct_search(s, mb_x, mb_y);
1562 //FIXME penalty stuff for non mpeg4
1564 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1567 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1568 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1571 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1572 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1574 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1575 const int xy = mb_y*s->mb_stride + mb_x;
1577 //FIXME mb type penalty
1579 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1580 fimin= interlaced_search(s, 0,
1581 s->b_field_mv_table[0], s->b_field_select_table[0],
1582 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1583 s->me.current_mv_penalty= s->me.mv_penalty[s->b_code] + MAX_MV;
1584 bimin= interlaced_search(s, 2,
1585 s->b_field_mv_table[1], s->b_field_select_table[1],
1586 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1]);
1588 fimin= bimin= INT_MAX;
1592 type = CANDIDATE_MB_TYPE_FORWARD;
1596 type = CANDIDATE_MB_TYPE_DIRECT;
1600 type= CANDIDATE_MB_TYPE_BACKWARD;
1604 type= CANDIDATE_MB_TYPE_BIDIR;
1608 type= CANDIDATE_MB_TYPE_FORWARD_I;
1612 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1615 score= ((unsigned)(score*score + 128*256))>>16;
1616 s->mc_mb_var_sum_temp += score;
1617 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1620 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1621 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1623 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1625 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1626 if(fimin < INT_MAX && bimin < INT_MAX){
1627 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1629 //FIXME something smarter
1630 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1632 if(s->out_format == FMT_MPEG1)
1633 type |= CANDIDATE_MB_TYPE_INTRA;
1637 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1640 /* find best f_code for ME which do unlimited searches */
1641 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1643 if(s->me_method>=ME_EPZS){
1646 uint8_t * fcode_tab= s->fcode_tab;
1648 int best_score=-10000000;
1650 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1652 for(y=0; y<s->mb_height; y++){
1654 int xy= y*s->mb_stride;
1655 for(x=0; x<s->mb_width; x++){
1656 if(s->mb_type[xy] & type){
1657 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1658 fcode_tab[mv_table[xy][1] + MAX_MV]);
1661 for(j=0; j<fcode && j<8; j++){
1662 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1671 if(score[i] > best_score){
1672 best_score= score[i];
1675 // printf("%d %d\n", i, score[i]);
1678 // printf("fcode: %d type: %d\n", i, s->pict_type);
1680 /* for(i=0; i<=MAX_FCODE; i++){
1681 printf("%d ", mv_num[i]);
1689 void ff_fix_long_p_mvs(MpegEncContext * s)
1691 const int f_code= s->f_code;
1693 assert(s->pict_type==P_TYPE);
1695 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1697 if(s->msmpeg4_version) range= 16;
1699 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1701 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1702 if(s->flags&CODEC_FLAG_4MV){
1703 const int wrap= s->b8_stride;
1705 /* clip / convert to intra 8x8 type MVs */
1706 for(y=0; y<s->mb_height; y++){
1708 int i= y*s->mb_stride;
1711 for(x=0; x<s->mb_width; x++){
1712 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1714 for(block=0; block<4; block++){
1715 int off= (block& 1) + (block>>1)*wrap;
1716 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1717 int my= s->current_picture.motion_val[0][ xy + off ][1];
1719 if( mx >=range || mx <-range
1720 || my >=range || my <-range){
1721 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1722 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1723 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1736 * @param truncate 1 for truncation, 0 for using intra
1738 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1739 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1741 int y, h_range, v_range;
1743 // RAL: 8 in MPEG-1, 16 in MPEG-4
1744 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1746 if(s->msmpeg4_version) range= 16;
1747 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1750 v_range= field_select_table ? range>>1 : range;
1752 /* clip / convert to intra 16x16 type MVs */
1753 for(y=0; y<s->mb_height; y++){
1755 int xy= y*s->mb_stride;
1756 for(x=0; x<s->mb_width; x++){
1757 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1758 if(field_select_table==NULL || field_select_table[xy] == field_select){
1759 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1760 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1763 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1764 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1765 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1766 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1768 s->mb_type[xy] &= ~type;
1769 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;