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 pred_x, int pred_y, uint8_t *src_data[3],
50 uint8_t *ref_data[6], int stride, int uvstride,
51 int size, int h, uint8_t * const mv_penalty);
53 static inline int update_map_generation(MpegEncContext * s)
55 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
56 if(s->me.map_generation==0){
57 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
58 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
60 return s->me.map_generation;
63 /* shape adaptive search stuff */
64 typedef struct Minima{
70 static int minima_cmp(const void *a, const void *b){
71 const Minima *da = (const Minima *) a;
72 const Minima *db = (const Minima *) b;
74 return da->height - db->height;
78 #define RENAME(a) simple_ ## a
80 #define CMP(d, x, y, size)\
81 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride, h);
83 #define CMP_HPEL(d, dx, dy, x, y, size)\
85 const int dxy= (dx) + 2*(dy);\
86 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, h);\
87 d = cmp_sub(s, s->me.scratchpad, src_y, stride, h);\
91 #define CMP_QPEL(d, dx, dy, x, y, size)\
93 const int dxy= (dx) + 4*(dy);\
94 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
95 d = cmp_sub(s, s->me.scratchpad, src_y, stride, h);\
98 #include "motion_est_template.c"
106 #define RENAME(a) simple_chroma_ ## a
108 #define CMP(d, x, y, size)\
109 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride, h);\
111 int dxy= ((x)&1) + 2*((y)&1);\
112 int c= ((x)>>1) + ((y)>>1)*uvstride;\
114 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, h>>1);\
115 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride, h>>1);\
116 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, h>>1);\
117 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride, h>>1);\
120 #define CMP_HPEL(d, dx, dy, x, y, size)\
122 const int dxy= (dx) + 2*(dy);\
123 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, h);\
124 d = cmp_sub(s, s->me.scratchpad, src_y, stride, h);\
126 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
127 int c= ((x)>>1) + ((y)>>1)*uvstride;\
128 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, h>>1);\
129 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride, h>>1);\
130 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, h>>1);\
131 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride, h>>1);\
135 #define CMP_QPEL(d, dx, dy, x, y, size)\
137 const int dxy= (dx) + 4*(dy);\
138 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
139 d = cmp_sub(s, s->me.scratchpad, src_y, stride, h);\
142 int cx= (4*(x) + (dx))/2;\
143 int cy= (4*(y) + (dy))/2;\
146 cxy= (cx&1) + 2*(cy&1);\
147 c= ((cx)>>1) + ((cy)>>1)*uvstride;\
148 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, h>>1);\
149 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride, h>>1);\
150 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, h>>1);\
151 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride, h>>1);\
155 #include "motion_est_template.c"
162 /* SIMPLE DIRECT HPEL */
163 #define RENAME(a) simple_direct_hpel_ ## a
164 //FIXME precalc divisions stuff
166 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
167 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
168 const int hx= 2*(x) + (dx);\
169 const int hy= 2*(y) + (dy);\
170 if(s->mv_type==MV_TYPE_8X8){\
173 int fx = s->me.direct_basis_mv[i][0] + hx;\
174 int fy = s->me.direct_basis_mv[i][1] + hy;\
175 int bx = hx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
176 int by = hy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
177 int fxy= (fx&1) + 2*(fy&1);\
178 int bxy= (bx&1) + 2*(by&1);\
180 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
181 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
182 hpel_avg[1][bxy](dst, (ref_data[3]) + (bx>>1) + (by>>1)*(stride), stride, 8);\
185 int fx = s->me.direct_basis_mv[0][0] + hx;\
186 int fy = s->me.direct_basis_mv[0][1] + hy;\
187 int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\
188 int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\
189 int fxy= (fx&1) + 2*(fy&1);\
190 int bxy= (bx&1) + 2*(by&1);\
192 assert((fx>>1) + 16*s->mb_x >= -16);\
193 assert((fy>>1) + 16*s->mb_y >= -16);\
194 assert((fx>>1) + 16*s->mb_x <= s->width);\
195 assert((fy>>1) + 16*s->mb_y <= s->height);\
196 assert((bx>>1) + 16*s->mb_x >= -16);\
197 assert((by>>1) + 16*s->mb_y >= -16);\
198 assert((bx>>1) + 16*s->mb_x <= s->width);\
199 assert((by>>1) + 16*s->mb_y <= s->height);\
201 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
202 hpel_avg[0][bxy](s->me.scratchpad, (ref_data[3]) + (bx>>1) + (by>>1)*(stride), stride, 16);\
204 d = cmp_func(s, s->me.scratchpad, src_y, stride, 16);\
209 #define CMP_HPEL(d, dx, dy, x, y, size)\
210 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
212 #define CMP(d, x, y, size)\
213 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
215 #include "motion_est_template.c"
223 /* SIMPLE DIRECT QPEL */
224 #define RENAME(a) simple_direct_qpel_ ## a
226 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
227 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
228 const int qx= 4*(x) + (dx);\
229 const int qy= 4*(y) + (dy);\
230 if(s->mv_type==MV_TYPE_8X8){\
233 int fx = s->me.direct_basis_mv[i][0] + qx;\
234 int fy = s->me.direct_basis_mv[i][1] + qy;\
235 int bx = qx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
236 int by = qy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
237 int fxy= (fx&3) + 4*(fy&3);\
238 int bxy= (bx&3) + 4*(by&3);\
240 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
241 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
242 qpel_avg[1][bxy](dst, (ref_data[3]) + (bx>>2) + (by>>2)*(stride), stride);\
245 int fx = s->me.direct_basis_mv[0][0] + qx;\
246 int fy = s->me.direct_basis_mv[0][1] + qy;\
247 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
248 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
249 int fxy= (fx&3) + 4*(fy&3);\
250 int bxy= (bx&3) + 4*(by&3);\
252 qpel_put[1][fxy](s->me.scratchpad , (ref_y ) + (fx>>2) + (fy>>2)*(stride) , stride);\
253 qpel_put[1][fxy](s->me.scratchpad + 8 , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 , stride);\
254 qpel_put[1][fxy](s->me.scratchpad + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8*stride, stride);\
255 qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
256 qpel_avg[1][bxy](s->me.scratchpad , (ref_data[3]) + (bx>>2) + (by>>2)*(stride) , stride);\
257 qpel_avg[1][bxy](s->me.scratchpad + 8 , (ref_data[3]) + (bx>>2) + (by>>2)*(stride) + 8 , stride);\
258 qpel_avg[1][bxy](s->me.scratchpad + 8*stride, (ref_data[3]) + (bx>>2) + (by>>2)*(stride) + 8*stride, stride);\
259 qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref_data[3]) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
261 d = cmp_func(s, s->me.scratchpad, src_y, stride, 16);\
266 #define CMP_QPEL(d, dx, dy, x, y, size)\
267 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
269 #define CMP(d, x, y, size)\
270 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
272 #include "motion_est_template.c"
280 static inline int get_penalty_factor(MpegEncContext *s, int type){
290 return s->qscale*s->qscale*2;
295 return (s->qscale*s->qscale*185 + 64)>>7;
299 void ff_init_me(MpegEncContext *s){
300 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
301 ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp);
302 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
303 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, s->avctx->mb_cmp);
305 if(s->flags&CODEC_FLAG_QPEL){
306 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
307 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
309 s->me.sub_motion_search= simple_qpel_motion_search;
311 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
312 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
313 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
314 && s->avctx-> me_cmp == FF_CMP_SAD
315 && s->avctx-> mb_cmp == FF_CMP_SAD)
316 s->me.sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
318 s->me.sub_motion_search= simple_hpel_motion_search;
321 if(s->avctx->me_cmp&FF_CMP_CHROMA){
322 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
323 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
324 s->me.motion_search[4]= simple_chroma_epzs_motion_search2;
326 s->me.motion_search[0]= simple_epzs_motion_search;
327 s->me.motion_search[1]= simple_epzs_motion_search4;
328 s->me.motion_search[4]= simple_epzs_motion_search2;
331 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
332 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
334 s->me.pre_motion_search= simple_epzs_motion_search;
337 if(s->flags&CODEC_FLAG_QPEL){
338 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
339 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
341 s->me.get_mb_score= simple_qpel_get_mb_score;
343 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
344 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
346 s->me.get_mb_score= simple_hpel_get_mb_score;
351 static int pix_dev(uint8_t * pix, int line_size, int mean)
356 for (i = 0; i < 16; i++) {
357 for (j = 0; j < 16; j += 8) {
358 s += ABS(pix[0]-mean);
359 s += ABS(pix[1]-mean);
360 s += ABS(pix[2]-mean);
361 s += ABS(pix[3]-mean);
362 s += ABS(pix[4]-mean);
363 s += ABS(pix[5]-mean);
364 s += ABS(pix[6]-mean);
365 s += ABS(pix[7]-mean);
368 pix += line_size - 16;
374 static inline void no_motion_search(MpegEncContext * s,
375 int *mx_ptr, int *my_ptr)
377 *mx_ptr = 16 * s->mb_x;
378 *my_ptr = 16 * s->mb_y;
381 static int full_motion_search(MpegEncContext * s,
382 int *mx_ptr, int *my_ptr, int range,
383 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
385 int x1, y1, x2, y2, xx, yy, x, y;
391 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
403 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
407 for (y = y1; y <= y2; y++) {
408 for (x = x1; x <= x2; x++) {
409 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
413 (abs(x - xx) + abs(y - yy)) <
414 (abs(mx - xx) + abs(my - yy)))) {
426 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
427 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
428 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
435 static int log_motion_search(MpegEncContext * s,
436 int *mx_ptr, int *my_ptr, int range,
437 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
439 int x1, y1, x2, y2, xx, yy, x, y;
466 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
472 for (y = y1; y <= y2; y += range) {
473 for (x = x1; x <= x2; x += range) {
474 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
475 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
501 } while (range >= 1);
504 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
511 static int phods_motion_search(MpegEncContext * s,
512 int *mx_ptr, int *my_ptr, int range,
513 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
515 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
516 int mx, my, dminx, dminy;
542 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
553 for (x = x1; x <= x2; x += range) {
554 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
555 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
562 for (y = y1; y <= y2; y += range) {
563 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
564 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
590 } while (range >= 1);
593 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
596 /* half pixel search */
603 #define Z_THRESHOLD 256
605 #define CHECK_SAD_HALF_MV(suffix, x, y) \
607 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
608 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
609 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
612 static inline int sad_hpel_motion_search(MpegEncContext * s,
613 int *mx_ptr, int *my_ptr, int dmin,
614 int pred_x, int pred_y, uint8_t *src_data[3],
615 uint8_t *ref_data[6], int stride, int uvstride,
616 int size, int h, uint8_t * const mv_penalty)
618 uint32_t *score_map= s->me.score_map;
619 const int penalty_factor= s->me.sub_penalty_factor;
622 const int xmin= s->me.xmin;
623 const int ymin= s->me.ymin;
624 const int xmax= s->me.xmax;
625 const int ymax= s->me.ymax;
639 ptr = ref_data[0] + (my * stride) + mx;
643 if (mx > xmin && mx < xmax &&
644 my > ymin && my < ymax) {
647 const int index= (my<<ME_MAP_SHIFT) + mx;
648 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
649 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
650 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
651 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
661 CHECK_SAD_HALF_MV(y2 , 0, -1)
663 CHECK_SAD_HALF_MV(xy2, -1, -1)
665 CHECK_SAD_HALF_MV(xy2, +1, -1)
669 CHECK_SAD_HALF_MV(xy2, -1, +1)
671 CHECK_SAD_HALF_MV(x2 , -1, 0)
673 CHECK_SAD_HALF_MV(xy2, +1, -1)
675 CHECK_SAD_HALF_MV(xy2, -1, -1)
679 CHECK_SAD_HALF_MV(xy2, +1, +1)
681 CHECK_SAD_HALF_MV(x2 , +1, 0)
686 CHECK_SAD_HALF_MV(xy2, -1, -1)
690 CHECK_SAD_HALF_MV(xy2, +1, +1)
692 CHECK_SAD_HALF_MV(x2 , -1, 0)
693 CHECK_SAD_HALF_MV(xy2, -1, +1)
696 CHECK_SAD_HALF_MV(xy2, +1, -1)
700 CHECK_SAD_HALF_MV(xy2, -1, +1)
702 CHECK_SAD_HALF_MV(x2 , +1, 0)
703 CHECK_SAD_HALF_MV(xy2, +1, +1)
705 CHECK_SAD_HALF_MV(y2 , 0, +1)
720 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
722 const int xy= s->mb_x + s->mb_y*s->mb_stride;
724 s->p_mv_table[xy][0] = mx;
725 s->p_mv_table[xy][1] = my;
727 /* has allready been set to the 4 MV if 4MV is done */
729 int mot_xy= s->block_index[0];
731 s->current_picture.motion_val[0][mot_xy ][0]= mx;
732 s->current_picture.motion_val[0][mot_xy ][1]= my;
733 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
734 s->current_picture.motion_val[0][mot_xy+1][1]= my;
736 mot_xy += s->block_wrap[0];
737 s->current_picture.motion_val[0][mot_xy ][0]= mx;
738 s->current_picture.motion_val[0][mot_xy ][1]= my;
739 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
740 s->current_picture.motion_val[0][mot_xy+1][1]= my;
745 * get fullpel ME search limits.
747 static inline void get_limits(MpegEncContext *s, int x, int y)
750 if(s->avctx->me_range) s->me.range= s->avctx->me_range >> 1;
751 else s->me.range= 16;
753 if (s->unrestricted_mv) {
754 s->me.xmin = - x - 16;
755 s->me.ymin = - y - 16;
756 s->me.xmax = - x + s->mb_width *16;
757 s->me.ymax = - y + s->mb_height*16;
761 s->me.xmax = - x + s->mb_width *16 - 16;
762 s->me.ymax = - y + s->mb_height*16 - 16;
766 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
772 int dmin_sum=0, mx4_sum=0, my4_sum=0;
773 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
775 const int stride= s->linesize;
776 const int uvstride= s->uvlinesize;
777 const int xmin= s->me.xmin;
778 const int ymin= s->me.ymin;
779 const int xmax= s->me.xmax;
780 const int ymax= s->me.ymax;
782 for(block=0; block<4; block++){
784 int pred_x4, pred_y4;
786 static const int off[4]= {2, 1, 1, -1};
787 const int mot_stride = s->block_wrap[0];
788 const int mot_xy = s->block_index[block];
789 const int block_x= (block&1);
790 const int block_y= (block>>1);
791 uint8_t *src_data[3]= {
792 s->new_picture.data[0] + 8*(2*s->mb_x + block_x) + stride *8*(2*s->mb_y + block_y), //FIXME chroma?
793 s->new_picture.data[1] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y),
794 s->new_picture.data[2] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y)
796 uint8_t *ref_data[3]= {
797 s->last_picture.data[0] + 8*(2*s->mb_x + block_x) + stride *8*(2*s->mb_y + block_y), //FIXME chroma?
798 s->last_picture.data[1] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y),
799 s->last_picture.data[2] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y)
802 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
803 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
805 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
807 /* special case for first line */
808 if (s->mb_y == 0 && block<2) {
812 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
813 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
814 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
815 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
816 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
817 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
818 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
819 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
821 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
822 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
824 // if(s->out_format == FMT_H263){
825 pred_x4 = P_MEDIAN[0];
826 pred_y4 = P_MEDIAN[1];
828 }else { /* mpeg1 at least */
837 dmin4 = s->me.motion_search[1](s, &mx4, &my4, P, pred_x4, pred_y4,
838 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
840 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4,
841 pred_x4, pred_y4, src_data, ref_data, stride, uvstride, size, h, mv_penalty);
843 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]
844 && s->avctx->mb_decision == FF_MB_DECISION_SIMPLE){
846 const int offset= ((block&1) + (block>>1)*stride)*8;
847 uint8_t *dest_y = s->me.scratchpad + offset;
849 if(s->quarter_sample){
850 uint8_t *ref= ref_data[0] + (mx4>>2) + (my4>>2)*stride;
851 dxy = ((my4 & 3) << 2) | (mx4 & 3);
854 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , s->linesize);
856 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
858 uint8_t *ref= ref_data[0] + (mx4>>1) + (my4>>1)*stride;
859 dxy = ((my4 & 1) << 1) | (mx4 & 1);
862 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
864 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
866 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
870 if(s->quarter_sample){
878 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
879 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
881 if(mx4 != mx || my4 != my) same=0;
887 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
888 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);
891 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
896 mx= ff_h263_round_chroma(mx4_sum);
897 my= ff_h263_round_chroma(my4_sum);
898 dxy = ((my & 1) << 1) | (mx & 1);
900 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
903 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
904 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
906 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
907 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
910 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);
911 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);
914 switch(s->avctx->mb_cmp&0xFF){
916 return dmin_sum+ 32*s->qscale*s->qscale;*/
920 return dmin_sum+ 11*s->me.mb_penalty_factor;
924 static int interlaced_search(MpegEncContext *s, uint8_t *frame_src_data[3], uint8_t *frame_ref_data[3],
925 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int f_code, int mx, int my)
931 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
933 const int stride= 2*s->linesize;
934 const int uvstride= 2*s->uvlinesize;
936 const int mot_stride= s->mb_stride;
937 const int xy= s->mb_x + s->mb_y*mot_stride;
942 for(block=0; block<2; block++){
944 int best_dmin= INT_MAX;
947 uint8_t *src_data[3]= {
948 frame_src_data[0] + s-> linesize*block,
949 frame_src_data[1] + s->uvlinesize*block,
950 frame_src_data[2] + s->uvlinesize*block
953 for(field_select=0; field_select<2; field_select++){
954 int dmin, mx_i, my_i, pred_x, pred_y;
955 uint8_t *ref_data[3]= {
956 frame_ref_data[0] + s-> linesize*field_select,
957 frame_ref_data[1] + s->uvlinesize*field_select,
958 frame_ref_data[2] + s->uvlinesize*field_select
960 int16_t (*mv_table)[2]= mv_tables[block][field_select];
962 P_LEFT[0] = mv_table[xy - 1][0];
963 P_LEFT[1] = mv_table[xy - 1][1];
964 if(P_LEFT[0] > (s->me.xmax<<1)) P_LEFT[0] = (s->me.xmax<<1);
970 P_TOP[0] = mv_table[xy - mot_stride][0];
971 P_TOP[1] = mv_table[xy - mot_stride][1];
972 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
973 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
974 if(P_TOP[1] > (s->me.ymax<<1)) P_TOP[1] = (s->me.ymax<<1);
975 if(P_TOPRIGHT[0] < (s->me.xmin<<1)) P_TOPRIGHT[0]= (s->me.xmin<<1);
976 if(P_TOPRIGHT[0] > (s->me.xmax<<1)) P_TOPRIGHT[0]= (s->me.xmax<<1);
977 if(P_TOPRIGHT[1] > (s->me.ymax<<1)) P_TOPRIGHT[1]= (s->me.ymax<<1);
979 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
980 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
982 P_MV1[0]= mx; //FIXME not correct if block != field_select
985 dmin = s->me.motion_search[4](s, &mx_i, &my_i, P, pred_x, pred_y,
986 src_data, ref_data, stride, uvstride, mv_table, (1<<16)>>1, mv_penalty);
988 dmin= s->me.sub_motion_search(s, &mx_i, &my_i, dmin,
989 pred_x, pred_y, src_data, ref_data, stride, uvstride, size, h, mv_penalty);
991 mv_table[xy][0]= mx_i;
992 mv_table[xy][1]= my_i;
994 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]
995 && s->avctx->mb_decision == FF_MB_DECISION_SIMPLE){
999 uint8_t *ref= ref_data[0] + (mx_i>>1) + (my_i>>1)*stride;
1000 dxy = ((my_i & 1) << 1) | (mx_i & 1);
1003 s->dsp.put_no_rnd_pixels_tab[size][dxy](s->me.scratchpad, ref , stride, h);
1005 s->dsp.put_pixels_tab [size][dxy](s->me.scratchpad, ref , stride, h);
1007 dmin= s->dsp.mb_cmp[size](s, src_data[0], s->me.scratchpad, stride, h);
1008 dmin+= (mv_penalty[mx_i-pred_x] + mv_penalty[my_i-pred_y] + 1)*s->me.mb_penalty_factor;
1010 dmin+= s->me.mb_penalty_factor; //field_select bits
1012 dmin += field_select != block; //slightly prefer same field
1014 if(dmin < best_dmin){
1016 best_field= field_select;
1020 int16_t (*mv_table)[2]= mv_tables[block][best_field];
1022 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
1023 if(mv_table[xy][1]&1) same=0;
1024 if(mv_table[xy][1]*2 != my) same=0;
1025 if(best_field != block) same=0;
1028 field_select_tables[block][xy]= best_field;
1029 dmin_sum += best_dmin;
1038 switch(s->avctx->mb_cmp&0xFF){
1040 return dmin_sum+ 32*s->qscale*s->qscale;*/
1044 return dmin_sum+ 11*s->me.mb_penalty_factor;
1048 void ff_estimate_p_frame_motion(MpegEncContext * s,
1051 uint8_t *pix, *ppix;
1052 int sum, varc, vard, mx, my, dmin, xx, yy;
1053 int pred_x=0, pred_y=0;
1055 const int shift= 1+s->quarter_sample;
1057 uint8_t *ref_picture= s->last_picture.data[0];
1058 Picture * const pic= &s->current_picture;
1059 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1060 const int stride= s->linesize;
1061 const int uvstride= s->uvlinesize;
1062 uint8_t *src_data[3]= {
1063 s->new_picture.data[0] + 16*(mb_x + stride*mb_y),
1064 s->new_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1065 s->new_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1067 uint8_t *ref_data[3]= {
1068 s->last_picture.data[0] + 16*(mb_x + stride*mb_y),
1069 s->last_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1070 s->last_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1073 assert(s->quarter_sample==0 || s->quarter_sample==1);
1075 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1076 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1077 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1079 get_limits(s, 16*mb_x, 16*mb_y);
1082 switch(s->me_method) {
1085 no_motion_search(s, &mx, &my);
1092 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1097 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1102 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1110 const int mot_stride = s->block_wrap[0];
1111 const int mot_xy = s->block_index[0];
1113 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1114 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1116 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1119 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1120 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1121 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1122 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1123 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
1124 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1125 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1127 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1128 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1130 if(s->out_format == FMT_H263){
1131 pred_x = P_MEDIAN[0];
1132 pred_y = P_MEDIAN[1];
1133 }else { /* mpeg1 at least */
1143 dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y,
1144 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1149 /* intra / predictive decision */
1154 /* At this point (mx,my) are full-pell and the relative displacement */
1155 ppix = ref_data[0] + (my * s->linesize) + mx;
1157 sum = s->dsp.pix_sum(pix, s->linesize);
1159 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1160 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
1162 //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);
1163 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1164 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1165 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1166 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1167 pic->mb_var_sum += varc;
1168 pic->mc_mb_var_sum += vard;
1169 //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);
1172 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1173 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1175 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1176 if (vard <= 64 || vard < varc)
1177 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1179 s->scene_change_score+= s->qscale;
1181 if (vard*2 + 200 > varc)
1182 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1183 if (varc*2 + 200 > vard){
1184 mb_type|= CANDIDATE_MB_TYPE_INTER;
1185 s->me.sub_motion_search(s, &mx, &my, dmin,
1186 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1187 if(s->flags&CODEC_FLAG_MV0)
1189 mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
1194 if((s->flags&CODEC_FLAG_4MV)
1195 && !s->me.skip && varc>50 && vard>10){
1196 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1197 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1199 set_p_mv_tables(s, mx, my, 0);
1201 set_p_mv_tables(s, mx, my, 1);
1202 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1203 && !s->me.skip){ //FIXME varc/d checks
1204 if(interlaced_search(s, src_data, ref_data, s->p_field_mv_table, s->p_field_select_table, s->f_code, mx, my) < INT_MAX)
1205 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1209 mb_type= CANDIDATE_MB_TYPE_INTER;
1211 dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
1212 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1213 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1214 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);
1216 if((s->flags&CODEC_FLAG_4MV)
1217 && !s->me.skip && varc>50 && vard>10){
1218 int dmin4= h263_mv4_search(s, mx, my, shift);
1220 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1224 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1225 && !s->me.skip){ //FIXME varc/d checks
1226 int dmin_i= interlaced_search(s, src_data, ref_data, s->p_field_mv_table, s->p_field_select_table, s->f_code, mx, my);
1228 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1233 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1234 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1236 /* get intra luma score */
1237 if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1238 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
1240 int mean= (sum+128)>>8;
1243 for(i=0; i<16; i++){
1244 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean;
1245 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean;
1246 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean;
1247 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean;
1250 intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize, 16);
1253 /* get chroma score */
1254 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1259 if(s->out_format == FMT_H263){
1260 mean= (s->dc_val[i][mb_x + (mb_y+1)*(s->mb_width+2)] + 4)>>3; //FIXME not exact but simple ;)
1262 mean= (s->last_dc[i] + 4)>>3;
1264 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1268 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean;
1269 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean;
1272 intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize);
1276 intra_score += s->me.mb_penalty_factor*16;
1278 if(intra_score < dmin){
1279 mb_type= CANDIDATE_MB_TYPE_INTRA;
1280 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1282 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1284 if (vard <= 64 || vard < varc) { //FIXME
1285 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1287 s->scene_change_score+= s->qscale;
1291 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1294 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1298 int pred_x=0, pred_y=0;
1300 const int shift= 1+s->quarter_sample;
1301 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1302 const int xy= mb_x + mb_y*s->mb_stride;
1303 const int stride= s->linesize;
1304 const int uvstride= s->uvlinesize;
1305 uint8_t *src_data[3]= {
1306 s->new_picture.data[0] + 16*(mb_x + stride*mb_y),
1307 s->new_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1308 s->new_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1310 uint8_t *ref_data[3]= {
1311 s->last_picture.data[0] + 16*(mb_x + stride*mb_y),
1312 s->last_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1313 s->last_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1316 assert(s->quarter_sample==0 || s->quarter_sample==1);
1318 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1320 get_limits(s, 16*mb_x, 16*mb_y);
1323 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1324 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1326 if(P_LEFT[0] < (s->me.xmin<<shift)) P_LEFT[0] = (s->me.xmin<<shift);
1328 /* special case for first line */
1329 if (mb_y == s->mb_height-1) {
1332 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1333 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1335 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1336 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1337 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1338 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1339 if(P_TOP[1] < (s->me.ymin<<shift)) P_TOP[1] = (s->me.ymin<<shift);
1340 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
1341 if(P_TOPRIGHT[1] < (s->me.ymin<<shift)) P_TOPRIGHT[1]= (s->me.ymin<<shift);
1343 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1344 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1346 pred_x = P_MEDIAN[0];
1347 pred_y = P_MEDIAN[1];
1349 dmin = s->me.pre_motion_search(s, &mx, &my, P, pred_x, pred_y,
1350 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1352 s->p_mv_table[xy][0] = mx<<shift;
1353 s->p_mv_table[xy][1] = my<<shift;
1358 static int ff_estimate_motion_b(MpegEncContext * s,
1359 int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *src_data[3],
1360 uint8_t *ref_data[3], int stride, int uvstride, int f_code)
1363 int pred_x=0, pred_y=0;
1365 const int shift= 1+s->quarter_sample;
1366 const int mot_stride = s->mb_stride;
1367 const int mot_xy = mb_y*mot_stride + mb_x;
1368 uint8_t * const ref_picture= ref_data[0] - 16*s->mb_x - 16*s->mb_y*s->linesize; //FIXME ugly
1369 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1372 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1373 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1374 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1376 get_limits(s, 16*mb_x, 16*mb_y);
1378 switch(s->me_method) {
1381 no_motion_search(s, &mx, &my);
1388 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1393 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1398 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1406 P_LEFT[0] = mv_table[mot_xy - 1][0];
1407 P_LEFT[1] = mv_table[mot_xy - 1][1];
1409 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1411 /* special case for first line */
1413 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1414 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1415 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1416 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1417 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1]= (s->me.ymax<<shift);
1418 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1419 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1421 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1422 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1428 if(mv_table == s->b_forw_mv_table){
1429 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1431 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1434 dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y,
1435 src_data, ref_data, stride, uvstride, s->p_mv_table, mv_scale, mv_penalty);
1440 dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
1441 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1443 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1444 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);
1446 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1447 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1448 mv_table[mot_xy][0]= mx;
1449 mv_table[mot_xy][1]= my;
1454 static inline int check_bidir_mv(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1455 int stride, int uvstride,
1456 int motion_fx, int motion_fy,
1457 int motion_bx, int motion_by,
1458 int pred_fx, int pred_fy,
1459 int pred_bx, int pred_by,
1463 //FIXME move into template?
1464 //FIXME better f_code prediction (max mv & distance)
1466 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1467 uint8_t *dest_y = s->me.scratchpad;
1473 if(s->quarter_sample){
1474 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1475 src_x = motion_fx >> 2;
1476 src_y = motion_fy >> 2;
1478 ptr = ref_data[0] + (src_y * stride) + src_x;
1479 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1481 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1482 src_x = motion_bx >> 2;
1483 src_y = motion_by >> 2;
1485 ptr = ref_data[3] + (src_y * stride) + src_x;
1486 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1488 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1489 src_x = motion_fx >> 1;
1490 src_y = motion_fy >> 1;
1492 ptr = ref_data[0] + (src_y * stride) + src_x;
1493 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1495 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1496 src_x = motion_bx >> 1;
1497 src_y = motion_by >> 1;
1499 ptr = ref_data[3] + (src_y * stride) + src_x;
1500 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1503 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1504 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1505 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1507 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1514 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1515 static inline int bidir_refine(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1516 int stride, int uvstride,
1519 const int mot_stride = s->mb_stride;
1520 const int xy = mb_y *mot_stride + mb_x;
1522 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1523 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1524 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1525 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1526 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1527 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1528 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1529 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1531 //FIXME do refinement and add flag
1533 fbmin= check_bidir_mv(s, src_data, ref_data, stride, uvstride,
1534 motion_fx, motion_fy,
1535 motion_bx, motion_by,
1543 static inline int direct_search(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1544 int stride, int uvstride,
1548 const int mot_stride = s->mb_stride;
1549 const int mot_xy = mb_y*mot_stride + mb_x;
1550 const int shift= 1+s->quarter_sample;
1552 const int time_pp= s->pp_time;
1553 const int time_pb= s->pb_time;
1554 int mx, my, xmin, xmax, ymin, ymax;
1555 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1556 uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1558 ymin= xmin=(-32)>>shift;
1559 ymax= xmax= 31>>shift;
1561 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1562 s->mv_type= MV_TYPE_8X8;
1564 s->mv_type= MV_TYPE_16X16;
1568 int index= s->block_index[i];
1571 s->me.co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1572 s->me.co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1573 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1574 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1575 // 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);
1576 // 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);
1578 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1579 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1580 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1582 xmax= FFMIN(xmax, s->width - max);
1583 xmin= FFMAX(xmin, - 16 - min);
1585 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1586 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1587 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1589 ymax= FFMIN(ymax, s->height - max);
1590 ymin= FFMAX(ymin, - 16 - min);
1592 if(s->mv_type == MV_TYPE_16X16) break;
1595 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1597 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1598 s->b_direct_mv_table[mot_xy][0]= 0;
1599 s->b_direct_mv_table[mot_xy][1]= 0;
1601 return 256*256*256*64;
1609 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1610 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1612 /* special case for first line */
1614 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1615 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1616 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1617 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1619 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1620 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1623 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1624 if(s->flags&CODEC_FLAG_QPEL){
1625 dmin = simple_direct_qpel_epzs_motion_search(s, &mx, &my, P, 0, 0,
1626 src_data, ref_data, stride, uvstride, mv_table, 1<<14, mv_penalty);
1627 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin,
1628 0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1630 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1631 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
1633 dmin = simple_direct_hpel_epzs_motion_search(s, &mx, &my, P, 0, 0,
1634 src_data, ref_data, stride, uvstride, mv_table, 1<<15, mv_penalty);
1635 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin,
1636 0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1638 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1639 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
1642 get_limits(s, 16*mb_x, 16*mb_y); //restore s->me.?min/max, maybe not needed
1644 s->b_direct_mv_table[mot_xy][0]= mx;
1645 s->b_direct_mv_table[mot_xy][1]= my;
1649 void ff_estimate_b_frame_motion(MpegEncContext * s,
1652 const int penalty_factor= s->me.mb_penalty_factor;
1653 int fmin, bmin, dmin, fbmin, bimin, fimin;
1655 const int stride= s->linesize;
1656 const int uvstride= s->uvlinesize;
1657 uint8_t *src_data[3]= {
1658 s->new_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1659 s->new_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1660 s->new_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
1662 uint8_t *ref_data[6]= {
1663 s->last_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1664 s->last_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1665 s->last_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y),
1666 s->next_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1667 s->next_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1668 s->next_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
1672 if (s->codec_id == CODEC_ID_MPEG4)
1673 dmin= direct_search(s, src_data, ref_data, stride, uvstride, mb_x, mb_y);
1676 //FIXME penalty stuff for non mpeg4
1678 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, src_data,
1679 ref_data, stride, uvstride, s->f_code) + 3*penalty_factor;
1682 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, src_data,
1683 ref_data+3, stride, uvstride, s->b_code) + 2*penalty_factor;
1684 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1687 fbmin= bidir_refine(s, src_data, ref_data, stride, uvstride, mb_x, mb_y) + penalty_factor;
1688 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1690 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1691 const int xy = mb_y*s->mb_stride + mb_x;
1693 //FIXME mb type penalty
1695 fimin= interlaced_search(s, src_data, ref_data ,
1696 s->b_field_mv_table[0], s->b_field_select_table[0], s->f_code,
1697 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1698 bimin= interlaced_search(s, src_data, ref_data+3,
1699 s->b_field_mv_table[1], s->b_field_select_table[1], s->b_code,
1700 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1]);
1702 fimin= bimin= INT_MAX;
1706 type = CANDIDATE_MB_TYPE_FORWARD;
1710 type = CANDIDATE_MB_TYPE_DIRECT;
1714 type= CANDIDATE_MB_TYPE_BACKWARD;
1718 type= CANDIDATE_MB_TYPE_BIDIR;
1722 type= CANDIDATE_MB_TYPE_FORWARD_I;
1726 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1729 score= ((unsigned)(score*score + 128*256))>>16;
1730 s->current_picture.mc_mb_var_sum += score;
1731 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1734 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1735 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1737 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1739 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1740 if(fimin < INT_MAX && bimin < INT_MAX){
1741 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1743 //FIXME something smarter
1744 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1746 if(s->out_format == FMT_MPEG1)
1747 type |= CANDIDATE_MB_TYPE_INTRA;
1751 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1754 /* find best f_code for ME which do unlimited searches */
1755 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1757 if(s->me_method>=ME_EPZS){
1760 uint8_t * fcode_tab= s->fcode_tab;
1762 int best_score=-10000000;
1764 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1766 for(y=0; y<s->mb_height; y++){
1768 int xy= y*s->mb_stride;
1769 for(x=0; x<s->mb_width; x++){
1770 if(s->mb_type[xy] & type){
1771 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1772 fcode_tab[mv_table[xy][1] + MAX_MV]);
1775 for(j=0; j<fcode && j<8; j++){
1776 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1785 if(score[i] > best_score){
1786 best_score= score[i];
1789 // printf("%d %d\n", i, score[i]);
1792 // printf("fcode: %d type: %d\n", i, s->pict_type);
1794 /* for(i=0; i<=MAX_FCODE; i++){
1795 printf("%d ", mv_num[i]);
1803 void ff_fix_long_p_mvs(MpegEncContext * s)
1805 const int f_code= s->f_code;
1807 assert(s->pict_type==P_TYPE);
1809 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1811 if(s->msmpeg4_version) range= 16;
1813 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1815 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1816 if(s->flags&CODEC_FLAG_4MV){
1817 const int wrap= 2+ s->mb_width*2;
1819 /* clip / convert to intra 8x8 type MVs */
1820 for(y=0; y<s->mb_height; y++){
1821 int xy= (y*2 + 1)*wrap + 1;
1822 int i= y*s->mb_stride;
1825 for(x=0; x<s->mb_width; x++){
1826 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1828 for(block=0; block<4; block++){
1829 int off= (block& 1) + (block>>1)*wrap;
1830 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1831 int my= s->current_picture.motion_val[0][ xy + off ][1];
1833 if( mx >=range || mx <-range
1834 || my >=range || my <-range){
1835 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1836 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1837 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1850 * @param truncate 1 for truncation, 0 for using intra
1852 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1853 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1855 int y, h_range, v_range;
1857 // RAL: 8 in MPEG-1, 16 in MPEG-4
1858 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1860 if(s->msmpeg4_version) range= 16;
1861 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1864 v_range= field_select_table ? range>>1 : range;
1866 /* clip / convert to intra 16x16 type MVs */
1867 for(y=0; y<s->mb_height; y++){
1869 int xy= y*s->mb_stride;
1870 for(x=0; x<s->mb_width; x++){
1871 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1872 if(field_select_table==NULL || field_select_table[xy] == field_select){
1873 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1874 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1877 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1878 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1879 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1880 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1882 s->mb_type[xy] &= ~type;
1883 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;