3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2003 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"
281 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
285 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
286 DSPContext* c= &s->dsp;
289 memset(cmp, 0, sizeof(void*)*5);
297 cmp[i]= c->hadamard8_diff[i];
303 cmp[i]= c->dct_sad[i];
306 cmp[i]= c->quant_psnr[i];
318 av_log(s->avctx, AV_LOG_ERROR,"internal error in cmp function selection\n");
323 static inline int get_penalty_factor(MpegEncContext *s, int type){
333 return s->qscale*s->qscale*2;
338 return (s->qscale*s->qscale*185 + 64)>>7;
342 void ff_init_me(MpegEncContext *s){
343 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
344 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
345 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
346 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
348 if(s->flags&CODEC_FLAG_QPEL){
349 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
350 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
352 s->me.sub_motion_search= simple_qpel_motion_search;
354 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
355 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
356 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
357 && s->avctx-> me_cmp == FF_CMP_SAD
358 && s->avctx-> mb_cmp == FF_CMP_SAD)
359 s->me.sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
361 s->me.sub_motion_search= simple_hpel_motion_search;
364 if(s->avctx->me_cmp&FF_CMP_CHROMA){
365 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
366 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
367 s->me.motion_search[4]= simple_chroma_epzs_motion_search2;
369 s->me.motion_search[0]= simple_epzs_motion_search;
370 s->me.motion_search[1]= simple_epzs_motion_search4;
371 s->me.motion_search[4]= simple_epzs_motion_search2;
374 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
375 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
377 s->me.pre_motion_search= simple_epzs_motion_search;
380 if(s->flags&CODEC_FLAG_QPEL){
381 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
382 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
384 s->me.get_mb_score= simple_qpel_get_mb_score;
386 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
387 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
389 s->me.get_mb_score= simple_hpel_get_mb_score;
394 static int pix_dev(uint8_t * pix, int line_size, int mean)
399 for (i = 0; i < 16; i++) {
400 for (j = 0; j < 16; j += 8) {
401 s += ABS(pix[0]-mean);
402 s += ABS(pix[1]-mean);
403 s += ABS(pix[2]-mean);
404 s += ABS(pix[3]-mean);
405 s += ABS(pix[4]-mean);
406 s += ABS(pix[5]-mean);
407 s += ABS(pix[6]-mean);
408 s += ABS(pix[7]-mean);
411 pix += line_size - 16;
417 static inline void no_motion_search(MpegEncContext * s,
418 int *mx_ptr, int *my_ptr)
420 *mx_ptr = 16 * s->mb_x;
421 *my_ptr = 16 * s->mb_y;
424 static int full_motion_search(MpegEncContext * s,
425 int *mx_ptr, int *my_ptr, int range,
426 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
428 int x1, y1, x2, y2, xx, yy, x, y;
434 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
446 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
450 for (y = y1; y <= y2; y++) {
451 for (x = x1; x <= x2; x++) {
452 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
456 (abs(x - xx) + abs(y - yy)) <
457 (abs(mx - xx) + abs(my - yy)))) {
469 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
470 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
471 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
478 static int log_motion_search(MpegEncContext * s,
479 int *mx_ptr, int *my_ptr, int range,
480 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
482 int x1, y1, x2, y2, xx, yy, x, y;
509 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
515 for (y = y1; y <= y2; y += range) {
516 for (x = x1; x <= x2; x += range) {
517 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
518 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
544 } while (range >= 1);
547 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
554 static int phods_motion_search(MpegEncContext * s,
555 int *mx_ptr, int *my_ptr, int range,
556 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
558 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
559 int mx, my, dminx, dminy;
585 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
596 for (x = x1; x <= x2; x += range) {
597 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
598 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
605 for (y = y1; y <= y2; y += range) {
606 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
607 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
633 } while (range >= 1);
636 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
639 /* half pixel search */
646 #define Z_THRESHOLD 256
648 #define CHECK_SAD_HALF_MV(suffix, x, y) \
650 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
651 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
652 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
655 static inline int sad_hpel_motion_search(MpegEncContext * s,
656 int *mx_ptr, int *my_ptr, int dmin,
657 int pred_x, int pred_y, uint8_t *src_data[3],
658 uint8_t *ref_data[6], int stride, int uvstride,
659 int size, int h, uint8_t * const mv_penalty)
661 uint32_t *score_map= s->me.score_map;
662 const int penalty_factor= s->me.sub_penalty_factor;
665 const int xmin= s->me.xmin;
666 const int ymin= s->me.ymin;
667 const int xmax= s->me.xmax;
668 const int ymax= s->me.ymax;
682 ptr = ref_data[0] + (my * stride) + mx;
686 if (mx > xmin && mx < xmax &&
687 my > ymin && my < ymax) {
690 const int index= (my<<ME_MAP_SHIFT) + mx;
691 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
692 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
693 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
694 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
704 CHECK_SAD_HALF_MV(y2 , 0, -1)
706 CHECK_SAD_HALF_MV(xy2, -1, -1)
708 CHECK_SAD_HALF_MV(xy2, +1, -1)
712 CHECK_SAD_HALF_MV(xy2, -1, +1)
714 CHECK_SAD_HALF_MV(x2 , -1, 0)
716 CHECK_SAD_HALF_MV(xy2, +1, -1)
718 CHECK_SAD_HALF_MV(xy2, -1, -1)
722 CHECK_SAD_HALF_MV(xy2, +1, +1)
724 CHECK_SAD_HALF_MV(x2 , +1, 0)
729 CHECK_SAD_HALF_MV(xy2, -1, -1)
733 CHECK_SAD_HALF_MV(xy2, +1, +1)
735 CHECK_SAD_HALF_MV(x2 , -1, 0)
736 CHECK_SAD_HALF_MV(xy2, -1, +1)
739 CHECK_SAD_HALF_MV(xy2, +1, -1)
743 CHECK_SAD_HALF_MV(xy2, -1, +1)
745 CHECK_SAD_HALF_MV(x2 , +1, 0)
746 CHECK_SAD_HALF_MV(xy2, +1, +1)
748 CHECK_SAD_HALF_MV(y2 , 0, +1)
763 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
765 const int xy= s->mb_x + s->mb_y*s->mb_stride;
767 s->p_mv_table[xy][0] = mx;
768 s->p_mv_table[xy][1] = my;
770 /* has allready been set to the 4 MV if 4MV is done */
772 int mot_xy= s->block_index[0];
774 s->current_picture.motion_val[0][mot_xy ][0]= mx;
775 s->current_picture.motion_val[0][mot_xy ][1]= my;
776 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
777 s->current_picture.motion_val[0][mot_xy+1][1]= my;
779 mot_xy += s->block_wrap[0];
780 s->current_picture.motion_val[0][mot_xy ][0]= mx;
781 s->current_picture.motion_val[0][mot_xy ][1]= my;
782 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
783 s->current_picture.motion_val[0][mot_xy+1][1]= my;
788 * get fullpel ME search limits.
790 static inline void get_limits(MpegEncContext *s, int x, int y)
793 if(s->avctx->me_range) s->me.range= s->avctx->me_range >> 1;
794 else s->me.range= 16;
796 if (s->unrestricted_mv) {
797 s->me.xmin = - x - 16;
798 s->me.ymin = - y - 16;
799 s->me.xmax = - x + s->mb_width *16;
800 s->me.ymax = - y + s->mb_height*16;
804 s->me.xmax = - x + s->mb_width *16 - 16;
805 s->me.ymax = - y + s->mb_height*16 - 16;
809 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
815 int dmin_sum=0, mx4_sum=0, my4_sum=0;
816 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
818 const int stride= s->linesize;
819 const int uvstride= s->uvlinesize;
820 const int xmin= s->me.xmin;
821 const int ymin= s->me.ymin;
822 const int xmax= s->me.xmax;
823 const int ymax= s->me.ymax;
825 for(block=0; block<4; block++){
827 int pred_x4, pred_y4;
829 static const int off[4]= {2, 1, 1, -1};
830 const int mot_stride = s->block_wrap[0];
831 const int mot_xy = s->block_index[block];
832 const int block_x= (block&1);
833 const int block_y= (block>>1);
834 uint8_t *src_data[3]= {
835 s->new_picture.data[0] + 8*(2*s->mb_x + block_x) + stride *8*(2*s->mb_y + block_y), //FIXME chroma?
836 s->new_picture.data[1] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y),
837 s->new_picture.data[2] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y)
839 uint8_t *ref_data[3]= {
840 s->last_picture.data[0] + 8*(2*s->mb_x + block_x) + stride *8*(2*s->mb_y + block_y), //FIXME chroma?
841 s->last_picture.data[1] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y),
842 s->last_picture.data[2] + 4*(2*s->mb_x + block_x) + uvstride*4*(2*s->mb_y + block_y)
845 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
846 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
848 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
850 /* special case for first line */
851 if (s->mb_y == 0 && block<2) {
855 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
856 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
857 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
858 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
859 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
860 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
861 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
862 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
864 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
865 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
867 // if(s->out_format == FMT_H263){
868 pred_x4 = P_MEDIAN[0];
869 pred_y4 = P_MEDIAN[1];
871 }else { /* mpeg1 at least */
880 dmin4 = s->me.motion_search[1](s, &mx4, &my4, P, pred_x4, pred_y4,
881 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
883 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4,
884 pred_x4, pred_y4, src_data, ref_data, stride, uvstride, size, h, mv_penalty);
886 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
888 const int offset= ((block&1) + (block>>1)*stride)*8;
889 uint8_t *dest_y = s->me.scratchpad + offset;
891 if(s->quarter_sample){
892 uint8_t *ref= ref_data[0] + (mx4>>2) + (my4>>2)*stride + offset;
893 dxy = ((my4 & 3) << 2) | (mx4 & 3);
896 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , s->linesize);
898 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
900 uint8_t *ref= ref_data[0] + (mx4>>1) + (my4>>1)*stride + offset;
901 dxy = ((my4 & 1) << 1) | (mx4 & 1);
904 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
906 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
908 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
912 if(s->quarter_sample){
920 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
921 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
923 if(mx4 != mx || my4 != my) same=0;
929 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
930 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);
933 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
938 mx= ff_h263_round_chroma(mx4_sum);
939 my= ff_h263_round_chroma(my4_sum);
940 dxy = ((my & 1) << 1) | (mx & 1);
942 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
945 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
946 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
948 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
949 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
952 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);
953 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);
956 switch(s->avctx->mb_cmp&0xFF){
958 return dmin_sum+ 32*s->qscale*s->qscale;*/
962 return dmin_sum+ 11*s->me.mb_penalty_factor;
966 static int interlaced_search(MpegEncContext *s, uint8_t *frame_src_data[3], uint8_t *frame_ref_data[3],
967 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int f_code, int mx, int my)
973 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
975 const int stride= 2*s->linesize;
976 const int uvstride= 2*s->uvlinesize;
978 const int mot_stride= s->mb_stride;
979 const int xy= s->mb_x + s->mb_y*mot_stride;
984 for(block=0; block<2; block++){
986 int best_dmin= INT_MAX;
989 uint8_t *src_data[3]= {
990 frame_src_data[0] + s-> linesize*block,
991 frame_src_data[1] + s->uvlinesize*block,
992 frame_src_data[2] + s->uvlinesize*block
995 for(field_select=0; field_select<2; field_select++){
996 int dmin, mx_i, my_i, pred_x, pred_y;
997 uint8_t *ref_data[3]= {
998 frame_ref_data[0] + s-> linesize*field_select,
999 frame_ref_data[1] + s->uvlinesize*field_select,
1000 frame_ref_data[2] + s->uvlinesize*field_select
1002 int16_t (*mv_table)[2]= mv_tables[block][field_select];
1004 P_LEFT[0] = mv_table[xy - 1][0];
1005 P_LEFT[1] = mv_table[xy - 1][1];
1006 if(P_LEFT[0] > (s->me.xmax<<1)) P_LEFT[0] = (s->me.xmax<<1);
1012 P_TOP[0] = mv_table[xy - mot_stride][0];
1013 P_TOP[1] = mv_table[xy - mot_stride][1];
1014 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
1015 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
1016 if(P_TOP[1] > (s->me.ymax<<1)) P_TOP[1] = (s->me.ymax<<1);
1017 if(P_TOPRIGHT[0] < (s->me.xmin<<1)) P_TOPRIGHT[0]= (s->me.xmin<<1);
1018 if(P_TOPRIGHT[0] > (s->me.xmax<<1)) P_TOPRIGHT[0]= (s->me.xmax<<1);
1019 if(P_TOPRIGHT[1] > (s->me.ymax<<1)) P_TOPRIGHT[1]= (s->me.ymax<<1);
1021 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1022 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1024 P_MV1[0]= mx; //FIXME not correct if block != field_select
1027 dmin = s->me.motion_search[4](s, &mx_i, &my_i, P, pred_x, pred_y,
1028 src_data, ref_data, stride, uvstride, mv_table, (1<<16)>>1, mv_penalty);
1030 dmin= s->me.sub_motion_search(s, &mx_i, &my_i, dmin,
1031 pred_x, pred_y, src_data, ref_data, stride, uvstride, size, h, mv_penalty);
1033 mv_table[xy][0]= mx_i;
1034 mv_table[xy][1]= my_i;
1036 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
1040 uint8_t *ref= ref_data[0] + (mx_i>>1) + (my_i>>1)*stride;
1041 dxy = ((my_i & 1) << 1) | (mx_i & 1);
1044 s->dsp.put_no_rnd_pixels_tab[size][dxy](s->me.scratchpad, ref , stride, h);
1046 s->dsp.put_pixels_tab [size][dxy](s->me.scratchpad, ref , stride, h);
1048 dmin= s->dsp.mb_cmp[size](s, src_data[0], s->me.scratchpad, stride, h);
1049 dmin+= (mv_penalty[mx_i-pred_x] + mv_penalty[my_i-pred_y] + 1)*s->me.mb_penalty_factor;
1051 dmin+= s->me.mb_penalty_factor; //field_select bits
1053 dmin += field_select != block; //slightly prefer same field
1055 if(dmin < best_dmin){
1057 best_field= field_select;
1061 int16_t (*mv_table)[2]= mv_tables[block][best_field];
1063 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
1064 if(mv_table[xy][1]&1) same=0;
1065 if(mv_table[xy][1]*2 != my) same=0;
1066 if(best_field != block) same=0;
1069 field_select_tables[block][xy]= best_field;
1070 dmin_sum += best_dmin;
1079 switch(s->avctx->mb_cmp&0xFF){
1081 return dmin_sum+ 32*s->qscale*s->qscale;*/
1085 return dmin_sum+ 11*s->me.mb_penalty_factor;
1089 void ff_estimate_p_frame_motion(MpegEncContext * s,
1092 uint8_t *pix, *ppix;
1093 int sum, varc, vard, mx, my, dmin, xx, yy;
1094 int pred_x=0, pred_y=0;
1096 const int shift= 1+s->quarter_sample;
1098 uint8_t *ref_picture= s->last_picture.data[0];
1099 Picture * const pic= &s->current_picture;
1100 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1101 const int stride= s->linesize;
1102 const int uvstride= s->uvlinesize;
1103 uint8_t *src_data[3]= {
1104 s->new_picture.data[0] + 16*(mb_x + stride*mb_y),
1105 s->new_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1106 s->new_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1108 uint8_t *ref_data[3]= {
1109 s->last_picture.data[0] + 16*(mb_x + stride*mb_y),
1110 s->last_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1111 s->last_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1114 assert(s->quarter_sample==0 || s->quarter_sample==1);
1116 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1117 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1118 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1120 get_limits(s, 16*mb_x, 16*mb_y);
1123 switch(s->me_method) {
1126 no_motion_search(s, &mx, &my);
1133 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1138 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1143 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1151 const int mot_stride = s->block_wrap[0];
1152 const int mot_xy = s->block_index[0];
1154 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1155 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1157 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1160 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1161 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1162 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1163 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1164 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift);
1165 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1166 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1168 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1169 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1171 if(s->out_format == FMT_H263){
1172 pred_x = P_MEDIAN[0];
1173 pred_y = P_MEDIAN[1];
1174 }else { /* mpeg1 at least */
1184 dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y,
1185 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1190 /* intra / predictive decision */
1195 /* At this point (mx,my) are full-pell and the relative displacement */
1196 ppix = ref_data[0] + (my * s->linesize) + mx;
1198 sum = s->dsp.pix_sum(pix, s->linesize);
1200 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1201 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
1203 //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);
1204 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1205 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1206 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1207 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1208 pic->mb_var_sum += varc;
1209 pic->mc_mb_var_sum += vard;
1210 //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);
1213 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1214 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1216 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1217 if (vard <= 64 || vard < varc)
1218 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1220 s->scene_change_score+= s->qscale;
1222 if (vard*2 + 200 > varc)
1223 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1224 if (varc*2 + 200 > vard){
1225 mb_type|= CANDIDATE_MB_TYPE_INTER;
1226 s->me.sub_motion_search(s, &mx, &my, dmin,
1227 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1228 if(s->flags&CODEC_FLAG_MV0)
1230 mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
1235 if((s->flags&CODEC_FLAG_4MV)
1236 && !s->me.skip && varc>50 && vard>10){
1237 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1238 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1240 set_p_mv_tables(s, mx, my, 0);
1242 set_p_mv_tables(s, mx, my, 1);
1243 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1244 && !s->me.skip){ //FIXME varc/d checks
1245 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)
1246 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1250 mb_type= CANDIDATE_MB_TYPE_INTER;
1252 dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
1253 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1254 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1255 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);
1257 if((s->flags&CODEC_FLAG_4MV)
1258 && !s->me.skip && varc>50 && vard>10){
1259 int dmin4= h263_mv4_search(s, mx, my, shift);
1261 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1265 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1266 && !s->me.skip){ //FIXME varc/d checks
1267 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);
1269 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1274 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1275 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1277 /* get intra luma score */
1278 if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1279 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
1281 int mean= (sum+128)>>8;
1284 for(i=0; i<16; i++){
1285 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean;
1286 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean;
1287 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean;
1288 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean;
1291 intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize, 16);
1294 /* get chroma score */
1295 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1300 if(s->out_format == FMT_H263){
1301 mean= (s->dc_val[i][mb_x + (mb_y+1)*(s->mb_width+2)] + 4)>>3; //FIXME not exact but simple ;)
1303 mean= (s->last_dc[i] + 4)>>3;
1305 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1309 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean;
1310 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean;
1313 intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize);
1317 intra_score += s->me.mb_penalty_factor*16;
1319 if(intra_score < dmin){
1320 mb_type= CANDIDATE_MB_TYPE_INTRA;
1321 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1323 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1325 if (vard <= 64 || vard < varc) { //FIXME
1326 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1328 s->scene_change_score+= s->qscale;
1332 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1335 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1339 int pred_x=0, pred_y=0;
1341 const int shift= 1+s->quarter_sample;
1342 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1343 const int xy= mb_x + mb_y*s->mb_stride;
1344 const int stride= s->linesize;
1345 const int uvstride= s->uvlinesize;
1346 uint8_t *src_data[3]= {
1347 s->new_picture.data[0] + 16*(mb_x + stride*mb_y),
1348 s->new_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1349 s->new_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1351 uint8_t *ref_data[3]= {
1352 s->last_picture.data[0] + 16*(mb_x + stride*mb_y),
1353 s->last_picture.data[1] + 8*(mb_x + uvstride*mb_y),
1354 s->last_picture.data[2] + 8*(mb_x + uvstride*mb_y)
1357 assert(s->quarter_sample==0 || s->quarter_sample==1);
1359 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1361 get_limits(s, 16*mb_x, 16*mb_y);
1364 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1365 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1367 if(P_LEFT[0] < (s->me.xmin<<shift)) P_LEFT[0] = (s->me.xmin<<shift);
1369 /* special case for first line */
1370 if (mb_y == s->mb_height-1) {
1373 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1374 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1376 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1377 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1378 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1379 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1380 if(P_TOP[1] < (s->me.ymin<<shift)) P_TOP[1] = (s->me.ymin<<shift);
1381 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift);
1382 if(P_TOPRIGHT[1] < (s->me.ymin<<shift)) P_TOPRIGHT[1]= (s->me.ymin<<shift);
1384 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1385 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1387 pred_x = P_MEDIAN[0];
1388 pred_y = P_MEDIAN[1];
1390 dmin = s->me.pre_motion_search(s, &mx, &my, P, pred_x, pred_y,
1391 src_data, ref_data, stride, uvstride, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1393 s->p_mv_table[xy][0] = mx<<shift;
1394 s->p_mv_table[xy][1] = my<<shift;
1399 static int ff_estimate_motion_b(MpegEncContext * s,
1400 int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *src_data[3],
1401 uint8_t *ref_data[3], int stride, int uvstride, int f_code)
1404 int pred_x=0, pred_y=0;
1406 const int shift= 1+s->quarter_sample;
1407 const int mot_stride = s->mb_stride;
1408 const int mot_xy = mb_y*mot_stride + mb_x;
1409 uint8_t * const ref_picture= ref_data[0] - 16*s->mb_x - 16*s->mb_y*s->linesize; //FIXME ugly
1410 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1413 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1414 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1415 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1417 get_limits(s, 16*mb_x, 16*mb_y);
1419 switch(s->me_method) {
1422 no_motion_search(s, &mx, &my);
1429 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1434 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1439 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1447 P_LEFT[0] = mv_table[mot_xy - 1][0];
1448 P_LEFT[1] = mv_table[mot_xy - 1][1];
1450 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift);
1452 /* special case for first line */
1454 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1455 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1456 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1457 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1458 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1]= (s->me.ymax<<shift);
1459 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift);
1460 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift);
1462 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1463 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1469 if(mv_table == s->b_forw_mv_table){
1470 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1472 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1475 dmin = s->me.motion_search[0](s, &mx, &my, P, pred_x, pred_y,
1476 src_data, ref_data, stride, uvstride, s->p_mv_table, mv_scale, mv_penalty);
1481 dmin= s->me.sub_motion_search(s, &mx, &my, dmin,
1482 pred_x, pred_y, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1484 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1485 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, src_data, ref_data, stride, uvstride, mv_penalty);
1487 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1488 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1489 mv_table[mot_xy][0]= mx;
1490 mv_table[mot_xy][1]= my;
1495 static inline int check_bidir_mv(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1496 int stride, int uvstride,
1497 int motion_fx, int motion_fy,
1498 int motion_bx, int motion_by,
1499 int pred_fx, int pred_fy,
1500 int pred_bx, int pred_by,
1504 //FIXME move into template?
1505 //FIXME better f_code prediction (max mv & distance)
1507 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1508 uint8_t *dest_y = s->me.scratchpad;
1514 if(s->quarter_sample){
1515 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1516 src_x = motion_fx >> 2;
1517 src_y = motion_fy >> 2;
1519 ptr = ref_data[0] + (src_y * stride) + src_x;
1520 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1522 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1523 src_x = motion_bx >> 2;
1524 src_y = motion_by >> 2;
1526 ptr = ref_data[3] + (src_y * stride) + src_x;
1527 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1529 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1530 src_x = motion_fx >> 1;
1531 src_y = motion_fy >> 1;
1533 ptr = ref_data[0] + (src_y * stride) + src_x;
1534 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1536 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1537 src_x = motion_bx >> 1;
1538 src_y = motion_by >> 1;
1540 ptr = ref_data[3] + (src_y * stride) + src_x;
1541 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1544 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1545 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1546 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1548 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1555 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1556 static inline int bidir_refine(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1557 int stride, int uvstride,
1560 const int mot_stride = s->mb_stride;
1561 const int xy = mb_y *mot_stride + mb_x;
1563 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1564 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1565 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1566 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1567 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1568 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1569 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1570 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1572 //FIXME do refinement and add flag
1574 fbmin= check_bidir_mv(s, src_data, ref_data, stride, uvstride,
1575 motion_fx, motion_fy,
1576 motion_bx, motion_by,
1584 static inline int direct_search(MpegEncContext * s, uint8_t *src_data[3], uint8_t *ref_data[6],
1585 int stride, int uvstride,
1589 const int mot_stride = s->mb_stride;
1590 const int mot_xy = mb_y*mot_stride + mb_x;
1591 const int shift= 1+s->quarter_sample;
1593 const int time_pp= s->pp_time;
1594 const int time_pb= s->pb_time;
1595 int mx, my, xmin, xmax, ymin, ymax;
1596 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1597 uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1599 ymin= xmin=(-32)>>shift;
1600 ymax= xmax= 31>>shift;
1602 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1603 s->mv_type= MV_TYPE_8X8;
1605 s->mv_type= MV_TYPE_16X16;
1609 int index= s->block_index[i];
1612 s->me.co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1613 s->me.co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1614 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1615 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1616 // 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);
1617 // 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);
1619 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1620 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1621 max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1622 min+= (2*mb_x + (i& 1))*8 - 1;
1623 xmax= FFMIN(xmax, s->width - max);
1624 xmin= FFMAX(xmin, - 16 - min);
1626 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1627 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1628 max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1629 min+= (2*mb_y + (i>>1))*8 - 1;
1630 ymax= FFMIN(ymax, s->height - max);
1631 ymin= FFMAX(ymin, - 16 - min);
1633 if(s->mv_type == MV_TYPE_16X16) break;
1636 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1638 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1639 s->b_direct_mv_table[mot_xy][0]= 0;
1640 s->b_direct_mv_table[mot_xy][1]= 0;
1642 return 256*256*256*64;
1650 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1651 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1653 /* special case for first line */
1655 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1656 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1657 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1658 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1660 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1661 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1664 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1665 if(s->flags&CODEC_FLAG_QPEL){
1666 dmin = simple_direct_qpel_epzs_motion_search(s, &mx, &my, P, 0, 0,
1667 src_data, ref_data, stride, uvstride, mv_table, 1<<14, mv_penalty);
1668 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin,
1669 0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1671 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1672 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
1674 dmin = simple_direct_hpel_epzs_motion_search(s, &mx, &my, P, 0, 0,
1675 src_data, ref_data, stride, uvstride, mv_table, 1<<15, mv_penalty);
1676 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin,
1677 0, 0, src_data, ref_data, stride, uvstride, 0, 16, mv_penalty);
1679 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1680 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, src_data, ref_data, stride, uvstride, mv_penalty);
1683 get_limits(s, 16*mb_x, 16*mb_y); //restore s->me.?min/max, maybe not needed
1685 s->b_direct_mv_table[mot_xy][0]= mx;
1686 s->b_direct_mv_table[mot_xy][1]= my;
1690 void ff_estimate_b_frame_motion(MpegEncContext * s,
1693 const int penalty_factor= s->me.mb_penalty_factor;
1694 int fmin, bmin, dmin, fbmin, bimin, fimin;
1696 const int stride= s->linesize;
1697 const int uvstride= s->uvlinesize;
1698 uint8_t *src_data[3]= {
1699 s->new_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1700 s->new_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1701 s->new_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
1703 uint8_t *ref_data[6]= {
1704 s->last_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1705 s->last_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1706 s->last_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y),
1707 s->next_picture.data[0] + 16*(s->mb_x + stride*s->mb_y),
1708 s->next_picture.data[1] + 8*(s->mb_x + uvstride*s->mb_y),
1709 s->next_picture.data[2] + 8*(s->mb_x + uvstride*s->mb_y)
1713 if (s->codec_id == CODEC_ID_MPEG4)
1714 dmin= direct_search(s, src_data, ref_data, stride, uvstride, mb_x, mb_y);
1717 //FIXME penalty stuff for non mpeg4
1719 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, src_data,
1720 ref_data, stride, uvstride, s->f_code) + 3*penalty_factor;
1723 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, src_data,
1724 ref_data+3, stride, uvstride, s->b_code) + 2*penalty_factor;
1725 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1728 fbmin= bidir_refine(s, src_data, ref_data, stride, uvstride, mb_x, mb_y) + penalty_factor;
1729 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1731 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1732 const int xy = mb_y*s->mb_stride + mb_x;
1734 //FIXME mb type penalty
1736 fimin= interlaced_search(s, src_data, ref_data ,
1737 s->b_field_mv_table[0], s->b_field_select_table[0], s->f_code,
1738 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1739 bimin= interlaced_search(s, src_data, ref_data+3,
1740 s->b_field_mv_table[1], s->b_field_select_table[1], s->b_code,
1741 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1]);
1743 fimin= bimin= INT_MAX;
1747 type = CANDIDATE_MB_TYPE_FORWARD;
1751 type = CANDIDATE_MB_TYPE_DIRECT;
1755 type= CANDIDATE_MB_TYPE_BACKWARD;
1759 type= CANDIDATE_MB_TYPE_BIDIR;
1763 type= CANDIDATE_MB_TYPE_FORWARD_I;
1767 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1770 score= ((unsigned)(score*score + 128*256))>>16;
1771 s->current_picture.mc_mb_var_sum += score;
1772 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1775 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1776 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1778 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1780 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1781 if(fimin < INT_MAX && bimin < INT_MAX){
1782 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1784 //FIXME something smarter
1785 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1788 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1791 /* find best f_code for ME which do unlimited searches */
1792 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1794 if(s->me_method>=ME_EPZS){
1797 uint8_t * fcode_tab= s->fcode_tab;
1799 int best_score=-10000000;
1801 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1803 for(y=0; y<s->mb_height; y++){
1805 int xy= y*s->mb_stride;
1806 for(x=0; x<s->mb_width; x++){
1807 if(s->mb_type[xy] & type){
1808 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1809 fcode_tab[mv_table[xy][1] + MAX_MV]);
1812 for(j=0; j<fcode && j<8; j++){
1813 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1822 if(score[i] > best_score){
1823 best_score= score[i];
1826 // printf("%d %d\n", i, score[i]);
1829 // printf("fcode: %d type: %d\n", i, s->pict_type);
1831 /* for(i=0; i<=MAX_FCODE; i++){
1832 printf("%d ", mv_num[i]);
1840 void ff_fix_long_p_mvs(MpegEncContext * s)
1842 const int f_code= s->f_code;
1844 assert(s->pict_type==P_TYPE);
1846 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1848 if(s->msmpeg4_version) range= 16;
1850 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1852 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1853 if(s->flags&CODEC_FLAG_4MV){
1854 const int wrap= 2+ s->mb_width*2;
1856 /* clip / convert to intra 8x8 type MVs */
1857 for(y=0; y<s->mb_height; y++){
1858 int xy= (y*2 + 1)*wrap + 1;
1859 int i= y*s->mb_stride;
1862 for(x=0; x<s->mb_width; x++){
1863 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1865 for(block=0; block<4; block++){
1866 int off= (block& 1) + (block>>1)*wrap;
1867 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1868 int my= s->current_picture.motion_val[0][ xy + off ][1];
1870 if( mx >=range || mx <-range
1871 || my >=range || my <-range){
1872 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1873 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1874 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1887 * @param truncate 1 for truncation, 0 for using intra
1889 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1890 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1892 int y, h_range, v_range;
1894 // RAL: 8 in MPEG-1, 16 in MPEG-4
1895 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1897 if(s->msmpeg4_version) range= 16;
1898 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1901 v_range= field_select_table ? range>>1 : range;
1903 /* clip / convert to intra 16x16 type MVs */
1904 for(y=0; y<s->mb_height; y++){
1906 int xy= y*s->mb_stride;
1907 for(x=0; x<s->mb_width; x++){
1908 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1909 if(field_select_table==NULL || field_select_table[xy] == field_select){
1910 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1911 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1914 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1915 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1916 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1917 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1919 s->mb_type[xy] &= ~type;
1920 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;