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 xmin, int ymin, int xmax, int ymax,
50 int pred_x, int pred_y, Picture *picture,
51 int n, int size, 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);
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, (16>>size));\
87 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
90 #define CMP_QPEL(d, dx, dy, x, y, size)\
92 const int dxy= (dx) + 4*(dy);\
93 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
94 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
97 #include "motion_est_template.c"
105 #define RENAME(a) simple_chroma_ ## a
107 #define CMP(d, x, y, size)\
108 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\
110 int dxy= ((x)&1) + 2*((y)&1);\
111 int c= ((x)>>1) + ((y)>>1)*uvstride;\
113 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
114 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\
115 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
116 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\
119 #define CMP_HPEL(d, dx, dy, x, y, size)\
121 const int dxy= (dx) + 2*(dy);\
122 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
123 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
125 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
126 int c= ((x)>>1) + ((y)>>1)*uvstride;\
127 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
128 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
129 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
130 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
134 #define CMP_QPEL(d, dx, dy, x, y, size)\
136 const int dxy= (dx) + 4*(dy);\
137 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
138 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
141 int cx= (4*(x) + (dx))/2;\
142 int cy= (4*(y) + (dy))/2;\
145 cxy= (cx&1) + 2*(cy&1);\
146 c= ((cx)>>1) + ((cy)>>1)*uvstride;\
147 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
148 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
149 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
150 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
154 #include "motion_est_template.c"
161 /* SIMPLE DIRECT HPEL */
162 #define RENAME(a) simple_direct_hpel_ ## a
163 //FIXME precalc divisions stuff
165 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
166 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
167 const int hx= 2*(x) + (dx);\
168 const int hy= 2*(y) + (dy);\
169 if(s->mv_type==MV_TYPE_8X8){\
172 int fx = s->me.direct_basis_mv[i][0] + hx;\
173 int fy = s->me.direct_basis_mv[i][1] + hy;\
174 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;\
175 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;\
176 int fxy= (fx&1) + 2*(fy&1);\
177 int bxy= (bx&1) + 2*(by&1);\
179 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
180 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
181 hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\
184 int fx = s->me.direct_basis_mv[0][0] + hx;\
185 int fy = s->me.direct_basis_mv[0][1] + hy;\
186 int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\
187 int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\
188 int fxy= (fx&1) + 2*(fy&1);\
189 int bxy= (bx&1) + 2*(by&1);\
191 assert((fx>>1) + 16*s->mb_x >= -16);\
192 assert((fy>>1) + 16*s->mb_y >= -16);\
193 assert((fx>>1) + 16*s->mb_x <= s->width);\
194 assert((fy>>1) + 16*s->mb_y <= s->height);\
195 assert((bx>>1) + 16*s->mb_x >= -16);\
196 assert((by>>1) + 16*s->mb_y >= -16);\
197 assert((bx>>1) + 16*s->mb_x <= s->width);\
198 assert((by>>1) + 16*s->mb_y <= s->height);\
200 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
201 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
203 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
208 #define CMP_HPEL(d, dx, dy, x, y, size)\
209 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
211 #define CMP(d, x, y, size)\
212 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
214 #include "motion_est_template.c"
222 /* SIMPLE DIRECT QPEL */
223 #define RENAME(a) simple_direct_qpel_ ## a
225 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
226 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
227 const int qx= 4*(x) + (dx);\
228 const int qy= 4*(y) + (dy);\
229 if(s->mv_type==MV_TYPE_8X8){\
232 int fx = s->me.direct_basis_mv[i][0] + qx;\
233 int fy = s->me.direct_basis_mv[i][1] + qy;\
234 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;\
235 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;\
236 int fxy= (fx&3) + 4*(fy&3);\
237 int bxy= (bx&3) + 4*(by&3);\
239 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
240 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
241 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
244 int fx = s->me.direct_basis_mv[0][0] + qx;\
245 int fy = s->me.direct_basis_mv[0][1] + qy;\
246 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
247 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
248 int fxy= (fx&3) + 4*(fy&3);\
249 int bxy= (bx&3) + 4*(by&3);\
251 qpel_put[1][fxy](s->me.scratchpad , (ref_y ) + (fx>>2) + (fy>>2)*(stride) , stride);\
252 qpel_put[1][fxy](s->me.scratchpad + 8 , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 , stride);\
253 qpel_put[1][fxy](s->me.scratchpad + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8*stride, stride);\
254 qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
255 qpel_avg[1][bxy](s->me.scratchpad , (ref2_y) + (bx>>2) + (by>>2)*(stride) , stride);\
256 qpel_avg[1][bxy](s->me.scratchpad + 8 , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 , stride);\
257 qpel_avg[1][bxy](s->me.scratchpad + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8*stride, stride);\
258 qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
260 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
265 #define CMP_QPEL(d, dx, dy, x, y, size)\
266 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
268 #define CMP(d, x, y, size)\
269 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
271 #include "motion_est_template.c"
280 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
284 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
285 DSPContext* c= &s->dsp;
288 memset(cmp, 0, sizeof(void*)*11);
296 cmp[0]= c->hadamard8_diff[0];
297 cmp[1]= c->hadamard8_diff[1];
304 cmp[0]= c->dct_sad[0];
305 cmp[1]= c->dct_sad[1];
308 cmp[0]= c->quant_psnr[0];
309 cmp[1]= c->quant_psnr[1];
325 fprintf(stderr,"internal error in cmp function selection\n");
329 static inline int get_penalty_factor(MpegEncContext *s, int type){
339 return s->qscale*s->qscale*2;
344 return (s->qscale*s->qscale*185 + 64)>>7;
348 void ff_init_me(MpegEncContext *s){
349 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
350 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
351 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
352 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
354 if(s->flags&CODEC_FLAG_QPEL){
355 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
356 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
358 s->me.sub_motion_search= simple_qpel_motion_search;
360 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
361 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
362 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
363 && s->avctx-> me_cmp == FF_CMP_SAD
364 && s->avctx-> mb_cmp == FF_CMP_SAD)
365 s->me.sub_motion_search= sad_hpel_motion_search;
367 s->me.sub_motion_search= simple_hpel_motion_search;
370 if(s->avctx->me_cmp&FF_CMP_CHROMA){
371 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
372 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
374 s->me.motion_search[0]= simple_epzs_motion_search;
375 s->me.motion_search[1]= simple_epzs_motion_search4;
378 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
379 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
381 s->me.pre_motion_search= simple_epzs_motion_search;
384 if(s->flags&CODEC_FLAG_QPEL){
385 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
386 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
388 s->me.get_mb_score= simple_qpel_get_mb_score;
390 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
391 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
393 s->me.get_mb_score= simple_hpel_get_mb_score;
398 static int pix_dev(uint8_t * pix, int line_size, int mean)
403 for (i = 0; i < 16; i++) {
404 for (j = 0; j < 16; j += 8) {
405 s += ABS(pix[0]-mean);
406 s += ABS(pix[1]-mean);
407 s += ABS(pix[2]-mean);
408 s += ABS(pix[3]-mean);
409 s += ABS(pix[4]-mean);
410 s += ABS(pix[5]-mean);
411 s += ABS(pix[6]-mean);
412 s += ABS(pix[7]-mean);
415 pix += line_size - 16;
421 static inline void no_motion_search(MpegEncContext * s,
422 int *mx_ptr, int *my_ptr)
424 *mx_ptr = 16 * s->mb_x;
425 *my_ptr = 16 * s->mb_y;
428 static int full_motion_search(MpegEncContext * s,
429 int *mx_ptr, int *my_ptr, int range,
430 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
432 int x1, y1, x2, y2, xx, yy, x, y;
438 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
450 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
454 for (y = y1; y <= y2; y++) {
455 for (x = x1; x <= x2; x++) {
456 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
460 (abs(x - xx) + abs(y - yy)) <
461 (abs(mx - xx) + abs(my - yy)))) {
473 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
474 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
475 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
482 static int log_motion_search(MpegEncContext * s,
483 int *mx_ptr, int *my_ptr, int range,
484 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
486 int x1, y1, x2, y2, xx, yy, x, y;
513 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
519 for (y = y1; y <= y2; y += range) {
520 for (x = x1; x <= x2; x += range) {
521 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
522 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
548 } while (range >= 1);
551 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
558 static int phods_motion_search(MpegEncContext * s,
559 int *mx_ptr, int *my_ptr, int range,
560 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
562 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
563 int mx, my, dminx, dminy;
589 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
600 for (x = x1; x <= x2; x += range) {
601 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
602 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
609 for (y = y1; y <= y2; y += range) {
610 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
611 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
637 } while (range >= 1);
640 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
643 /* half pixel search */
650 #define Z_THRESHOLD 256
652 #define CHECK_SAD_HALF_MV(suffix, x, y) \
654 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
655 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
656 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
659 static inline int sad_hpel_motion_search(MpegEncContext * s,
660 int *mx_ptr, int *my_ptr, int dmin,
661 int xmin, int ymin, int xmax, int ymax,
662 int pred_x, int pred_y, Picture *picture,
663 int n, int size, uint8_t * const mv_penalty)
665 uint8_t *ref_picture= picture->data[0];
666 uint32_t *score_map= s->me.score_map;
667 const int penalty_factor= s->me.sub_penalty_factor;
668 int mx, my, xx, yy, dminh;
670 op_pixels_abs_func pix_abs_x2;
671 op_pixels_abs_func pix_abs_y2;
672 op_pixels_abs_func pix_abs_xy2;
675 pix_abs_x2 = s->dsp.pix_abs16x16_x2;
676 pix_abs_y2 = s->dsp.pix_abs16x16_y2;
677 pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
679 pix_abs_x2 = s->dsp.pix_abs8x8_x2;
680 pix_abs_y2 = s->dsp.pix_abs8x8_y2;
681 pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
692 xx = 16 * s->mb_x + 8*(n&1);
693 yy = 16 * s->mb_y + 8*(n>>1);
694 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
698 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
702 if (mx > xmin && mx < xmax &&
703 my > ymin && my < ymax) {
706 const int index= (my<<ME_MAP_SHIFT) + mx;
707 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
708 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
709 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
710 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
720 CHECK_SAD_HALF_MV(y2 , 0, -1)
722 CHECK_SAD_HALF_MV(xy2, -1, -1)
724 CHECK_SAD_HALF_MV(xy2, +1, -1)
728 CHECK_SAD_HALF_MV(xy2, -1, +1)
730 CHECK_SAD_HALF_MV(x2 , -1, 0)
732 CHECK_SAD_HALF_MV(xy2, +1, -1)
734 CHECK_SAD_HALF_MV(xy2, -1, -1)
738 CHECK_SAD_HALF_MV(xy2, +1, +1)
740 CHECK_SAD_HALF_MV(x2 , +1, 0)
745 CHECK_SAD_HALF_MV(xy2, -1, -1)
749 CHECK_SAD_HALF_MV(xy2, +1, +1)
751 CHECK_SAD_HALF_MV(x2 , -1, 0)
752 CHECK_SAD_HALF_MV(xy2, -1, +1)
755 CHECK_SAD_HALF_MV(xy2, +1, -1)
759 CHECK_SAD_HALF_MV(xy2, -1, +1)
761 CHECK_SAD_HALF_MV(x2 , +1, 0)
762 CHECK_SAD_HALF_MV(xy2, +1, +1)
764 CHECK_SAD_HALF_MV(y2 , 0, +1)
779 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
781 const int xy= s->mb_x + s->mb_y*s->mb_stride;
783 s->p_mv_table[xy][0] = mx;
784 s->p_mv_table[xy][1] = my;
786 /* has allready been set to the 4 MV if 4MV is done */
788 int mot_xy= s->block_index[0];
790 s->motion_val[mot_xy ][0]= mx;
791 s->motion_val[mot_xy ][1]= my;
792 s->motion_val[mot_xy+1][0]= mx;
793 s->motion_val[mot_xy+1][1]= my;
795 mot_xy += s->block_wrap[0];
796 s->motion_val[mot_xy ][0]= mx;
797 s->motion_val[mot_xy ][1]= my;
798 s->motion_val[mot_xy+1][0]= mx;
799 s->motion_val[mot_xy+1][1]= my;
804 * get fullpel ME search limits.
805 * @param range the approximate search range for the old ME code, unused for EPZS and newer
807 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax)
809 if(s->avctx->me_range) *range= s->avctx->me_range >> 1;
812 if (s->unrestricted_mv) {
815 *xmax = s->mb_width*16;
816 *ymax = s->mb_height*16;
820 *xmax = s->mb_width*16 - 16;
821 *ymax = s->mb_height*16 - 16;
824 //FIXME try to limit x/y min/max if me_range is set
827 static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
831 int dmin_sum=0, mx4_sum=0, my4_sum=0;
832 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
834 for(block=0; block<4; block++){
836 int pred_x4, pred_y4;
838 static const int off[4]= {2, 1, 1, -1};
839 const int mot_stride = s->block_wrap[0];
840 const int mot_xy = s->block_index[block];
841 // const int block_x= (block&1);
842 // const int block_y= (block>>1);
843 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
844 const int rel_xmin4= xmin;
845 const int rel_xmax4= xmax;
846 const int rel_ymin4= ymin;
847 const int rel_ymax4= ymax;
849 const int rel_xmin4= xmin - block_x*8;
850 const int rel_xmax4= xmax - block_x*8 + 8;
851 const int rel_ymin4= ymin - block_y*8;
852 const int rel_ymax4= ymax - block_y*8 + 8;
854 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
855 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
857 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift);
859 /* special case for first line */
860 if (s->mb_y == 0 && block<2) {
864 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
865 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
866 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
867 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
868 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift);
869 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
870 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
871 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
873 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
874 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
876 // if(s->out_format == FMT_H263){
877 pred_x4 = P_MEDIAN[0];
878 pred_y4 = P_MEDIAN[1];
880 }else { /* mpeg1 at least */
889 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
890 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
892 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
893 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
895 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
897 const int offset= ((block&1) + (block>>1)*s->linesize)*8;
898 uint8_t *dest_y = s->me.scratchpad + offset;
900 if(s->quarter_sample){
901 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
902 dxy = ((my4 & 3) << 2) | (mx4 & 3);
905 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , s->linesize);
907 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , s->linesize);
909 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
910 dxy = ((my4 & 1) << 1) | (mx4 & 1);
913 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , s->linesize, 8);
915 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , s->linesize, 8);
917 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
921 if(s->quarter_sample){
929 s->motion_val[ s->block_index[block] ][0]= mx4;
930 s->motion_val[ s->block_index[block] ][1]= my4;
933 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
934 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*s->linesize, s->me.scratchpad, s->linesize);
937 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
942 mx= ff_h263_round_chroma(mx4_sum);
943 my= ff_h263_round_chroma(my4_sum);
944 dxy = ((my & 1) << 1) | (mx & 1);
946 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
949 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
950 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
952 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
953 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
956 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);
957 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);
960 switch(s->avctx->mb_cmp&0xFF){
962 return dmin_sum+ 32*s->qscale*s->qscale;*/
966 return dmin_sum+ 11*s->me.mb_penalty_factor;
970 void ff_estimate_p_frame_motion(MpegEncContext * s,
974 int sum, varc, vard, mx, my, range, dmin, xx, yy;
975 int xmin, ymin, xmax, ymax;
976 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
977 int pred_x=0, pred_y=0;
979 const int shift= 1+s->quarter_sample;
981 uint8_t *ref_picture= s->last_picture.data[0];
982 Picture * const pic= &s->current_picture;
983 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
985 assert(s->quarter_sample==0 || s->quarter_sample==1);
987 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
988 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
989 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
991 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
992 rel_xmin= xmin - mb_x*16;
993 rel_xmax= xmax - mb_x*16;
994 rel_ymin= ymin - mb_y*16;
995 rel_ymax= ymax - mb_y*16;
998 switch(s->me_method) {
1001 no_motion_search(s, &mx, &my);
1007 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1012 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1017 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1024 const int mot_stride = s->block_wrap[0];
1025 const int mot_xy = s->block_index[0];
1027 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
1028 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
1030 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1033 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
1034 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
1035 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1036 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1037 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift);
1038 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1039 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1041 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1042 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1044 if(s->out_format == FMT_H263){
1045 pred_x = P_MEDIAN[0];
1046 pred_y = P_MEDIAN[1];
1047 }else { /* mpeg1 at least */
1057 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1058 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1063 /* intra / predictive decision */
1067 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1068 /* At this point (mx,my) are full-pell and the relative displacement */
1069 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1071 sum = s->dsp.pix_sum(pix, s->linesize);
1073 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1074 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1076 //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);
1077 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1078 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1079 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1080 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1081 pic->mb_var_sum += varc;
1082 pic->mc_mb_var_sum += vard;
1083 //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);
1086 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1087 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1089 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1090 if (vard <= 64 || vard < varc)
1091 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1093 s->scene_change_score+= s->qscale;
1095 if (vard*2 + 200 > varc)
1096 mb_type|= MB_TYPE_INTRA;
1097 if (varc*2 + 200 > vard){
1098 mb_type|= MB_TYPE_INTER;
1099 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1100 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1105 if((s->flags&CODEC_FLAG_4MV)
1106 && !s->me.skip && varc>50 && vard>10){
1107 h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1108 mb_type|=MB_TYPE_INTER4V;
1110 set_p_mv_tables(s, mx, my, 0);
1112 set_p_mv_tables(s, mx, my, 1);
1114 mb_type= MB_TYPE_INTER;
1116 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1117 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1119 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1120 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1122 if((s->flags&CODEC_FLAG_4MV)
1123 && !s->me.skip && varc>50 && vard>10){
1124 int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1126 mb_type= MB_TYPE_INTER4V;
1130 pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1131 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1133 if (vard <= 64 || vard < varc) {
1134 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1136 s->scene_change_score+= s->qscale;
1140 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1143 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1146 int mx, my, range, dmin;
1147 int xmin, ymin, xmax, ymax;
1148 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1149 int pred_x=0, pred_y=0;
1151 const int shift= 1+s->quarter_sample;
1152 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1153 const int xy= mb_x + mb_y*s->mb_stride;
1155 assert(s->quarter_sample==0 || s->quarter_sample==1);
1157 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1159 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1160 rel_xmin= xmin - mb_x*16;
1161 rel_xmax= xmax - mb_x*16;
1162 rel_ymin= ymin - mb_y*16;
1163 rel_ymax= ymax - mb_y*16;
1166 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1167 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1169 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift);
1171 /* special case for first line */
1172 if (mb_y == s->mb_height-1) {
1175 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1176 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1178 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1179 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1180 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1181 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1182 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
1183 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1184 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1186 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1187 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1189 pred_x = P_MEDIAN[0];
1190 pred_y = P_MEDIAN[1];
1192 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1193 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1195 s->p_mv_table[xy][0] = mx<<shift;
1196 s->p_mv_table[xy][1] = my<<shift;
1201 static int ff_estimate_motion_b(MpegEncContext * s,
1202 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1204 int mx, my, range, dmin;
1205 int xmin, ymin, xmax, ymax;
1206 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1207 int pred_x=0, pred_y=0;
1209 const int shift= 1+s->quarter_sample;
1210 const int mot_stride = s->mb_stride;
1211 const int mot_xy = mb_y*mot_stride + mb_x;
1212 uint8_t * const ref_picture= picture->data[0];
1213 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1216 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1217 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1218 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1220 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1221 rel_xmin= xmin - mb_x*16;
1222 rel_xmax= xmax - mb_x*16;
1223 rel_ymin= ymin - mb_y*16;
1224 rel_ymax= ymax - mb_y*16;
1226 switch(s->me_method) {
1229 no_motion_search(s, &mx, &my);
1235 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1240 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1245 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1252 P_LEFT[0] = mv_table[mot_xy - 1][0];
1253 P_LEFT[1] = mv_table[mot_xy - 1][1];
1255 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1257 /* special case for first line */
1259 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1260 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1261 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1262 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1263 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1264 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1265 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1267 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1268 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1274 if(mv_table == s->b_forw_mv_table){
1275 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1277 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1280 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1281 picture, s->p_mv_table, mv_scale, mv_penalty);
1286 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1287 pred_x, pred_y, picture, 0, 0, mv_penalty);
1289 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1290 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1292 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1293 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1294 mv_table[mot_xy][0]= mx;
1295 mv_table[mot_xy][1]= my;
1300 static inline int check_bidir_mv(MpegEncContext * s,
1302 int motion_fx, int motion_fy,
1303 int motion_bx, int motion_by,
1304 int pred_fx, int pred_fy,
1305 int pred_bx, int pred_by)
1308 //FIXME move into template?
1309 //FIXME better f_code prediction (max mv & distance)
1310 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1311 uint8_t *dest_y = s->me.scratchpad;
1317 if(s->quarter_sample){
1318 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1319 src_x = mb_x * 16 + (motion_fx >> 2);
1320 src_y = mb_y * 16 + (motion_fy >> 2);
1321 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1322 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1324 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1325 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1327 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1328 src_x = mb_x * 16 + (motion_bx >> 2);
1329 src_y = mb_y * 16 + (motion_by >> 2);
1330 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1331 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1333 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1334 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1336 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1337 src_x = mb_x * 16 + (motion_fx >> 1);
1338 src_y = mb_y * 16 + (motion_fy >> 1);
1339 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1340 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1342 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1343 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1345 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1346 src_x = mb_x * 16 + (motion_bx >> 1);
1347 src_y = mb_y * 16 + (motion_by >> 1);
1348 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1349 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1351 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1352 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1355 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1356 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1357 + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1359 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1366 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1367 static inline int bidir_refine(MpegEncContext * s,
1370 const int mot_stride = s->mb_stride;
1371 const int xy = mb_y *mot_stride + mb_x;
1373 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1374 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1375 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1376 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1377 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1378 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1379 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1380 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1382 //FIXME do refinement and add flag
1384 fbmin= check_bidir_mv(s, mb_x, mb_y,
1385 motion_fx, motion_fy,
1386 motion_bx, motion_by,
1393 static inline int direct_search(MpegEncContext * s,
1397 const int mot_stride = s->mb_stride;
1398 const int mot_xy = mb_y*mot_stride + mb_x;
1399 const int shift= 1+s->quarter_sample;
1401 const int time_pp= s->pp_time;
1402 const int time_pb= s->pb_time;
1403 int mx, my, xmin, xmax, ymin, ymax;
1404 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1405 uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1407 ymin= xmin=(-32)>>shift;
1408 ymax= xmax= 31>>shift;
1410 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1411 s->mv_type= MV_TYPE_8X8;
1413 s->mv_type= MV_TYPE_16X16;
1417 int index= s->block_index[i];
1420 s->me.co_located_mv[i][0]= s->motion_val[index][0];
1421 s->me.co_located_mv[i][1]= s->motion_val[index][1];
1422 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1423 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1424 // 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);
1425 // 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);
1427 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1428 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1429 max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1430 min+= (2*mb_x + (i& 1))*8 - 1;
1431 xmax= FFMIN(xmax, s->width - max);
1432 xmin= FFMAX(xmin, - 16 - min);
1434 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1435 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1436 max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1437 min+= (2*mb_y + (i>>1))*8 - 1;
1438 ymax= FFMIN(ymax, s->height - max);
1439 ymin= FFMAX(ymin, - 16 - min);
1441 if(s->mv_type == MV_TYPE_16X16) break;
1444 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1446 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1447 s->b_direct_mv_table[mot_xy][0]= 0;
1448 s->b_direct_mv_table[mot_xy][1]= 0;
1450 return 256*256*256*64;
1453 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1454 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1456 /* special case for first line */
1458 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1459 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1460 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1461 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1463 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1464 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1467 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1468 if(s->flags&CODEC_FLAG_QPEL){
1469 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1470 &s->last_picture, mv_table, 1<<14, mv_penalty);
1471 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1472 0, 0, &s->last_picture, 0, 0, mv_penalty);
1474 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1475 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1477 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1478 &s->last_picture, mv_table, 1<<15, mv_penalty);
1479 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1480 0, 0, &s->last_picture, 0, 0, mv_penalty);
1482 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1483 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1486 s->b_direct_mv_table[mot_xy][0]= mx;
1487 s->b_direct_mv_table[mot_xy][1]= my;
1491 void ff_estimate_b_frame_motion(MpegEncContext * s,
1494 const int penalty_factor= s->me.mb_penalty_factor;
1495 int fmin, bmin, dmin, fbmin;
1499 if (s->codec_id == CODEC_ID_MPEG4)
1500 dmin= direct_search(s, mb_x, mb_y);
1505 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1508 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1509 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1512 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1513 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1516 type = MB_TYPE_FORWARD;
1520 type = MB_TYPE_DIRECT;
1524 type= MB_TYPE_BACKWARD;
1528 type= MB_TYPE_BIDIR;
1531 score= ((unsigned)(score*score + 128*256))>>16;
1532 s->current_picture.mc_mb_var_sum += score;
1533 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1536 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1537 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1538 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1541 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1544 /* find best f_code for ME which do unlimited searches */
1545 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1547 if(s->me_method>=ME_EPZS){
1550 uint8_t * fcode_tab= s->fcode_tab;
1552 int best_score=-10000000;
1554 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1556 for(y=0; y<s->mb_height; y++){
1558 int xy= y*s->mb_stride;
1559 for(x=0; x<s->mb_width; x++){
1560 if(s->mb_type[xy] & type){
1561 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1562 fcode_tab[mv_table[xy][1] + MAX_MV]);
1565 for(j=0; j<fcode && j<8; j++){
1566 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1575 if(score[i] > best_score){
1576 best_score= score[i];
1579 // printf("%d %d\n", i, score[i]);
1582 // printf("fcode: %d type: %d\n", i, s->pict_type);
1584 /* for(i=0; i<=MAX_FCODE; i++){
1585 printf("%d ", mv_num[i]);
1593 void ff_fix_long_p_mvs(MpegEncContext * s)
1595 const int f_code= s->f_code;
1597 assert(s->pict_type==P_TYPE);
1599 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1601 if(s->msmpeg4_version) range= 16;
1603 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1605 /* clip / convert to intra 16x16 type MVs */
1606 for(y=0; y<s->mb_height; y++){
1608 int xy= y*s->mb_stride;
1609 for(x=0; x<s->mb_width; x++){
1610 if(s->mb_type[xy]&MB_TYPE_INTER){
1611 if( s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
1612 || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
1613 s->mb_type[xy] &= ~MB_TYPE_INTER;
1614 s->mb_type[xy] |= MB_TYPE_INTRA;
1615 s->p_mv_table[xy][0] = 0;
1616 s->p_mv_table[xy][1] = 0;
1622 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1623 if(s->flags&CODEC_FLAG_4MV){
1624 const int wrap= 2+ s->mb_width*2;
1626 /* clip / convert to intra 8x8 type MVs */
1627 for(y=0; y<s->mb_height; y++){
1628 int xy= (y*2 + 1)*wrap + 1;
1629 int i= y*s->mb_stride;
1632 for(x=0; x<s->mb_width; x++){
1633 if(s->mb_type[i]&MB_TYPE_INTER4V){
1635 for(block=0; block<4; block++){
1636 int off= (block& 1) + (block>>1)*wrap;
1637 int mx= s->motion_val[ xy + off ][0];
1638 int my= s->motion_val[ xy + off ][1];
1640 if( mx >=range || mx <-range
1641 || my >=range || my <-range){
1642 s->mb_type[i] &= ~MB_TYPE_INTER4V;
1643 s->mb_type[i] |= MB_TYPE_INTRA;
1654 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1658 // RAL: 8 in MPEG-1, 16 in MPEG-4
1659 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1661 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1663 /* clip / convert to intra 16x16 type MVs */
1664 for(y=0; y<s->mb_height; y++){
1666 int xy= y*s->mb_stride;
1667 for(x=0; x<s->mb_width; x++){
1668 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1669 if( mv_table[xy][0] >=range || mv_table[xy][0] <-range
1670 || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
1672 if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
1674 if (mv_table[xy][0] > range-1) mv_table[xy][0]= range-1;
1675 else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
1676 if (mv_table[xy][1] > range-1) mv_table[xy][1]= range-1;
1677 else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;