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>
27 #include "mpegvideo.h"
32 #define SQ(a) ((a)*(a))
36 #define P_TOPRIGHT P[3]
40 static inline int sad_hpel_motion_search(MpegEncContext * s,
41 int *mx_ptr, int *my_ptr, int dmin,
42 int xmin, int ymin, int xmax, int ymax,
43 int pred_x, int pred_y, Picture *picture,
44 int n, int size, uint16_t * const mv_penalty);
46 static inline int update_map_generation(MpegEncContext * s)
48 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
49 if(s->me.map_generation==0){
50 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
51 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
53 return s->me.map_generation;
56 /* shape adaptive search stuff */
57 typedef struct Minima{
63 static int minima_cmp(const void *a, const void *b){
64 Minima *da = (Minima *) a;
65 Minima *db = (Minima *) b;
67 return da->height - db->height;
71 #define RENAME(a) simple_ ## a
73 #define CMP(d, x, y, size)\
74 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);
76 #define CMP_HPEL(d, dx, dy, x, y, size)\
78 const int dxy= (dx) + 2*(dy);\
79 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
80 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
83 #define CMP_QPEL(d, dx, dy, x, y, size)\
85 const int dxy= (dx) + 4*(dy);\
86 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
87 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
90 #include "motion_est_template.c"
98 #define RENAME(a) simple_chroma_ ## a
100 #define CMP(d, x, y, size)\
101 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\
103 int dxy= ((x)&1) + 2*((y)&1);\
104 int c= ((x)>>1) + ((y)>>1)*uvstride;\
106 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
107 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\
108 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
109 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\
112 #define CMP_HPEL(d, dx, dy, x, y, size)\
114 const int dxy= (dx) + 2*(dy);\
115 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
116 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
118 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
119 int c= ((x)>>1) + ((y)>>1)*uvstride;\
120 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
121 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
122 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
123 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
127 #define CMP_QPEL(d, dx, dy, x, y, size)\
129 const int dxy= (dx) + 4*(dy);\
130 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
131 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
134 int cx= (4*(x) + (dx))/2;\
135 int cy= (4*(y) + (dy))/2;\
138 cxy= (cx&1) + 2*(cy&1);\
139 c= ((cx)>>1) + ((cy)>>1)*uvstride;\
140 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
141 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
142 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
143 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
147 #include "motion_est_template.c"
154 /* SIMPLE DIRECT HPEL */
155 #define RENAME(a) simple_direct_hpel_ ## a
156 //FIXME precalc divisions stuff
158 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
159 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
160 const int hx= 2*(x) + (dx);\
161 const int hy= 2*(y) + (dy);\
162 if(s->mv_type==MV_TYPE_8X8){\
165 int fx = s->me.direct_basis_mv[i][0] + hx;\
166 int fy = s->me.direct_basis_mv[i][1] + hy;\
167 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;\
168 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;\
169 int fxy= (fx&1) + 2*(fy&1);\
170 int bxy= (bx&1) + 2*(by&1);\
172 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
173 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
174 hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\
177 int fx = s->me.direct_basis_mv[0][0] + hx;\
178 int fy = s->me.direct_basis_mv[0][1] + hy;\
179 int bx = hx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
180 int by = hy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
181 int fxy= (fx&1) + 2*(fy&1);\
182 int bxy= (bx&1) + 2*(by&1);\
184 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
185 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
187 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
192 #define CMP_HPEL(d, dx, dy, x, y, size)\
193 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
195 #define CMP(d, x, y, size)\
196 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
198 #include "motion_est_template.c"
206 /* SIMPLE DIRECT QPEL */
207 #define RENAME(a) simple_direct_qpel_ ## a
209 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
210 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
211 const int qx= 4*(x) + (dx);\
212 const int qy= 4*(y) + (dy);\
213 if(s->mv_type==MV_TYPE_8X8){\
216 int fx = s->me.direct_basis_mv[i][0] + qx;\
217 int fy = s->me.direct_basis_mv[i][1] + qy;\
218 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;\
219 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;\
220 int fxy= (fx&3) + 4*(fy&3);\
221 int bxy= (bx&3) + 4*(by&3);\
223 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
224 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
225 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
228 int fx = s->me.direct_basis_mv[0][0] + qx;\
229 int fy = s->me.direct_basis_mv[0][1] + qy;\
230 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
231 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
232 int fxy= (fx&3) + 4*(fy&3);\
233 int bxy= (bx&3) + 4*(by&3);\
235 qpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
236 qpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
238 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
243 #define CMP_QPEL(d, dx, dy, x, y, size)\
244 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
246 #define CMP(d, x, y, size)\
247 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
249 #include "motion_est_template.c"
258 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
262 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
263 DSPContext* c= &s->dsp;
266 memset(cmp, 0, sizeof(void*)*11);
274 cmp[0]= c->hadamard8_diff[0];
275 cmp[1]= c->hadamard8_diff[1];
282 cmp[0]= c->dct_sad[0];
283 cmp[1]= c->dct_sad[1];
286 cmp[0]= c->quant_psnr[0];
287 cmp[1]= c->quant_psnr[1];
303 fprintf(stderr,"internal error in cmp function selection\n");
307 static inline int get_penalty_factor(MpegEncContext *s, int type){
321 return (s->qscale*s->qscale*105 + 64)>>7;
325 void ff_init_me(MpegEncContext *s){
326 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
327 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
328 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
329 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
331 if(s->flags&CODEC_FLAG_QPEL){
332 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
333 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
335 s->me.sub_motion_search= simple_qpel_motion_search;
337 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
338 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
339 else if(s->avctx->me_sub_cmp == FF_CMP_SAD && s->avctx->me_cmp == FF_CMP_SAD)
340 s->me.sub_motion_search= sad_hpel_motion_search;
342 s->me.sub_motion_search= simple_hpel_motion_search;
345 if(s->avctx->me_cmp&FF_CMP_CHROMA){
346 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
347 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
349 s->me.motion_search[0]= simple_epzs_motion_search;
350 s->me.motion_search[1]= simple_epzs_motion_search4;
353 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
354 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
356 s->me.pre_motion_search= simple_epzs_motion_search;
360 static int pix_dev(UINT8 * pix, int line_size, int mean)
365 for (i = 0; i < 16; i++) {
366 for (j = 0; j < 16; j += 8) {
367 s += ABS(pix[0]-mean);
368 s += ABS(pix[1]-mean);
369 s += ABS(pix[2]-mean);
370 s += ABS(pix[3]-mean);
371 s += ABS(pix[4]-mean);
372 s += ABS(pix[5]-mean);
373 s += ABS(pix[6]-mean);
374 s += ABS(pix[7]-mean);
377 pix += line_size - 16;
382 static inline void no_motion_search(MpegEncContext * s,
383 int *mx_ptr, int *my_ptr)
385 *mx_ptr = 16 * s->mb_x;
386 *my_ptr = 16 * s->mb_y;
389 static int full_motion_search(MpegEncContext * s,
390 int *mx_ptr, int *my_ptr, int range,
391 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
393 int x1, y1, x2, y2, xx, yy, x, y;
399 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
411 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
415 for (y = y1; y <= y2; y++) {
416 for (x = x1; x <= x2; x++) {
417 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
421 (abs(x - xx) + abs(y - yy)) <
422 (abs(mx - xx) + abs(my - yy)))) {
434 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
435 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
436 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
443 static int log_motion_search(MpegEncContext * s,
444 int *mx_ptr, int *my_ptr, int range,
445 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
447 int x1, y1, x2, y2, xx, yy, x, y;
474 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
480 for (y = y1; y <= y2; y += range) {
481 for (x = x1; x <= x2; x += range) {
482 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
483 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
509 } while (range >= 1);
512 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
519 static int phods_motion_search(MpegEncContext * s,
520 int *mx_ptr, int *my_ptr, int range,
521 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
523 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
524 int mx, my, dminx, dminy;
550 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
561 for (x = x1; x <= x2; x += range) {
562 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
563 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
570 for (y = y1; y <= y2; y += range) {
571 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
572 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
598 } while (range >= 1);
601 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
604 /* half pixel search */
611 #define Z_THRESHOLD 256
613 #define CHECK_SAD_HALF_MV(suffix, x, y) \
615 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
616 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
617 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
620 static inline int sad_hpel_motion_search(MpegEncContext * s,
621 int *mx_ptr, int *my_ptr, int dmin,
622 int xmin, int ymin, int xmax, int ymax,
623 int pred_x, int pred_y, Picture *picture,
624 int n, int size, uint16_t * const mv_penalty)
626 uint8_t *ref_picture= picture->data[0];
627 uint32_t *score_map= s->me.score_map;
628 const int penalty_factor= s->me.sub_penalty_factor;
629 int mx, my, xx, yy, dminh;
631 op_pixels_abs_func pix_abs_x2;
632 op_pixels_abs_func pix_abs_y2;
633 op_pixels_abs_func pix_abs_xy2;
636 pix_abs_x2 = s->dsp.pix_abs16x16_x2;
637 pix_abs_y2 = s->dsp.pix_abs16x16_y2;
638 pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
640 pix_abs_x2 = s->dsp.pix_abs8x8_x2;
641 pix_abs_y2 = s->dsp.pix_abs8x8_y2;
642 pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
653 xx = 16 * s->mb_x + 8*(n&1);
654 yy = 16 * s->mb_y + 8*(n>>1);
655 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
659 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
663 if (mx > xmin && mx < xmax &&
664 my > ymin && my < ymax) {
667 const int index= (my<<ME_MAP_SHIFT) + mx;
668 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
669 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
670 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
671 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
681 CHECK_SAD_HALF_MV(y2 , 0, -1)
683 CHECK_SAD_HALF_MV(xy2, -1, -1)
685 CHECK_SAD_HALF_MV(xy2, +1, -1)
689 CHECK_SAD_HALF_MV(xy2, -1, +1)
691 CHECK_SAD_HALF_MV(x2 , -1, 0)
693 CHECK_SAD_HALF_MV(xy2, +1, -1)
695 CHECK_SAD_HALF_MV(xy2, -1, -1)
699 CHECK_SAD_HALF_MV(xy2, +1, +1)
701 CHECK_SAD_HALF_MV(x2 , +1, 0)
706 CHECK_SAD_HALF_MV(xy2, -1, -1)
710 CHECK_SAD_HALF_MV(xy2, +1, +1)
712 CHECK_SAD_HALF_MV(x2 , -1, 0)
713 CHECK_SAD_HALF_MV(xy2, -1, +1)
716 CHECK_SAD_HALF_MV(xy2, +1, -1)
720 CHECK_SAD_HALF_MV(xy2, -1, +1)
722 CHECK_SAD_HALF_MV(x2 , +1, 0)
723 CHECK_SAD_HALF_MV(xy2, +1, +1)
725 CHECK_SAD_HALF_MV(y2 , 0, +1)
740 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
742 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
744 s->p_mv_table[xy][0] = mx;
745 s->p_mv_table[xy][1] = my;
747 /* has allready been set to the 4 MV if 4MV is done */
749 int mot_xy= s->block_index[0];
751 s->motion_val[mot_xy ][0]= mx;
752 s->motion_val[mot_xy ][1]= my;
753 s->motion_val[mot_xy+1][0]= mx;
754 s->motion_val[mot_xy+1][1]= my;
756 mot_xy += s->block_wrap[0];
757 s->motion_val[mot_xy ][0]= mx;
758 s->motion_val[mot_xy ][1]= my;
759 s->motion_val[mot_xy+1][0]= mx;
760 s->motion_val[mot_xy+1][1]= my;
764 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
766 *range = 8 * (1 << (f_code - 1));
767 /* XXX: temporary kludge to avoid overflow for msmpeg4 */
768 if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
771 if (s->unrestricted_mv) {
776 if(s->avctx->codec->id!=CODEC_ID_MPEG4){
777 *xmax = s->mb_width*16;
778 *ymax = s->mb_height*16;
786 *xmax = s->mb_width*16 - 16;
787 *ymax = s->mb_height*16 - 16;
791 static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
795 uint8_t *ref_picture= s->last_picture.data[0];
797 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
799 for(block=0; block<4; block++){
801 int pred_x4, pred_y4;
803 static const int off[4]= {2, 1, 1, -1};
804 const int mot_stride = s->block_wrap[0];
805 const int mot_xy = s->block_index[block];
806 // const int block_x= (block&1);
807 // const int block_y= (block>>1);
808 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
809 const int rel_xmin4= xmin;
810 const int rel_xmax4= xmax;
811 const int rel_ymin4= ymin;
812 const int rel_ymax4= ymax;
814 const int rel_xmin4= xmin - block_x*8;
815 const int rel_xmax4= xmax - block_x*8 + 8;
816 const int rel_ymin4= ymin - block_y*8;
817 const int rel_ymax4= ymax - block_y*8 + 8;
819 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
820 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
822 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift);
824 /* special case for first line */
825 if (s->mb_y == 0 && block<2) {
829 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
830 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
831 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
832 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
833 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift);
834 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
835 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
836 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
838 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
839 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
841 if(s->out_format == FMT_H263){
842 pred_x4 = P_MEDIAN[0];
843 pred_y4 = P_MEDIAN[1];
844 }else { /* mpeg1 at least */
852 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
853 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
855 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
856 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
858 s->motion_val[ s->block_index[block] ][0]= mx4;
859 s->motion_val[ s->block_index[block] ][1]= my4;
865 void ff_estimate_p_frame_motion(MpegEncContext * s,
869 int sum, varc, vard, mx, my, range, dmin, xx, yy;
870 int xmin, ymin, xmax, ymax;
871 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
872 int pred_x=0, pred_y=0;
874 const int shift= 1+s->quarter_sample;
876 uint8_t *ref_picture= s->last_picture.data[0];
877 Picture * const pic= &s->current_picture;
878 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
880 assert(s->quarter_sample==0 || s->quarter_sample==1);
882 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
883 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
885 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
886 rel_xmin= xmin - mb_x*16;
887 rel_xmax= xmax - mb_x*16;
888 rel_ymin= ymin - mb_y*16;
889 rel_ymax= ymax - mb_y*16;
892 switch(s->me_method) {
895 no_motion_search(s, &mx, &my);
901 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
906 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
911 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
918 const int mot_stride = s->block_wrap[0];
919 const int mot_xy = s->block_index[0];
921 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
922 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
924 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
927 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
928 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
929 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
930 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
931 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift);
932 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
933 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
935 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
936 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
938 if(s->out_format == FMT_H263){
939 pred_x = P_MEDIAN[0];
940 pred_y = P_MEDIAN[1];
941 }else { /* mpeg1 at least */
951 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
952 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
957 /* intra / predictive decision */
961 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
962 /* At this point (mx,my) are full-pell and the relative displacement */
963 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
965 sum = s->dsp.pix_sum(pix, s->linesize);
967 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
968 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
970 //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);
971 pic->mb_var [s->mb_width * mb_y + mb_x] = varc;
972 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
973 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
974 pic->mb_var_sum += varc;
975 pic->mc_mb_var_sum += vard;
976 //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);
979 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
980 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
982 if(s->flags&CODEC_FLAG_HQ){
983 if (vard <= 64 || vard < varc)
984 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
986 s->scene_change_score+= s->qscale;
988 if (vard*2 + 200 > varc)
989 mb_type|= MB_TYPE_INTRA;
990 if (varc*2 + 200 > vard){
991 mb_type|= MB_TYPE_INTER;
992 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
993 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
998 if((s->flags&CODEC_FLAG_4MV)
999 && !s->me.skip && varc>50 && vard>10){
1000 mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1001 mb_type|=MB_TYPE_INTER4V;
1003 set_p_mv_tables(s, mx, my, 0);
1005 set_p_mv_tables(s, mx, my, 1);
1007 if (vard <= 64 || vard < varc) {
1008 // if (sadP <= 32 || sadP < sadI + 500) {
1009 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1010 mb_type|= MB_TYPE_INTER;
1011 if (s->me_method != ME_ZERO) {
1012 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1013 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1014 if((s->flags&CODEC_FLAG_4MV)
1015 && !s->me.skip && varc>50 && vard>10){
1016 int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1017 if(dmin4 + 128 <dmin)
1018 mb_type= MB_TYPE_INTER4V;
1020 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1029 fprintf(stderr,"\nEarly skip: %d vard: %2d varc: %5d dmin: %d",
1030 skip, vard, varc, dmin);
1034 s->scene_change_score+= s->qscale;
1035 mb_type|= MB_TYPE_INTRA;
1041 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1044 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1047 int mx, my, range, dmin;
1048 int xmin, ymin, xmax, ymax;
1049 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1050 int pred_x=0, pred_y=0;
1052 const int shift= 1+s->quarter_sample;
1053 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1054 const int mv_stride= s->mb_width + 2;
1055 const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;
1057 assert(s->quarter_sample==0 || s->quarter_sample==1);
1059 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1061 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
1062 rel_xmin= xmin - mb_x*16;
1063 rel_xmax= xmax - mb_x*16;
1064 rel_ymin= ymin - mb_y*16;
1065 rel_ymax= ymax - mb_y*16;
1068 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1069 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1071 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift);
1073 /* special case for first line */
1074 if (mb_y == s->mb_height-1) {
1077 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1078 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1080 P_TOP[0] = s->p_mv_table[xy + mv_stride ][0];
1081 P_TOP[1] = s->p_mv_table[xy + mv_stride ][1];
1082 P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];
1083 P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];
1084 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
1085 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1086 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1088 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1089 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1091 pred_x = P_MEDIAN[0];
1092 pred_y = P_MEDIAN[1];
1094 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1095 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1097 s->p_mv_table[xy][0] = mx<<shift;
1098 s->p_mv_table[xy][1] = my<<shift;
1103 int ff_estimate_motion_b(MpegEncContext * s,
1104 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1106 int mx, my, range, dmin;
1107 int xmin, ymin, xmax, ymax;
1108 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1109 int pred_x=0, pred_y=0;
1111 const int shift= 1+s->quarter_sample;
1112 const int mot_stride = s->mb_width + 2;
1113 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1114 uint8_t * const ref_picture= picture->data[0];
1115 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1118 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1119 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1121 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
1122 rel_xmin= xmin - mb_x*16;
1123 rel_xmax= xmax - mb_x*16;
1124 rel_ymin= ymin - mb_y*16;
1125 rel_ymax= ymax - mb_y*16;
1127 switch(s->me_method) {
1130 no_motion_search(s, &mx, &my);
1136 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1141 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1146 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1153 P_LEFT[0] = mv_table[mot_xy - 1][0];
1154 P_LEFT[1] = mv_table[mot_xy - 1][1];
1156 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1158 /* special case for first line */
1160 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1161 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1162 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1163 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1164 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1165 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1166 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_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]);
1175 if(mv_table == s->b_forw_mv_table){
1176 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1178 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1181 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1182 picture, s->p_mv_table, mv_scale, mv_penalty);
1187 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1188 pred_x, pred_y, picture, 0, 0, mv_penalty);
1189 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1190 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1191 mv_table[mot_xy][0]= mx;
1192 mv_table[mot_xy][1]= my;
1197 static inline int check_bidir_mv(MpegEncContext * s,
1199 int motion_fx, int motion_fy,
1200 int motion_bx, int motion_by,
1201 int pred_fx, int pred_fy,
1202 int pred_bx, int pred_by)
1205 //FIXME move into template?
1206 //FIXME better f_code prediction (max mv & distance)
1207 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1208 uint8_t *dest_y = s->me.scratchpad;
1214 if(s->quarter_sample){
1215 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1216 src_x = mb_x * 16 + (motion_fx >> 2);
1217 src_y = mb_y * 16 + (motion_fy >> 2);
1218 assert(src_x >=-16 && src_x<=s->width);
1219 assert(src_y >=-16 && src_y<=s->height);
1221 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1222 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1224 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1225 src_x = mb_x * 16 + (motion_bx >> 2);
1226 src_y = mb_y * 16 + (motion_by >> 2);
1227 assert(src_x >=-16 && src_x<=s->width);
1228 assert(src_y >=-16 && src_y<=s->height);
1230 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1231 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1233 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1234 src_x = mb_x * 16 + (motion_fx >> 1);
1235 src_y = mb_y * 16 + (motion_fy >> 1);
1236 assert(src_x >=-16 && src_x<=s->width);
1237 assert(src_y >=-16 && src_y<=s->height);
1239 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1240 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1242 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1243 src_x = mb_x * 16 + (motion_bx >> 1);
1244 src_y = mb_y * 16 + (motion_by >> 1);
1245 assert(src_x >=-16 && src_x<=s->width);
1246 assert(src_y >=-16 && src_y<=s->height);
1248 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1249 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1252 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.sub_penalty_factor
1253 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.sub_penalty_factor;
1254 + s->dsp.me_sub_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1259 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1260 static inline int bidir_refine(MpegEncContext * s,
1263 const int mot_stride = s->mb_width + 2;
1264 const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
1266 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1267 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1268 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1269 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1270 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1271 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1272 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1273 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1275 //FIXME do refinement and add flag
1277 fbmin= check_bidir_mv(s, mb_x, mb_y,
1278 motion_fx, motion_fy,
1279 motion_bx, motion_by,
1286 static inline int direct_search(MpegEncContext * s,
1290 const int mot_stride = s->mb_width + 2;
1291 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1292 const int shift= 1+s->quarter_sample;
1294 const int time_pp= s->pp_time;
1295 const int time_pb= s->pb_time;
1296 int mx, my, xmin, xmax, ymin, ymax;
1297 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1298 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1300 ymin= xmin=(-32)>>shift;
1301 ymax= xmax= 31>>shift;
1303 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){
1304 s->mv_type= MV_TYPE_8X8;
1306 s->mv_type= MV_TYPE_16X16;
1310 int index= s->block_index[i];
1313 s->me.co_located_mv[i][0]= s->motion_val[index][0];
1314 s->me.co_located_mv[i][1]= s->motion_val[index][1];
1315 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1316 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1317 // 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);
1318 // 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);
1320 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1321 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1322 max+= (2*mb_x + (i& 1))*8 - 1; // +-1 is for the simpler rounding
1323 min+= (2*mb_x + (i& 1))*8 + 1;
1324 xmax= FFMIN(xmax, s->width - max);
1325 xmin= FFMAX(xmin, - 16 - min);
1327 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1328 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1329 max+= (2*mb_y + (i>>1))*8 - 1; // +-1 is for the simpler rounding
1330 min+= (2*mb_y + (i>>1))*8 + 1;
1331 ymax= FFMIN(ymax, s->height - max);
1332 ymin= FFMAX(ymin, - 16 - min);
1334 if(s->mv_type == MV_TYPE_16X16) break;
1337 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1339 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1340 s->b_direct_mv_table[mot_xy][0]= 0;
1341 s->b_direct_mv_table[mot_xy][1]= 0;
1343 return 256*256*256*64;
1346 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1347 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1349 /* special case for first line */
1351 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1352 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1353 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1354 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1356 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1357 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1360 if(s->flags&CODEC_FLAG_QPEL){
1361 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1362 &s->last_picture, mv_table, 1<<14, mv_penalty);
1363 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1364 0, 0, &s->last_picture, 0, 0, mv_penalty);
1366 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1367 &s->last_picture, mv_table, 1<<15, mv_penalty);
1368 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1369 0, 0, &s->last_picture, 0, 0, mv_penalty);
1372 s->b_direct_mv_table[mot_xy][0]= mx;
1373 s->b_direct_mv_table[mot_xy][1]= my;
1377 void ff_estimate_b_frame_motion(MpegEncContext * s,
1380 const int penalty_factor= s->me.penalty_factor;
1381 int fmin, bmin, dmin, fbmin;
1384 dmin= direct_search(s, mb_x, mb_y);
1386 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code);
1387 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) - penalty_factor;
1388 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1390 fbmin= bidir_refine(s, mb_x, mb_y);
1394 type=MB_TYPE_DIRECT;
1398 type= MB_TYPE_FORWARD;
1402 type= MB_TYPE_BACKWARD;
1406 type= MB_TYPE_BIDIR;
1408 score= ((unsigned)(score*score + 128*256))>>16;
1409 s->current_picture.mc_mb_var_sum += score;
1410 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD
1413 if(s->flags&CODEC_FLAG_HQ){
1414 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1415 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1418 s->mb_type[mb_y*s->mb_width + mb_x]= type;
1421 /* find best f_code for ME which do unlimited searches */
1422 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1424 if(s->me_method>=ME_EPZS){
1427 UINT8 * fcode_tab= s->fcode_tab;
1429 int best_score=-10000000;
1431 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1433 for(y=0; y<s->mb_height; y++){
1435 int xy= (y+1)* (s->mb_width+2) + 1;
1437 for(x=0; x<s->mb_width; x++){
1438 if(s->mb_type[i] & type){
1439 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1440 fcode_tab[mv_table[xy][1] + MAX_MV]);
1443 for(j=0; j<fcode && j<8; j++){
1444 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])
1454 if(score[i] > best_score){
1455 best_score= score[i];
1458 // printf("%d %d\n", i, score[i]);
1461 // printf("fcode: %d type: %d\n", i, s->pict_type);
1463 /* for(i=0; i<=MAX_FCODE; i++){
1464 printf("%d ", mv_num[i]);
1472 void ff_fix_long_p_mvs(MpegEncContext * s)
1474 const int f_code= s->f_code;
1476 UINT8 * fcode_tab= s->fcode_tab;
1479 /* clip / convert to intra 16x16 type MVs */
1480 for(y=0; y<s->mb_height; y++){
1482 int xy= (y+1)* (s->mb_width+2)+1;
1483 int i= y*s->mb_width;
1484 for(x=0; x<s->mb_width; x++){
1485 if(s->mb_type[i]&MB_TYPE_INTER){
1486 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
1487 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
1488 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
1489 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
1490 s->mb_type[i] &= ~MB_TYPE_INTER;
1491 s->mb_type[i] |= MB_TYPE_INTRA;
1492 s->p_mv_table[xy][0] = 0;
1493 s->p_mv_table[xy][1] = 0;
1503 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1504 if(s->flags&CODEC_FLAG_4MV){
1505 const int wrap= 2+ s->mb_width*2;
1507 /* clip / convert to intra 8x8 type MVs */
1508 for(y=0; y<s->mb_height; y++){
1509 int xy= (y*2 + 1)*wrap + 1;
1510 int i= y*s->mb_width;
1513 for(x=0; x<s->mb_width; x++){
1514 if(s->mb_type[i]&MB_TYPE_INTER4V){
1516 for(block=0; block<4; block++){
1517 int off= (block& 1) + (block>>1)*wrap;
1518 int mx= s->motion_val[ xy + off ][0];
1519 int my= s->motion_val[ xy + off ][1];
1521 if( fcode_tab[mx + MAX_MV] > f_code
1522 || fcode_tab[mx + MAX_MV] == 0
1523 || fcode_tab[my + MAX_MV] > f_code
1524 || fcode_tab[my + MAX_MV] == 0 ){
1525 s->mb_type[i] &= ~MB_TYPE_INTER4V;
1526 s->mb_type[i] |= MB_TYPE_INTRA;
1537 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1540 UINT8 * fcode_tab= s->fcode_tab;
1542 /* clip / convert to intra 16x16 type MVs */
1543 for(y=0; y<s->mb_height; y++){
1545 int xy= (y+1)* (s->mb_width+2)+1;
1546 int i= y*s->mb_width;
1547 for(x=0; x<s->mb_width; x++){
1548 if( fcode_tab[mv_table[xy][0] + MAX_MV] > f_code
1549 || fcode_tab[mv_table[xy][0] + MAX_MV] == 0){
1550 if(mv_table[xy][0]>0) mv_table[xy][0]= (16<<f_code)-1;
1551 else mv_table[xy][0]= -(16<<f_code);
1553 if( fcode_tab[mv_table[xy][1] + MAX_MV] > f_code
1554 || fcode_tab[mv_table[xy][1] + MAX_MV] == 0){
1555 if(mv_table[xy][1]>0) mv_table[xy][1]= (16<<f_code)-1;
1556 else mv_table[xy][1]= -(16<<f_code);