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){
317 return s->qscale*s->qscale*2;
322 return (s->qscale*s->qscale*185 + 64)>>7;
326 void ff_init_me(MpegEncContext *s){
327 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
328 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
329 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
330 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
332 if(s->flags&CODEC_FLAG_QPEL){
333 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
334 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
336 s->me.sub_motion_search= simple_qpel_motion_search;
338 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
339 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
340 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
341 && s->avctx-> me_cmp == FF_CMP_SAD
342 && s->avctx-> mb_cmp == FF_CMP_SAD)
343 s->me.sub_motion_search= sad_hpel_motion_search;
345 s->me.sub_motion_search= simple_hpel_motion_search;
348 if(s->avctx->me_cmp&FF_CMP_CHROMA){
349 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
350 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
352 s->me.motion_search[0]= simple_epzs_motion_search;
353 s->me.motion_search[1]= simple_epzs_motion_search4;
356 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
357 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
359 s->me.pre_motion_search= simple_epzs_motion_search;
362 if(s->flags&CODEC_FLAG_QPEL){
363 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
364 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
366 s->me.get_mb_score= simple_qpel_get_mb_score;
368 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
369 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
371 s->me.get_mb_score= simple_hpel_get_mb_score;
375 static int pix_dev(UINT8 * pix, int line_size, int mean)
380 for (i = 0; i < 16; i++) {
381 for (j = 0; j < 16; j += 8) {
382 s += ABS(pix[0]-mean);
383 s += ABS(pix[1]-mean);
384 s += ABS(pix[2]-mean);
385 s += ABS(pix[3]-mean);
386 s += ABS(pix[4]-mean);
387 s += ABS(pix[5]-mean);
388 s += ABS(pix[6]-mean);
389 s += ABS(pix[7]-mean);
392 pix += line_size - 16;
397 static inline void no_motion_search(MpegEncContext * s,
398 int *mx_ptr, int *my_ptr)
400 *mx_ptr = 16 * s->mb_x;
401 *my_ptr = 16 * s->mb_y;
404 static int full_motion_search(MpegEncContext * s,
405 int *mx_ptr, int *my_ptr, int range,
406 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
408 int x1, y1, x2, y2, xx, yy, x, y;
414 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
426 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
430 for (y = y1; y <= y2; y++) {
431 for (x = x1; x <= x2; x++) {
432 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
436 (abs(x - xx) + abs(y - yy)) <
437 (abs(mx - xx) + abs(my - yy)))) {
449 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
450 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
451 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
458 static int log_motion_search(MpegEncContext * s,
459 int *mx_ptr, int *my_ptr, int range,
460 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
462 int x1, y1, x2, y2, xx, yy, x, y;
489 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
495 for (y = y1; y <= y2; y += range) {
496 for (x = x1; x <= x2; x += range) {
497 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
498 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
524 } while (range >= 1);
527 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
534 static int phods_motion_search(MpegEncContext * s,
535 int *mx_ptr, int *my_ptr, int range,
536 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
538 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
539 int mx, my, dminx, dminy;
565 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
576 for (x = x1; x <= x2; x += range) {
577 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
578 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
585 for (y = y1; y <= y2; y += range) {
586 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
587 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
613 } while (range >= 1);
616 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
619 /* half pixel search */
626 #define Z_THRESHOLD 256
628 #define CHECK_SAD_HALF_MV(suffix, x, y) \
630 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
631 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
632 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
635 static inline int sad_hpel_motion_search(MpegEncContext * s,
636 int *mx_ptr, int *my_ptr, int dmin,
637 int xmin, int ymin, int xmax, int ymax,
638 int pred_x, int pred_y, Picture *picture,
639 int n, int size, uint16_t * const mv_penalty)
641 uint8_t *ref_picture= picture->data[0];
642 uint32_t *score_map= s->me.score_map;
643 const int penalty_factor= s->me.sub_penalty_factor;
644 int mx, my, xx, yy, dminh;
646 op_pixels_abs_func pix_abs_x2;
647 op_pixels_abs_func pix_abs_y2;
648 op_pixels_abs_func pix_abs_xy2;
651 pix_abs_x2 = s->dsp.pix_abs16x16_x2;
652 pix_abs_y2 = s->dsp.pix_abs16x16_y2;
653 pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
655 pix_abs_x2 = s->dsp.pix_abs8x8_x2;
656 pix_abs_y2 = s->dsp.pix_abs8x8_y2;
657 pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
668 xx = 16 * s->mb_x + 8*(n&1);
669 yy = 16 * s->mb_y + 8*(n>>1);
670 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
674 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
678 if (mx > xmin && mx < xmax &&
679 my > ymin && my < ymax) {
682 const int index= (my<<ME_MAP_SHIFT) + mx;
683 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
684 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
685 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
686 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
696 CHECK_SAD_HALF_MV(y2 , 0, -1)
698 CHECK_SAD_HALF_MV(xy2, -1, -1)
700 CHECK_SAD_HALF_MV(xy2, +1, -1)
704 CHECK_SAD_HALF_MV(xy2, -1, +1)
706 CHECK_SAD_HALF_MV(x2 , -1, 0)
708 CHECK_SAD_HALF_MV(xy2, +1, -1)
710 CHECK_SAD_HALF_MV(xy2, -1, -1)
714 CHECK_SAD_HALF_MV(xy2, +1, +1)
716 CHECK_SAD_HALF_MV(x2 , +1, 0)
721 CHECK_SAD_HALF_MV(xy2, -1, -1)
725 CHECK_SAD_HALF_MV(xy2, +1, +1)
727 CHECK_SAD_HALF_MV(x2 , -1, 0)
728 CHECK_SAD_HALF_MV(xy2, -1, +1)
731 CHECK_SAD_HALF_MV(xy2, +1, -1)
735 CHECK_SAD_HALF_MV(xy2, -1, +1)
737 CHECK_SAD_HALF_MV(x2 , +1, 0)
738 CHECK_SAD_HALF_MV(xy2, +1, +1)
740 CHECK_SAD_HALF_MV(y2 , 0, +1)
755 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
757 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
759 s->p_mv_table[xy][0] = mx;
760 s->p_mv_table[xy][1] = my;
762 /* has allready been set to the 4 MV if 4MV is done */
764 int mot_xy= s->block_index[0];
766 s->motion_val[mot_xy ][0]= mx;
767 s->motion_val[mot_xy ][1]= my;
768 s->motion_val[mot_xy+1][0]= mx;
769 s->motion_val[mot_xy+1][1]= my;
771 mot_xy += s->block_wrap[0];
772 s->motion_val[mot_xy ][0]= mx;
773 s->motion_val[mot_xy ][1]= my;
774 s->motion_val[mot_xy+1][0]= mx;
775 s->motion_val[mot_xy+1][1]= my;
779 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
781 *range = 8 * (1 << (f_code - 1));
782 /* XXX: temporary kludge to avoid overflow for msmpeg4 */
783 if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
786 if (s->unrestricted_mv) {
791 if(s->avctx->codec->id!=CODEC_ID_MPEG4){
792 *xmax = s->mb_width*16;
793 *ymax = s->mb_height*16;
801 *xmax = s->mb_width*16 - 16;
802 *ymax = s->mb_height*16 - 16;
806 static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
810 int dmin_sum=0, mx4_sum=0, my4_sum=0;
811 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
813 for(block=0; block<4; block++){
815 int pred_x4, pred_y4;
817 static const int off[4]= {2, 1, 1, -1};
818 const int mot_stride = s->block_wrap[0];
819 const int mot_xy = s->block_index[block];
820 // const int block_x= (block&1);
821 // const int block_y= (block>>1);
822 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
823 const int rel_xmin4= xmin;
824 const int rel_xmax4= xmax;
825 const int rel_ymin4= ymin;
826 const int rel_ymax4= ymax;
828 const int rel_xmin4= xmin - block_x*8;
829 const int rel_xmax4= xmax - block_x*8 + 8;
830 const int rel_ymin4= ymin - block_y*8;
831 const int rel_ymax4= ymax - block_y*8 + 8;
833 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
834 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
836 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift);
838 /* special case for first line */
839 if (s->mb_y == 0 && block<2) {
843 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
844 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
845 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
846 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
847 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift);
848 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
849 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
850 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
852 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
853 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
855 // if(s->out_format == FMT_H263){
856 pred_x4 = P_MEDIAN[0];
857 pred_y4 = P_MEDIAN[1];
859 }else { /* mpeg1 at least */
868 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
869 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
871 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
872 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
874 if(s->dsp.me_sub_cmp != s->dsp.mb_cmp){
876 const int offset= ((block&1) + (block>>1)*s->linesize)*8;
877 uint8_t *dest_y = s->me.scratchpad + offset;
879 if(s->quarter_sample){
880 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
881 dxy = ((my4 & 3) << 2) | (mx4 & 3);
884 s->dsp.put_no_rnd_qpel_pixels_tab[0][dxy](dest_y , ref , s->linesize);
886 s->dsp.put_qpel_pixels_tab [0][dxy](dest_y , ref , s->linesize);
888 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
889 dxy = ((my4 & 1) << 1) | (mx4 & 1);
892 s->dsp.put_no_rnd_pixels_tab[0][dxy](dest_y , ref , s->linesize, 16);
894 s->dsp.put_pixels_tab [0][dxy](dest_y , ref , s->linesize, 16);
896 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
900 if(s->quarter_sample){
908 s->motion_val[ s->block_index[block] ][0]= mx4;
909 s->motion_val[ s->block_index[block] ][1]= my4;
912 if(s->dsp.me_sub_cmp != s->dsp.mb_cmp){
913 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);
916 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
921 mx= ff_h263_round_chroma(mx4_sum);
922 my= ff_h263_round_chroma(my4_sum);
923 dxy = ((my & 1) << 1) | (mx & 1);
925 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
928 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
929 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
931 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
932 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
935 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);
936 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);
939 switch(s->avctx->mb_cmp&0xFF){
941 return dmin_sum+ 32*s->qscale*s->qscale;*/
945 return dmin_sum+ 11*s->me.mb_penalty_factor;
949 void ff_estimate_p_frame_motion(MpegEncContext * s,
953 int sum, varc, vard, mx, my, range, dmin, xx, yy;
954 int xmin, ymin, xmax, ymax;
955 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
956 int pred_x=0, pred_y=0;
958 const int shift= 1+s->quarter_sample;
960 uint8_t *ref_picture= s->last_picture.data[0];
961 Picture * const pic= &s->current_picture;
962 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
964 assert(s->quarter_sample==0 || s->quarter_sample==1);
966 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
967 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
968 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
970 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
971 rel_xmin= xmin - mb_x*16;
972 rel_xmax= xmax - mb_x*16;
973 rel_ymin= ymin - mb_y*16;
974 rel_ymax= ymax - mb_y*16;
977 switch(s->me_method) {
980 no_motion_search(s, &mx, &my);
986 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
991 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
996 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1003 const int mot_stride = s->block_wrap[0];
1004 const int mot_xy = s->block_index[0];
1006 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
1007 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
1009 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1012 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
1013 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
1014 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1015 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1016 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift);
1017 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1018 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1020 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1021 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1023 if(s->out_format == FMT_H263){
1024 pred_x = P_MEDIAN[0];
1025 pred_y = P_MEDIAN[1];
1026 }else { /* mpeg1 at least */
1036 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1037 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1042 /* intra / predictive decision */
1046 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1047 /* At this point (mx,my) are full-pell and the relative displacement */
1048 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1050 sum = s->dsp.pix_sum(pix, s->linesize);
1052 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1053 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1055 //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);
1056 pic->mb_var [s->mb_width * mb_y + mb_x] = varc;
1057 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
1058 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
1059 // pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
1060 pic->mb_var_sum += varc;
1061 pic->mc_mb_var_sum += vard;
1062 //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);
1065 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1066 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1068 if(s->flags&CODEC_FLAG_HQ){
1069 if (vard <= 64 || vard < varc)
1070 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1072 s->scene_change_score+= s->qscale;
1074 if (vard*2 + 200 > varc)
1075 mb_type|= MB_TYPE_INTRA;
1076 if (varc*2 + 200 > vard){
1077 mb_type|= MB_TYPE_INTER;
1078 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1079 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1084 if((s->flags&CODEC_FLAG_4MV)
1085 && !s->me.skip && varc>50 && vard>10){
1086 h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1087 mb_type|=MB_TYPE_INTER4V;
1089 set_p_mv_tables(s, mx, my, 0);
1091 set_p_mv_tables(s, mx, my, 1);
1093 mb_type= MB_TYPE_INTER;
1095 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1096 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1098 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1099 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1101 if((s->flags&CODEC_FLAG_4MV)
1102 && !s->me.skip && varc>50 && vard>10){
1103 int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1105 mb_type= MB_TYPE_INTER4V;
1109 pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
1110 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1112 if (vard <= 64 || vard < varc) {
1113 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1115 s->scene_change_score+= s->qscale;
1119 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1122 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1125 int mx, my, range, dmin;
1126 int xmin, ymin, xmax, ymax;
1127 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1128 int pred_x=0, pred_y=0;
1130 const int shift= 1+s->quarter_sample;
1131 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1132 const int mv_stride= s->mb_width + 2;
1133 const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;
1135 assert(s->quarter_sample==0 || s->quarter_sample==1);
1137 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1139 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
1140 rel_xmin= xmin - mb_x*16;
1141 rel_xmax= xmax - mb_x*16;
1142 rel_ymin= ymin - mb_y*16;
1143 rel_ymax= ymax - mb_y*16;
1146 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1147 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1149 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift);
1151 /* special case for first line */
1152 if (mb_y == s->mb_height-1) {
1155 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1156 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1158 P_TOP[0] = s->p_mv_table[xy + mv_stride ][0];
1159 P_TOP[1] = s->p_mv_table[xy + mv_stride ][1];
1160 P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];
1161 P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];
1162 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
1163 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1164 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1166 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1167 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1169 pred_x = P_MEDIAN[0];
1170 pred_y = P_MEDIAN[1];
1172 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1173 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1175 s->p_mv_table[xy][0] = mx<<shift;
1176 s->p_mv_table[xy][1] = my<<shift;
1181 int ff_estimate_motion_b(MpegEncContext * s,
1182 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1184 int mx, my, range, dmin;
1185 int xmin, ymin, xmax, ymax;
1186 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1187 int pred_x=0, pred_y=0;
1189 const int shift= 1+s->quarter_sample;
1190 const int mot_stride = s->mb_width + 2;
1191 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1192 uint8_t * const ref_picture= picture->data[0];
1193 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1196 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1197 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1198 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1200 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
1201 rel_xmin= xmin - mb_x*16;
1202 rel_xmax= xmax - mb_x*16;
1203 rel_ymin= ymin - mb_y*16;
1204 rel_ymax= ymax - mb_y*16;
1206 switch(s->me_method) {
1209 no_motion_search(s, &mx, &my);
1215 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1220 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1225 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1232 P_LEFT[0] = mv_table[mot_xy - 1][0];
1233 P_LEFT[1] = mv_table[mot_xy - 1][1];
1235 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1237 /* special case for first line */
1239 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1240 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1241 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1242 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1243 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1244 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1245 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1247 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1248 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1254 if(mv_table == s->b_forw_mv_table){
1255 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1257 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1260 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1261 picture, s->p_mv_table, mv_scale, mv_penalty);
1266 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1267 pred_x, pred_y, picture, 0, 0, mv_penalty);
1269 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1270 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1272 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1273 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1274 mv_table[mot_xy][0]= mx;
1275 mv_table[mot_xy][1]= my;
1280 static inline int check_bidir_mv(MpegEncContext * s,
1282 int motion_fx, int motion_fy,
1283 int motion_bx, int motion_by,
1284 int pred_fx, int pred_fy,
1285 int pred_bx, int pred_by)
1288 //FIXME move into template?
1289 //FIXME better f_code prediction (max mv & distance)
1290 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1291 uint8_t *dest_y = s->me.scratchpad;
1297 if(s->quarter_sample){
1298 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1299 src_x = mb_x * 16 + (motion_fx >> 2);
1300 src_y = mb_y * 16 + (motion_fy >> 2);
1301 assert(src_x >=-16 && src_x<=s->width);
1302 assert(src_y >=-16 && src_y<=s->height);
1304 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1305 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1307 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1308 src_x = mb_x * 16 + (motion_bx >> 2);
1309 src_y = mb_y * 16 + (motion_by >> 2);
1310 assert(src_x >=-16 && src_x<=s->width);
1311 assert(src_y >=-16 && src_y<=s->height);
1313 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1314 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1316 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1317 src_x = mb_x * 16 + (motion_fx >> 1);
1318 src_y = mb_y * 16 + (motion_fy >> 1);
1319 assert(src_x >=-16 && src_x<=s->width);
1320 assert(src_y >=-16 && src_y<=s->height);
1322 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1323 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1325 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1326 src_x = mb_x * 16 + (motion_bx >> 1);
1327 src_y = mb_y * 16 + (motion_by >> 1);
1328 assert(src_x >=-16 && src_x<=s->width);
1329 assert(src_y >=-16 && src_y<=s->height);
1331 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1332 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1335 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1336 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1337 + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1339 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1346 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1347 static inline int bidir_refine(MpegEncContext * s,
1350 const int mot_stride = s->mb_width + 2;
1351 const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
1353 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1354 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1355 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1356 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1357 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1358 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1359 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1360 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1362 //FIXME do refinement and add flag
1364 fbmin= check_bidir_mv(s, mb_x, mb_y,
1365 motion_fx, motion_fy,
1366 motion_bx, motion_by,
1373 static inline int direct_search(MpegEncContext * s,
1377 const int mot_stride = s->mb_width + 2;
1378 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1379 const int shift= 1+s->quarter_sample;
1381 const int time_pp= s->pp_time;
1382 const int time_pb= s->pb_time;
1383 int mx, my, xmin, xmax, ymin, ymax;
1384 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1385 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1387 ymin= xmin=(-32)>>shift;
1388 ymax= xmax= 31>>shift;
1390 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){
1391 s->mv_type= MV_TYPE_8X8;
1393 s->mv_type= MV_TYPE_16X16;
1397 int index= s->block_index[i];
1400 s->me.co_located_mv[i][0]= s->motion_val[index][0];
1401 s->me.co_located_mv[i][1]= s->motion_val[index][1];
1402 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1403 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1404 // 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);
1405 // 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);
1407 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1408 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1409 max+= (2*mb_x + (i& 1))*8 - 1; // +-1 is for the simpler rounding
1410 min+= (2*mb_x + (i& 1))*8 + 1;
1411 xmax= FFMIN(xmax, s->width - max);
1412 xmin= FFMAX(xmin, - 16 - min);
1414 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1415 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1416 max+= (2*mb_y + (i>>1))*8 - 1; // +-1 is for the simpler rounding
1417 min+= (2*mb_y + (i>>1))*8 + 1;
1418 ymax= FFMIN(ymax, s->height - max);
1419 ymin= FFMAX(ymin, - 16 - min);
1421 if(s->mv_type == MV_TYPE_16X16) break;
1424 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1426 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1427 s->b_direct_mv_table[mot_xy][0]= 0;
1428 s->b_direct_mv_table[mot_xy][1]= 0;
1430 return 256*256*256*64;
1433 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1434 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1436 /* special case for first line */
1438 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1439 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1440 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1441 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1443 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1444 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1447 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1448 if(s->flags&CODEC_FLAG_QPEL){
1449 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1450 &s->last_picture, mv_table, 1<<14, mv_penalty);
1451 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1452 0, 0, &s->last_picture, 0, 0, mv_penalty);
1454 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1455 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1457 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1458 &s->last_picture, mv_table, 1<<15, mv_penalty);
1459 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1460 0, 0, &s->last_picture, 0, 0, mv_penalty);
1462 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1463 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1466 s->b_direct_mv_table[mot_xy][0]= mx;
1467 s->b_direct_mv_table[mot_xy][1]= my;
1471 void ff_estimate_b_frame_motion(MpegEncContext * s,
1474 const int penalty_factor= s->me.mb_penalty_factor;
1475 int fmin, bmin, dmin, fbmin;
1478 dmin= direct_search(s, mb_x, mb_y);
1480 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1481 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1482 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1484 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1485 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1488 type=MB_TYPE_DIRECT;
1492 type= MB_TYPE_FORWARD;
1496 type= MB_TYPE_BACKWARD;
1500 type= MB_TYPE_BIDIR;
1503 score= ((unsigned)(score*score + 128*256))>>16;
1504 s->current_picture.mc_mb_var_sum += score;
1505 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSE
1508 if(s->flags&CODEC_FLAG_HQ){
1509 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1510 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1513 s->mb_type[mb_y*s->mb_width + mb_x]= type;
1516 /* find best f_code for ME which do unlimited searches */
1517 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1519 if(s->me_method>=ME_EPZS){
1522 UINT8 * fcode_tab= s->fcode_tab;
1524 int best_score=-10000000;
1526 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1528 for(y=0; y<s->mb_height; y++){
1530 int xy= (y+1)* (s->mb_width+2) + 1;
1532 for(x=0; x<s->mb_width; x++){
1533 if(s->mb_type[i] & type){
1534 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1535 fcode_tab[mv_table[xy][1] + MAX_MV]);
1538 for(j=0; j<fcode && j<8; j++){
1539 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])
1549 if(score[i] > best_score){
1550 best_score= score[i];
1553 // printf("%d %d\n", i, score[i]);
1556 // printf("fcode: %d type: %d\n", i, s->pict_type);
1558 /* for(i=0; i<=MAX_FCODE; i++){
1559 printf("%d ", mv_num[i]);
1567 void ff_fix_long_p_mvs(MpegEncContext * s)
1569 const int f_code= s->f_code;
1571 UINT8 * fcode_tab= s->fcode_tab;
1574 /* clip / convert to intra 16x16 type MVs */
1575 for(y=0; y<s->mb_height; y++){
1577 int xy= (y+1)* (s->mb_width+2)+1;
1578 int i= y*s->mb_width;
1579 for(x=0; x<s->mb_width; x++){
1580 if(s->mb_type[i]&MB_TYPE_INTER){
1581 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
1582 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
1583 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
1584 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
1585 s->mb_type[i] &= ~MB_TYPE_INTER;
1586 s->mb_type[i] |= MB_TYPE_INTRA;
1587 s->p_mv_table[xy][0] = 0;
1588 s->p_mv_table[xy][1] = 0;
1598 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1599 if(s->flags&CODEC_FLAG_4MV){
1600 const int wrap= 2+ s->mb_width*2;
1602 /* clip / convert to intra 8x8 type MVs */
1603 for(y=0; y<s->mb_height; y++){
1604 int xy= (y*2 + 1)*wrap + 1;
1605 int i= y*s->mb_width;
1608 for(x=0; x<s->mb_width; x++){
1609 if(s->mb_type[i]&MB_TYPE_INTER4V){
1611 for(block=0; block<4; block++){
1612 int off= (block& 1) + (block>>1)*wrap;
1613 int mx= s->motion_val[ xy + off ][0];
1614 int my= s->motion_val[ xy + off ][1];
1616 if( fcode_tab[mx + MAX_MV] > f_code
1617 || fcode_tab[mx + MAX_MV] == 0
1618 || fcode_tab[my + MAX_MV] > f_code
1619 || fcode_tab[my + MAX_MV] == 0 ){
1620 s->mb_type[i] &= ~MB_TYPE_INTER4V;
1621 s->mb_type[i] |= MB_TYPE_INTRA;
1632 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1635 UINT8 * fcode_tab= s->fcode_tab;
1637 /* clip / convert to intra 16x16 type MVs */
1638 for(y=0; y<s->mb_height; y++){
1640 int xy= (y+1)* (s->mb_width+2)+1;
1641 int i= y*s->mb_width;
1642 for(x=0; x<s->mb_width; x++){
1643 if( fcode_tab[mv_table[xy][0] + MAX_MV] > f_code
1644 || fcode_tab[mv_table[xy][0] + MAX_MV] == 0){
1645 if(mv_table[xy][0]>0) mv_table[xy][0]= (16<<f_code)-1;
1646 else mv_table[xy][0]= -(16<<f_code);
1648 if( fcode_tab[mv_table[xy][1] + MAX_MV] > f_code
1649 || fcode_tab[mv_table[xy][1] + MAX_MV] == 0){
1650 if(mv_table[xy][1]>0) mv_table[xy][1]= (16<<f_code)-1;
1651 else mv_table[xy][1]= -(16<<f_code);