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 const Minima *da = (const Minima *) a;
65 const Minima *db = (const 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 assert((fx>>1) + 16*s->mb_x >= -16);\
185 assert((fy>>1) + 16*s->mb_y >= -16);\
186 assert((fx>>1) + 16*s->mb_x <= s->width);\
187 assert((fy>>1) + 16*s->mb_y <= s->height);\
188 assert((bx>>1) + 16*s->mb_x >= -16);\
189 assert((by>>1) + 16*s->mb_y >= -16);\
190 assert((bx>>1) + 16*s->mb_x <= s->width);\
191 assert((by>>1) + 16*s->mb_y <= s->height);\
193 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
194 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
196 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
201 #define CMP_HPEL(d, dx, dy, x, y, size)\
202 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
204 #define CMP(d, x, y, size)\
205 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
207 #include "motion_est_template.c"
215 /* SIMPLE DIRECT QPEL */
216 #define RENAME(a) simple_direct_qpel_ ## a
218 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
219 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
220 const int qx= 4*(x) + (dx);\
221 const int qy= 4*(y) + (dy);\
222 if(s->mv_type==MV_TYPE_8X8){\
225 int fx = s->me.direct_basis_mv[i][0] + qx;\
226 int fy = s->me.direct_basis_mv[i][1] + qy;\
227 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;\
228 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;\
229 int fxy= (fx&3) + 4*(fy&3);\
230 int bxy= (bx&3) + 4*(by&3);\
232 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
233 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
234 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
237 int fx = s->me.direct_basis_mv[0][0] + qx;\
238 int fy = s->me.direct_basis_mv[0][1] + qy;\
239 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
240 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
241 int fxy= (fx&3) + 4*(fy&3);\
242 int bxy= (bx&3) + 4*(by&3);\
244 qpel_put[1][fxy](s->me.scratchpad , (ref_y ) + (fx>>2) + (fy>>2)*(stride) , stride);\
245 qpel_put[1][fxy](s->me.scratchpad + 8 , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 , stride);\
246 qpel_put[1][fxy](s->me.scratchpad + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8*stride, stride);\
247 qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
248 qpel_avg[1][bxy](s->me.scratchpad , (ref2_y) + (bx>>2) + (by>>2)*(stride) , stride);\
249 qpel_avg[1][bxy](s->me.scratchpad + 8 , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 , stride);\
250 qpel_avg[1][bxy](s->me.scratchpad + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8*stride, stride);\
251 qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
253 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
258 #define CMP_QPEL(d, dx, dy, x, y, size)\
259 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
261 #define CMP(d, x, y, size)\
262 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
264 #include "motion_est_template.c"
273 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
277 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
278 DSPContext* c= &s->dsp;
281 memset(cmp, 0, sizeof(void*)*11);
289 cmp[0]= c->hadamard8_diff[0];
290 cmp[1]= c->hadamard8_diff[1];
297 cmp[0]= c->dct_sad[0];
298 cmp[1]= c->dct_sad[1];
301 cmp[0]= c->quant_psnr[0];
302 cmp[1]= c->quant_psnr[1];
318 fprintf(stderr,"internal error in cmp function selection\n");
322 static inline int get_penalty_factor(MpegEncContext *s, int type){
332 return s->qscale*s->qscale*2;
337 return (s->qscale*s->qscale*185 + 64)>>7;
341 void ff_init_me(MpegEncContext *s){
342 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
343 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
344 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
345 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
347 if(s->flags&CODEC_FLAG_QPEL){
348 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
349 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
351 s->me.sub_motion_search= simple_qpel_motion_search;
353 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
354 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
355 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
356 && s->avctx-> me_cmp == FF_CMP_SAD
357 && s->avctx-> mb_cmp == FF_CMP_SAD)
358 s->me.sub_motion_search= sad_hpel_motion_search;
360 s->me.sub_motion_search= simple_hpel_motion_search;
363 if(s->avctx->me_cmp&FF_CMP_CHROMA){
364 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
365 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
367 s->me.motion_search[0]= simple_epzs_motion_search;
368 s->me.motion_search[1]= simple_epzs_motion_search4;
371 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
372 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
374 s->me.pre_motion_search= simple_epzs_motion_search;
377 if(s->flags&CODEC_FLAG_QPEL){
378 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
379 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
381 s->me.get_mb_score= simple_qpel_get_mb_score;
383 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
384 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
386 s->me.get_mb_score= simple_hpel_get_mb_score;
390 static int pix_dev(UINT8 * pix, int line_size, int mean)
395 for (i = 0; i < 16; i++) {
396 for (j = 0; j < 16; j += 8) {
397 s += ABS(pix[0]-mean);
398 s += ABS(pix[1]-mean);
399 s += ABS(pix[2]-mean);
400 s += ABS(pix[3]-mean);
401 s += ABS(pix[4]-mean);
402 s += ABS(pix[5]-mean);
403 s += ABS(pix[6]-mean);
404 s += ABS(pix[7]-mean);
407 pix += line_size - 16;
412 static inline void no_motion_search(MpegEncContext * s,
413 int *mx_ptr, int *my_ptr)
415 *mx_ptr = 16 * s->mb_x;
416 *my_ptr = 16 * s->mb_y;
419 static int full_motion_search(MpegEncContext * s,
420 int *mx_ptr, int *my_ptr, int range,
421 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
423 int x1, y1, x2, y2, xx, yy, x, y;
429 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
441 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
445 for (y = y1; y <= y2; y++) {
446 for (x = x1; x <= x2; x++) {
447 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
451 (abs(x - xx) + abs(y - yy)) <
452 (abs(mx - xx) + abs(my - yy)))) {
464 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
465 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
466 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
473 static int log_motion_search(MpegEncContext * s,
474 int *mx_ptr, int *my_ptr, int range,
475 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
477 int x1, y1, x2, y2, xx, yy, x, y;
504 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
510 for (y = y1; y <= y2; y += range) {
511 for (x = x1; x <= x2; x += range) {
512 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
513 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
539 } while (range >= 1);
542 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
549 static int phods_motion_search(MpegEncContext * s,
550 int *mx_ptr, int *my_ptr, int range,
551 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
553 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
554 int mx, my, dminx, dminy;
580 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
591 for (x = x1; x <= x2; x += range) {
592 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
593 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
600 for (y = y1; y <= y2; y += range) {
601 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
602 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
628 } while (range >= 1);
631 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
634 /* half pixel search */
641 #define Z_THRESHOLD 256
643 #define CHECK_SAD_HALF_MV(suffix, x, y) \
645 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
646 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
647 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
650 static inline int sad_hpel_motion_search(MpegEncContext * s,
651 int *mx_ptr, int *my_ptr, int dmin,
652 int xmin, int ymin, int xmax, int ymax,
653 int pred_x, int pred_y, Picture *picture,
654 int n, int size, uint16_t * const mv_penalty)
656 uint8_t *ref_picture= picture->data[0];
657 uint32_t *score_map= s->me.score_map;
658 const int penalty_factor= s->me.sub_penalty_factor;
659 int mx, my, xx, yy, dminh;
661 op_pixels_abs_func pix_abs_x2;
662 op_pixels_abs_func pix_abs_y2;
663 op_pixels_abs_func pix_abs_xy2;
666 pix_abs_x2 = s->dsp.pix_abs16x16_x2;
667 pix_abs_y2 = s->dsp.pix_abs16x16_y2;
668 pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
670 pix_abs_x2 = s->dsp.pix_abs8x8_x2;
671 pix_abs_y2 = s->dsp.pix_abs8x8_y2;
672 pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
683 xx = 16 * s->mb_x + 8*(n&1);
684 yy = 16 * s->mb_y + 8*(n>>1);
685 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
689 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
693 if (mx > xmin && mx < xmax &&
694 my > ymin && my < ymax) {
697 const int index= (my<<ME_MAP_SHIFT) + mx;
698 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
699 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
700 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
701 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
711 CHECK_SAD_HALF_MV(y2 , 0, -1)
713 CHECK_SAD_HALF_MV(xy2, -1, -1)
715 CHECK_SAD_HALF_MV(xy2, +1, -1)
719 CHECK_SAD_HALF_MV(xy2, -1, +1)
721 CHECK_SAD_HALF_MV(x2 , -1, 0)
723 CHECK_SAD_HALF_MV(xy2, +1, -1)
725 CHECK_SAD_HALF_MV(xy2, -1, -1)
729 CHECK_SAD_HALF_MV(xy2, +1, +1)
731 CHECK_SAD_HALF_MV(x2 , +1, 0)
736 CHECK_SAD_HALF_MV(xy2, -1, -1)
740 CHECK_SAD_HALF_MV(xy2, +1, +1)
742 CHECK_SAD_HALF_MV(x2 , -1, 0)
743 CHECK_SAD_HALF_MV(xy2, -1, +1)
746 CHECK_SAD_HALF_MV(xy2, +1, -1)
750 CHECK_SAD_HALF_MV(xy2, -1, +1)
752 CHECK_SAD_HALF_MV(x2 , +1, 0)
753 CHECK_SAD_HALF_MV(xy2, +1, +1)
755 CHECK_SAD_HALF_MV(y2 , 0, +1)
770 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
772 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
774 s->p_mv_table[xy][0] = mx;
775 s->p_mv_table[xy][1] = my;
777 /* has allready been set to the 4 MV if 4MV is done */
779 int mot_xy= s->block_index[0];
781 s->motion_val[mot_xy ][0]= mx;
782 s->motion_val[mot_xy ][1]= my;
783 s->motion_val[mot_xy+1][0]= mx;
784 s->motion_val[mot_xy+1][1]= my;
786 mot_xy += s->block_wrap[0];
787 s->motion_val[mot_xy ][0]= mx;
788 s->motion_val[mot_xy ][1]= my;
789 s->motion_val[mot_xy+1][0]= mx;
790 s->motion_val[mot_xy+1][1]= my;
794 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
796 *range = 8 * (1 << (f_code - 1));
797 /* XXX: temporary kludge to avoid overflow for msmpeg4 */
798 if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
801 if (s->unrestricted_mv) {
806 if(s->avctx->codec->id!=CODEC_ID_MPEG4){
807 *xmax = s->mb_width*16;
808 *ymax = s->mb_height*16;
816 *xmax = s->mb_width*16 - 16;
817 *ymax = s->mb_height*16 - 16;
821 static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
825 int dmin_sum=0, mx4_sum=0, my4_sum=0;
826 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
828 for(block=0; block<4; block++){
830 int pred_x4, pred_y4;
832 static const int off[4]= {2, 1, 1, -1};
833 const int mot_stride = s->block_wrap[0];
834 const int mot_xy = s->block_index[block];
835 // const int block_x= (block&1);
836 // const int block_y= (block>>1);
837 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
838 const int rel_xmin4= xmin;
839 const int rel_xmax4= xmax;
840 const int rel_ymin4= ymin;
841 const int rel_ymax4= ymax;
843 const int rel_xmin4= xmin - block_x*8;
844 const int rel_xmax4= xmax - block_x*8 + 8;
845 const int rel_ymin4= ymin - block_y*8;
846 const int rel_ymax4= ymax - block_y*8 + 8;
848 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
849 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
851 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift);
853 /* special case for first line */
854 if (s->mb_y == 0 && block<2) {
858 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
859 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
860 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
861 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
862 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift);
863 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
864 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
865 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
867 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
868 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
870 // if(s->out_format == FMT_H263){
871 pred_x4 = P_MEDIAN[0];
872 pred_y4 = P_MEDIAN[1];
874 }else { /* mpeg1 at least */
883 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
884 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
886 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
887 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
889 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
891 const int offset= ((block&1) + (block>>1)*s->linesize)*8;
892 uint8_t *dest_y = s->me.scratchpad + offset;
894 if(s->quarter_sample){
895 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
896 dxy = ((my4 & 3) << 2) | (mx4 & 3);
899 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , s->linesize);
901 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , s->linesize);
903 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
904 dxy = ((my4 & 1) << 1) | (mx4 & 1);
907 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , s->linesize, 8);
909 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , s->linesize, 8);
911 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
915 if(s->quarter_sample){
923 s->motion_val[ s->block_index[block] ][0]= mx4;
924 s->motion_val[ s->block_index[block] ][1]= my4;
927 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
928 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);
931 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
936 mx= ff_h263_round_chroma(mx4_sum);
937 my= ff_h263_round_chroma(my4_sum);
938 dxy = ((my & 1) << 1) | (mx & 1);
940 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
943 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
944 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
946 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
947 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
950 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);
951 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);
954 switch(s->avctx->mb_cmp&0xFF){
956 return dmin_sum+ 32*s->qscale*s->qscale;*/
960 return dmin_sum+ 11*s->me.mb_penalty_factor;
964 void ff_estimate_p_frame_motion(MpegEncContext * s,
968 int sum, varc, vard, mx, my, range, dmin, xx, yy;
969 int xmin, ymin, xmax, ymax;
970 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
971 int pred_x=0, pred_y=0;
973 const int shift= 1+s->quarter_sample;
975 uint8_t *ref_picture= s->last_picture.data[0];
976 Picture * const pic= &s->current_picture;
977 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
979 assert(s->quarter_sample==0 || s->quarter_sample==1);
981 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
982 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
983 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
985 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
986 rel_xmin= xmin - mb_x*16;
987 rel_xmax= xmax - mb_x*16;
988 rel_ymin= ymin - mb_y*16;
989 rel_ymax= ymax - mb_y*16;
992 switch(s->me_method) {
995 no_motion_search(s, &mx, &my);
1001 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1006 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1011 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1018 const int mot_stride = s->block_wrap[0];
1019 const int mot_xy = s->block_index[0];
1021 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
1022 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
1024 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1027 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
1028 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
1029 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1030 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1031 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift);
1032 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1033 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1035 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1036 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1038 if(s->out_format == FMT_H263){
1039 pred_x = P_MEDIAN[0];
1040 pred_y = P_MEDIAN[1];
1041 }else { /* mpeg1 at least */
1051 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1052 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1057 /* intra / predictive decision */
1061 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1062 /* At this point (mx,my) are full-pell and the relative displacement */
1063 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1065 sum = s->dsp.pix_sum(pix, s->linesize);
1067 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1068 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1070 //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);
1071 pic->mb_var [s->mb_width * mb_y + mb_x] = varc;
1072 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
1073 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
1074 // pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
1075 pic->mb_var_sum += varc;
1076 pic->mc_mb_var_sum += vard;
1077 //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);
1080 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1081 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1083 if(s->flags&CODEC_FLAG_HQ){
1084 if (vard <= 64 || vard < varc)
1085 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1087 s->scene_change_score+= s->qscale;
1089 if (vard*2 + 200 > varc)
1090 mb_type|= MB_TYPE_INTRA;
1091 if (varc*2 + 200 > vard){
1092 mb_type|= MB_TYPE_INTER;
1093 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1094 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1099 if((s->flags&CODEC_FLAG_4MV)
1100 && !s->me.skip && varc>50 && vard>10){
1101 h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1102 mb_type|=MB_TYPE_INTER4V;
1104 set_p_mv_tables(s, mx, my, 0);
1106 set_p_mv_tables(s, mx, my, 1);
1108 mb_type= MB_TYPE_INTER;
1110 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1111 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1113 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1114 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1116 if((s->flags&CODEC_FLAG_4MV)
1117 && !s->me.skip && varc>50 && vard>10){
1118 int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1120 mb_type= MB_TYPE_INTER4V;
1124 pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
1125 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1127 if (vard <= 64 || vard < varc) {
1128 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1130 s->scene_change_score+= s->qscale;
1134 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1137 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1140 int mx, my, range, dmin;
1141 int xmin, ymin, xmax, ymax;
1142 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1143 int pred_x=0, pred_y=0;
1145 const int shift= 1+s->quarter_sample;
1146 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1147 const int mv_stride= s->mb_width + 2;
1148 const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;
1150 assert(s->quarter_sample==0 || s->quarter_sample==1);
1152 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1154 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
1155 rel_xmin= xmin - mb_x*16;
1156 rel_xmax= xmax - mb_x*16;
1157 rel_ymin= ymin - mb_y*16;
1158 rel_ymax= ymax - mb_y*16;
1161 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1162 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1164 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift);
1166 /* special case for first line */
1167 if (mb_y == s->mb_height-1) {
1170 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1171 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1173 P_TOP[0] = s->p_mv_table[xy + mv_stride ][0];
1174 P_TOP[1] = s->p_mv_table[xy + mv_stride ][1];
1175 P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];
1176 P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];
1177 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
1178 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1179 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1181 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1182 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1184 pred_x = P_MEDIAN[0];
1185 pred_y = P_MEDIAN[1];
1187 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1188 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1190 s->p_mv_table[xy][0] = mx<<shift;
1191 s->p_mv_table[xy][1] = my<<shift;
1196 static int ff_estimate_motion_b(MpegEncContext * s,
1197 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1199 int mx, my, range, dmin;
1200 int xmin, ymin, xmax, ymax;
1201 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1202 int pred_x=0, pred_y=0;
1204 const int shift= 1+s->quarter_sample;
1205 const int mot_stride = s->mb_width + 2;
1206 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1207 uint8_t * const ref_picture= picture->data[0];
1208 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1211 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1212 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1213 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1215 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
1216 rel_xmin= xmin - mb_x*16;
1217 rel_xmax= xmax - mb_x*16;
1218 rel_ymin= ymin - mb_y*16;
1219 rel_ymax= ymax - mb_y*16;
1221 switch(s->me_method) {
1224 no_motion_search(s, &mx, &my);
1230 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1235 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1240 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1247 P_LEFT[0] = mv_table[mot_xy - 1][0];
1248 P_LEFT[1] = mv_table[mot_xy - 1][1];
1250 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1252 /* special case for first line */
1254 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1255 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1256 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1257 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1258 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1259 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1260 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1262 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1263 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1269 if(mv_table == s->b_forw_mv_table){
1270 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1272 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1275 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1276 picture, s->p_mv_table, mv_scale, mv_penalty);
1281 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1282 pred_x, pred_y, picture, 0, 0, mv_penalty);
1284 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1285 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1287 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1288 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1289 mv_table[mot_xy][0]= mx;
1290 mv_table[mot_xy][1]= my;
1295 static inline int check_bidir_mv(MpegEncContext * s,
1297 int motion_fx, int motion_fy,
1298 int motion_bx, int motion_by,
1299 int pred_fx, int pred_fy,
1300 int pred_bx, int pred_by)
1303 //FIXME move into template?
1304 //FIXME better f_code prediction (max mv & distance)
1305 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1306 uint8_t *dest_y = s->me.scratchpad;
1312 if(s->quarter_sample){
1313 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1314 src_x = mb_x * 16 + (motion_fx >> 2);
1315 src_y = mb_y * 16 + (motion_fy >> 2);
1316 assert(src_x >=-16 && src_x<=s->width);
1317 assert(src_y >=-16 && src_y<=s->height);
1319 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1320 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1322 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1323 src_x = mb_x * 16 + (motion_bx >> 2);
1324 src_y = mb_y * 16 + (motion_by >> 2);
1325 assert(src_x >=-16 && src_x<=s->width);
1326 assert(src_y >=-16 && src_y<=s->height);
1328 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1329 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1331 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1332 src_x = mb_x * 16 + (motion_fx >> 1);
1333 src_y = mb_y * 16 + (motion_fy >> 1);
1334 assert(src_x >=-16 && src_x<=s->width);
1335 assert(src_y >=-16 && src_y<=s->height);
1337 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1338 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1340 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1341 src_x = mb_x * 16 + (motion_bx >> 1);
1342 src_y = mb_y * 16 + (motion_by >> 1);
1343 assert(src_x >=-16 && src_x<=s->width);
1344 assert(src_y >=-16 && src_y<=s->height);
1346 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1347 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1350 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1351 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1352 + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1354 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1361 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1362 static inline int bidir_refine(MpegEncContext * s,
1365 const int mot_stride = s->mb_width + 2;
1366 const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
1368 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1369 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1370 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1371 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1372 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1373 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1374 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1375 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1377 //FIXME do refinement and add flag
1379 fbmin= check_bidir_mv(s, mb_x, mb_y,
1380 motion_fx, motion_fy,
1381 motion_bx, motion_by,
1388 static inline int direct_search(MpegEncContext * s,
1392 const int mot_stride = s->mb_width + 2;
1393 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1394 const int shift= 1+s->quarter_sample;
1396 const int time_pp= s->pp_time;
1397 const int time_pb= s->pb_time;
1398 int mx, my, xmin, xmax, ymin, ymax;
1399 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1400 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1402 ymin= xmin=(-32)>>shift;
1403 ymax= xmax= 31>>shift;
1405 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){
1406 s->mv_type= MV_TYPE_8X8;
1408 s->mv_type= MV_TYPE_16X16;
1412 int index= s->block_index[i];
1415 s->me.co_located_mv[i][0]= s->motion_val[index][0];
1416 s->me.co_located_mv[i][1]= s->motion_val[index][1];
1417 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1418 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1419 // 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);
1420 // 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);
1422 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1423 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1424 max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1425 min+= (2*mb_x + (i& 1))*8 - 1;
1426 xmax= FFMIN(xmax, s->width - max);
1427 xmin= FFMAX(xmin, - 16 - min);
1429 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1430 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1431 max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1432 min+= (2*mb_y + (i>>1))*8 - 1;
1433 ymax= FFMIN(ymax, s->height - max);
1434 ymin= FFMAX(ymin, - 16 - min);
1436 if(s->mv_type == MV_TYPE_16X16) break;
1439 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1441 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1442 s->b_direct_mv_table[mot_xy][0]= 0;
1443 s->b_direct_mv_table[mot_xy][1]= 0;
1445 return 256*256*256*64;
1448 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1449 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1451 /* special case for first line */
1453 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1454 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1455 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1456 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1458 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1459 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1462 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1463 if(s->flags&CODEC_FLAG_QPEL){
1464 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1465 &s->last_picture, mv_table, 1<<14, mv_penalty);
1466 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1467 0, 0, &s->last_picture, 0, 0, mv_penalty);
1469 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1470 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1472 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1473 &s->last_picture, mv_table, 1<<15, mv_penalty);
1474 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1475 0, 0, &s->last_picture, 0, 0, mv_penalty);
1477 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1478 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1481 s->b_direct_mv_table[mot_xy][0]= mx;
1482 s->b_direct_mv_table[mot_xy][1]= my;
1486 void ff_estimate_b_frame_motion(MpegEncContext * s,
1489 const int penalty_factor= s->me.mb_penalty_factor;
1490 int fmin, bmin, dmin, fbmin;
1493 dmin= direct_search(s, mb_x, mb_y);
1495 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1496 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1497 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1499 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1500 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1503 type = MB_TYPE_FORWARD;
1505 // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)
1506 if (s->codec_id != CODEC_ID_MPEG1VIDEO && dmin <= score){
1508 type = MB_TYPE_DIRECT;
1512 type= MB_TYPE_BACKWARD;
1516 type= MB_TYPE_BIDIR;
1519 score= ((unsigned)(score*score + 128*256))>>16;
1520 s->current_picture.mc_mb_var_sum += score;
1521 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSE
1524 if(s->flags&CODEC_FLAG_HQ){
1525 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1526 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1529 s->mb_type[mb_y*s->mb_width + mb_x]= type;
1532 /* find best f_code for ME which do unlimited searches */
1533 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1535 if(s->me_method>=ME_EPZS){
1538 UINT8 * fcode_tab= s->fcode_tab;
1540 int best_score=-10000000;
1542 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1544 for(y=0; y<s->mb_height; y++){
1546 int xy= (y+1)* (s->mb_width+2) + 1;
1548 for(x=0; x<s->mb_width; x++){
1549 if(s->mb_type[i] & type){
1550 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1551 fcode_tab[mv_table[xy][1] + MAX_MV]);
1554 for(j=0; j<fcode && j<8; j++){
1555 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])
1565 if(score[i] > best_score){
1566 best_score= score[i];
1569 // printf("%d %d\n", i, score[i]);
1572 // printf("fcode: %d type: %d\n", i, s->pict_type);
1574 /* for(i=0; i<=MAX_FCODE; i++){
1575 printf("%d ", mv_num[i]);
1583 void ff_fix_long_p_mvs(MpegEncContext * s)
1585 const int f_code= s->f_code;
1587 UINT8 * fcode_tab= s->fcode_tab;
1590 /* clip / convert to intra 16x16 type MVs */
1591 for(y=0; y<s->mb_height; y++){
1593 int xy= (y+1)* (s->mb_width+2)+1;
1594 int i= y*s->mb_width;
1595 for(x=0; x<s->mb_width; x++){
1596 if(s->mb_type[i]&MB_TYPE_INTER){
1597 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
1598 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
1599 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
1600 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
1601 s->mb_type[i] &= ~MB_TYPE_INTER;
1602 s->mb_type[i] |= MB_TYPE_INTRA;
1603 s->p_mv_table[xy][0] = 0;
1604 s->p_mv_table[xy][1] = 0;
1614 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1615 if(s->flags&CODEC_FLAG_4MV){
1616 const int wrap= 2+ s->mb_width*2;
1618 /* clip / convert to intra 8x8 type MVs */
1619 for(y=0; y<s->mb_height; y++){
1620 int xy= (y*2 + 1)*wrap + 1;
1621 int i= y*s->mb_width;
1624 for(x=0; x<s->mb_width; x++){
1625 if(s->mb_type[i]&MB_TYPE_INTER4V){
1627 for(block=0; block<4; block++){
1628 int off= (block& 1) + (block>>1)*wrap;
1629 int mx= s->motion_val[ xy + off ][0];
1630 int my= s->motion_val[ xy + off ][1];
1632 if( fcode_tab[mx + MAX_MV] > f_code
1633 || fcode_tab[mx + MAX_MV] == 0
1634 || fcode_tab[my + MAX_MV] > f_code
1635 || fcode_tab[my + MAX_MV] == 0 ){
1636 s->mb_type[i] &= ~MB_TYPE_INTER4V;
1637 s->mb_type[i] |= MB_TYPE_INTRA;
1648 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1651 UINT8 * fcode_tab= s->fcode_tab;
1653 // RAL: 8 in MPEG-1, 16 in MPEG-4
1654 int range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
1656 /* clip / convert to intra 16x16 type MVs */
1657 for(y=0; y<s->mb_height; y++){
1659 int xy= (y+1)* (s->mb_width+2)+1;
1660 int i= y*s->mb_width;
1661 for(x=0; x<s->mb_width; x++)
1663 if (s->mb_type[i] & type) // RAL: "type" test added...
1665 if (fcode_tab[mv_table[xy][0] + MAX_MV] > f_code || fcode_tab[mv_table[xy][0] + MAX_MV] == 0)
1667 if(mv_table[xy][0]>0)
1668 mv_table[xy][0]= range-1;
1670 mv_table[xy][0]= -range;
1672 if (fcode_tab[mv_table[xy][1] + MAX_MV] > f_code || fcode_tab[mv_table[xy][1] + MAX_MV] == 0)
1674 if(mv_table[xy][1]>0)
1675 mv_table[xy][1]= range-1;
1677 mv_table[xy][1]= -range;