3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2 of the License, or (at your option) any later version.
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
34 #include "mpegvideo.h"
39 #define SQ(a) ((a)*(a))
43 #define P_TOPRIGHT P[3]
47 static inline int sad_hpel_motion_search(MpegEncContext * s,
48 int *mx_ptr, int *my_ptr, int dmin,
49 int src_index, int ref_index,
52 static inline int update_map_generation(MotionEstContext *c)
54 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
55 if(c->map_generation==0){
56 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
57 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59 return c->map_generation;
62 /* shape adaptive search stuff */
63 typedef struct Minima{
69 static int minima_cmp(const void *a, const void *b){
70 const Minima *da = (const Minima *) a;
71 const Minima *db = (const Minima *) b;
73 return da->height - db->height;
76 #define FLAG_QPEL 1 //must be 1
80 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
81 const int offset[3]= {
83 ((y*c->uvstride + x)>>1),
84 ((y*c->uvstride + x)>>1),
88 c->src[0][i]= src [i] + offset[i];
89 c->ref[0][i]= ref [i] + offset[i];
93 c->ref[ref_index][i]= ref2[i] + offset[i];
98 static int get_flags(MotionEstContext *c, int direct, int chroma){
99 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
100 + (direct ? FLAG_DIRECT : 0)
101 + (chroma ? FLAG_CHROMA : 0);
104 static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
105 const int size, const int h, int ref_index, int src_index,
106 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
107 MotionEstContext * const c= &s->me;
108 const int stride= c->stride;
109 const int uvstride= c->uvstride;
110 const int qpel= flags&FLAG_QPEL;
111 const int chroma= flags&FLAG_CHROMA;
112 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
113 const int hx= subx + (x<<(1+qpel));
114 const int hy= suby + (y<<(1+qpel));
115 uint8_t * const * const ref= c->ref[ref_index];
116 uint8_t * const * const src= c->src[src_index];
118 //FIXME check chroma 4mv, (no crashes ...)
119 if(flags&FLAG_DIRECT){
120 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
121 const int time_pp= s->pp_time;
122 const int time_pb= s->pb_time;
123 const int mask= 2*qpel+1;
124 if(s->mv_type==MV_TYPE_8X8){
127 int fx = c->direct_basis_mv[i][0] + hx;
128 int fy = c->direct_basis_mv[i][1] + hy;
129 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
130 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
131 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
132 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
134 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
136 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
137 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
139 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
140 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
144 int fx = c->direct_basis_mv[0][0] + hx;
145 int fy = c->direct_basis_mv[0][1] + hy;
146 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
147 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
148 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
149 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
152 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
153 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
154 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
155 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
156 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
157 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
158 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
159 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
161 assert((fx>>1) + 16*s->mb_x >= -16);
162 assert((fy>>1) + 16*s->mb_y >= -16);
163 assert((fx>>1) + 16*s->mb_x <= s->width);
164 assert((fy>>1) + 16*s->mb_y <= s->height);
165 assert((bx>>1) + 16*s->mb_x >= -16);
166 assert((by>>1) + 16*s->mb_y >= -16);
167 assert((bx>>1) + 16*s->mb_x <= s->width);
168 assert((by>>1) + 16*s->mb_y <= s->height);
170 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
171 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
174 d = cmp_func(s, c->temp, src[0], stride, 16);
178 int uvdxy; /* no, it might not be used uninitialized */
181 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
187 uvdxy= (cx&1) + 2*(cy&1);
188 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
191 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
193 uvdxy= dxy | (x&1) | (2*(y&1));
195 d = cmp_func(s, c->temp, src[0], stride, h);
197 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
199 uvdxy= (x&1) + 2*(y&1);
202 uint8_t * const uvtemp= c->temp + 16*stride;
203 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
204 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
205 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
206 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
211 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
215 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
220 #include "motion_est_template.c"
222 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
226 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
229 void ff_init_me(MpegEncContext *s){
230 MotionEstContext * const c= &s->me;
233 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
234 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
235 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
236 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
238 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
239 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
240 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
242 /*FIXME s->no_rounding b_type*/
243 if(s->flags&CODEC_FLAG_QPEL){
244 c->sub_motion_search= qpel_motion_search;
245 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
246 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
247 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
249 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
250 c->sub_motion_search= hpel_motion_search;
251 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
252 && c->avctx-> me_cmp == FF_CMP_SAD
253 && c->avctx-> mb_cmp == FF_CMP_SAD)
254 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
256 c->sub_motion_search= hpel_motion_search;
258 c->hpel_avg= s->dsp.avg_pixels_tab;
259 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
260 else c->hpel_put= s->dsp.put_pixels_tab;
263 c->stride = s->linesize;
264 c->uvstride= s->uvlinesize;
266 c->stride = 16*s->mb_width + 32;
267 c->uvstride= 8*s->mb_width + 16;
270 // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
271 if(s->codec_id != CODEC_ID_SNOW){
272 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
273 s->dsp.me_cmp[2]= zero_cmp;
275 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
276 s->dsp.me_sub_cmp[2]= zero_cmp;
278 c->hpel_put[2][0]= c->hpel_put[2][1]=
279 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
282 if(s->codec_id == CODEC_ID_H261){
283 c->sub_motion_search= no_sub_motion_search;
286 c->temp= c->scratchpad;
290 static int pix_dev(uint8_t * pix, int line_size, int mean)
295 for (i = 0; i < 16; i++) {
296 for (j = 0; j < 16; j += 8) {
297 s += ABS(pix[0]-mean);
298 s += ABS(pix[1]-mean);
299 s += ABS(pix[2]-mean);
300 s += ABS(pix[3]-mean);
301 s += ABS(pix[4]-mean);
302 s += ABS(pix[5]-mean);
303 s += ABS(pix[6]-mean);
304 s += ABS(pix[7]-mean);
307 pix += line_size - 16;
313 static inline void no_motion_search(MpegEncContext * s,
314 int *mx_ptr, int *my_ptr)
316 *mx_ptr = 16 * s->mb_x;
317 *my_ptr = 16 * s->mb_y;
320 #if 0 /* the use of these functions is inside #if 0 */
321 static int full_motion_search(MpegEncContext * s,
322 int *mx_ptr, int *my_ptr, int range,
323 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
325 int x1, y1, x2, y2, xx, yy, x, y;
331 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
343 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
347 for (y = y1; y <= y2; y++) {
348 for (x = x1; x <= x2; x++) {
349 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
353 (abs(x - xx) + abs(y - yy)) <
354 (abs(mx - xx) + abs(my - yy)))) {
366 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
367 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
368 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
375 static int log_motion_search(MpegEncContext * s,
376 int *mx_ptr, int *my_ptr, int range,
377 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
379 int x1, y1, x2, y2, xx, yy, x, y;
406 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
412 for (y = y1; y <= y2; y += range) {
413 for (x = x1; x <= x2; x += range) {
414 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
415 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
441 } while (range >= 1);
444 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
451 static int phods_motion_search(MpegEncContext * s,
452 int *mx_ptr, int *my_ptr, int range,
453 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
455 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
456 int mx, my, dminx, dminy;
482 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
493 for (x = x1; x <= x2; x += range) {
494 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
495 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
502 for (y = y1; y <= y2; y += range) {
503 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
504 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
530 } while (range >= 1);
533 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
536 /* half pixel search */
543 #define Z_THRESHOLD 256
545 #define CHECK_SAD_HALF_MV(suffix, x, y) \
547 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
548 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
549 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
552 static inline int sad_hpel_motion_search(MpegEncContext * s,
553 int *mx_ptr, int *my_ptr, int dmin,
554 int src_index, int ref_index,
557 MotionEstContext * const c= &s->me;
558 const int penalty_factor= c->sub_penalty_factor;
561 int stride= c->stride;
562 const int flags= c->sub_flags;
575 pix = c->src[src_index][0];
579 ptr = c->ref[ref_index][0] + (my * stride) + mx;
583 if (mx > xmin && mx < xmax &&
584 my > ymin && my < ymax) {
587 const int index= (my<<ME_MAP_SHIFT) + mx;
588 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
589 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
590 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
591 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
601 CHECK_SAD_HALF_MV(y2 , 0, -1)
603 CHECK_SAD_HALF_MV(xy2, -1, -1)
605 CHECK_SAD_HALF_MV(xy2, +1, -1)
609 CHECK_SAD_HALF_MV(xy2, -1, +1)
611 CHECK_SAD_HALF_MV(x2 , -1, 0)
613 CHECK_SAD_HALF_MV(xy2, +1, -1)
615 CHECK_SAD_HALF_MV(xy2, -1, -1)
619 CHECK_SAD_HALF_MV(xy2, +1, +1)
621 CHECK_SAD_HALF_MV(x2 , +1, 0)
626 CHECK_SAD_HALF_MV(xy2, -1, -1)
630 CHECK_SAD_HALF_MV(xy2, +1, +1)
632 CHECK_SAD_HALF_MV(x2 , -1, 0)
633 CHECK_SAD_HALF_MV(xy2, -1, +1)
636 CHECK_SAD_HALF_MV(xy2, +1, -1)
640 CHECK_SAD_HALF_MV(xy2, -1, +1)
642 CHECK_SAD_HALF_MV(x2 , +1, 0)
643 CHECK_SAD_HALF_MV(xy2, +1, +1)
645 CHECK_SAD_HALF_MV(y2 , 0, +1)
660 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
662 const int xy= s->mb_x + s->mb_y*s->mb_stride;
664 s->p_mv_table[xy][0] = mx;
665 s->p_mv_table[xy][1] = my;
667 /* has already been set to the 4 MV if 4MV is done */
669 int mot_xy= s->block_index[0];
671 s->current_picture.motion_val[0][mot_xy ][0]= mx;
672 s->current_picture.motion_val[0][mot_xy ][1]= my;
673 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
674 s->current_picture.motion_val[0][mot_xy+1][1]= my;
676 mot_xy += s->b8_stride;
677 s->current_picture.motion_val[0][mot_xy ][0]= mx;
678 s->current_picture.motion_val[0][mot_xy ][1]= my;
679 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
680 s->current_picture.motion_val[0][mot_xy+1][1]= my;
685 * get fullpel ME search limits.
687 static inline void get_limits(MpegEncContext *s, int x, int y)
689 MotionEstContext * const c= &s->me;
691 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
694 if (s->unrestricted_mv) {
697 c->xmax = - x + s->mb_width *16;
698 c->ymax = - y + s->mb_height*16;
699 } else if (s->out_format == FMT_H261){
700 // Search range of H261 is different from other codec standards
701 c->xmin = (x > 15) ? - 15 : 0;
702 c->ymin = (y > 15) ? - 15 : 0;
703 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
704 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
708 c->xmax = - x + s->mb_width *16 - 16;
709 c->ymax = - y + s->mb_height*16 - 16;
713 static inline void init_mv4_ref(MotionEstContext *c){
714 const int stride= c->stride;
716 c->ref[1][0] = c->ref[0][0] + 8;
717 c->ref[2][0] = c->ref[0][0] + 8*stride;
718 c->ref[3][0] = c->ref[2][0] + 8;
719 c->src[1][0] = c->src[0][0] + 8;
720 c->src[2][0] = c->src[0][0] + 8*stride;
721 c->src[3][0] = c->src[2][0] + 8;
724 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
726 MotionEstContext * const c= &s->me;
731 int dmin_sum=0, mx4_sum=0, my4_sum=0;
733 const int stride= c->stride;
734 uint8_t *mv_penalty= c->current_mv_penalty;
738 for(block=0; block<4; block++){
740 int pred_x4, pred_y4;
742 static const int off[4]= {2, 1, 1, -1};
743 const int mot_stride = s->b8_stride;
744 const int mot_xy = s->block_index[block];
746 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
747 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
749 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
751 /* special case for first line */
752 if (s->first_slice_line && block<2) {
753 c->pred_x= pred_x4= P_LEFT[0];
754 c->pred_y= pred_y4= P_LEFT[1];
756 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
757 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
758 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
759 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
760 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
761 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
762 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
763 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
765 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
766 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
768 c->pred_x= pred_x4 = P_MEDIAN[0];
769 c->pred_y= pred_y4 = P_MEDIAN[1];
774 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
776 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
778 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
780 const int offset= ((block&1) + (block>>1)*stride)*8;
781 uint8_t *dest_y = c->scratchpad + offset;
782 if(s->quarter_sample){
783 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
784 dxy = ((my4 & 3) << 2) | (mx4 & 3);
787 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
789 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
791 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
792 dxy = ((my4 & 1) << 1) | (mx4 & 1);
795 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
797 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
799 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
803 if(s->quarter_sample){
811 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
812 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
814 if(mx4 != mx || my4 != my) same=0;
820 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
821 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
824 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
829 mx= ff_h263_round_chroma(mx4_sum);
830 my= ff_h263_round_chroma(my4_sum);
831 dxy = ((my & 1) << 1) | (mx & 1);
833 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
836 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
837 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
839 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
840 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
843 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
844 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
850 switch(c->avctx->mb_cmp&0xFF){
852 return dmin_sum+ 32*s->qscale*s->qscale;*/
856 return dmin_sum+ 11*c->mb_penalty_factor;
860 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
861 MotionEstContext * const c= &s->me;
863 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
864 c->src[1][0] = c->src[0][0] + s->linesize;
865 if(c->flags & FLAG_CHROMA){
866 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
867 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
868 c->src[1][1] = c->src[0][1] + s->uvlinesize;
869 c->src[1][2] = c->src[0][2] + s->uvlinesize;
873 static int interlaced_search(MpegEncContext *s, int ref_index,
874 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
876 MotionEstContext * const c= &s->me;
881 uint8_t * const mv_penalty= c->current_mv_penalty;
883 const int stride= 2*s->linesize;
885 const int mot_stride= s->mb_stride;
886 const int xy= s->mb_x + s->mb_y*mot_stride;
892 init_interlaced_ref(s, ref_index);
894 for(block=0; block<2; block++){
896 int best_dmin= INT_MAX;
899 for(field_select=0; field_select<2; field_select++){
900 int dmin, mx_i, my_i;
901 int16_t (*mv_table)[2]= mv_tables[block][field_select];
903 if(user_field_select){
904 if(field_select_tables[block][xy] != field_select)
908 P_LEFT[0] = mv_table[xy - 1][0];
909 P_LEFT[1] = mv_table[xy - 1][1];
910 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
912 c->pred_x= P_LEFT[0];
913 c->pred_y= P_LEFT[1];
915 if(!s->first_slice_line){
916 P_TOP[0] = mv_table[xy - mot_stride][0];
917 P_TOP[1] = mv_table[xy - mot_stride][1];
918 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
919 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
920 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
921 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
922 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
923 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
925 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
926 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
928 P_MV1[0]= mx; //FIXME not correct if block != field_select
931 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
933 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
935 mv_table[xy][0]= mx_i;
936 mv_table[xy][1]= my_i;
938 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
942 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
943 dxy = ((my_i & 1) << 1) | (mx_i & 1);
946 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
948 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
950 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
951 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
953 dmin+= c->mb_penalty_factor; //field_select bits
955 dmin += field_select != block; //slightly prefer same field
957 if(dmin < best_dmin){
959 best_field= field_select;
963 int16_t (*mv_table)[2]= mv_tables[block][best_field];
965 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
966 if(mv_table[xy][1]&1) same=0;
967 if(mv_table[xy][1]*2 != my) same=0;
968 if(best_field != block) same=0;
971 field_select_tables[block][xy]= best_field;
972 dmin_sum += best_dmin;
983 switch(c->avctx->mb_cmp&0xFF){
985 return dmin_sum+ 32*s->qscale*s->qscale;*/
989 return dmin_sum+ 11*c->mb_penalty_factor;
993 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
994 int ymax= s->me.ymax>>interlaced;
995 int ymin= s->me.ymin>>interlaced;
997 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
998 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
999 if(mv[1] < ymin) mv[1] = ymin;
1000 if(mv[1] > ymax) mv[1] = ymax;
1003 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1004 MotionEstContext * const c= &s->me;
1005 Picture *p= s->current_picture_ptr;
1006 int mb_xy= mb_x + mb_y*s->mb_stride;
1007 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1008 int mb_type= s->current_picture.mb_type[mb_xy];
1009 int flags= c->flags;
1010 int shift= (flags&FLAG_QPEL) + 1;
1011 int mask= (1<<shift)-1;
1014 me_cmp_func cmpf= s->dsp.sse[0];
1015 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1017 if(p_type && USES_LIST(mb_type, 1)){
1018 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1021 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1024 int xy= s->block_index[i];
1025 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1026 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1029 if(IS_INTERLACED(mb_type)){
1030 int xy2= xy + s->b8_stride;
1031 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1035 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1036 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1040 if(USES_LIST(mb_type, 0)){
1041 int field_select0= p->ref_index[0][xy ];
1042 int field_select1= p->ref_index[0][xy2];
1043 assert(field_select0==0 ||field_select0==1);
1044 assert(field_select1==0 ||field_select1==1);
1045 init_interlaced_ref(s, 0);
1048 s->p_field_select_table[0][mb_xy]= field_select0;
1049 s->p_field_select_table[1][mb_xy]= field_select1;
1050 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1051 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1052 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1054 s->b_field_select_table[0][0][mb_xy]= field_select0;
1055 s->b_field_select_table[0][1][mb_xy]= field_select1;
1056 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1057 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1058 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1061 x= p->motion_val[0][xy ][0];
1062 y= p->motion_val[0][xy ][1];
1063 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1064 x= p->motion_val[0][xy2][0];
1065 y= p->motion_val[0][xy2][1];
1066 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1068 if(USES_LIST(mb_type, 1)){
1069 int field_select0= p->ref_index[1][xy ];
1070 int field_select1= p->ref_index[1][xy2];
1071 assert(field_select0==0 ||field_select0==1);
1072 assert(field_select1==0 ||field_select1==1);
1073 init_interlaced_ref(s, 2);
1075 s->b_field_select_table[1][0][mb_xy]= field_select0;
1076 s->b_field_select_table[1][1][mb_xy]= field_select1;
1077 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1078 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1079 if(USES_LIST(mb_type, 0)){
1080 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1082 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1085 x= p->motion_val[1][xy ][0];
1086 y= p->motion_val[1][xy ][1];
1087 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1088 x= p->motion_val[1][xy2][0];
1089 y= p->motion_val[1][xy2][1];
1090 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1091 //FIXME bidir scores
1095 }else if(IS_8X8(mb_type)){
1096 if(!(s->flags & CODEC_FLAG_4MV)){
1097 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1100 cmpf= s->dsp.sse[1];
1101 chroma_cmpf= s->dsp.sse[1];
1104 xy= s->block_index[i];
1105 x= p->motion_val[0][xy][0];
1106 y= p->motion_val[0][xy][1];
1107 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1109 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1111 if(USES_LIST(mb_type, 0)){
1113 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1114 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1115 }else if(USES_LIST(mb_type, 1)){
1116 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1117 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1118 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1120 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1121 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1123 x= p->motion_val[0][xy][0];
1124 y= p->motion_val[0][xy][1];
1125 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1126 }else if(USES_LIST(mb_type, 1)){
1127 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1128 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1130 x= p->motion_val[1][xy][0];
1131 y= p->motion_val[1][xy][1];
1132 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1134 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1139 void ff_estimate_p_frame_motion(MpegEncContext * s,
1142 MotionEstContext * const c= &s->me;
1143 uint8_t *pix, *ppix;
1144 int sum, varc, vard, mx, my, dmin;
1146 const int shift= 1+s->quarter_sample;
1148 Picture * const pic= &s->current_picture;
1150 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1152 assert(s->quarter_sample==0 || s->quarter_sample==1);
1153 assert(s->linesize == c->stride);
1154 assert(s->uvlinesize == c->uvstride);
1156 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1157 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1158 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1159 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1161 get_limits(s, 16*mb_x, 16*mb_y);
1164 /* intra / predictive decision */
1166 sum = s->dsp.pix_sum(pix, s->linesize);
1167 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1169 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1170 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1171 c->mb_var_sum_temp += varc;
1173 if(c->avctx->me_threshold){
1174 vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
1176 if(vard<c->avctx->me_threshold){
1177 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1178 c->mc_mb_var_sum_temp += vard;
1179 if (vard <= 64 || vard < varc) { //FIXME
1180 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1182 c->scene_change_score+= s->qscale;
1186 if(vard<c->avctx->mb_threshold)
1187 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1190 switch(s->me_method) {
1193 no_motion_search(s, &mx, &my);
1200 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1205 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1210 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1218 const int mot_stride = s->b8_stride;
1219 const int mot_xy = s->block_index[0];
1221 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1222 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1224 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1226 if(!s->first_slice_line) {
1227 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1228 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1229 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1230 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1231 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1232 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1233 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1235 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1236 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1238 if(s->out_format == FMT_H263){
1239 c->pred_x = P_MEDIAN[0];
1240 c->pred_y = P_MEDIAN[1];
1241 }else { /* mpeg1 at least */
1242 c->pred_x= P_LEFT[0];
1243 c->pred_y= P_LEFT[1];
1246 c->pred_x= P_LEFT[0];
1247 c->pred_y= P_LEFT[1];
1251 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1256 /* At this point (mx,my) are full-pell and the relative displacement */
1257 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1259 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
1261 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1262 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1263 c->mc_mb_var_sum_temp += vard;
1266 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1267 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1270 if (vard <= 64 || vard < varc)
1271 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1273 c->scene_change_score+= s->qscale;
1275 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1276 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1277 set_p_mv_tables(s, mx, my, 1);
1282 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1283 h263_mv4_search(s, mx, my, shift);
1285 set_p_mv_tables(s, mx, my, 0);
1287 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1288 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1290 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1291 if (vard <= 64 || vard < varc)
1292 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1294 c->scene_change_score+= s->qscale;
1296 if (vard*2 + 200 > varc)
1297 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1298 if (varc*2 + 200 > vard){
1299 mb_type|= CANDIDATE_MB_TYPE_INTER;
1300 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1301 if(s->flags&CODEC_FLAG_MV0)
1303 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1308 if((s->flags&CODEC_FLAG_4MV)
1309 && !c->skip && varc>50 && vard>10){
1310 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1311 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1313 set_p_mv_tables(s, mx, my, 0);
1315 set_p_mv_tables(s, mx, my, 1);
1316 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1317 && !c->skip){ //FIXME varc/d checks
1318 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1319 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1323 mb_type= CANDIDATE_MB_TYPE_INTER;
1325 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1326 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1327 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1329 if((s->flags&CODEC_FLAG_4MV)
1330 && !c->skip && varc>50 && vard>10){
1331 int dmin4= h263_mv4_search(s, mx, my, shift);
1333 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1337 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1338 && !c->skip){ //FIXME varc/d checks
1339 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1341 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1346 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1347 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1349 /* get intra luma score */
1350 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1351 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
1353 int mean= (sum+128)>>8;
1356 for(i=0; i<16; i++){
1357 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1358 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1359 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1360 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1363 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1366 /* get chroma score */
1367 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1372 if(s->out_format == FMT_H263){
1373 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1375 mean= (s->last_dc[i] + 4)>>3;
1377 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1381 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1382 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1385 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1389 intra_score += c->mb_penalty_factor*16;
1391 if(intra_score < dmin){
1392 mb_type= CANDIDATE_MB_TYPE_INTRA;
1393 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1395 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1397 if (vard <= 64 || vard < varc) { //FIXME
1398 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1400 c->scene_change_score+= s->qscale;
1404 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1407 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1410 MotionEstContext * const c= &s->me;
1413 const int shift= 1+s->quarter_sample;
1414 const int xy= mb_x + mb_y*s->mb_stride;
1415 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1417 assert(s->quarter_sample==0 || s->quarter_sample==1);
1419 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1420 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1422 get_limits(s, 16*mb_x, 16*mb_y);
1425 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1426 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1428 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1430 /* special case for first line */
1431 if (s->first_slice_line) {
1432 c->pred_x= P_LEFT[0];
1433 c->pred_y= P_LEFT[1];
1434 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1435 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1437 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1438 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1439 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1440 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1441 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1442 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1443 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1445 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1446 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1448 c->pred_x = P_MEDIAN[0];
1449 c->pred_y = P_MEDIAN[1];
1452 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1454 s->p_mv_table[xy][0] = mx<<shift;
1455 s->p_mv_table[xy][1] = my<<shift;
1460 static int ff_estimate_motion_b(MpegEncContext * s,
1461 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1463 MotionEstContext * const c= &s->me;
1466 const int shift= 1+s->quarter_sample;
1467 const int mot_stride = s->mb_stride;
1468 const int mot_xy = mb_y*mot_stride + mb_x;
1469 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1472 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1473 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1474 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1475 c->current_mv_penalty= mv_penalty;
1477 get_limits(s, 16*mb_x, 16*mb_y);
1479 switch(s->me_method) {
1482 no_motion_search(s, &mx, &my);
1489 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1494 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1499 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1507 P_LEFT[0] = mv_table[mot_xy - 1][0];
1508 P_LEFT[1] = mv_table[mot_xy - 1][1];
1510 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1512 /* special case for first line */
1513 if (!s->first_slice_line) {
1514 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1515 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1516 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1517 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1518 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1519 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1520 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1522 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1523 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1525 c->pred_x= P_LEFT[0];
1526 c->pred_y= P_LEFT[1];
1529 if(mv_table == s->b_forw_mv_table){
1530 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1532 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1535 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1540 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1542 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1543 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1545 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1546 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1547 mv_table[mot_xy][0]= mx;
1548 mv_table[mot_xy][1]= my;
1553 static inline int check_bidir_mv(MpegEncContext * s,
1554 int motion_fx, int motion_fy,
1555 int motion_bx, int motion_by,
1556 int pred_fx, int pred_fy,
1557 int pred_bx, int pred_by,
1561 //FIXME better f_code prediction (max mv & distance)
1563 MotionEstContext * const c= &s->me;
1564 uint8_t * const mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1565 int stride= c->stride;
1566 uint8_t *dest_y = c->scratchpad;
1571 uint8_t **src_data= c->src[0];
1572 uint8_t **ref_data= c->ref[0];
1573 uint8_t **ref2_data= c->ref[2];
1575 if(s->quarter_sample){
1576 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1577 src_x = motion_fx >> 2;
1578 src_y = motion_fy >> 2;
1580 ptr = ref_data[0] + (src_y * stride) + src_x;
1581 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1583 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1584 src_x = motion_bx >> 2;
1585 src_y = motion_by >> 2;
1587 ptr = ref2_data[0] + (src_y * stride) + src_x;
1588 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1590 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1591 src_x = motion_fx >> 1;
1592 src_y = motion_fy >> 1;
1594 ptr = ref_data[0] + (src_y * stride) + src_x;
1595 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1597 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1598 src_x = motion_bx >> 1;
1599 src_y = motion_by >> 1;
1601 ptr = ref2_data[0] + (src_y * stride) + src_x;
1602 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1605 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*c->mb_penalty_factor
1606 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*c->mb_penalty_factor
1607 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1609 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1616 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1617 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1619 const int mot_stride = s->mb_stride;
1620 const int xy = mb_y *mot_stride + mb_x;
1622 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1623 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1624 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1625 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1626 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1627 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1628 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1629 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1631 //FIXME do refinement and add flag
1633 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1634 motion_bx, motion_by,
1642 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1644 MotionEstContext * const c= &s->me;
1646 const int mot_stride = s->mb_stride;
1647 const int mot_xy = mb_y*mot_stride + mb_x;
1648 const int shift= 1+s->quarter_sample;
1650 const int time_pp= s->pp_time;
1651 const int time_pb= s->pb_time;
1652 int mx, my, xmin, xmax, ymin, ymax;
1653 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1655 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1656 ymin= xmin=(-32)>>shift;
1657 ymax= xmax= 31>>shift;
1659 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1660 s->mv_type= MV_TYPE_8X8;
1662 s->mv_type= MV_TYPE_16X16;
1666 int index= s->block_index[i];
1669 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1670 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1671 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1672 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1673 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1674 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1676 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1677 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1678 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1680 xmax= FFMIN(xmax, s->width - max);
1681 xmin= FFMAX(xmin, - 16 - min);
1683 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1684 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1685 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1687 ymax= FFMIN(ymax, s->height - max);
1688 ymin= FFMAX(ymin, - 16 - min);
1690 if(s->mv_type == MV_TYPE_16X16) break;
1693 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1695 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1696 s->b_direct_mv_table[mot_xy][0]= 0;
1697 s->b_direct_mv_table[mot_xy][1]= 0;
1699 return 256*256*256*64;
1706 c->flags |= FLAG_DIRECT;
1707 c->sub_flags |= FLAG_DIRECT;
1711 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1712 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1714 /* special case for first line */
1715 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
1716 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1717 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1718 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1719 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1721 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1722 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1725 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1726 if(c->sub_flags&FLAG_QPEL)
1727 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1729 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1731 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1732 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1734 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1736 s->b_direct_mv_table[mot_xy][0]= mx;
1737 s->b_direct_mv_table[mot_xy][1]= my;
1738 c->flags &= ~FLAG_DIRECT;
1739 c->sub_flags &= ~FLAG_DIRECT;
1744 void ff_estimate_b_frame_motion(MpegEncContext * s,
1747 MotionEstContext * const c= &s->me;
1748 const int penalty_factor= c->mb_penalty_factor;
1749 int fmin, bmin, dmin, fbmin, bimin, fimin;
1751 const int xy = mb_y*s->mb_stride + mb_x;
1752 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1754 get_limits(s, 16*mb_x, 16*mb_y);
1757 if(c->avctx->me_threshold){
1758 int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
1760 if(vard<c->avctx->me_threshold){
1761 // pix = c->src[0][0];
1762 // sum = s->dsp.pix_sum(pix, s->linesize);
1763 // varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1765 // pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1766 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1767 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1768 c->mb_var_sum_temp += varc;*/
1769 c->mc_mb_var_sum_temp += vard;
1770 /* if (vard <= 64 || vard < varc) {
1771 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1773 c->scene_change_score+= s->qscale;
1777 if(vard<c->avctx->mb_threshold){
1778 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1779 if(type == CANDIDATE_MB_TYPE_DIRECT){
1780 direct_search(s, mb_x, mb_y);
1782 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1784 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1786 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1788 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1790 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1792 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1793 interlaced_search(s, 0,
1794 s->b_field_mv_table[0], s->b_field_select_table[0],
1795 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1797 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1799 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1800 interlaced_search(s, 2,
1801 s->b_field_mv_table[1], s->b_field_select_table[1],
1802 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1808 if (s->codec_id == CODEC_ID_MPEG4)
1809 dmin= direct_search(s, mb_x, mb_y);
1812 //FIXME penalty stuff for non mpeg4
1814 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1817 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1818 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1821 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1822 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1824 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1825 //FIXME mb type penalty
1827 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1828 fimin= interlaced_search(s, 0,
1829 s->b_field_mv_table[0], s->b_field_select_table[0],
1830 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1831 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1832 bimin= interlaced_search(s, 2,
1833 s->b_field_mv_table[1], s->b_field_select_table[1],
1834 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1836 fimin= bimin= INT_MAX;
1840 type = CANDIDATE_MB_TYPE_FORWARD;
1844 type = CANDIDATE_MB_TYPE_DIRECT;
1848 type= CANDIDATE_MB_TYPE_BACKWARD;
1852 type= CANDIDATE_MB_TYPE_BIDIR;
1856 type= CANDIDATE_MB_TYPE_FORWARD_I;
1860 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1863 score= ((unsigned)(score*score + 128*256))>>16;
1864 c->mc_mb_var_sum_temp += score;
1865 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1868 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1869 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1871 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1873 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1874 if(fimin < INT_MAX && bimin < INT_MAX){
1875 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1877 //FIXME something smarter
1878 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1880 if(s->out_format == FMT_MPEG1)
1881 type |= CANDIDATE_MB_TYPE_INTRA;
1885 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1888 /* find best f_code for ME which do unlimited searches */
1889 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1891 if(s->me_method>=ME_EPZS){
1893 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1894 uint8_t * fcode_tab= s->fcode_tab;
1896 int best_score=-10000000;
1898 if(s->msmpeg4_version)
1899 range= FFMIN(range, 16);
1900 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1901 range= FFMIN(range, 256);
1903 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1905 for(y=0; y<s->mb_height; y++){
1907 int xy= y*s->mb_stride;
1908 for(x=0; x<s->mb_width; x++){
1909 if(s->mb_type[xy] & type){
1910 int mx= mv_table[xy][0];
1911 int my= mv_table[xy][1];
1912 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1913 fcode_tab[my + MAX_MV]);
1916 if(mx >= range || mx < -range ||
1917 my >= range || my < -range)
1920 for(j=0; j<fcode && j<8; j++){
1921 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1930 if(score[i] > best_score){
1931 best_score= score[i];
1934 // printf("%d %d\n", i, score[i]);
1937 // printf("fcode: %d type: %d\n", i, s->pict_type);
1939 /* for(i=0; i<=MAX_FCODE; i++){
1940 printf("%d ", mv_num[i]);
1948 void ff_fix_long_p_mvs(MpegEncContext * s)
1950 MotionEstContext * const c= &s->me;
1951 const int f_code= s->f_code;
1953 assert(s->pict_type==P_TYPE);
1955 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1957 assert(range <= 16 || !s->msmpeg4_version);
1958 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1960 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1962 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1963 if(s->flags&CODEC_FLAG_4MV){
1964 const int wrap= s->b8_stride;
1966 /* clip / convert to intra 8x8 type MVs */
1967 for(y=0; y<s->mb_height; y++){
1969 int i= y*s->mb_stride;
1972 for(x=0; x<s->mb_width; x++){
1973 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1975 for(block=0; block<4; block++){
1976 int off= (block& 1) + (block>>1)*wrap;
1977 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1978 int my= s->current_picture.motion_val[0][ xy + off ][1];
1980 if( mx >=range || mx <-range
1981 || my >=range || my <-range){
1982 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1983 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1984 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1997 * @param truncate 1 for truncation, 0 for using intra
1999 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2000 int16_t (*mv_table)[2], int f_code, int type, int truncate)
2002 MotionEstContext * const c= &s->me;
2003 int y, h_range, v_range;
2005 // RAL: 8 in MPEG-1, 16 in MPEG-4
2006 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2008 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2011 v_range= field_select_table ? range>>1 : range;
2013 /* clip / convert to intra 16x16 type MVs */
2014 for(y=0; y<s->mb_height; y++){
2016 int xy= y*s->mb_stride;
2017 for(x=0; x<s->mb_width; x++){
2018 if (s->mb_type[xy] & type){ // RAL: "type" test added...
2019 if(field_select_table==NULL || field_select_table[xy] == field_select){
2020 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2021 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2024 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2025 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2026 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2027 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2029 s->mb_type[xy] &= ~type;
2030 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;