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);
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((c->avctx->me_cmp&FF_CMP_CHROMA) && !s->dsp.me_cmp[2]){
272 s->dsp.me_cmp[2]= zero_cmp;
274 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
275 s->dsp.me_sub_cmp[2]= zero_cmp;
277 c->hpel_put[2][0]= c->hpel_put[2][1]=
278 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
280 c->temp= c->scratchpad;
284 static int pix_dev(uint8_t * pix, int line_size, int mean)
289 for (i = 0; i < 16; i++) {
290 for (j = 0; j < 16; j += 8) {
291 s += ABS(pix[0]-mean);
292 s += ABS(pix[1]-mean);
293 s += ABS(pix[2]-mean);
294 s += ABS(pix[3]-mean);
295 s += ABS(pix[4]-mean);
296 s += ABS(pix[5]-mean);
297 s += ABS(pix[6]-mean);
298 s += ABS(pix[7]-mean);
301 pix += line_size - 16;
307 static inline void no_motion_search(MpegEncContext * s,
308 int *mx_ptr, int *my_ptr)
310 *mx_ptr = 16 * s->mb_x;
311 *my_ptr = 16 * s->mb_y;
314 static int full_motion_search(MpegEncContext * s,
315 int *mx_ptr, int *my_ptr, int range,
316 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
318 int x1, y1, x2, y2, xx, yy, x, y;
324 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
336 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
340 for (y = y1; y <= y2; y++) {
341 for (x = x1; x <= x2; x++) {
342 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
346 (abs(x - xx) + abs(y - yy)) <
347 (abs(mx - xx) + abs(my - yy)))) {
359 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
360 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
361 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
368 static int log_motion_search(MpegEncContext * s,
369 int *mx_ptr, int *my_ptr, int range,
370 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
372 int x1, y1, x2, y2, xx, yy, x, y;
399 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
405 for (y = y1; y <= y2; y += range) {
406 for (x = x1; x <= x2; x += range) {
407 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
408 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
434 } while (range >= 1);
437 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
444 static int phods_motion_search(MpegEncContext * s,
445 int *mx_ptr, int *my_ptr, int range,
446 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
448 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
449 int mx, my, dminx, dminy;
475 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
486 for (x = x1; x <= x2; x += range) {
487 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
488 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
495 for (y = y1; y <= y2; y += range) {
496 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
497 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
523 } while (range >= 1);
526 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
529 /* half pixel search */
536 #define Z_THRESHOLD 256
538 #define CHECK_SAD_HALF_MV(suffix, x, y) \
540 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
541 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
542 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
545 static inline int sad_hpel_motion_search(MpegEncContext * s,
546 int *mx_ptr, int *my_ptr, int dmin,
547 int src_index, int ref_index,
550 MotionEstContext * const c= &s->me;
551 const int penalty_factor= c->sub_penalty_factor;
554 int stride= c->stride;
555 const int flags= c->sub_flags;
568 pix = c->src[src_index][0];
572 ptr = c->ref[ref_index][0] + (my * stride) + mx;
576 if (mx > xmin && mx < xmax &&
577 my > ymin && my < ymax) {
580 const int index= (my<<ME_MAP_SHIFT) + mx;
581 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
582 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
583 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
584 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
594 CHECK_SAD_HALF_MV(y2 , 0, -1)
596 CHECK_SAD_HALF_MV(xy2, -1, -1)
598 CHECK_SAD_HALF_MV(xy2, +1, -1)
602 CHECK_SAD_HALF_MV(xy2, -1, +1)
604 CHECK_SAD_HALF_MV(x2 , -1, 0)
606 CHECK_SAD_HALF_MV(xy2, +1, -1)
608 CHECK_SAD_HALF_MV(xy2, -1, -1)
612 CHECK_SAD_HALF_MV(xy2, +1, +1)
614 CHECK_SAD_HALF_MV(x2 , +1, 0)
619 CHECK_SAD_HALF_MV(xy2, -1, -1)
623 CHECK_SAD_HALF_MV(xy2, +1, +1)
625 CHECK_SAD_HALF_MV(x2 , -1, 0)
626 CHECK_SAD_HALF_MV(xy2, -1, +1)
629 CHECK_SAD_HALF_MV(xy2, +1, -1)
633 CHECK_SAD_HALF_MV(xy2, -1, +1)
635 CHECK_SAD_HALF_MV(x2 , +1, 0)
636 CHECK_SAD_HALF_MV(xy2, +1, +1)
638 CHECK_SAD_HALF_MV(y2 , 0, +1)
653 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
655 const int xy= s->mb_x + s->mb_y*s->mb_stride;
657 s->p_mv_table[xy][0] = mx;
658 s->p_mv_table[xy][1] = my;
660 /* has allready been set to the 4 MV if 4MV is done */
662 int mot_xy= s->block_index[0];
664 s->current_picture.motion_val[0][mot_xy ][0]= mx;
665 s->current_picture.motion_val[0][mot_xy ][1]= my;
666 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
667 s->current_picture.motion_val[0][mot_xy+1][1]= my;
669 mot_xy += s->b8_stride;
670 s->current_picture.motion_val[0][mot_xy ][0]= mx;
671 s->current_picture.motion_val[0][mot_xy ][1]= my;
672 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
673 s->current_picture.motion_val[0][mot_xy+1][1]= my;
678 * get fullpel ME search limits.
680 static inline void get_limits(MpegEncContext *s, int x, int y)
682 MotionEstContext * const c= &s->me;
684 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
687 if (s->unrestricted_mv) {
690 c->xmax = - x + s->mb_width *16;
691 c->ymax = - y + s->mb_height*16;
695 c->xmax = - x + s->mb_width *16 - 16;
696 c->ymax = - y + s->mb_height*16 - 16;
700 static inline void init_mv4_ref(MotionEstContext *c){
701 const int stride= c->stride;
703 c->ref[1][0] = c->ref[0][0] + 8;
704 c->ref[2][0] = c->ref[0][0] + 8*stride;
705 c->ref[3][0] = c->ref[2][0] + 8;
706 c->src[1][0] = c->src[0][0] + 8;
707 c->src[2][0] = c->src[0][0] + 8*stride;
708 c->src[3][0] = c->src[2][0] + 8;
711 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
713 MotionEstContext * const c= &s->me;
718 int dmin_sum=0, mx4_sum=0, my4_sum=0;
720 const int stride= c->stride;
721 const int uvstride= c->uvstride;
722 uint8_t *mv_penalty= c->current_mv_penalty;
726 for(block=0; block<4; block++){
728 int pred_x4, pred_y4;
730 static const int off[4]= {2, 1, 1, -1};
731 const int mot_stride = s->b8_stride;
732 const int mot_xy = s->block_index[block];
734 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
735 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
737 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
739 /* special case for first line */
740 if (s->first_slice_line && block<2) {
741 c->pred_x= pred_x4= P_LEFT[0];
742 c->pred_y= pred_y4= P_LEFT[1];
744 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
745 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
746 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
747 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
748 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
749 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
750 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
751 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
753 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
754 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
756 c->pred_x= pred_x4 = P_MEDIAN[0];
757 c->pred_y= pred_y4 = P_MEDIAN[1];
762 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
764 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
766 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
768 const int offset= ((block&1) + (block>>1)*stride)*8;
769 uint8_t *dest_y = c->scratchpad + offset;
770 if(s->quarter_sample){
771 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
772 dxy = ((my4 & 3) << 2) | (mx4 & 3);
775 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
777 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
779 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
780 dxy = ((my4 & 1) << 1) | (mx4 & 1);
783 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
785 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
787 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
791 if(s->quarter_sample){
799 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
800 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
802 if(mx4 != mx || my4 != my) same=0;
808 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
809 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);
812 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
817 mx= ff_h263_round_chroma(mx4_sum);
818 my= ff_h263_round_chroma(my4_sum);
819 dxy = ((my & 1) << 1) | (mx & 1);
821 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
824 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
825 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
827 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
828 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
831 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);
832 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);
838 switch(c->avctx->mb_cmp&0xFF){
840 return dmin_sum+ 32*s->qscale*s->qscale;*/
844 return dmin_sum+ 11*c->mb_penalty_factor;
848 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
849 MotionEstContext * const c= &s->me;
851 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
852 c->src[1][0] = c->src[0][0] + s->linesize;
853 if(c->flags & FLAG_CHROMA){
854 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
855 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
856 c->src[1][1] = c->src[0][1] + s->uvlinesize;
857 c->src[1][2] = c->src[0][2] + s->uvlinesize;
861 static int interlaced_search(MpegEncContext *s, int ref_index,
862 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
864 MotionEstContext * const c= &s->me;
869 uint8_t * const mv_penalty= c->current_mv_penalty;
871 const int stride= 2*s->linesize;
872 const int uvstride= 2*s->uvlinesize;
874 const int mot_stride= s->mb_stride;
875 const int xy= s->mb_x + s->mb_y*mot_stride;
881 init_interlaced_ref(s, ref_index);
883 for(block=0; block<2; block++){
885 int best_dmin= INT_MAX;
888 for(field_select=0; field_select<2; field_select++){
889 int dmin, mx_i, my_i;
890 int16_t (*mv_table)[2]= mv_tables[block][field_select];
892 if(user_field_select){
893 if(field_select_tables[block][xy] != field_select)
897 P_LEFT[0] = mv_table[xy - 1][0];
898 P_LEFT[1] = mv_table[xy - 1][1];
899 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
901 c->pred_x= P_LEFT[0];
902 c->pred_y= P_LEFT[1];
904 if(!s->first_slice_line){
905 P_TOP[0] = mv_table[xy - mot_stride][0];
906 P_TOP[1] = mv_table[xy - mot_stride][1];
907 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
908 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
909 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
910 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
911 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
912 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
914 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
915 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
917 P_MV1[0]= mx; //FIXME not correct if block != field_select
920 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
922 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
924 mv_table[xy][0]= mx_i;
925 mv_table[xy][1]= my_i;
927 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
931 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
932 dxy = ((my_i & 1) << 1) | (mx_i & 1);
935 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
937 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
939 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
940 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
942 dmin+= c->mb_penalty_factor; //field_select bits
944 dmin += field_select != block; //slightly prefer same field
946 if(dmin < best_dmin){
948 best_field= field_select;
952 int16_t (*mv_table)[2]= mv_tables[block][best_field];
954 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
955 if(mv_table[xy][1]&1) same=0;
956 if(mv_table[xy][1]*2 != my) same=0;
957 if(best_field != block) same=0;
960 field_select_tables[block][xy]= best_field;
961 dmin_sum += best_dmin;
972 switch(c->avctx->mb_cmp&0xFF){
974 return dmin_sum+ 32*s->qscale*s->qscale;*/
978 return dmin_sum+ 11*c->mb_penalty_factor;
982 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
983 int ymax= s->me.ymax>>interlaced;
984 int ymin= s->me.ymin>>interlaced;
986 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
987 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
988 if(mv[1] < ymin) mv[1] = ymin;
989 if(mv[1] > ymax) mv[1] = ymax;
992 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
993 MotionEstContext * const c= &s->me;
994 Picture *p= s->current_picture_ptr;
995 int mb_xy= mb_x + mb_y*s->mb_stride;
996 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
997 int mb_type= s->current_picture.mb_type[mb_xy];
999 int shift= (flags&FLAG_QPEL) + 1;
1000 int mask= (1<<shift)-1;
1003 me_cmp_func cmpf= s->dsp.sse[0];
1004 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1006 if(p_type && USES_LIST(mb_type, 1)){
1007 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1010 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1013 int xy= s->block_index[i];
1014 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1015 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1018 if(IS_INTERLACED(mb_type)){
1019 int xy2= xy + s->b8_stride;
1020 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1024 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1025 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1029 if(USES_LIST(mb_type, 0)){
1030 int field_select0= p->ref_index[0][xy ];
1031 int field_select1= p->ref_index[0][xy2];
1032 assert(field_select0==0 ||field_select0==1);
1033 assert(field_select1==0 ||field_select1==1);
1034 init_interlaced_ref(s, 0);
1037 s->p_field_select_table[0][mb_xy]= field_select0;
1038 s->p_field_select_table[1][mb_xy]= field_select1;
1039 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1040 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1041 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1043 s->b_field_select_table[0][0][mb_xy]= field_select0;
1044 s->b_field_select_table[0][1][mb_xy]= field_select1;
1045 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1046 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1047 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1050 x= p->motion_val[0][xy ][0];
1051 y= p->motion_val[0][xy ][1];
1052 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1053 x= p->motion_val[0][xy2][0];
1054 y= p->motion_val[0][xy2][1];
1055 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1057 if(USES_LIST(mb_type, 1)){
1058 int field_select0= p->ref_index[1][xy ];
1059 int field_select1= p->ref_index[1][xy2];
1060 assert(field_select0==0 ||field_select0==1);
1061 assert(field_select1==0 ||field_select1==1);
1062 init_interlaced_ref(s, 2);
1064 s->b_field_select_table[1][0][mb_xy]= field_select0;
1065 s->b_field_select_table[1][1][mb_xy]= field_select1;
1066 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1067 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1068 if(USES_LIST(mb_type, 0)){
1069 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1071 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1074 x= p->motion_val[1][xy ][0];
1075 y= p->motion_val[1][xy ][1];
1076 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1077 x= p->motion_val[1][xy2][0];
1078 y= p->motion_val[1][xy2][1];
1079 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1080 //FIXME bidir scores
1084 }else if(IS_8X8(mb_type)){
1085 if(!(s->flags & CODEC_FLAG_4MV)){
1086 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1089 cmpf= s->dsp.sse[1];
1090 chroma_cmpf= s->dsp.sse[1];
1093 xy= s->block_index[i];
1094 x= p->motion_val[0][xy][0];
1095 y= p->motion_val[0][xy][1];
1096 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1098 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1100 if(USES_LIST(mb_type, 0)){
1102 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1103 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1104 }else if(USES_LIST(mb_type, 1)){
1105 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1106 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1107 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1109 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1110 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1112 x= p->motion_val[0][xy][0];
1113 y= p->motion_val[0][xy][1];
1114 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1115 }else if(USES_LIST(mb_type, 1)){
1116 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1117 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1119 x= p->motion_val[1][xy][0];
1120 y= p->motion_val[1][xy][1];
1121 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1123 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1128 void ff_estimate_p_frame_motion(MpegEncContext * s,
1131 MotionEstContext * const c= &s->me;
1132 uint8_t *pix, *ppix;
1133 int sum, varc, vard, mx, my, dmin;
1135 const int shift= 1+s->quarter_sample;
1137 Picture * const pic= &s->current_picture;
1139 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1141 assert(s->quarter_sample==0 || s->quarter_sample==1);
1142 assert(s->linesize == c->stride);
1143 assert(s->uvlinesize == c->uvstride);
1145 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1146 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1147 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1148 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1150 get_limits(s, 16*mb_x, 16*mb_y);
1153 /* intra / predictive decision */
1155 sum = s->dsp.pix_sum(pix, s->linesize);
1156 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1158 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1159 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1160 c->mb_var_sum_temp += varc;
1162 if(c->avctx->me_threshold){
1163 vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
1165 if(vard<c->avctx->me_threshold){
1166 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1167 c->mc_mb_var_sum_temp += vard;
1168 if (vard <= 64 || vard < varc) { //FIXME
1169 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1171 c->scene_change_score+= s->qscale;
1175 if(vard<c->avctx->mb_threshold)
1176 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1179 switch(s->me_method) {
1182 no_motion_search(s, &mx, &my);
1189 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1194 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1199 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1207 const int mot_stride = s->b8_stride;
1208 const int mot_xy = s->block_index[0];
1210 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1211 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1213 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1215 if(!s->first_slice_line) {
1216 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1217 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1218 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1219 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1220 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1221 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1222 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1224 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1225 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1227 if(s->out_format == FMT_H263){
1228 c->pred_x = P_MEDIAN[0];
1229 c->pred_y = P_MEDIAN[1];
1230 }else { /* mpeg1 at least */
1231 c->pred_x= P_LEFT[0];
1232 c->pred_y= P_LEFT[1];
1235 c->pred_x= P_LEFT[0];
1236 c->pred_y= P_LEFT[1];
1240 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1245 /* At this point (mx,my) are full-pell and the relative displacement */
1246 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1248 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8;
1250 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1251 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1252 c->mc_mb_var_sum_temp += vard;
1255 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1256 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1259 if (vard <= 64 || vard < varc)
1260 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1262 c->scene_change_score+= s->qscale;
1264 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1265 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1266 set_p_mv_tables(s, mx, my, 1);
1271 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1272 h263_mv4_search(s, mx, my, shift);
1274 set_p_mv_tables(s, mx, my, 0);
1276 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1277 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1279 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1280 if (vard <= 64 || vard < varc)
1281 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1283 c->scene_change_score+= s->qscale;
1285 if (vard*2 + 200 > varc)
1286 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1287 if (varc*2 + 200 > vard){
1288 mb_type|= CANDIDATE_MB_TYPE_INTER;
1289 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1290 if(s->flags&CODEC_FLAG_MV0)
1292 mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference
1297 if((s->flags&CODEC_FLAG_4MV)
1298 && !c->skip && varc>50 && vard>10){
1299 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1300 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1302 set_p_mv_tables(s, mx, my, 0);
1304 set_p_mv_tables(s, mx, my, 1);
1305 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1306 && !c->skip){ //FIXME varc/d checks
1307 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1308 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1312 mb_type= CANDIDATE_MB_TYPE_INTER;
1314 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1315 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1316 dmin= get_mb_score(s, mx, my, 0, 0);
1318 if((s->flags&CODEC_FLAG_4MV)
1319 && !c->skip && varc>50 && vard>10){
1320 int dmin4= h263_mv4_search(s, mx, my, shift);
1322 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1326 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1327 && !c->skip){ //FIXME varc/d checks
1328 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1330 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1335 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1336 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1338 /* get intra luma score */
1339 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1340 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it
1342 int mean= (sum+128)>>8;
1345 for(i=0; i<16; i++){
1346 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1347 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1348 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1349 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1352 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1355 /* get chroma score */
1356 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1361 if(s->out_format == FMT_H263){
1362 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1364 mean= (s->last_dc[i] + 4)>>3;
1366 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1370 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1371 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1374 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1378 intra_score += c->mb_penalty_factor*16;
1380 if(intra_score < dmin){
1381 mb_type= CANDIDATE_MB_TYPE_INTRA;
1382 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1384 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1386 if (vard <= 64 || vard < varc) { //FIXME
1387 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1389 c->scene_change_score+= s->qscale;
1393 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1396 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1399 MotionEstContext * const c= &s->me;
1402 const int shift= 1+s->quarter_sample;
1403 const int xy= mb_x + mb_y*s->mb_stride;
1404 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1406 assert(s->quarter_sample==0 || s->quarter_sample==1);
1408 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1409 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1411 get_limits(s, 16*mb_x, 16*mb_y);
1414 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1415 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1417 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1419 /* special case for first line */
1420 if (s->first_slice_line) {
1421 c->pred_x= P_LEFT[0];
1422 c->pred_y= P_LEFT[1];
1423 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1424 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1426 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1427 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1428 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1429 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1430 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1431 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1432 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1434 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1435 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1437 c->pred_x = P_MEDIAN[0];
1438 c->pred_y = P_MEDIAN[1];
1441 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1443 s->p_mv_table[xy][0] = mx<<shift;
1444 s->p_mv_table[xy][1] = my<<shift;
1449 static int ff_estimate_motion_b(MpegEncContext * s,
1450 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1452 MotionEstContext * const c= &s->me;
1455 const int shift= 1+s->quarter_sample;
1456 const int mot_stride = s->mb_stride;
1457 const int mot_xy = mb_y*mot_stride + mb_x;
1458 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1461 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1462 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1463 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1464 c->current_mv_penalty= mv_penalty;
1466 get_limits(s, 16*mb_x, 16*mb_y);
1468 switch(s->me_method) {
1471 no_motion_search(s, &mx, &my);
1478 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1483 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1488 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1496 P_LEFT[0] = mv_table[mot_xy - 1][0];
1497 P_LEFT[1] = mv_table[mot_xy - 1][1];
1499 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1501 /* special case for first line */
1502 if (!s->first_slice_line) {
1503 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1504 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1505 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1506 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1507 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1508 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1509 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1511 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1512 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1514 c->pred_x= P_LEFT[0];
1515 c->pred_y= P_LEFT[1];
1518 if(mv_table == s->b_forw_mv_table){
1519 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1521 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1524 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1529 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1531 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1532 dmin= get_mb_score(s, mx, my, 0, ref_index);
1534 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1535 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1536 mv_table[mot_xy][0]= mx;
1537 mv_table[mot_xy][1]= my;
1542 static inline int check_bidir_mv(MpegEncContext * s,
1543 int motion_fx, int motion_fy,
1544 int motion_bx, int motion_by,
1545 int pred_fx, int pred_fy,
1546 int pred_bx, int pred_by,
1550 //FIXME better f_code prediction (max mv & distance)
1552 MotionEstContext * const c= &s->me;
1553 uint8_t * const mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1554 int stride= c->stride;
1555 int uvstride= c->uvstride;
1556 uint8_t *dest_y = c->scratchpad;
1561 uint8_t **src_data= c->src[0];
1562 uint8_t **ref_data= c->ref[0];
1563 uint8_t **ref2_data= c->ref[2];
1565 if(s->quarter_sample){
1566 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1567 src_x = motion_fx >> 2;
1568 src_y = motion_fy >> 2;
1570 ptr = ref_data[0] + (src_y * stride) + src_x;
1571 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1573 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1574 src_x = motion_bx >> 2;
1575 src_y = motion_by >> 2;
1577 ptr = ref2_data[0] + (src_y * stride) + src_x;
1578 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1580 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1581 src_x = motion_fx >> 1;
1582 src_y = motion_fy >> 1;
1584 ptr = ref_data[0] + (src_y * stride) + src_x;
1585 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1587 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1588 src_x = motion_bx >> 1;
1589 src_y = motion_by >> 1;
1591 ptr = ref2_data[0] + (src_y * stride) + src_x;
1592 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1595 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*c->mb_penalty_factor
1596 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*c->mb_penalty_factor
1597 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1599 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1606 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1607 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1609 const int mot_stride = s->mb_stride;
1610 const int xy = mb_y *mot_stride + mb_x;
1612 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1613 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1614 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1615 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1616 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1617 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1618 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1619 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1621 //FIXME do refinement and add flag
1623 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1624 motion_bx, motion_by,
1632 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1634 MotionEstContext * const c= &s->me;
1636 const int mot_stride = s->mb_stride;
1637 const int mot_xy = mb_y*mot_stride + mb_x;
1638 const int shift= 1+s->quarter_sample;
1640 const int time_pp= s->pp_time;
1641 const int time_pb= s->pb_time;
1642 int mx, my, xmin, xmax, ymin, ymax;
1643 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1645 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1646 ymin= xmin=(-32)>>shift;
1647 ymax= xmax= 31>>shift;
1649 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1650 s->mv_type= MV_TYPE_8X8;
1652 s->mv_type= MV_TYPE_16X16;
1656 int index= s->block_index[i];
1659 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1660 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1661 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1662 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1663 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1664 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1666 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1667 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1668 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1670 xmax= FFMIN(xmax, s->width - max);
1671 xmin= FFMAX(xmin, - 16 - min);
1673 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1674 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1675 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1677 ymax= FFMIN(ymax, s->height - max);
1678 ymin= FFMAX(ymin, - 16 - min);
1680 if(s->mv_type == MV_TYPE_16X16) break;
1683 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1685 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1686 s->b_direct_mv_table[mot_xy][0]= 0;
1687 s->b_direct_mv_table[mot_xy][1]= 0;
1689 return 256*256*256*64;
1696 c->flags |= FLAG_DIRECT;
1697 c->sub_flags |= FLAG_DIRECT;
1701 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1702 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1704 /* special case for first line */
1705 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
1706 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1707 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1708 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1709 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1711 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1712 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1715 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1716 if(c->sub_flags&FLAG_QPEL)
1717 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1719 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1721 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1722 dmin= get_mb_score(s, mx, my, 0, 0);
1724 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1726 s->b_direct_mv_table[mot_xy][0]= mx;
1727 s->b_direct_mv_table[mot_xy][1]= my;
1728 c->flags &= ~FLAG_DIRECT;
1729 c->sub_flags &= ~FLAG_DIRECT;
1734 void ff_estimate_b_frame_motion(MpegEncContext * s,
1737 MotionEstContext * const c= &s->me;
1738 const int penalty_factor= c->mb_penalty_factor;
1739 int fmin, bmin, dmin, fbmin, bimin, fimin;
1741 const int xy = mb_y*s->mb_stride + mb_x;
1742 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1744 get_limits(s, 16*mb_x, 16*mb_y);
1747 if(c->avctx->me_threshold){
1748 int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
1750 if(vard<c->avctx->me_threshold){
1751 // pix = c->src[0][0];
1752 // sum = s->dsp.pix_sum(pix, s->linesize);
1753 // varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1755 // pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1756 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1757 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1758 c->mb_var_sum_temp += varc;*/
1759 c->mc_mb_var_sum_temp += vard;
1760 /* if (vard <= 64 || vard < varc) {
1761 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1763 c->scene_change_score+= s->qscale;
1767 if(vard<c->avctx->mb_threshold){
1768 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1769 if(type == CANDIDATE_MB_TYPE_DIRECT){
1770 direct_search(s, mb_x, mb_y);
1772 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1774 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1776 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1778 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1780 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1782 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1783 interlaced_search(s, 0,
1784 s->b_field_mv_table[0], s->b_field_select_table[0],
1785 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1787 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1789 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1790 interlaced_search(s, 2,
1791 s->b_field_mv_table[1], s->b_field_select_table[1],
1792 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1798 if (s->codec_id == CODEC_ID_MPEG4)
1799 dmin= direct_search(s, mb_x, mb_y);
1802 //FIXME penalty stuff for non mpeg4
1804 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1807 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1808 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1811 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1812 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1814 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1815 //FIXME mb type penalty
1817 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1818 fimin= interlaced_search(s, 0,
1819 s->b_field_mv_table[0], s->b_field_select_table[0],
1820 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1821 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1822 bimin= interlaced_search(s, 2,
1823 s->b_field_mv_table[1], s->b_field_select_table[1],
1824 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1826 fimin= bimin= INT_MAX;
1830 type = CANDIDATE_MB_TYPE_FORWARD;
1834 type = CANDIDATE_MB_TYPE_DIRECT;
1838 type= CANDIDATE_MB_TYPE_BACKWARD;
1842 type= CANDIDATE_MB_TYPE_BIDIR;
1846 type= CANDIDATE_MB_TYPE_FORWARD_I;
1850 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1853 score= ((unsigned)(score*score + 128*256))>>16;
1854 c->mc_mb_var_sum_temp += score;
1855 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1858 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1859 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1861 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1863 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1864 if(fimin < INT_MAX && bimin < INT_MAX){
1865 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1867 //FIXME something smarter
1868 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1870 if(s->out_format == FMT_MPEG1)
1871 type |= CANDIDATE_MB_TYPE_INTRA;
1875 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1878 /* find best f_code for ME which do unlimited searches */
1879 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1881 if(s->me_method>=ME_EPZS){
1884 uint8_t * fcode_tab= s->fcode_tab;
1886 int best_score=-10000000;
1888 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1890 for(y=0; y<s->mb_height; y++){
1892 int xy= y*s->mb_stride;
1893 for(x=0; x<s->mb_width; x++){
1894 if(s->mb_type[xy] & type){
1895 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1896 fcode_tab[mv_table[xy][1] + MAX_MV]);
1899 for(j=0; j<fcode && j<8; j++){
1900 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1909 if(score[i] > best_score){
1910 best_score= score[i];
1913 // printf("%d %d\n", i, score[i]);
1916 // printf("fcode: %d type: %d\n", i, s->pict_type);
1918 /* for(i=0; i<=MAX_FCODE; i++){
1919 printf("%d ", mv_num[i]);
1927 void ff_fix_long_p_mvs(MpegEncContext * s)
1929 MotionEstContext * const c= &s->me;
1930 const int f_code= s->f_code;
1932 assert(s->pict_type==P_TYPE);
1934 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1936 if(s->msmpeg4_version) range= 16;
1938 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1940 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1941 if(s->flags&CODEC_FLAG_4MV){
1942 const int wrap= s->b8_stride;
1944 /* clip / convert to intra 8x8 type MVs */
1945 for(y=0; y<s->mb_height; y++){
1947 int i= y*s->mb_stride;
1950 for(x=0; x<s->mb_width; x++){
1951 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1953 for(block=0; block<4; block++){
1954 int off= (block& 1) + (block>>1)*wrap;
1955 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1956 int my= s->current_picture.motion_val[0][ xy + off ][1];
1958 if( mx >=range || mx <-range
1959 || my >=range || my <-range){
1960 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1961 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1962 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
1975 * @param truncate 1 for truncation, 0 for using intra
1977 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1978 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1980 MotionEstContext * const c= &s->me;
1981 int y, h_range, v_range;
1983 // RAL: 8 in MPEG-1, 16 in MPEG-4
1984 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1986 if(s->msmpeg4_version) range= 16;
1987 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1990 v_range= field_select_table ? range>>1 : range;
1992 /* clip / convert to intra 16x16 type MVs */
1993 for(y=0; y<s->mb_height; y++){
1995 int xy= y*s->mb_stride;
1996 for(x=0; x<s->mb_width; x++){
1997 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1998 if(field_select_table==NULL || field_select_table[xy] == field_select){
1999 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2000 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2003 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2004 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2005 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2006 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2008 s->mb_type[xy] &= ~type;
2009 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;