3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
36 #include "mpegvideo.h"
41 #define SQ(a) ((a)*(a))
45 #define P_TOPRIGHT P[3]
49 static inline int sad_hpel_motion_search(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
51 int src_index, int ref_index,
54 static inline int update_map_generation(MotionEstContext *c)
56 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57 if(c->map_generation==0){
58 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
61 return c->map_generation;
64 /* shape adaptive search stuff */
65 typedef struct Minima{
71 static int minima_cmp(const void *a, const void *b){
72 const Minima *da = (const Minima *) a;
73 const Minima *db = (const Minima *) b;
75 return da->height - db->height;
78 #define FLAG_QPEL 1 //must be 1
82 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){
83 const int offset[3]= {
85 ((y*c->uvstride + x)>>1),
86 ((y*c->uvstride + x)>>1),
90 c->src[0][i]= src [i] + offset[i];
91 c->ref[0][i]= ref [i] + offset[i];
95 c->ref[ref_index][i]= ref2[i] + offset[i];
100 static int get_flags(MotionEstContext *c, int direct, int chroma){
101 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102 + (direct ? FLAG_DIRECT : 0)
103 + (chroma ? FLAG_CHROMA : 0);
106 /*! \brief compares a block (either a full macroblock or a partition thereof)
107 against a proposed motion-compensated prediction of that block
109 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
110 const int size, const int h, int ref_index, int src_index,
111 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
112 MotionEstContext * const c= &s->me;
113 const int stride= c->stride;
114 const int uvstride= c->uvstride;
115 const int qpel= flags&FLAG_QPEL;
116 const int chroma= flags&FLAG_CHROMA;
117 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
118 const int hx= subx + (x<<(1+qpel));
119 const int hy= suby + (y<<(1+qpel));
120 uint8_t * const * const ref= c->ref[ref_index];
121 uint8_t * const * const src= c->src[src_index];
123 //FIXME check chroma 4mv, (no crashes ...)
124 if(flags&FLAG_DIRECT){
125 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
126 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
127 const int time_pp= s->pp_time;
128 const int time_pb= s->pb_time;
129 const int mask= 2*qpel+1;
130 if(s->mv_type==MV_TYPE_8X8){
133 int fx = c->direct_basis_mv[i][0] + hx;
134 int fy = c->direct_basis_mv[i][1] + hy;
135 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));
136 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));
137 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
138 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
140 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
142 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
143 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
145 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
146 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
150 int fx = c->direct_basis_mv[0][0] + hx;
151 int fy = c->direct_basis_mv[0][1] + hy;
152 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
153 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
154 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
155 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
158 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
159 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
160 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
161 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
162 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
163 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
164 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
165 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
167 assert((fx>>1) + 16*s->mb_x >= -16);
168 assert((fy>>1) + 16*s->mb_y >= -16);
169 assert((fx>>1) + 16*s->mb_x <= s->width);
170 assert((fy>>1) + 16*s->mb_y <= s->height);
171 assert((bx>>1) + 16*s->mb_x >= -16);
172 assert((by>>1) + 16*s->mb_y >= -16);
173 assert((bx>>1) + 16*s->mb_x <= s->width);
174 assert((by>>1) + 16*s->mb_y <= s->height);
176 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
177 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
180 d = cmp_func(s, c->temp, src[0], stride, 16);
184 int uvdxy; /* no, it might not be used uninitialized */
187 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
193 uvdxy= (cx&1) + 2*(cy&1);
194 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
197 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
199 uvdxy= dxy | (x&1) | (2*(y&1));
201 d = cmp_func(s, c->temp, src[0], stride, h);
203 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
205 uvdxy= (x&1) + 2*(y&1);
208 uint8_t * const uvtemp= c->temp + 16*stride;
209 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
211 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
212 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
217 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
221 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
226 #include "motion_est_template.c"
228 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
232 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
235 void ff_init_me(MpegEncContext *s){
236 MotionEstContext * const c= &s->me;
237 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
238 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
241 if(cache_size < 2*dia_size && !c->stride){
242 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
245 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
246 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
247 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
248 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
250 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
251 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
252 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
254 /*FIXME s->no_rounding b_type*/
255 if(s->flags&CODEC_FLAG_QPEL){
256 c->sub_motion_search= qpel_motion_search;
257 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
258 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
259 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
261 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
262 c->sub_motion_search= hpel_motion_search;
263 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
264 && c->avctx-> me_cmp == FF_CMP_SAD
265 && c->avctx-> mb_cmp == FF_CMP_SAD)
266 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
268 c->sub_motion_search= hpel_motion_search;
270 c->hpel_avg= s->dsp.avg_pixels_tab;
271 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
272 else c->hpel_put= s->dsp.put_pixels_tab;
275 c->stride = s->linesize;
276 c->uvstride= s->uvlinesize;
278 c->stride = 16*s->mb_width + 32;
279 c->uvstride= 8*s->mb_width + 16;
282 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
283 * not have yet, and even if we had, the motion estimation code
284 * does not expect it. */
285 if(s->codec_id != CODEC_ID_SNOW){
286 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
287 s->dsp.me_cmp[2]= zero_cmp;
289 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
290 s->dsp.me_sub_cmp[2]= zero_cmp;
292 c->hpel_put[2][0]= c->hpel_put[2][1]=
293 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
296 if(s->codec_id == CODEC_ID_H261){
297 c->sub_motion_search= no_sub_motion_search;
300 c->temp= c->scratchpad;
304 static int pix_dev(uint8_t * pix, int line_size, int mean)
309 for (i = 0; i < 16; i++) {
310 for (j = 0; j < 16; j += 8) {
311 s += FFABS(pix[0]-mean);
312 s += FFABS(pix[1]-mean);
313 s += FFABS(pix[2]-mean);
314 s += FFABS(pix[3]-mean);
315 s += FFABS(pix[4]-mean);
316 s += FFABS(pix[5]-mean);
317 s += FFABS(pix[6]-mean);
318 s += FFABS(pix[7]-mean);
321 pix += line_size - 16;
327 static inline void no_motion_search(MpegEncContext * s,
328 int *mx_ptr, int *my_ptr)
330 *mx_ptr = 16 * s->mb_x;
331 *my_ptr = 16 * s->mb_y;
334 #if 0 /* the use of these functions is inside #if 0 */
335 static int full_motion_search(MpegEncContext * s,
336 int *mx_ptr, int *my_ptr, int range,
337 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
339 int x1, y1, x2, y2, xx, yy, x, y;
345 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
357 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
361 for (y = y1; y <= y2; y++) {
362 for (x = x1; x <= x2; x++) {
363 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
367 (abs(x - xx) + abs(y - yy)) <
368 (abs(mx - xx) + abs(my - yy)))) {
380 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
381 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
382 av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
389 static int log_motion_search(MpegEncContext * s,
390 int *mx_ptr, int *my_ptr, int range,
391 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
393 int x1, y1, x2, y2, xx, yy, x, y;
420 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
426 for (y = y1; y <= y2; y += range) {
427 for (x = x1; x <= x2; x += range) {
428 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
429 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
455 } while (range >= 1);
458 av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
465 static int phods_motion_search(MpegEncContext * s,
466 int *mx_ptr, int *my_ptr, int range,
467 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
469 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
470 int mx, my, dminx, dminy;
496 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
507 for (x = x1; x <= x2; x += range) {
508 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
509 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
516 for (y = y1; y <= y2; y += range) {
517 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
518 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
544 } while (range >= 1);
547 av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
550 /* half pixel search */
557 #define Z_THRESHOLD 256
559 #define CHECK_SAD_HALF_MV(suffix, x, y) \
561 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
562 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
563 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
566 static inline int sad_hpel_motion_search(MpegEncContext * s,
567 int *mx_ptr, int *my_ptr, int dmin,
568 int src_index, int ref_index,
571 MotionEstContext * const c= &s->me;
572 const int penalty_factor= c->sub_penalty_factor;
575 int stride= c->stride;
576 const int flags= c->sub_flags;
589 pix = c->src[src_index][0];
593 ptr = c->ref[ref_index][0] + (my * stride) + mx;
597 if (mx > xmin && mx < xmax &&
598 my > ymin && my < ymax) {
601 const int index= (my<<ME_MAP_SHIFT) + mx;
602 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
603 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
604 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
605 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
615 CHECK_SAD_HALF_MV(y2 , 0, -1)
617 CHECK_SAD_HALF_MV(xy2, -1, -1)
619 CHECK_SAD_HALF_MV(xy2, +1, -1)
623 CHECK_SAD_HALF_MV(xy2, -1, +1)
625 CHECK_SAD_HALF_MV(x2 , -1, 0)
627 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)
640 CHECK_SAD_HALF_MV(xy2, -1, -1)
644 CHECK_SAD_HALF_MV(xy2, +1, +1)
646 CHECK_SAD_HALF_MV(x2 , -1, 0)
647 CHECK_SAD_HALF_MV(xy2, -1, +1)
650 CHECK_SAD_HALF_MV(xy2, +1, -1)
654 CHECK_SAD_HALF_MV(xy2, -1, +1)
656 CHECK_SAD_HALF_MV(x2 , +1, 0)
657 CHECK_SAD_HALF_MV(xy2, +1, +1)
659 CHECK_SAD_HALF_MV(y2 , 0, +1)
674 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
676 const int xy= s->mb_x + s->mb_y*s->mb_stride;
678 s->p_mv_table[xy][0] = mx;
679 s->p_mv_table[xy][1] = my;
681 /* has already been set to the 4 MV if 4MV is done */
683 int mot_xy= s->block_index[0];
685 s->current_picture.motion_val[0][mot_xy ][0]= mx;
686 s->current_picture.motion_val[0][mot_xy ][1]= my;
687 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
688 s->current_picture.motion_val[0][mot_xy+1][1]= my;
690 mot_xy += s->b8_stride;
691 s->current_picture.motion_val[0][mot_xy ][0]= mx;
692 s->current_picture.motion_val[0][mot_xy ][1]= my;
693 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
694 s->current_picture.motion_val[0][mot_xy+1][1]= my;
699 * get fullpel ME search limits.
701 static inline void get_limits(MpegEncContext *s, int x, int y)
703 MotionEstContext * const c= &s->me;
704 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
706 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
709 if (s->unrestricted_mv) {
712 c->xmax = - x + s->mb_width *16;
713 c->ymax = - y + s->mb_height*16;
714 } else if (s->out_format == FMT_H261){
715 // Search range of H261 is different from other codec standards
716 c->xmin = (x > 15) ? - 15 : 0;
717 c->ymin = (y > 15) ? - 15 : 0;
718 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
719 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
723 c->xmax = - x + s->mb_width *16 - 16;
724 c->ymax = - y + s->mb_height*16 - 16;
727 c->xmin = FFMAX(c->xmin,-range);
728 c->xmax = FFMIN(c->xmax, range);
729 c->ymin = FFMAX(c->ymin,-range);
730 c->ymax = FFMIN(c->ymax, range);
734 static inline void init_mv4_ref(MotionEstContext *c){
735 const int stride= c->stride;
737 c->ref[1][0] = c->ref[0][0] + 8;
738 c->ref[2][0] = c->ref[0][0] + 8*stride;
739 c->ref[3][0] = c->ref[2][0] + 8;
740 c->src[1][0] = c->src[0][0] + 8;
741 c->src[2][0] = c->src[0][0] + 8*stride;
742 c->src[3][0] = c->src[2][0] + 8;
745 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
747 MotionEstContext * const c= &s->me;
752 int dmin_sum=0, mx4_sum=0, my4_sum=0;
754 const int stride= c->stride;
755 uint8_t *mv_penalty= c->current_mv_penalty;
759 for(block=0; block<4; block++){
761 int pred_x4, pred_y4;
763 static const int off[4]= {2, 1, 1, -1};
764 const int mot_stride = s->b8_stride;
765 const int mot_xy = s->block_index[block];
767 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
768 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
770 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
772 /* special case for first line */
773 if (s->first_slice_line && block<2) {
774 c->pred_x= pred_x4= P_LEFT[0];
775 c->pred_y= pred_y4= P_LEFT[1];
777 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
778 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
779 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
780 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
781 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
782 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
783 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
784 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
786 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
787 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
789 c->pred_x= pred_x4 = P_MEDIAN[0];
790 c->pred_y= pred_y4 = P_MEDIAN[1];
795 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
797 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
799 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
801 const int offset= ((block&1) + (block>>1)*stride)*8;
802 uint8_t *dest_y = c->scratchpad + offset;
803 if(s->quarter_sample){
804 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
805 dxy = ((my4 & 3) << 2) | (mx4 & 3);
808 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
810 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
812 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
813 dxy = ((my4 & 1) << 1) | (mx4 & 1);
816 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
818 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
820 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
824 if(s->quarter_sample){
832 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
833 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
835 if(mx4 != mx || my4 != my) same=0;
841 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
842 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);
845 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
850 mx= ff_h263_round_chroma(mx4_sum);
851 my= ff_h263_round_chroma(my4_sum);
852 dxy = ((my & 1) << 1) | (mx & 1);
854 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
857 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
858 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
860 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
861 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
864 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);
865 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);
871 switch(c->avctx->mb_cmp&0xFF){
873 return dmin_sum+ 32*s->qscale*s->qscale;*/
877 return dmin_sum+ 11*c->mb_penalty_factor;
881 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
882 MotionEstContext * const c= &s->me;
884 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
885 c->src[1][0] = c->src[0][0] + s->linesize;
886 if(c->flags & FLAG_CHROMA){
887 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
888 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
889 c->src[1][1] = c->src[0][1] + s->uvlinesize;
890 c->src[1][2] = c->src[0][2] + s->uvlinesize;
894 static int interlaced_search(MpegEncContext *s, int ref_index,
895 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
897 MotionEstContext * const c= &s->me;
902 uint8_t * const mv_penalty= c->current_mv_penalty;
904 const int stride= 2*s->linesize;
906 const int mot_stride= s->mb_stride;
907 const int xy= s->mb_x + s->mb_y*mot_stride;
913 init_interlaced_ref(s, ref_index);
915 for(block=0; block<2; block++){
917 int best_dmin= INT_MAX;
920 for(field_select=0; field_select<2; field_select++){
921 int dmin, mx_i, my_i;
922 int16_t (*mv_table)[2]= mv_tables[block][field_select];
924 if(user_field_select){
925 assert(field_select==0 || field_select==1);
926 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
927 if(field_select_tables[block][xy] != field_select)
931 P_LEFT[0] = mv_table[xy - 1][0];
932 P_LEFT[1] = mv_table[xy - 1][1];
933 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
935 c->pred_x= P_LEFT[0];
936 c->pred_y= P_LEFT[1];
938 if(!s->first_slice_line){
939 P_TOP[0] = mv_table[xy - mot_stride][0];
940 P_TOP[1] = mv_table[xy - mot_stride][1];
941 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
942 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
943 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
944 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
945 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
946 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
948 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
949 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
951 P_MV1[0]= mx; //FIXME not correct if block != field_select
954 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
956 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
958 mv_table[xy][0]= mx_i;
959 mv_table[xy][1]= my_i;
961 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
965 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
966 dxy = ((my_i & 1) << 1) | (mx_i & 1);
969 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
971 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
973 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
974 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
976 dmin+= c->mb_penalty_factor; //field_select bits
978 dmin += field_select != block; //slightly prefer same field
980 if(dmin < best_dmin){
982 best_field= field_select;
986 int16_t (*mv_table)[2]= mv_tables[block][best_field];
988 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
989 if(mv_table[xy][1]&1) same=0;
990 if(mv_table[xy][1]*2 != my) same=0;
991 if(best_field != block) same=0;
994 field_select_tables[block][xy]= best_field;
995 dmin_sum += best_dmin;
1006 switch(c->avctx->mb_cmp&0xFF){
1008 return dmin_sum+ 32*s->qscale*s->qscale;*/
1012 return dmin_sum+ 11*c->mb_penalty_factor;
1016 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1017 int ymax= s->me.ymax>>interlaced;
1018 int ymin= s->me.ymin>>interlaced;
1020 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1021 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1022 if(mv[1] < ymin) mv[1] = ymin;
1023 if(mv[1] > ymax) mv[1] = ymax;
1026 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1027 MotionEstContext * const c= &s->me;
1028 Picture *p= s->current_picture_ptr;
1029 int mb_xy= mb_x + mb_y*s->mb_stride;
1030 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1031 int mb_type= s->current_picture.mb_type[mb_xy];
1032 int flags= c->flags;
1033 int shift= (flags&FLAG_QPEL) + 1;
1034 int mask= (1<<shift)-1;
1037 me_cmp_func cmpf= s->dsp.sse[0];
1038 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1040 if(p_type && USES_LIST(mb_type, 1)){
1041 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1044 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1047 int xy= s->block_index[i];
1048 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1049 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1052 if(IS_INTERLACED(mb_type)){
1053 int xy2= xy + s->b8_stride;
1054 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1058 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1059 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1063 if(USES_LIST(mb_type, 0)){
1064 int field_select0= p->ref_index[0][xy ];
1065 int field_select1= p->ref_index[0][xy2];
1066 assert(field_select0==0 ||field_select0==1);
1067 assert(field_select1==0 ||field_select1==1);
1068 init_interlaced_ref(s, 0);
1071 s->p_field_select_table[0][mb_xy]= field_select0;
1072 s->p_field_select_table[1][mb_xy]= field_select1;
1073 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1074 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1075 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1077 s->b_field_select_table[0][0][mb_xy]= field_select0;
1078 s->b_field_select_table[0][1][mb_xy]= field_select1;
1079 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1080 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1081 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1084 x= p->motion_val[0][xy ][0];
1085 y= p->motion_val[0][xy ][1];
1086 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1087 x= p->motion_val[0][xy2][0];
1088 y= p->motion_val[0][xy2][1];
1089 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1091 if(USES_LIST(mb_type, 1)){
1092 int field_select0= p->ref_index[1][xy ];
1093 int field_select1= p->ref_index[1][xy2];
1094 assert(field_select0==0 ||field_select0==1);
1095 assert(field_select1==0 ||field_select1==1);
1096 init_interlaced_ref(s, 2);
1098 s->b_field_select_table[1][0][mb_xy]= field_select0;
1099 s->b_field_select_table[1][1][mb_xy]= field_select1;
1100 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1101 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1102 if(USES_LIST(mb_type, 0)){
1103 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1105 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1108 x= p->motion_val[1][xy ][0];
1109 y= p->motion_val[1][xy ][1];
1110 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1111 x= p->motion_val[1][xy2][0];
1112 y= p->motion_val[1][xy2][1];
1113 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1114 //FIXME bidir scores
1118 }else if(IS_8X8(mb_type)){
1119 if(!(s->flags & CODEC_FLAG_4MV)){
1120 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1123 cmpf= s->dsp.sse[1];
1124 chroma_cmpf= s->dsp.sse[1];
1127 xy= s->block_index[i];
1128 x= p->motion_val[0][xy][0];
1129 y= p->motion_val[0][xy][1];
1130 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1132 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1134 if(USES_LIST(mb_type, 0)){
1136 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1137 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1138 }else if(USES_LIST(mb_type, 1)){
1139 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1140 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1141 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1143 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1144 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1146 x= p->motion_val[0][xy][0];
1147 y= p->motion_val[0][xy][1];
1148 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1149 }else if(USES_LIST(mb_type, 1)){
1150 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1151 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1153 x= p->motion_val[1][xy][0];
1154 y= p->motion_val[1][xy][1];
1155 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1157 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1162 void ff_estimate_p_frame_motion(MpegEncContext * s,
1165 MotionEstContext * const c= &s->me;
1166 uint8_t *pix, *ppix;
1167 int sum, mx, my, dmin;
1168 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1169 int vard; ///< sum of squared differences with the estimated motion vector
1171 const int shift= 1+s->quarter_sample;
1173 Picture * const pic= &s->current_picture;
1175 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1177 assert(s->quarter_sample==0 || s->quarter_sample==1);
1178 assert(s->linesize == c->stride);
1179 assert(s->uvlinesize == c->uvstride);
1181 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1182 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1183 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1184 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1186 get_limits(s, 16*mb_x, 16*mb_y);
1189 /* intra / predictive decision */
1191 sum = s->dsp.pix_sum(pix, s->linesize);
1192 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1194 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1195 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1196 c->mb_var_sum_temp += (varc+128)>>8;
1198 if(c->avctx->me_threshold){
1199 vard= check_input_motion(s, mb_x, mb_y, 1);
1201 if((vard+128)>>8 < c->avctx->me_threshold){
1202 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1203 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1204 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1205 c->mc_mb_var_sum_temp += (vard+128)>>8;
1206 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1209 if((vard+128)>>8 < c->avctx->mb_threshold)
1210 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1213 switch(s->me_method) {
1216 no_motion_search(s, &mx, &my);
1223 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1228 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1233 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1241 const int mot_stride = s->b8_stride;
1242 const int mot_xy = s->block_index[0];
1244 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1245 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1247 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1249 if(!s->first_slice_line) {
1250 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1251 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1252 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1253 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1254 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1255 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1256 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1258 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1259 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1261 if(s->out_format == FMT_H263){
1262 c->pred_x = P_MEDIAN[0];
1263 c->pred_y = P_MEDIAN[1];
1264 }else { /* mpeg1 at least */
1265 c->pred_x= P_LEFT[0];
1266 c->pred_y= P_LEFT[1];
1269 c->pred_x= P_LEFT[0];
1270 c->pred_y= P_LEFT[1];
1274 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1279 /* At this point (mx,my) are full-pell and the relative displacement */
1280 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1282 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1284 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1285 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1286 c->mc_mb_var_sum_temp += (vard+128)>>8;
1289 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1290 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1293 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1294 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1295 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1297 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1298 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1299 set_p_mv_tables(s, mx, my, 1);
1304 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1305 h263_mv4_search(s, mx, my, shift);
1307 set_p_mv_tables(s, mx, my, 0);
1309 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1310 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1312 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1313 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1314 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1315 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1317 if (vard*2 + 200*256 > varc)
1318 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1319 if (varc*2 + 200*256 > vard || s->qscale > 24){
1320 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1321 mb_type|= CANDIDATE_MB_TYPE_INTER;
1322 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1323 if(s->flags&CODEC_FLAG_MV0)
1325 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1330 if((s->flags&CODEC_FLAG_4MV)
1331 && !c->skip && varc>50<<8 && vard>10<<8){
1332 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1333 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1335 set_p_mv_tables(s, mx, my, 0);
1337 set_p_mv_tables(s, mx, my, 1);
1338 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1339 && !c->skip){ //FIXME varc/d checks
1340 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1341 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1345 mb_type= CANDIDATE_MB_TYPE_INTER;
1347 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1348 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1349 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1351 if((s->flags&CODEC_FLAG_4MV)
1352 && !c->skip && varc>50<<8 && vard>10<<8){
1353 int dmin4= h263_mv4_search(s, mx, my, shift);
1355 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1359 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1360 && !c->skip){ //FIXME varc/d checks
1361 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1363 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1368 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1369 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1371 /* get intra luma score */
1372 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1373 intra_score= varc - 500;
1375 int mean= (sum+128)>>8;
1378 for(i=0; i<16; i++){
1379 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1380 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1381 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1382 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1385 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1388 /* get chroma score */
1389 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1394 if(s->out_format == FMT_H263){
1395 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1397 mean= (s->last_dc[i] + 4)>>3;
1399 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1403 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1404 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1407 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1411 intra_score += c->mb_penalty_factor*16;
1413 if(intra_score < dmin){
1414 mb_type= CANDIDATE_MB_TYPE_INTRA;
1415 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1417 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1420 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1421 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1422 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1426 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1429 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1432 MotionEstContext * const c= &s->me;
1435 const int shift= 1+s->quarter_sample;
1436 const int xy= mb_x + mb_y*s->mb_stride;
1437 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1439 assert(s->quarter_sample==0 || s->quarter_sample==1);
1441 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1442 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1444 get_limits(s, 16*mb_x, 16*mb_y);
1447 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1448 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1450 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1452 /* special case for first line */
1453 if (s->first_slice_line) {
1454 c->pred_x= P_LEFT[0];
1455 c->pred_y= P_LEFT[1];
1456 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1457 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1459 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1460 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1461 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1462 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1463 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1464 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1465 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1467 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1468 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1470 c->pred_x = P_MEDIAN[0];
1471 c->pred_y = P_MEDIAN[1];
1474 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1476 s->p_mv_table[xy][0] = mx<<shift;
1477 s->p_mv_table[xy][1] = my<<shift;
1482 static int ff_estimate_motion_b(MpegEncContext * s,
1483 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1485 MotionEstContext * const c= &s->me;
1488 const int shift= 1+s->quarter_sample;
1489 const int mot_stride = s->mb_stride;
1490 const int mot_xy = mb_y*mot_stride + mb_x;
1491 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1494 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1495 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1496 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1497 c->current_mv_penalty= mv_penalty;
1499 get_limits(s, 16*mb_x, 16*mb_y);
1501 switch(s->me_method) {
1504 no_motion_search(s, &mx, &my);
1511 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1516 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1521 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1529 P_LEFT[0] = mv_table[mot_xy - 1][0];
1530 P_LEFT[1] = mv_table[mot_xy - 1][1];
1532 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1534 /* special case for first line */
1535 if (!s->first_slice_line) {
1536 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1537 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1538 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1539 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1540 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1541 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1542 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1544 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1545 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1547 c->pred_x= P_LEFT[0];
1548 c->pred_y= P_LEFT[1];
1551 if(mv_table == s->b_forw_mv_table){
1552 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1554 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1557 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1562 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1564 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1565 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1567 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1568 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1569 mv_table[mot_xy][0]= mx;
1570 mv_table[mot_xy][1]= my;
1575 static inline int check_bidir_mv(MpegEncContext * s,
1576 int motion_fx, int motion_fy,
1577 int motion_bx, int motion_by,
1578 int pred_fx, int pred_fy,
1579 int pred_bx, int pred_by,
1583 //FIXME better f_code prediction (max mv & distance)
1585 MotionEstContext * const c= &s->me;
1586 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1587 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1588 int stride= c->stride;
1589 uint8_t *dest_y = c->scratchpad;
1594 uint8_t **src_data= c->src[0];
1595 uint8_t **ref_data= c->ref[0];
1596 uint8_t **ref2_data= c->ref[2];
1598 if(s->quarter_sample){
1599 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1600 src_x = motion_fx >> 2;
1601 src_y = motion_fy >> 2;
1603 ptr = ref_data[0] + (src_y * stride) + src_x;
1604 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1606 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1607 src_x = motion_bx >> 2;
1608 src_y = motion_by >> 2;
1610 ptr = ref2_data[0] + (src_y * stride) + src_x;
1611 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1613 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1614 src_x = motion_fx >> 1;
1615 src_y = motion_fy >> 1;
1617 ptr = ref_data[0] + (src_y * stride) + src_x;
1618 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1620 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1621 src_x = motion_bx >> 1;
1622 src_y = motion_by >> 1;
1624 ptr = ref2_data[0] + (src_y * stride) + src_x;
1625 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1628 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1629 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1630 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1632 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1639 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1640 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1642 MotionEstContext * const c= &s->me;
1643 const int mot_stride = s->mb_stride;
1644 const int xy = mb_y *mot_stride + mb_x;
1646 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1647 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1648 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1649 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1650 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1651 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1652 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1653 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1654 const int flags= c->sub_flags;
1655 const int qpel= flags&FLAG_QPEL;
1656 const int shift= 1+qpel;
1657 const int xmin= c->xmin<<shift;
1658 const int ymin= c->ymin<<shift;
1659 const int xmax= c->xmax<<shift;
1660 const int ymax= c->ymax<<shift;
1661 uint8_t map[8][8][8][8];
1663 memset(map,0,sizeof(map));
1664 #define BIDIR_MAP(fx,fy,bx,by) \
1665 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1666 BIDIR_MAP(0,0,0,0) = 1;
1668 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1669 motion_bx, motion_by,
1674 if(s->avctx->bidir_refine){
1676 #define CHECK_BIDIR(fx,fy,bx,by)\
1677 if( !BIDIR_MAP(fx,fy,bx,by)\
1678 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1679 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1680 BIDIR_MAP(fx,fy,bx,by) = 1;\
1681 score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
1691 #define CHECK_BIDIR2(a,b,c,d)\
1692 CHECK_BIDIR(a,b,c,d)\
1693 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1695 #define CHECK_BIDIRR(a,b,c,d)\
1696 CHECK_BIDIR2(a,b,c,d)\
1697 CHECK_BIDIR2(b,c,d,a)\
1698 CHECK_BIDIR2(c,d,a,b)\
1699 CHECK_BIDIR2(d,a,b,c)
1704 CHECK_BIDIRR( 0, 0, 0, 1)
1705 if(s->avctx->bidir_refine > 1){
1706 CHECK_BIDIRR( 0, 0, 1, 1)
1707 CHECK_BIDIR2( 0, 1, 0, 1)
1708 CHECK_BIDIR2( 1, 0, 1, 0)
1709 CHECK_BIDIRR( 0, 0,-1, 1)
1710 CHECK_BIDIR2( 0,-1, 0, 1)
1711 CHECK_BIDIR2(-1, 0, 1, 0)
1712 if(s->avctx->bidir_refine > 2){
1713 CHECK_BIDIRR( 0, 1, 1, 1)
1714 CHECK_BIDIRR( 0,-1, 1, 1)
1715 CHECK_BIDIRR( 0, 1,-1, 1)
1716 CHECK_BIDIRR( 0, 1, 1,-1)
1717 if(s->avctx->bidir_refine > 3){
1718 CHECK_BIDIR2( 1, 1, 1, 1)
1719 CHECK_BIDIRR( 1, 1, 1,-1)
1720 CHECK_BIDIR2( 1, 1,-1,-1)
1721 CHECK_BIDIR2( 1,-1,-1, 1)
1722 CHECK_BIDIR2( 1,-1, 1,-1)
1729 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1730 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1731 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1732 s->b_bidir_back_mv_table[xy][1]= motion_by;
1737 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1739 MotionEstContext * const c= &s->me;
1741 const int mot_stride = s->mb_stride;
1742 const int mot_xy = mb_y*mot_stride + mb_x;
1743 const int shift= 1+s->quarter_sample;
1745 const int time_pp= s->pp_time;
1746 const int time_pb= s->pb_time;
1747 int mx, my, xmin, xmax, ymin, ymax;
1748 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1750 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1751 ymin= xmin=(-32)>>shift;
1752 ymax= xmax= 31>>shift;
1754 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1755 s->mv_type= MV_TYPE_8X8;
1757 s->mv_type= MV_TYPE_16X16;
1761 int index= s->block_index[i];
1764 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1765 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1766 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1767 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1768 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1769 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1771 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1772 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1773 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1775 xmax= FFMIN(xmax, s->width - max);
1776 xmin= FFMAX(xmin, - 16 - min);
1778 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1779 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1780 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1782 ymax= FFMIN(ymax, s->height - max);
1783 ymin= FFMAX(ymin, - 16 - min);
1785 if(s->mv_type == MV_TYPE_16X16) break;
1788 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1790 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1791 s->b_direct_mv_table[mot_xy][0]= 0;
1792 s->b_direct_mv_table[mot_xy][1]= 0;
1794 return 256*256*256*64;
1801 c->flags |= FLAG_DIRECT;
1802 c->sub_flags |= FLAG_DIRECT;
1806 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1807 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1809 /* special case for first line */
1810 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1811 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1812 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1813 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1814 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1816 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1817 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1820 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1821 if(c->sub_flags&FLAG_QPEL)
1822 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1824 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1826 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1827 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1829 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1831 mv_table[mot_xy][0]= mx;
1832 mv_table[mot_xy][1]= my;
1833 c->flags &= ~FLAG_DIRECT;
1834 c->sub_flags &= ~FLAG_DIRECT;
1839 void ff_estimate_b_frame_motion(MpegEncContext * s,
1842 MotionEstContext * const c= &s->me;
1843 const int penalty_factor= c->mb_penalty_factor;
1844 int fmin, bmin, dmin, fbmin, bimin, fimin;
1846 const int xy = mb_y*s->mb_stride + mb_x;
1847 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1849 get_limits(s, 16*mb_x, 16*mb_y);
1853 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1854 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1856 score= ((unsigned)(score*score + 128*256))>>16;
1857 c->mc_mb_var_sum_temp += score;
1858 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1859 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1864 if(c->avctx->me_threshold){
1865 int vard= check_input_motion(s, mb_x, mb_y, 0);
1867 if((vard+128)>>8 < c->avctx->me_threshold){
1868 // pix = c->src[0][0];
1869 // sum = s->dsp.pix_sum(pix, s->linesize);
1870 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1872 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1873 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1874 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1875 c->mb_var_sum_temp += (varc+128)>>8;*/
1876 c->mc_mb_var_sum_temp += (vard+128)>>8;
1877 /* if (vard <= 64<<8 || vard < varc) {
1878 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1880 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1884 if((vard+128)>>8 < c->avctx->mb_threshold){
1885 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1886 if(type == CANDIDATE_MB_TYPE_DIRECT){
1887 direct_search(s, mb_x, mb_y);
1889 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1891 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1893 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1895 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1897 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1899 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1900 interlaced_search(s, 0,
1901 s->b_field_mv_table[0], s->b_field_select_table[0],
1902 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1904 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1906 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1907 interlaced_search(s, 2,
1908 s->b_field_mv_table[1], s->b_field_select_table[1],
1909 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1915 if (s->codec_id == CODEC_ID_MPEG4)
1916 dmin= direct_search(s, mb_x, mb_y);
1919 //FIXME penalty stuff for non mpeg4
1921 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1924 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1925 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1928 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1929 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1931 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1932 //FIXME mb type penalty
1934 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1935 fimin= interlaced_search(s, 0,
1936 s->b_field_mv_table[0], s->b_field_select_table[0],
1937 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1938 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1939 bimin= interlaced_search(s, 2,
1940 s->b_field_mv_table[1], s->b_field_select_table[1],
1941 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1943 fimin= bimin= INT_MAX;
1947 type = CANDIDATE_MB_TYPE_FORWARD;
1951 type = CANDIDATE_MB_TYPE_DIRECT;
1955 type= CANDIDATE_MB_TYPE_BACKWARD;
1959 type= CANDIDATE_MB_TYPE_BIDIR;
1963 type= CANDIDATE_MB_TYPE_FORWARD_I;
1967 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1970 score= ((unsigned)(score*score + 128*256))>>16;
1971 c->mc_mb_var_sum_temp += score;
1972 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1975 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1976 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1978 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1980 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1981 if(fimin < INT_MAX && bimin < INT_MAX){
1982 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1984 //FIXME something smarter
1985 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1986 if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1987 type |= CANDIDATE_MB_TYPE_DIRECT0;
1989 if(s->out_format == FMT_MPEG1)
1990 type |= CANDIDATE_MB_TYPE_INTRA;
1994 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1997 /* find best f_code for ME which do unlimited searches */
1998 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
2000 if(s->me_method>=ME_EPZS){
2002 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
2003 uint8_t * fcode_tab= s->fcode_tab;
2005 int best_score=-10000000;
2007 if(s->msmpeg4_version)
2008 range= FFMIN(range, 16);
2009 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2010 range= FFMIN(range, 256);
2012 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
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){
2019 int mx= mv_table[xy][0];
2020 int my= mv_table[xy][1];
2021 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2022 fcode_tab[my + MAX_MV]);
2025 if(mx >= range || mx < -range ||
2026 my >= range || my < -range)
2029 for(j=0; j<fcode && j<8; j++){
2030 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2039 if(score[i] > best_score){
2040 best_score= score[i];
2043 // printf("%d %d\n", i, score[i]);
2046 // printf("fcode: %d type: %d\n", i, s->pict_type);
2048 /* for(i=0; i<=MAX_FCODE; i++){
2049 printf("%d ", mv_num[i]);
2057 void ff_fix_long_p_mvs(MpegEncContext * s)
2059 MotionEstContext * const c= &s->me;
2060 const int f_code= s->f_code;
2062 assert(s->pict_type==P_TYPE);
2064 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2066 assert(range <= 16 || !s->msmpeg4_version);
2067 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2069 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2071 //printf("%d no:%d %d//\n", clip, noclip, f_code);
2072 if(s->flags&CODEC_FLAG_4MV){
2073 const int wrap= s->b8_stride;
2075 /* clip / convert to intra 8x8 type MVs */
2076 for(y=0; y<s->mb_height; y++){
2078 int i= y*s->mb_stride;
2081 for(x=0; x<s->mb_width; x++){
2082 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2084 for(block=0; block<4; block++){
2085 int off= (block& 1) + (block>>1)*wrap;
2086 int mx= s->current_picture.motion_val[0][ xy + off ][0];
2087 int my= s->current_picture.motion_val[0][ xy + off ][1];
2089 if( mx >=range || mx <-range
2090 || my >=range || my <-range){
2091 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2092 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2093 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2106 * @param truncate 1 for truncation, 0 for using intra
2108 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2109 int16_t (*mv_table)[2], int f_code, int type, int truncate)
2111 MotionEstContext * const c= &s->me;
2112 int y, h_range, v_range;
2114 // RAL: 8 in MPEG-1, 16 in MPEG-4
2115 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2117 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2120 v_range= field_select_table ? range>>1 : range;
2122 /* clip / convert to intra 16x16 type MVs */
2123 for(y=0; y<s->mb_height; y++){
2125 int xy= y*s->mb_stride;
2126 for(x=0; x<s->mb_width; x++){
2127 if (s->mb_type[xy] & type){ // RAL: "type" test added...
2128 if(field_select_table==NULL || field_select_table[xy] == field_select){
2129 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2130 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2133 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2134 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2135 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2136 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2138 s->mb_type[xy] &= ~type;
2139 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;