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 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
107 const int size, const int h, int ref_index, int src_index,
108 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
109 MotionEstContext * const c= &s->me;
110 const int stride= c->stride;
111 const int uvstride= c->uvstride;
112 const int qpel= flags&FLAG_QPEL;
113 const int chroma= flags&FLAG_CHROMA;
114 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
115 const int hx= subx + (x<<(1+qpel));
116 const int hy= suby + (y<<(1+qpel));
117 uint8_t * const * const ref= c->ref[ref_index];
118 uint8_t * const * const src= c->src[src_index];
120 //FIXME check chroma 4mv, (no crashes ...)
121 if(flags&FLAG_DIRECT){
122 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
123 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
124 const int time_pp= s->pp_time;
125 const int time_pb= s->pb_time;
126 const int mask= 2*qpel+1;
127 if(s->mv_type==MV_TYPE_8X8){
130 int fx = c->direct_basis_mv[i][0] + hx;
131 int fy = c->direct_basis_mv[i][1] + hy;
132 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));
133 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));
134 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
135 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
137 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
139 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
140 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
142 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
143 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
147 int fx = c->direct_basis_mv[0][0] + hx;
148 int fy = c->direct_basis_mv[0][1] + hy;
149 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
150 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
151 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
152 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
155 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
156 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
157 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
158 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
159 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
160 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
161 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
162 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
164 assert((fx>>1) + 16*s->mb_x >= -16);
165 assert((fy>>1) + 16*s->mb_y >= -16);
166 assert((fx>>1) + 16*s->mb_x <= s->width);
167 assert((fy>>1) + 16*s->mb_y <= s->height);
168 assert((bx>>1) + 16*s->mb_x >= -16);
169 assert((by>>1) + 16*s->mb_y >= -16);
170 assert((bx>>1) + 16*s->mb_x <= s->width);
171 assert((by>>1) + 16*s->mb_y <= s->height);
173 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
174 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
177 d = cmp_func(s, c->temp, src[0], stride, 16);
181 int uvdxy; /* no, it might not be used uninitialized */
184 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
190 uvdxy= (cx&1) + 2*(cy&1);
191 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
194 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
196 uvdxy= dxy | (x&1) | (2*(y&1));
198 d = cmp_func(s, c->temp, src[0], stride, h);
200 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
202 uvdxy= (x&1) + 2*(y&1);
205 uint8_t * const uvtemp= c->temp + 16*stride;
206 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
207 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
208 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
209 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
214 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
218 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
223 #include "motion_est_template.c"
225 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
229 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
232 void ff_init_me(MpegEncContext *s){
233 MotionEstContext * const c= &s->me;
234 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
235 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
238 if(cache_size < 2*dia_size && !c->stride){
239 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
242 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
243 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
244 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
245 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
247 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
248 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
249 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
251 /*FIXME s->no_rounding b_type*/
252 if(s->flags&CODEC_FLAG_QPEL){
253 c->sub_motion_search= qpel_motion_search;
254 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
255 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
256 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
258 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
259 c->sub_motion_search= hpel_motion_search;
260 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
261 && c->avctx-> me_cmp == FF_CMP_SAD
262 && c->avctx-> mb_cmp == FF_CMP_SAD)
263 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
265 c->sub_motion_search= hpel_motion_search;
267 c->hpel_avg= s->dsp.avg_pixels_tab;
268 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
269 else c->hpel_put= s->dsp.put_pixels_tab;
272 c->stride = s->linesize;
273 c->uvstride= s->uvlinesize;
275 c->stride = 16*s->mb_width + 32;
276 c->uvstride= 8*s->mb_width + 16;
279 // 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
280 if(s->codec_id != CODEC_ID_SNOW){
281 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
282 s->dsp.me_cmp[2]= zero_cmp;
284 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
285 s->dsp.me_sub_cmp[2]= zero_cmp;
287 c->hpel_put[2][0]= c->hpel_put[2][1]=
288 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
291 if(s->codec_id == CODEC_ID_H261){
292 c->sub_motion_search= no_sub_motion_search;
295 c->temp= c->scratchpad;
299 static int pix_dev(uint8_t * pix, int line_size, int mean)
304 for (i = 0; i < 16; i++) {
305 for (j = 0; j < 16; j += 8) {
306 s += FFABS(pix[0]-mean);
307 s += FFABS(pix[1]-mean);
308 s += FFABS(pix[2]-mean);
309 s += FFABS(pix[3]-mean);
310 s += FFABS(pix[4]-mean);
311 s += FFABS(pix[5]-mean);
312 s += FFABS(pix[6]-mean);
313 s += FFABS(pix[7]-mean);
316 pix += line_size - 16;
322 static inline void no_motion_search(MpegEncContext * s,
323 int *mx_ptr, int *my_ptr)
325 *mx_ptr = 16 * s->mb_x;
326 *my_ptr = 16 * s->mb_y;
329 #if 0 /* the use of these functions is inside #if 0 */
330 static int full_motion_search(MpegEncContext * s,
331 int *mx_ptr, int *my_ptr, int range,
332 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
334 int x1, y1, x2, y2, xx, yy, x, y;
340 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
352 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
356 for (y = y1; y <= y2; y++) {
357 for (x = x1; x <= x2; x++) {
358 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
362 (abs(x - xx) + abs(y - yy)) <
363 (abs(mx - xx) + abs(my - yy)))) {
375 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
376 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
377 av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
384 static int log_motion_search(MpegEncContext * s,
385 int *mx_ptr, int *my_ptr, int range,
386 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
388 int x1, y1, x2, y2, xx, yy, x, y;
415 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
421 for (y = y1; y <= y2; y += range) {
422 for (x = x1; x <= x2; x += range) {
423 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
424 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
450 } while (range >= 1);
453 av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
460 static int phods_motion_search(MpegEncContext * s,
461 int *mx_ptr, int *my_ptr, int range,
462 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
464 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
465 int mx, my, dminx, dminy;
491 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
502 for (x = x1; x <= x2; x += range) {
503 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
504 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
511 for (y = y1; y <= y2; y += range) {
512 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
513 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
539 } while (range >= 1);
542 av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
545 /* half pixel search */
552 #define Z_THRESHOLD 256
554 #define CHECK_SAD_HALF_MV(suffix, x, y) \
556 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
557 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
558 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
561 static inline int sad_hpel_motion_search(MpegEncContext * s,
562 int *mx_ptr, int *my_ptr, int dmin,
563 int src_index, int ref_index,
566 MotionEstContext * const c= &s->me;
567 const int penalty_factor= c->sub_penalty_factor;
570 int stride= c->stride;
571 const int flags= c->sub_flags;
584 pix = c->src[src_index][0];
588 ptr = c->ref[ref_index][0] + (my * stride) + mx;
592 if (mx > xmin && mx < xmax &&
593 my > ymin && my < ymax) {
596 const int index= (my<<ME_MAP_SHIFT) + mx;
597 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
598 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
599 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
600 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
610 CHECK_SAD_HALF_MV(y2 , 0, -1)
612 CHECK_SAD_HALF_MV(xy2, -1, -1)
614 CHECK_SAD_HALF_MV(xy2, +1, -1)
618 CHECK_SAD_HALF_MV(xy2, -1, +1)
620 CHECK_SAD_HALF_MV(x2 , -1, 0)
622 CHECK_SAD_HALF_MV(xy2, +1, -1)
624 CHECK_SAD_HALF_MV(xy2, -1, -1)
628 CHECK_SAD_HALF_MV(xy2, +1, +1)
630 CHECK_SAD_HALF_MV(x2 , +1, 0)
635 CHECK_SAD_HALF_MV(xy2, -1, -1)
639 CHECK_SAD_HALF_MV(xy2, +1, +1)
641 CHECK_SAD_HALF_MV(x2 , -1, 0)
642 CHECK_SAD_HALF_MV(xy2, -1, +1)
645 CHECK_SAD_HALF_MV(xy2, +1, -1)
649 CHECK_SAD_HALF_MV(xy2, -1, +1)
651 CHECK_SAD_HALF_MV(x2 , +1, 0)
652 CHECK_SAD_HALF_MV(xy2, +1, +1)
654 CHECK_SAD_HALF_MV(y2 , 0, +1)
669 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
671 const int xy= s->mb_x + s->mb_y*s->mb_stride;
673 s->p_mv_table[xy][0] = mx;
674 s->p_mv_table[xy][1] = my;
676 /* has already been set to the 4 MV if 4MV is done */
678 int mot_xy= s->block_index[0];
680 s->current_picture.motion_val[0][mot_xy ][0]= mx;
681 s->current_picture.motion_val[0][mot_xy ][1]= my;
682 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
683 s->current_picture.motion_val[0][mot_xy+1][1]= my;
685 mot_xy += s->b8_stride;
686 s->current_picture.motion_val[0][mot_xy ][0]= mx;
687 s->current_picture.motion_val[0][mot_xy ][1]= my;
688 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
689 s->current_picture.motion_val[0][mot_xy+1][1]= my;
694 * get fullpel ME search limits.
696 static inline void get_limits(MpegEncContext *s, int x, int y)
698 MotionEstContext * const c= &s->me;
699 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
701 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
704 if (s->unrestricted_mv) {
707 c->xmax = - x + s->mb_width *16;
708 c->ymax = - y + s->mb_height*16;
709 } else if (s->out_format == FMT_H261){
710 // Search range of H261 is different from other codec standards
711 c->xmin = (x > 15) ? - 15 : 0;
712 c->ymin = (y > 15) ? - 15 : 0;
713 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
714 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
718 c->xmax = - x + s->mb_width *16 - 16;
719 c->ymax = - y + s->mb_height*16 - 16;
722 c->xmin = FFMAX(c->xmin,-range);
723 c->xmax = FFMIN(c->xmax, range);
724 c->ymin = FFMAX(c->ymin,-range);
725 c->ymax = FFMIN(c->ymax, range);
729 static inline void init_mv4_ref(MotionEstContext *c){
730 const int stride= c->stride;
732 c->ref[1][0] = c->ref[0][0] + 8;
733 c->ref[2][0] = c->ref[0][0] + 8*stride;
734 c->ref[3][0] = c->ref[2][0] + 8;
735 c->src[1][0] = c->src[0][0] + 8;
736 c->src[2][0] = c->src[0][0] + 8*stride;
737 c->src[3][0] = c->src[2][0] + 8;
740 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
742 MotionEstContext * const c= &s->me;
747 int dmin_sum=0, mx4_sum=0, my4_sum=0;
749 const int stride= c->stride;
750 uint8_t *mv_penalty= c->current_mv_penalty;
754 for(block=0; block<4; block++){
756 int pred_x4, pred_y4;
758 static const int off[4]= {2, 1, 1, -1};
759 const int mot_stride = s->b8_stride;
760 const int mot_xy = s->block_index[block];
762 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
763 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
765 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
767 /* special case for first line */
768 if (s->first_slice_line && block<2) {
769 c->pred_x= pred_x4= P_LEFT[0];
770 c->pred_y= pred_y4= P_LEFT[1];
772 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
773 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
774 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
775 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
776 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
777 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
778 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
779 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
781 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
782 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
784 c->pred_x= pred_x4 = P_MEDIAN[0];
785 c->pred_y= pred_y4 = P_MEDIAN[1];
790 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
792 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
794 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
796 const int offset= ((block&1) + (block>>1)*stride)*8;
797 uint8_t *dest_y = c->scratchpad + offset;
798 if(s->quarter_sample){
799 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
800 dxy = ((my4 & 3) << 2) | (mx4 & 3);
803 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
805 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
807 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
808 dxy = ((my4 & 1) << 1) | (mx4 & 1);
811 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
813 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
815 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
819 if(s->quarter_sample){
827 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
828 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
830 if(mx4 != mx || my4 != my) same=0;
836 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
837 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);
840 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
845 mx= ff_h263_round_chroma(mx4_sum);
846 my= ff_h263_round_chroma(my4_sum);
847 dxy = ((my & 1) << 1) | (mx & 1);
849 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
852 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
853 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
855 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
856 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
859 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);
860 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);
866 switch(c->avctx->mb_cmp&0xFF){
868 return dmin_sum+ 32*s->qscale*s->qscale;*/
872 return dmin_sum+ 11*c->mb_penalty_factor;
876 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
877 MotionEstContext * const c= &s->me;
879 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
880 c->src[1][0] = c->src[0][0] + s->linesize;
881 if(c->flags & FLAG_CHROMA){
882 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
883 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
884 c->src[1][1] = c->src[0][1] + s->uvlinesize;
885 c->src[1][2] = c->src[0][2] + s->uvlinesize;
889 static int interlaced_search(MpegEncContext *s, int ref_index,
890 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
892 MotionEstContext * const c= &s->me;
897 uint8_t * const mv_penalty= c->current_mv_penalty;
899 const int stride= 2*s->linesize;
901 const int mot_stride= s->mb_stride;
902 const int xy= s->mb_x + s->mb_y*mot_stride;
908 init_interlaced_ref(s, ref_index);
910 for(block=0; block<2; block++){
912 int best_dmin= INT_MAX;
915 for(field_select=0; field_select<2; field_select++){
916 int dmin, mx_i, my_i;
917 int16_t (*mv_table)[2]= mv_tables[block][field_select];
919 if(user_field_select){
920 assert(field_select==0 || field_select==1);
921 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
922 if(field_select_tables[block][xy] != field_select)
926 P_LEFT[0] = mv_table[xy - 1][0];
927 P_LEFT[1] = mv_table[xy - 1][1];
928 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
930 c->pred_x= P_LEFT[0];
931 c->pred_y= P_LEFT[1];
933 if(!s->first_slice_line){
934 P_TOP[0] = mv_table[xy - mot_stride][0];
935 P_TOP[1] = mv_table[xy - mot_stride][1];
936 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
937 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
938 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
939 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
940 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
941 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
943 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
944 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
946 P_MV1[0]= mx; //FIXME not correct if block != field_select
949 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
951 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
953 mv_table[xy][0]= mx_i;
954 mv_table[xy][1]= my_i;
956 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
960 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
961 dxy = ((my_i & 1) << 1) | (mx_i & 1);
964 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
966 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
968 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
969 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
971 dmin+= c->mb_penalty_factor; //field_select bits
973 dmin += field_select != block; //slightly prefer same field
975 if(dmin < best_dmin){
977 best_field= field_select;
981 int16_t (*mv_table)[2]= mv_tables[block][best_field];
983 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
984 if(mv_table[xy][1]&1) same=0;
985 if(mv_table[xy][1]*2 != my) same=0;
986 if(best_field != block) same=0;
989 field_select_tables[block][xy]= best_field;
990 dmin_sum += best_dmin;
1001 switch(c->avctx->mb_cmp&0xFF){
1003 return dmin_sum+ 32*s->qscale*s->qscale;*/
1007 return dmin_sum+ 11*c->mb_penalty_factor;
1011 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1012 int ymax= s->me.ymax>>interlaced;
1013 int ymin= s->me.ymin>>interlaced;
1015 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1016 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1017 if(mv[1] < ymin) mv[1] = ymin;
1018 if(mv[1] > ymax) mv[1] = ymax;
1021 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1022 MotionEstContext * const c= &s->me;
1023 Picture *p= s->current_picture_ptr;
1024 int mb_xy= mb_x + mb_y*s->mb_stride;
1025 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1026 int mb_type= s->current_picture.mb_type[mb_xy];
1027 int flags= c->flags;
1028 int shift= (flags&FLAG_QPEL) + 1;
1029 int mask= (1<<shift)-1;
1032 me_cmp_func cmpf= s->dsp.sse[0];
1033 me_cmp_func chroma_cmpf= s->dsp.sse[1];
1035 if(p_type && USES_LIST(mb_type, 1)){
1036 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1039 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1042 int xy= s->block_index[i];
1043 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1044 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1047 if(IS_INTERLACED(mb_type)){
1048 int xy2= xy + s->b8_stride;
1049 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1053 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1054 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1058 if(USES_LIST(mb_type, 0)){
1059 int field_select0= p->ref_index[0][xy ];
1060 int field_select1= p->ref_index[0][xy2];
1061 assert(field_select0==0 ||field_select0==1);
1062 assert(field_select1==0 ||field_select1==1);
1063 init_interlaced_ref(s, 0);
1066 s->p_field_select_table[0][mb_xy]= field_select0;
1067 s->p_field_select_table[1][mb_xy]= field_select1;
1068 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1069 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1070 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1072 s->b_field_select_table[0][0][mb_xy]= field_select0;
1073 s->b_field_select_table[0][1][mb_xy]= field_select1;
1074 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1075 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1076 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1079 x= p->motion_val[0][xy ][0];
1080 y= p->motion_val[0][xy ][1];
1081 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1082 x= p->motion_val[0][xy2][0];
1083 y= p->motion_val[0][xy2][1];
1084 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1086 if(USES_LIST(mb_type, 1)){
1087 int field_select0= p->ref_index[1][xy ];
1088 int field_select1= p->ref_index[1][xy2];
1089 assert(field_select0==0 ||field_select0==1);
1090 assert(field_select1==0 ||field_select1==1);
1091 init_interlaced_ref(s, 2);
1093 s->b_field_select_table[1][0][mb_xy]= field_select0;
1094 s->b_field_select_table[1][1][mb_xy]= field_select1;
1095 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1096 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1097 if(USES_LIST(mb_type, 0)){
1098 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1100 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1103 x= p->motion_val[1][xy ][0];
1104 y= p->motion_val[1][xy ][1];
1105 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1106 x= p->motion_val[1][xy2][0];
1107 y= p->motion_val[1][xy2][1];
1108 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1109 //FIXME bidir scores
1113 }else if(IS_8X8(mb_type)){
1114 if(!(s->flags & CODEC_FLAG_4MV)){
1115 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1118 cmpf= s->dsp.sse[1];
1119 chroma_cmpf= s->dsp.sse[1];
1122 xy= s->block_index[i];
1123 x= p->motion_val[0][xy][0];
1124 y= p->motion_val[0][xy][1];
1125 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1127 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1129 if(USES_LIST(mb_type, 0)){
1131 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1132 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1133 }else if(USES_LIST(mb_type, 1)){
1134 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1135 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1136 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1138 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1139 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1141 x= p->motion_val[0][xy][0];
1142 y= p->motion_val[0][xy][1];
1143 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1144 }else if(USES_LIST(mb_type, 1)){
1145 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1146 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1148 x= p->motion_val[1][xy][0];
1149 y= p->motion_val[1][xy][1];
1150 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1152 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1157 void ff_estimate_p_frame_motion(MpegEncContext * s,
1160 MotionEstContext * const c= &s->me;
1161 uint8_t *pix, *ppix;
1162 int sum, mx, my, dmin;
1163 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1164 int vard; ///< sum of squared differences with the estimated motion vector
1166 const int shift= 1+s->quarter_sample;
1168 Picture * const pic= &s->current_picture;
1170 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1172 assert(s->quarter_sample==0 || s->quarter_sample==1);
1173 assert(s->linesize == c->stride);
1174 assert(s->uvlinesize == c->uvstride);
1176 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1177 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1178 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1179 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1181 get_limits(s, 16*mb_x, 16*mb_y);
1184 /* intra / predictive decision */
1186 sum = s->dsp.pix_sum(pix, s->linesize);
1187 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1189 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1190 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1191 c->mb_var_sum_temp += (varc+128)>>8;
1193 if(c->avctx->me_threshold){
1194 vard= check_input_motion(s, mb_x, mb_y, 1);
1196 if((vard+128)>>8 < c->avctx->me_threshold){
1197 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1198 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1199 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1200 c->mc_mb_var_sum_temp += (vard+128)>>8;
1201 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1204 if((vard+128)>>8 < c->avctx->mb_threshold)
1205 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1208 switch(s->me_method) {
1211 no_motion_search(s, &mx, &my);
1218 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1223 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1228 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1236 const int mot_stride = s->b8_stride;
1237 const int mot_xy = s->block_index[0];
1239 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1240 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
1242 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1244 if(!s->first_slice_line) {
1245 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1246 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1247 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1248 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1249 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1250 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1251 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1253 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1254 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1256 if(s->out_format == FMT_H263){
1257 c->pred_x = P_MEDIAN[0];
1258 c->pred_y = P_MEDIAN[1];
1259 }else { /* mpeg1 at least */
1260 c->pred_x= P_LEFT[0];
1261 c->pred_y= P_LEFT[1];
1264 c->pred_x= P_LEFT[0];
1265 c->pred_y= P_LEFT[1];
1269 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1274 /* At this point (mx,my) are full-pell and the relative displacement */
1275 ppix = c->ref[0][0] + (my * s->linesize) + mx;
1277 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1279 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1280 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1281 c->mc_mb_var_sum_temp += (vard+128)>>8;
1284 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1285 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1288 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1289 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1290 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1292 if(mb_type == CANDIDATE_MB_TYPE_INTER){
1293 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1294 set_p_mv_tables(s, mx, my, 1);
1299 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1300 h263_mv4_search(s, mx, my, shift);
1302 set_p_mv_tables(s, mx, my, 0);
1304 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1305 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1307 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1308 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1309 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1310 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1312 if (vard*2 + 200*256 > varc)
1313 mb_type|= CANDIDATE_MB_TYPE_INTRA;
1314 if (varc*2 + 200*256 > vard || s->qscale > 24){
1315 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1316 mb_type|= CANDIDATE_MB_TYPE_INTER;
1317 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1318 if(s->flags&CODEC_FLAG_MV0)
1320 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1325 if((s->flags&CODEC_FLAG_4MV)
1326 && !c->skip && varc>50<<8 && vard>10<<8){
1327 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1328 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1330 set_p_mv_tables(s, mx, my, 0);
1332 set_p_mv_tables(s, mx, my, 1);
1333 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1334 && !c->skip){ //FIXME varc/d checks
1335 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1336 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1340 mb_type= CANDIDATE_MB_TYPE_INTER;
1342 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1343 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1344 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1346 if((s->flags&CODEC_FLAG_4MV)
1347 && !c->skip && varc>50<<8 && vard>10<<8){
1348 int dmin4= h263_mv4_search(s, mx, my, shift);
1350 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1354 if((s->flags&CODEC_FLAG_INTERLACED_ME)
1355 && !c->skip){ //FIXME varc/d checks
1356 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1358 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1363 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1364 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1366 /* get intra luma score */
1367 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1368 intra_score= varc - 500;
1370 int mean= (sum+128)>>8;
1373 for(i=0; i<16; i++){
1374 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1375 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1376 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1377 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1380 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1383 /* get chroma score */
1384 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1389 if(s->out_format == FMT_H263){
1390 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1392 mean= (s->last_dc[i] + 4)>>3;
1394 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
1398 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1399 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1402 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1406 intra_score += c->mb_penalty_factor*16;
1408 if(intra_score < dmin){
1409 mb_type= CANDIDATE_MB_TYPE_INTRA;
1410 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1412 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1415 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1416 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1417 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1421 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1424 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1427 MotionEstContext * const c= &s->me;
1430 const int shift= 1+s->quarter_sample;
1431 const int xy= mb_x + mb_y*s->mb_stride;
1432 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1434 assert(s->quarter_sample==0 || s->quarter_sample==1);
1436 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1437 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1439 get_limits(s, 16*mb_x, 16*mb_y);
1442 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1443 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1445 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1447 /* special case for first line */
1448 if (s->first_slice_line) {
1449 c->pred_x= P_LEFT[0];
1450 c->pred_y= P_LEFT[1];
1451 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1452 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1454 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1455 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1456 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1457 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1458 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1459 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1460 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1462 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1463 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1465 c->pred_x = P_MEDIAN[0];
1466 c->pred_y = P_MEDIAN[1];
1469 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1471 s->p_mv_table[xy][0] = mx<<shift;
1472 s->p_mv_table[xy][1] = my<<shift;
1477 static int ff_estimate_motion_b(MpegEncContext * s,
1478 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1480 MotionEstContext * const c= &s->me;
1483 const int shift= 1+s->quarter_sample;
1484 const int mot_stride = s->mb_stride;
1485 const int mot_xy = mb_y*mot_stride + mb_x;
1486 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1489 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1490 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1491 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1492 c->current_mv_penalty= mv_penalty;
1494 get_limits(s, 16*mb_x, 16*mb_y);
1496 switch(s->me_method) {
1499 no_motion_search(s, &mx, &my);
1506 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1511 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1516 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1524 P_LEFT[0] = mv_table[mot_xy - 1][0];
1525 P_LEFT[1] = mv_table[mot_xy - 1][1];
1527 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
1529 /* special case for first line */
1530 if (!s->first_slice_line) {
1531 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1532 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1533 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1534 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1535 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1536 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1537 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1539 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1540 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1542 c->pred_x= P_LEFT[0];
1543 c->pred_y= P_LEFT[1];
1546 if(mv_table == s->b_forw_mv_table){
1547 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1549 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1552 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1557 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1559 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1560 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1562 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1563 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1564 mv_table[mot_xy][0]= mx;
1565 mv_table[mot_xy][1]= my;
1570 static inline int check_bidir_mv(MpegEncContext * s,
1571 int motion_fx, int motion_fy,
1572 int motion_bx, int motion_by,
1573 int pred_fx, int pred_fy,
1574 int pred_bx, int pred_by,
1578 //FIXME better f_code prediction (max mv & distance)
1580 MotionEstContext * const c= &s->me;
1581 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1582 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1583 int stride= c->stride;
1584 uint8_t *dest_y = c->scratchpad;
1589 uint8_t **src_data= c->src[0];
1590 uint8_t **ref_data= c->ref[0];
1591 uint8_t **ref2_data= c->ref[2];
1593 if(s->quarter_sample){
1594 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1595 src_x = motion_fx >> 2;
1596 src_y = motion_fy >> 2;
1598 ptr = ref_data[0] + (src_y * stride) + src_x;
1599 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1601 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1602 src_x = motion_bx >> 2;
1603 src_y = motion_by >> 2;
1605 ptr = ref2_data[0] + (src_y * stride) + src_x;
1606 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1608 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1609 src_x = motion_fx >> 1;
1610 src_y = motion_fy >> 1;
1612 ptr = ref_data[0] + (src_y * stride) + src_x;
1613 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1615 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1616 src_x = motion_bx >> 1;
1617 src_y = motion_by >> 1;
1619 ptr = ref2_data[0] + (src_y * stride) + src_x;
1620 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1623 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1624 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1625 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1627 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1634 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1635 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1637 MotionEstContext * const c= &s->me;
1638 const int mot_stride = s->mb_stride;
1639 const int xy = mb_y *mot_stride + mb_x;
1641 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1642 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1643 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1644 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1645 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1646 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1647 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1648 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1649 const int flags= c->sub_flags;
1650 const int qpel= flags&FLAG_QPEL;
1651 const int shift= 1+qpel;
1652 const int xmin= c->xmin<<shift;
1653 const int ymin= c->ymin<<shift;
1654 const int xmax= c->xmax<<shift;
1655 const int ymax= c->ymax<<shift;
1656 uint8_t map[8][8][8][8];
1658 memset(map,0,sizeof(map));
1659 #define BIDIR_MAP(fx,fy,bx,by) \
1660 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1661 BIDIR_MAP(0,0,0,0) = 1;
1663 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1664 motion_bx, motion_by,
1669 if(s->avctx->bidir_refine){
1671 #define CHECK_BIDIR(fx,fy,bx,by)\
1672 if( !BIDIR_MAP(fx,fy,bx,by)\
1673 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1674 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1675 BIDIR_MAP(fx,fy,bx,by) = 1;\
1676 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);\
1686 #define CHECK_BIDIR2(a,b,c,d)\
1687 CHECK_BIDIR(a,b,c,d)\
1688 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1690 #define CHECK_BIDIRR(a,b,c,d)\
1691 CHECK_BIDIR2(a,b,c,d)\
1692 CHECK_BIDIR2(b,c,d,a)\
1693 CHECK_BIDIR2(c,d,a,b)\
1694 CHECK_BIDIR2(d,a,b,c)
1699 CHECK_BIDIRR( 0, 0, 0, 1)
1700 if(s->avctx->bidir_refine > 1){
1701 CHECK_BIDIRR( 0, 0, 1, 1)
1702 CHECK_BIDIR2( 0, 1, 0, 1)
1703 CHECK_BIDIR2( 1, 0, 1, 0)
1704 CHECK_BIDIRR( 0, 0,-1, 1)
1705 CHECK_BIDIR2( 0,-1, 0, 1)
1706 CHECK_BIDIR2(-1, 0, 1, 0)
1707 if(s->avctx->bidir_refine > 2){
1708 CHECK_BIDIRR( 0, 1, 1, 1)
1709 CHECK_BIDIRR( 0,-1, 1, 1)
1710 CHECK_BIDIRR( 0, 1,-1, 1)
1711 CHECK_BIDIRR( 0, 1, 1,-1)
1712 if(s->avctx->bidir_refine > 3){
1713 CHECK_BIDIR2( 1, 1, 1, 1)
1714 CHECK_BIDIRR( 1, 1, 1,-1)
1715 CHECK_BIDIR2( 1, 1,-1,-1)
1716 CHECK_BIDIR2( 1,-1,-1, 1)
1717 CHECK_BIDIR2( 1,-1, 1,-1)
1724 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1725 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1726 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1727 s->b_bidir_back_mv_table[xy][1]= motion_by;
1732 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1734 MotionEstContext * const c= &s->me;
1736 const int mot_stride = s->mb_stride;
1737 const int mot_xy = mb_y*mot_stride + mb_x;
1738 const int shift= 1+s->quarter_sample;
1740 const int time_pp= s->pp_time;
1741 const int time_pb= s->pb_time;
1742 int mx, my, xmin, xmax, ymin, ymax;
1743 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1745 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1746 ymin= xmin=(-32)>>shift;
1747 ymax= xmax= 31>>shift;
1749 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1750 s->mv_type= MV_TYPE_8X8;
1752 s->mv_type= MV_TYPE_16X16;
1756 int index= s->block_index[i];
1759 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1760 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1761 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1762 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1763 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1764 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1766 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1767 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1768 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1770 xmax= FFMIN(xmax, s->width - max);
1771 xmin= FFMAX(xmin, - 16 - min);
1773 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1774 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1775 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1777 ymax= FFMIN(ymax, s->height - max);
1778 ymin= FFMAX(ymin, - 16 - min);
1780 if(s->mv_type == MV_TYPE_16X16) break;
1783 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1785 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1786 s->b_direct_mv_table[mot_xy][0]= 0;
1787 s->b_direct_mv_table[mot_xy][1]= 0;
1789 return 256*256*256*64;
1796 c->flags |= FLAG_DIRECT;
1797 c->sub_flags |= FLAG_DIRECT;
1801 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1802 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1804 /* special case for first line */
1805 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its clipped
1806 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1807 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1808 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1809 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1811 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1812 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1815 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1816 if(c->sub_flags&FLAG_QPEL)
1817 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1819 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1821 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1822 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1824 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1826 mv_table[mot_xy][0]= mx;
1827 mv_table[mot_xy][1]= my;
1828 c->flags &= ~FLAG_DIRECT;
1829 c->sub_flags &= ~FLAG_DIRECT;
1834 void ff_estimate_b_frame_motion(MpegEncContext * s,
1837 MotionEstContext * const c= &s->me;
1838 const int penalty_factor= c->mb_penalty_factor;
1839 int fmin, bmin, dmin, fbmin, bimin, fimin;
1841 const int xy = mb_y*s->mb_stride + mb_x;
1842 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1844 get_limits(s, 16*mb_x, 16*mb_y);
1848 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1849 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1851 score= ((unsigned)(score*score + 128*256))>>16;
1852 c->mc_mb_var_sum_temp += score;
1853 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1854 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1859 if(c->avctx->me_threshold){
1860 int vard= check_input_motion(s, mb_x, mb_y, 0);
1862 if((vard+128)>>8 < c->avctx->me_threshold){
1863 // pix = c->src[0][0];
1864 // sum = s->dsp.pix_sum(pix, s->linesize);
1865 // varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1867 // pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1868 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1869 /* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1870 c->mb_var_sum_temp += (varc+128)>>8;*/
1871 c->mc_mb_var_sum_temp += (vard+128)>>8;
1872 /* if (vard <= 64<<8 || vard < varc) {
1873 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1875 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1879 if((vard+128)>>8 < c->avctx->mb_threshold){
1880 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1881 if(type == CANDIDATE_MB_TYPE_DIRECT){
1882 direct_search(s, mb_x, mb_y);
1884 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1886 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1888 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1890 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1892 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1894 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1895 interlaced_search(s, 0,
1896 s->b_field_mv_table[0], s->b_field_select_table[0],
1897 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1899 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1901 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1902 interlaced_search(s, 2,
1903 s->b_field_mv_table[1], s->b_field_select_table[1],
1904 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1910 if (s->codec_id == CODEC_ID_MPEG4)
1911 dmin= direct_search(s, mb_x, mb_y);
1914 //FIXME penalty stuff for non mpeg4
1916 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1919 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1920 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1923 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1924 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1926 if(s->flags & CODEC_FLAG_INTERLACED_ME){
1927 //FIXME mb type penalty
1929 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1930 fimin= interlaced_search(s, 0,
1931 s->b_field_mv_table[0], s->b_field_select_table[0],
1932 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1933 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1934 bimin= interlaced_search(s, 2,
1935 s->b_field_mv_table[1], s->b_field_select_table[1],
1936 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1938 fimin= bimin= INT_MAX;
1942 type = CANDIDATE_MB_TYPE_FORWARD;
1946 type = CANDIDATE_MB_TYPE_DIRECT;
1950 type= CANDIDATE_MB_TYPE_BACKWARD;
1954 type= CANDIDATE_MB_TYPE_BIDIR;
1958 type= CANDIDATE_MB_TYPE_FORWARD_I;
1962 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1965 score= ((unsigned)(score*score + 128*256))>>16;
1966 c->mc_mb_var_sum_temp += score;
1967 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1970 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1971 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1973 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1975 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1976 if(fimin < INT_MAX && bimin < INT_MAX){
1977 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1979 //FIXME something smarter
1980 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1981 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])
1982 type |= CANDIDATE_MB_TYPE_DIRECT0;
1984 if(s->out_format == FMT_MPEG1)
1985 type |= CANDIDATE_MB_TYPE_INTRA;
1989 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1992 /* find best f_code for ME which do unlimited searches */
1993 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1995 if(s->me_method>=ME_EPZS){
1997 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1998 uint8_t * fcode_tab= s->fcode_tab;
2000 int best_score=-10000000;
2002 if(s->msmpeg4_version)
2003 range= FFMIN(range, 16);
2004 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2005 range= FFMIN(range, 256);
2007 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
2009 for(y=0; y<s->mb_height; y++){
2011 int xy= y*s->mb_stride;
2012 for(x=0; x<s->mb_width; x++){
2013 if(s->mb_type[xy] & type){
2014 int mx= mv_table[xy][0];
2015 int my= mv_table[xy][1];
2016 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2017 fcode_tab[my + MAX_MV]);
2020 if(mx >= range || mx < -range ||
2021 my >= range || my < -range)
2024 for(j=0; j<fcode && j<8; j++){
2025 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2034 if(score[i] > best_score){
2035 best_score= score[i];
2038 // printf("%d %d\n", i, score[i]);
2041 // printf("fcode: %d type: %d\n", i, s->pict_type);
2043 /* for(i=0; i<=MAX_FCODE; i++){
2044 printf("%d ", mv_num[i]);
2052 void ff_fix_long_p_mvs(MpegEncContext * s)
2054 MotionEstContext * const c= &s->me;
2055 const int f_code= s->f_code;
2057 assert(s->pict_type==P_TYPE);
2059 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2061 assert(range <= 16 || !s->msmpeg4_version);
2062 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2064 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2066 //printf("%d no:%d %d//\n", clip, noclip, f_code);
2067 if(s->flags&CODEC_FLAG_4MV){
2068 const int wrap= s->b8_stride;
2070 /* clip / convert to intra 8x8 type MVs */
2071 for(y=0; y<s->mb_height; y++){
2073 int i= y*s->mb_stride;
2076 for(x=0; x<s->mb_width; x++){
2077 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2079 for(block=0; block<4; block++){
2080 int off= (block& 1) + (block>>1)*wrap;
2081 int mx= s->current_picture.motion_val[0][ xy + off ][0];
2082 int my= s->current_picture.motion_val[0][ xy + off ][1];
2084 if( mx >=range || mx <-range
2085 || my >=range || my <-range){
2086 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2087 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2088 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2101 * @param truncate 1 for truncation, 0 for using intra
2103 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2104 int16_t (*mv_table)[2], int f_code, int type, int truncate)
2106 MotionEstContext * const c= &s->me;
2107 int y, h_range, v_range;
2109 // RAL: 8 in MPEG-1, 16 in MPEG-4
2110 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2112 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2115 v_range= field_select_table ? range>>1 : range;
2117 /* clip / convert to intra 16x16 type MVs */
2118 for(y=0; y<s->mb_height; y++){
2120 int xy= y*s->mb_stride;
2121 for(x=0; x<s->mb_width; x++){
2122 if (s->mb_type[xy] & type){ // RAL: "type" test added...
2123 if(field_select_table==NULL || field_select_table[xy] == field_select){
2124 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2125 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2128 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2129 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2130 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2131 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2133 s->mb_type[xy] &= ~type;
2134 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;