3 * Copyright (c) 2000,2001 Fabrice Bellard.
4 * Copyright (c) 2002-2003 Michael Niedermayer
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2 of the License, or (at your option) any later version.
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
33 #include "mpegvideo.h"
38 #define SQ(a) ((a)*(a))
42 #define P_TOPRIGHT P[3]
46 static inline int sad_hpel_motion_search(MpegEncContext * s,
47 int *mx_ptr, int *my_ptr, int dmin,
48 int xmin, int ymin, int xmax, int ymax,
49 int pred_x, int pred_y, Picture *picture,
50 int n, int size, uint8_t * const mv_penalty);
52 static inline int update_map_generation(MpegEncContext * s)
54 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
55 if(s->me.map_generation==0){
56 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
57 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59 return s->me.map_generation;
62 /* shape adaptive search stuff */
63 typedef struct Minima{
69 static int minima_cmp(const void *a, const void *b){
70 const Minima *da = (const Minima *) a;
71 const Minima *db = (const Minima *) b;
73 return da->height - db->height;
77 #define RENAME(a) simple_ ## a
79 #define CMP(d, x, y, size)\
80 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);
82 #define CMP_HPEL(d, dx, dy, x, y, size)\
84 const int dxy= (dx) + 2*(dy);\
85 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
86 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
89 #define CMP_QPEL(d, dx, dy, x, y, size)\
91 const int dxy= (dx) + 4*(dy);\
92 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
93 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
96 #include "motion_est_template.c"
104 #define RENAME(a) simple_chroma_ ## a
106 #define CMP(d, x, y, size)\
107 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\
109 int dxy= ((x)&1) + 2*((y)&1);\
110 int c= ((x)>>1) + ((y)>>1)*uvstride;\
112 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
113 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\
114 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
115 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\
118 #define CMP_HPEL(d, dx, dy, x, y, size)\
120 const int dxy= (dx) + 2*(dy);\
121 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
122 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
124 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
125 int c= ((x)>>1) + ((y)>>1)*uvstride;\
126 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
127 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
128 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
129 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
133 #define CMP_QPEL(d, dx, dy, x, y, size)\
135 const int dxy= (dx) + 4*(dy);\
136 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
137 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
140 int cx= (4*(x) + (dx))/2;\
141 int cy= (4*(y) + (dy))/2;\
144 cxy= (cx&1) + 2*(cy&1);\
145 c= ((cx)>>1) + ((cy)>>1)*uvstride;\
146 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
147 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
148 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
149 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
153 #include "motion_est_template.c"
160 /* SIMPLE DIRECT HPEL */
161 #define RENAME(a) simple_direct_hpel_ ## a
162 //FIXME precalc divisions stuff
164 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
165 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
166 const int hx= 2*(x) + (dx);\
167 const int hy= 2*(y) + (dy);\
168 if(s->mv_type==MV_TYPE_8X8){\
171 int fx = s->me.direct_basis_mv[i][0] + hx;\
172 int fy = s->me.direct_basis_mv[i][1] + hy;\
173 int bx = hx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
174 int by = hy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
175 int fxy= (fx&1) + 2*(fy&1);\
176 int bxy= (bx&1) + 2*(by&1);\
178 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
179 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
180 hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\
183 int fx = s->me.direct_basis_mv[0][0] + hx;\
184 int fy = s->me.direct_basis_mv[0][1] + hy;\
185 int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\
186 int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\
187 int fxy= (fx&1) + 2*(fy&1);\
188 int bxy= (bx&1) + 2*(by&1);\
190 assert((fx>>1) + 16*s->mb_x >= -16);\
191 assert((fy>>1) + 16*s->mb_y >= -16);\
192 assert((fx>>1) + 16*s->mb_x <= s->width);\
193 assert((fy>>1) + 16*s->mb_y <= s->height);\
194 assert((bx>>1) + 16*s->mb_x >= -16);\
195 assert((by>>1) + 16*s->mb_y >= -16);\
196 assert((bx>>1) + 16*s->mb_x <= s->width);\
197 assert((by>>1) + 16*s->mb_y <= s->height);\
199 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
200 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
202 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
207 #define CMP_HPEL(d, dx, dy, x, y, size)\
208 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
210 #define CMP(d, x, y, size)\
211 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
213 #include "motion_est_template.c"
221 /* SIMPLE DIRECT QPEL */
222 #define RENAME(a) simple_direct_qpel_ ## a
224 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
225 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
226 const int qx= 4*(x) + (dx);\
227 const int qy= 4*(y) + (dy);\
228 if(s->mv_type==MV_TYPE_8X8){\
231 int fx = s->me.direct_basis_mv[i][0] + qx;\
232 int fy = s->me.direct_basis_mv[i][1] + qy;\
233 int bx = qx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
234 int by = qy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
235 int fxy= (fx&3) + 4*(fy&3);\
236 int bxy= (bx&3) + 4*(by&3);\
238 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
239 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
240 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
243 int fx = s->me.direct_basis_mv[0][0] + qx;\
244 int fy = s->me.direct_basis_mv[0][1] + qy;\
245 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
246 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
247 int fxy= (fx&3) + 4*(fy&3);\
248 int bxy= (bx&3) + 4*(by&3);\
250 qpel_put[1][fxy](s->me.scratchpad , (ref_y ) + (fx>>2) + (fy>>2)*(stride) , stride);\
251 qpel_put[1][fxy](s->me.scratchpad + 8 , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 , stride);\
252 qpel_put[1][fxy](s->me.scratchpad + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8*stride, stride);\
253 qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
254 qpel_avg[1][bxy](s->me.scratchpad , (ref2_y) + (bx>>2) + (by>>2)*(stride) , stride);\
255 qpel_avg[1][bxy](s->me.scratchpad + 8 , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 , stride);\
256 qpel_avg[1][bxy](s->me.scratchpad + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8*stride, stride);\
257 qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
259 d = cmp_func(s, s->me.scratchpad, src_y, stride);\
264 #define CMP_QPEL(d, dx, dy, x, y, size)\
265 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
267 #define CMP(d, x, y, size)\
268 CMP_DIRECT(d, 0, 0, x, y, size, cmp)
270 #include "motion_est_template.c"
279 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
283 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
284 DSPContext* c= &s->dsp;
287 memset(cmp, 0, sizeof(void*)*11);
295 cmp[0]= c->hadamard8_diff[0];
296 cmp[1]= c->hadamard8_diff[1];
303 cmp[0]= c->dct_sad[0];
304 cmp[1]= c->dct_sad[1];
307 cmp[0]= c->quant_psnr[0];
308 cmp[1]= c->quant_psnr[1];
324 fprintf(stderr,"internal error in cmp function selection\n");
328 static inline int get_penalty_factor(MpegEncContext *s, int type){
338 return s->qscale*s->qscale*2;
343 return (s->qscale*s->qscale*185 + 64)>>7;
347 void ff_init_me(MpegEncContext *s){
348 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
349 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
350 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
351 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
353 if(s->flags&CODEC_FLAG_QPEL){
354 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
355 s->me.sub_motion_search= simple_chroma_qpel_motion_search;
357 s->me.sub_motion_search= simple_qpel_motion_search;
359 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
360 s->me.sub_motion_search= simple_chroma_hpel_motion_search;
361 else if( s->avctx->me_sub_cmp == FF_CMP_SAD
362 && s->avctx-> me_cmp == FF_CMP_SAD
363 && s->avctx-> mb_cmp == FF_CMP_SAD)
364 s->me.sub_motion_search= sad_hpel_motion_search;
366 s->me.sub_motion_search= simple_hpel_motion_search;
369 if(s->avctx->me_cmp&FF_CMP_CHROMA){
370 s->me.motion_search[0]= simple_chroma_epzs_motion_search;
371 s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
373 s->me.motion_search[0]= simple_epzs_motion_search;
374 s->me.motion_search[1]= simple_epzs_motion_search4;
377 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
378 s->me.pre_motion_search= simple_chroma_epzs_motion_search;
380 s->me.pre_motion_search= simple_epzs_motion_search;
383 if(s->flags&CODEC_FLAG_QPEL){
384 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
385 s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
387 s->me.get_mb_score= simple_qpel_get_mb_score;
389 if(s->avctx->mb_cmp&FF_CMP_CHROMA)
390 s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
392 s->me.get_mb_score= simple_hpel_get_mb_score;
397 static int pix_dev(uint8_t * pix, int line_size, int mean)
402 for (i = 0; i < 16; i++) {
403 for (j = 0; j < 16; j += 8) {
404 s += ABS(pix[0]-mean);
405 s += ABS(pix[1]-mean);
406 s += ABS(pix[2]-mean);
407 s += ABS(pix[3]-mean);
408 s += ABS(pix[4]-mean);
409 s += ABS(pix[5]-mean);
410 s += ABS(pix[6]-mean);
411 s += ABS(pix[7]-mean);
414 pix += line_size - 16;
420 static inline void no_motion_search(MpegEncContext * s,
421 int *mx_ptr, int *my_ptr)
423 *mx_ptr = 16 * s->mb_x;
424 *my_ptr = 16 * s->mb_y;
427 static int full_motion_search(MpegEncContext * s,
428 int *mx_ptr, int *my_ptr, int range,
429 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
431 int x1, y1, x2, y2, xx, yy, x, y;
437 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
449 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
453 for (y = y1; y <= y2; y++) {
454 for (x = x1; x <= x2; x++) {
455 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
459 (abs(x - xx) + abs(y - yy)) <
460 (abs(mx - xx) + abs(my - yy)))) {
472 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
473 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
474 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
481 static int log_motion_search(MpegEncContext * s,
482 int *mx_ptr, int *my_ptr, int range,
483 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
485 int x1, y1, x2, y2, xx, yy, x, y;
512 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
518 for (y = y1; y <= y2; y += range) {
519 for (x = x1; x <= x2; x += range) {
520 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
521 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
547 } while (range >= 1);
550 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
557 static int phods_motion_search(MpegEncContext * s,
558 int *mx_ptr, int *my_ptr, int range,
559 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
561 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
562 int mx, my, dminx, dminy;
588 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
599 for (x = x1; x <= x2; x += range) {
600 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
601 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
608 for (y = y1; y <= y2; y += range) {
609 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
610 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
636 } while (range >= 1);
639 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
642 /* half pixel search */
649 #define Z_THRESHOLD 256
651 #define CHECK_SAD_HALF_MV(suffix, x, y) \
653 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
654 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
655 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
658 static inline int sad_hpel_motion_search(MpegEncContext * s,
659 int *mx_ptr, int *my_ptr, int dmin,
660 int xmin, int ymin, int xmax, int ymax,
661 int pred_x, int pred_y, Picture *picture,
662 int n, int size, uint8_t * const mv_penalty)
664 uint8_t *ref_picture= picture->data[0];
665 uint32_t *score_map= s->me.score_map;
666 const int penalty_factor= s->me.sub_penalty_factor;
667 int mx, my, xx, yy, dminh;
669 op_pixels_abs_func pix_abs_x2;
670 op_pixels_abs_func pix_abs_y2;
671 op_pixels_abs_func pix_abs_xy2;
674 pix_abs_x2 = s->dsp.pix_abs16x16_x2;
675 pix_abs_y2 = s->dsp.pix_abs16x16_y2;
676 pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
678 pix_abs_x2 = s->dsp.pix_abs8x8_x2;
679 pix_abs_y2 = s->dsp.pix_abs8x8_y2;
680 pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
691 xx = 16 * s->mb_x + 8*(n&1);
692 yy = 16 * s->mb_y + 8*(n>>1);
693 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
697 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
701 if (mx > xmin && mx < xmax &&
702 my > ymin && my < ymax) {
705 const int index= (my<<ME_MAP_SHIFT) + mx;
706 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
707 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
708 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
709 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
719 CHECK_SAD_HALF_MV(y2 , 0, -1)
721 CHECK_SAD_HALF_MV(xy2, -1, -1)
723 CHECK_SAD_HALF_MV(xy2, +1, -1)
727 CHECK_SAD_HALF_MV(xy2, -1, +1)
729 CHECK_SAD_HALF_MV(x2 , -1, 0)
731 CHECK_SAD_HALF_MV(xy2, +1, -1)
733 CHECK_SAD_HALF_MV(xy2, -1, -1)
737 CHECK_SAD_HALF_MV(xy2, +1, +1)
739 CHECK_SAD_HALF_MV(x2 , +1, 0)
744 CHECK_SAD_HALF_MV(xy2, -1, -1)
748 CHECK_SAD_HALF_MV(xy2, +1, +1)
750 CHECK_SAD_HALF_MV(x2 , -1, 0)
751 CHECK_SAD_HALF_MV(xy2, -1, +1)
754 CHECK_SAD_HALF_MV(xy2, +1, -1)
758 CHECK_SAD_HALF_MV(xy2, -1, +1)
760 CHECK_SAD_HALF_MV(x2 , +1, 0)
761 CHECK_SAD_HALF_MV(xy2, +1, +1)
763 CHECK_SAD_HALF_MV(y2 , 0, +1)
778 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
780 const int xy= s->mb_x + s->mb_y*s->mb_stride;
782 s->p_mv_table[xy][0] = mx;
783 s->p_mv_table[xy][1] = my;
785 /* has allready been set to the 4 MV if 4MV is done */
787 int mot_xy= s->block_index[0];
789 s->motion_val[mot_xy ][0]= mx;
790 s->motion_val[mot_xy ][1]= my;
791 s->motion_val[mot_xy+1][0]= mx;
792 s->motion_val[mot_xy+1][1]= my;
794 mot_xy += s->block_wrap[0];
795 s->motion_val[mot_xy ][0]= mx;
796 s->motion_val[mot_xy ][1]= my;
797 s->motion_val[mot_xy+1][0]= mx;
798 s->motion_val[mot_xy+1][1]= my;
803 * get fullpel ME search limits.
804 * @param range the approximate search range for the old ME code, unused for EPZS and newer
806 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax)
808 if(s->avctx->me_range) *range= s->avctx->me_range >> 1;
811 if (s->unrestricted_mv) {
814 *xmax = s->mb_width*16;
815 *ymax = s->mb_height*16;
819 *xmax = s->mb_width*16 - 16;
820 *ymax = s->mb_height*16 - 16;
823 //FIXME try to limit x/y min/max if me_range is set
826 static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
830 int dmin_sum=0, mx4_sum=0, my4_sum=0;
831 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
833 for(block=0; block<4; block++){
835 int pred_x4, pred_y4;
837 static const int off[4]= {2, 1, 1, -1};
838 const int mot_stride = s->block_wrap[0];
839 const int mot_xy = s->block_index[block];
840 // const int block_x= (block&1);
841 // const int block_y= (block>>1);
842 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
843 const int rel_xmin4= xmin;
844 const int rel_xmax4= xmax;
845 const int rel_ymin4= ymin;
846 const int rel_ymax4= ymax;
848 const int rel_xmin4= xmin - block_x*8;
849 const int rel_xmax4= xmax - block_x*8 + 8;
850 const int rel_ymin4= ymin - block_y*8;
851 const int rel_ymax4= ymax - block_y*8 + 8;
853 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
854 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
856 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift);
858 /* special case for first line */
859 if (s->mb_y == 0 && block<2) {
863 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
864 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
865 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
866 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
867 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift);
868 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
869 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
870 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
872 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
873 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
875 // if(s->out_format == FMT_H263){
876 pred_x4 = P_MEDIAN[0];
877 pred_y4 = P_MEDIAN[1];
879 }else { /* mpeg1 at least */
888 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
889 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
891 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
892 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
894 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
896 const int offset= ((block&1) + (block>>1)*s->linesize)*8;
897 uint8_t *dest_y = s->me.scratchpad + offset;
899 if(s->quarter_sample){
900 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
901 dxy = ((my4 & 3) << 2) | (mx4 & 3);
904 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , s->linesize);
906 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , s->linesize);
908 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
909 dxy = ((my4 & 1) << 1) | (mx4 & 1);
912 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , s->linesize, 8);
914 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , s->linesize, 8);
916 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
920 if(s->quarter_sample){
928 s->motion_val[ s->block_index[block] ][0]= mx4;
929 s->motion_val[ s->block_index[block] ][1]= my4;
932 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
933 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*s->linesize, s->me.scratchpad, s->linesize);
936 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
941 mx= ff_h263_round_chroma(mx4_sum);
942 my= ff_h263_round_chroma(my4_sum);
943 dxy = ((my & 1) << 1) | (mx & 1);
945 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
948 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
949 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
951 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
952 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
955 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad , s->uvlinesize);
956 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad+8, s->uvlinesize);
959 switch(s->avctx->mb_cmp&0xFF){
961 return dmin_sum+ 32*s->qscale*s->qscale;*/
965 return dmin_sum+ 11*s->me.mb_penalty_factor;
969 void ff_estimate_p_frame_motion(MpegEncContext * s,
973 int sum, varc, vard, mx, my, range, dmin, xx, yy;
974 int xmin, ymin, xmax, ymax;
975 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
976 int pred_x=0, pred_y=0;
978 const int shift= 1+s->quarter_sample;
980 uint8_t *ref_picture= s->last_picture.data[0];
981 Picture * const pic= &s->current_picture;
982 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
984 assert(s->quarter_sample==0 || s->quarter_sample==1);
986 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
987 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
988 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
990 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
991 rel_xmin= xmin - mb_x*16;
992 rel_xmax= xmax - mb_x*16;
993 rel_ymin= ymin - mb_y*16;
994 rel_ymax= ymax - mb_y*16;
997 switch(s->me_method) {
1000 no_motion_search(s, &mx, &my);
1006 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1011 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1016 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1023 const int mot_stride = s->block_wrap[0];
1024 const int mot_xy = s->block_index[0];
1026 P_LEFT[0] = s->motion_val[mot_xy - 1][0];
1027 P_LEFT[1] = s->motion_val[mot_xy - 1][1];
1029 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1032 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0];
1033 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1];
1034 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1035 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1036 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift);
1037 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1038 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1040 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1041 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1043 if(s->out_format == FMT_H263){
1044 pred_x = P_MEDIAN[0];
1045 pred_y = P_MEDIAN[1];
1046 }else { /* mpeg1 at least */
1056 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1057 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1062 /* intra / predictive decision */
1066 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1067 /* At this point (mx,my) are full-pell and the relative displacement */
1068 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1070 sum = s->dsp.pix_sum(pix, s->linesize);
1072 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1073 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1075 //printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
1076 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
1077 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1078 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1079 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1080 pic->mb_var_sum += varc;
1081 pic->mc_mb_var_sum += vard;
1082 //printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
1085 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1086 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1088 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1089 if (vard <= 64 || vard < varc)
1090 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1092 s->scene_change_score+= s->qscale;
1094 if (vard*2 + 200 > varc)
1095 mb_type|= MB_TYPE_INTRA;
1096 if (varc*2 + 200 > vard){
1097 mb_type|= MB_TYPE_INTER;
1098 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1099 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1104 if((s->flags&CODEC_FLAG_4MV)
1105 && !s->me.skip && varc>50 && vard>10){
1106 h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1107 mb_type|=MB_TYPE_INTER4V;
1109 set_p_mv_tables(s, mx, my, 0);
1111 set_p_mv_tables(s, mx, my, 1);
1113 mb_type= MB_TYPE_INTER;
1115 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1116 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1118 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1119 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1121 if((s->flags&CODEC_FLAG_4MV)
1122 && !s->me.skip && varc>50 && vard>10){
1123 int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1125 mb_type= MB_TYPE_INTER4V;
1129 pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1130 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1132 if (vard <= 64 || vard < varc) {
1133 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1135 s->scene_change_score+= s->qscale;
1139 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1142 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1145 int mx, my, range, dmin;
1146 int xmin, ymin, xmax, ymax;
1147 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1148 int pred_x=0, pred_y=0;
1150 const int shift= 1+s->quarter_sample;
1151 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1152 const int xy= mb_x + mb_y*s->mb_stride;
1154 assert(s->quarter_sample==0 || s->quarter_sample==1);
1156 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
1158 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1159 rel_xmin= xmin - mb_x*16;
1160 rel_xmax= xmax - mb_x*16;
1161 rel_ymin= ymin - mb_y*16;
1162 rel_ymax= ymax - mb_y*16;
1165 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1166 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1168 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift);
1170 /* special case for first line */
1171 if (mb_y == s->mb_height-1) {
1174 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1175 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1177 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1178 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1179 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1180 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1181 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
1182 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1183 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1185 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1186 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1188 pred_x = P_MEDIAN[0];
1189 pred_y = P_MEDIAN[1];
1191 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1192 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1194 s->p_mv_table[xy][0] = mx<<shift;
1195 s->p_mv_table[xy][1] = my<<shift;
1200 static int ff_estimate_motion_b(MpegEncContext * s,
1201 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1203 int mx, my, range, dmin;
1204 int xmin, ymin, xmax, ymax;
1205 int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1206 int pred_x=0, pred_y=0;
1208 const int shift= 1+s->quarter_sample;
1209 const int mot_stride = s->mb_stride;
1210 const int mot_xy = mb_y*mot_stride + mb_x;
1211 uint8_t * const ref_picture= picture->data[0];
1212 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1215 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
1216 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1217 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1219 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1220 rel_xmin= xmin - mb_x*16;
1221 rel_xmax= xmax - mb_x*16;
1222 rel_ymin= ymin - mb_y*16;
1223 rel_ymax= ymax - mb_y*16;
1225 switch(s->me_method) {
1228 no_motion_search(s, &mx, &my);
1234 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1239 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1244 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1251 P_LEFT[0] = mv_table[mot_xy - 1][0];
1252 P_LEFT[1] = mv_table[mot_xy - 1][1];
1254 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift);
1256 /* special case for first line */
1258 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1259 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1260 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1261 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
1262 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1263 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1264 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1266 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1267 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1273 if(mv_table == s->b_forw_mv_table){
1274 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1276 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1279 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1280 picture, s->p_mv_table, mv_scale, mv_penalty);
1285 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1286 pred_x, pred_y, picture, 0, 0, mv_penalty);
1288 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1289 dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1291 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1292 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1293 mv_table[mot_xy][0]= mx;
1294 mv_table[mot_xy][1]= my;
1299 static inline int check_bidir_mv(MpegEncContext * s,
1301 int motion_fx, int motion_fy,
1302 int motion_bx, int motion_by,
1303 int pred_fx, int pred_fy,
1304 int pred_bx, int pred_by)
1307 //FIXME move into template?
1308 //FIXME better f_code prediction (max mv & distance)
1309 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1310 uint8_t *dest_y = s->me.scratchpad;
1316 if(s->quarter_sample){
1317 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1318 src_x = mb_x * 16 + (motion_fx >> 2);
1319 src_y = mb_y * 16 + (motion_fy >> 2);
1320 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1321 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1323 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1324 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1326 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1327 src_x = mb_x * 16 + (motion_bx >> 2);
1328 src_y = mb_y * 16 + (motion_by >> 2);
1329 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1330 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1332 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1333 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
1335 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1336 src_x = mb_x * 16 + (motion_fx >> 1);
1337 src_y = mb_y * 16 + (motion_fy >> 1);
1338 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1339 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1341 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1342 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1344 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1345 src_x = mb_x * 16 + (motion_bx >> 1);
1346 src_y = mb_y * 16 + (motion_by >> 1);
1347 assert(src_x >=-16 && src_x<=s->h_edge_pos);
1348 assert(src_y >=-16 && src_y<=s->v_edge_pos);
1350 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1351 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
1354 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1355 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1356 + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1358 if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1365 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1366 static inline int bidir_refine(MpegEncContext * s,
1369 const int mot_stride = s->mb_stride;
1370 const int xy = mb_y *mot_stride + mb_x;
1372 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1373 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1374 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1375 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1376 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1377 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1378 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1379 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1381 //FIXME do refinement and add flag
1383 fbmin= check_bidir_mv(s, mb_x, mb_y,
1384 motion_fx, motion_fy,
1385 motion_bx, motion_by,
1392 static inline int direct_search(MpegEncContext * s,
1396 const int mot_stride = s->mb_stride;
1397 const int mot_xy = mb_y*mot_stride + mb_x;
1398 const int shift= 1+s->quarter_sample;
1400 const int time_pp= s->pp_time;
1401 const int time_pb= s->pb_time;
1402 int mx, my, xmin, xmax, ymin, ymax;
1403 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1404 uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1406 ymin= xmin=(-32)>>shift;
1407 ymax= xmax= 31>>shift;
1409 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1410 s->mv_type= MV_TYPE_8X8;
1412 s->mv_type= MV_TYPE_16X16;
1416 int index= s->block_index[i];
1419 s->me.co_located_mv[i][0]= s->motion_val[index][0];
1420 s->me.co_located_mv[i][1]= s->motion_val[index][1];
1421 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1422 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1423 // s->me.direct_basis_mv[1][i][0]= s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1424 // s->me.direct_basis_mv[1][i][1]= s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1426 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1427 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1428 max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1429 min+= (2*mb_x + (i& 1))*8 - 1;
1430 xmax= FFMIN(xmax, s->width - max);
1431 xmin= FFMAX(xmin, - 16 - min);
1433 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1434 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1435 max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1436 min+= (2*mb_y + (i>>1))*8 - 1;
1437 ymax= FFMIN(ymax, s->height - max);
1438 ymin= FFMAX(ymin, - 16 - min);
1440 if(s->mv_type == MV_TYPE_16X16) break;
1443 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1445 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1446 s->b_direct_mv_table[mot_xy][0]= 0;
1447 s->b_direct_mv_table[mot_xy][1]= 0;
1449 return 256*256*256*64;
1452 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1453 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1455 /* special case for first line */
1457 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1458 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1459 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1460 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<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]);
1466 //FIXME direct_search ptr in context!!! (needed for chroma anyway or this will get messy)
1467 if(s->flags&CODEC_FLAG_QPEL){
1468 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1469 &s->last_picture, mv_table, 1<<14, mv_penalty);
1470 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1471 0, 0, &s->last_picture, 0, 0, mv_penalty);
1473 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1474 dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1476 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax,
1477 &s->last_picture, mv_table, 1<<15, mv_penalty);
1478 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1479 0, 0, &s->last_picture, 0, 0, mv_penalty);
1481 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1482 dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1485 s->b_direct_mv_table[mot_xy][0]= mx;
1486 s->b_direct_mv_table[mot_xy][1]= my;
1490 void ff_estimate_b_frame_motion(MpegEncContext * s,
1493 const int penalty_factor= s->me.mb_penalty_factor;
1494 int fmin, bmin, dmin, fbmin;
1497 dmin= direct_search(s, mb_x, mb_y);
1499 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1500 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1501 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1503 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1504 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1507 type = MB_TYPE_FORWARD;
1509 // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)
1510 if (s->codec_id == CODEC_ID_MPEG4 && dmin <= score){
1512 type = MB_TYPE_DIRECT;
1516 type= MB_TYPE_BACKWARD;
1520 type= MB_TYPE_BIDIR;
1523 score= ((unsigned)(score*score + 128*256))>>16;
1524 s->current_picture.mc_mb_var_sum += score;
1525 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1528 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1529 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1530 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1533 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1536 /* find best f_code for ME which do unlimited searches */
1537 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1539 if(s->me_method>=ME_EPZS){
1542 uint8_t * fcode_tab= s->fcode_tab;
1544 int best_score=-10000000;
1546 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1548 for(y=0; y<s->mb_height; y++){
1550 int xy= y*s->mb_stride;
1551 for(x=0; x<s->mb_width; x++){
1552 if(s->mb_type[xy] & type){
1553 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1554 fcode_tab[mv_table[xy][1] + MAX_MV]);
1557 for(j=0; j<fcode && j<8; j++){
1558 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1567 if(score[i] > best_score){
1568 best_score= score[i];
1571 // printf("%d %d\n", i, score[i]);
1574 // printf("fcode: %d type: %d\n", i, s->pict_type);
1576 /* for(i=0; i<=MAX_FCODE; i++){
1577 printf("%d ", mv_num[i]);
1585 void ff_fix_long_p_mvs(MpegEncContext * s)
1587 const int f_code= s->f_code;
1589 assert(s->pict_type==P_TYPE);
1591 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1593 if(s->msmpeg4_version) range= 16;
1595 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1597 /* clip / convert to intra 16x16 type MVs */
1598 for(y=0; y<s->mb_height; y++){
1600 int xy= y*s->mb_stride;
1601 for(x=0; x<s->mb_width; x++){
1602 if(s->mb_type[xy]&MB_TYPE_INTER){
1603 if( s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
1604 || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
1605 s->mb_type[xy] &= ~MB_TYPE_INTER;
1606 s->mb_type[xy] |= MB_TYPE_INTRA;
1607 s->p_mv_table[xy][0] = 0;
1608 s->p_mv_table[xy][1] = 0;
1614 //printf("%d no:%d %d//\n", clip, noclip, f_code);
1615 if(s->flags&CODEC_FLAG_4MV){
1616 const int wrap= 2+ s->mb_width*2;
1618 /* clip / convert to intra 8x8 type MVs */
1619 for(y=0; y<s->mb_height; y++){
1620 int xy= (y*2 + 1)*wrap + 1;
1621 int i= y*s->mb_stride;
1624 for(x=0; x<s->mb_width; x++){
1625 if(s->mb_type[i]&MB_TYPE_INTER4V){
1627 for(block=0; block<4; block++){
1628 int off= (block& 1) + (block>>1)*wrap;
1629 int mx= s->motion_val[ xy + off ][0];
1630 int my= s->motion_val[ xy + off ][1];
1632 if( mx >=range || mx <-range
1633 || my >=range || my <-range){
1634 s->mb_type[i] &= ~MB_TYPE_INTER4V;
1635 s->mb_type[i] |= MB_TYPE_INTRA;
1646 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1650 // RAL: 8 in MPEG-1, 16 in MPEG-4
1651 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1653 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1655 /* clip / convert to intra 16x16 type MVs */
1656 for(y=0; y<s->mb_height; y++){
1658 int xy= y*s->mb_stride;
1659 for(x=0; x<s->mb_width; x++){
1660 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1661 if( mv_table[xy][0] >=range || mv_table[xy][0] <-range
1662 || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
1664 if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
1666 if (mv_table[xy][0] > range-1) mv_table[xy][0]= range-1;
1667 else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
1668 if (mv_table[xy][1] > range-1) mv_table[xy][1]= range-1;
1669 else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;
1679 * estimates global motion and inits sprite_ref
1681 void ff_estimate_global_motion(MpegEncContext *s, int sprite_ref[3][2]){
1683 int num= 16<<s->f_code;
1685 int best_i[2]={0,0};
1686 int best_score[2]={0,0};
1688 memset(score, 0, 2*num*sizeof(int));
1690 for(y=0; y<s->mb_height; y++){
1692 int xy= (y+1)* (s->mb_width+2)+1;
1693 int i= y*s->mb_width;
1694 for(x=0; x<s->mb_width; x++){
1697 if(!(s->mb_type[i]&MB_TYPE_INTER))
1700 mv[0]= s->p_mv_table[xy][0];
1701 mv[1]= s->p_mv_table[xy][1];
1703 if(mv[0]==0 && mv[1]==0) continue;
1705 score[0][mv[0] + num/2]++;
1706 score[1][mv[1] + num/2]++;
1711 for(i=1; i<num-1; i++){
1712 int s= score[n][i-1] + score[n][i]*2 + score[n][i+1];
1714 if(s > best_score[n]){
1721 sprite_ref[0][0]= best_i[0] - num/2;
1722 sprite_ref[0][1]= best_i[1] - num/2;
1724 // decide block type