10l: bidir_refine didn't save the new mvs. also improve speed.
[libav.git] / libavcodec / motion_est.c
CommitLineData
de6d9b64 1/*
115329f1 2 * Motion estimation
ff4ec49e 3 * Copyright (c) 2000,2001 Fabrice Bellard.
8f2ab833 4 * Copyright (c) 2002-2004 Michael Niedermayer
115329f1 5 *
de6d9b64 6 *
ff4ec49e
FB
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.
de6d9b64 11 *
ff4ec49e 12 * This library is distributed in the hope that it will be useful,
de6d9b64 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ff4ec49e
FB
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
de6d9b64 16 *
ff4ec49e
FB
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
45870f57
MN
20 *
21 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
de6d9b64 22 */
115329f1 23
983e3246
MN
24/**
25 * @file motion_est.c
26 * Motion estimation.
27 */
115329f1 28
de6d9b64
FB
29#include <stdlib.h>
30#include <stdio.h>
2d5e962b 31#include <limits.h>
de6d9b64
FB
32#include "avcodec.h"
33#include "dsputil.h"
34#include "mpegvideo.h"
35
2750b827
MN
36#undef NDEBUG
37#include <assert.h>
1457ab52 38
0d21a846 39#define SQ(a) ((a)*(a))
b3987771 40
0d21a846
MN
41#define P_LEFT P[1]
42#define P_TOP P[2]
43#define P_TOPRIGHT P[3]
44#define P_MEDIAN P[4]
0d21a846
MN
45#define P_MV1 P[9]
46
1457ab52 47static inline int sad_hpel_motion_search(MpegEncContext * s,
bb270c08 48 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN
49 int src_index, int ref_index,
50 int size, int h);
1457ab52 51
af4091f1 52static inline int update_map_generation(MotionEstContext *c)
1457ab52 53{
af4091f1
MN
54 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
55 if(c->map_generation==0){
56 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
57 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
1457ab52 58 }
af4091f1 59 return c->map_generation;
1457ab52
MN
60}
61
b07a5980
MN
62/* shape adaptive search stuff */
63typedef struct Minima{
64 int height;
65 int x, y;
66 int checked;
67}Minima;
1457ab52 68
b07a5980 69static int minima_cmp(const void *a, const void *b){
5c91a675
ZK
70 const Minima *da = (const Minima *) a;
71 const Minima *db = (const Minima *) b;
115329f1 72
b07a5980
MN
73 return da->height - db->height;
74}
1457ab52 75
2750b827
MN
76#define FLAG_QPEL 1 //must be 1
77#define FLAG_CHROMA 2
78#define FLAG_DIRECT 4
1457ab52 79
af4091f1 80static 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){
2750b827
MN
81 const int offset[3]= {
82 y*c-> stride + x,
83 ((y*c->uvstride + x)>>1),
84 ((y*c->uvstride + x)>>1),
85 };
86 int i;
87 for(i=0; i<3; i++){
88 c->src[0][i]= src [i] + offset[i];
89 c->ref[0][i]= ref [i] + offset[i];
90 }
91 if(ref_index){
92 for(i=0; i<3; i++){
93 c->ref[ref_index][i]= ref2[i] + offset[i];
94 }
95 }
1457ab52
MN
96}
97
1f202b0d
MN
98static int get_flags(MotionEstContext *c, int direct, int chroma){
99 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
115329f1 100 + (direct ? FLAG_DIRECT : 0)
2750b827 101 + (chroma ? FLAG_CHROMA : 0);
1457ab52
MN
102}
103
2750b827
MN
104static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
105 const int size, const int h, int ref_index, int src_index,
106 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
107 MotionEstContext * const c= &s->me;
108 const int stride= c->stride;
109 const int uvstride= c->uvstride;
110 const int qpel= flags&FLAG_QPEL;
111 const int chroma= flags&FLAG_CHROMA;
112 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
113 const int hx= subx + (x<<(1+qpel));
114 const int hy= suby + (y<<(1+qpel));
115 uint8_t * const * const ref= c->ref[ref_index];
116 uint8_t * const * const src= c->src[src_index];
117 int d;
118 //FIXME check chroma 4mv, (no crashes ...)
119 if(flags&FLAG_DIRECT){
120 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
121 const int time_pp= s->pp_time;
122 const int time_pb= s->pb_time;
123 const int mask= 2*qpel+1;
124 if(s->mv_type==MV_TYPE_8X8){
125 int i;
126 for(i=0; i<4; i++){
127 int fx = c->direct_basis_mv[i][0] + hx;
128 int fy = c->direct_basis_mv[i][1] + hy;
129 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
130 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
131 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
132 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
115329f1 133
2750b827
MN
134 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
135 if(qpel){
136 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
137 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
138 }else{
139 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
140 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
141 }
142 }
143 }else{
144 int fx = c->direct_basis_mv[0][0] + hx;
145 int fy = c->direct_basis_mv[0][1] + hy;
146 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
147 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
148 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
149 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
115329f1 150
2750b827
MN
151 if(qpel){
152 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
153 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
154 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
155 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
156 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
157 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
158 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
159 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
115329f1 160 }else{
2750b827
MN
161 assert((fx>>1) + 16*s->mb_x >= -16);
162 assert((fy>>1) + 16*s->mb_y >= -16);
163 assert((fx>>1) + 16*s->mb_x <= s->width);
164 assert((fy>>1) + 16*s->mb_y <= s->height);
165 assert((bx>>1) + 16*s->mb_x >= -16);
166 assert((by>>1) + 16*s->mb_y >= -16);
167 assert((bx>>1) + 16*s->mb_x <= s->width);
168 assert((by>>1) + 16*s->mb_y <= s->height);
169
170 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
171 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
172 }
173 }
174 d = cmp_func(s, c->temp, src[0], stride, 16);
175 }else
176 d= 256*256*256*32;
177 }else{
bf4e3bd2 178 int uvdxy; /* no, it might not be used uninitialized */
2750b827
MN
179 if(dxy){
180 if(qpel){
181 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
182 if(chroma){
183 int cx= hx/2;
184 int cy= hy/2;
185 cx= (cx>>1)|(cx&1);
186 cy= (cy>>1)|(cy&1);
187 uvdxy= (cx&1) + 2*(cy&1);
188 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
189 }
190 }else{
191 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
192 if(chroma)
193 uvdxy= dxy | (x&1) | (2*(y&1));
194 }
115329f1 195 d = cmp_func(s, c->temp, src[0], stride, h);
2750b827 196 }else{
115329f1 197 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
2750b827
MN
198 if(chroma)
199 uvdxy= (x&1) + 2*(y&1);
200 }
201 if(chroma){
202 uint8_t * const uvtemp= c->temp + 16*stride;
203 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
204 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
115329f1
DB
205 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
206 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
2750b827
MN
207 }
208 }
209#if 0
210 if(full_pel){
211 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
212 score_map[index]= d;
213 }
214
215 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
216#endif
217 return d;
218}
1457ab52
MN
219
220#include "motion_est_template.c"
1457ab52 221
dbc56b39
MN
222static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
223 return 0;
224}
225
226static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
227}
228
1457ab52 229void ff_init_me(MpegEncContext *s){
9964df63 230 MotionEstContext * const c= &s->me;
1f202b0d 231 c->avctx= s->avctx;
9964df63 232
1f202b0d
MN
233 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
234 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
235 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
236 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
115329f1 237
1f202b0d
MN
238 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
239 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
240 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
de6d9b64 241
9964df63 242/*FIXME s->no_rounding b_type*/
1457ab52 243 if(s->flags&CODEC_FLAG_QPEL){
af4091f1 244 c->sub_motion_search= qpel_motion_search;
9964df63
MN
245 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
246 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
247 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
1457ab52 248 }else{
1f202b0d 249 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
af4091f1 250 c->sub_motion_search= hpel_motion_search;
115329f1
DB
251 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
252 && c->avctx-> me_cmp == FF_CMP_SAD
1f202b0d 253 && c->avctx-> mb_cmp == FF_CMP_SAD)
af4091f1 254 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
1457ab52 255 else
af4091f1 256 c->sub_motion_search= hpel_motion_search;
826f429a 257 }
dbc56b39
MN
258 c->hpel_avg= s->dsp.avg_pixels_tab;
259 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
260 else c->hpel_put= s->dsp.put_pixels_tab;
261
2750b827 262 if(s->linesize){
115329f1 263 c->stride = s->linesize;
af4091f1 264 c->uvstride= s->uvlinesize;
67725183 265 }else{
af4091f1
MN
266 c->stride = 16*s->mb_width + 32;
267 c->uvstride= 8*s->mb_width + 16;
67725183 268 }
9964df63 269
dbc56b39 270 // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
155ec6ed 271 if(s->codec_id != CODEC_ID_SNOW){
725de25c 272 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
155ec6ed
MN
273 s->dsp.me_cmp[2]= zero_cmp;
274 }
275 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
276 s->dsp.me_sub_cmp[2]= zero_cmp;
277 }
278 c->hpel_put[2][0]= c->hpel_put[2][1]=
279 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
dbc56b39 280 }
dbc56b39 281
1c3990db
MN
282 if(s->codec_id == CODEC_ID_H261){
283 c->sub_motion_search= no_sub_motion_search;
284 }
285
9964df63 286 c->temp= c->scratchpad;
1457ab52 287}
115329f1 288
802f454e 289#if 0
0c1a9eda 290static int pix_dev(uint8_t * pix, int line_size, int mean)
b3987771
MN
291{
292 int s, i, j;
293
294 s = 0;
295 for (i = 0; i < 16; i++) {
bb270c08
DB
296 for (j = 0; j < 16; j += 8) {
297 s += ABS(pix[0]-mean);
298 s += ABS(pix[1]-mean);
299 s += ABS(pix[2]-mean);
300 s += ABS(pix[3]-mean);
301 s += ABS(pix[4]-mean);
302 s += ABS(pix[5]-mean);
303 s += ABS(pix[6]-mean);
304 s += ABS(pix[7]-mean);
305 pix += 8;
306 }
307 pix += line_size - 16;
b3987771
MN
308 }
309 return s;
310}
802f454e 311#endif
b3987771 312
eb4b3dd3 313static inline void no_motion_search(MpegEncContext * s,
bb270c08 314 int *mx_ptr, int *my_ptr)
de6d9b64
FB
315{
316 *mx_ptr = 16 * s->mb_x;
317 *my_ptr = 16 * s->mb_y;
318}
319
88730be6 320#if 0 /* the use of these functions is inside #if 0 */
de6d9b64
FB
321static int full_motion_search(MpegEncContext * s,
322 int *mx_ptr, int *my_ptr, int range,
9dbcbd92 323 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
de6d9b64
FB
324{
325 int x1, y1, x2, y2, xx, yy, x, y;
326 int mx, my, dmin, d;
0c1a9eda 327 uint8_t *pix;
de6d9b64
FB
328
329 xx = 16 * s->mb_x;
330 yy = 16 * s->mb_y;
bb270c08 331 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
de6d9b64 332 if (x1 < xmin)
bb270c08 333 x1 = xmin;
de6d9b64
FB
334 x2 = xx + range - 1;
335 if (x2 > xmax)
bb270c08 336 x2 = xmax;
de6d9b64
FB
337 y1 = yy - range + 1;
338 if (y1 < ymin)
bb270c08 339 y1 = ymin;
de6d9b64
FB
340 y2 = yy + range - 1;
341 if (y2 > ymax)
bb270c08 342 y2 = ymax;
1e491e29 343 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
de6d9b64
FB
344 dmin = 0x7fffffff;
345 mx = 0;
346 my = 0;
347 for (y = y1; y <= y2; y++) {
bb270c08
DB
348 for (x = x1; x <= x2; x++) {
349 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
350 s->linesize, 16);
351 if (d < dmin ||
352 (d == dmin &&
353 (abs(x - xx) + abs(y - yy)) <
354 (abs(mx - xx) + abs(my - yy)))) {
355 dmin = d;
356 mx = x;
357 my = y;
358 }
359 }
de6d9b64
FB
360 }
361
362 *mx_ptr = mx;
363 *my_ptr = my;
364
365#if 0
366 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
bb270c08
DB
367 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
368 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
de6d9b64
FB
369 }
370#endif
371 return dmin;
372}
373
374
375static int log_motion_search(MpegEncContext * s,
376 int *mx_ptr, int *my_ptr, int range,
9dbcbd92 377 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
de6d9b64
FB
378{
379 int x1, y1, x2, y2, xx, yy, x, y;
380 int mx, my, dmin, d;
0c1a9eda 381 uint8_t *pix;
de6d9b64
FB
382
383 xx = s->mb_x << 4;
384 yy = s->mb_y << 4;
385
386 /* Left limit */
387 x1 = xx - range;
388 if (x1 < xmin)
bb270c08 389 x1 = xmin;
de6d9b64
FB
390
391 /* Right limit */
392 x2 = xx + range;
393 if (x2 > xmax)
bb270c08 394 x2 = xmax;
de6d9b64
FB
395
396 /* Upper limit */
397 y1 = yy - range;
398 if (y1 < ymin)
bb270c08 399 y1 = ymin;
de6d9b64
FB
400
401 /* Lower limit */
402 y2 = yy + range;
403 if (y2 > ymax)
bb270c08 404 y2 = ymax;
de6d9b64 405
1e491e29 406 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
de6d9b64
FB
407 dmin = 0x7fffffff;
408 mx = 0;
409 my = 0;
410
411 do {
bb270c08
DB
412 for (y = y1; y <= y2; y += range) {
413 for (x = x1; x <= x2; x += range) {
414 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
415 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
416 dmin = d;
417 mx = x;
418 my = y;
419 }
420 }
421 }
422
423 range = range >> 1;
424
425 x1 = mx - range;
426 if (x1 < xmin)
427 x1 = xmin;
428
429 x2 = mx + range;
430 if (x2 > xmax)
431 x2 = xmax;
432
433 y1 = my - range;
434 if (y1 < ymin)
435 y1 = ymin;
436
437 y2 = my + range;
438 if (y2 > ymax)
439 y2 = ymax;
de6d9b64
FB
440
441 } while (range >= 1);
442
443#ifdef DEBUG
95ae72b7 444 av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
de6d9b64
FB
445#endif
446 *mx_ptr = mx;
447 *my_ptr = my;
448 return dmin;
449}
450
451static int phods_motion_search(MpegEncContext * s,
452 int *mx_ptr, int *my_ptr, int range,
9dbcbd92 453 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
de6d9b64
FB
454{
455 int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
456 int mx, my, dminx, dminy;
0c1a9eda 457 uint8_t *pix;
de6d9b64
FB
458
459 xx = s->mb_x << 4;
460 yy = s->mb_y << 4;
461
462 /* Left limit */
463 x1 = xx - range;
464 if (x1 < xmin)
bb270c08 465 x1 = xmin;
de6d9b64
FB
466
467 /* Right limit */
468 x2 = xx + range;
469 if (x2 > xmax)
bb270c08 470 x2 = xmax;
de6d9b64
FB
471
472 /* Upper limit */
473 y1 = yy - range;
474 if (y1 < ymin)
bb270c08 475 y1 = ymin;
de6d9b64
FB
476
477 /* Lower limit */
478 y2 = yy + range;
479 if (y2 > ymax)
bb270c08 480 y2 = ymax;
de6d9b64 481
1e491e29 482 pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
de6d9b64
FB
483 mx = 0;
484 my = 0;
485
486 x = xx;
487 y = yy;
488 do {
489 dminx = 0x7fffffff;
490 dminy = 0x7fffffff;
491
bb270c08
DB
492 lastx = x;
493 for (x = x1; x <= x2; x += range) {
494 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
495 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
496 dminx = d;
497 mx = x;
498 }
499 }
500
501 x = lastx;
502 for (y = y1; y <= y2; y += range) {
503 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
504 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
505 dminy = d;
506 my = y;
507 }
508 }
509
510 range = range >> 1;
511
512 x = mx;
513 y = my;
514 x1 = mx - range;
515 if (x1 < xmin)
516 x1 = xmin;
517
518 x2 = mx + range;
519 if (x2 > xmax)
520 x2 = xmax;
521
522 y1 = my - range;
523 if (y1 < ymin)
524 y1 = ymin;
525
526 y2 = my + range;
527 if (y2 > ymax)
528 y2 = ymax;
de6d9b64
FB
529
530 } while (range >= 1);
531
532#ifdef DEBUG
95ae72b7 533 av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
de6d9b64
FB
534#endif
535
536 /* half pixel search */
537 *mx_ptr = mx;
538 *my_ptr = my;
539 return dminy;
540}
88730be6 541#endif /* 0 */
45870f57
MN
542
543#define Z_THRESHOLD 256
544
1457ab52 545#define CHECK_SAD_HALF_MV(suffix, x, y) \
0d21a846 546{\
bb198e19 547 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
1457ab52 548 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
0d21a846
MN
549 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
550}
45870f57 551
1457ab52 552static inline int sad_hpel_motion_search(MpegEncContext * s,
bb270c08 553 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN
554 int src_index, int ref_index,
555 int size, int h)
de6d9b64 556{
af4091f1
MN
557 MotionEstContext * const c= &s->me;
558 const int penalty_factor= c->sub_penalty_factor;
bb198e19 559 int mx, my, dminh;
0c1a9eda 560 uint8_t *pix, *ptr;
af4091f1
MN
561 int stride= c->stride;
562 const int flags= c->sub_flags;
2750b827 563 LOAD_COMMON
115329f1 564
2750b827 565 assert(flags == 0);
de6d9b64 566
af4091f1 567 if(c->skip){
0d21a846
MN
568// printf("S");
569 *mx_ptr = 0;
570 *my_ptr = 0;
571 return dmin;
572 }
573// printf("N");
115329f1 574
af4091f1 575 pix = c->src[src_index][0];
0d21a846 576
ba6802de
MN
577 mx = *mx_ptr;
578 my = *my_ptr;
af4091f1 579 ptr = c->ref[ref_index][0] + (my * stride) + mx;
115329f1 580
ba6802de
MN
581 dminh = dmin;
582
115329f1 583 if (mx > xmin && mx < xmax &&
ba6802de 584 my > ymin && my < ymax) {
0d21a846 585 int dx=0, dy=0;
115329f1 586 int d, pen_x, pen_y;
0d21a846
MN
587 const int index= (my<<ME_MAP_SHIFT) + mx;
588 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
589 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
590 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
591 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
592 mx<<=1;
593 my<<=1;
ba6802de 594
115329f1 595
ba6802de
MN
596 pen_x= pred_x + mx;
597 pen_y= pred_y + my;
598
bb198e19 599 ptr-= stride;
0d21a846 600 if(t<=b){
1457ab52 601 CHECK_SAD_HALF_MV(y2 , 0, -1)
0d21a846 602 if(l<=r){
1457ab52 603 CHECK_SAD_HALF_MV(xy2, -1, -1)
0d21a846 604 if(t+r<=b+l){
1457ab52 605 CHECK_SAD_HALF_MV(xy2, +1, -1)
bb198e19 606 ptr+= stride;
0d21a846 607 }else{
bb198e19 608 ptr+= stride;
1457ab52 609 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846 610 }
1457ab52 611 CHECK_SAD_HALF_MV(x2 , -1, 0)
0d21a846 612 }else{
1457ab52 613 CHECK_SAD_HALF_MV(xy2, +1, -1)
0d21a846 614 if(t+l<=b+r){
1457ab52 615 CHECK_SAD_HALF_MV(xy2, -1, -1)
bb198e19 616 ptr+= stride;
0d21a846 617 }else{
bb198e19 618 ptr+= stride;
1457ab52 619 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 620 }
1457ab52 621 CHECK_SAD_HALF_MV(x2 , +1, 0)
0d21a846
MN
622 }
623 }else{
624 if(l<=r){
625 if(t+l<=b+r){
1457ab52 626 CHECK_SAD_HALF_MV(xy2, -1, -1)
bb198e19 627 ptr+= stride;
0d21a846 628 }else{
bb198e19 629 ptr+= stride;
1457ab52 630 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 631 }
1457ab52
MN
632 CHECK_SAD_HALF_MV(x2 , -1, 0)
633 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846
MN
634 }else{
635 if(t+r<=b+l){
1457ab52 636 CHECK_SAD_HALF_MV(xy2, +1, -1)
bb198e19 637 ptr+= stride;
0d21a846 638 }else{
bb198e19 639 ptr+= stride;
1457ab52 640 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846 641 }
1457ab52
MN
642 CHECK_SAD_HALF_MV(x2 , +1, 0)
643 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 644 }
1457ab52 645 CHECK_SAD_HALF_MV(y2 , 0, +1)
0d21a846
MN
646 }
647 mx+=dx;
648 my+=dy;
ba6802de
MN
649
650 }else{
0d21a846
MN
651 mx<<=1;
652 my<<=1;
ba6802de
MN
653 }
654
655 *mx_ptr = mx;
656 *my_ptr = my;
0d21a846 657 return dminh;
ba6802de
MN
658}
659
0d21a846 660static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
ba6802de 661{
7bc9090a 662 const int xy= s->mb_x + s->mb_y*s->mb_stride;
115329f1 663
9dbcbd92
MN
664 s->p_mv_table[xy][0] = mx;
665 s->p_mv_table[xy][1] = my;
ba6802de 666
bb628dae 667 /* has already been set to the 4 MV if 4MV is done */
0d21a846 668 if(mv4){
ba6802de
MN
669 int mot_xy= s->block_index[0];
670
8d7ec294
WH
671 s->current_picture.motion_val[0][mot_xy ][0]= mx;
672 s->current_picture.motion_val[0][mot_xy ][1]= my;
673 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
674 s->current_picture.motion_val[0][mot_xy+1][1]= my;
ba6802de 675
137c8468 676 mot_xy += s->b8_stride;
8d7ec294
WH
677 s->current_picture.motion_val[0][mot_xy ][0]= mx;
678 s->current_picture.motion_val[0][mot_xy ][1]= my;
679 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
680 s->current_picture.motion_val[0][mot_xy+1][1]= my;
ba6802de
MN
681 }
682}
683
ebbcdc9a
MN
684/**
685 * get fullpel ME search limits.
ebbcdc9a 686 */
bb198e19 687static inline void get_limits(MpegEncContext *s, int x, int y)
de6d9b64 688{
af4091f1 689 MotionEstContext * const c= &s->me;
bb198e19 690/*
1f202b0d 691 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
af4091f1 692 else c->range= 16;
bb198e19 693*/
de6d9b64 694 if (s->unrestricted_mv) {
af4091f1
MN
695 c->xmin = - x - 16;
696 c->ymin = - y - 16;
697 c->xmax = - x + s->mb_width *16;
698 c->ymax = - y + s->mb_height*16;
1c3990db
MN
699 } else if (s->out_format == FMT_H261){
700 // Search range of H261 is different from other codec standards
701 c->xmin = (x > 15) ? - 15 : 0;
702 c->ymin = (y > 15) ? - 15 : 0;
115329f1 703 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
1c3990db 704 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
de6d9b64 705 } else {
af4091f1
MN
706 c->xmin = - x;
707 c->ymin = - y;
708 c->xmax = - x + s->mb_width *16 - 16;
709 c->ymax = - y + s->mb_height*16 - 16;
de6d9b64 710 }
9dbcbd92
MN
711}
712
1f202b0d
MN
713static inline void init_mv4_ref(MotionEstContext *c){
714 const int stride= c->stride;
9964df63
MN
715
716 c->ref[1][0] = c->ref[0][0] + 8;
717 c->ref[2][0] = c->ref[0][0] + 8*stride;
718 c->ref[3][0] = c->ref[2][0] + 8;
719 c->src[1][0] = c->src[0][0] + 8;
720 c->src[2][0] = c->src[0][0] + 8*stride;
721 c->src[3][0] = c->src[2][0] + 8;
722}
723
bb198e19 724static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
0d21a846 725{
2750b827 726 MotionEstContext * const c= &s->me;
bb198e19
MN
727 const int size= 1;
728 const int h=8;
0d21a846
MN
729 int block;
730 int P[10][2];
67725183 731 int dmin_sum=0, mx4_sum=0, my4_sum=0;
ca7d05d5 732 int same=1;
1f202b0d 733 const int stride= c->stride;
af4091f1 734 uint8_t *mv_penalty= c->current_mv_penalty;
2750b827 735
1f202b0d 736 init_mv4_ref(c);
115329f1 737
0d21a846
MN
738 for(block=0; block<4; block++){
739 int mx4, my4;
740 int pred_x4, pred_y4;
741 int dmin4;
742 static const int off[4]= {2, 1, 1, -1};
137c8468 743 const int mot_stride = s->b8_stride;
0d21a846 744 const int mot_xy = s->block_index[block];
bb198e19 745
8d7ec294
WH
746 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
747 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
0d21a846 748
af4091f1 749 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
0d21a846
MN
750
751 /* special case for first line */
9c3d33d6 752 if (s->first_slice_line && block<2) {
af4091f1
MN
753 c->pred_x= pred_x4= P_LEFT[0];
754 c->pred_y= pred_y4= P_LEFT[1];
0d21a846 755 } else {
8d7ec294
WH
756 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
757 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
758 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
759 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
af4091f1
MN
760 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
761 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
762 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
763 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 764
0d21a846
MN
765 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
766 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
767
af4091f1
MN
768 c->pred_x= pred_x4 = P_MEDIAN[0];
769 c->pred_y= pred_y4 = P_MEDIAN[1];
0d21a846
MN
770 }
771 P_MV1[0]= mx;
772 P_MV1[1]= my;
773
2750b827 774 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
0d21a846 775
af4091f1 776 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
115329f1 777
2750b827 778 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
67725183 779 int dxy;
bb198e19 780 const int offset= ((block&1) + (block>>1)*stride)*8;
af4091f1 781 uint8_t *dest_y = c->scratchpad + offset;
67725183 782 if(s->quarter_sample){
2750b827 783 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
67725183
MN
784 dxy = ((my4 & 3) << 2) | (mx4 & 3);
785
786 if(s->no_rounding)
9c3d33d6 787 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
67725183 788 else
bb198e19 789 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
67725183 790 }else{
2750b827 791 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
67725183
MN
792 dxy = ((my4 & 1) << 1) | (mx4 & 1);
793
794 if(s->no_rounding)
bb198e19 795 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
67725183 796 else
bb198e19 797 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
67725183 798 }
af4091f1 799 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
67725183
MN
800 }else
801 dmin_sum+= dmin4;
802
803 if(s->quarter_sample){
804 mx4_sum+= mx4/2;
805 my4_sum+= my4/2;
806 }else{
807 mx4_sum+= mx4;
808 my4_sum+= my4;
809 }
115329f1 810
8d7ec294
WH
811 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
812 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
ca7d05d5
MN
813
814 if(mx4 != mx || my4 != my) same=0;
0d21a846 815 }
115329f1 816
ca7d05d5
MN
817 if(same)
818 return INT_MAX;
115329f1 819
cb009a8c 820 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
af4091f1 821 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
67725183 822 }
115329f1 823
1f202b0d 824 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
67725183
MN
825 int dxy;
826 int mx, my;
827 int offset;
828
829 mx= ff_h263_round_chroma(mx4_sum);
830 my= ff_h263_round_chroma(my4_sum);
831 dxy = ((my & 1) << 1) | (mx & 1);
115329f1 832
67725183 833 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
115329f1 834
67725183 835 if(s->no_rounding){
af4091f1
MN
836 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
837 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
67725183 838 }else{
af4091f1
MN
839 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
840 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
67725183
MN
841 }
842
af4091f1
MN
843 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
844 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
67725183 845 }
115329f1 846
af4091f1
MN
847 c->pred_x= mx;
848 c->pred_y= my;
67725183 849
1f202b0d 850 switch(c->avctx->mb_cmp&0xFF){
67725183
MN
851 /*case FF_CMP_SSE:
852 return dmin_sum+ 32*s->qscale*s->qscale;*/
853 case FF_CMP_RD:
854 return dmin_sum;
855 default:
af4091f1 856 return dmin_sum+ 11*c->mb_penalty_factor;
67725183 857 }
0d21a846
MN
858}
859
9964df63
MN
860static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
861 MotionEstContext * const c= &s->me;
862
863 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
864 c->src[1][0] = c->src[0][0] + s->linesize;
865 if(c->flags & FLAG_CHROMA){
866 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
867 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
868 c->src[1][1] = c->src[0][1] + s->uvlinesize;
869 c->src[1][2] = c->src[0][2] + s->uvlinesize;
870 }
871}
872
115329f1 873static int interlaced_search(MpegEncContext *s, int ref_index,
f20f8a8b 874 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
bb198e19 875{
2750b827 876 MotionEstContext * const c= &s->me;
bb198e19
MN
877 const int size=0;
878 const int h=8;
879 int block;
880 int P[10][2];
2750b827 881 uint8_t * const mv_penalty= c->current_mv_penalty;
bb198e19
MN
882 int same=1;
883 const int stride= 2*s->linesize;
bb198e19
MN
884 int dmin_sum= 0;
885 const int mot_stride= s->mb_stride;
886 const int xy= s->mb_x + s->mb_y*mot_stride;
115329f1 887
2750b827
MN
888 c->ymin>>=1;
889 c->ymax>>=1;
890 c->stride<<=1;
891 c->uvstride<<=1;
9964df63 892 init_interlaced_ref(s, ref_index);
115329f1 893
bb198e19
MN
894 for(block=0; block<2; block++){
895 int field_select;
896 int best_dmin= INT_MAX;
897 int best_field= -1;
898
bb198e19 899 for(field_select=0; field_select<2; field_select++){
2750b827 900 int dmin, mx_i, my_i;
bb198e19 901 int16_t (*mv_table)[2]= mv_tables[block][field_select];
115329f1 902
f20f8a8b
MN
903 if(user_field_select){
904 if(field_select_tables[block][xy] != field_select)
905 continue;
906 }
115329f1 907
bb198e19
MN
908 P_LEFT[0] = mv_table[xy - 1][0];
909 P_LEFT[1] = mv_table[xy - 1][1];
2750b827 910 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
115329f1 911
af4091f1
MN
912 c->pred_x= P_LEFT[0];
913 c->pred_y= P_LEFT[1];
115329f1 914
9c3d33d6 915 if(!s->first_slice_line){
bb198e19
MN
916 P_TOP[0] = mv_table[xy - mot_stride][0];
917 P_TOP[1] = mv_table[xy - mot_stride][1];
918 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
919 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
2750b827
MN
920 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
921 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
922 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
923 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
115329f1 924
bb198e19
MN
925 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
926 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
927 }
928 P_MV1[0]= mx; //FIXME not correct if block != field_select
929 P_MV1[1]= my / 2;
115329f1 930
2750b827 931 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
bb198e19 932
2750b827 933 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
115329f1 934
bb198e19
MN
935 mv_table[xy][0]= mx_i;
936 mv_table[xy][1]= my_i;
115329f1 937
2750b827 938 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
bb198e19
MN
939 int dxy;
940
941 //FIXME chroma ME
2750b827 942 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
bb198e19
MN
943 dxy = ((my_i & 1) << 1) | (mx_i & 1);
944
945 if(s->no_rounding){
2750b827 946 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
bb198e19 947 }else{
2750b827 948 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
bb198e19 949 }
2750b827 950 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
af4091f1 951 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
bb198e19 952 }else
2750b827 953 dmin+= c->mb_penalty_factor; //field_select bits
115329f1 954
bb198e19 955 dmin += field_select != block; //slightly prefer same field
115329f1 956
bb198e19
MN
957 if(dmin < best_dmin){
958 best_dmin= dmin;
959 best_field= field_select;
960 }
961 }
962 {
963 int16_t (*mv_table)[2]= mv_tables[block][best_field];
964
965 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
966 if(mv_table[xy][1]&1) same=0;
115329f1 967 if(mv_table[xy][1]*2 != my) same=0;
bb198e19
MN
968 if(best_field != block) same=0;
969 }
970
971 field_select_tables[block][xy]= best_field;
972 dmin_sum += best_dmin;
973 }
115329f1 974
2750b827
MN
975 c->ymin<<=1;
976 c->ymax<<=1;
977 c->stride>>=1;
978 c->uvstride>>=1;
bb198e19
MN
979
980 if(same)
981 return INT_MAX;
115329f1 982
1f202b0d 983 switch(c->avctx->mb_cmp&0xFF){
bb198e19
MN
984 /*case FF_CMP_SSE:
985 return dmin_sum+ 32*s->qscale*s->qscale;*/
986 case FF_CMP_RD:
987 return dmin_sum;
988 default:
2750b827 989 return dmin_sum+ 11*c->mb_penalty_factor;
bb198e19
MN
990 }
991}
992
ae1dbde1
MN
993static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
994 int ymax= s->me.ymax>>interlaced;
995 int ymin= s->me.ymin>>interlaced;
115329f1 996
ae1dbde1
MN
997 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
998 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
999 if(mv[1] < ymin) mv[1] = ymin;
1000 if(mv[1] > ymax) mv[1] = ymax;
1001}
1002
f4f3223f
MN
1003static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1004 MotionEstContext * const c= &s->me;
1005 Picture *p= s->current_picture_ptr;
1006 int mb_xy= mb_x + mb_y*s->mb_stride;
1007 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1008 int mb_type= s->current_picture.mb_type[mb_xy];
1009 int flags= c->flags;
1010 int shift= (flags&FLAG_QPEL) + 1;
1011 int mask= (1<<shift)-1;
9964df63 1012 int x, y, i;
f4f3223f
MN
1013 int d=0;
1014 me_cmp_func cmpf= s->dsp.sse[0];
1015 me_cmp_func chroma_cmpf= s->dsp.sse[1];
115329f1 1016
ae1dbde1
MN
1017 if(p_type && USES_LIST(mb_type, 1)){
1018 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
cec1f05f 1019 return INT_MAX/2;
ae1dbde1 1020 }
f4f3223f 1021 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
115329f1 1022
ae1dbde1
MN
1023 for(i=0; i<4; i++){
1024 int xy= s->block_index[i];
1025 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1026 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1027 }
1028
f4f3223f
MN
1029 if(IS_INTERLACED(mb_type)){
1030 int xy2= xy + s->b8_stride;
1031 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1032 c->stride<<=1;
1033 c->uvstride<<=1;
115329f1 1034
1545acea 1035 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1f202b0d 1036 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
cec1f05f 1037 return INT_MAX/2;
1545acea 1038 }
9964df63 1039
f4f3223f
MN
1040 if(USES_LIST(mb_type, 0)){
1041 int field_select0= p->ref_index[0][xy ];
1042 int field_select1= p->ref_index[0][xy2];
1043 assert(field_select0==0 ||field_select0==1);
1044 assert(field_select1==0 ||field_select1==1);
f20f8a8b
MN
1045 init_interlaced_ref(s, 0);
1046
f4f3223f
MN
1047 if(p_type){
1048 s->p_field_select_table[0][mb_xy]= field_select0;
1049 s->p_field_select_table[1][mb_xy]= field_select1;
1050 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1051 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1052 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1053 }else{
1054 s->b_field_select_table[0][0][mb_xy]= field_select0;
1055 s->b_field_select_table[0][1][mb_xy]= field_select1;
1056 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1057 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1058 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1059 }
1060
115329f1 1061 x= p->motion_val[0][xy ][0];
f4f3223f
MN
1062 y= p->motion_val[0][xy ][1];
1063 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
115329f1 1064 x= p->motion_val[0][xy2][0];
f4f3223f
MN
1065 y= p->motion_val[0][xy2][1];
1066 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1067 }
1068 if(USES_LIST(mb_type, 1)){
1069 int field_select0= p->ref_index[1][xy ];
1070 int field_select1= p->ref_index[1][xy2];
1071 assert(field_select0==0 ||field_select0==1);
1072 assert(field_select1==0 ||field_select1==1);
f20f8a8b
MN
1073 init_interlaced_ref(s, 2);
1074
f4f3223f
MN
1075 s->b_field_select_table[1][0][mb_xy]= field_select0;
1076 s->b_field_select_table[1][1][mb_xy]= field_select1;
1077 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1078 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1079 if(USES_LIST(mb_type, 0)){
1080 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1081 }else{
1082 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1083 }
1084
115329f1 1085 x= p->motion_val[1][xy ][0];
f4f3223f
MN
1086 y= p->motion_val[1][xy ][1];
1087 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
115329f1 1088 x= p->motion_val[1][xy2][0];
f4f3223f
MN
1089 y= p->motion_val[1][xy2][1];
1090 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1091 //FIXME bidir scores
1092 }
1093 c->stride>>=1;
1094 c->uvstride>>=1;
9964df63 1095 }else if(IS_8X8(mb_type)){
1545acea 1096 if(!(s->flags & CODEC_FLAG_4MV)){
1f202b0d 1097 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
cec1f05f 1098 return INT_MAX/2;
1545acea 1099 }
9964df63
MN
1100 cmpf= s->dsp.sse[1];
1101 chroma_cmpf= s->dsp.sse[1];
1f202b0d 1102 init_mv4_ref(c);
9964df63
MN
1103 for(i=0; i<4; i++){
1104 xy= s->block_index[i];
115329f1 1105 x= p->motion_val[0][xy][0];
9964df63
MN
1106 y= p->motion_val[0][xy][1];
1107 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1108 }
1109 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
f4f3223f
MN
1110 }else{
1111 if(USES_LIST(mb_type, 0)){
1112 if(p_type){
1113 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1114 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1115 }else if(USES_LIST(mb_type, 1)){
1116 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1117 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1118 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1119 }else{
1120 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1121 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1122 }
115329f1 1123 x= p->motion_val[0][xy][0];
f4f3223f
MN
1124 y= p->motion_val[0][xy][1];
1125 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1126 }else if(USES_LIST(mb_type, 1)){
1127 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1128 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
115329f1
DB
1129
1130 x= p->motion_val[1][xy][0];
f4f3223f
MN
1131 y= p->motion_val[1][xy][1];
1132 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1133 }else
1134 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1135 }
1136 return d;
1137}
1138
9dbcbd92
MN
1139void ff_estimate_p_frame_motion(MpegEncContext * s,
1140 int mb_x, int mb_y)
1141{
2750b827 1142 MotionEstContext * const c= &s->me;
0c1a9eda 1143 uint8_t *pix, *ppix;
f20f8a8b 1144 int sum, varc, vard, mx, my, dmin;
0d21a846 1145 int P[10][2];
9dbcbd92
MN
1146 const int shift= 1+s->quarter_sample;
1147 int mb_type=0;
1e491e29 1148 Picture * const pic= &s->current_picture;
115329f1 1149
af4091f1 1150 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
bb198e19 1151
1457ab52 1152 assert(s->quarter_sample==0 || s->quarter_sample==1);
af4091f1
MN
1153 assert(s->linesize == c->stride);
1154 assert(s->uvlinesize == c->uvstride);
1457ab52 1155
26efc54e
MN
1156 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1157 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1158 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
af4091f1 1159 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
9dbcbd92 1160
bb198e19 1161 get_limits(s, 16*mb_x, 16*mb_y);
af4091f1 1162 c->skip=0;
9dbcbd92 1163
f20f8a8b
MN
1164 /* intra / predictive decision */
1165 pix = c->src[0][0];
1166 sum = s->dsp.pix_sum(pix, s->linesize);
aaaa6f15 1167 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
f20f8a8b
MN
1168
1169 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
aaaa6f15
MN
1170 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1171 c->mb_var_sum_temp += (varc+128)>>8;
f20f8a8b 1172
1f202b0d 1173 if(c->avctx->me_threshold){
aaaa6f15 1174 vard= check_input_motion(s, mb_x, mb_y, 1);
115329f1 1175
aaaa6f15
MN
1176 if((vard+128)>>8 < c->avctx->me_threshold){
1177 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1178 c->mc_mb_var_sum_temp += (vard+128)>>8;
1179 if (vard <= 64<<8 || vard < varc) { //FIXME
1180 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
f4f3223f 1181 }else{
1f202b0d 1182 c->scene_change_score+= s->qscale;
f4f3223f
MN
1183 }
1184 return;
1185 }
aaaa6f15 1186 if((vard+128)>>8 < c->avctx->mb_threshold)
f20f8a8b 1187 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
f4f3223f
MN
1188 }
1189
e4986da9 1190 switch(s->me_method) {
de6d9b64
FB
1191 case ME_ZERO:
1192 default:
bb270c08 1193 no_motion_search(s, &mx, &my);
0d21a846
MN
1194 mx-= mb_x*16;
1195 my-= mb_y*16;
de6d9b64
FB
1196 dmin = 0;
1197 break;
bb198e19 1198#if 0
de6d9b64 1199 case ME_FULL:
bb270c08 1200 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
0d21a846
MN
1201 mx-= mb_x*16;
1202 my-= mb_y*16;
de6d9b64
FB
1203 break;
1204 case ME_LOG:
bb270c08 1205 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
0d21a846
MN
1206 mx-= mb_x*16;
1207 my-= mb_y*16;
de6d9b64
FB
1208 break;
1209 case ME_PHODS:
bb270c08 1210 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
0d21a846
MN
1211 mx-= mb_x*16;
1212 my-= mb_y*16;
de6d9b64 1213 break;
bb198e19 1214#endif
174489bd 1215 case ME_X1:
45870f57 1216 case ME_EPZS:
174489bd 1217 {
137c8468 1218 const int mot_stride = s->b8_stride;
ba6802de 1219 const int mot_xy = s->block_index[0];
174489bd 1220
8d7ec294
WH
1221 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1222 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
174489bd 1223
af4091f1 1224 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
c5b1c10a 1225
9c3d33d6 1226 if(!s->first_slice_line) {
8d7ec294
WH
1227 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1228 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1229 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1230 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
af4091f1
MN
1231 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1232 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1233 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 1234
0d21a846
MN
1235 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1236 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1237
1238 if(s->out_format == FMT_H263){
2750b827
MN
1239 c->pred_x = P_MEDIAN[0];
1240 c->pred_y = P_MEDIAN[1];
0d21a846 1241 }else { /* mpeg1 at least */
2750b827
MN
1242 c->pred_x= P_LEFT[0];
1243 c->pred_y= P_LEFT[1];
0d21a846 1244 }
f931ff7b 1245 }else{
2750b827
MN
1246 c->pred_x= P_LEFT[0];
1247 c->pred_y= P_LEFT[1];
174489bd 1248 }
f931ff7b 1249
c5b1c10a 1250 }
115329f1 1251 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
2750b827 1252
45870f57 1253 break;
de6d9b64 1254 }
de6d9b64 1255
0d21a846 1256 /* At this point (mx,my) are full-pell and the relative displacement */
2750b827 1257 ppix = c->ref[0][0] + (my * s->linesize) + mx;
115329f1 1258
aaaa6f15 1259 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1fb4890b 1260
aaaa6f15 1261 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
115329f1 1262// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
aaaa6f15 1263 c->mc_mb_var_sum_temp += (vard+128)>>8;
115329f1 1264
de6d9b64 1265#if 0
e03c341e 1266 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
bb270c08 1267 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
de6d9b64 1268#endif
f20f8a8b 1269 if(mb_type){
aaaa6f15
MN
1270 if (vard <= 64<<8 || vard < varc)
1271 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
f20f8a8b 1272 else
1f202b0d 1273 c->scene_change_score+= s->qscale;
f20f8a8b
MN
1274
1275 if(mb_type == CANDIDATE_MB_TYPE_INTER){
af4091f1 1276 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
f20f8a8b
MN
1277 set_p_mv_tables(s, mx, my, 1);
1278 }else{
1279 mx <<=shift;
1280 my <<=shift;
1281 }
1282 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1283 h263_mv4_search(s, mx, my, shift);
1284
1285 set_p_mv_tables(s, mx, my, 0);
1286 }
1287 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1288 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1289 }
1f202b0d 1290 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
aaaa6f15
MN
1291 if (vard <= 64<<8 || vard < varc)
1292 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1fb4890b 1293 else
1f202b0d 1294 c->scene_change_score+= s->qscale;
1fb4890b 1295
aaaa6f15 1296 if (vard*2 + 200*256 > varc)
bb198e19 1297 mb_type|= CANDIDATE_MB_TYPE_INTRA;
aaaa6f15 1298 if (varc*2 + 200*256 > vard){
bb198e19 1299 mb_type|= CANDIDATE_MB_TYPE_INTER;
af4091f1 1300 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
ca7d05d5
MN
1301 if(s->flags&CODEC_FLAG_MV0)
1302 if(mx || my)
160d679c 1303 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
c60cf138 1304 }else{
1457ab52
MN
1305 mx <<=shift;
1306 my <<=shift;
de6d9b64 1307 }
0d21a846 1308 if((s->flags&CODEC_FLAG_4MV)
aaaa6f15 1309 && !c->skip && varc>50<<8 && vard>10<<8){
bb198e19
MN
1310 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1311 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
0d21a846
MN
1312
1313 set_p_mv_tables(s, mx, my, 0);
1314 }else
1315 set_p_mv_tables(s, mx, my, 1);
bb198e19 1316 if((s->flags&CODEC_FLAG_INTERLACED_ME)
af4091f1 1317 && !c->skip){ //FIXME varc/d checks
f20f8a8b 1318 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
bb198e19
MN
1319 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1320 }
ba6802de 1321 }else{
f7190f73 1322 int intra_score, i;
bb198e19 1323 mb_type= CANDIDATE_MB_TYPE_INTER;
0d21a846 1324
af4091f1 1325 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1f202b0d 1326 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1327 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
67725183
MN
1328
1329 if((s->flags&CODEC_FLAG_4MV)
aaaa6f15 1330 && !c->skip && varc>50<<8 && vard>10<<8){
bb198e19 1331 int dmin4= h263_mv4_search(s, mx, my, shift);
67725183 1332 if(dmin4 < dmin){
bb198e19 1333 mb_type= CANDIDATE_MB_TYPE_INTER4V;
67725183 1334 dmin=dmin4;
e4986da9 1335 }
67725183 1336 }
bb198e19 1337 if((s->flags&CODEC_FLAG_INTERLACED_ME)
af4091f1 1338 && !c->skip){ //FIXME varc/d checks
f20f8a8b 1339 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
bb198e19
MN
1340 if(dmin_i < dmin){
1341 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1342 dmin= dmin_i;
1343 }
1344 }
115329f1
DB
1345
1346// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
bb198e19 1347 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
f7190f73
MN
1348
1349 /* get intra luma score */
1f202b0d 1350 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
aaaa6f15 1351 intra_score= varc - 500;
f7190f73
MN
1352 }else{
1353 int mean= (sum+128)>>8;
1354 mean*= 0x01010101;
115329f1 1355
f7190f73 1356 for(i=0; i<16; i++){
af4091f1
MN
1357 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1358 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1359 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1360 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
f7190f73
MN
1361 }
1362
af4091f1 1363 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
f7190f73
MN
1364 }
1365#if 0 //FIXME
1366 /* get chroma score */
1f202b0d 1367 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
f7190f73
MN
1368 for(i=1; i<3; i++){
1369 uint8_t *dest_c;
1370 int mean;
115329f1 1371
f7190f73 1372 if(s->out_format == FMT_H263){
137c8468 1373 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
f7190f73
MN
1374 }else{
1375 mean= (s->last_dc[i] + 4)>>3;
1376 }
1377 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
115329f1 1378
f7190f73
MN
1379 mean*= 0x01010101;
1380 for(i=0; i<8; i++){
af4091f1
MN
1381 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1382 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
f7190f73 1383 }
115329f1 1384
af4091f1 1385 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
115329f1 1386 }
f7190f73
MN
1387 }
1388#endif
af4091f1 1389 intra_score += c->mb_penalty_factor*16;
115329f1 1390
f7190f73 1391 if(intra_score < dmin){
bb198e19
MN
1392 mb_type= CANDIDATE_MB_TYPE_INTRA;
1393 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
f7190f73
MN
1394 }else
1395 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
115329f1 1396
aaaa6f15
MN
1397 if (vard <= 64<<8 || vard < varc) { //FIXME
1398 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
ba6802de 1399 }else{
1f202b0d 1400 c->scene_change_score+= s->qscale;
ba6802de 1401 }
de6d9b64 1402 }
ba6802de 1403
7bc9090a 1404 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
9dbcbd92
MN
1405}
1406
f5fb6b34
MN
1407int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1408 int mb_x, int mb_y)
1409{
2750b827 1410 MotionEstContext * const c= &s->me;
bb198e19 1411 int mx, my, dmin;
f5fb6b34
MN
1412 int P[10][2];
1413 const int shift= 1+s->quarter_sample;
7bc9090a 1414 const int xy= mb_x + mb_y*s->mb_stride;
af4091f1 1415 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
115329f1 1416
f5fb6b34
MN
1417 assert(s->quarter_sample==0 || s->quarter_sample==1);
1418
26efc54e 1419 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
af4091f1 1420 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
f5fb6b34 1421
bb198e19 1422 get_limits(s, 16*mb_x, 16*mb_y);
af4091f1 1423 c->skip=0;
f5fb6b34
MN
1424
1425 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1426 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1427
af4091f1 1428 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
f5fb6b34
MN
1429
1430 /* special case for first line */
9c3d33d6 1431 if (s->first_slice_line) {
2750b827
MN
1432 c->pred_x= P_LEFT[0];
1433 c->pred_y= P_LEFT[1];
f931ff7b 1434 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
115329f1 1435 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
f5fb6b34 1436 } else {
7bc9090a
MN
1437 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1438 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1439 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1440 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
af4091f1
MN
1441 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1442 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1443 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
115329f1 1444
f5fb6b34
MN
1445 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1446 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1447
2750b827
MN
1448 c->pred_x = P_MEDIAN[0];
1449 c->pred_y = P_MEDIAN[1];
f5fb6b34 1450 }
2750b827 1451
115329f1 1452 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
f931ff7b 1453
f5fb6b34
MN
1454 s->p_mv_table[xy][0] = mx<<shift;
1455 s->p_mv_table[xy][1] = my<<shift;
115329f1 1456
f5fb6b34
MN
1457 return dmin;
1458}
1459
5c91a675 1460static int ff_estimate_motion_b(MpegEncContext * s,
2750b827 1461 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
9dbcbd92 1462{
af4091f1 1463 MotionEstContext * const c= &s->me;
bb198e19 1464 int mx, my, dmin;
0d21a846 1465 int P[10][2];
9dbcbd92 1466 const int shift= 1+s->quarter_sample;
7bc9090a
MN
1467 const int mot_stride = s->mb_stride;
1468 const int mot_xy = mb_y*mot_stride + mb_x;
af4091f1 1469 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
b07a5980 1470 int mv_scale;
115329f1 1471
26efc54e
MN
1472 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1473 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1474 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
af4091f1 1475 c->current_mv_penalty= mv_penalty;
1457ab52 1476
bb198e19 1477 get_limits(s, 16*mb_x, 16*mb_y);
9dbcbd92
MN
1478
1479 switch(s->me_method) {
1480 case ME_ZERO:
1481 default:
bb270c08 1482 no_motion_search(s, &mx, &my);
9dbcbd92 1483 dmin = 0;
0d21a846
MN
1484 mx-= mb_x*16;
1485 my-= mb_y*16;
9dbcbd92 1486 break;
bb198e19 1487#if 0
9dbcbd92 1488 case ME_FULL:
bb270c08 1489 dmin = full_motion_search(s, &mx, &my, range, ref_picture);
0d21a846
MN
1490 mx-= mb_x*16;
1491 my-= mb_y*16;
9dbcbd92
MN
1492 break;
1493 case ME_LOG:
bb270c08 1494 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
0d21a846
MN
1495 mx-= mb_x*16;
1496 my-= mb_y*16;
9dbcbd92
MN
1497 break;
1498 case ME_PHODS:
bb270c08 1499 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
0d21a846
MN
1500 mx-= mb_x*16;
1501 my-= mb_y*16;
9dbcbd92 1502 break;
bb198e19 1503#endif
9dbcbd92
MN
1504 case ME_X1:
1505 case ME_EPZS:
1506 {
0d21a846
MN
1507 P_LEFT[0] = mv_table[mot_xy - 1][0];
1508 P_LEFT[1] = mv_table[mot_xy - 1][1];
9dbcbd92 1509
af4091f1 1510 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
9dbcbd92
MN
1511
1512 /* special case for first line */
9c3d33d6 1513 if (!s->first_slice_line) {
0d21a846
MN
1514 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1515 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1516 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1517 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
af4091f1
MN
1518 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1519 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1520 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 1521
0d21a846
MN
1522 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1523 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
9dbcbd92 1524 }
af4091f1
MN
1525 c->pred_x= P_LEFT[0];
1526 c->pred_y= P_LEFT[1];
9dbcbd92 1527 }
115329f1 1528
b07a5980
MN
1529 if(mv_table == s->b_forw_mv_table){
1530 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1531 }else{
1532 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1533 }
115329f1 1534
26efc54e 1535 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
115329f1 1536
9dbcbd92
MN
1537 break;
1538 }
115329f1 1539
af4091f1 1540 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
115329f1 1541
1f202b0d 1542 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1543 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
67725183 1544
0d21a846 1545//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
9dbcbd92
MN
1546// s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1547 mv_table[mot_xy][0]= mx;
1548 mv_table[mot_xy][1]= my;
1457ab52 1549
91029be7 1550 return dmin;
de6d9b64
FB
1551}
1552
2750b827 1553static inline int check_bidir_mv(MpegEncContext * s,
91029be7
MN
1554 int motion_fx, int motion_fy,
1555 int motion_bx, int motion_by,
1556 int pred_fx, int pred_fy,
bb198e19
MN
1557 int pred_bx, int pred_by,
1558 int size, int h)
de6d9b64 1559{
91029be7 1560 //FIXME optimize?
1457ab52 1561 //FIXME better f_code prediction (max mv & distance)
bb198e19 1562 //FIXME pointers
2750b827 1563 MotionEstContext * const c= &s->me;
377798d6
MN
1564 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1565 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
af4091f1 1566 int stride= c->stride;
af4091f1 1567 uint8_t *dest_y = c->scratchpad;
91029be7
MN
1568 uint8_t *ptr;
1569 int dxy;
1570 int src_x, src_y;
1571 int fbmin;
2750b827
MN
1572 uint8_t **src_data= c->src[0];
1573 uint8_t **ref_data= c->ref[0];
1574 uint8_t **ref2_data= c->ref[2];
91029be7 1575
1457ab52
MN
1576 if(s->quarter_sample){
1577 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
bb198e19
MN
1578 src_x = motion_fx >> 2;
1579 src_y = motion_fy >> 2;
1457ab52 1580
bb198e19
MN
1581 ptr = ref_data[0] + (src_y * stride) + src_x;
1582 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1457ab52
MN
1583
1584 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
bb198e19
MN
1585 src_x = motion_bx >> 2;
1586 src_y = motion_by >> 2;
115329f1 1587
2750b827 1588 ptr = ref2_data[0] + (src_y * stride) + src_x;
bb198e19 1589 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1457ab52
MN
1590 }else{
1591 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
bb198e19
MN
1592 src_x = motion_fx >> 1;
1593 src_y = motion_fy >> 1;
1457ab52 1594
bb198e19
MN
1595 ptr = ref_data[0] + (src_y * stride) + src_x;
1596 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1457ab52
MN
1597
1598 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
bb198e19
MN
1599 src_x = motion_bx >> 1;
1600 src_y = motion_by >> 1;
115329f1 1601
2750b827 1602 ptr = ref2_data[0] + (src_y * stride) + src_x;
bb198e19 1603 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1457ab52
MN
1604 }
1605
377798d6
MN
1606 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1607 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
bb198e19 1608 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
115329f1 1609
1f202b0d 1610 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
67725183
MN
1611 }
1612 //FIXME CHROMA !!!
115329f1 1613
91029be7
MN
1614 return fbmin;
1615}
9dbcbd92 1616
91029be7 1617/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
2750b827 1618static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
91029be7 1619{
316a2ec8 1620 MotionEstContext * const c= &s->me;
7bc9090a
MN
1621 const int mot_stride = s->mb_stride;
1622 const int xy = mb_y *mot_stride + mb_x;
91029be7
MN
1623 int fbmin;
1624 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1625 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1626 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1627 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1628 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1629 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1630 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1631 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
316a2ec8
MN
1632 const int flags= c->sub_flags;
1633 const int qpel= flags&FLAG_QPEL;
1634 const int shift= 1+qpel;
1635 const int xmin= c->xmin<<shift;
1636 const int ymin= c->ymin<<shift;
1637 const int xmax= c->xmax<<shift;
1638 const int ymax= c->ymax<<shift;
8bf755d4
LM
1639 uint8_t map[8][8][8][8];
1640
1641 memset(map,0,sizeof(map));
1642#define BIDIR_MAP(fx,fy,bx,by) \
1643 map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1644 BIDIR_MAP(0,0,0,0) = 1;
115329f1 1645
2750b827 1646 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
91029be7
MN
1647 motion_bx, motion_by,
1648 pred_fx, pred_fy,
bb198e19
MN
1649 pred_bx, pred_by,
1650 0, 16);
91029be7 1651
316a2ec8
MN
1652 if(s->avctx->bidir_refine){
1653 int score, end;
1654#define CHECK_BIDIR(fx,fy,bx,by)\
8bf755d4
LM
1655 if( !BIDIR_MAP(fx,fy,bx,by)\
1656 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
d4f0c2fc 1657 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
8bf755d4 1658 BIDIR_MAP(fx,fy,bx,by) = 1;\
d4f0c2fc
MN
1659 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);\
1660 if(score < fbmin){\
1661 fbmin= score;\
1662 motion_fx+=fx;\
1663 motion_fy+=fy;\
1664 motion_bx+=bx;\
1665 motion_by+=by;\
1666 end=0;\
1667 }\
316a2ec8
MN
1668 }
1669#define CHECK_BIDIR2(a,b,c,d)\
1670CHECK_BIDIR(a,b,c,d)\
1671CHECK_BIDIR(-a,-b,-c,-d)
1672
1673#define CHECK_BIDIRR(a,b,c,d)\
1674CHECK_BIDIR2(a,b,c,d)\
1675CHECK_BIDIR2(b,c,d,a)\
1676CHECK_BIDIR2(c,d,a,b)\
1677CHECK_BIDIR2(d,a,b,c)
1678
1679 do{
1680 end=1;
1681
316a2ec8
MN
1682 CHECK_BIDIRR( 0, 0, 0, 1)
1683 if(s->avctx->bidir_refine > 1){
1684 CHECK_BIDIRR( 0, 0, 1, 1)
1685 CHECK_BIDIR2( 0, 1, 0, 1)
1686 CHECK_BIDIR2( 1, 0, 1, 0)
1687 CHECK_BIDIRR( 0, 0,-1, 1)
1688 CHECK_BIDIR2( 0,-1, 0, 1)
1689 CHECK_BIDIR2(-1, 0, 1, 0)
1690 if(s->avctx->bidir_refine > 2){
1691 CHECK_BIDIRR( 0, 1, 1, 1)
1692 CHECK_BIDIRR( 0,-1, 1, 1)
1693 CHECK_BIDIRR( 0, 1,-1, 1)
1694 CHECK_BIDIRR( 0, 1, 1,-1)
1695 if(s->avctx->bidir_refine > 3){
1696 CHECK_BIDIR2( 1, 1, 1, 1)
1697 CHECK_BIDIRR( 1, 1, 1,-1)
1698 CHECK_BIDIR2( 1, 1,-1,-1)
1699 CHECK_BIDIR2( 1,-1,-1, 1)
1700 CHECK_BIDIR2( 1,-1, 1,-1)
1701 }
1702 }
1703 }
1704 }while(!end);
1705 }
1706
8bf755d4
LM
1707 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1708 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1709 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1710 s->b_bidir_back_mv_table[xy][1]= motion_by;
1711
316a2ec8 1712 return fbmin;
91029be7
MN
1713}
1714
2750b827 1715static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
91029be7 1716{
af4091f1 1717 MotionEstContext * const c= &s->me;
0d21a846 1718 int P[10][2];
7bc9090a
MN
1719 const int mot_stride = s->mb_stride;
1720 const int mot_xy = mb_y*mot_stride + mb_x;
1457ab52
MN
1721 const int shift= 1+s->quarter_sample;
1722 int dmin, i;
91029be7 1723 const int time_pp= s->pp_time;
bea669e5 1724 const int time_pb= s->pb_time;
1457ab52 1725 int mx, my, xmin, xmax, ymin, ymax;
91029be7 1726 int16_t (*mv_table)[2]= s->b_direct_mv_table;
115329f1 1727
af4091f1 1728 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1457ab52
MN
1729 ymin= xmin=(-32)>>shift;
1730 ymax= xmax= 31>>shift;
1731
7bc9090a 1732 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1457ab52
MN
1733 s->mv_type= MV_TYPE_8X8;
1734 }else{
1735 s->mv_type= MV_TYPE_16X16;
91029be7 1736 }
1457ab52
MN
1737
1738 for(i=0; i<4; i++){
1739 int index= s->block_index[i];
1740 int min, max;
115329f1 1741
af4091f1
MN
1742 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1743 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1744 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1745 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1746// c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1747// c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1748
1749 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1750 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
5a603607
MN
1751 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1752 min+= 16*mb_x - 1;
4a711c33
MN
1753 xmax= FFMIN(xmax, s->width - max);
1754 xmin= FFMAX(xmin, - 16 - min);
1457ab52 1755
af4091f1
MN
1756 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1757 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
5a603607
MN
1758 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1759 min+= 16*mb_y - 1;
4a711c33
MN
1760 ymax= FFMIN(ymax, s->height - max);
1761 ymin= FFMAX(ymin, - 16 - min);
115329f1 1762
1457ab52 1763 if(s->mv_type == MV_TYPE_16X16) break;
91029be7 1764 }
115329f1 1765
1457ab52 1766 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
115329f1 1767
1457ab52
MN
1768 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1769 s->b_direct_mv_table[mot_xy][0]= 0;
1770 s->b_direct_mv_table[mot_xy][1]= 0;
1771
1772 return 256*256*256*64;
1773 }
115329f1 1774
af4091f1
MN
1775 c->xmin= xmin;
1776 c->ymin= ymin;
1777 c->xmax= xmax;
1778 c->ymax= ymax;
1779 c->flags |= FLAG_DIRECT;
1780 c->sub_flags |= FLAG_DIRECT;
1781 c->pred_x=0;
1782 c->pred_y=0;
1457ab52 1783
0a13093d
MN
1784 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1785 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1786
1787 /* special case for first line */
9c3d33d6 1788 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
0a13093d
MN
1789 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1790 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1791 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1792 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
115329f1 1793
0a13093d
MN
1794 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1795 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1796 }
115329f1 1797
26efc54e 1798 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
115329f1 1799 if(c->sub_flags&FLAG_QPEL)
2750b827
MN
1800 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1801 else
1802 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
115329f1 1803
1f202b0d 1804 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1805 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
115329f1 1806
af4091f1 1807 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
91029be7
MN
1808
1809 s->b_direct_mv_table[mot_xy][0]= mx;
1810 s->b_direct_mv_table[mot_xy][1]= my;
af4091f1
MN
1811 c->flags &= ~FLAG_DIRECT;
1812 c->sub_flags &= ~FLAG_DIRECT;
2750b827 1813
91029be7 1814 return dmin;
9dbcbd92
MN
1815}
1816
1817void ff_estimate_b_frame_motion(MpegEncContext * s,
1818 int mb_x, int mb_y)
1819{
af4091f1
MN
1820 MotionEstContext * const c= &s->me;
1821 const int penalty_factor= c->mb_penalty_factor;
bb198e19 1822 int fmin, bmin, dmin, fbmin, bimin, fimin;
91029be7 1823 int type=0;
f20f8a8b 1824 const int xy = mb_y*s->mb_stride + mb_x;
af4091f1 1825 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
f20f8a8b 1826
ae1dbde1 1827 get_limits(s, 16*mb_x, 16*mb_y);
115329f1 1828
af4091f1 1829 c->skip=0;
1f202b0d 1830 if(c->avctx->me_threshold){
aaaa6f15 1831 int vard= check_input_motion(s, mb_x, mb_y, 0);
115329f1 1832
aaaa6f15 1833 if((vard+128)>>8 < c->avctx->me_threshold){
f4f3223f
MN
1834// pix = c->src[0][0];
1835// sum = s->dsp.pix_sum(pix, s->linesize);
aaaa6f15 1836// varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
115329f1 1837
aaaa6f15
MN
1838// pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1839 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
f4f3223f 1840/* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
aaaa6f15
MN
1841 c->mb_var_sum_temp += (varc+128)>>8;*/
1842 c->mc_mb_var_sum_temp += (vard+128)>>8;
1843/* if (vard <= 64<<8 || vard < varc) {
1844 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
f4f3223f 1845 }else{
1f202b0d 1846 c->scene_change_score+= s->qscale;
f4f3223f
MN
1847 }*/
1848 return;
1849 }
aaaa6f15 1850 if((vard+128)>>8 < c->avctx->mb_threshold){
f20f8a8b
MN
1851 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1852 if(type == CANDIDATE_MB_TYPE_DIRECT){
1853 direct_search(s, mb_x, mb_y);
1854 }
1855 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
af4091f1 1856 c->skip=0;
f20f8a8b
MN
1857 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1858 }
1859 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
af4091f1 1860 c->skip=0;
f20f8a8b
MN
1861 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1862 }
1863 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
af4091f1
MN
1864 c->skip=0;
1865 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
f20f8a8b
MN
1866 interlaced_search(s, 0,
1867 s->b_field_mv_table[0], s->b_field_select_table[0],
1868 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1869 }
1870 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
af4091f1
MN
1871 c->skip=0;
1872 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
f20f8a8b
MN
1873 interlaced_search(s, 2,
1874 s->b_field_mv_table[1], s->b_field_select_table[1],
1875 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1876 }
1877 return;
1878 }
f4f3223f
MN
1879 }
1880
2d5e962b 1881 if (s->codec_id == CODEC_ID_MPEG4)
2750b827 1882 dmin= direct_search(s, mb_x, mb_y);
2d5e962b
MN
1883 else
1884 dmin= INT_MAX;
bb198e19 1885//FIXME penalty stuff for non mpeg4
af4091f1 1886 c->skip=0;
2750b827 1887 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
115329f1 1888
af4091f1 1889 c->skip=0;
2750b827 1890 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
9dbcbd92 1891//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
91029be7 1892
af4091f1 1893 c->skip=0;
2750b827 1894 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
67725183 1895//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
115329f1 1896
bb198e19 1897 if(s->flags & CODEC_FLAG_INTERLACED_ME){
bb198e19 1898//FIXME mb type penalty
af4091f1
MN
1899 c->skip=0;
1900 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
2750b827
MN
1901 fimin= interlaced_search(s, 0,
1902 s->b_field_mv_table[0], s->b_field_select_table[0],
f20f8a8b 1903 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
af4091f1 1904 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
2750b827
MN
1905 bimin= interlaced_search(s, 2,
1906 s->b_field_mv_table[1], s->b_field_select_table[1],
f20f8a8b 1907 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
bb198e19
MN
1908 }else
1909 fimin= bimin= INT_MAX;
1910
3aa102be 1911 {
63b15e55 1912 int score= fmin;
bb198e19 1913 type = CANDIDATE_MB_TYPE_FORWARD;
115329f1 1914
2d5e962b 1915 if (dmin <= score){
63b15e55 1916 score = dmin;
bb198e19 1917 type = CANDIDATE_MB_TYPE_DIRECT;
91029be7
MN
1918 }
1919 if(bmin<score){
1920 score=bmin;
115329f1 1921 type= CANDIDATE_MB_TYPE_BACKWARD;
91029be7
MN
1922 }
1923 if(fbmin<score){
1924 score=fbmin;
bb198e19
MN
1925 type= CANDIDATE_MB_TYPE_BIDIR;
1926 }
1927 if(fimin<score){
1928 score=fimin;
1929 type= CANDIDATE_MB_TYPE_FORWARD_I;
1930 }
1931 if(bimin<score){
1932 score=bimin;
1933 type= CANDIDATE_MB_TYPE_BACKWARD_I;
91029be7 1934 }
115329f1 1935
0bfacb95 1936 score= ((unsigned)(score*score + 128*256))>>16;
1f202b0d 1937 c->mc_mb_var_sum_temp += score;
7bc9090a 1938 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
91029be7 1939 }
3aa102be 1940
1f202b0d 1941 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
bb198e19
MN
1942 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1943 if(fimin < INT_MAX)
1944 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1945 if(bimin < INT_MAX)
1946 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1947 if(fimin < INT_MAX && bimin < INT_MAX){
1948 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1949 }
1950 //FIXME something smarter
1951 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
115329f1 1952#if 0
622348f9
MN
1953 if(s->out_format == FMT_MPEG1)
1954 type |= CANDIDATE_MB_TYPE_INTRA;
1955#endif
3aa102be
MN
1956 }
1957
7bc9090a 1958 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
9dbcbd92
MN
1959}
1960
1961/* find best f_code for ME which do unlimited searches */
1962int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1963{
9dbcbd92 1964 if(s->me_method>=ME_EPZS){
0d21a846 1965 int score[8];
0cc64d3d 1966 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
0c1a9eda 1967 uint8_t * fcode_tab= s->fcode_tab;
0d21a846
MN
1968 int best_fcode=-1;
1969 int best_score=-10000000;
9dbcbd92 1970
115329f1 1971 if(s->msmpeg4_version)
2f300f89
MN
1972 range= FFMIN(range, 16);
1973 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1974 range= FFMIN(range, 256);
1975
1457ab52 1976 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
9dbcbd92
MN
1977
1978 for(y=0; y<s->mb_height; y++){
1979 int x;
7bc9090a 1980 int xy= y*s->mb_stride;
9dbcbd92 1981 for(x=0; x<s->mb_width; x++){
d3985de7 1982 if(s->mb_type[xy] & type){
7494cac0
MN
1983 int mx= mv_table[xy][0];
1984 int my= mv_table[xy][1];
1985 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1986 fcode_tab[my + MAX_MV]);
0d21a846 1987 int j;
115329f1
DB
1988
1989 if(mx >= range || mx < -range ||
7494cac0
MN
1990 my >= range || my < -range)
1991 continue;
115329f1 1992
0d21a846 1993 for(j=0; j<fcode && j<8; j++){
7bc9090a 1994 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
0d21a846
MN
1995 score[j]-= 170;
1996 }
9dbcbd92 1997 }
9dbcbd92
MN
1998 xy++;
1999 }
2000 }
115329f1 2001
0d21a846
MN
2002 for(i=1; i<8; i++){
2003 if(score[i] > best_score){
2004 best_score= score[i];
2005 best_fcode= i;
2006 }
2007// printf("%d %d\n", i, score[i]);
9dbcbd92 2008 }
0d21a846 2009
9dbcbd92 2010// printf("fcode: %d type: %d\n", i, s->pict_type);
0d21a846 2011 return best_fcode;
9dbcbd92
MN
2012/* for(i=0; i<=MAX_FCODE; i++){
2013 printf("%d ", mv_num[i]);
2014 }
2015 printf("\n");*/
2016 }else{
2017 return 1;
de6d9b64
FB
2018 }
2019}
2020
9dbcbd92
MN
2021void ff_fix_long_p_mvs(MpegEncContext * s)
2022{
1f202b0d 2023 MotionEstContext * const c= &s->me;
9dbcbd92 2024 const int f_code= s->f_code;
ebbcdc9a 2025 int y, range;
029911d1 2026 assert(s->pict_type==P_TYPE);
ebbcdc9a 2027
2f300f89
MN
2028 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2029
2030 assert(range <= 16 || !s->msmpeg4_version);
2031 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
115329f1 2032
1f202b0d 2033 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
115329f1 2034
0d21a846 2035//printf("%d no:%d %d//\n", clip, noclip, f_code);
9dbcbd92 2036 if(s->flags&CODEC_FLAG_4MV){
137c8468 2037 const int wrap= s->b8_stride;
9dbcbd92
MN
2038
2039 /* clip / convert to intra 8x8 type MVs */
2040 for(y=0; y<s->mb_height; y++){
137c8468 2041 int xy= y*2*wrap;
7bc9090a 2042 int i= y*s->mb_stride;
9dbcbd92
MN
2043 int x;
2044
2045 for(x=0; x<s->mb_width; x++){
bb198e19 2046 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
9dbcbd92
MN
2047 int block;
2048 for(block=0; block<4; block++){
2049 int off= (block& 1) + (block>>1)*wrap;
8d7ec294
WH
2050 int mx= s->current_picture.motion_val[0][ xy + off ][0];
2051 int my= s->current_picture.motion_val[0][ xy + off ][1];
9dbcbd92 2052
ebbcdc9a
MN
2053 if( mx >=range || mx <-range
2054 || my >=range || my <-range){
bb198e19
MN
2055 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2056 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2057 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
9dbcbd92
MN
2058 }
2059 }
9dbcbd92 2060 }
f10bd870
MN
2061 xy+=2;
2062 i++;
9dbcbd92
MN
2063 }
2064 }
2065 }
2066}
2067
bb198e19
MN
2068/**
2069 *
2070 * @param truncate 1 for truncation, 0 for using intra
2071 */
115329f1 2072void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
bb198e19 2073 int16_t (*mv_table)[2], int f_code, int type, int truncate)
9dbcbd92 2074{
1f202b0d 2075 MotionEstContext * const c= &s->me;
bb198e19 2076 int y, h_range, v_range;
9dbcbd92 2077
63b15e55 2078 // RAL: 8 in MPEG-1, 16 in MPEG-4
2f300f89 2079 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
bb198e19 2080
1f202b0d 2081 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
63b15e55 2082
bb198e19
MN
2083 h_range= range;
2084 v_range= field_select_table ? range>>1 : range;
2085
9dbcbd92
MN
2086 /* clip / convert to intra 16x16 type MVs */
2087 for(y=0; y<s->mb_height; y++){
2088 int x;
7bc9090a 2089 int xy= y*s->mb_stride;
ebbcdc9a 2090 for(x=0; x<s->mb_width; x++){
7bc9090a 2091 if (s->mb_type[xy] & type){ // RAL: "type" test added...
bb198e19
MN
2092 if(field_select_table==NULL || field_select_table[xy] == field_select){
2093 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2094 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2095
2096 if(truncate){
2097 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
2098 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2099 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
2100 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2101 }else{
2102 s->mb_type[xy] &= ~type;
2103 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2104 mv_table[xy][0]=
2105 mv_table[xy][1]= 0;
2106 }
63b15e55 2107 }
ebbcdc9a 2108 }
9dbcbd92
MN
2109 }
2110 xy++;
9dbcbd92
MN
2111 }
2112 }
2113}