summaryrefslogtreecommitdiff
path: root/libavcodec/motion_est.c (plain)
blob: 25b606f81957c1e126e6309ac119541baa0df40c
1/*
2 * Motion estimation
3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
5 *
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
7 *
8 * This file is part of FFmpeg.
9 *
10 * FFmpeg is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * FFmpeg is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with FFmpeg; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 */
24
25/**
26 * @file
27 * Motion estimation.
28 */
29
30#include <stdlib.h>
31#include <stdio.h>
32#include <limits.h>
33
34#include "avcodec.h"
35#include "internal.h"
36#include "mathops.h"
37#include "motion_est.h"
38#include "mpegutils.h"
39#include "mpegvideo.h"
40
41#define P_LEFT P[1]
42#define P_TOP P[2]
43#define P_TOPRIGHT P[3]
44#define P_MEDIAN P[4]
45#define P_MV1 P[9]
46
47#define ME_MAP_SHIFT 3
48#define ME_MAP_MV_BITS 11
49
50static int sad_hpel_motion_search(MpegEncContext * s,
51 int *mx_ptr, int *my_ptr, int dmin,
52 int src_index, int ref_index,
53 int size, int h);
54
55static inline unsigned update_map_generation(MotionEstContext *c)
56{
57 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
58 if(c->map_generation==0){
59 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
60 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
61 }
62 return c->map_generation;
63}
64
65/* shape adaptive search stuff */
66typedef struct Minima{
67 int height;
68 int x, y;
69 int checked;
70}Minima;
71
72static int minima_cmp(const void *a, const void *b){
73 const Minima *da = (const Minima *) a;
74 const Minima *db = (const Minima *) b;
75
76 return da->height - db->height;
77}
78
79#define FLAG_QPEL 1 //must be 1
80#define FLAG_CHROMA 2
81#define FLAG_DIRECT 4
82
83static 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){
84 const int offset[3]= {
85 y*c-> stride + x,
86 ((y*c->uvstride + x)>>1),
87 ((y*c->uvstride + x)>>1),
88 };
89 int i;
90 for(i=0; i<3; i++){
91 c->src[0][i]= src [i] + offset[i];
92 c->ref[0][i]= ref [i] + offset[i];
93 }
94 if(ref_index){
95 for(i=0; i<3; i++){
96 c->ref[ref_index][i]= ref2[i] + offset[i];
97 }
98 }
99}
100
101static int get_flags(MotionEstContext *c, int direct, int chroma){
102 return ((c->avctx->flags&AV_CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
103 + (direct ? FLAG_DIRECT : 0)
104 + (chroma ? FLAG_CHROMA : 0);
105}
106
107static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
108 const int size, const int h, int ref_index, int src_index,
109 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
110 MotionEstContext * const c= &s->me;
111 const int stride= c->stride;
112 const int hx= subx + (x<<(1+qpel));
113 const int hy= suby + (y<<(1+qpel));
114 uint8_t * const * const ref= c->ref[ref_index];
115 uint8_t * const * const src= c->src[src_index];
116 int d;
117 //FIXME check chroma 4mv, (no crashes ...)
118 av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
119 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
120 const int time_pp= s->pp_time;
121 const int time_pb= s->pb_time;
122 const int mask= 2*qpel+1;
123 if(s->mv_type==MV_TYPE_8X8){
124 int i;
125 for(i=0; i<4; i++){
126 int fx = c->direct_basis_mv[i][0] + hx;
127 int fy = c->direct_basis_mv[i][1] + hy;
128 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));
129 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));
130 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
131 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
132
133 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
134 if(qpel){
135 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
136 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
137 }else{
138 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
139 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
140 }
141 }
142 }else{
143 int fx = c->direct_basis_mv[0][0] + hx;
144 int fy = c->direct_basis_mv[0][1] + hy;
145 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
146 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
147 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
148 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
149
150 if(qpel){
151 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
152 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
153 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
154 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
155 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
156 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
157 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
158 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
159 }else{
160 av_assert2((fx>>1) + 16*s->mb_x >= -16);
161 av_assert2((fy>>1) + 16*s->mb_y >= -16);
162 av_assert2((fx>>1) + 16*s->mb_x <= s->width);
163 av_assert2((fy>>1) + 16*s->mb_y <= s->height);
164 av_assert2((bx>>1) + 16*s->mb_x >= -16);
165 av_assert2((by>>1) + 16*s->mb_y >= -16);
166 av_assert2((bx>>1) + 16*s->mb_x <= s->width);
167 av_assert2((by>>1) + 16*s->mb_y <= s->height);
168
169 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
170 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
171 }
172 }
173 d = cmp_func(s, c->temp, src[0], stride, 16);
174 }else
175 d= 256*256*256*32;
176 return d;
177}
178
179static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
180 const int size, const int h, int ref_index, int src_index,
181 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
182 MotionEstContext * const c= &s->me;
183 const int stride= c->stride;
184 const int uvstride= c->uvstride;
185 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
186 const int hx= subx + x*(1<<(1+qpel));
187 const int hy= suby + y*(1<<(1+qpel));
188 uint8_t * const * const ref= c->ref[ref_index];
189 uint8_t * const * const src= c->src[src_index];
190 int d;
191 //FIXME check chroma 4mv, (no crashes ...)
192 int uvdxy; /* no, it might not be used uninitialized */
193 if(dxy){
194 if(qpel){
195 if (h << size == 16) {
196 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
197 } else if (size == 0 && h == 8) {
198 c->qpel_put[1][dxy](c->temp , ref[0] + x + y*stride , stride);
199 c->qpel_put[1][dxy](c->temp + 8, ref[0] + x + y*stride + 8, stride);
200 } else
201 av_assert2(0);
202 if(chroma){
203 int cx= hx/2;
204 int cy= hy/2;
205 cx= (cx>>1)|(cx&1);
206 cy= (cy>>1)|(cy&1);
207 uvdxy= (cx&1) + 2*(cy&1);
208 // FIXME x/y wrong, but MPEG-4 qpel is sick anyway, we should drop as much of it as possible in favor for H.264
209 }
210 }else{
211 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
212 if(chroma)
213 uvdxy= dxy | (x&1) | (2*(y&1));
214 }
215 d = cmp_func(s, c->temp, src[0], stride, h);
216 }else{
217 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
218 if(chroma)
219 uvdxy= (x&1) + 2*(y&1);
220 }
221 if(chroma){
222 uint8_t * const uvtemp= c->temp + 16*stride;
223 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
224 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
225 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
226 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
227 }
228 return d;
229}
230
231static int cmp_simple(MpegEncContext *s, const int x, const int y,
232 int ref_index, int src_index,
233 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
234 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
235}
236
237static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
238 const int size, const int h, int ref_index, int src_index,
239 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
240 if(flags&FLAG_DIRECT){
241 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
242 }else{
243 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
244 }
245}
246
247static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
248 const int size, const int h, int ref_index, int src_index,
249 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
250 if(flags&FLAG_DIRECT){
251 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
252 }else{
253 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA);
254 }
255}
256
257/** @brief compares a block (either a full macroblock or a partition thereof)
258 against a proposed motion-compensated prediction of that block
259 */
260static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
261 const int size, const int h, int ref_index, int src_index,
262 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
263 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
264 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
265 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
266 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
267 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
268 && subx==0 && suby==0){
269 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
270 }else{
271 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
272 }
273}
274
275static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
276 const int size, const int h, int ref_index, int src_index,
277 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
278 if(flags&FLAG_DIRECT){
279 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
280 }else{
281 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
282 }
283}
284
285static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
286 const int size, const int h, int ref_index, int src_index,
287 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
288 if(flags&FLAG_DIRECT){
289 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
290 }else{
291 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
292 }
293}
294
295#include "motion_est_template.c"
296
297static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
298 ptrdiff_t stride, int h)
299{
300 return 0;
301}
302
303static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
304}
305
306int ff_init_me(MpegEncContext *s){
307 MotionEstContext * const c= &s->me;
308 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
309 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
310
311 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
312 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
313 return -1;
314 }
315
316#if FF_API_MOTION_EST
317 //special case of snow is needed because snow uses its own iterative ME code
318FF_DISABLE_DEPRECATION_WARNINGS
319 if (s->motion_est == FF_ME_EPZS) {
320 if (s->me_method == ME_ZERO)
321 s->motion_est = FF_ME_ZERO;
322 else if (s->me_method == ME_EPZS)
323 s->motion_est = FF_ME_EPZS;
324 else if (s->me_method == ME_X1)
325 s->motion_est = FF_ME_XONE;
326 else if (s->avctx->codec_id != AV_CODEC_ID_SNOW) {
327 av_log(s->avctx, AV_LOG_ERROR,
328 "me_method is only allowed to be set to zero and epzs; "
329 "for hex,umh,full and others see dia_size\n");
330 return -1;
331 }
332 }
333FF_ENABLE_DEPRECATION_WARNINGS
334#endif
335
336 c->avctx= s->avctx;
337
338 if(s->codec_id == AV_CODEC_ID_H261)
339 c->avctx->me_sub_cmp = c->avctx->me_cmp;
340
341 if(cache_size < 2*dia_size && !c->stride){
342 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
343 }
344
345 ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
346 ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
347 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
348 ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
349
350 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
351 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
352 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
353
354/*FIXME s->no_rounding b_type*/
355 if (s->avctx->flags & AV_CODEC_FLAG_QPEL) {
356 c->sub_motion_search= qpel_motion_search;
357 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
358 if (s->no_rounding)
359 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
360 else
361 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
362 }else{
363 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
364 c->sub_motion_search= hpel_motion_search;
365 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
366 && c->avctx-> me_cmp == FF_CMP_SAD
367 && c->avctx-> mb_cmp == FF_CMP_SAD)
368 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
369 else
370 c->sub_motion_search= hpel_motion_search;
371 }
372 c->hpel_avg = s->hdsp.avg_pixels_tab;
373 if (s->no_rounding)
374 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
375 else
376 c->hpel_put = s->hdsp.put_pixels_tab;
377
378 if(s->linesize){
379 c->stride = s->linesize;
380 c->uvstride= s->uvlinesize;
381 }else{
382 c->stride = 16*s->mb_width + 32;
383 c->uvstride= 8*s->mb_width + 16;
384 }
385
386 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
387 * not have yet, and even if we had, the motion estimation code
388 * does not expect it. */
389 if (s->codec_id != AV_CODEC_ID_SNOW) {
390 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
391 s->mecc.me_cmp[2] = zero_cmp;
392 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
393 s->mecc.me_sub_cmp[2] = zero_cmp;
394 c->hpel_put[2][0]= c->hpel_put[2][1]=
395 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
396 }
397
398 if(s->codec_id == AV_CODEC_ID_H261){
399 c->sub_motion_search= no_sub_motion_search;
400 }
401
402 return 0;
403}
404
405#define CHECK_SAD_HALF_MV(suffix, x, y) \
406{\
407 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
408 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
409 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
410}
411
412static int sad_hpel_motion_search(MpegEncContext * s,
413 int *mx_ptr, int *my_ptr, int dmin,
414 int src_index, int ref_index,
415 int size, int h)
416{
417 MotionEstContext * const c= &s->me;
418 const int penalty_factor= c->sub_penalty_factor;
419 int mx, my, dminh;
420 uint8_t *pix, *ptr;
421 int stride= c->stride;
422 LOAD_COMMON
423
424 av_assert2(c->sub_flags == 0);
425
426 if(c->skip){
427 *mx_ptr = 0;
428 *my_ptr = 0;
429 return dmin;
430 }
431
432 pix = c->src[src_index][0];
433
434 mx = *mx_ptr;
435 my = *my_ptr;
436 ptr = c->ref[ref_index][0] + (my * stride) + mx;
437
438 dminh = dmin;
439
440 if (mx > xmin && mx < xmax &&
441 my > ymin && my < ymax) {
442 int dx=0, dy=0;
443 int d, pen_x, pen_y;
444 const int index= my*(1<<ME_MAP_SHIFT) + mx;
445 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
446 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
447 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
448 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
449 mx += mx;
450 my += my;
451
452
453 pen_x= pred_x + mx;
454 pen_y= pred_y + my;
455
456 ptr-= stride;
457 if(t<=b){
458 CHECK_SAD_HALF_MV(y2 , 0, -1)
459 if(l<=r){
460 CHECK_SAD_HALF_MV(xy2, -1, -1)
461 if(t+r<=b+l){
462 CHECK_SAD_HALF_MV(xy2, +1, -1)
463 ptr+= stride;
464 }else{
465 ptr+= stride;
466 CHECK_SAD_HALF_MV(xy2, -1, +1)
467 }
468 CHECK_SAD_HALF_MV(x2 , -1, 0)
469 }else{
470 CHECK_SAD_HALF_MV(xy2, +1, -1)
471 if(t+l<=b+r){
472 CHECK_SAD_HALF_MV(xy2, -1, -1)
473 ptr+= stride;
474 }else{
475 ptr+= stride;
476 CHECK_SAD_HALF_MV(xy2, +1, +1)
477 }
478 CHECK_SAD_HALF_MV(x2 , +1, 0)
479 }
480 }else{
481 if(l<=r){
482 if(t+l<=b+r){
483 CHECK_SAD_HALF_MV(xy2, -1, -1)
484 ptr+= stride;
485 }else{
486 ptr+= stride;
487 CHECK_SAD_HALF_MV(xy2, +1, +1)
488 }
489 CHECK_SAD_HALF_MV(x2 , -1, 0)
490 CHECK_SAD_HALF_MV(xy2, -1, +1)
491 }else{
492 if(t+r<=b+l){
493 CHECK_SAD_HALF_MV(xy2, +1, -1)
494 ptr+= stride;
495 }else{
496 ptr+= stride;
497 CHECK_SAD_HALF_MV(xy2, -1, +1)
498 }
499 CHECK_SAD_HALF_MV(x2 , +1, 0)
500 CHECK_SAD_HALF_MV(xy2, +1, +1)
501 }
502 CHECK_SAD_HALF_MV(y2 , 0, +1)
503 }
504 mx+=dx;
505 my+=dy;
506
507 }else{
508 mx += mx;
509 my += my;
510 }
511
512 *mx_ptr = mx;
513 *my_ptr = my;
514 return dminh;
515}
516
517static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
518{
519 const int xy= s->mb_x + s->mb_y*s->mb_stride;
520
521 s->p_mv_table[xy][0] = mx;
522 s->p_mv_table[xy][1] = my;
523
524 /* has already been set to the 4 MV if 4MV is done */
525 if(mv4){
526 int mot_xy= s->block_index[0];
527
528 s->current_picture.motion_val[0][mot_xy ][0] = mx;
529 s->current_picture.motion_val[0][mot_xy ][1] = my;
530 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
531 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
532
533 mot_xy += s->b8_stride;
534 s->current_picture.motion_val[0][mot_xy ][0] = mx;
535 s->current_picture.motion_val[0][mot_xy ][1] = my;
536 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
537 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
538 }
539}
540
541/**
542 * get fullpel ME search limits.
543 */
544static inline void get_limits(MpegEncContext *s, int x, int y)
545{
546 MotionEstContext * const c= &s->me;
547 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
548 int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
549/*
550 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
551 else c->range= 16;
552*/
553 if (s->unrestricted_mv) {
554 c->xmin = - x - 16;
555 c->ymin = - y - 16;
556 c->xmax = - x + s->width;
557 c->ymax = - y + s->height;
558 } else if (s->out_format == FMT_H261){
559 // Search range of H.261 is different from other codec standards
560 c->xmin = (x > 15) ? - 15 : 0;
561 c->ymin = (y > 15) ? - 15 : 0;
562 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
563 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
564 } else {
565 c->xmin = - x;
566 c->ymin = - y;
567 c->xmax = - x + s->mb_width *16 - 16;
568 c->ymax = - y + s->mb_height*16 - 16;
569 }
570 if(!range || range > max_range)
571 range = max_range;
572 if(range){
573 c->xmin = FFMAX(c->xmin,-range);
574 c->xmax = FFMIN(c->xmax, range);
575 c->ymin = FFMAX(c->ymin,-range);
576 c->ymax = FFMIN(c->ymax, range);
577 }
578}
579
580static inline void init_mv4_ref(MotionEstContext *c){
581 const int stride= c->stride;
582
583 c->ref[1][0] = c->ref[0][0] + 8;
584 c->ref[2][0] = c->ref[0][0] + 8*stride;
585 c->ref[3][0] = c->ref[2][0] + 8;
586 c->src[1][0] = c->src[0][0] + 8;
587 c->src[2][0] = c->src[0][0] + 8*stride;
588 c->src[3][0] = c->src[2][0] + 8;
589}
590
591static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
592{
593 MotionEstContext * const c= &s->me;
594 const int size= 1;
595 const int h=8;
596 int block;
597 int P[10][2];
598 int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
599 int same=1;
600 const int stride= c->stride;
601 uint8_t *mv_penalty= c->current_mv_penalty;
602 int safety_clipping= s->unrestricted_mv && (s->width&15) && (s->height&15);
603
604 init_mv4_ref(c);
605
606 for(block=0; block<4; block++){
607 int mx4, my4;
608 int pred_x4, pred_y4;
609 int dmin4;
610 static const int off[4]= {2, 1, 1, -1};
611 const int mot_stride = s->b8_stride;
612 const int mot_xy = s->block_index[block];
613
614 if(safety_clipping){
615 c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
616 c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
617 }
618
619 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
620 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
621
622 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
623
624 /* special case for first line */
625 if (s->first_slice_line && block<2) {
626 c->pred_x= pred_x4= P_LEFT[0];
627 c->pred_y= pred_y4= P_LEFT[1];
628 } else {
629 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
630 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
631 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
632 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
633 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
634 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
635 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
636 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
637
638 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
639 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
640
641 c->pred_x= pred_x4 = P_MEDIAN[0];
642 c->pred_y= pred_y4 = P_MEDIAN[1];
643 }
644 P_MV1[0]= mx;
645 P_MV1[1]= my;
646 if(safety_clipping)
647 for(i=1; i<10; i++){
648 if (s->first_slice_line && block<2 && i>1 && i<9)
649 continue;
650 if (i>4 && i<9)
651 continue;
652 if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift);
653 if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift);
654 }
655
656 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
657
658 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
659
660 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
661 int dxy;
662 const int offset= ((block&1) + (block>>1)*stride)*8;
663 uint8_t *dest_y = c->scratchpad + offset;
664 if(s->quarter_sample){
665 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
666 dxy = ((my4 & 3) << 2) | (mx4 & 3);
667
668 if(s->no_rounding)
669 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
670 else
671 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
672 }else{
673 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
674 dxy = ((my4 & 1) << 1) | (mx4 & 1);
675
676 if(s->no_rounding)
677 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
678 else
679 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
680 }
681 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
682 }else
683 dmin_sum+= dmin4;
684
685 if(s->quarter_sample){
686 mx4_sum+= mx4/2;
687 my4_sum+= my4/2;
688 }else{
689 mx4_sum+= mx4;
690 my4_sum+= my4;
691 }
692
693 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
694 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
695
696 if(mx4 != mx || my4 != my) same=0;
697 }
698
699 if(same)
700 return INT_MAX;
701
702 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
703 dmin_sum += s->mecc.mb_cmp[0](s,
704 s->new_picture.f->data[0] +
705 s->mb_x * 16 + s->mb_y * 16 * stride,
706 c->scratchpad, stride, 16);
707 }
708
709 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
710 int dxy;
711 int mx, my;
712 int offset;
713
714 mx= ff_h263_round_chroma(mx4_sum);
715 my= ff_h263_round_chroma(my4_sum);
716 dxy = ((my & 1) << 1) | (mx & 1);
717
718 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
719
720 if(s->no_rounding){
721 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
722 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
723 }else{
724 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
725 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
726 }
727
728 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[1] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad, s->uvlinesize, 8);
729 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[2] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad + 8, s->uvlinesize, 8);
730 }
731
732 c->pred_x= mx;
733 c->pred_y= my;
734
735 switch(c->avctx->mb_cmp&0xFF){
736 /*case FF_CMP_SSE:
737 return dmin_sum+ 32*s->qscale*s->qscale;*/
738 case FF_CMP_RD:
739 return dmin_sum;
740 default:
741 return dmin_sum+ 11*c->mb_penalty_factor;
742 }
743}
744
745static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
746 MotionEstContext * const c= &s->me;
747
748 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
749 c->src[1][0] = c->src[0][0] + s->linesize;
750 if(c->flags & FLAG_CHROMA){
751 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
752 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
753 c->src[1][1] = c->src[0][1] + s->uvlinesize;
754 c->src[1][2] = c->src[0][2] + s->uvlinesize;
755 }
756}
757
758static int interlaced_search(MpegEncContext *s, int ref_index,
759 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
760{
761 MotionEstContext * const c= &s->me;
762 const int size=0;
763 const int h=8;
764 int block;
765 int P[10][2];
766 uint8_t * const mv_penalty= c->current_mv_penalty;
767 int same=1;
768 const int stride= 2*s->linesize;
769 int dmin_sum= 0;
770 const int mot_stride= s->mb_stride;
771 const int xy= s->mb_x + s->mb_y*mot_stride;
772
773 c->ymin>>=1;
774 c->ymax>>=1;
775 c->stride<<=1;
776 c->uvstride<<=1;
777 init_interlaced_ref(s, ref_index);
778
779 for(block=0; block<2; block++){
780 int field_select;
781 int best_dmin= INT_MAX;
782 int best_field= -1;
783
784 for(field_select=0; field_select<2; field_select++){
785 int dmin, mx_i, my_i;
786 int16_t (*mv_table)[2]= mv_tables[block][field_select];
787
788 if(user_field_select){
789 av_assert1(field_select==0 || field_select==1);
790 av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
791 if(field_select_tables[block][xy] != field_select)
792 continue;
793 }
794
795 P_LEFT[0] = mv_table[xy - 1][0];
796 P_LEFT[1] = mv_table[xy - 1][1];
797 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
798
799 c->pred_x= P_LEFT[0];
800 c->pred_y= P_LEFT[1];
801
802 if(!s->first_slice_line){
803 P_TOP[0] = mv_table[xy - mot_stride][0];
804 P_TOP[1] = mv_table[xy - mot_stride][1];
805 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
806 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
807 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
808 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
809 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
810 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
811
812 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
813 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
814 }
815 P_MV1[0]= mx; //FIXME not correct if block != field_select
816 P_MV1[1]= my / 2;
817
818 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
819
820 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
821
822 mv_table[xy][0]= mx_i;
823 mv_table[xy][1]= my_i;
824
825 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
826 int dxy;
827
828 //FIXME chroma ME
829 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
830 dxy = ((my_i & 1) << 1) | (mx_i & 1);
831
832 if(s->no_rounding){
833 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
834 }else{
835 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
836 }
837 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
838 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
839 }else
840 dmin+= c->mb_penalty_factor; //field_select bits
841
842 dmin += field_select != block; //slightly prefer same field
843
844 if(dmin < best_dmin){
845 best_dmin= dmin;
846 best_field= field_select;
847 }
848 }
849 {
850 int16_t (*mv_table)[2]= mv_tables[block][best_field];
851
852 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
853 if(mv_table[xy][1]&1) same=0;
854 if(mv_table[xy][1]*2 != my) same=0;
855 if(best_field != block) same=0;
856 }
857
858 field_select_tables[block][xy]= best_field;
859 dmin_sum += best_dmin;
860 }
861
862 c->ymin<<=1;
863 c->ymax<<=1;
864 c->stride>>=1;
865 c->uvstride>>=1;
866
867 if(same)
868 return INT_MAX;
869
870 switch(c->avctx->mb_cmp&0xFF){
871 /*case FF_CMP_SSE:
872 return dmin_sum+ 32*s->qscale*s->qscale;*/
873 case FF_CMP_RD:
874 return dmin_sum;
875 default:
876 return dmin_sum+ 11*c->mb_penalty_factor;
877 }
878}
879
880static inline int get_penalty_factor(int lambda, int lambda2, int type){
881 switch(type&0xFF){
882 default:
883 case FF_CMP_SAD:
884 return lambda>>FF_LAMBDA_SHIFT;
885 case FF_CMP_DCT:
886 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
887 case FF_CMP_W53:
888 return (4*lambda)>>(FF_LAMBDA_SHIFT);
889 case FF_CMP_W97:
890 return (2*lambda)>>(FF_LAMBDA_SHIFT);
891 case FF_CMP_SATD:
892 case FF_CMP_DCT264:
893 return (2*lambda)>>FF_LAMBDA_SHIFT;
894 case FF_CMP_RD:
895 case FF_CMP_PSNR:
896 case FF_CMP_SSE:
897 case FF_CMP_NSSE:
898 return lambda2>>FF_LAMBDA_SHIFT;
899 case FF_CMP_BIT:
900 case FF_CMP_MEDIAN_SAD:
901 return 1;
902 }
903}
904
905void ff_estimate_p_frame_motion(MpegEncContext * s,
906 int mb_x, int mb_y)
907{
908 MotionEstContext * const c= &s->me;
909 uint8_t *pix, *ppix;
910 int sum, mx = 0, my = 0, dmin = 0;
911 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
912 int vard; ///< sum of squared differences with the estimated motion vector
913 int P[10][2];
914 const int shift= 1+s->quarter_sample;
915 int mb_type=0;
916 Picture * const pic= &s->current_picture;
917
918 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
919
920 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
921 av_assert0(s->linesize == c->stride);
922 av_assert0(s->uvlinesize == c->uvstride);
923
924 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
925 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
926 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
927 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
928
929 get_limits(s, 16*mb_x, 16*mb_y);
930 c->skip=0;
931
932 /* intra / predictive decision */
933 pix = c->src[0][0];
934 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
935 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
936 (((unsigned) sum * sum) >> 8) + 500;
937
938 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
939 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
940 c->mb_var_sum_temp += (varc+128)>>8;
941
942 if (s->motion_est != FF_ME_ZERO) {
943 const int mot_stride = s->b8_stride;
944 const int mot_xy = s->block_index[0];
945
946 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
947 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
948
949 if (P_LEFT[0] > (c->xmax << shift))
950 P_LEFT[0] = c->xmax << shift;
951
952 if (!s->first_slice_line) {
953 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
954 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
955 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
956 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
957 if (P_TOP[1] > (c->ymax << shift))
958 P_TOP[1] = c->ymax << shift;
959 if (P_TOPRIGHT[0] < (c->xmin << shift))
960 P_TOPRIGHT[0] = c->xmin << shift;
961 if (P_TOPRIGHT[1] > (c->ymax << shift))
962 P_TOPRIGHT[1] = c->ymax << shift;
963
964 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
965 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
966
967 if (s->out_format == FMT_H263) {
968 c->pred_x = P_MEDIAN[0];
969 c->pred_y = P_MEDIAN[1];
970 } else { /* MPEG-1 at least */
971 c->pred_x = P_LEFT[0];
972 c->pred_y = P_LEFT[1];
973 }
974 } else {
975 c->pred_x = P_LEFT[0];
976 c->pred_y = P_LEFT[1];
977 }
978 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
979 }
980
981 /* At this point (mx,my) are full-pell and the relative displacement */
982 ppix = c->ref[0][0] + (my * s->linesize) + mx;
983
984 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
985
986 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
987 c->mc_mb_var_sum_temp += (vard+128)>>8;
988
989 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
990 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
991 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
992 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
993
994 if (vard*2 + 200*256 > varc)
995 mb_type|= CANDIDATE_MB_TYPE_INTRA;
996 if (varc*2 + 200*256 > vard || s->qscale > 24){
997// if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
998 mb_type|= CANDIDATE_MB_TYPE_INTER;
999 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1000 if (s->mpv_flags & FF_MPV_FLAG_MV0)
1001 if(mx || my)
1002 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1003 }else{
1004 mx <<=shift;
1005 my <<=shift;
1006 }
1007 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
1008 && !c->skip && varc>50<<8 && vard>10<<8){
1009 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1010 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1011
1012 set_p_mv_tables(s, mx, my, 0);
1013 }else
1014 set_p_mv_tables(s, mx, my, 1);
1015 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
1016 && !c->skip){ //FIXME varc/d checks
1017 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1018 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1019 }
1020 }else{
1021 int intra_score, i;
1022 mb_type= CANDIDATE_MB_TYPE_INTER;
1023
1024 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1025 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1026 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1027
1028 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
1029 && !c->skip && varc>50<<8 && vard>10<<8){
1030 int dmin4= h263_mv4_search(s, mx, my, shift);
1031 if(dmin4 < dmin){
1032 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1033 dmin=dmin4;
1034 }
1035 }
1036 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
1037 && !c->skip){ //FIXME varc/d checks
1038 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1039 if(dmin_i < dmin){
1040 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1041 dmin= dmin_i;
1042 }
1043 }
1044
1045 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1046
1047 /* get intra luma score */
1048 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1049 intra_score= varc - 500;
1050 }else{
1051 unsigned mean = (sum+128)>>8;
1052 mean*= 0x01010101;
1053
1054 for(i=0; i<16; i++){
1055 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1056 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1057 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1058 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1059 }
1060
1061 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1062 }
1063 intra_score += c->mb_penalty_factor*16;
1064
1065 if(intra_score < dmin){
1066 mb_type= CANDIDATE_MB_TYPE_INTRA;
1067 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1068 }else
1069 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1070
1071 {
1072 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1073 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1074 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1075 }
1076 }
1077
1078 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1079}
1080
1081int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1082 int mb_x, int mb_y)
1083{
1084 MotionEstContext * const c= &s->me;
1085 int mx, my, dmin;
1086 int P[10][2];
1087 const int shift= 1+s->quarter_sample;
1088 const int xy= mb_x + mb_y*s->mb_stride;
1089 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1090
1091 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1092
1093 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1094 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
1095
1096 get_limits(s, 16*mb_x, 16*mb_y);
1097 c->skip=0;
1098
1099 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1100 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1101
1102 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1103
1104 /* special case for first line */
1105 if (s->first_slice_line) {
1106 c->pred_x= P_LEFT[0];
1107 c->pred_y= P_LEFT[1];
1108 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1109 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1110 } else {
1111 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1112 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1113 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1114 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1115 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1116 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1117 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1118
1119 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1120 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1121
1122 c->pred_x = P_MEDIAN[0];
1123 c->pred_y = P_MEDIAN[1];
1124 }
1125
1126 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1127
1128 s->p_mv_table[xy][0] = mx<<shift;
1129 s->p_mv_table[xy][1] = my<<shift;
1130
1131 return dmin;
1132}
1133
1134static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1135 int16_t (*mv_table)[2], int ref_index, int f_code)
1136{
1137 MotionEstContext * const c= &s->me;
1138 int mx = 0, my = 0, dmin = 0;
1139 int P[10][2];
1140 const int shift= 1+s->quarter_sample;
1141 const int mot_stride = s->mb_stride;
1142 const int mot_xy = mb_y*mot_stride + mb_x;
1143 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_DMV;
1144 int mv_scale;
1145
1146 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1147 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1148 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1149 c->current_mv_penalty= mv_penalty;
1150
1151 get_limits(s, 16*mb_x, 16*mb_y);
1152
1153 if (s->motion_est != FF_ME_ZERO) {
1154 P_LEFT[0] = mv_table[mot_xy - 1][0];
1155 P_LEFT[1] = mv_table[mot_xy - 1][1];
1156
1157 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1158
1159 /* special case for first line */
1160 if (!s->first_slice_line) {
1161 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1162 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1163 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1164 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1165 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1166 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1167 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1168
1169 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1170 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1171 }
1172 c->pred_x = P_LEFT[0];
1173 c->pred_y = P_LEFT[1];
1174
1175 if(mv_table == s->b_forw_mv_table){
1176 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1177 }else{
1178 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1179 }
1180
1181 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1182 }
1183
1184 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1185
1186 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1187 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1188
1189// s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1190 mv_table[mot_xy][0]= mx;
1191 mv_table[mot_xy][1]= my;
1192
1193 return dmin;
1194}
1195
1196static inline int check_bidir_mv(MpegEncContext * s,
1197 int motion_fx, int motion_fy,
1198 int motion_bx, int motion_by,
1199 int pred_fx, int pred_fy,
1200 int pred_bx, int pred_by,
1201 int size, int h)
1202{
1203 //FIXME optimize?
1204 //FIXME better f_code prediction (max mv & distance)
1205 //FIXME pointers
1206 MotionEstContext * const c= &s->me;
1207 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_DMV; // f_code of the prev frame
1208 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_DMV; // f_code of the prev frame
1209 int stride= c->stride;
1210 uint8_t *dest_y = c->scratchpad;
1211 uint8_t *ptr;
1212 int dxy;
1213 int src_x, src_y;
1214 int fbmin;
1215 uint8_t **src_data= c->src[0];
1216 uint8_t **ref_data= c->ref[0];
1217 uint8_t **ref2_data= c->ref[2];
1218
1219 if(s->quarter_sample){
1220 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1221 src_x = motion_fx >> 2;
1222 src_y = motion_fy >> 2;
1223
1224 ptr = ref_data[0] + (src_y * stride) + src_x;
1225 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1226
1227 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1228 src_x = motion_bx >> 2;
1229 src_y = motion_by >> 2;
1230
1231 ptr = ref2_data[0] + (src_y * stride) + src_x;
1232 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1233 }else{
1234 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1235 src_x = motion_fx >> 1;
1236 src_y = motion_fy >> 1;
1237
1238 ptr = ref_data[0] + (src_y * stride) + src_x;
1239 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1240
1241 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1242 src_x = motion_bx >> 1;
1243 src_y = motion_by >> 1;
1244
1245 ptr = ref2_data[0] + (src_y * stride) + src_x;
1246 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1247 }
1248
1249 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1250 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1251 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1252
1253 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1254 }
1255 //FIXME CHROMA !!!
1256
1257 return fbmin;
1258}
1259
1260/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1261static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1262{
1263 MotionEstContext * const c= &s->me;
1264 const int mot_stride = s->mb_stride;
1265 const int xy = mb_y *mot_stride + mb_x;
1266 int fbmin;
1267 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1268 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1269 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1270 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1271 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1272 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1273 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1274 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1275 const int flags= c->sub_flags;
1276 const int qpel= flags&FLAG_QPEL;
1277 const int shift= 1+qpel;
1278 const int xmin= c->xmin<<shift;
1279 const int ymin= c->ymin<<shift;
1280 const int xmax= c->xmax<<shift;
1281 const int ymax= c->ymax<<shift;
1282#define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1283#define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1284 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1285 uint8_t map[256] = { 0 };
1286
1287 map[hashidx&255] = 1;
1288
1289 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1290 motion_bx, motion_by,
1291 pred_fx, pred_fy,
1292 pred_bx, pred_by,
1293 0, 16);
1294
1295 if(s->avctx->bidir_refine){
1296 int end;
1297 static const uint8_t limittab[5]={0,8,32,64,80};
1298 const int limit= limittab[s->avctx->bidir_refine];
1299 static const int8_t vect[][4]={
1300{ 0, 0, 0, 1}, { 0, 0, 0,-1}, { 0, 0, 1, 0}, { 0, 0,-1, 0}, { 0, 1, 0, 0}, { 0,-1, 0, 0}, { 1, 0, 0, 0}, {-1, 0, 0, 0},
1301
1302{ 0, 0, 1, 1}, { 0, 0,-1,-1}, { 0, 1, 1, 0}, { 0,-1,-1, 0}, { 1, 1, 0, 0}, {-1,-1, 0, 0}, { 1, 0, 0, 1}, {-1, 0, 0,-1},
1303{ 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1304{ 0, 0,-1, 1}, { 0, 0, 1,-1}, { 0,-1, 1, 0}, { 0, 1,-1, 0}, {-1, 1, 0, 0}, { 1,-1, 0, 0}, { 1, 0, 0,-1}, {-1, 0, 0, 1},
1305{ 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1306
1307{ 0, 1, 1, 1}, { 0,-1,-1,-1}, { 1, 1, 1, 0}, {-1,-1,-1, 0}, { 1, 1, 0, 1}, {-1,-1, 0,-1}, { 1, 0, 1, 1}, {-1, 0,-1,-1},
1308{ 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1},
1309{ 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1},
1310{ 0, 1, 1,-1}, { 0,-1,-1, 1}, { 1, 1,-1, 0}, {-1,-1, 1, 0}, { 1,-1, 0, 1}, {-1, 1, 0,-1}, {-1, 0, 1, 1}, { 1, 0,-1,-1},
1311
1312{ 1, 1, 1, 1}, {-1,-1,-1,-1},
1313{ 1, 1, 1,-1}, {-1,-1,-1, 1}, { 1, 1,-1, 1}, {-1,-1, 1,-1}, { 1,-1, 1, 1}, {-1, 1,-1,-1}, {-1, 1, 1, 1}, { 1,-1,-1,-1},
1314{ 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1315 };
1316 static const uint8_t hash[]={
1317HASH8( 0, 0, 0, 1), HASH8( 0, 0, 0,-1), HASH8( 0, 0, 1, 0), HASH8( 0, 0,-1, 0), HASH8( 0, 1, 0, 0), HASH8( 0,-1, 0, 0), HASH8( 1, 0, 0, 0), HASH8(-1, 0, 0, 0),
1318
1319HASH8( 0, 0, 1, 1), HASH8( 0, 0,-1,-1), HASH8( 0, 1, 1, 0), HASH8( 0,-1,-1, 0), HASH8( 1, 1, 0, 0), HASH8(-1,-1, 0, 0), HASH8( 1, 0, 0, 1), HASH8(-1, 0, 0,-1),
1320HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1321HASH8( 0, 0,-1, 1), HASH8( 0, 0, 1,-1), HASH8( 0,-1, 1, 0), HASH8( 0, 1,-1, 0), HASH8(-1, 1, 0, 0), HASH8( 1,-1, 0, 0), HASH8( 1, 0, 0,-1), HASH8(-1, 0, 0, 1),
1322HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1323
1324HASH8( 0, 1, 1, 1), HASH8( 0,-1,-1,-1), HASH8( 1, 1, 1, 0), HASH8(-1,-1,-1, 0), HASH8( 1, 1, 0, 1), HASH8(-1,-1, 0,-1), HASH8( 1, 0, 1, 1), HASH8(-1, 0,-1,-1),
1325HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1),
1326HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1),
1327HASH8( 0, 1, 1,-1), HASH8( 0,-1,-1, 1), HASH8( 1, 1,-1, 0), HASH8(-1,-1, 1, 0), HASH8( 1,-1, 0, 1), HASH8(-1, 1, 0,-1), HASH8(-1, 0, 1, 1), HASH8( 1, 0,-1,-1),
1328
1329HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1330HASH8( 1, 1, 1,-1), HASH8(-1,-1,-1, 1), HASH8( 1, 1,-1, 1), HASH8(-1,-1, 1,-1), HASH8( 1,-1, 1, 1), HASH8(-1, 1,-1,-1), HASH8(-1, 1, 1, 1), HASH8( 1,-1,-1,-1),
1331HASH8( 1, 1,-1,-1), HASH8(-1,-1, 1, 1), HASH8( 1,-1,-1, 1), HASH8(-1, 1, 1,-1), HASH8( 1,-1, 1,-1), HASH8(-1, 1,-1, 1),
1332};
1333
1334#define CHECK_BIDIR(fx,fy,bx,by)\
1335 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1336 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1337 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1338 int score;\
1339 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1340 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);\
1341 if(score < fbmin){\
1342 hashidx += HASH(fx,fy,bx,by);\
1343 fbmin= score;\
1344 motion_fx+=fx;\
1345 motion_fy+=fy;\
1346 motion_bx+=bx;\
1347 motion_by+=by;\
1348 end=0;\
1349 }\
1350 }
1351#define CHECK_BIDIR2(a,b,c,d)\
1352CHECK_BIDIR(a,b,c,d)\
1353CHECK_BIDIR(-(a),-(b),-(c),-(d))
1354
1355 do{
1356 int i;
1357 int borderdist=0;
1358 end=1;
1359
1360 CHECK_BIDIR2(0,0,0,1)
1361 CHECK_BIDIR2(0,0,1,0)
1362 CHECK_BIDIR2(0,1,0,0)
1363 CHECK_BIDIR2(1,0,0,0)
1364
1365 for(i=8; i<limit; i++){
1366 int fx= motion_fx+vect[i][0];
1367 int fy= motion_fy+vect[i][1];
1368 int bx= motion_bx+vect[i][2];
1369 int by= motion_by+vect[i][3];
1370 if(borderdist<=0){
1371 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1372 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1373 if((a|b) < 0)
1374 map[(hashidx+hash[i])&255] = 1;
1375 }
1376 if(!map[(hashidx+hash[i])&255]){
1377 int score;
1378 map[(hashidx+hash[i])&255] = 1;
1379 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1380 if(score < fbmin){
1381 hashidx += hash[i];
1382 fbmin= score;
1383 motion_fx=fx;
1384 motion_fy=fy;
1385 motion_bx=bx;
1386 motion_by=by;
1387 end=0;
1388 borderdist--;
1389 if(borderdist<=0){
1390 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1391 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1392 borderdist= FFMIN(a,b);
1393 }
1394 }
1395 }
1396 }
1397 }while(!end);
1398 }
1399
1400 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1401 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1402 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1403 s->b_bidir_back_mv_table[xy][1]= motion_by;
1404
1405 return fbmin;
1406}
1407
1408static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1409{
1410 MotionEstContext * const c= &s->me;
1411 int P[10][2];
1412 const int mot_stride = s->mb_stride;
1413 const int mot_xy = mb_y*mot_stride + mb_x;
1414 const int shift= 1+s->quarter_sample;
1415 int dmin, i;
1416 const int time_pp= s->pp_time;
1417 const int time_pb= s->pb_time;
1418 int mx, my, xmin, xmax, ymin, ymax;
1419 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1420
1421 c->current_mv_penalty= c->mv_penalty[1] + MAX_DMV;
1422 ymin= xmin=(-32)>>shift;
1423 ymax= xmax= 31>>shift;
1424
1425 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1426 s->mv_type= MV_TYPE_8X8;
1427 }else{
1428 s->mv_type= MV_TYPE_16X16;
1429 }
1430
1431 for(i=0; i<4; i++){
1432 int index= s->block_index[i];
1433 int min, max;
1434
1435 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1436 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1437 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1438 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1439// c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1440// c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1441
1442 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1443 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1444 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1445 min+= 16*mb_x - 1;
1446 xmax= FFMIN(xmax, s->width - max);
1447 xmin= FFMAX(xmin, - 16 - min);
1448
1449 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1450 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1451 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1452 min+= 16*mb_y - 1;
1453 ymax= FFMIN(ymax, s->height - max);
1454 ymin= FFMAX(ymin, - 16 - min);
1455
1456 if(s->mv_type == MV_TYPE_16X16) break;
1457 }
1458
1459 av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1460
1461 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1462 s->b_direct_mv_table[mot_xy][0]= 0;
1463 s->b_direct_mv_table[mot_xy][1]= 0;
1464
1465 return 256*256*256*64;
1466 }
1467
1468 c->xmin= xmin;
1469 c->ymin= ymin;
1470 c->xmax= xmax;
1471 c->ymax= ymax;
1472 c->flags |= FLAG_DIRECT;
1473 c->sub_flags |= FLAG_DIRECT;
1474 c->pred_x=0;
1475 c->pred_y=0;
1476
1477 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1478 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1479
1480 /* special case for first line */
1481 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1482 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1483 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1484 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1485 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1486
1487 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1488 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1489 }
1490
1491 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1492 if(c->sub_flags&FLAG_QPEL)
1493 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1494 else
1495 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1496
1497 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1498 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1499
1500 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1501
1502 mv_table[mot_xy][0]= mx;
1503 mv_table[mot_xy][1]= my;
1504 c->flags &= ~FLAG_DIRECT;
1505 c->sub_flags &= ~FLAG_DIRECT;
1506
1507 return dmin;
1508}
1509
1510void ff_estimate_b_frame_motion(MpegEncContext * s,
1511 int mb_x, int mb_y)
1512{
1513 MotionEstContext * const c= &s->me;
1514 const int penalty_factor= c->mb_penalty_factor;
1515 int fmin, bmin, dmin, fbmin, bimin, fimin;
1516 int type=0;
1517 const int xy = mb_y*s->mb_stride + mb_x;
1518 init_ref(c, s->new_picture.f->data, s->last_picture.f->data,
1519 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1520
1521 get_limits(s, 16*mb_x, 16*mb_y);
1522
1523 c->skip=0;
1524
1525 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1526 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1527
1528 score= ((unsigned)(score*score + 128*256))>>16;
1529 c->mc_mb_var_sum_temp += score;
1530 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1531 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1532
1533 return;
1534 }
1535
1536 if (s->codec_id == AV_CODEC_ID_MPEG4)
1537 dmin= direct_search(s, mb_x, mb_y);
1538 else
1539 dmin= INT_MAX;
1540// FIXME penalty stuff for non-MPEG-4
1541 c->skip=0;
1542 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1543 3 * penalty_factor;
1544
1545 c->skip=0;
1546 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1547 2 * penalty_factor;
1548 ff_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1549
1550 c->skip=0;
1551 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1552 ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1553
1554 if (s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME) {
1555//FIXME mb type penalty
1556 c->skip=0;
1557 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
1558 fimin= interlaced_search(s, 0,
1559 s->b_field_mv_table[0], s->b_field_select_table[0],
1560 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1561 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_DMV;
1562 bimin= interlaced_search(s, 2,
1563 s->b_field_mv_table[1], s->b_field_select_table[1],
1564 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1565 }else
1566 fimin= bimin= INT_MAX;
1567
1568 {
1569 int score= fmin;
1570 type = CANDIDATE_MB_TYPE_FORWARD;
1571
1572 if (dmin <= score){
1573 score = dmin;
1574 type = CANDIDATE_MB_TYPE_DIRECT;
1575 }
1576 if(bmin<score){
1577 score=bmin;
1578 type= CANDIDATE_MB_TYPE_BACKWARD;
1579 }
1580 if(fbmin<score){
1581 score=fbmin;
1582 type= CANDIDATE_MB_TYPE_BIDIR;
1583 }
1584 if(fimin<score){
1585 score=fimin;
1586 type= CANDIDATE_MB_TYPE_FORWARD_I;
1587 }
1588 if(bimin<score){
1589 score=bimin;
1590 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1591 }
1592
1593 score= ((unsigned)(score*score + 128*256))>>16;
1594 c->mc_mb_var_sum_temp += score;
1595 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1596 }
1597
1598 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1599 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1600 if(fimin < INT_MAX)
1601 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1602 if(bimin < INT_MAX)
1603 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1604 if(fimin < INT_MAX && bimin < INT_MAX){
1605 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1606 }
1607 //FIXME something smarter
1608 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1609 if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1610 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1611 type |= CANDIDATE_MB_TYPE_DIRECT0;
1612 }
1613
1614 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1615}
1616
1617/* find best f_code for ME which do unlimited searches */
1618int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1619{
1620 if (s->motion_est != FF_ME_ZERO) {
1621 int score[8];
1622 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1623 uint8_t * fcode_tab= s->fcode_tab;
1624 int best_fcode=-1;
1625 int best_score=-10000000;
1626
1627 if(s->msmpeg4_version)
1628 range= FFMIN(range, 16);
1629 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1630 range= FFMIN(range, 256);
1631
1632 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1633
1634 for(y=0; y<s->mb_height; y++){
1635 int x;
1636 int xy= y*s->mb_stride;
1637 for(x=0; x<s->mb_width; x++){
1638 if(s->mb_type[xy] & type){
1639 int mx= mv_table[xy][0];
1640 int my= mv_table[xy][1];
1641 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1642 fcode_tab[my + MAX_MV]);
1643 int j;
1644
1645 if(mx >= range || mx < -range ||
1646 my >= range || my < -range)
1647 continue;
1648
1649 for(j=0; j<fcode && j<8; j++){
1650 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1651 score[j]-= 170;
1652 }
1653 }
1654 xy++;
1655 }
1656 }
1657
1658 for(i=1; i<8; i++){
1659 if(score[i] > best_score){
1660 best_score= score[i];
1661 best_fcode= i;
1662 }
1663 }
1664
1665 return best_fcode;
1666 }else{
1667 return 1;
1668 }
1669}
1670
1671void ff_fix_long_p_mvs(MpegEncContext * s)
1672{
1673 MotionEstContext * const c= &s->me;
1674 const int f_code= s->f_code;
1675 int y, range;
1676 av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1677
1678 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1679
1680 av_assert0(range <= 16 || !s->msmpeg4_version);
1681 av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1682
1683 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1684
1685 if (s->avctx->flags & AV_CODEC_FLAG_4MV) {
1686 const int wrap= s->b8_stride;
1687
1688 /* clip / convert to intra 8x8 type MVs */
1689 for(y=0; y<s->mb_height; y++){
1690 int xy= y*2*wrap;
1691 int i= y*s->mb_stride;
1692 int x;
1693
1694 for(x=0; x<s->mb_width; x++){
1695 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1696 int block;
1697 for(block=0; block<4; block++){
1698 int off= (block& 1) + (block>>1)*wrap;
1699 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1700 int my = s->current_picture.motion_val[0][ xy + off ][1];
1701
1702 if( mx >=range || mx <-range
1703 || my >=range || my <-range){
1704 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1705 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1706 s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1707 }
1708 }
1709 }
1710 xy+=2;
1711 i++;
1712 }
1713 }
1714 }
1715}
1716
1717/**
1718 * @param truncate 1 for truncation, 0 for using intra
1719 */
1720void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1721 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1722{
1723 MotionEstContext * const c= &s->me;
1724 int y, h_range, v_range;
1725
1726 // RAL: 8 in MPEG-1, 16 in MPEG-4
1727 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1728
1729 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1730
1731 h_range= range;
1732 v_range= field_select_table ? range>>1 : range;
1733
1734 /* clip / convert to intra 16x16 type MVs */
1735 for(y=0; y<s->mb_height; y++){
1736 int x;
1737 int xy= y*s->mb_stride;
1738 for(x=0; x<s->mb_width; x++){
1739 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1740 if (!field_select_table || field_select_table[xy] == field_select) {
1741 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1742 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1743
1744 if(truncate){
1745 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1746 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1747 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1748 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1749 }else{
1750 s->mb_type[xy] &= ~type;
1751 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1752 mv_table[xy][0]=
1753 mv_table[xy][1]= 0;
1754 }
1755 }
1756 }
1757 }
1758 xy++;
1759 }
1760 }
1761}
1762
generated by cgit at 2024-05-17 05:03:33 (GMT)