?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
aarch64: Add assembly support for -fsanitize=hwaddress tagged globals.
[libav.git] / libavcodec / motion_est.c
blobc6dd3b86afb2bc311052c8a24a036f9403204d3e
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 Libav.
9 *
10 * Libav 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 * Libav 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 Libav; 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 #undef NDEBUG
42 #include <assert.h>
43
44 #define P_LEFT P[1]
45 #define P_TOP P[2]
46 #define P_TOPRIGHT P[3]
47 #define P_MEDIAN P[4]
48 #define P_MV1 P[9]
49
50 #define ME_MAP_SHIFT 3
51 #define ME_MAP_MV_BITS 11
52
53 static int sad_hpel_motion_search(MpegEncContext * s,
54 int *mx_ptr, int *my_ptr, int dmin,
55 int src_index, int ref_index,
56 int size, int h);
57
58 static inline unsigned update_map_generation(MotionEstContext *c)
59 {
60 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
61 if(c->map_generation==0){
62 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
63 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
64 }
65 return c->map_generation;
66 }
67
68 /* shape adaptive search stuff */
69 typedef struct Minima{
70 int height;
71 int x, y;
72 int checked;
73 }Minima;
74
75 static int minima_cmp(const void *a, const void *b){
76 const Minima *da = (const Minima *) a;
77 const Minima *db = (const Minima *) b;
78
79 return da->height - db->height;
80 }
81
82 #define FLAG_QPEL 1 //must be 1
83 #define FLAG_CHROMA 2
84 #define FLAG_DIRECT 4
85
86 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
87 const int offset[3]= {
88 y*c-> stride + x,
89 ((y*c->uvstride + x)>>1),
90 ((y*c->uvstride + x)>>1),
91 };
92 int i;
93 for(i=0; i<3; i++){
94 c->src[0][i]= src [i] + offset[i];
95 c->ref[0][i]= ref [i] + offset[i];
96 }
97 if(ref_index){
98 for(i=0; i<3; i++){
99 c->ref[ref_index][i]= ref2[i] + offset[i];
100 }
101 }
102 }
103
104 static int get_flags(MotionEstContext *c, int direct, int chroma){
105 return ((c->avctx->flags&AV_CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
106 + (direct ? FLAG_DIRECT : 0)
107 + (chroma ? FLAG_CHROMA : 0);
108 }
109
110 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
111 const int size, const int h, int ref_index, int src_index,
112 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
113 MotionEstContext * const c= &s->me;
114 const int stride= c->stride;
115 const int hx= subx + (x<<(1+qpel));
116 const int hy= suby + (y<<(1+qpel));
117 uint8_t * const * const ref= c->ref[ref_index];
118 uint8_t * const * const src= c->src[src_index];
119 int d;
120 //FIXME check chroma 4mv, (no crashes ...)
121 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
122 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
123 const int time_pp= s->pp_time;
124 const int time_pb= s->pb_time;
125 const int mask= 2*qpel+1;
126 if(s->mv_type==MV_TYPE_8X8){
127 int i;
128 for(i=0; i<4; i++){
129 int fx = c->direct_basis_mv[i][0] + hx;
130 int fy = c->direct_basis_mv[i][1] + hy;
131 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));
132 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));
133 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
134 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
135
136 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
137 if(qpel){
138 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
139 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
140 }else{
141 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
142 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
143 }
144 }
145 }else{
146 int fx = c->direct_basis_mv[0][0] + hx;
147 int fy = c->direct_basis_mv[0][1] + hy;
148 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
149 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
150 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
151 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
152
153 if(qpel){
154 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
155 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
156 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
157 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
158 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
159 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
160 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
161 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
162 }else{
163 assert((fx>>1) + 16*s->mb_x >= -16);
164 assert((fy>>1) + 16*s->mb_y >= -16);
165 assert((fx>>1) + 16*s->mb_x <= s->width);
166 assert((fy>>1) + 16*s->mb_y <= s->height);
167 assert((bx>>1) + 16*s->mb_x >= -16);
168 assert((by>>1) + 16*s->mb_y >= -16);
169 assert((bx>>1) + 16*s->mb_x <= s->width);
170 assert((by>>1) + 16*s->mb_y <= s->height);
171
172 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
173 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
174 }
175 }
176 d = cmp_func(s, c->temp, src[0], stride, 16);
177 }else
178 d= 256*256*256*32;
179 return d;
180 }
181
182 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
183 const int size, const int h, int ref_index, int src_index,
184 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
185 MotionEstContext * const c= &s->me;
186 const int stride= c->stride;
187 const int uvstride= c->uvstride;
188 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
189 const int hx= subx + (x<<(1+qpel));
190 const int hy= suby + (y<<(1+qpel));
191 uint8_t * const * const ref= c->ref[ref_index];
192 uint8_t * const * const src= c->src[src_index];
193 int d;
194 //FIXME check chroma 4mv, (no crashes ...)
195 int uvdxy; /* no, it might not be used uninitialized */
196 if(dxy){
197 if(qpel){
198 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
199 if(chroma){
200 int cx= hx/2;
201 int cy= hy/2;
202 cx= (cx>>1)|(cx&1);
203 cy= (cy>>1)|(cy&1);
204 uvdxy= (cx&1) + 2*(cy&1);
205 // 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
206 }
207 }else{
208 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
209 if(chroma)
210 uvdxy= dxy | (x&1) | (2*(y&1));
211 }
212 d = cmp_func(s, c->temp, src[0], stride, h);
213 }else{
214 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
215 if(chroma)
216 uvdxy= (x&1) + 2*(y&1);
217 }
218 if(chroma){
219 uint8_t * const uvtemp= c->temp + 16*stride;
220 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
221 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
222 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
223 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
224 }
225 return d;
226 }
227
228 static int cmp_simple(MpegEncContext *s, const int x, const int y,
229 int ref_index, int src_index,
230 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
231 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
232 }
233
234 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
235 const int size, const int h, int ref_index, int src_index,
236 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
237 if(flags&FLAG_DIRECT){
238 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
239 }else{
240 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
241 }
242 }
243
244 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
245 const int size, const int h, int ref_index, int src_index,
246 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
247 if(flags&FLAG_DIRECT){
248 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
249 }else{
250 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);
251 }
252 }
253
254 /** @brief compares a block (either a full macroblock or a partition thereof)
255 against a proposed motion-compensated prediction of that block
256 */
257 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
258 const int size, const int h, int ref_index, int src_index,
259 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
260 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
261 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
262 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
263 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
264 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
265 && subx==0 && suby==0){
266 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
267 }else{
268 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
269 }
270 }
271
272 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
273 const int size, const int h, int ref_index, int src_index,
274 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
275 if(flags&FLAG_DIRECT){
276 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
277 }else{
278 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
279 }
280 }
281
282 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
283 const int size, const int h, int ref_index, int src_index,
284 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
285 if(flags&FLAG_DIRECT){
286 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
287 }else{
288 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
289 }
290 }
291
292 #include "motion_est_template.c"
293
294 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
295 ptrdiff_t stride, int h)
296 {
297 return 0;
298 }
299
300 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
301 }
302
303 int ff_init_me(MpegEncContext *s){
304 MotionEstContext * const c= &s->me;
305 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
306 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
307
308 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
309 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
310 return -1;
311 }
312
313 c->avctx= s->avctx;
314
315 if(cache_size < 2*dia_size && !c->stride){
316 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
317 }
318
319 ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
320 ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
321 ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
322 ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
323
324 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
325 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
326 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
327
328 /*FIXME s->no_rounding b_type*/
329 if (s->avctx->flags & AV_CODEC_FLAG_QPEL) {
330 c->sub_motion_search= qpel_motion_search;
331 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
332 if (s->no_rounding)
333 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
334 else
335 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
336 }else{
337 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
338 c->sub_motion_search= hpel_motion_search;
339 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
340 && c->avctx-> me_cmp == FF_CMP_SAD
341 && c->avctx-> mb_cmp == FF_CMP_SAD)
342 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
343 else
344 c->sub_motion_search= hpel_motion_search;
345 }
346 c->hpel_avg = s->hdsp.avg_pixels_tab;
347 if (s->no_rounding)
348 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
349 else
350 c->hpel_put = s->hdsp.put_pixels_tab;
351
352 if(s->linesize){
353 c->stride = s->linesize;
354 c->uvstride= s->uvlinesize;
355 }else{
356 c->stride = 16*s->mb_width + 32;
357 c->uvstride= 8*s->mb_width + 16;
358 }
359
360 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
361 * not have yet, and even if we had, the motion estimation code
362 * does not expect it. */
363 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
364 s->mecc.me_cmp[2] = zero_cmp;
365 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
366 s->mecc.me_sub_cmp[2] = zero_cmp;
367 c->hpel_put[2][0]= c->hpel_put[2][1]=
368 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
369
370 if(s->codec_id == AV_CODEC_ID_H261){
371 c->sub_motion_search= no_sub_motion_search;
372 }
373
374 return 0;
375 }
376
377 #define CHECK_SAD_HALF_MV(suffix, x, y) \
378 {\
379 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
380 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
381 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
382 }
383
384 static int sad_hpel_motion_search(MpegEncContext * s,
385 int *mx_ptr, int *my_ptr, int dmin,
386 int src_index, int ref_index,
387 int size, int h)
388 {
389 MotionEstContext * const c= &s->me;
390 const int penalty_factor= c->sub_penalty_factor;
391 int mx, my, dminh;
392 uint8_t *pix, *ptr;
393 int stride= c->stride;
394 const int flags= c->sub_flags;
395 LOAD_COMMON
396
397 assert(flags == 0);
398
399 if(c->skip){
400 *mx_ptr = 0;
401 *my_ptr = 0;
402 return dmin;
403 }
404
405 pix = c->src[src_index][0];
406
407 mx = *mx_ptr;
408 my = *my_ptr;
409 ptr = c->ref[ref_index][0] + (my * stride) + mx;
410
411 dminh = dmin;
412
413 if (mx > xmin && mx < xmax &&
414 my > ymin && my < ymax) {
415 int dx=0, dy=0;
416 int d, pen_x, pen_y;
417 const int index= (my<<ME_MAP_SHIFT) + mx;
418 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
419 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
420 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
421 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
422 mx<<=1;
423 my<<=1;
424
425
426 pen_x= pred_x + mx;
427 pen_y= pred_y + my;
428
429 ptr-= stride;
430 if(t<=b){
431 CHECK_SAD_HALF_MV(y2 , 0, -1)
432 if(l<=r){
433 CHECK_SAD_HALF_MV(xy2, -1, -1)
434 if(t+r<=b+l){
435 CHECK_SAD_HALF_MV(xy2, +1, -1)
436 ptr+= stride;
437 }else{
438 ptr+= stride;
439 CHECK_SAD_HALF_MV(xy2, -1, +1)
440 }
441 CHECK_SAD_HALF_MV(x2 , -1, 0)
442 }else{
443 CHECK_SAD_HALF_MV(xy2, +1, -1)
444 if(t+l<=b+r){
445 CHECK_SAD_HALF_MV(xy2, -1, -1)
446 ptr+= stride;
447 }else{
448 ptr+= stride;
449 CHECK_SAD_HALF_MV(xy2, +1, +1)
450 }
451 CHECK_SAD_HALF_MV(x2 , +1, 0)
452 }
453 }else{
454 if(l<=r){
455 if(t+l<=b+r){
456 CHECK_SAD_HALF_MV(xy2, -1, -1)
457 ptr+= stride;
458 }else{
459 ptr+= stride;
460 CHECK_SAD_HALF_MV(xy2, +1, +1)
461 }
462 CHECK_SAD_HALF_MV(x2 , -1, 0)
463 CHECK_SAD_HALF_MV(xy2, -1, +1)
464 }else{
465 if(t+r<=b+l){
466 CHECK_SAD_HALF_MV(xy2, +1, -1)
467 ptr+= stride;
468 }else{
469 ptr+= stride;
470 CHECK_SAD_HALF_MV(xy2, -1, +1)
471 }
472 CHECK_SAD_HALF_MV(x2 , +1, 0)
473 CHECK_SAD_HALF_MV(xy2, +1, +1)
474 }
475 CHECK_SAD_HALF_MV(y2 , 0, +1)
476 }
477 mx+=dx;
478 my+=dy;
479
480 }else{
481 mx<<=1;
482 my<<=1;
483 }
484
485 *mx_ptr = mx;
486 *my_ptr = my;
487 return dminh;
488 }
489
490 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
491 {
492 const int xy= s->mb_x + s->mb_y*s->mb_stride;
493
494 s->p_mv_table[xy][0] = mx;
495 s->p_mv_table[xy][1] = my;
496
497 /* has already been set to the 4 MV if 4MV is done */
498 if(mv4){
499 int mot_xy= s->block_index[0];
500
501 s->current_picture.motion_val[0][mot_xy ][0] = mx;
502 s->current_picture.motion_val[0][mot_xy ][1] = my;
503 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
504 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
505
506 mot_xy += s->b8_stride;
507 s->current_picture.motion_val[0][mot_xy ][0] = mx;
508 s->current_picture.motion_val[0][mot_xy ][1] = my;
509 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
510 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
511 }
512 }
513
514 /**
515 * get fullpel ME search limits.
516 */
517 static inline void get_limits(MpegEncContext *s, int x, int y)
518 {
519 MotionEstContext * const c= &s->me;
520 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
521 /*
522 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
523 else c->range= 16;
524 */
525 if (s->unrestricted_mv) {
526 c->xmin = - x - 16;
527 c->ymin = - y - 16;
528 c->xmax = - x + s->mb_width *16;
529 c->ymax = - y + s->mb_height*16;
530 } else if (s->out_format == FMT_H261){
531 // Search range of H.261 is different from other codec standards
532 c->xmin = (x > 15) ? - 15 : 0;
533 c->ymin = (y > 15) ? - 15 : 0;
534 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
535 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
536 } else {
537 c->xmin = - x;
538 c->ymin = - y;
539 c->xmax = - x + s->mb_width *16 - 16;
540 c->ymax = - y + s->mb_height*16 - 16;
541 }
542 if(range){
543 c->xmin = FFMAX(c->xmin,-range);
544 c->xmax = FFMIN(c->xmax, range);
545 c->ymin = FFMAX(c->ymin,-range);
546 c->ymax = FFMIN(c->ymax, range);
547 }
548 }
549
550 static inline void init_mv4_ref(MotionEstContext *c){
551 const int stride= c->stride;
552
553 c->ref[1][0] = c->ref[0][0] + 8;
554 c->ref[2][0] = c->ref[0][0] + 8*stride;
555 c->ref[3][0] = c->ref[2][0] + 8;
556 c->src[1][0] = c->src[0][0] + 8;
557 c->src[2][0] = c->src[0][0] + 8*stride;
558 c->src[3][0] = c->src[2][0] + 8;
559 }
560
561 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
562 {
563 MotionEstContext * const c= &s->me;
564 const int size= 1;
565 const int h=8;
566 int block;
567 int P[10][2];
568 int dmin_sum=0, mx4_sum=0, my4_sum=0;
569 int same=1;
570 const int stride= c->stride;
571 uint8_t *mv_penalty= c->current_mv_penalty;
572
573 init_mv4_ref(c);
574
575 for(block=0; block<4; block++){
576 int mx4, my4;
577 int pred_x4, pred_y4;
578 int dmin4;
579 static const int off[4]= {2, 1, 1, -1};
580 const int mot_stride = s->b8_stride;
581 const int mot_xy = s->block_index[block];
582
583 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
584 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
585
586 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
587
588 /* special case for first line */
589 if (s->first_slice_line && block<2) {
590 c->pred_x= pred_x4= P_LEFT[0];
591 c->pred_y= pred_y4= P_LEFT[1];
592 } else {
593 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
594 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
595 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
596 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
597 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
598 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
599 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
600 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
601
602 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
603 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
604
605 c->pred_x= pred_x4 = P_MEDIAN[0];
606 c->pred_y= pred_y4 = P_MEDIAN[1];
607 }
608 P_MV1[0]= mx;
609 P_MV1[1]= my;
610
611 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
612
613 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
614
615 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
616 int dxy;
617 const int offset= ((block&1) + (block>>1)*stride)*8;
618 uint8_t *dest_y = c->scratchpad + offset;
619 if(s->quarter_sample){
620 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
621 dxy = ((my4 & 3) << 2) | (mx4 & 3);
622
623 if(s->no_rounding)
624 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
625 else
626 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
627 }else{
628 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
629 dxy = ((my4 & 1) << 1) | (mx4 & 1);
630
631 if(s->no_rounding)
632 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
633 else
634 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
635 }
636 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
637 }else
638 dmin_sum+= dmin4;
639
640 if(s->quarter_sample){
641 mx4_sum+= mx4/2;
642 my4_sum+= my4/2;
643 }else{
644 mx4_sum+= mx4;
645 my4_sum+= my4;
646 }
647
648 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
649 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
650
651 if(mx4 != mx || my4 != my) same=0;
652 }
653
654 if(same)
655 return INT_MAX;
656
657 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
658 dmin_sum += s->mecc.mb_cmp[0](s,
659 s->new_picture.f->data[0] +
660 s->mb_x * 16 + s->mb_y * 16 * stride,
661 c->scratchpad, stride, 16);
662 }
663
664 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
665 int dxy;
666 int mx, my;
667 int offset;
668
669 mx= ff_h263_round_chroma(mx4_sum);
670 my= ff_h263_round_chroma(my4_sum);
671 dxy = ((my & 1) << 1) | (mx & 1);
672
673 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
674
675 if(s->no_rounding){
676 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
677 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
678 }else{
679 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
680 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
681 }
682
683 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);
684 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);
685 }
686
687 c->pred_x= mx;
688 c->pred_y= my;
689
690 switch(c->avctx->mb_cmp&0xFF){
691 /*case FF_CMP_SSE:
692 return dmin_sum+ 32*s->qscale*s->qscale;*/
693 case FF_CMP_RD:
694 return dmin_sum;
695 default:
696 return dmin_sum+ 11*c->mb_penalty_factor;
697 }
698 }
699
700 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
701 MotionEstContext * const c= &s->me;
702
703 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
704 c->src[1][0] = c->src[0][0] + s->linesize;
705 if(c->flags & FLAG_CHROMA){
706 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
707 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
708 c->src[1][1] = c->src[0][1] + s->uvlinesize;
709 c->src[1][2] = c->src[0][2] + s->uvlinesize;
710 }
711 }
712
713 static int interlaced_search(MpegEncContext *s, int ref_index,
714 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
715 {
716 MotionEstContext * const c= &s->me;
717 const int size=0;
718 const int h=8;
719 int block;
720 int P[10][2];
721 uint8_t * const mv_penalty= c->current_mv_penalty;
722 int same=1;
723 const int stride= 2*s->linesize;
724 int dmin_sum= 0;
725 const int mot_stride= s->mb_stride;
726 const int xy= s->mb_x + s->mb_y*mot_stride;
727
728 c->ymin>>=1;
729 c->ymax>>=1;
730 c->stride<<=1;
731 c->uvstride<<=1;
732 init_interlaced_ref(s, ref_index);
733
734 for(block=0; block<2; block++){
735 int field_select;
736 int best_dmin= INT_MAX;
737 int best_field= -1;
738
739 for(field_select=0; field_select<2; field_select++){
740 int dmin, mx_i, my_i;
741 int16_t (*mv_table)[2]= mv_tables[block][field_select];
742
743 if(user_field_select){
744 assert(field_select==0 || field_select==1);
745 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
746 if(field_select_tables[block][xy] != field_select)
747 continue;
748 }
749
750 P_LEFT[0] = mv_table[xy - 1][0];
751 P_LEFT[1] = mv_table[xy - 1][1];
752 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
753
754 c->pred_x= P_LEFT[0];
755 c->pred_y= P_LEFT[1];
756
757 if(!s->first_slice_line){
758 P_TOP[0] = mv_table[xy - mot_stride][0];
759 P_TOP[1] = mv_table[xy - mot_stride][1];
760 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
761 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
762 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
763 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
764 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
765 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
766
767 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
768 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
769 }
770 P_MV1[0]= mx; //FIXME not correct if block != field_select
771 P_MV1[1]= my / 2;
772
773 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
774
775 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
776
777 mv_table[xy][0]= mx_i;
778 mv_table[xy][1]= my_i;
779
780 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
781 int dxy;
782
783 //FIXME chroma ME
784 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
785 dxy = ((my_i & 1) << 1) | (mx_i & 1);
786
787 if(s->no_rounding){
788 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
789 }else{
790 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
791 }
792 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
793 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
794 }else
795 dmin+= c->mb_penalty_factor; //field_select bits
796
797 dmin += field_select != block; //slightly prefer same field
798
799 if(dmin < best_dmin){
800 best_dmin= dmin;
801 best_field= field_select;
802 }
803 }
804 {
805 int16_t (*mv_table)[2]= mv_tables[block][best_field];
806
807 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
808 if(mv_table[xy][1]&1) same=0;
809 if(mv_table[xy][1]*2 != my) same=0;
810 if(best_field != block) same=0;
811 }
812
813 field_select_tables[block][xy]= best_field;
814 dmin_sum += best_dmin;
815 }
816
817 c->ymin<<=1;
818 c->ymax<<=1;
819 c->stride>>=1;
820 c->uvstride>>=1;
821
822 if(same)
823 return INT_MAX;
824
825 switch(c->avctx->mb_cmp&0xFF){
826 /*case FF_CMP_SSE:
827 return dmin_sum+ 32*s->qscale*s->qscale;*/
828 case FF_CMP_RD:
829 return dmin_sum;
830 default:
831 return dmin_sum+ 11*c->mb_penalty_factor;
832 }
833 }
834
835 static inline int get_penalty_factor(int lambda, int lambda2, int type){
836 switch(type&0xFF){
837 default:
838 case FF_CMP_SAD:
839 return lambda>>FF_LAMBDA_SHIFT;
840 case FF_CMP_DCT:
841 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
842 case FF_CMP_SATD:
843 case FF_CMP_DCT264:
844 return (2*lambda)>>FF_LAMBDA_SHIFT;
845 case FF_CMP_RD:
846 case FF_CMP_PSNR:
847 case FF_CMP_SSE:
848 case FF_CMP_NSSE:
849 return lambda2>>FF_LAMBDA_SHIFT;
850 case FF_CMP_BIT:
851 return 1;
852 }
853 }
854
855 void ff_estimate_p_frame_motion(MpegEncContext * s,
856 int mb_x, int mb_y)
857 {
858 MotionEstContext * const c= &s->me;
859 uint8_t *pix, *ppix;
860 int sum, mx = 0, my = 0, dmin = 0;
861 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
862 int vard; ///< sum of squared differences with the estimated motion vector
863 int P[10][2];
864 const int shift= 1+s->quarter_sample;
865 int mb_type=0;
866 Picture * const pic= &s->current_picture;
867
868 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
869
870 assert(s->quarter_sample==0 || s->quarter_sample==1);
871 assert(s->linesize == c->stride);
872 assert(s->uvlinesize == c->uvstride);
873
874 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
875 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
876 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
877 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
878
879 get_limits(s, 16*mb_x, 16*mb_y);
880 c->skip=0;
881
882 /* intra / predictive decision */
883 pix = c->src[0][0];
884 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
885 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
886 (((unsigned) sum * sum) >> 8) + 500;
887
888 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
889 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
890 c->mb_var_sum_temp += (varc+128)>>8;
891
892 if (s->motion_est != FF_ME_ZERO) {
893 const int mot_stride = s->b8_stride;
894 const int mot_xy = s->block_index[0];
895
896 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
897 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
898
899 if (P_LEFT[0] > (c->xmax << shift))
900 P_LEFT[0] = c->xmax << shift;
901
902 if (!s->first_slice_line) {
903 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
904 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
905 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
906 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
907 if (P_TOP[1] > (c->ymax << shift))
908 P_TOP[1] = c->ymax << shift;
909 if (P_TOPRIGHT[0] < (c->xmin << shift))
910 P_TOPRIGHT[0] = c->xmin << shift;
911 if (P_TOPRIGHT[1] > (c->ymax << shift))
912 P_TOPRIGHT[1] = c->ymax << shift;
913
914 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
915 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
916
917 if (s->out_format == FMT_H263) {
918 c->pred_x = P_MEDIAN[0];
919 c->pred_y = P_MEDIAN[1];
920 } else { /* MPEG-1 at least */
921 c->pred_x = P_LEFT[0];
922 c->pred_y = P_LEFT[1];
923 }
924 } else {
925 c->pred_x = P_LEFT[0];
926 c->pred_y = P_LEFT[1];
927 }
928 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
929 }
930
931 /* At this point (mx,my) are full-pell and the relative displacement */
932 ppix = c->ref[0][0] + (my * s->linesize) + mx;
933
934 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
935
936 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
937 c->mc_mb_var_sum_temp += (vard+128)>>8;
938
939 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
940 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
941 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
942 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
943
944 if (vard*2 + 200*256 > varc)
945 mb_type|= CANDIDATE_MB_TYPE_INTRA;
946 if (varc*2 + 200*256 > vard || s->qscale > 24){
947 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
948 mb_type|= CANDIDATE_MB_TYPE_INTER;
949 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
950 if (s->mpv_flags & FF_MPV_FLAG_MV0)
951 if(mx || my)
952 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
953 }else{
954 mx <<=shift;
955 my <<=shift;
956 }
957 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
958 && !c->skip && varc>50<<8 && vard>10<<8){
959 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
960 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
961
962 set_p_mv_tables(s, mx, my, 0);
963 }else
964 set_p_mv_tables(s, mx, my, 1);
965 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
966 && !c->skip){ //FIXME varc/d checks
967 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
968 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
969 }
970 }else{
971 int intra_score, i;
972 mb_type= CANDIDATE_MB_TYPE_INTER;
973
974 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
975 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
976 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
977
978 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
979 && !c->skip && varc>50<<8 && vard>10<<8){
980 int dmin4= h263_mv4_search(s, mx, my, shift);
981 if(dmin4 < dmin){
982 mb_type= CANDIDATE_MB_TYPE_INTER4V;
983 dmin=dmin4;
984 }
985 }
986 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
987 && !c->skip){ //FIXME varc/d checks
988 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
989 if(dmin_i < dmin){
990 mb_type = CANDIDATE_MB_TYPE_INTER_I;
991 dmin= dmin_i;
992 }
993 }
994
995 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
996
997 /* get intra luma score */
998 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
999 intra_score= varc - 500;
1000 }else{
1001 unsigned mean = (sum+128)>>8;
1002 mean*= 0x01010101;
1003
1004 for(i=0; i<16; i++){
1005 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1006 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1007 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1008 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1009 }
1010
1011 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1012 }
1013 intra_score += c->mb_penalty_factor*16;
1014
1015 if(intra_score < dmin){
1016 mb_type= CANDIDATE_MB_TYPE_INTRA;
1017 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1018 }else
1019 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1020
1021 {
1022 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1023 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1024 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1025 }
1026 }
1027
1028 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1029 }
1030
1031 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1032 int mb_x, int mb_y)
1033 {
1034 MotionEstContext * const c= &s->me;
1035 int mx, my, dmin;
1036 int P[10][2];
1037 const int shift= 1+s->quarter_sample;
1038 const int xy= mb_x + mb_y*s->mb_stride;
1039 init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1040
1041 assert(s->quarter_sample==0 || s->quarter_sample==1);
1042
1043 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1044 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1045
1046 get_limits(s, 16*mb_x, 16*mb_y);
1047 c->skip=0;
1048
1049 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1050 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1051
1052 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1053
1054 /* special case for first line */
1055 if (s->first_slice_line) {
1056 c->pred_x= P_LEFT[0];
1057 c->pred_y= P_LEFT[1];
1058 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1059 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1060 } else {
1061 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1062 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1063 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1064 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1065 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1066 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1067 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1068
1069 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1070 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1071
1072 c->pred_x = P_MEDIAN[0];
1073 c->pred_y = P_MEDIAN[1];
1074 }
1075
1076 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1077
1078 s->p_mv_table[xy][0] = mx<<shift;
1079 s->p_mv_table[xy][1] = my<<shift;
1080
1081 return dmin;
1082 }
1083
1084 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1085 int16_t (*mv_table)[2], int ref_index, int f_code)
1086 {
1087 MotionEstContext * const c= &s->me;
1088 int mx = 0, my = 0, dmin = 0;
1089 int P[10][2];
1090 const int shift= 1+s->quarter_sample;
1091 const int mot_stride = s->mb_stride;
1092 const int mot_xy = mb_y*mot_stride + mb_x;
1093 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1094 int mv_scale;
1095
1096 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1097 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1098 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1099 c->current_mv_penalty= mv_penalty;
1100
1101 get_limits(s, 16*mb_x, 16*mb_y);
1102
1103 if (s->motion_est != FF_ME_ZERO) {
1104 P_LEFT[0] = mv_table[mot_xy - 1][0];
1105 P_LEFT[1] = mv_table[mot_xy - 1][1];
1106
1107 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1108
1109 /* special case for first line */
1110 if (!s->first_slice_line) {
1111 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1112 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1113 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1114 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1115 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1116 if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);
1117 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << 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_LEFT[0];
1123 c->pred_y = P_LEFT[1];
1124
1125 if(mv_table == s->b_forw_mv_table){
1126 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1127 }else{
1128 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1129 }
1130
1131 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1132 }
1133
1134 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1135
1136 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1137 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1138
1139 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1140 mv_table[mot_xy][0]= mx;
1141 mv_table[mot_xy][1]= my;
1142
1143 return dmin;
1144 }
1145
1146 static inline int check_bidir_mv(MpegEncContext * s,
1147 int motion_fx, int motion_fy,
1148 int motion_bx, int motion_by,
1149 int pred_fx, int pred_fy,
1150 int pred_bx, int pred_by,
1151 int size, int h)
1152 {
1153 //FIXME optimize?
1154 //FIXME better f_code prediction (max mv & distance)
1155 //FIXME pointers
1156 MotionEstContext * const c= &s->me;
1157 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1158 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1159 int stride= c->stride;
1160 uint8_t *dest_y = c->scratchpad;
1161 uint8_t *ptr;
1162 int dxy;
1163 int src_x, src_y;
1164 int fbmin;
1165 uint8_t **src_data= c->src[0];
1166 uint8_t **ref_data= c->ref[0];
1167 uint8_t **ref2_data= c->ref[2];
1168
1169 if(s->quarter_sample){
1170 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1171 src_x = motion_fx >> 2;
1172 src_y = motion_fy >> 2;
1173
1174 ptr = ref_data[0] + (src_y * stride) + src_x;
1175 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1176
1177 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1178 src_x = motion_bx >> 2;
1179 src_y = motion_by >> 2;
1180
1181 ptr = ref2_data[0] + (src_y * stride) + src_x;
1182 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1183 }else{
1184 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1185 src_x = motion_fx >> 1;
1186 src_y = motion_fy >> 1;
1187
1188 ptr = ref_data[0] + (src_y * stride) + src_x;
1189 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1190
1191 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1192 src_x = motion_bx >> 1;
1193 src_y = motion_by >> 1;
1194
1195 ptr = ref2_data[0] + (src_y * stride) + src_x;
1196 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1197 }
1198
1199 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1200 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1201 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1202
1203 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1204 }
1205 //FIXME CHROMA !!!
1206
1207 return fbmin;
1208 }
1209
1210 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1211 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1212 {
1213 MotionEstContext * const c= &s->me;
1214 const int mot_stride = s->mb_stride;
1215 const int xy = mb_y *mot_stride + mb_x;
1216 int fbmin;
1217 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1218 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1219 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1220 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1221 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1222 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1223 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1224 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1225 const int flags= c->sub_flags;
1226 const int qpel= flags&FLAG_QPEL;
1227 const int shift= 1+qpel;
1228 const int xmin= c->xmin<<shift;
1229 const int ymin= c->ymin<<shift;
1230 const int xmax= c->xmax<<shift;
1231 const int ymax= c->ymax<<shift;
1232 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1233 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1234 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1235 uint8_t map[256] = { 0 };
1236
1237 map[hashidx&255] = 1;
1238
1239 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1240 motion_bx, motion_by,
1241 pred_fx, pred_fy,
1242 pred_bx, pred_by,
1243 0, 16);
1244
1245 if(s->avctx->bidir_refine){
1246 int end;
1247 static const uint8_t limittab[5]={0,8,32,64,80};
1248 const int limit= limittab[s->avctx->bidir_refine];
1249 static const int8_t vect[][4]={
1250 { 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},
1251
1252 { 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},
1253 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1254 { 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},
1255 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1256
1257 { 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},
1258 { 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},
1259 { 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},
1260 { 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},
1261
1262 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1263 { 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},
1264 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1265 };
1266 static const uint8_t hash[]={
1267 HASH8( 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),
1268
1269 HASH8( 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),
1270 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1271 HASH8( 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),
1272 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1273
1274 HASH8( 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),
1275 HASH8( 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),
1276 HASH8( 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),
1277 HASH8( 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),
1278
1279 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1280 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), HASH8( 1,-1,-1,-1),
1281 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),
1282 };
1283
1284 #define CHECK_BIDIR(fx,fy,bx,by)\
1285 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1286 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1287 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1288 int score;\
1289 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1290 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);\
1291 if(score < fbmin){\
1292 hashidx += HASH(fx,fy,bx,by);\
1293 fbmin= score;\
1294 motion_fx+=fx;\
1295 motion_fy+=fy;\
1296 motion_bx+=bx;\
1297 motion_by+=by;\
1298 end=0;\
1299 }\
1300 }
1301 #define CHECK_BIDIR2(a,b,c,d)\
1302 CHECK_BIDIR(a,b,c,d)\
1303 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1304
1305 do{
1306 int i;
1307 int borderdist=0;
1308 end=1;
1309
1310 CHECK_BIDIR2(0,0,0,1)
1311 CHECK_BIDIR2(0,0,1,0)
1312 CHECK_BIDIR2(0,1,0,0)
1313 CHECK_BIDIR2(1,0,0,0)
1314
1315 for(i=8; i<limit; i++){
1316 int fx= motion_fx+vect[i][0];
1317 int fy= motion_fy+vect[i][1];
1318 int bx= motion_bx+vect[i][2];
1319 int by= motion_by+vect[i][3];
1320 if(borderdist<=0){
1321 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1322 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1323 if((a|b) < 0)
1324 map[(hashidx+hash[i])&255] = 1;
1325 }
1326 if(!map[(hashidx+hash[i])&255]){
1327 int score;
1328 map[(hashidx+hash[i])&255] = 1;
1329 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1330 if(score < fbmin){
1331 hashidx += hash[i];
1332 fbmin= score;
1333 motion_fx=fx;
1334 motion_fy=fy;
1335 motion_bx=bx;
1336 motion_by=by;
1337 end=0;
1338 borderdist--;
1339 if(borderdist<=0){
1340 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1341 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1342 borderdist= FFMIN(a,b);
1343 }
1344 }
1345 }
1346 }
1347 }while(!end);
1348 }
1349
1350 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1351 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1352 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1353 s->b_bidir_back_mv_table[xy][1]= motion_by;
1354
1355 return fbmin;
1356 }
1357
1358 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1359 {
1360 MotionEstContext * const c= &s->me;
1361 int P[10][2];
1362 const int mot_stride = s->mb_stride;
1363 const int mot_xy = mb_y*mot_stride + mb_x;
1364 const int shift= 1+s->quarter_sample;
1365 int dmin, i;
1366 const int time_pp= s->pp_time;
1367 const int time_pb= s->pb_time;
1368 int mx, my, xmin, xmax, ymin, ymax;
1369 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1370
1371 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1372 ymin= xmin=(-32)>>shift;
1373 ymax= xmax= 31>>shift;
1374
1375 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1376 s->mv_type= MV_TYPE_8X8;
1377 }else{
1378 s->mv_type= MV_TYPE_16X16;
1379 }
1380
1381 for(i=0; i<4; i++){
1382 int index= s->block_index[i];
1383 int min, max;
1384
1385 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1386 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1387 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1388 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1389 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1390 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1391
1392 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1393 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1394 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1395 min+= 16*mb_x - 1;
1396 xmax= FFMIN(xmax, s->width - max);
1397 xmin= FFMAX(xmin, - 16 - min);
1398
1399 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1400 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1401 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1402 min+= 16*mb_y - 1;
1403 ymax= FFMIN(ymax, s->height - max);
1404 ymin= FFMAX(ymin, - 16 - min);
1405
1406 if(s->mv_type == MV_TYPE_16X16) break;
1407 }
1408
1409 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1410
1411 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1412 s->b_direct_mv_table[mot_xy][0]= 0;
1413 s->b_direct_mv_table[mot_xy][1]= 0;
1414
1415 return 256*256*256*64;
1416 }
1417
1418 c->xmin= xmin;
1419 c->ymin= ymin;
1420 c->xmax= xmax;
1421 c->ymax= ymax;
1422 c->flags |= FLAG_DIRECT;
1423 c->sub_flags |= FLAG_DIRECT;
1424 c->pred_x=0;
1425 c->pred_y=0;
1426
1427 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1428 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1429
1430 /* special case for first line */
1431 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1432 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1433 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1434 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1435 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
1436
1437 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1438 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1439 }
1440
1441 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1442 if(c->sub_flags&FLAG_QPEL)
1443 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1444 else
1445 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1446
1447 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1448 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1449
1450 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1451
1452 mv_table[mot_xy][0]= mx;
1453 mv_table[mot_xy][1]= my;
1454 c->flags &= ~FLAG_DIRECT;
1455 c->sub_flags &= ~FLAG_DIRECT;
1456
1457 return dmin;
1458 }
1459
1460 void ff_estimate_b_frame_motion(MpegEncContext * s,
1461 int mb_x, int mb_y)
1462 {
1463 MotionEstContext * const c= &s->me;
1464 const int penalty_factor= c->mb_penalty_factor;
1465 int fmin, bmin, dmin, fbmin, bimin, fimin;
1466 int type=0;
1467 const int xy = mb_y*s->mb_stride + mb_x;
1468 init_ref(c, s->new_picture.f->data, s->last_picture.f->data,
1469 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1470
1471 get_limits(s, 16*mb_x, 16*mb_y);
1472
1473 c->skip=0;
1474
1475 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1476 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1477
1478 score= ((unsigned)(score*score + 128*256))>>16;
1479 c->mc_mb_var_sum_temp += score;
1480 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1481 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1482
1483 return;
1484 }
1485
1486 if (s->codec_id == AV_CODEC_ID_MPEG4)
1487 dmin= direct_search(s, mb_x, mb_y);
1488 else
1489 dmin= INT_MAX;
1490 // FIXME penalty stuff for non-MPEG-4
1491 c->skip=0;
1492 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1493 3 * penalty_factor;
1494
1495 c->skip=0;
1496 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1497 2 * penalty_factor;
1498 ff_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1499
1500 c->skip=0;
1501 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1502 ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1503
1504 if (s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME) {
1505 //FIXME mb type penalty
1506 c->skip=0;
1507 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1508 fimin= interlaced_search(s, 0,
1509 s->b_field_mv_table[0], s->b_field_select_table[0],
1510 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1511 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1512 bimin= interlaced_search(s, 2,
1513 s->b_field_mv_table[1], s->b_field_select_table[1],
1514 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1515 }else
1516 fimin= bimin= INT_MAX;
1517
1518 {
1519 int score= fmin;
1520 type = CANDIDATE_MB_TYPE_FORWARD;
1521
1522 if (dmin <= score){
1523 score = dmin;
1524 type = CANDIDATE_MB_TYPE_DIRECT;
1525 }
1526 if(bmin<score){
1527 score=bmin;
1528 type= CANDIDATE_MB_TYPE_BACKWARD;
1529 }
1530 if(fbmin<score){
1531 score=fbmin;
1532 type= CANDIDATE_MB_TYPE_BIDIR;
1533 }
1534 if(fimin<score){
1535 score=fimin;
1536 type= CANDIDATE_MB_TYPE_FORWARD_I;
1537 }
1538 if(bimin<score){
1539 score=bimin;
1540 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1541 }
1542
1543 score= ((unsigned)(score*score + 128*256))>>16;
1544 c->mc_mb_var_sum_temp += score;
1545 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1546 }
1547
1548 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1549 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1550 if(fimin < INT_MAX)
1551 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1552 if(bimin < INT_MAX)
1553 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1554 if(fimin < INT_MAX && bimin < INT_MAX){
1555 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1556 }
1557 //FIXME something smarter
1558 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1559 if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1560 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1561 type |= CANDIDATE_MB_TYPE_DIRECT0;
1562 }
1563
1564 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1565 }
1566
1567 /* find best f_code for ME which do unlimited searches */
1568 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1569 {
1570 if (s->motion_est != FF_ME_ZERO) {
1571 int score[8];
1572 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1573 uint8_t * fcode_tab= s->fcode_tab;
1574 int best_fcode=-1;
1575 int best_score=-10000000;
1576
1577 if(s->msmpeg4_version)
1578 range= FFMIN(range, 16);
1579 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1580 range= FFMIN(range, 256);
1581
1582 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1583
1584 for(y=0; y<s->mb_height; y++){
1585 int x;
1586 int xy= y*s->mb_stride;
1587 for(x=0; x<s->mb_width; x++){
1588 if(s->mb_type[xy] & type){
1589 int mx= mv_table[xy][0];
1590 int my= mv_table[xy][1];
1591 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1592 fcode_tab[my + MAX_MV]);
1593 int j;
1594
1595 if(mx >= range || mx < -range ||
1596 my >= range || my < -range)
1597 continue;
1598
1599 for(j=0; j<fcode && j<8; j++){
1600 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1601 score[j]-= 170;
1602 }
1603 }
1604 xy++;
1605 }
1606 }
1607
1608 for(i=1; i<8; i++){
1609 if(score[i] > best_score){
1610 best_score= score[i];
1611 best_fcode= i;
1612 }
1613 }
1614
1615 return best_fcode;
1616 }else{
1617 return 1;
1618 }
1619 }
1620
1621 void ff_fix_long_p_mvs(MpegEncContext * s)
1622 {
1623 MotionEstContext * const c= &s->me;
1624 const int f_code= s->f_code;
1625 int y, range;
1626 assert(s->pict_type==AV_PICTURE_TYPE_P);
1627
1628 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1629
1630 assert(range <= 16 || !s->msmpeg4_version);
1631 assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1632
1633 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1634
1635 if (s->avctx->flags & AV_CODEC_FLAG_4MV) {
1636 const int wrap= s->b8_stride;
1637
1638 /* clip / convert to intra 8x8 type MVs */
1639 for(y=0; y<s->mb_height; y++){
1640 int xy= y*2*wrap;
1641 int i= y*s->mb_stride;
1642 int x;
1643
1644 for(x=0; x<s->mb_width; x++){
1645 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1646 int block;
1647 for(block=0; block<4; block++){
1648 int off= (block& 1) + (block>>1)*wrap;
1649 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1650 int my = s->current_picture.motion_val[0][ xy + off ][1];
1651
1652 if( mx >=range || mx <-range
1653 || my >=range || my <-range){
1654 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1655 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1656 s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
1657 }
1658 }
1659 }
1660 xy+=2;
1661 i++;
1662 }
1663 }
1664 }
1665 }
1666
1667 /**
1668 * @param truncate 1 for truncation, 0 for using intra
1669 */
1670 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1671 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1672 {
1673 MotionEstContext * const c= &s->me;
1674 int y, h_range, v_range;
1675
1676 // RAL: 8 in MPEG-1, 16 in MPEG-4
1677 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1678
1679 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1680
1681 h_range= range;
1682 v_range= field_select_table ? range>>1 : range;
1683
1684 /* clip / convert to intra 16x16 type MVs */
1685 for(y=0; y<s->mb_height; y++){
1686 int x;
1687 int xy= y*s->mb_stride;
1688 for(x=0; x<s->mb_width; x++){
1689 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1690 if (!field_select_table || field_select_table[xy] == field_select) {
1691 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1692 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1693
1694 if(truncate){
1695 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1696 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1697 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1698 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1699 }else{
1700 s->mb_type[xy] &= ~type;
1701 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1702 mv_table[xy][0]=
1703 mv_table[xy][1]= 0;
1704 }
1705 }
1706 }
1707 }
1708 xy++;
1709 }
1710 }
1711 }
Modern FFMPEG