Mercurial > libavcodec.hg
annotate motion_est.c @ 1007:b2cf2a1d9a51 libavcodec
more compare functions (rd & bit)
author | michaelni |
---|---|
date | Wed, 15 Jan 2003 18:05:23 +0000 |
parents | dd421045b4ce |
children | 3b7fcfb9c551 |
rev | line source |
---|---|
0 | 1 /* |
2 * Motion estimation | |
429 | 3 * Copyright (c) 2000,2001 Fabrice Bellard. |
455 | 4 * Copyright (c) 2002 Michael Niedermayer |
0 | 5 * |
6 * | |
429 | 7 * This library is free software; you can redistribute it and/or |
8 * modify it under the terms of the GNU Lesser General Public | |
9 * License as published by the Free Software Foundation; either | |
10 * version 2 of the License, or (at your option) any later version. | |
0 | 11 * |
429 | 12 * This library is distributed in the hope that it will be useful, |
0 | 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
429 | 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
15 * Lesser General Public License for more details. | |
0 | 16 * |
429 | 17 * You should have received a copy of the GNU Lesser General Public |
18 * License along with this library; if not, write to the Free Software | |
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
20 * |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
21 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at> |
0 | 22 */ |
23 #include <stdlib.h> | |
24 #include <stdio.h> | |
25 #include "avcodec.h" | |
26 #include "dsputil.h" | |
27 #include "mpegvideo.h" | |
28 | |
936 | 29 //#undef NDEBUG |
30 //#include <assert.h> | |
31 | |
455 | 32 #define SQ(a) ((a)*(a)) |
284 | 33 |
455 | 34 #define P_LEFT P[1] |
35 #define P_TOP P[2] | |
36 #define P_TOPRIGHT P[3] | |
37 #define P_MEDIAN P[4] | |
38 #define P_MV1 P[9] | |
39 | |
936 | 40 static inline int sad_hpel_motion_search(MpegEncContext * s, |
41 int *mx_ptr, int *my_ptr, int dmin, | |
42 int xmin, int ymin, int xmax, int ymax, | |
43 int pred_x, int pred_y, Picture *picture, | |
44 int n, int size, uint16_t * const mv_penalty); | |
0 | 45 |
936 | 46 static inline int update_map_generation(MpegEncContext * s) |
47 { | |
48 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2); | |
49 if(s->me.map_generation==0){ | |
50 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2); | |
51 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE); | |
52 } | |
53 return s->me.map_generation; | |
54 } | |
55 | |
948 | 56 /* shape adaptive search stuff */ |
57 typedef struct Minima{ | |
58 int height; | |
59 int x, y; | |
60 int checked; | |
61 }Minima; | |
936 | 62 |
948 | 63 static int minima_cmp(const void *a, const void *b){ |
64 Minima *da = (Minima *) a; | |
65 Minima *db = (Minima *) b; | |
66 | |
67 return da->height - db->height; | |
68 } | |
936 | 69 |
70 /* SIMPLE */ | |
71 #define RENAME(a) simple_ ## a | |
72 | |
73 #define CMP(d, x, y, size)\ | |
74 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride); | |
75 | |
76 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
77 {\ | |
78 const int dxy= (dx) + 2*(dy);\ | |
79 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\ | |
80 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
81 } | |
82 | |
83 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
84 {\ | |
85 const int dxy= (dx) + 4*(dy);\ | |
86 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\ | |
87 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
88 } | |
89 | |
90 #include "motion_est_template.c" | |
91 #undef RENAME | |
92 #undef CMP | |
93 #undef CMP_HPEL | |
94 #undef CMP_QPEL | |
95 #undef INIT | |
96 | |
97 /* SIMPLE CHROMA */ | |
98 #define RENAME(a) simple_chroma_ ## a | |
99 | |
100 #define CMP(d, x, y, size)\ | |
101 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\ | |
102 if(chroma_cmp){\ | |
103 int dxy= ((x)&1) + 2*((y)&1);\ | |
104 int c= ((x)>>1) + ((y)>>1)*uvstride;\ | |
105 \ | |
106 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
107 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\ | |
108 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
109 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\ | |
110 } | |
111 | |
112 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
113 {\ | |
114 const int dxy= (dx) + 2*(dy);\ | |
115 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\ | |
116 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
117 if(chroma_cmp_sub){\ | |
118 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\ | |
119 int c= ((x)>>1) + ((y)>>1)*uvstride;\ | |
120 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
121 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\ | |
122 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
123 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\ | |
124 }\ | |
125 } | |
126 | |
127 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
128 {\ | |
129 const int dxy= (dx) + 4*(dy);\ | |
130 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\ | |
131 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
132 if(chroma_cmp_sub){\ | |
133 int cxy, c;\ | |
134 int cx= (4*(x) + (dx))/2;\ | |
135 int cy= (4*(y) + (dy))/2;\ | |
136 cx= (cx>>1)|(cx&1);\ | |
137 cy= (cy>>1)|(cy&1);\ | |
138 cxy= (cx&1) + 2*(cy&1);\ | |
139 c= ((cx)>>1) + ((cy)>>1)*uvstride;\ | |
140 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
141 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\ | |
142 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
143 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\ | |
144 }\ | |
145 } | |
146 | |
147 #include "motion_est_template.c" | |
148 #undef RENAME | |
149 #undef CMP | |
150 #undef CMP_HPEL | |
151 #undef CMP_QPEL | |
152 #undef INIT | |
153 | |
154 /* SIMPLE DIRECT HPEL */ | |
155 #define RENAME(a) simple_direct_hpel_ ## a | |
156 //FIXME precalc divisions stuff | |
157 | |
158 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\ | |
159 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\ | |
160 const int hx= 2*(x) + (dx);\ | |
161 const int hy= 2*(y) + (dy);\ | |
162 if(s->mv_type==MV_TYPE_8X8){\ | |
163 int i;\ | |
164 for(i=0; i<4; i++){\ | |
165 int fx = s->me.direct_basis_mv[i][0] + hx;\ | |
166 int fy = s->me.direct_basis_mv[i][1] + hy;\ | |
167 int bx = hx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\ | |
168 int by = hy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\ | |
169 int fxy= (fx&1) + 2*(fy&1);\ | |
170 int bxy= (bx&1) + 2*(by&1);\ | |
171 \ | |
172 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\ | |
173 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\ | |
174 hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\ | |
175 }\ | |
176 }else{\ | |
177 int fx = s->me.direct_basis_mv[0][0] + hx;\ | |
178 int fy = s->me.direct_basis_mv[0][1] + hy;\ | |
179 int bx = hx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\ | |
180 int by = hy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\ | |
181 int fxy= (fx&1) + 2*(fy&1);\ | |
182 int bxy= (bx&1) + 2*(by&1);\ | |
183 \ | |
184 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\ | |
185 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\ | |
186 }\ | |
187 d = cmp_func(s, s->me.scratchpad, src_y, stride);\ | |
188 }else\ | |
189 d= 256*256*256*32; | |
190 | |
191 | |
192 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
193 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub) | |
194 | |
195 #define CMP(d, x, y, size)\ | |
196 CMP_DIRECT(d, 0, 0, x, y, size, cmp) | |
197 | |
198 #include "motion_est_template.c" | |
199 #undef RENAME | |
200 #undef CMP | |
201 #undef CMP_HPEL | |
202 #undef CMP_QPEL | |
203 #undef INIT | |
204 #undef CMP_DIRECT | |
205 | |
206 /* SIMPLE DIRECT QPEL */ | |
207 #define RENAME(a) simple_direct_qpel_ ## a | |
208 | |
209 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\ | |
210 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\ | |
211 const int qx= 4*(x) + (dx);\ | |
212 const int qy= 4*(y) + (dy);\ | |
213 if(s->mv_type==MV_TYPE_8X8){\ | |
214 int i;\ | |
215 for(i=0; i<4; i++){\ | |
216 int fx = s->me.direct_basis_mv[i][0] + qx;\ | |
217 int fy = s->me.direct_basis_mv[i][1] + qy;\ | |
218 int bx = qx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\ | |
219 int by = qy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\ | |
220 int fxy= (fx&3) + 4*(fy&3);\ | |
221 int bxy= (bx&3) + 4*(by&3);\ | |
222 \ | |
223 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\ | |
224 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\ | |
225 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\ | |
226 }\ | |
227 }else{\ | |
228 int fx = s->me.direct_basis_mv[0][0] + qx;\ | |
229 int fy = s->me.direct_basis_mv[0][1] + qy;\ | |
230 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\ | |
231 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\ | |
232 int fxy= (fx&3) + 4*(fy&3);\ | |
233 int bxy= (bx&3) + 4*(by&3);\ | |
234 \ | |
235 qpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\ | |
236 qpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\ | |
237 }\ | |
238 d = cmp_func(s, s->me.scratchpad, src_y, stride);\ | |
239 }else\ | |
240 d= 256*256*256*32; | |
241 | |
242 | |
243 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
244 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub) | |
245 | |
246 #define CMP(d, x, y, size)\ | |
247 CMP_DIRECT(d, 0, 0, x, y, size, cmp) | |
248 | |
249 #include "motion_est_template.c" | |
250 #undef RENAME | |
251 #undef CMP | |
252 #undef CMP_HPEL | |
253 #undef CMP_QPEL | |
254 #undef INIT | |
255 #undef CMP__DIRECT | |
256 | |
257 | |
258 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){ | |
259 return 0; | |
260 } | |
261 | |
262 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){ | |
263 DSPContext* c= &s->dsp; | |
264 int i; | |
265 | |
266 memset(cmp, 0, sizeof(void*)*11); | |
267 | |
268 switch(type&0xFF){ | |
269 case FF_CMP_SAD: | |
270 cmp[0]= c->sad[0]; | |
271 cmp[1]= c->sad[1]; | |
272 break; | |
273 case FF_CMP_SATD: | |
274 cmp[0]= c->hadamard8_diff[0]; | |
275 cmp[1]= c->hadamard8_diff[1]; | |
276 break; | |
277 case FF_CMP_SSE: | |
278 cmp[0]= c->sse[0]; | |
279 cmp[1]= c->sse[1]; | |
280 break; | |
281 case FF_CMP_DCT: | |
282 cmp[0]= c->dct_sad[0]; | |
283 cmp[1]= c->dct_sad[1]; | |
284 break; | |
285 case FF_CMP_PSNR: | |
286 cmp[0]= c->quant_psnr[0]; | |
287 cmp[1]= c->quant_psnr[1]; | |
288 break; | |
1007 | 289 case FF_CMP_BIT: |
290 cmp[0]= c->bit[0]; | |
291 cmp[1]= c->bit[1]; | |
292 break; | |
293 case FF_CMP_RD: | |
294 cmp[0]= c->rd[0]; | |
295 cmp[1]= c->rd[1]; | |
296 break; | |
936 | 297 case FF_CMP_ZERO: |
298 for(i=0; i<7; i++){ | |
299 cmp[i]= zero_cmp; | |
300 } | |
301 break; | |
302 default: | |
303 fprintf(stderr,"internal error in cmp function selection\n"); | |
304 } | |
305 }; | |
306 | |
307 static inline int get_penalty_factor(MpegEncContext *s, int type){ | |
308 | |
309 switch(type){ | |
310 default: | |
311 case FF_CMP_SAD: | |
312 return s->qscale; | |
313 case FF_CMP_DCT: | |
314 case FF_CMP_SATD: | |
1007 | 315 case FF_CMP_SSE: |
936 | 316 return s->qscale*8; |
1007 | 317 case FF_CMP_BIT: |
318 return 1; | |
319 case FF_CMP_RD: | |
320 return (s->qscale*s->qscale*105 + 64)>>7; | |
936 | 321 } |
322 } | |
323 | |
324 void ff_init_me(MpegEncContext *s){ | |
954 | 325 set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp); |
936 | 326 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp); |
327 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp); | |
328 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp); | |
329 | |
330 if(s->flags&CODEC_FLAG_QPEL){ | |
331 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
332 s->me.sub_motion_search= simple_chroma_qpel_motion_search; | |
333 else | |
334 s->me.sub_motion_search= simple_qpel_motion_search; | |
335 }else{ | |
336 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
337 s->me.sub_motion_search= simple_chroma_hpel_motion_search; | |
338 else if(s->avctx->me_sub_cmp == FF_CMP_SAD && s->avctx->me_cmp == FF_CMP_SAD) | |
339 s->me.sub_motion_search= sad_hpel_motion_search; | |
340 else | |
341 s->me.sub_motion_search= simple_hpel_motion_search; | |
342 } | |
343 | |
344 if(s->avctx->me_cmp&FF_CMP_CHROMA){ | |
345 s->me.motion_search[0]= simple_chroma_epzs_motion_search; | |
346 s->me.motion_search[1]= simple_chroma_epzs_motion_search4; | |
347 }else{ | |
348 s->me.motion_search[0]= simple_epzs_motion_search; | |
349 s->me.motion_search[1]= simple_epzs_motion_search4; | |
350 } | |
954 | 351 |
352 if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){ | |
353 s->me.pre_motion_search= simple_chroma_epzs_motion_search; | |
354 }else{ | |
355 s->me.pre_motion_search= simple_epzs_motion_search; | |
356 } | |
936 | 357 } |
358 | |
284 | 359 static int pix_dev(UINT8 * pix, int line_size, int mean) |
360 { | |
361 int s, i, j; | |
362 | |
363 s = 0; | |
364 for (i = 0; i < 16; i++) { | |
365 for (j = 0; j < 16; j += 8) { | |
366 s += ABS(pix[0]-mean); | |
367 s += ABS(pix[1]-mean); | |
368 s += ABS(pix[2]-mean); | |
369 s += ABS(pix[3]-mean); | |
370 s += ABS(pix[4]-mean); | |
371 s += ABS(pix[5]-mean); | |
372 s += ABS(pix[6]-mean); | |
373 s += ABS(pix[7]-mean); | |
374 pix += 8; | |
375 } | |
376 pix += line_size - 16; | |
377 } | |
378 return s; | |
379 } | |
380 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
381 static inline void no_motion_search(MpegEncContext * s, |
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
382 int *mx_ptr, int *my_ptr) |
0 | 383 { |
384 *mx_ptr = 16 * s->mb_x; | |
385 *my_ptr = 16 * s->mb_y; | |
386 } | |
387 | |
388 static int full_motion_search(MpegEncContext * s, | |
389 int *mx_ptr, int *my_ptr, int range, | |
324 | 390 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 391 { |
392 int x1, y1, x2, y2, xx, yy, x, y; | |
393 int mx, my, dmin, d; | |
394 UINT8 *pix; | |
395 | |
396 xx = 16 * s->mb_x; | |
397 yy = 16 * s->mb_y; | |
398 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */ | |
399 if (x1 < xmin) | |
400 x1 = xmin; | |
401 x2 = xx + range - 1; | |
402 if (x2 > xmax) | |
403 x2 = xmax; | |
404 y1 = yy - range + 1; | |
405 if (y1 < ymin) | |
406 y1 = ymin; | |
407 y2 = yy + range - 1; | |
408 if (y2 > ymax) | |
409 y2 = ymax; | |
903 | 410 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 411 dmin = 0x7fffffff; |
412 mx = 0; | |
413 my = 0; | |
414 for (y = y1; y <= y2; y++) { | |
415 for (x = x1; x <= x2; x++) { | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
416 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, |
294 | 417 s->linesize); |
0 | 418 if (d < dmin || |
419 (d == dmin && | |
420 (abs(x - xx) + abs(y - yy)) < | |
421 (abs(mx - xx) + abs(my - yy)))) { | |
422 dmin = d; | |
423 mx = x; | |
424 my = y; | |
425 } | |
426 } | |
427 } | |
428 | |
429 *mx_ptr = mx; | |
430 *my_ptr = my; | |
431 | |
432 #if 0 | |
433 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) || | |
434 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) { | |
435 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr); | |
436 } | |
437 #endif | |
438 return dmin; | |
439 } | |
440 | |
441 | |
442 static int log_motion_search(MpegEncContext * s, | |
443 int *mx_ptr, int *my_ptr, int range, | |
324 | 444 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 445 { |
446 int x1, y1, x2, y2, xx, yy, x, y; | |
447 int mx, my, dmin, d; | |
448 UINT8 *pix; | |
449 | |
450 xx = s->mb_x << 4; | |
451 yy = s->mb_y << 4; | |
452 | |
453 /* Left limit */ | |
454 x1 = xx - range; | |
455 if (x1 < xmin) | |
456 x1 = xmin; | |
457 | |
458 /* Right limit */ | |
459 x2 = xx + range; | |
460 if (x2 > xmax) | |
461 x2 = xmax; | |
462 | |
463 /* Upper limit */ | |
464 y1 = yy - range; | |
465 if (y1 < ymin) | |
466 y1 = ymin; | |
467 | |
468 /* Lower limit */ | |
469 y2 = yy + range; | |
470 if (y2 > ymax) | |
471 y2 = ymax; | |
472 | |
903 | 473 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 474 dmin = 0x7fffffff; |
475 mx = 0; | |
476 my = 0; | |
477 | |
478 do { | |
479 for (y = y1; y <= y2; y += range) { | |
480 for (x = x1; x <= x2; x += range) { | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
481 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 482 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
483 dmin = d; | |
484 mx = x; | |
485 my = y; | |
486 } | |
487 } | |
488 } | |
489 | |
490 range = range >> 1; | |
491 | |
492 x1 = mx - range; | |
493 if (x1 < xmin) | |
494 x1 = xmin; | |
495 | |
496 x2 = mx + range; | |
497 if (x2 > xmax) | |
498 x2 = xmax; | |
499 | |
500 y1 = my - range; | |
501 if (y1 < ymin) | |
502 y1 = ymin; | |
503 | |
504 y2 = my + range; | |
505 if (y2 > ymax) | |
506 y2 = ymax; | |
507 | |
508 } while (range >= 1); | |
509 | |
510 #ifdef DEBUG | |
511 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my); | |
512 #endif | |
513 *mx_ptr = mx; | |
514 *my_ptr = my; | |
515 return dmin; | |
516 } | |
517 | |
518 static int phods_motion_search(MpegEncContext * s, | |
519 int *mx_ptr, int *my_ptr, int range, | |
324 | 520 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 521 { |
522 int x1, y1, x2, y2, xx, yy, x, y, lastx, d; | |
523 int mx, my, dminx, dminy; | |
524 UINT8 *pix; | |
525 | |
526 xx = s->mb_x << 4; | |
527 yy = s->mb_y << 4; | |
528 | |
529 /* Left limit */ | |
530 x1 = xx - range; | |
531 if (x1 < xmin) | |
532 x1 = xmin; | |
533 | |
534 /* Right limit */ | |
535 x2 = xx + range; | |
536 if (x2 > xmax) | |
537 x2 = xmax; | |
538 | |
539 /* Upper limit */ | |
540 y1 = yy - range; | |
541 if (y1 < ymin) | |
542 y1 = ymin; | |
543 | |
544 /* Lower limit */ | |
545 y2 = yy + range; | |
546 if (y2 > ymax) | |
547 y2 = ymax; | |
548 | |
903 | 549 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 550 mx = 0; |
551 my = 0; | |
552 | |
553 x = xx; | |
554 y = yy; | |
555 do { | |
556 dminx = 0x7fffffff; | |
557 dminy = 0x7fffffff; | |
558 | |
559 lastx = x; | |
560 for (x = x1; x <= x2; x += range) { | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
561 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 562 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
563 dminx = d; | |
564 mx = x; | |
565 } | |
566 } | |
567 | |
568 x = lastx; | |
569 for (y = y1; y <= y2; y += range) { | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
570 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 571 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
572 dminy = d; | |
573 my = y; | |
574 } | |
575 } | |
576 | |
577 range = range >> 1; | |
578 | |
579 x = mx; | |
580 y = my; | |
581 x1 = mx - range; | |
582 if (x1 < xmin) | |
583 x1 = xmin; | |
584 | |
585 x2 = mx + range; | |
586 if (x2 > xmax) | |
587 x2 = xmax; | |
588 | |
589 y1 = my - range; | |
590 if (y1 < ymin) | |
591 y1 = ymin; | |
592 | |
593 y2 = my + range; | |
594 if (y2 > ymax) | |
595 y2 = ymax; | |
596 | |
597 } while (range >= 1); | |
598 | |
599 #ifdef DEBUG | |
600 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my); | |
601 #endif | |
602 | |
603 /* half pixel search */ | |
604 *mx_ptr = mx; | |
605 *my_ptr = my; | |
606 return dminy; | |
607 } | |
608 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
609 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
610 #define Z_THRESHOLD 256 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
611 |
936 | 612 #define CHECK_SAD_HALF_MV(suffix, x, y) \ |
455 | 613 {\ |
614 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\ | |
936 | 615 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\ |
455 | 616 COPY3_IF_LT(dminh, d, dx, x, dy, y)\ |
617 } | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
618 |
936 | 619 static inline int sad_hpel_motion_search(MpegEncContext * s, |
0 | 620 int *mx_ptr, int *my_ptr, int dmin, |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
621 int xmin, int ymin, int xmax, int ymax, |
936 | 622 int pred_x, int pred_y, Picture *picture, |
623 int n, int size, uint16_t * const mv_penalty) | |
0 | 624 { |
936 | 625 uint8_t *ref_picture= picture->data[0]; |
626 uint32_t *score_map= s->me.score_map; | |
627 const int penalty_factor= s->me.sub_penalty_factor; | |
455 | 628 int mx, my, xx, yy, dminh; |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
629 UINT8 *pix, *ptr; |
936 | 630 op_pixels_abs_func pix_abs_x2; |
631 op_pixels_abs_func pix_abs_y2; | |
632 op_pixels_abs_func pix_abs_xy2; | |
633 | |
634 if(size==0){ | |
635 pix_abs_x2 = s->dsp.pix_abs16x16_x2; | |
636 pix_abs_y2 = s->dsp.pix_abs16x16_y2; | |
637 pix_abs_xy2= s->dsp.pix_abs16x16_xy2; | |
638 }else{ | |
639 pix_abs_x2 = s->dsp.pix_abs8x8_x2; | |
640 pix_abs_y2 = s->dsp.pix_abs8x8_y2; | |
641 pix_abs_xy2= s->dsp.pix_abs8x8_xy2; | |
863 | 642 } |
455 | 643 |
936 | 644 if(s->me.skip){ |
455 | 645 // printf("S"); |
646 *mx_ptr = 0; | |
647 *my_ptr = 0; | |
648 return dmin; | |
649 } | |
650 // printf("N"); | |
651 | |
652 xx = 16 * s->mb_x + 8*(n&1); | |
653 yy = 16 * s->mb_y + 8*(n>>1); | |
903 | 654 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 655 |
294 | 656 mx = *mx_ptr; |
657 my = *my_ptr; | |
455 | 658 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx); |
659 | |
294 | 660 dminh = dmin; |
661 | |
662 if (mx > xmin && mx < xmax && | |
663 my > ymin && my < ymax) { | |
455 | 664 int dx=0, dy=0; |
665 int d, pen_x, pen_y; | |
666 const int index= (my<<ME_MAP_SHIFT) + mx; | |
667 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
668 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
669 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
670 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
671 mx<<=1; | |
672 my<<=1; | |
294 | 673 |
674 | |
675 pen_x= pred_x + mx; | |
676 pen_y= pred_y + my; | |
677 | |
678 ptr-= s->linesize; | |
455 | 679 if(t<=b){ |
936 | 680 CHECK_SAD_HALF_MV(y2 , 0, -1) |
455 | 681 if(l<=r){ |
936 | 682 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 683 if(t+r<=b+l){ |
936 | 684 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 685 ptr+= s->linesize; |
686 }else{ | |
687 ptr+= s->linesize; | |
936 | 688 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 689 } |
936 | 690 CHECK_SAD_HALF_MV(x2 , -1, 0) |
455 | 691 }else{ |
936 | 692 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 693 if(t+l<=b+r){ |
936 | 694 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 695 ptr+= s->linesize; |
696 }else{ | |
697 ptr+= s->linesize; | |
936 | 698 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 699 } |
936 | 700 CHECK_SAD_HALF_MV(x2 , +1, 0) |
455 | 701 } |
702 }else{ | |
703 if(l<=r){ | |
704 if(t+l<=b+r){ | |
936 | 705 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 706 ptr+= s->linesize; |
707 }else{ | |
708 ptr+= s->linesize; | |
936 | 709 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 710 } |
936 | 711 CHECK_SAD_HALF_MV(x2 , -1, 0) |
712 CHECK_SAD_HALF_MV(xy2, -1, +1) | |
455 | 713 }else{ |
714 if(t+r<=b+l){ | |
936 | 715 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 716 ptr+= s->linesize; |
717 }else{ | |
718 ptr+= s->linesize; | |
936 | 719 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 720 } |
936 | 721 CHECK_SAD_HALF_MV(x2 , +1, 0) |
722 CHECK_SAD_HALF_MV(xy2, +1, +1) | |
455 | 723 } |
936 | 724 CHECK_SAD_HALF_MV(y2 , 0, +1) |
455 | 725 } |
726 mx+=dx; | |
727 my+=dy; | |
294 | 728 |
729 }else{ | |
455 | 730 mx<<=1; |
731 my<<=1; | |
294 | 732 } |
733 | |
734 *mx_ptr = mx; | |
735 *my_ptr = my; | |
455 | 736 return dminh; |
294 | 737 } |
738 | |
455 | 739 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4) |
294 | 740 { |
324 | 741 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2); |
294 | 742 |
324 | 743 s->p_mv_table[xy][0] = mx; |
744 s->p_mv_table[xy][1] = my; | |
294 | 745 |
746 /* has allready been set to the 4 MV if 4MV is done */ | |
455 | 747 if(mv4){ |
294 | 748 int mot_xy= s->block_index[0]; |
749 | |
750 s->motion_val[mot_xy ][0]= mx; | |
751 s->motion_val[mot_xy ][1]= my; | |
752 s->motion_val[mot_xy+1][0]= mx; | |
753 s->motion_val[mot_xy+1][1]= my; | |
754 | |
755 mot_xy += s->block_wrap[0]; | |
756 s->motion_val[mot_xy ][0]= mx; | |
757 s->motion_val[mot_xy ][1]= my; | |
758 s->motion_val[mot_xy+1][0]= mx; | |
759 s->motion_val[mot_xy+1][1]= my; | |
760 } | |
761 } | |
762 | |
324 | 763 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code) |
764 { | |
765 *range = 8 * (1 << (f_code - 1)); | |
766 /* XXX: temporary kludge to avoid overflow for msmpeg4 */ | |
767 if (s->out_format == FMT_H263 && !s->h263_msmpeg4) | |
768 *range *= 2; | |
0 | 769 |
324 | 770 if (s->unrestricted_mv) { |
771 *xmin = -16; | |
772 *ymin = -16; | |
773 if (s->h263_plus) | |
774 *range *= 2; | |
951 | 775 if(s->avctx->codec->id!=CODEC_ID_MPEG4){ |
324 | 776 *xmax = s->mb_width*16; |
777 *ymax = s->mb_height*16; | |
778 }else { | |
779 *xmax = s->width; | |
780 *ymax = s->height; | |
781 } | |
782 } else { | |
783 *xmin = 0; | |
784 *ymin = 0; | |
785 *xmax = s->mb_width*16 - 16; | |
786 *ymax = s->mb_height*16 - 16; | |
787 } | |
788 } | |
789 | |
455 | 790 static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift) |
791 { | |
792 int block; | |
793 int P[10][2]; | |
903 | 794 uint8_t *ref_picture= s->last_picture.data[0]; |
455 | 795 int dmin_sum=0; |
936 | 796 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
455 | 797 |
798 for(block=0; block<4; block++){ | |
799 int mx4, my4; | |
800 int pred_x4, pred_y4; | |
801 int dmin4; | |
802 static const int off[4]= {2, 1, 1, -1}; | |
803 const int mot_stride = s->block_wrap[0]; | |
804 const int mot_xy = s->block_index[block]; | |
805 // const int block_x= (block&1); | |
806 // const int block_y= (block>>1); | |
807 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way | |
808 const int rel_xmin4= xmin; | |
809 const int rel_xmax4= xmax; | |
810 const int rel_ymin4= ymin; | |
811 const int rel_ymax4= ymax; | |
812 #else | |
813 const int rel_xmin4= xmin - block_x*8; | |
814 const int rel_xmax4= xmax - block_x*8 + 8; | |
815 const int rel_ymin4= ymin - block_y*8; | |
816 const int rel_ymax4= ymax - block_y*8 + 8; | |
817 #endif | |
818 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; | |
819 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
820 | |
821 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift); | |
822 | |
823 /* special case for first line */ | |
952 | 824 if (s->mb_y == 0 && block<2) { |
455 | 825 pred_x4= P_LEFT[0]; |
826 pred_y4= P_LEFT[1]; | |
827 } else { | |
828 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; | |
829 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
830 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0]; | |
831 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1]; | |
832 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift); | |
833 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift); | |
834 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift); | |
835 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift); | |
836 | |
837 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
838 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
839 | |
840 if(s->out_format == FMT_H263){ | |
841 pred_x4 = P_MEDIAN[0]; | |
842 pred_y4 = P_MEDIAN[1]; | |
843 }else { /* mpeg1 at least */ | |
844 pred_x4= P_LEFT[0]; | |
845 pred_y4= P_LEFT[1]; | |
846 } | |
847 } | |
848 P_MV1[0]= mx; | |
849 P_MV1[1]= my; | |
850 | |
936 | 851 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
948 | 852 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty); |
455 | 853 |
936 | 854 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
855 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty); | |
455 | 856 |
857 s->motion_val[ s->block_index[block] ][0]= mx4; | |
858 s->motion_val[ s->block_index[block] ][1]= my4; | |
859 dmin_sum+= dmin4; | |
860 } | |
861 return dmin_sum; | |
862 } | |
863 | |
324 | 864 void ff_estimate_p_frame_motion(MpegEncContext * s, |
865 int mb_x, int mb_y) | |
0 | 866 { |
867 UINT8 *pix, *ppix; | |
868 int sum, varc, vard, mx, my, range, dmin, xx, yy; | |
869 int xmin, ymin, xmax, ymax; | |
280 | 870 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
871 int pred_x=0, pred_y=0; |
455 | 872 int P[10][2]; |
280 | 873 const int shift= 1+s->quarter_sample; |
294 | 874 int mb_type=0; |
903 | 875 uint8_t *ref_picture= s->last_picture.data[0]; |
876 Picture * const pic= &s->current_picture; | |
936 | 877 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
878 | |
879 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
880 | |
881 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
882 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
0 | 883 |
324 | 884 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code); |
455 | 885 rel_xmin= xmin - mb_x*16; |
886 rel_xmax= xmax - mb_x*16; | |
887 rel_ymin= ymin - mb_y*16; | |
888 rel_ymax= ymax - mb_y*16; | |
936 | 889 s->me.skip=0; |
324 | 890 |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
891 switch(s->me_method) { |
0 | 892 case ME_ZERO: |
893 default: | |
894 no_motion_search(s, &mx, &my); | |
455 | 895 mx-= mb_x*16; |
896 my-= mb_y*16; | |
0 | 897 dmin = 0; |
898 break; | |
899 case ME_FULL: | |
324 | 900 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); |
455 | 901 mx-= mb_x*16; |
902 my-= mb_y*16; | |
0 | 903 break; |
904 case ME_LOG: | |
324 | 905 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 906 mx-= mb_x*16; |
907 my-= mb_y*16; | |
0 | 908 break; |
909 case ME_PHODS: | |
324 | 910 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 911 mx-= mb_x*16; |
912 my-= mb_y*16; | |
0 | 913 break; |
288 | 914 case ME_X1: |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
915 case ME_EPZS: |
288 | 916 { |
294 | 917 const int mot_stride = s->block_wrap[0]; |
918 const int mot_xy = s->block_index[0]; | |
288 | 919 |
455 | 920 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; |
921 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
288 | 922 |
455 | 923 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
280 | 924 |
952 | 925 if(mb_y) { |
455 | 926 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; |
927 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
928 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0]; | |
929 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1]; | |
930 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift); | |
931 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
932 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
280 | 933 |
455 | 934 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
935 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
936 | |
937 if(s->out_format == FMT_H263){ | |
938 pred_x = P_MEDIAN[0]; | |
939 pred_y = P_MEDIAN[1]; | |
940 }else { /* mpeg1 at least */ | |
941 pred_x= P_LEFT[0]; | |
942 pred_y= P_LEFT[1]; | |
943 } | |
952 | 944 }else{ |
945 pred_x= P_LEFT[0]; | |
946 pred_y= P_LEFT[1]; | |
288 | 947 } |
952 | 948 |
280 | 949 } |
936 | 950 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
948 | 951 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty); |
281
1fc96b02142e
mpeg4 aspect_ratio_info in AVCodecContext (requested by alex)
michaelni
parents:
280
diff
changeset
|
952 |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
953 break; |
0 | 954 } |
955 | |
956 /* intra / predictive decision */ | |
957 xx = mb_x * 16; | |
958 yy = mb_y * 16; | |
959 | |
903 | 960 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 961 /* At this point (mx,my) are full-pell and the relative displacement */ |
962 ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx); | |
289
648e9245546d
seems the old intra/inter decission is slightly better with a threshold, than the new one
michaelni
parents:
288
diff
changeset
|
963 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
964 sum = s->dsp.pix_sum(pix, s->linesize); |
455 | 965 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
966 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8; |
936 | 967 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8; |
608 | 968 |
455 | 969 //printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout); |
903 | 970 pic->mb_var [s->mb_width * mb_y + mb_x] = varc; |
971 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard; | |
972 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8; | |
973 pic->mb_var_sum += varc; | |
974 pic->mc_mb_var_sum += vard; | |
455 | 975 //printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout); |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
976 |
0 | 977 #if 0 |
233
3f5b72726118
- More work on preliminary bit rate control, just to be able to get an
pulento
parents:
232
diff
changeset
|
978 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n", |
3f5b72726118
- More work on preliminary bit rate control, just to be able to get an
pulento
parents:
232
diff
changeset
|
979 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy); |
0 | 980 #endif |
294 | 981 if(s->flags&CODEC_FLAG_HQ){ |
608 | 982 if (vard <= 64 || vard < varc) |
983 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); | |
984 else | |
913 | 985 s->scene_change_score+= s->qscale; |
608 | 986 |
294 | 987 if (vard*2 + 200 > varc) |
988 mb_type|= MB_TYPE_INTRA; | |
989 if (varc*2 + 200 > vard){ | |
990 mb_type|= MB_TYPE_INTER; | |
936 | 991 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
992 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
304 | 993 }else{ |
936 | 994 mx <<=shift; |
995 my <<=shift; | |
0 | 996 } |
455 | 997 if((s->flags&CODEC_FLAG_4MV) |
936 | 998 && !s->me.skip && varc>50 && vard>10){ |
455 | 999 mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
1000 mb_type|=MB_TYPE_INTER4V; | |
1001 | |
1002 set_p_mv_tables(s, mx, my, 0); | |
1003 }else | |
1004 set_p_mv_tables(s, mx, my, 1); | |
294 | 1005 }else{ |
1006 if (vard <= 64 || vard < varc) { | |
936 | 1007 // if (sadP <= 32 || sadP < sadI + 500) { |
608 | 1008 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); |
294 | 1009 mb_type|= MB_TYPE_INTER; |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1010 if (s->me_method != ME_ZERO) { |
936 | 1011 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1012 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
455 | 1013 if((s->flags&CODEC_FLAG_4MV) |
936 | 1014 && !s->me.skip && varc>50 && vard>10){ |
455 | 1015 int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
1016 if(dmin4 + 128 <dmin) | |
1017 mb_type= MB_TYPE_INTER4V; | |
1018 } | |
1019 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V); | |
1020 | |
294 | 1021 } else { |
936 | 1022 mx <<=shift; |
1023 my <<=shift; | |
294 | 1024 } |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1025 #if 0 |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1026 if (vard < 10) { |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1027 skip++; |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1028 fprintf(stderr,"\nEarly skip: %d vard: %2d varc: %5d dmin: %d", |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1029 skip, vard, varc, dmin); |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1030 } |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1031 #endif |
294 | 1032 }else{ |
608 | 1033 s->scene_change_score+= 20; |
294 | 1034 mb_type|= MB_TYPE_INTRA; |
455 | 1035 mx = 0; |
1036 my = 0; | |
294 | 1037 } |
1038 } | |
284 | 1039 |
294 | 1040 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
0 | 1041 } |
1042 | |
951 | 1043 int ff_pre_estimate_p_frame_motion(MpegEncContext * s, |
1044 int mb_x, int mb_y) | |
1045 { | |
1046 int mx, my, range, dmin; | |
1047 int xmin, ymin, xmax, ymax; | |
1048 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; | |
1049 int pred_x=0, pred_y=0; | |
1050 int P[10][2]; | |
1051 const int shift= 1+s->quarter_sample; | |
1052 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; | |
1053 const int mv_stride= s->mb_width + 2; | |
1054 const int xy= mb_x + 1 + (mb_y + 1)*mv_stride; | |
1055 | |
1056 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
1057 | |
954 | 1058 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp); |
951 | 1059 |
1060 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code); | |
1061 rel_xmin= xmin - mb_x*16; | |
1062 rel_xmax= xmax - mb_x*16; | |
1063 rel_ymin= ymin - mb_y*16; | |
1064 rel_ymax= ymax - mb_y*16; | |
1065 s->me.skip=0; | |
1066 | |
1067 P_LEFT[0] = s->p_mv_table[xy + 1][0]; | |
1068 P_LEFT[1] = s->p_mv_table[xy + 1][1]; | |
1069 | |
1070 if(P_LEFT[0] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift); | |
1071 | |
1072 /* special case for first line */ | |
1073 if (mb_y == s->mb_height-1) { | |
1074 pred_x= P_LEFT[0]; | |
1075 pred_y= P_LEFT[1]; | |
952 | 1076 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]= |
1077 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME | |
951 | 1078 } else { |
1079 P_TOP[0] = s->p_mv_table[xy + mv_stride ][0]; | |
1080 P_TOP[1] = s->p_mv_table[xy + mv_stride ][1]; | |
1081 P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0]; | |
1082 P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1]; | |
1083 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift); | |
1084 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift); | |
1085 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift); | |
1086 | |
1087 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1088 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1089 | |
952 | 1090 pred_x = P_MEDIAN[0]; |
1091 pred_y = P_MEDIAN[1]; | |
951 | 1092 } |
954 | 1093 dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1094 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty); | |
952 | 1095 |
951 | 1096 s->p_mv_table[xy][0] = mx<<shift; |
1097 s->p_mv_table[xy][1] = my<<shift; | |
1098 | |
1099 return dmin; | |
1100 } | |
1101 | |
327 | 1102 int ff_estimate_motion_b(MpegEncContext * s, |
936 | 1103 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code) |
0 | 1104 { |
327 | 1105 int mx, my, range, dmin; |
324 | 1106 int xmin, ymin, xmax, ymax; |
1107 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; | |
1108 int pred_x=0, pred_y=0; | |
455 | 1109 int P[10][2]; |
324 | 1110 const int shift= 1+s->quarter_sample; |
1111 const int mot_stride = s->mb_width + 2; | |
1112 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1113 uint8_t * const ref_picture= picture->data[0]; |
1114 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV; | |
948 | 1115 int mv_scale; |
936 | 1116 |
1117 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
1118 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
1119 | |
324 | 1120 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code); |
455 | 1121 rel_xmin= xmin - mb_x*16; |
1122 rel_xmax= xmax - mb_x*16; | |
1123 rel_ymin= ymin - mb_y*16; | |
1124 rel_ymax= ymax - mb_y*16; | |
324 | 1125 |
1126 switch(s->me_method) { | |
1127 case ME_ZERO: | |
1128 default: | |
1129 no_motion_search(s, &mx, &my); | |
1130 dmin = 0; | |
455 | 1131 mx-= mb_x*16; |
1132 my-= mb_y*16; | |
324 | 1133 break; |
1134 case ME_FULL: | |
1135 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1136 mx-= mb_x*16; |
1137 my-= mb_y*16; | |
324 | 1138 break; |
1139 case ME_LOG: | |
1140 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1141 mx-= mb_x*16; |
1142 my-= mb_y*16; | |
324 | 1143 break; |
1144 case ME_PHODS: | |
1145 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1146 mx-= mb_x*16; |
1147 my-= mb_y*16; | |
324 | 1148 break; |
1149 case ME_X1: | |
1150 case ME_EPZS: | |
1151 { | |
455 | 1152 P_LEFT[0] = mv_table[mot_xy - 1][0]; |
1153 P_LEFT[1] = mv_table[mot_xy - 1][1]; | |
324 | 1154 |
455 | 1155 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
324 | 1156 |
1157 /* special case for first line */ | |
952 | 1158 if (mb_y) { |
455 | 1159 P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
1160 P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; | |
1161 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0]; | |
1162 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1]; | |
1163 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift); | |
1164 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
1165 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
324 | 1166 |
455 | 1167 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1168 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
324 | 1169 } |
455 | 1170 pred_x= P_LEFT[0]; |
1171 pred_y= P_LEFT[1]; | |
324 | 1172 } |
948 | 1173 |
1174 if(mv_table == s->b_forw_mv_table){ | |
1175 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift); | |
1176 }else{ | |
1177 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift); | |
1178 } | |
1179 | |
936 | 1180 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
948 | 1181 picture, s->p_mv_table, mv_scale, mv_penalty); |
324 | 1182 |
1183 break; | |
1184 } | |
1185 | |
936 | 1186 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1187 pred_x, pred_y, picture, 0, 0, mv_penalty); | |
455 | 1188 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my); |
324 | 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; | |
936 | 1192 |
327 | 1193 return dmin; |
324 | 1194 } |
1195 | |
327 | 1196 static inline int check_bidir_mv(MpegEncContext * s, |
1197 int mb_x, int mb_y, | |
1198 int motion_fx, int motion_fy, | |
1199 int motion_bx, int motion_by, | |
1200 int pred_fx, int pred_fy, | |
1201 int pred_bx, int pred_by) | |
1202 { | |
1203 //FIXME optimize? | |
936 | 1204 //FIXME move into template? |
1205 //FIXME better f_code prediction (max mv & distance) | |
1206 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame | |
1207 uint8_t *dest_y = s->me.scratchpad; | |
327 | 1208 uint8_t *ptr; |
1209 int dxy; | |
1210 int src_x, src_y; | |
1211 int fbmin; | |
1212 | |
936 | 1213 if(s->quarter_sample){ |
1214 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3); | |
1215 src_x = mb_x * 16 + (motion_fx >> 2); | |
1216 src_y = mb_y * 16 + (motion_fy >> 2); | |
1217 assert(src_x >=-16 && src_x<=s->width); | |
1218 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1219 |
936 | 1220 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1221 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize); | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1222 |
936 | 1223 dxy = ((motion_by & 3) << 2) | (motion_bx & 3); |
1224 src_x = mb_x * 16 + (motion_bx >> 2); | |
1225 src_y = mb_y * 16 + (motion_by >> 2); | |
1226 assert(src_x >=-16 && src_x<=s->width); | |
1227 assert(src_y >=-16 && src_y<=s->height); | |
1228 | |
1229 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1230 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize); | |
1231 }else{ | |
1232 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1); | |
1233 src_x = mb_x * 16 + (motion_fx >> 1); | |
1234 src_y = mb_y * 16 + (motion_fy >> 1); | |
1235 assert(src_x >=-16 && src_x<=s->width); | |
1236 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1237 |
936 | 1238 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1239 s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16); | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1240 |
936 | 1241 dxy = ((motion_by & 1) << 1) | (motion_bx & 1); |
1242 src_x = mb_x * 16 + (motion_bx >> 1); | |
1243 src_y = mb_y * 16 + (motion_by >> 1); | |
1244 assert(src_x >=-16 && src_x<=s->width); | |
1245 assert(src_y >=-16 && src_y<=s->height); | |
1246 | |
1247 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1248 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16); | |
1249 } | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1250 |
936 | 1251 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.sub_penalty_factor |
1252 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.sub_penalty_factor; | |
1253 + s->dsp.me_sub_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize); | |
1254 | |
327 | 1255 return fbmin; |
1256 } | |
1257 | |
1258 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/ | |
1259 static inline int bidir_refine(MpegEncContext * s, | |
1260 int mb_x, int mb_y) | |
1261 { | |
1262 const int mot_stride = s->mb_width + 2; | |
1263 const int xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
1264 int fbmin; | |
1265 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0]; | |
1266 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1]; | |
1267 int pred_bx= s->b_bidir_back_mv_table[xy-1][0]; | |
1268 int pred_by= s->b_bidir_back_mv_table[xy-1][1]; | |
1269 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0]; | |
1270 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1]; | |
1271 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0]; | |
1272 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1]; | |
1273 | |
1274 //FIXME do refinement and add flag | |
1275 | |
1276 fbmin= check_bidir_mv(s, mb_x, mb_y, | |
1277 motion_fx, motion_fy, | |
1278 motion_bx, motion_by, | |
1279 pred_fx, pred_fy, | |
1280 pred_bx, pred_by); | |
1281 | |
1282 return fbmin; | |
1283 } | |
1284 | |
1285 static inline int direct_search(MpegEncContext * s, | |
1286 int mb_x, int mb_y) | |
324 | 1287 { |
455 | 1288 int P[10][2]; |
327 | 1289 const int mot_stride = s->mb_width + 2; |
1290 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1291 const int shift= 1+s->quarter_sample; |
1292 int dmin, i; | |
327 | 1293 const int time_pp= s->pp_time; |
664 | 1294 const int time_pb= s->pb_time; |
936 | 1295 int mx, my, xmin, xmax, ymin, ymax; |
327 | 1296 int16_t (*mv_table)[2]= s->b_direct_mv_table; |
936 | 1297 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV; |
1298 | |
1299 ymin= xmin=(-32)>>shift; | |
1300 ymax= xmax= 31>>shift; | |
1301 | |
1302 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){ | |
1303 s->mv_type= MV_TYPE_8X8; | |
1304 }else{ | |
1305 s->mv_type= MV_TYPE_16X16; | |
327 | 1306 } |
936 | 1307 |
1308 for(i=0; i<4; i++){ | |
1309 int index= s->block_index[i]; | |
1310 int min, max; | |
1311 | |
1312 s->me.co_located_mv[i][0]= s->motion_val[index][0]; | |
1313 s->me.co_located_mv[i][1]= s->motion_val[index][1]; | |
1314 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3)); | |
1315 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3)); | |
1316 // s->me.direct_basis_mv[1][i][0]= s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3); | |
1317 // s->me.direct_basis_mv[1][i][1]= s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3); | |
1318 | |
1319 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1320 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1321 max+= (2*mb_x + (i& 1))*8 - 1; // +-1 is for the simpler rounding | |
1322 min+= (2*mb_x + (i& 1))*8 + 1; | |
960 | 1323 xmax= FFMIN(xmax, s->width - max); |
1324 xmin= FFMAX(xmin, - 16 - min); | |
936 | 1325 |
1326 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1327 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1328 max+= (2*mb_y + (i>>1))*8 - 1; // +-1 is for the simpler rounding | |
1329 min+= (2*mb_y + (i>>1))*8 + 1; | |
960 | 1330 ymax= FFMIN(ymax, s->height - max); |
1331 ymin= FFMAX(ymin, - 16 - min); | |
936 | 1332 |
1333 if(s->mv_type == MV_TYPE_16X16) break; | |
327 | 1334 } |
936 | 1335 |
1336 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16); | |
1337 | |
1338 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){ | |
1339 s->b_direct_mv_table[mot_xy][0]= 0; | |
1340 s->b_direct_mv_table[mot_xy][1]= 0; | |
1341 | |
1342 return 256*256*256*64; | |
1343 } | |
1344 | |
950 | 1345 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift); |
1346 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift); | |
1347 | |
1348 /* special case for first line */ | |
952 | 1349 if (mb_y) { |
950 | 1350 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift); |
1351 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift); | |
1352 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift); | |
1353 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift); | |
1354 | |
1355 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1356 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1357 } | |
1358 | |
936 | 1359 if(s->flags&CODEC_FLAG_QPEL){ |
1360 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
948 | 1361 &s->last_picture, mv_table, 1<<14, mv_penalty); |
936 | 1362 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, |
1363 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
1364 }else{ | |
1365 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
948 | 1366 &s->last_picture, mv_table, 1<<15, mv_penalty); |
936 | 1367 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, |
1368 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
327 | 1369 } |
1370 | |
1371 s->b_direct_mv_table[mot_xy][0]= mx; | |
1372 s->b_direct_mv_table[mot_xy][1]= my; | |
1373 return dmin; | |
324 | 1374 } |
1375 | |
1376 void ff_estimate_b_frame_motion(MpegEncContext * s, | |
1377 int mb_x, int mb_y) | |
1378 { | |
936 | 1379 const int penalty_factor= s->me.penalty_factor; |
327 | 1380 int fmin, bmin, dmin, fbmin; |
1381 int type=0; | |
1382 | |
1383 dmin= direct_search(s, mb_x, mb_y); | |
324 | 1384 |
936 | 1385 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code); |
1386 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) - penalty_factor; | |
324 | 1387 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
327 | 1388 |
1389 fbmin= bidir_refine(s, mb_x, mb_y); | |
1390 | |
612 | 1391 { |
327 | 1392 int score= dmin; |
1393 type=MB_TYPE_DIRECT; | |
1394 | |
1395 if(fmin<score){ | |
1396 score=fmin; | |
1397 type= MB_TYPE_FORWARD; | |
1398 } | |
1399 if(bmin<score){ | |
1400 score=bmin; | |
1401 type= MB_TYPE_BACKWARD; | |
1402 } | |
1403 if(fbmin<score){ | |
1404 score=fbmin; | |
1405 type= MB_TYPE_BIDIR; | |
1406 } | |
807
0e1d375c537f
fixing q>0.0 assert failure caused by overflow of variance for b frames
michaelni
parents:
765
diff
changeset
|
1407 score= ((unsigned)(score*score + 128*256))>>16; |
903 | 1408 s->current_picture.mc_mb_var_sum += score; |
1409 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD | |
327 | 1410 } |
612 | 1411 |
1412 if(s->flags&CODEC_FLAG_HQ){ | |
1413 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter | |
936 | 1414 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB |
612 | 1415 } |
1416 | |
327 | 1417 s->mb_type[mb_y*s->mb_width + mb_x]= type; |
324 | 1418 } |
0 | 1419 |
324 | 1420 /* find best f_code for ME which do unlimited searches */ |
1421 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type) | |
1422 { | |
1423 if(s->me_method>=ME_EPZS){ | |
455 | 1424 int score[8]; |
324 | 1425 int i, y; |
1426 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1427 int best_fcode=-1; |
1428 int best_score=-10000000; | |
324 | 1429 |
936 | 1430 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); |
324 | 1431 |
1432 for(y=0; y<s->mb_height; y++){ | |
1433 int x; | |
1434 int xy= (y+1)* (s->mb_width+2) + 1; | |
1435 i= y*s->mb_width; | |
1436 for(x=0; x<s->mb_width; x++){ | |
1437 if(s->mb_type[i] & type){ | |
847 | 1438 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV], |
1439 fcode_tab[mv_table[xy][1] + MAX_MV]); | |
455 | 1440 int j; |
1441 | |
1442 for(j=0; j<fcode && j<8; j++){ | |
903 | 1443 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i]) |
455 | 1444 score[j]-= 170; |
1445 } | |
324 | 1446 } |
1447 i++; | |
1448 xy++; | |
1449 } | |
1450 } | |
455 | 1451 |
1452 for(i=1; i<8; i++){ | |
1453 if(score[i] > best_score){ | |
1454 best_score= score[i]; | |
1455 best_fcode= i; | |
1456 } | |
1457 // printf("%d %d\n", i, score[i]); | |
1458 } | |
327 | 1459 |
324 | 1460 // printf("fcode: %d type: %d\n", i, s->pict_type); |
455 | 1461 return best_fcode; |
324 | 1462 /* for(i=0; i<=MAX_FCODE; i++){ |
1463 printf("%d ", mv_num[i]); | |
1464 } | |
1465 printf("\n");*/ | |
1466 }else{ | |
1467 return 1; | |
0 | 1468 } |
1469 } | |
1470 | |
324 | 1471 void ff_fix_long_p_mvs(MpegEncContext * s) |
1472 { | |
1473 const int f_code= s->f_code; | |
1474 int y; | |
1475 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1476 //int clip=0; |
1477 //int noclip=0; | |
324 | 1478 /* clip / convert to intra 16x16 type MVs */ |
1479 for(y=0; y<s->mb_height; y++){ | |
1480 int x; | |
1481 int xy= (y+1)* (s->mb_width+2)+1; | |
1482 int i= y*s->mb_width; | |
1483 for(x=0; x<s->mb_width; x++){ | |
1484 if(s->mb_type[i]&MB_TYPE_INTER){ | |
1485 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code | |
1486 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0 | |
1487 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code | |
1488 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){ | |
1489 s->mb_type[i] &= ~MB_TYPE_INTER; | |
1490 s->mb_type[i] |= MB_TYPE_INTRA; | |
1491 s->p_mv_table[xy][0] = 0; | |
1492 s->p_mv_table[xy][1] = 0; | |
455 | 1493 //clip++; |
324 | 1494 } |
455 | 1495 //else |
1496 // noclip++; | |
324 | 1497 } |
1498 xy++; | |
1499 i++; | |
1500 } | |
1501 } | |
455 | 1502 //printf("%d no:%d %d//\n", clip, noclip, f_code); |
324 | 1503 if(s->flags&CODEC_FLAG_4MV){ |
1504 const int wrap= 2+ s->mb_width*2; | |
1505 | |
1506 /* clip / convert to intra 8x8 type MVs */ | |
1507 for(y=0; y<s->mb_height; y++){ | |
1508 int xy= (y*2 + 1)*wrap + 1; | |
1509 int i= y*s->mb_width; | |
1510 int x; | |
1511 | |
1512 for(x=0; x<s->mb_width; x++){ | |
1513 if(s->mb_type[i]&MB_TYPE_INTER4V){ | |
1514 int block; | |
1515 for(block=0; block<4; block++){ | |
1516 int off= (block& 1) + (block>>1)*wrap; | |
1517 int mx= s->motion_val[ xy + off ][0]; | |
1518 int my= s->motion_val[ xy + off ][1]; | |
1519 | |
1520 if( fcode_tab[mx + MAX_MV] > f_code | |
1521 || fcode_tab[mx + MAX_MV] == 0 | |
1522 || fcode_tab[my + MAX_MV] > f_code | |
1523 || fcode_tab[my + MAX_MV] == 0 ){ | |
1524 s->mb_type[i] &= ~MB_TYPE_INTER4V; | |
1525 s->mb_type[i] |= MB_TYPE_INTRA; | |
1526 } | |
1527 } | |
1528 } | |
502 | 1529 xy+=2; |
1530 i++; | |
324 | 1531 } |
1532 } | |
1533 } | |
1534 } | |
1535 | |
1536 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type) | |
1537 { | |
1538 int y; | |
1539 UINT8 * fcode_tab= s->fcode_tab; | |
1540 | |
1541 /* clip / convert to intra 16x16 type MVs */ | |
1542 for(y=0; y<s->mb_height; y++){ | |
1543 int x; | |
1544 int xy= (y+1)* (s->mb_width+2)+1; | |
1545 int i= y*s->mb_width; | |
1546 for(x=0; x<s->mb_width; x++){ | |
691
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1547 if( fcode_tab[mv_table[xy][0] + MAX_MV] > f_code |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1548 || fcode_tab[mv_table[xy][0] + MAX_MV] == 0){ |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1549 if(mv_table[xy][0]>0) mv_table[xy][0]= (16<<f_code)-1; |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1550 else mv_table[xy][0]= -(16<<f_code); |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1551 } |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1552 if( fcode_tab[mv_table[xy][1] + MAX_MV] > f_code |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1553 || fcode_tab[mv_table[xy][1] + MAX_MV] == 0){ |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1554 if(mv_table[xy][1]>0) mv_table[xy][1]= (16<<f_code)-1; |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1555 else mv_table[xy][1]= -(16<<f_code); |
324 | 1556 } |
1557 xy++; | |
1558 i++; | |
1559 } | |
1560 } | |
1561 } |