Mercurial > libavcodec.hg
annotate motion_est.c @ 951:ad264a7d4f94 libavcodec
motion estimation pre pass
author | michaelni |
---|---|
date | Thu, 02 Jan 2003 01:29:35 +0000 |
parents | d4d714493faa |
children | f348d302a51e |
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; | |
289 case FF_CMP_ZERO: | |
290 for(i=0; i<7; i++){ | |
291 cmp[i]= zero_cmp; | |
292 } | |
293 break; | |
294 default: | |
295 fprintf(stderr,"internal error in cmp function selection\n"); | |
296 } | |
297 }; | |
298 | |
299 static inline int get_penalty_factor(MpegEncContext *s, int type){ | |
300 | |
301 switch(type){ | |
302 default: | |
303 case FF_CMP_SAD: | |
304 return s->qscale; | |
305 case FF_CMP_SSE: | |
306 // return s->qscale*8; | |
307 case FF_CMP_DCT: | |
308 case FF_CMP_SATD: | |
309 return s->qscale*8; | |
310 } | |
311 } | |
312 | |
313 void ff_init_me(MpegEncContext *s){ | |
314 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp); | |
315 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp); | |
316 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp); | |
317 | |
318 if(s->flags&CODEC_FLAG_QPEL){ | |
319 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
320 s->me.sub_motion_search= simple_chroma_qpel_motion_search; | |
321 else | |
322 s->me.sub_motion_search= simple_qpel_motion_search; | |
323 }else{ | |
324 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
325 s->me.sub_motion_search= simple_chroma_hpel_motion_search; | |
326 else if(s->avctx->me_sub_cmp == FF_CMP_SAD && s->avctx->me_cmp == FF_CMP_SAD) | |
327 s->me.sub_motion_search= sad_hpel_motion_search; | |
328 else | |
329 s->me.sub_motion_search= simple_hpel_motion_search; | |
330 } | |
331 | |
332 if(s->avctx->me_cmp&FF_CMP_CHROMA){ | |
333 s->me.motion_search[0]= simple_chroma_epzs_motion_search; | |
334 s->me.motion_search[1]= simple_chroma_epzs_motion_search4; | |
335 }else{ | |
336 s->me.motion_search[0]= simple_epzs_motion_search; | |
337 s->me.motion_search[1]= simple_epzs_motion_search4; | |
338 } | |
339 } | |
340 | |
284 | 341 static int pix_dev(UINT8 * pix, int line_size, int mean) |
342 { | |
343 int s, i, j; | |
344 | |
345 s = 0; | |
346 for (i = 0; i < 16; i++) { | |
347 for (j = 0; j < 16; j += 8) { | |
348 s += ABS(pix[0]-mean); | |
349 s += ABS(pix[1]-mean); | |
350 s += ABS(pix[2]-mean); | |
351 s += ABS(pix[3]-mean); | |
352 s += ABS(pix[4]-mean); | |
353 s += ABS(pix[5]-mean); | |
354 s += ABS(pix[6]-mean); | |
355 s += ABS(pix[7]-mean); | |
356 pix += 8; | |
357 } | |
358 pix += line_size - 16; | |
359 } | |
360 return s; | |
361 } | |
362 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
363 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
|
364 int *mx_ptr, int *my_ptr) |
0 | 365 { |
366 *mx_ptr = 16 * s->mb_x; | |
367 *my_ptr = 16 * s->mb_y; | |
368 } | |
369 | |
370 static int full_motion_search(MpegEncContext * s, | |
371 int *mx_ptr, int *my_ptr, int range, | |
324 | 372 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 373 { |
374 int x1, y1, x2, y2, xx, yy, x, y; | |
375 int mx, my, dmin, d; | |
376 UINT8 *pix; | |
377 | |
378 xx = 16 * s->mb_x; | |
379 yy = 16 * s->mb_y; | |
380 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */ | |
381 if (x1 < xmin) | |
382 x1 = xmin; | |
383 x2 = xx + range - 1; | |
384 if (x2 > xmax) | |
385 x2 = xmax; | |
386 y1 = yy - range + 1; | |
387 if (y1 < ymin) | |
388 y1 = ymin; | |
389 y2 = yy + range - 1; | |
390 if (y2 > ymax) | |
391 y2 = ymax; | |
903 | 392 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 393 dmin = 0x7fffffff; |
394 mx = 0; | |
395 my = 0; | |
396 for (y = y1; y <= y2; y++) { | |
397 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
|
398 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, |
294 | 399 s->linesize); |
0 | 400 if (d < dmin || |
401 (d == dmin && | |
402 (abs(x - xx) + abs(y - yy)) < | |
403 (abs(mx - xx) + abs(my - yy)))) { | |
404 dmin = d; | |
405 mx = x; | |
406 my = y; | |
407 } | |
408 } | |
409 } | |
410 | |
411 *mx_ptr = mx; | |
412 *my_ptr = my; | |
413 | |
414 #if 0 | |
415 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) || | |
416 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) { | |
417 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr); | |
418 } | |
419 #endif | |
420 return dmin; | |
421 } | |
422 | |
423 | |
424 static int log_motion_search(MpegEncContext * s, | |
425 int *mx_ptr, int *my_ptr, int range, | |
324 | 426 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 427 { |
428 int x1, y1, x2, y2, xx, yy, x, y; | |
429 int mx, my, dmin, d; | |
430 UINT8 *pix; | |
431 | |
432 xx = s->mb_x << 4; | |
433 yy = s->mb_y << 4; | |
434 | |
435 /* Left limit */ | |
436 x1 = xx - range; | |
437 if (x1 < xmin) | |
438 x1 = xmin; | |
439 | |
440 /* Right limit */ | |
441 x2 = xx + range; | |
442 if (x2 > xmax) | |
443 x2 = xmax; | |
444 | |
445 /* Upper limit */ | |
446 y1 = yy - range; | |
447 if (y1 < ymin) | |
448 y1 = ymin; | |
449 | |
450 /* Lower limit */ | |
451 y2 = yy + range; | |
452 if (y2 > ymax) | |
453 y2 = ymax; | |
454 | |
903 | 455 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 456 dmin = 0x7fffffff; |
457 mx = 0; | |
458 my = 0; | |
459 | |
460 do { | |
461 for (y = y1; y <= y2; y += range) { | |
462 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
|
463 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 464 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
465 dmin = d; | |
466 mx = x; | |
467 my = y; | |
468 } | |
469 } | |
470 } | |
471 | |
472 range = range >> 1; | |
473 | |
474 x1 = mx - range; | |
475 if (x1 < xmin) | |
476 x1 = xmin; | |
477 | |
478 x2 = mx + range; | |
479 if (x2 > xmax) | |
480 x2 = xmax; | |
481 | |
482 y1 = my - range; | |
483 if (y1 < ymin) | |
484 y1 = ymin; | |
485 | |
486 y2 = my + range; | |
487 if (y2 > ymax) | |
488 y2 = ymax; | |
489 | |
490 } while (range >= 1); | |
491 | |
492 #ifdef DEBUG | |
493 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my); | |
494 #endif | |
495 *mx_ptr = mx; | |
496 *my_ptr = my; | |
497 return dmin; | |
498 } | |
499 | |
500 static int phods_motion_search(MpegEncContext * s, | |
501 int *mx_ptr, int *my_ptr, int range, | |
324 | 502 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 503 { |
504 int x1, y1, x2, y2, xx, yy, x, y, lastx, d; | |
505 int mx, my, dminx, dminy; | |
506 UINT8 *pix; | |
507 | |
508 xx = s->mb_x << 4; | |
509 yy = s->mb_y << 4; | |
510 | |
511 /* Left limit */ | |
512 x1 = xx - range; | |
513 if (x1 < xmin) | |
514 x1 = xmin; | |
515 | |
516 /* Right limit */ | |
517 x2 = xx + range; | |
518 if (x2 > xmax) | |
519 x2 = xmax; | |
520 | |
521 /* Upper limit */ | |
522 y1 = yy - range; | |
523 if (y1 < ymin) | |
524 y1 = ymin; | |
525 | |
526 /* Lower limit */ | |
527 y2 = yy + range; | |
528 if (y2 > ymax) | |
529 y2 = ymax; | |
530 | |
903 | 531 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 532 mx = 0; |
533 my = 0; | |
534 | |
535 x = xx; | |
536 y = yy; | |
537 do { | |
538 dminx = 0x7fffffff; | |
539 dminy = 0x7fffffff; | |
540 | |
541 lastx = x; | |
542 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
|
543 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 544 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
545 dminx = d; | |
546 mx = x; | |
547 } | |
548 } | |
549 | |
550 x = lastx; | |
551 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
|
552 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 553 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
554 dminy = d; | |
555 my = y; | |
556 } | |
557 } | |
558 | |
559 range = range >> 1; | |
560 | |
561 x = mx; | |
562 y = my; | |
563 x1 = mx - range; | |
564 if (x1 < xmin) | |
565 x1 = xmin; | |
566 | |
567 x2 = mx + range; | |
568 if (x2 > xmax) | |
569 x2 = xmax; | |
570 | |
571 y1 = my - range; | |
572 if (y1 < ymin) | |
573 y1 = ymin; | |
574 | |
575 y2 = my + range; | |
576 if (y2 > ymax) | |
577 y2 = ymax; | |
578 | |
579 } while (range >= 1); | |
580 | |
581 #ifdef DEBUG | |
582 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my); | |
583 #endif | |
584 | |
585 /* half pixel search */ | |
586 *mx_ptr = mx; | |
587 *my_ptr = my; | |
588 return dminy; | |
589 } | |
590 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
591 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
592 #define Z_THRESHOLD 256 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
593 |
936 | 594 #define CHECK_SAD_HALF_MV(suffix, x, y) \ |
455 | 595 {\ |
596 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\ | |
936 | 597 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\ |
455 | 598 COPY3_IF_LT(dminh, d, dx, x, dy, y)\ |
599 } | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
600 |
936 | 601 static inline int sad_hpel_motion_search(MpegEncContext * s, |
0 | 602 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
|
603 int xmin, int ymin, int xmax, int ymax, |
936 | 604 int pred_x, int pred_y, Picture *picture, |
605 int n, int size, uint16_t * const mv_penalty) | |
0 | 606 { |
936 | 607 uint8_t *ref_picture= picture->data[0]; |
608 uint32_t *score_map= s->me.score_map; | |
609 const int penalty_factor= s->me.sub_penalty_factor; | |
455 | 610 int mx, my, xx, yy, dminh; |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
611 UINT8 *pix, *ptr; |
936 | 612 op_pixels_abs_func pix_abs_x2; |
613 op_pixels_abs_func pix_abs_y2; | |
614 op_pixels_abs_func pix_abs_xy2; | |
615 | |
616 if(size==0){ | |
617 pix_abs_x2 = s->dsp.pix_abs16x16_x2; | |
618 pix_abs_y2 = s->dsp.pix_abs16x16_y2; | |
619 pix_abs_xy2= s->dsp.pix_abs16x16_xy2; | |
620 }else{ | |
621 pix_abs_x2 = s->dsp.pix_abs8x8_x2; | |
622 pix_abs_y2 = s->dsp.pix_abs8x8_y2; | |
623 pix_abs_xy2= s->dsp.pix_abs8x8_xy2; | |
863 | 624 } |
455 | 625 |
936 | 626 if(s->me.skip){ |
455 | 627 // printf("S"); |
628 *mx_ptr = 0; | |
629 *my_ptr = 0; | |
630 return dmin; | |
631 } | |
632 // printf("N"); | |
633 | |
634 xx = 16 * s->mb_x + 8*(n&1); | |
635 yy = 16 * s->mb_y + 8*(n>>1); | |
903 | 636 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 637 |
294 | 638 mx = *mx_ptr; |
639 my = *my_ptr; | |
455 | 640 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx); |
641 | |
294 | 642 dminh = dmin; |
643 | |
644 if (mx > xmin && mx < xmax && | |
645 my > ymin && my < ymax) { | |
455 | 646 int dx=0, dy=0; |
647 int d, pen_x, pen_y; | |
648 const int index= (my<<ME_MAP_SHIFT) + mx; | |
649 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
650 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
651 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
652 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
653 mx<<=1; | |
654 my<<=1; | |
294 | 655 |
656 | |
657 pen_x= pred_x + mx; | |
658 pen_y= pred_y + my; | |
659 | |
660 ptr-= s->linesize; | |
455 | 661 if(t<=b){ |
936 | 662 CHECK_SAD_HALF_MV(y2 , 0, -1) |
455 | 663 if(l<=r){ |
936 | 664 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 665 if(t+r<=b+l){ |
936 | 666 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 667 ptr+= s->linesize; |
668 }else{ | |
669 ptr+= s->linesize; | |
936 | 670 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 671 } |
936 | 672 CHECK_SAD_HALF_MV(x2 , -1, 0) |
455 | 673 }else{ |
936 | 674 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 675 if(t+l<=b+r){ |
936 | 676 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 677 ptr+= s->linesize; |
678 }else{ | |
679 ptr+= s->linesize; | |
936 | 680 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 681 } |
936 | 682 CHECK_SAD_HALF_MV(x2 , +1, 0) |
455 | 683 } |
684 }else{ | |
685 if(l<=r){ | |
686 if(t+l<=b+r){ | |
936 | 687 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 688 ptr+= s->linesize; |
689 }else{ | |
690 ptr+= s->linesize; | |
936 | 691 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 692 } |
936 | 693 CHECK_SAD_HALF_MV(x2 , -1, 0) |
694 CHECK_SAD_HALF_MV(xy2, -1, +1) | |
455 | 695 }else{ |
696 if(t+r<=b+l){ | |
936 | 697 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 698 ptr+= s->linesize; |
699 }else{ | |
700 ptr+= s->linesize; | |
936 | 701 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 702 } |
936 | 703 CHECK_SAD_HALF_MV(x2 , +1, 0) |
704 CHECK_SAD_HALF_MV(xy2, +1, +1) | |
455 | 705 } |
936 | 706 CHECK_SAD_HALF_MV(y2 , 0, +1) |
455 | 707 } |
708 mx+=dx; | |
709 my+=dy; | |
294 | 710 |
711 }else{ | |
455 | 712 mx<<=1; |
713 my<<=1; | |
294 | 714 } |
715 | |
716 *mx_ptr = mx; | |
717 *my_ptr = my; | |
455 | 718 return dminh; |
294 | 719 } |
720 | |
455 | 721 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4) |
294 | 722 { |
324 | 723 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2); |
294 | 724 |
324 | 725 s->p_mv_table[xy][0] = mx; |
726 s->p_mv_table[xy][1] = my; | |
294 | 727 |
728 /* has allready been set to the 4 MV if 4MV is done */ | |
455 | 729 if(mv4){ |
294 | 730 int mot_xy= s->block_index[0]; |
731 | |
732 s->motion_val[mot_xy ][0]= mx; | |
733 s->motion_val[mot_xy ][1]= my; | |
734 s->motion_val[mot_xy+1][0]= mx; | |
735 s->motion_val[mot_xy+1][1]= my; | |
736 | |
737 mot_xy += s->block_wrap[0]; | |
738 s->motion_val[mot_xy ][0]= mx; | |
739 s->motion_val[mot_xy ][1]= my; | |
740 s->motion_val[mot_xy+1][0]= mx; | |
741 s->motion_val[mot_xy+1][1]= my; | |
742 } | |
743 } | |
744 | |
324 | 745 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code) |
746 { | |
747 *range = 8 * (1 << (f_code - 1)); | |
748 /* XXX: temporary kludge to avoid overflow for msmpeg4 */ | |
749 if (s->out_format == FMT_H263 && !s->h263_msmpeg4) | |
750 *range *= 2; | |
0 | 751 |
324 | 752 if (s->unrestricted_mv) { |
753 *xmin = -16; | |
754 *ymin = -16; | |
755 if (s->h263_plus) | |
756 *range *= 2; | |
951 | 757 if(s->avctx->codec->id!=CODEC_ID_MPEG4){ |
324 | 758 *xmax = s->mb_width*16; |
759 *ymax = s->mb_height*16; | |
760 }else { | |
761 *xmax = s->width; | |
762 *ymax = s->height; | |
763 } | |
764 } else { | |
765 *xmin = 0; | |
766 *ymin = 0; | |
767 *xmax = s->mb_width*16 - 16; | |
768 *ymax = s->mb_height*16 - 16; | |
769 } | |
770 } | |
771 | |
455 | 772 static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift) |
773 { | |
774 int block; | |
775 int P[10][2]; | |
903 | 776 uint8_t *ref_picture= s->last_picture.data[0]; |
455 | 777 int dmin_sum=0; |
936 | 778 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
455 | 779 |
780 for(block=0; block<4; block++){ | |
781 int mx4, my4; | |
782 int pred_x4, pred_y4; | |
783 int dmin4; | |
784 static const int off[4]= {2, 1, 1, -1}; | |
785 const int mot_stride = s->block_wrap[0]; | |
786 const int mot_xy = s->block_index[block]; | |
787 // const int block_x= (block&1); | |
788 // const int block_y= (block>>1); | |
789 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way | |
790 const int rel_xmin4= xmin; | |
791 const int rel_xmax4= xmax; | |
792 const int rel_ymin4= ymin; | |
793 const int rel_ymax4= ymax; | |
794 #else | |
795 const int rel_xmin4= xmin - block_x*8; | |
796 const int rel_xmax4= xmax - block_x*8 + 8; | |
797 const int rel_ymin4= ymin - block_y*8; | |
798 const int rel_ymax4= ymax - block_y*8 + 8; | |
799 #endif | |
800 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; | |
801 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
802 | |
803 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift); | |
804 | |
805 /* special case for first line */ | |
806 if ((s->mb_y == 0 || s->first_slice_line) && block<2) { | |
807 pred_x4= P_LEFT[0]; | |
808 pred_y4= P_LEFT[1]; | |
809 } else { | |
810 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; | |
811 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
812 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0]; | |
813 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1]; | |
814 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift); | |
815 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift); | |
816 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift); | |
817 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift); | |
818 | |
819 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
820 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
821 | |
822 if(s->out_format == FMT_H263){ | |
823 pred_x4 = P_MEDIAN[0]; | |
824 pred_y4 = P_MEDIAN[1]; | |
825 }else { /* mpeg1 at least */ | |
826 pred_x4= P_LEFT[0]; | |
827 pred_y4= P_LEFT[1]; | |
828 } | |
829 } | |
830 P_MV1[0]= mx; | |
831 P_MV1[1]= my; | |
832 | |
936 | 833 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
948 | 834 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty); |
455 | 835 |
936 | 836 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
837 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty); | |
455 | 838 |
839 s->motion_val[ s->block_index[block] ][0]= mx4; | |
840 s->motion_val[ s->block_index[block] ][1]= my4; | |
841 dmin_sum+= dmin4; | |
842 } | |
843 return dmin_sum; | |
844 } | |
845 | |
324 | 846 void ff_estimate_p_frame_motion(MpegEncContext * s, |
847 int mb_x, int mb_y) | |
0 | 848 { |
849 UINT8 *pix, *ppix; | |
850 int sum, varc, vard, mx, my, range, dmin, xx, yy; | |
851 int xmin, ymin, xmax, ymax; | |
280 | 852 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
|
853 int pred_x=0, pred_y=0; |
455 | 854 int P[10][2]; |
280 | 855 const int shift= 1+s->quarter_sample; |
294 | 856 int mb_type=0; |
903 | 857 uint8_t *ref_picture= s->last_picture.data[0]; |
858 Picture * const pic= &s->current_picture; | |
936 | 859 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
860 | |
861 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
862 | |
863 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
864 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
0 | 865 |
324 | 866 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code); |
455 | 867 rel_xmin= xmin - mb_x*16; |
868 rel_xmax= xmax - mb_x*16; | |
869 rel_ymin= ymin - mb_y*16; | |
870 rel_ymax= ymax - mb_y*16; | |
936 | 871 s->me.skip=0; |
324 | 872 |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
873 switch(s->me_method) { |
0 | 874 case ME_ZERO: |
875 default: | |
876 no_motion_search(s, &mx, &my); | |
455 | 877 mx-= mb_x*16; |
878 my-= mb_y*16; | |
0 | 879 dmin = 0; |
880 break; | |
881 case ME_FULL: | |
324 | 882 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); |
455 | 883 mx-= mb_x*16; |
884 my-= mb_y*16; | |
0 | 885 break; |
886 case ME_LOG: | |
324 | 887 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 888 mx-= mb_x*16; |
889 my-= mb_y*16; | |
0 | 890 break; |
891 case ME_PHODS: | |
324 | 892 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 893 mx-= mb_x*16; |
894 my-= mb_y*16; | |
0 | 895 break; |
288 | 896 case ME_X1: |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
897 case ME_EPZS: |
288 | 898 { |
294 | 899 const int mot_stride = s->block_wrap[0]; |
900 const int mot_xy = s->block_index[0]; | |
288 | 901 |
455 | 902 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; |
903 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
288 | 904 |
455 | 905 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
280 | 906 |
907 /* special case for first line */ | |
455 | 908 if ((mb_y == 0 || s->first_slice_line)) { |
909 pred_x= P_LEFT[0]; | |
910 pred_y= P_LEFT[1]; | |
280 | 911 } else { |
455 | 912 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; |
913 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
914 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0]; | |
915 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1]; | |
916 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift); | |
917 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
918 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
280 | 919 |
455 | 920 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
921 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
922 | |
923 if(s->out_format == FMT_H263){ | |
924 pred_x = P_MEDIAN[0]; | |
925 pred_y = P_MEDIAN[1]; | |
926 }else { /* mpeg1 at least */ | |
927 pred_x= P_LEFT[0]; | |
928 pred_y= P_LEFT[1]; | |
929 } | |
288 | 930 } |
280 | 931 } |
936 | 932 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
948 | 933 &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
|
934 |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
935 break; |
0 | 936 } |
937 | |
938 /* intra / predictive decision */ | |
939 xx = mb_x * 16; | |
940 yy = mb_y * 16; | |
941 | |
903 | 942 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 943 /* At this point (mx,my) are full-pell and the relative displacement */ |
944 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
|
945 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
946 sum = s->dsp.pix_sum(pix, s->linesize); |
455 | 947 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
948 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8; |
936 | 949 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8; |
608 | 950 |
455 | 951 //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 | 952 pic->mb_var [s->mb_width * mb_y + mb_x] = varc; |
953 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard; | |
954 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8; | |
955 pic->mb_var_sum += varc; | |
956 pic->mc_mb_var_sum += vard; | |
455 | 957 //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
|
958 |
0 | 959 #if 0 |
233
3f5b72726118
- More work on preliminary bit rate control, just to be able to get an
pulento
parents:
232
diff
changeset
|
960 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
|
961 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy); |
0 | 962 #endif |
294 | 963 if(s->flags&CODEC_FLAG_HQ){ |
608 | 964 if (vard <= 64 || vard < varc) |
965 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); | |
966 else | |
913 | 967 s->scene_change_score+= s->qscale; |
608 | 968 |
294 | 969 if (vard*2 + 200 > varc) |
970 mb_type|= MB_TYPE_INTRA; | |
971 if (varc*2 + 200 > vard){ | |
972 mb_type|= MB_TYPE_INTER; | |
936 | 973 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
974 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
304 | 975 }else{ |
936 | 976 mx <<=shift; |
977 my <<=shift; | |
0 | 978 } |
455 | 979 if((s->flags&CODEC_FLAG_4MV) |
936 | 980 && !s->me.skip && varc>50 && vard>10){ |
455 | 981 mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
982 mb_type|=MB_TYPE_INTER4V; | |
983 | |
984 set_p_mv_tables(s, mx, my, 0); | |
985 }else | |
986 set_p_mv_tables(s, mx, my, 1); | |
294 | 987 }else{ |
988 if (vard <= 64 || vard < varc) { | |
936 | 989 // if (sadP <= 32 || sadP < sadI + 500) { |
608 | 990 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); |
294 | 991 mb_type|= MB_TYPE_INTER; |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
992 if (s->me_method != ME_ZERO) { |
936 | 993 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
994 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
455 | 995 if((s->flags&CODEC_FLAG_4MV) |
936 | 996 && !s->me.skip && varc>50 && vard>10){ |
455 | 997 int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
998 if(dmin4 + 128 <dmin) | |
999 mb_type= MB_TYPE_INTER4V; | |
1000 } | |
1001 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V); | |
1002 | |
294 | 1003 } else { |
936 | 1004 mx <<=shift; |
1005 my <<=shift; | |
294 | 1006 } |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1007 #if 0 |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1008 if (vard < 10) { |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1009 skip++; |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1010 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
|
1011 skip, vard, varc, dmin); |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1012 } |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1013 #endif |
294 | 1014 }else{ |
608 | 1015 s->scene_change_score+= 20; |
294 | 1016 mb_type|= MB_TYPE_INTRA; |
455 | 1017 mx = 0; |
1018 my = 0; | |
294 | 1019 } |
1020 } | |
284 | 1021 |
294 | 1022 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
0 | 1023 } |
1024 | |
951 | 1025 int ff_pre_estimate_p_frame_motion(MpegEncContext * s, |
1026 int mb_x, int mb_y) | |
1027 { | |
1028 int mx, my, range, dmin; | |
1029 int xmin, ymin, xmax, ymax; | |
1030 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; | |
1031 int pred_x=0, pred_y=0; | |
1032 int P[10][2]; | |
1033 const int shift= 1+s->quarter_sample; | |
1034 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; | |
1035 const int mv_stride= s->mb_width + 2; | |
1036 const int xy= mb_x + 1 + (mb_y + 1)*mv_stride; | |
1037 | |
1038 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
1039 | |
1040 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
1041 | |
1042 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code); | |
1043 rel_xmin= xmin - mb_x*16; | |
1044 rel_xmax= xmax - mb_x*16; | |
1045 rel_ymin= ymin - mb_y*16; | |
1046 rel_ymax= ymax - mb_y*16; | |
1047 s->me.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] < (rel_xmin<<shift)) P_LEFT[0] = (rel_xmin<<shift); | |
1053 | |
1054 /* special case for first line */ | |
1055 if (mb_y == s->mb_height-1) { | |
1056 pred_x= P_LEFT[0]; | |
1057 pred_y= P_LEFT[1]; | |
1058 } else { | |
1059 P_TOP[0] = s->p_mv_table[xy + mv_stride ][0]; | |
1060 P_TOP[1] = s->p_mv_table[xy + mv_stride ][1]; | |
1061 P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0]; | |
1062 P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1]; | |
1063 if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift); | |
1064 if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift); | |
1065 if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift); | |
1066 | |
1067 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1068 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1069 | |
1070 if(s->out_format == FMT_H263){ | |
1071 pred_x = P_MEDIAN[0]; | |
1072 pred_y = P_MEDIAN[1]; | |
1073 }else { /* mpeg1 at least */ | |
1074 pred_x= P_LEFT[0]; | |
1075 pred_y= P_LEFT[1]; | |
1076 } | |
1077 } | |
1078 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, | |
1079 &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty); | |
1080 | |
1081 s->p_mv_table[xy][0] = mx<<shift; | |
1082 s->p_mv_table[xy][1] = my<<shift; | |
1083 | |
1084 return dmin; | |
1085 } | |
1086 | |
327 | 1087 int ff_estimate_motion_b(MpegEncContext * s, |
936 | 1088 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code) |
0 | 1089 { |
327 | 1090 int mx, my, range, dmin; |
324 | 1091 int xmin, ymin, xmax, ymax; |
1092 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; | |
1093 int pred_x=0, pred_y=0; | |
455 | 1094 int P[10][2]; |
324 | 1095 const int shift= 1+s->quarter_sample; |
1096 const int mot_stride = s->mb_width + 2; | |
1097 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1098 uint8_t * const ref_picture= picture->data[0]; |
1099 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV; | |
948 | 1100 int mv_scale; |
936 | 1101 |
1102 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
1103 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
1104 | |
324 | 1105 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code); |
455 | 1106 rel_xmin= xmin - mb_x*16; |
1107 rel_xmax= xmax - mb_x*16; | |
1108 rel_ymin= ymin - mb_y*16; | |
1109 rel_ymax= ymax - mb_y*16; | |
324 | 1110 |
1111 switch(s->me_method) { | |
1112 case ME_ZERO: | |
1113 default: | |
1114 no_motion_search(s, &mx, &my); | |
1115 dmin = 0; | |
455 | 1116 mx-= mb_x*16; |
1117 my-= mb_y*16; | |
324 | 1118 break; |
1119 case ME_FULL: | |
1120 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1121 mx-= mb_x*16; |
1122 my-= mb_y*16; | |
324 | 1123 break; |
1124 case ME_LOG: | |
1125 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1126 mx-= mb_x*16; |
1127 my-= mb_y*16; | |
324 | 1128 break; |
1129 case ME_PHODS: | |
1130 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1131 mx-= mb_x*16; |
1132 my-= mb_y*16; | |
324 | 1133 break; |
1134 case ME_X1: | |
1135 case ME_EPZS: | |
1136 { | |
455 | 1137 P_LEFT[0] = mv_table[mot_xy - 1][0]; |
1138 P_LEFT[1] = mv_table[mot_xy - 1][1]; | |
324 | 1139 |
455 | 1140 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
324 | 1141 |
1142 /* special case for first line */ | |
455 | 1143 if ((mb_y == 0 || s->first_slice_line)) { |
324 | 1144 } else { |
455 | 1145 P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
1146 P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; | |
1147 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0]; | |
1148 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1]; | |
1149 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift); | |
1150 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
1151 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
324 | 1152 |
455 | 1153 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1154 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
324 | 1155 } |
455 | 1156 pred_x= P_LEFT[0]; |
1157 pred_y= P_LEFT[1]; | |
324 | 1158 } |
948 | 1159 |
1160 if(mv_table == s->b_forw_mv_table){ | |
1161 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift); | |
1162 }else{ | |
1163 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift); | |
1164 } | |
1165 | |
936 | 1166 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
948 | 1167 picture, s->p_mv_table, mv_scale, mv_penalty); |
324 | 1168 |
1169 break; | |
1170 } | |
1171 | |
936 | 1172 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1173 pred_x, pred_y, picture, 0, 0, mv_penalty); | |
455 | 1174 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my); |
324 | 1175 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
1176 mv_table[mot_xy][0]= mx; | |
1177 mv_table[mot_xy][1]= my; | |
936 | 1178 |
327 | 1179 return dmin; |
324 | 1180 } |
1181 | |
327 | 1182 static inline int check_bidir_mv(MpegEncContext * s, |
1183 int mb_x, int mb_y, | |
1184 int motion_fx, int motion_fy, | |
1185 int motion_bx, int motion_by, | |
1186 int pred_fx, int pred_fy, | |
1187 int pred_bx, int pred_by) | |
1188 { | |
1189 //FIXME optimize? | |
936 | 1190 //FIXME move into template? |
1191 //FIXME better f_code prediction (max mv & distance) | |
1192 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame | |
1193 uint8_t *dest_y = s->me.scratchpad; | |
327 | 1194 uint8_t *ptr; |
1195 int dxy; | |
1196 int src_x, src_y; | |
1197 int fbmin; | |
1198 | |
936 | 1199 if(s->quarter_sample){ |
1200 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3); | |
1201 src_x = mb_x * 16 + (motion_fx >> 2); | |
1202 src_y = mb_y * 16 + (motion_fy >> 2); | |
1203 assert(src_x >=-16 && src_x<=s->width); | |
1204 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1205 |
936 | 1206 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1207 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
|
1208 |
936 | 1209 dxy = ((motion_by & 3) << 2) | (motion_bx & 3); |
1210 src_x = mb_x * 16 + (motion_bx >> 2); | |
1211 src_y = mb_y * 16 + (motion_by >> 2); | |
1212 assert(src_x >=-16 && src_x<=s->width); | |
1213 assert(src_y >=-16 && src_y<=s->height); | |
1214 | |
1215 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1216 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize); | |
1217 }else{ | |
1218 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1); | |
1219 src_x = mb_x * 16 + (motion_fx >> 1); | |
1220 src_y = mb_y * 16 + (motion_fy >> 1); | |
1221 assert(src_x >=-16 && src_x<=s->width); | |
1222 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1223 |
936 | 1224 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1225 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
|
1226 |
936 | 1227 dxy = ((motion_by & 1) << 1) | (motion_bx & 1); |
1228 src_x = mb_x * 16 + (motion_bx >> 1); | |
1229 src_y = mb_y * 16 + (motion_by >> 1); | |
1230 assert(src_x >=-16 && src_x<=s->width); | |
1231 assert(src_y >=-16 && src_y<=s->height); | |
1232 | |
1233 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1234 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16); | |
1235 } | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1236 |
936 | 1237 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.sub_penalty_factor |
1238 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.sub_penalty_factor; | |
1239 + s->dsp.me_sub_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize); | |
1240 | |
327 | 1241 return fbmin; |
1242 } | |
1243 | |
1244 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/ | |
1245 static inline int bidir_refine(MpegEncContext * s, | |
1246 int mb_x, int mb_y) | |
1247 { | |
1248 const int mot_stride = s->mb_width + 2; | |
1249 const int xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
1250 int fbmin; | |
1251 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0]; | |
1252 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1]; | |
1253 int pred_bx= s->b_bidir_back_mv_table[xy-1][0]; | |
1254 int pred_by= s->b_bidir_back_mv_table[xy-1][1]; | |
1255 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0]; | |
1256 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1]; | |
1257 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0]; | |
1258 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1]; | |
1259 | |
1260 //FIXME do refinement and add flag | |
1261 | |
1262 fbmin= check_bidir_mv(s, mb_x, mb_y, | |
1263 motion_fx, motion_fy, | |
1264 motion_bx, motion_by, | |
1265 pred_fx, pred_fy, | |
1266 pred_bx, pred_by); | |
1267 | |
1268 return fbmin; | |
1269 } | |
1270 | |
1271 static inline int direct_search(MpegEncContext * s, | |
1272 int mb_x, int mb_y) | |
324 | 1273 { |
455 | 1274 int P[10][2]; |
327 | 1275 const int mot_stride = s->mb_width + 2; |
1276 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1277 const int shift= 1+s->quarter_sample; |
1278 int dmin, i; | |
327 | 1279 const int time_pp= s->pp_time; |
664 | 1280 const int time_pb= s->pb_time; |
936 | 1281 int mx, my, xmin, xmax, ymin, ymax; |
327 | 1282 int16_t (*mv_table)[2]= s->b_direct_mv_table; |
936 | 1283 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV; |
1284 | |
1285 ymin= xmin=(-32)>>shift; | |
1286 ymax= xmax= 31>>shift; | |
1287 | |
1288 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){ | |
1289 s->mv_type= MV_TYPE_8X8; | |
1290 }else{ | |
1291 s->mv_type= MV_TYPE_16X16; | |
327 | 1292 } |
936 | 1293 |
1294 for(i=0; i<4; i++){ | |
1295 int index= s->block_index[i]; | |
1296 int min, max; | |
1297 | |
1298 s->me.co_located_mv[i][0]= s->motion_val[index][0]; | |
1299 s->me.co_located_mv[i][1]= s->motion_val[index][1]; | |
1300 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3)); | |
1301 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3)); | |
1302 // 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); | |
1303 // 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); | |
1304 | |
1305 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1306 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1307 max+= (2*mb_x + (i& 1))*8 - 1; // +-1 is for the simpler rounding | |
1308 min+= (2*mb_x + (i& 1))*8 + 1; | |
1309 if(max >= s->width) xmax= s->width - max - 1; | |
1310 if(min < -16 ) xmin= - 32 - min; | |
1311 | |
1312 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1313 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1314 max+= (2*mb_y + (i>>1))*8 - 1; // +-1 is for the simpler rounding | |
1315 min+= (2*mb_y + (i>>1))*8 + 1; | |
1316 if(max >= s->height) ymax= s->height - max - 1; | |
1317 if(min < -16 ) ymin= - 32 - min; | |
1318 | |
1319 if(s->mv_type == MV_TYPE_16X16) break; | |
327 | 1320 } |
936 | 1321 |
1322 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16); | |
1323 | |
1324 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){ | |
1325 s->b_direct_mv_table[mot_xy][0]= 0; | |
1326 s->b_direct_mv_table[mot_xy][1]= 0; | |
1327 | |
1328 return 256*256*256*64; | |
1329 } | |
1330 | |
950 | 1331 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift); |
1332 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift); | |
1333 | |
1334 /* special case for first line */ | |
1335 if ((mb_y == 0 || s->first_slice_line)) { | |
1336 } else { | |
1337 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift); | |
1338 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift); | |
1339 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift); | |
1340 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift); | |
1341 | |
1342 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1343 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1344 } | |
1345 | |
936 | 1346 if(s->flags&CODEC_FLAG_QPEL){ |
1347 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
948 | 1348 &s->last_picture, mv_table, 1<<14, mv_penalty); |
936 | 1349 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, |
1350 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
1351 }else{ | |
1352 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
948 | 1353 &s->last_picture, mv_table, 1<<15, mv_penalty); |
936 | 1354 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, |
1355 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
327 | 1356 } |
1357 | |
1358 s->b_direct_mv_table[mot_xy][0]= mx; | |
1359 s->b_direct_mv_table[mot_xy][1]= my; | |
1360 return dmin; | |
324 | 1361 } |
1362 | |
1363 void ff_estimate_b_frame_motion(MpegEncContext * s, | |
1364 int mb_x, int mb_y) | |
1365 { | |
936 | 1366 const int penalty_factor= s->me.penalty_factor; |
327 | 1367 int fmin, bmin, dmin, fbmin; |
1368 int type=0; | |
1369 | |
1370 dmin= direct_search(s, mb_x, mb_y); | |
324 | 1371 |
936 | 1372 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code); |
1373 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) - penalty_factor; | |
324 | 1374 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
327 | 1375 |
1376 fbmin= bidir_refine(s, mb_x, mb_y); | |
1377 | |
612 | 1378 { |
327 | 1379 int score= dmin; |
1380 type=MB_TYPE_DIRECT; | |
1381 | |
1382 if(fmin<score){ | |
1383 score=fmin; | |
1384 type= MB_TYPE_FORWARD; | |
1385 } | |
1386 if(bmin<score){ | |
1387 score=bmin; | |
1388 type= MB_TYPE_BACKWARD; | |
1389 } | |
1390 if(fbmin<score){ | |
1391 score=fbmin; | |
1392 type= MB_TYPE_BIDIR; | |
1393 } | |
807
0e1d375c537f
fixing q>0.0 assert failure caused by overflow of variance for b frames
michaelni
parents:
765
diff
changeset
|
1394 score= ((unsigned)(score*score + 128*256))>>16; |
903 | 1395 s->current_picture.mc_mb_var_sum += score; |
1396 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD | |
327 | 1397 } |
612 | 1398 |
1399 if(s->flags&CODEC_FLAG_HQ){ | |
1400 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter | |
936 | 1401 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB |
612 | 1402 } |
1403 | |
327 | 1404 s->mb_type[mb_y*s->mb_width + mb_x]= type; |
324 | 1405 } |
0 | 1406 |
324 | 1407 /* find best f_code for ME which do unlimited searches */ |
1408 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type) | |
1409 { | |
1410 if(s->me_method>=ME_EPZS){ | |
455 | 1411 int score[8]; |
324 | 1412 int i, y; |
1413 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1414 int best_fcode=-1; |
1415 int best_score=-10000000; | |
324 | 1416 |
936 | 1417 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); |
324 | 1418 |
1419 for(y=0; y<s->mb_height; y++){ | |
1420 int x; | |
1421 int xy= (y+1)* (s->mb_width+2) + 1; | |
1422 i= y*s->mb_width; | |
1423 for(x=0; x<s->mb_width; x++){ | |
1424 if(s->mb_type[i] & type){ | |
847 | 1425 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV], |
1426 fcode_tab[mv_table[xy][1] + MAX_MV]); | |
455 | 1427 int j; |
1428 | |
1429 for(j=0; j<fcode && j<8; j++){ | |
903 | 1430 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i]) |
455 | 1431 score[j]-= 170; |
1432 } | |
324 | 1433 } |
1434 i++; | |
1435 xy++; | |
1436 } | |
1437 } | |
455 | 1438 |
1439 for(i=1; i<8; i++){ | |
1440 if(score[i] > best_score){ | |
1441 best_score= score[i]; | |
1442 best_fcode= i; | |
1443 } | |
1444 // printf("%d %d\n", i, score[i]); | |
1445 } | |
327 | 1446 |
324 | 1447 // printf("fcode: %d type: %d\n", i, s->pict_type); |
455 | 1448 return best_fcode; |
324 | 1449 /* for(i=0; i<=MAX_FCODE; i++){ |
1450 printf("%d ", mv_num[i]); | |
1451 } | |
1452 printf("\n");*/ | |
1453 }else{ | |
1454 return 1; | |
0 | 1455 } |
1456 } | |
1457 | |
324 | 1458 void ff_fix_long_p_mvs(MpegEncContext * s) |
1459 { | |
1460 const int f_code= s->f_code; | |
1461 int y; | |
1462 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1463 //int clip=0; |
1464 //int noclip=0; | |
324 | 1465 /* clip / convert to intra 16x16 type MVs */ |
1466 for(y=0; y<s->mb_height; y++){ | |
1467 int x; | |
1468 int xy= (y+1)* (s->mb_width+2)+1; | |
1469 int i= y*s->mb_width; | |
1470 for(x=0; x<s->mb_width; x++){ | |
1471 if(s->mb_type[i]&MB_TYPE_INTER){ | |
1472 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code | |
1473 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0 | |
1474 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code | |
1475 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){ | |
1476 s->mb_type[i] &= ~MB_TYPE_INTER; | |
1477 s->mb_type[i] |= MB_TYPE_INTRA; | |
1478 s->p_mv_table[xy][0] = 0; | |
1479 s->p_mv_table[xy][1] = 0; | |
455 | 1480 //clip++; |
324 | 1481 } |
455 | 1482 //else |
1483 // noclip++; | |
324 | 1484 } |
1485 xy++; | |
1486 i++; | |
1487 } | |
1488 } | |
455 | 1489 //printf("%d no:%d %d//\n", clip, noclip, f_code); |
324 | 1490 if(s->flags&CODEC_FLAG_4MV){ |
1491 const int wrap= 2+ s->mb_width*2; | |
1492 | |
1493 /* clip / convert to intra 8x8 type MVs */ | |
1494 for(y=0; y<s->mb_height; y++){ | |
1495 int xy= (y*2 + 1)*wrap + 1; | |
1496 int i= y*s->mb_width; | |
1497 int x; | |
1498 | |
1499 for(x=0; x<s->mb_width; x++){ | |
1500 if(s->mb_type[i]&MB_TYPE_INTER4V){ | |
1501 int block; | |
1502 for(block=0; block<4; block++){ | |
1503 int off= (block& 1) + (block>>1)*wrap; | |
1504 int mx= s->motion_val[ xy + off ][0]; | |
1505 int my= s->motion_val[ xy + off ][1]; | |
1506 | |
1507 if( fcode_tab[mx + MAX_MV] > f_code | |
1508 || fcode_tab[mx + MAX_MV] == 0 | |
1509 || fcode_tab[my + MAX_MV] > f_code | |
1510 || fcode_tab[my + MAX_MV] == 0 ){ | |
1511 s->mb_type[i] &= ~MB_TYPE_INTER4V; | |
1512 s->mb_type[i] |= MB_TYPE_INTRA; | |
1513 } | |
1514 } | |
1515 } | |
502 | 1516 xy+=2; |
1517 i++; | |
324 | 1518 } |
1519 } | |
1520 } | |
1521 } | |
1522 | |
1523 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type) | |
1524 { | |
1525 int y; | |
1526 UINT8 * fcode_tab= s->fcode_tab; | |
1527 | |
1528 /* clip / convert to intra 16x16 type MVs */ | |
1529 for(y=0; y<s->mb_height; y++){ | |
1530 int x; | |
1531 int xy= (y+1)* (s->mb_width+2)+1; | |
1532 int i= y*s->mb_width; | |
1533 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
|
1534 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
|
1535 || 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
|
1536 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
|
1537 else mv_table[xy][0]= -(16<<f_code); |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1538 } |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1539 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
|
1540 || 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
|
1541 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
|
1542 else mv_table[xy][1]= -(16<<f_code); |
324 | 1543 } |
1544 xy++; | |
1545 i++; | |
1546 } | |
1547 } | |
1548 } |