Mercurial > libavcodec.hg
annotate motion_est.c @ 936:caa77cd960c0 libavcodec
qpel encoding
4mv+b frames encoding finally fixed
chroma ME
5 comparission functions for ME
b frame encoding speedup
wmv2 codec (unfinished)
user specified diamond size for EPZS
author | michaelni |
---|---|
date | Fri, 27 Dec 2002 23:51:46 +0000 |
parents | 51f3b644ae30 |
children | 371bc36a9c5c |
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_LAST P[0] |
35 #define P_LEFT P[1] | |
36 #define P_TOP P[2] | |
37 #define P_TOPRIGHT P[3] | |
38 #define P_MEDIAN P[4] | |
39 #define P_LAST_LEFT P[5] | |
40 #define P_LAST_RIGHT P[6] | |
41 #define P_LAST_TOP P[7] | |
42 #define P_LAST_BOTTOM P[8] | |
43 #define P_MV1 P[9] | |
44 | |
936 | 45 static inline int sad_hpel_motion_search(MpegEncContext * s, |
46 int *mx_ptr, int *my_ptr, int dmin, | |
47 int xmin, int ymin, int xmax, int ymax, | |
48 int pred_x, int pred_y, Picture *picture, | |
49 int n, int size, uint16_t * const mv_penalty); | |
0 | 50 |
936 | 51 static inline int update_map_generation(MpegEncContext * s) |
52 { | |
53 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2); | |
54 if(s->me.map_generation==0){ | |
55 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2); | |
56 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE); | |
57 } | |
58 return s->me.map_generation; | |
59 } | |
60 | |
61 | |
62 | |
63 /* SIMPLE */ | |
64 #define RENAME(a) simple_ ## a | |
65 | |
66 #define CMP(d, x, y, size)\ | |
67 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride); | |
68 | |
69 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
70 {\ | |
71 const int dxy= (dx) + 2*(dy);\ | |
72 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\ | |
73 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
74 } | |
75 | |
76 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
77 {\ | |
78 const int dxy= (dx) + 4*(dy);\ | |
79 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\ | |
80 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
81 } | |
82 | |
83 #include "motion_est_template.c" | |
84 #undef RENAME | |
85 #undef CMP | |
86 #undef CMP_HPEL | |
87 #undef CMP_QPEL | |
88 #undef INIT | |
89 | |
90 /* SIMPLE CHROMA */ | |
91 #define RENAME(a) simple_chroma_ ## a | |
92 | |
93 #define CMP(d, x, y, size)\ | |
94 d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\ | |
95 if(chroma_cmp){\ | |
96 int dxy= ((x)&1) + 2*((y)&1);\ | |
97 int c= ((x)>>1) + ((y)>>1)*uvstride;\ | |
98 \ | |
99 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
100 d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\ | |
101 chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
102 d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\ | |
103 } | |
104 | |
105 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
106 {\ | |
107 const int dxy= (dx) + 2*(dy);\ | |
108 hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\ | |
109 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
110 if(chroma_cmp_sub){\ | |
111 int cxy= (dxy) | ((x)&1) | (2*((y)&1));\ | |
112 int c= ((x)>>1) + ((y)>>1)*uvstride;\ | |
113 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
114 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\ | |
115 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
116 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\ | |
117 }\ | |
118 } | |
119 | |
120 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
121 {\ | |
122 const int dxy= (dx) + 4*(dy);\ | |
123 qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\ | |
124 d = cmp_sub(s, s->me.scratchpad, src_y, stride);\ | |
125 if(chroma_cmp_sub){\ | |
126 int cxy, c;\ | |
127 int cx= (4*(x) + (dx))/2;\ | |
128 int cy= (4*(y) + (dy))/2;\ | |
129 cx= (cx>>1)|(cx&1);\ | |
130 cy= (cy>>1)|(cy&1);\ | |
131 cxy= (cx&1) + 2*(cy&1);\ | |
132 c= ((cx)>>1) + ((cy)>>1)*uvstride;\ | |
133 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\ | |
134 d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\ | |
135 chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\ | |
136 d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\ | |
137 }\ | |
138 } | |
139 | |
140 #include "motion_est_template.c" | |
141 #undef RENAME | |
142 #undef CMP | |
143 #undef CMP_HPEL | |
144 #undef CMP_QPEL | |
145 #undef INIT | |
146 | |
147 /* SIMPLE DIRECT HPEL */ | |
148 #define RENAME(a) simple_direct_hpel_ ## a | |
149 //FIXME precalc divisions stuff | |
150 | |
151 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\ | |
152 if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\ | |
153 const int hx= 2*(x) + (dx);\ | |
154 const int hy= 2*(y) + (dy);\ | |
155 if(s->mv_type==MV_TYPE_8X8){\ | |
156 int i;\ | |
157 for(i=0; i<4; i++){\ | |
158 int fx = s->me.direct_basis_mv[i][0] + hx;\ | |
159 int fy = s->me.direct_basis_mv[i][1] + hy;\ | |
160 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;\ | |
161 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;\ | |
162 int fxy= (fx&1) + 2*(fy&1);\ | |
163 int bxy= (bx&1) + 2*(by&1);\ | |
164 \ | |
165 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\ | |
166 hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\ | |
167 hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\ | |
168 }\ | |
169 }else{\ | |
170 int fx = s->me.direct_basis_mv[0][0] + hx;\ | |
171 int fy = s->me.direct_basis_mv[0][1] + hy;\ | |
172 int bx = hx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\ | |
173 int by = hy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\ | |
174 int fxy= (fx&1) + 2*(fy&1);\ | |
175 int bxy= (bx&1) + 2*(by&1);\ | |
176 \ | |
177 hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\ | |
178 hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\ | |
179 }\ | |
180 d = cmp_func(s, s->me.scratchpad, src_y, stride);\ | |
181 }else\ | |
182 d= 256*256*256*32; | |
183 | |
184 | |
185 #define CMP_HPEL(d, dx, dy, x, y, size)\ | |
186 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub) | |
187 | |
188 #define CMP(d, x, y, size)\ | |
189 CMP_DIRECT(d, 0, 0, x, y, size, cmp) | |
190 | |
191 #include "motion_est_template.c" | |
192 #undef RENAME | |
193 #undef CMP | |
194 #undef CMP_HPEL | |
195 #undef CMP_QPEL | |
196 #undef INIT | |
197 #undef CMP_DIRECT | |
198 | |
199 /* SIMPLE DIRECT QPEL */ | |
200 #define RENAME(a) simple_direct_qpel_ ## a | |
201 | |
202 #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\ | |
203 if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\ | |
204 const int qx= 4*(x) + (dx);\ | |
205 const int qy= 4*(y) + (dy);\ | |
206 if(s->mv_type==MV_TYPE_8X8){\ | |
207 int i;\ | |
208 for(i=0; i<4; i++){\ | |
209 int fx = s->me.direct_basis_mv[i][0] + qx;\ | |
210 int fy = s->me.direct_basis_mv[i][1] + qy;\ | |
211 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;\ | |
212 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;\ | |
213 int fxy= (fx&3) + 4*(fy&3);\ | |
214 int bxy= (bx&3) + 4*(by&3);\ | |
215 \ | |
216 uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\ | |
217 qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\ | |
218 qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\ | |
219 }\ | |
220 }else{\ | |
221 int fx = s->me.direct_basis_mv[0][0] + qx;\ | |
222 int fy = s->me.direct_basis_mv[0][1] + qy;\ | |
223 int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\ | |
224 int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\ | |
225 int fxy= (fx&3) + 4*(fy&3);\ | |
226 int bxy= (bx&3) + 4*(by&3);\ | |
227 \ | |
228 qpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\ | |
229 qpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\ | |
230 }\ | |
231 d = cmp_func(s, s->me.scratchpad, src_y, stride);\ | |
232 }else\ | |
233 d= 256*256*256*32; | |
234 | |
235 | |
236 #define CMP_QPEL(d, dx, dy, x, y, size)\ | |
237 CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub) | |
238 | |
239 #define CMP(d, x, y, size)\ | |
240 CMP_DIRECT(d, 0, 0, x, y, size, cmp) | |
241 | |
242 #include "motion_est_template.c" | |
243 #undef RENAME | |
244 #undef CMP | |
245 #undef CMP_HPEL | |
246 #undef CMP_QPEL | |
247 #undef INIT | |
248 #undef CMP__DIRECT | |
249 | |
250 | |
251 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){ | |
252 return 0; | |
253 } | |
254 | |
255 static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){ | |
256 DSPContext* c= &s->dsp; | |
257 int i; | |
258 | |
259 memset(cmp, 0, sizeof(void*)*11); | |
260 | |
261 switch(type&0xFF){ | |
262 case FF_CMP_SAD: | |
263 cmp[0]= c->sad[0]; | |
264 cmp[1]= c->sad[1]; | |
265 break; | |
266 case FF_CMP_SATD: | |
267 cmp[0]= c->hadamard8_diff[0]; | |
268 cmp[1]= c->hadamard8_diff[1]; | |
269 break; | |
270 case FF_CMP_SSE: | |
271 cmp[0]= c->sse[0]; | |
272 cmp[1]= c->sse[1]; | |
273 break; | |
274 case FF_CMP_DCT: | |
275 cmp[0]= c->dct_sad[0]; | |
276 cmp[1]= c->dct_sad[1]; | |
277 break; | |
278 case FF_CMP_PSNR: | |
279 cmp[0]= c->quant_psnr[0]; | |
280 cmp[1]= c->quant_psnr[1]; | |
281 break; | |
282 case FF_CMP_ZERO: | |
283 for(i=0; i<7; i++){ | |
284 cmp[i]= zero_cmp; | |
285 } | |
286 break; | |
287 default: | |
288 fprintf(stderr,"internal error in cmp function selection\n"); | |
289 } | |
290 }; | |
291 | |
292 static inline int get_penalty_factor(MpegEncContext *s, int type){ | |
293 | |
294 switch(type){ | |
295 default: | |
296 case FF_CMP_SAD: | |
297 return s->qscale; | |
298 case FF_CMP_SSE: | |
299 // return s->qscale*8; | |
300 case FF_CMP_DCT: | |
301 case FF_CMP_SATD: | |
302 return s->qscale*8; | |
303 } | |
304 } | |
305 | |
306 void ff_init_me(MpegEncContext *s){ | |
307 set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp); | |
308 set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp); | |
309 set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp); | |
310 | |
311 if(s->flags&CODEC_FLAG_QPEL){ | |
312 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
313 s->me.sub_motion_search= simple_chroma_qpel_motion_search; | |
314 else | |
315 s->me.sub_motion_search= simple_qpel_motion_search; | |
316 }else{ | |
317 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
318 s->me.sub_motion_search= simple_chroma_hpel_motion_search; | |
319 else if(s->avctx->me_sub_cmp == FF_CMP_SAD && s->avctx->me_cmp == FF_CMP_SAD) | |
320 s->me.sub_motion_search= sad_hpel_motion_search; | |
321 else | |
322 s->me.sub_motion_search= simple_hpel_motion_search; | |
323 } | |
324 | |
325 if(s->avctx->me_cmp&FF_CMP_CHROMA){ | |
326 s->me.motion_search[0]= simple_chroma_epzs_motion_search; | |
327 s->me.motion_search[1]= simple_chroma_epzs_motion_search4; | |
328 }else{ | |
329 s->me.motion_search[0]= simple_epzs_motion_search; | |
330 s->me.motion_search[1]= simple_epzs_motion_search4; | |
331 } | |
332 } | |
333 | |
284 | 334 static int pix_dev(UINT8 * pix, int line_size, int mean) |
335 { | |
336 int s, i, j; | |
337 | |
338 s = 0; | |
339 for (i = 0; i < 16; i++) { | |
340 for (j = 0; j < 16; j += 8) { | |
341 s += ABS(pix[0]-mean); | |
342 s += ABS(pix[1]-mean); | |
343 s += ABS(pix[2]-mean); | |
344 s += ABS(pix[3]-mean); | |
345 s += ABS(pix[4]-mean); | |
346 s += ABS(pix[5]-mean); | |
347 s += ABS(pix[6]-mean); | |
348 s += ABS(pix[7]-mean); | |
349 pix += 8; | |
350 } | |
351 pix += line_size - 16; | |
352 } | |
353 return s; | |
354 } | |
355 | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
356 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
|
357 int *mx_ptr, int *my_ptr) |
0 | 358 { |
359 *mx_ptr = 16 * s->mb_x; | |
360 *my_ptr = 16 * s->mb_y; | |
361 } | |
362 | |
363 static int full_motion_search(MpegEncContext * s, | |
364 int *mx_ptr, int *my_ptr, int range, | |
324 | 365 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 366 { |
367 int x1, y1, x2, y2, xx, yy, x, y; | |
368 int mx, my, dmin, d; | |
369 UINT8 *pix; | |
370 | |
371 xx = 16 * s->mb_x; | |
372 yy = 16 * s->mb_y; | |
373 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */ | |
374 if (x1 < xmin) | |
375 x1 = xmin; | |
376 x2 = xx + range - 1; | |
377 if (x2 > xmax) | |
378 x2 = xmax; | |
379 y1 = yy - range + 1; | |
380 if (y1 < ymin) | |
381 y1 = ymin; | |
382 y2 = yy + range - 1; | |
383 if (y2 > ymax) | |
384 y2 = ymax; | |
903 | 385 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 386 dmin = 0x7fffffff; |
387 mx = 0; | |
388 my = 0; | |
389 for (y = y1; y <= y2; y++) { | |
390 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
|
391 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, |
294 | 392 s->linesize); |
0 | 393 if (d < dmin || |
394 (d == dmin && | |
395 (abs(x - xx) + abs(y - yy)) < | |
396 (abs(mx - xx) + abs(my - yy)))) { | |
397 dmin = d; | |
398 mx = x; | |
399 my = y; | |
400 } | |
401 } | |
402 } | |
403 | |
404 *mx_ptr = mx; | |
405 *my_ptr = my; | |
406 | |
407 #if 0 | |
408 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) || | |
409 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) { | |
410 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr); | |
411 } | |
412 #endif | |
413 return dmin; | |
414 } | |
415 | |
416 | |
417 static int log_motion_search(MpegEncContext * s, | |
418 int *mx_ptr, int *my_ptr, int range, | |
324 | 419 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 420 { |
421 int x1, y1, x2, y2, xx, yy, x, y; | |
422 int mx, my, dmin, d; | |
423 UINT8 *pix; | |
424 | |
425 xx = s->mb_x << 4; | |
426 yy = s->mb_y << 4; | |
427 | |
428 /* Left limit */ | |
429 x1 = xx - range; | |
430 if (x1 < xmin) | |
431 x1 = xmin; | |
432 | |
433 /* Right limit */ | |
434 x2 = xx + range; | |
435 if (x2 > xmax) | |
436 x2 = xmax; | |
437 | |
438 /* Upper limit */ | |
439 y1 = yy - range; | |
440 if (y1 < ymin) | |
441 y1 = ymin; | |
442 | |
443 /* Lower limit */ | |
444 y2 = yy + range; | |
445 if (y2 > ymax) | |
446 y2 = ymax; | |
447 | |
903 | 448 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 449 dmin = 0x7fffffff; |
450 mx = 0; | |
451 my = 0; | |
452 | |
453 do { | |
454 for (y = y1; y <= y2; y += range) { | |
455 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
|
456 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 457 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
458 dmin = d; | |
459 mx = x; | |
460 my = y; | |
461 } | |
462 } | |
463 } | |
464 | |
465 range = range >> 1; | |
466 | |
467 x1 = mx - range; | |
468 if (x1 < xmin) | |
469 x1 = xmin; | |
470 | |
471 x2 = mx + range; | |
472 if (x2 > xmax) | |
473 x2 = xmax; | |
474 | |
475 y1 = my - range; | |
476 if (y1 < ymin) | |
477 y1 = ymin; | |
478 | |
479 y2 = my + range; | |
480 if (y2 > ymax) | |
481 y2 = ymax; | |
482 | |
483 } while (range >= 1); | |
484 | |
485 #ifdef DEBUG | |
486 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my); | |
487 #endif | |
488 *mx_ptr = mx; | |
489 *my_ptr = my; | |
490 return dmin; | |
491 } | |
492 | |
493 static int phods_motion_search(MpegEncContext * s, | |
494 int *mx_ptr, int *my_ptr, int range, | |
324 | 495 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 496 { |
497 int x1, y1, x2, y2, xx, yy, x, y, lastx, d; | |
498 int mx, my, dminx, dminy; | |
499 UINT8 *pix; | |
500 | |
501 xx = s->mb_x << 4; | |
502 yy = s->mb_y << 4; | |
503 | |
504 /* Left limit */ | |
505 x1 = xx - range; | |
506 if (x1 < xmin) | |
507 x1 = xmin; | |
508 | |
509 /* Right limit */ | |
510 x2 = xx + range; | |
511 if (x2 > xmax) | |
512 x2 = xmax; | |
513 | |
514 /* Upper limit */ | |
515 y1 = yy - range; | |
516 if (y1 < ymin) | |
517 y1 = ymin; | |
518 | |
519 /* Lower limit */ | |
520 y2 = yy + range; | |
521 if (y2 > ymax) | |
522 y2 = ymax; | |
523 | |
903 | 524 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 525 mx = 0; |
526 my = 0; | |
527 | |
528 x = xx; | |
529 y = yy; | |
530 do { | |
531 dminx = 0x7fffffff; | |
532 dminy = 0x7fffffff; | |
533 | |
534 lastx = x; | |
535 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
|
536 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 537 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
538 dminx = d; | |
539 mx = x; | |
540 } | |
541 } | |
542 | |
543 x = lastx; | |
544 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
|
545 d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize); |
0 | 546 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
547 dminy = d; | |
548 my = y; | |
549 } | |
550 } | |
551 | |
552 range = range >> 1; | |
553 | |
554 x = mx; | |
555 y = my; | |
556 x1 = mx - range; | |
557 if (x1 < xmin) | |
558 x1 = xmin; | |
559 | |
560 x2 = mx + range; | |
561 if (x2 > xmax) | |
562 x2 = xmax; | |
563 | |
564 y1 = my - range; | |
565 if (y1 < ymin) | |
566 y1 = ymin; | |
567 | |
568 y2 = my + range; | |
569 if (y2 > ymax) | |
570 y2 = ymax; | |
571 | |
572 } while (range >= 1); | |
573 | |
574 #ifdef DEBUG | |
575 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my); | |
576 #endif | |
577 | |
578 /* half pixel search */ | |
579 *mx_ptr = mx; | |
580 *my_ptr = my; | |
581 return dminy; | |
582 } | |
583 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
584 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
585 #define Z_THRESHOLD 256 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
586 |
936 | 587 #define CHECK_SAD_HALF_MV(suffix, x, y) \ |
455 | 588 {\ |
589 d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\ | |
936 | 590 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\ |
455 | 591 COPY3_IF_LT(dminh, d, dx, x, dy, y)\ |
592 } | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
593 |
936 | 594 static inline int sad_hpel_motion_search(MpegEncContext * s, |
0 | 595 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
|
596 int xmin, int ymin, int xmax, int ymax, |
936 | 597 int pred_x, int pred_y, Picture *picture, |
598 int n, int size, uint16_t * const mv_penalty) | |
0 | 599 { |
936 | 600 uint8_t *ref_picture= picture->data[0]; |
601 uint32_t *score_map= s->me.score_map; | |
602 const int penalty_factor= s->me.sub_penalty_factor; | |
455 | 603 int mx, my, xx, yy, dminh; |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
604 UINT8 *pix, *ptr; |
936 | 605 op_pixels_abs_func pix_abs_x2; |
606 op_pixels_abs_func pix_abs_y2; | |
607 op_pixels_abs_func pix_abs_xy2; | |
608 | |
609 if(size==0){ | |
610 pix_abs_x2 = s->dsp.pix_abs16x16_x2; | |
611 pix_abs_y2 = s->dsp.pix_abs16x16_y2; | |
612 pix_abs_xy2= s->dsp.pix_abs16x16_xy2; | |
613 }else{ | |
614 pix_abs_x2 = s->dsp.pix_abs8x8_x2; | |
615 pix_abs_y2 = s->dsp.pix_abs8x8_y2; | |
616 pix_abs_xy2= s->dsp.pix_abs8x8_xy2; | |
863 | 617 } |
455 | 618 |
936 | 619 if(s->me.skip){ |
455 | 620 // printf("S"); |
621 *mx_ptr = 0; | |
622 *my_ptr = 0; | |
623 return dmin; | |
624 } | |
625 // printf("N"); | |
626 | |
627 xx = 16 * s->mb_x + 8*(n&1); | |
628 yy = 16 * s->mb_y + 8*(n>>1); | |
903 | 629 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 630 |
294 | 631 mx = *mx_ptr; |
632 my = *my_ptr; | |
455 | 633 ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx); |
634 | |
294 | 635 dminh = dmin; |
636 | |
637 if (mx > xmin && mx < xmax && | |
638 my > ymin && my < ymax) { | |
455 | 639 int dx=0, dy=0; |
640 int d, pen_x, pen_y; | |
641 const int index= (my<<ME_MAP_SHIFT) + mx; | |
642 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
643 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
644 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
645 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
646 mx<<=1; | |
647 my<<=1; | |
294 | 648 |
649 | |
650 pen_x= pred_x + mx; | |
651 pen_y= pred_y + my; | |
652 | |
653 ptr-= s->linesize; | |
455 | 654 if(t<=b){ |
936 | 655 CHECK_SAD_HALF_MV(y2 , 0, -1) |
455 | 656 if(l<=r){ |
936 | 657 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 658 if(t+r<=b+l){ |
936 | 659 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 660 ptr+= s->linesize; |
661 }else{ | |
662 ptr+= s->linesize; | |
936 | 663 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 664 } |
936 | 665 CHECK_SAD_HALF_MV(x2 , -1, 0) |
455 | 666 }else{ |
936 | 667 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 668 if(t+l<=b+r){ |
936 | 669 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 670 ptr+= s->linesize; |
671 }else{ | |
672 ptr+= s->linesize; | |
936 | 673 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 674 } |
936 | 675 CHECK_SAD_HALF_MV(x2 , +1, 0) |
455 | 676 } |
677 }else{ | |
678 if(l<=r){ | |
679 if(t+l<=b+r){ | |
936 | 680 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 681 ptr+= s->linesize; |
682 }else{ | |
683 ptr+= s->linesize; | |
936 | 684 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 685 } |
936 | 686 CHECK_SAD_HALF_MV(x2 , -1, 0) |
687 CHECK_SAD_HALF_MV(xy2, -1, +1) | |
455 | 688 }else{ |
689 if(t+r<=b+l){ | |
936 | 690 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 691 ptr+= s->linesize; |
692 }else{ | |
693 ptr+= s->linesize; | |
936 | 694 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 695 } |
936 | 696 CHECK_SAD_HALF_MV(x2 , +1, 0) |
697 CHECK_SAD_HALF_MV(xy2, +1, +1) | |
455 | 698 } |
936 | 699 CHECK_SAD_HALF_MV(y2 , 0, +1) |
455 | 700 } |
701 mx+=dx; | |
702 my+=dy; | |
294 | 703 |
704 }else{ | |
455 | 705 mx<<=1; |
706 my<<=1; | |
294 | 707 } |
708 | |
709 *mx_ptr = mx; | |
710 *my_ptr = my; | |
455 | 711 return dminh; |
294 | 712 } |
713 | |
455 | 714 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4) |
294 | 715 { |
324 | 716 const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2); |
294 | 717 |
324 | 718 s->p_mv_table[xy][0] = mx; |
719 s->p_mv_table[xy][1] = my; | |
294 | 720 |
721 /* has allready been set to the 4 MV if 4MV is done */ | |
455 | 722 if(mv4){ |
294 | 723 int mot_xy= s->block_index[0]; |
724 | |
725 s->motion_val[mot_xy ][0]= mx; | |
726 s->motion_val[mot_xy ][1]= my; | |
727 s->motion_val[mot_xy+1][0]= mx; | |
728 s->motion_val[mot_xy+1][1]= my; | |
729 | |
730 mot_xy += s->block_wrap[0]; | |
731 s->motion_val[mot_xy ][0]= mx; | |
732 s->motion_val[mot_xy ][1]= my; | |
733 s->motion_val[mot_xy+1][0]= mx; | |
734 s->motion_val[mot_xy+1][1]= my; | |
735 } | |
736 } | |
737 | |
324 | 738 static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code) |
739 { | |
740 *range = 8 * (1 << (f_code - 1)); | |
741 /* XXX: temporary kludge to avoid overflow for msmpeg4 */ | |
742 if (s->out_format == FMT_H263 && !s->h263_msmpeg4) | |
743 *range *= 2; | |
0 | 744 |
324 | 745 if (s->unrestricted_mv) { |
746 *xmin = -16; | |
747 *ymin = -16; | |
748 if (s->h263_plus) | |
749 *range *= 2; | |
750 if(s->avctx==NULL || s->avctx->codec->id!=CODEC_ID_MPEG4){ | |
751 *xmax = s->mb_width*16; | |
752 *ymax = s->mb_height*16; | |
753 }else { | |
754 /* XXX: dunno if this is correct but ffmpeg4 decoder wont like it otherwise | |
755 (cuz the drawn edge isnt large enough))*/ | |
756 *xmax = s->width; | |
757 *ymax = s->height; | |
758 } | |
759 } else { | |
760 *xmin = 0; | |
761 *ymin = 0; | |
762 *xmax = s->mb_width*16 - 16; | |
763 *ymax = s->mb_height*16 - 16; | |
764 } | |
765 } | |
766 | |
455 | 767 static inline int mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift) |
768 { | |
769 int block; | |
770 int P[10][2]; | |
903 | 771 uint8_t *ref_picture= s->last_picture.data[0]; |
455 | 772 int dmin_sum=0; |
936 | 773 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
455 | 774 |
775 for(block=0; block<4; block++){ | |
776 int mx4, my4; | |
777 int pred_x4, pred_y4; | |
778 int dmin4; | |
779 static const int off[4]= {2, 1, 1, -1}; | |
780 const int mot_stride = s->block_wrap[0]; | |
781 const int mot_xy = s->block_index[block]; | |
782 // const int block_x= (block&1); | |
783 // const int block_y= (block>>1); | |
784 #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way | |
785 const int rel_xmin4= xmin; | |
786 const int rel_xmax4= xmax; | |
787 const int rel_ymin4= ymin; | |
788 const int rel_ymax4= ymax; | |
789 #else | |
790 const int rel_xmin4= xmin - block_x*8; | |
791 const int rel_xmax4= xmax - block_x*8 + 8; | |
792 const int rel_ymin4= ymin - block_y*8; | |
793 const int rel_ymax4= ymax - block_y*8 + 8; | |
794 #endif | |
795 P_LAST[0] = s->motion_val[mot_xy ][0]; | |
796 P_LAST[1] = s->motion_val[mot_xy ][1]; | |
797 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; | |
798 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
799 P_LAST_RIGHT[0] = s->motion_val[mot_xy + 1][0]; | |
800 P_LAST_RIGHT[1] = s->motion_val[mot_xy + 1][1]; | |
801 P_LAST_BOTTOM[0]= s->motion_val[mot_xy + 1*mot_stride][0]; | |
802 P_LAST_BOTTOM[1]= s->motion_val[mot_xy + 1*mot_stride][1]; | |
803 | |
804 if(P_LEFT[0] > (rel_xmax4<<shift)) P_LEFT[0] = (rel_xmax4<<shift); | |
805 if(P_LAST_RIGHT[0] < (rel_xmin4<<shift)) P_LAST_RIGHT[0] = (rel_xmin4<<shift); | |
806 if(P_LAST_BOTTOM[1]< (rel_ymin4<<shift)) P_LAST_BOTTOM[1]= (rel_ymin4<<shift); | |
807 | |
808 /* special case for first line */ | |
809 if ((s->mb_y == 0 || s->first_slice_line) && block<2) { | |
810 pred_x4= P_LEFT[0]; | |
811 pred_y4= P_LEFT[1]; | |
812 } else { | |
813 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; | |
814 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
815 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0]; | |
816 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1]; | |
817 if(P_TOP[1] > (rel_ymax4<<shift)) P_TOP[1] = (rel_ymax4<<shift); | |
818 if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift); | |
819 if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift); | |
820 if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift); | |
821 | |
822 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
823 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
824 | |
825 if(s->out_format == FMT_H263){ | |
826 pred_x4 = P_MEDIAN[0]; | |
827 pred_y4 = P_MEDIAN[1]; | |
828 }else { /* mpeg1 at least */ | |
829 pred_x4= P_LEFT[0]; | |
830 pred_y4= P_LEFT[1]; | |
831 } | |
832 } | |
833 P_MV1[0]= mx; | |
834 P_MV1[1]= my; | |
835 | |
936 | 836 dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
837 &s->last_picture, mv_penalty); | |
455 | 838 |
936 | 839 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, |
840 pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty); | |
455 | 841 |
842 s->motion_val[ s->block_index[block] ][0]= mx4; | |
843 s->motion_val[ s->block_index[block] ][1]= my4; | |
844 dmin_sum+= dmin4; | |
845 } | |
846 return dmin_sum; | |
847 } | |
848 | |
324 | 849 void ff_estimate_p_frame_motion(MpegEncContext * s, |
850 int mb_x, int mb_y) | |
0 | 851 { |
852 UINT8 *pix, *ppix; | |
853 int sum, varc, vard, mx, my, range, dmin, xx, yy; | |
854 int xmin, ymin, xmax, ymax; | |
280 | 855 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
|
856 int pred_x=0, pred_y=0; |
455 | 857 int P[10][2]; |
280 | 858 const int shift= 1+s->quarter_sample; |
294 | 859 int mb_type=0; |
903 | 860 uint8_t *ref_picture= s->last_picture.data[0]; |
861 Picture * const pic= &s->current_picture; | |
936 | 862 uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
863 | |
864 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
865 | |
866 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
867 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
0 | 868 |
324 | 869 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code); |
455 | 870 rel_xmin= xmin - mb_x*16; |
871 rel_xmax= xmax - mb_x*16; | |
872 rel_ymin= ymin - mb_y*16; | |
873 rel_ymax= ymax - mb_y*16; | |
936 | 874 s->me.skip=0; |
324 | 875 |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
876 switch(s->me_method) { |
0 | 877 case ME_ZERO: |
878 default: | |
879 no_motion_search(s, &mx, &my); | |
455 | 880 mx-= mb_x*16; |
881 my-= mb_y*16; | |
0 | 882 dmin = 0; |
883 break; | |
884 case ME_FULL: | |
324 | 885 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); |
455 | 886 mx-= mb_x*16; |
887 my-= mb_y*16; | |
0 | 888 break; |
889 case ME_LOG: | |
324 | 890 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 891 mx-= mb_x*16; |
892 my-= mb_y*16; | |
0 | 893 break; |
894 case ME_PHODS: | |
324 | 895 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); |
455 | 896 mx-= mb_x*16; |
897 my-= mb_y*16; | |
0 | 898 break; |
288 | 899 case ME_X1: |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
900 case ME_EPZS: |
288 | 901 { |
294 | 902 const int mot_stride = s->block_wrap[0]; |
903 const int mot_xy = s->block_index[0]; | |
288 | 904 |
455 | 905 P_LAST[0] = s->motion_val[mot_xy ][0]; |
906 P_LAST[1] = s->motion_val[mot_xy ][1]; | |
907 P_LEFT[0] = s->motion_val[mot_xy - 1][0]; | |
908 P_LEFT[1] = s->motion_val[mot_xy - 1][1]; | |
909 P_LAST_RIGHT[0] = s->motion_val[mot_xy + 2][0]; | |
910 P_LAST_RIGHT[1] = s->motion_val[mot_xy + 2][1]; | |
911 P_LAST_BOTTOM[0]= s->motion_val[mot_xy + 2*mot_stride][0]; | |
912 P_LAST_BOTTOM[1]= s->motion_val[mot_xy + 2*mot_stride][1]; | |
288 | 913 |
455 | 914 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
915 if(P_LAST_RIGHT[0] < (rel_xmin<<shift)) P_LAST_RIGHT[0] = (rel_xmin<<shift); | |
916 if(P_LAST_BOTTOM[1]< (rel_ymin<<shift)) P_LAST_BOTTOM[1]= (rel_ymin<<shift); | |
280 | 917 |
918 /* special case for first line */ | |
455 | 919 if ((mb_y == 0 || s->first_slice_line)) { |
920 pred_x= P_LEFT[0]; | |
921 pred_y= P_LEFT[1]; | |
280 | 922 } else { |
455 | 923 P_TOP[0] = s->motion_val[mot_xy - mot_stride ][0]; |
924 P_TOP[1] = s->motion_val[mot_xy - mot_stride ][1]; | |
925 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0]; | |
926 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1]; | |
927 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1] = (rel_ymax<<shift); | |
928 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
929 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
280 | 930 |
455 | 931 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
932 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
933 | |
934 if(s->out_format == FMT_H263){ | |
935 pred_x = P_MEDIAN[0]; | |
936 pred_y = P_MEDIAN[1]; | |
937 }else { /* mpeg1 at least */ | |
938 pred_x= P_LEFT[0]; | |
939 pred_y= P_LEFT[1]; | |
940 } | |
288 | 941 } |
280 | 942 } |
936 | 943 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
944 &s->last_picture, mv_penalty); | |
281
1fc96b02142e
mpeg4 aspect_ratio_info in AVCodecContext (requested by alex)
michaelni
parents:
280
diff
changeset
|
945 |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
946 break; |
0 | 947 } |
948 | |
949 /* intra / predictive decision */ | |
950 xx = mb_x * 16; | |
951 yy = mb_y * 16; | |
952 | |
903 | 953 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
455 | 954 /* At this point (mx,my) are full-pell and the relative displacement */ |
955 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
|
956 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
957 sum = s->dsp.pix_sum(pix, s->linesize); |
455 | 958 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
959 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8; |
936 | 960 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8; |
608 | 961 |
455 | 962 //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 | 963 pic->mb_var [s->mb_width * mb_y + mb_x] = varc; |
964 pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard; | |
965 pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8; | |
966 pic->mb_var_sum += varc; | |
967 pic->mc_mb_var_sum += vard; | |
455 | 968 //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
|
969 |
0 | 970 #if 0 |
233
3f5b72726118
- More work on preliminary bit rate control, just to be able to get an
pulento
parents:
232
diff
changeset
|
971 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
|
972 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy); |
0 | 973 #endif |
294 | 974 if(s->flags&CODEC_FLAG_HQ){ |
608 | 975 if (vard <= 64 || vard < varc) |
976 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); | |
977 else | |
913 | 978 s->scene_change_score+= s->qscale; |
608 | 979 |
294 | 980 if (vard*2 + 200 > varc) |
981 mb_type|= MB_TYPE_INTRA; | |
982 if (varc*2 + 200 > vard){ | |
983 mb_type|= MB_TYPE_INTER; | |
936 | 984 s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
985 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
304 | 986 }else{ |
936 | 987 mx <<=shift; |
988 my <<=shift; | |
0 | 989 } |
455 | 990 if((s->flags&CODEC_FLAG_4MV) |
936 | 991 && !s->me.skip && varc>50 && vard>10){ |
455 | 992 mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
993 mb_type|=MB_TYPE_INTER4V; | |
994 | |
995 set_p_mv_tables(s, mx, my, 0); | |
996 }else | |
997 set_p_mv_tables(s, mx, my, 1); | |
294 | 998 }else{ |
999 if (vard <= 64 || vard < varc) { | |
936 | 1000 // if (sadP <= 32 || sadP < sadI + 500) { |
608 | 1001 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); |
294 | 1002 mb_type|= MB_TYPE_INTER; |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1003 if (s->me_method != ME_ZERO) { |
936 | 1004 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1005 pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty); | |
455 | 1006 if((s->flags&CODEC_FLAG_4MV) |
936 | 1007 && !s->me.skip && varc>50 && vard>10){ |
455 | 1008 int dmin4= mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift); |
1009 if(dmin4 + 128 <dmin) | |
1010 mb_type= MB_TYPE_INTER4V; | |
1011 } | |
1012 set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V); | |
1013 | |
294 | 1014 } else { |
936 | 1015 mx <<=shift; |
1016 my <<=shift; | |
294 | 1017 } |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1018 #if 0 |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1019 if (vard < 10) { |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1020 skip++; |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1021 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
|
1022 skip, vard, varc, dmin); |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1023 } |
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1024 #endif |
294 | 1025 }else{ |
608 | 1026 s->scene_change_score+= 20; |
294 | 1027 mb_type|= MB_TYPE_INTRA; |
455 | 1028 mx = 0; |
1029 my = 0; | |
294 | 1030 } |
1031 } | |
284 | 1032 |
294 | 1033 s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
0 | 1034 } |
1035 | |
327 | 1036 int ff_estimate_motion_b(MpegEncContext * s, |
936 | 1037 int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code) |
0 | 1038 { |
327 | 1039 int mx, my, range, dmin; |
324 | 1040 int xmin, ymin, xmax, ymax; |
1041 int rel_xmin, rel_ymin, rel_xmax, rel_ymax; | |
1042 int pred_x=0, pred_y=0; | |
455 | 1043 int P[10][2]; |
324 | 1044 const int shift= 1+s->quarter_sample; |
1045 const int mot_stride = s->mb_width + 2; | |
1046 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1047 uint8_t * const ref_picture= picture->data[0]; |
1048 uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV; | |
1049 | |
1050 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
1051 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
1052 | |
324 | 1053 get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code); |
455 | 1054 rel_xmin= xmin - mb_x*16; |
1055 rel_xmax= xmax - mb_x*16; | |
1056 rel_ymin= ymin - mb_y*16; | |
1057 rel_ymax= ymax - mb_y*16; | |
324 | 1058 |
1059 switch(s->me_method) { | |
1060 case ME_ZERO: | |
1061 default: | |
1062 no_motion_search(s, &mx, &my); | |
1063 dmin = 0; | |
455 | 1064 mx-= mb_x*16; |
1065 my-= mb_y*16; | |
324 | 1066 break; |
1067 case ME_FULL: | |
1068 dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1069 mx-= mb_x*16; |
1070 my-= mb_y*16; | |
324 | 1071 break; |
1072 case ME_LOG: | |
1073 dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1074 mx-= mb_x*16; |
1075 my-= mb_y*16; | |
324 | 1076 break; |
1077 case ME_PHODS: | |
1078 dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture); | |
455 | 1079 mx-= mb_x*16; |
1080 my-= mb_y*16; | |
324 | 1081 break; |
1082 case ME_X1: | |
1083 case ME_EPZS: | |
1084 { | |
0 | 1085 |
455 | 1086 P_LAST[0] = mv_table[mot_xy ][0]; |
1087 P_LAST[1] = mv_table[mot_xy ][1]; | |
1088 P_LEFT[0] = mv_table[mot_xy - 1][0]; | |
1089 P_LEFT[1] = mv_table[mot_xy - 1][1]; | |
1090 P_LAST_RIGHT[0] = mv_table[mot_xy + 1][0]; | |
1091 P_LAST_RIGHT[1] = mv_table[mot_xy + 1][1]; | |
1092 P_LAST_BOTTOM[0] = mv_table[mot_xy + mot_stride][0]; | |
1093 P_LAST_BOTTOM[1] = mv_table[mot_xy + mot_stride][1]; | |
324 | 1094 |
455 | 1095 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); |
1096 if(P_LAST_RIGHT[0] < (rel_xmin<<shift)) P_LAST_RIGHT[0] = (rel_xmin<<shift); | |
1097 if(P_LAST_BOTTOM[1]< (rel_ymin<<shift)) P_LAST_BOTTOM[1]= (rel_ymin<<shift); | |
324 | 1098 |
1099 /* special case for first line */ | |
455 | 1100 if ((mb_y == 0 || s->first_slice_line)) { |
324 | 1101 } else { |
455 | 1102 P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
1103 P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; | |
1104 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0]; | |
1105 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1]; | |
1106 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift); | |
1107 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift); | |
1108 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift); | |
324 | 1109 |
455 | 1110 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1111 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
324 | 1112 } |
455 | 1113 pred_x= P_LEFT[0]; |
1114 pred_y= P_LEFT[1]; | |
324 | 1115 } |
936 | 1116 dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1117 picture, mv_penalty); | |
324 | 1118 |
1119 break; | |
1120 } | |
1121 | |
936 | 1122 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax, |
1123 pred_x, pred_y, picture, 0, 0, mv_penalty); | |
455 | 1124 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my); |
324 | 1125 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
1126 mv_table[mot_xy][0]= mx; | |
1127 mv_table[mot_xy][1]= my; | |
936 | 1128 |
327 | 1129 return dmin; |
324 | 1130 } |
1131 | |
327 | 1132 static inline int check_bidir_mv(MpegEncContext * s, |
1133 int mb_x, int mb_y, | |
1134 int motion_fx, int motion_fy, | |
1135 int motion_bx, int motion_by, | |
1136 int pred_fx, int pred_fy, | |
1137 int pred_bx, int pred_by) | |
1138 { | |
1139 //FIXME optimize? | |
936 | 1140 //FIXME move into template? |
1141 //FIXME better f_code prediction (max mv & distance) | |
1142 UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame | |
1143 uint8_t *dest_y = s->me.scratchpad; | |
327 | 1144 uint8_t *ptr; |
1145 int dxy; | |
1146 int src_x, src_y; | |
1147 int fbmin; | |
1148 | |
936 | 1149 if(s->quarter_sample){ |
1150 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3); | |
1151 src_x = mb_x * 16 + (motion_fx >> 2); | |
1152 src_y = mb_y * 16 + (motion_fy >> 2); | |
1153 assert(src_x >=-16 && src_x<=s->width); | |
1154 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1155 |
936 | 1156 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1157 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
|
1158 |
936 | 1159 dxy = ((motion_by & 3) << 2) | (motion_bx & 3); |
1160 src_x = mb_x * 16 + (motion_bx >> 2); | |
1161 src_y = mb_y * 16 + (motion_by >> 2); | |
1162 assert(src_x >=-16 && src_x<=s->width); | |
1163 assert(src_y >=-16 && src_y<=s->height); | |
1164 | |
1165 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1166 s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize); | |
1167 }else{ | |
1168 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1); | |
1169 src_x = mb_x * 16 + (motion_fx >> 1); | |
1170 src_y = mb_y * 16 + (motion_fy >> 1); | |
1171 assert(src_x >=-16 && src_x<=s->width); | |
1172 assert(src_y >=-16 && src_y<=s->height); | |
327 | 1173 |
936 | 1174 ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x; |
1175 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
|
1176 |
936 | 1177 dxy = ((motion_by & 1) << 1) | (motion_bx & 1); |
1178 src_x = mb_x * 16 + (motion_bx >> 1); | |
1179 src_y = mb_y * 16 + (motion_by >> 1); | |
1180 assert(src_x >=-16 && src_x<=s->width); | |
1181 assert(src_y >=-16 && src_y<=s->height); | |
1182 | |
1183 ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x; | |
1184 s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16); | |
1185 } | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1186 |
936 | 1187 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.sub_penalty_factor |
1188 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.sub_penalty_factor; | |
1189 + s->dsp.me_sub_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize); | |
1190 | |
327 | 1191 return fbmin; |
1192 } | |
1193 | |
1194 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/ | |
1195 static inline int bidir_refine(MpegEncContext * s, | |
1196 int mb_x, int mb_y) | |
1197 { | |
1198 const int mot_stride = s->mb_width + 2; | |
1199 const int xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
1200 int fbmin; | |
1201 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0]; | |
1202 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1]; | |
1203 int pred_bx= s->b_bidir_back_mv_table[xy-1][0]; | |
1204 int pred_by= s->b_bidir_back_mv_table[xy-1][1]; | |
1205 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0]; | |
1206 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1]; | |
1207 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0]; | |
1208 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1]; | |
1209 | |
1210 //FIXME do refinement and add flag | |
1211 | |
1212 fbmin= check_bidir_mv(s, mb_x, mb_y, | |
1213 motion_fx, motion_fy, | |
1214 motion_bx, motion_by, | |
1215 pred_fx, pred_fy, | |
1216 pred_bx, pred_by); | |
1217 | |
1218 return fbmin; | |
1219 } | |
1220 | |
1221 static inline int direct_search(MpegEncContext * s, | |
1222 int mb_x, int mb_y) | |
324 | 1223 { |
455 | 1224 int P[10][2]; |
327 | 1225 const int mot_stride = s->mb_width + 2; |
1226 const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1; | |
936 | 1227 const int shift= 1+s->quarter_sample; |
1228 int dmin, i; | |
327 | 1229 const int time_pp= s->pp_time; |
664 | 1230 const int time_pb= s->pb_time; |
936 | 1231 int mx, my, xmin, xmax, ymin, ymax; |
327 | 1232 int16_t (*mv_table)[2]= s->b_direct_mv_table; |
936 | 1233 uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV; |
1234 | |
455 | 1235 P_LAST[0] = mv_table[mot_xy ][0]; |
1236 P_LAST[1] = mv_table[mot_xy ][1]; | |
1237 P_LEFT[0] = mv_table[mot_xy - 1][0]; | |
1238 P_LEFT[1] = mv_table[mot_xy - 1][1]; | |
1239 P_LAST_RIGHT[0] = mv_table[mot_xy + 1][0]; | |
1240 P_LAST_RIGHT[1] = mv_table[mot_xy + 1][1]; | |
1241 P_LAST_BOTTOM[0] = mv_table[mot_xy + mot_stride][0]; | |
1242 P_LAST_BOTTOM[1] = mv_table[mot_xy + mot_stride][1]; | |
1243 /* | |
1244 if(P_LEFT[0] > (rel_xmax<<shift)) P_LEFT[0] = (rel_xmax<<shift); | |
1245 if(P_LAST_RIGHT[0] < (rel_xmin<<shift)) P_LAST_RIGHT[0] = (rel_xmin<<shift); | |
1246 if(P_LAST_BOTTOM[1]< (rel_ymin<<shift)) P_LAST_BOTTOM[1]= (rel_ymin<<shift); | |
1247 */ | |
327 | 1248 /* special case for first line */ |
455 | 1249 if ((mb_y == 0 || s->first_slice_line)) { |
327 | 1250 } else { |
455 | 1251 P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
1252 P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; | |
1253 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0]; | |
1254 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1]; | |
327 | 1255 |
455 | 1256 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1257 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
327 | 1258 } |
936 | 1259 |
1260 ymin= xmin=(-32)>>shift; | |
1261 ymax= xmax= 31>>shift; | |
1262 | |
1263 if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){ | |
1264 s->mv_type= MV_TYPE_8X8; | |
1265 }else{ | |
1266 s->mv_type= MV_TYPE_16X16; | |
327 | 1267 } |
936 | 1268 |
1269 for(i=0; i<4; i++){ | |
1270 int index= s->block_index[i]; | |
1271 int min, max; | |
1272 | |
1273 s->me.co_located_mv[i][0]= s->motion_val[index][0]; | |
1274 s->me.co_located_mv[i][1]= s->motion_val[index][1]; | |
1275 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3)); | |
1276 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3)); | |
1277 // 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); | |
1278 // 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); | |
1279 | |
1280 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1281 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1282 max+= (2*mb_x + (i& 1))*8 - 1; // +-1 is for the simpler rounding | |
1283 min+= (2*mb_x + (i& 1))*8 + 1; | |
1284 if(max >= s->width) xmax= s->width - max - 1; | |
1285 if(min < -16 ) xmin= - 32 - min; | |
1286 | |
1287 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1288 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1289 max+= (2*mb_y + (i>>1))*8 - 1; // +-1 is for the simpler rounding | |
1290 min+= (2*mb_y + (i>>1))*8 + 1; | |
1291 if(max >= s->height) ymax= s->height - max - 1; | |
1292 if(min < -16 ) ymin= - 32 - min; | |
1293 | |
1294 if(s->mv_type == MV_TYPE_16X16) break; | |
327 | 1295 } |
936 | 1296 |
1297 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16); | |
1298 | |
1299 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){ | |
1300 s->b_direct_mv_table[mot_xy][0]= 0; | |
1301 s->b_direct_mv_table[mot_xy][1]= 0; | |
1302 | |
1303 return 256*256*256*64; | |
1304 } | |
1305 | |
1306 if(s->flags&CODEC_FLAG_QPEL){ | |
1307 dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
1308 &s->last_picture, mv_penalty); | |
1309 dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, | |
1310 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
1311 }else{ | |
1312 dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, | |
1313 &s->last_picture, mv_penalty); | |
1314 dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, | |
1315 0, 0, &s->last_picture, 0, 0, mv_penalty); | |
327 | 1316 } |
1317 | |
1318 s->b_direct_mv_table[mot_xy][0]= mx; | |
1319 s->b_direct_mv_table[mot_xy][1]= my; | |
1320 return dmin; | |
324 | 1321 } |
1322 | |
1323 void ff_estimate_b_frame_motion(MpegEncContext * s, | |
1324 int mb_x, int mb_y) | |
1325 { | |
936 | 1326 const int penalty_factor= s->me.penalty_factor; |
327 | 1327 int fmin, bmin, dmin, fbmin; |
1328 int type=0; | |
1329 | |
1330 dmin= direct_search(s, mb_x, mb_y); | |
324 | 1331 |
936 | 1332 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code); |
1333 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) - penalty_factor; | |
324 | 1334 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
327 | 1335 |
1336 fbmin= bidir_refine(s, mb_x, mb_y); | |
1337 | |
612 | 1338 { |
327 | 1339 int score= dmin; |
1340 type=MB_TYPE_DIRECT; | |
1341 | |
1342 if(fmin<score){ | |
1343 score=fmin; | |
1344 type= MB_TYPE_FORWARD; | |
1345 } | |
1346 if(bmin<score){ | |
1347 score=bmin; | |
1348 type= MB_TYPE_BACKWARD; | |
1349 } | |
1350 if(fbmin<score){ | |
1351 score=fbmin; | |
1352 type= MB_TYPE_BIDIR; | |
1353 } | |
807
0e1d375c537f
fixing q>0.0 assert failure caused by overflow of variance for b frames
michaelni
parents:
765
diff
changeset
|
1354 score= ((unsigned)(score*score + 128*256))>>16; |
903 | 1355 s->current_picture.mc_mb_var_sum += score; |
1356 s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSD | |
327 | 1357 } |
612 | 1358 |
1359 if(s->flags&CODEC_FLAG_HQ){ | |
1360 type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter | |
936 | 1361 if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB |
612 | 1362 } |
1363 | |
327 | 1364 s->mb_type[mb_y*s->mb_width + mb_x]= type; |
324 | 1365 } |
0 | 1366 |
324 | 1367 /* find best f_code for ME which do unlimited searches */ |
1368 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type) | |
1369 { | |
1370 if(s->me_method>=ME_EPZS){ | |
455 | 1371 int score[8]; |
324 | 1372 int i, y; |
1373 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1374 int best_fcode=-1; |
1375 int best_score=-10000000; | |
324 | 1376 |
936 | 1377 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); |
324 | 1378 |
1379 for(y=0; y<s->mb_height; y++){ | |
1380 int x; | |
1381 int xy= (y+1)* (s->mb_width+2) + 1; | |
1382 i= y*s->mb_width; | |
1383 for(x=0; x<s->mb_width; x++){ | |
1384 if(s->mb_type[i] & type){ | |
847 | 1385 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV], |
1386 fcode_tab[mv_table[xy][1] + MAX_MV]); | |
455 | 1387 int j; |
1388 | |
1389 for(j=0; j<fcode && j<8; j++){ | |
903 | 1390 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i]) |
455 | 1391 score[j]-= 170; |
1392 } | |
324 | 1393 } |
1394 i++; | |
1395 xy++; | |
1396 } | |
1397 } | |
455 | 1398 |
1399 for(i=1; i<8; i++){ | |
1400 if(score[i] > best_score){ | |
1401 best_score= score[i]; | |
1402 best_fcode= i; | |
1403 } | |
1404 // printf("%d %d\n", i, score[i]); | |
1405 } | |
327 | 1406 |
324 | 1407 // printf("fcode: %d type: %d\n", i, s->pict_type); |
455 | 1408 return best_fcode; |
324 | 1409 /* for(i=0; i<=MAX_FCODE; i++){ |
1410 printf("%d ", mv_num[i]); | |
1411 } | |
1412 printf("\n");*/ | |
1413 }else{ | |
1414 return 1; | |
0 | 1415 } |
1416 } | |
1417 | |
324 | 1418 void ff_fix_long_p_mvs(MpegEncContext * s) |
1419 { | |
1420 const int f_code= s->f_code; | |
1421 int y; | |
1422 UINT8 * fcode_tab= s->fcode_tab; | |
455 | 1423 //int clip=0; |
1424 //int noclip=0; | |
324 | 1425 /* clip / convert to intra 16x16 type MVs */ |
1426 for(y=0; y<s->mb_height; y++){ | |
1427 int x; | |
1428 int xy= (y+1)* (s->mb_width+2)+1; | |
1429 int i= y*s->mb_width; | |
1430 for(x=0; x<s->mb_width; x++){ | |
1431 if(s->mb_type[i]&MB_TYPE_INTER){ | |
1432 if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code | |
1433 || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0 | |
1434 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code | |
1435 || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){ | |
1436 s->mb_type[i] &= ~MB_TYPE_INTER; | |
1437 s->mb_type[i] |= MB_TYPE_INTRA; | |
1438 s->p_mv_table[xy][0] = 0; | |
1439 s->p_mv_table[xy][1] = 0; | |
455 | 1440 //clip++; |
324 | 1441 } |
455 | 1442 //else |
1443 // noclip++; | |
324 | 1444 } |
1445 xy++; | |
1446 i++; | |
1447 } | |
1448 } | |
455 | 1449 //printf("%d no:%d %d//\n", clip, noclip, f_code); |
324 | 1450 if(s->flags&CODEC_FLAG_4MV){ |
1451 const int wrap= 2+ s->mb_width*2; | |
1452 | |
1453 /* clip / convert to intra 8x8 type MVs */ | |
1454 for(y=0; y<s->mb_height; y++){ | |
1455 int xy= (y*2 + 1)*wrap + 1; | |
1456 int i= y*s->mb_width; | |
1457 int x; | |
1458 | |
1459 for(x=0; x<s->mb_width; x++){ | |
1460 if(s->mb_type[i]&MB_TYPE_INTER4V){ | |
1461 int block; | |
1462 for(block=0; block<4; block++){ | |
1463 int off= (block& 1) + (block>>1)*wrap; | |
1464 int mx= s->motion_val[ xy + off ][0]; | |
1465 int my= s->motion_val[ xy + off ][1]; | |
1466 | |
1467 if( fcode_tab[mx + MAX_MV] > f_code | |
1468 || fcode_tab[mx + MAX_MV] == 0 | |
1469 || fcode_tab[my + MAX_MV] > f_code | |
1470 || fcode_tab[my + MAX_MV] == 0 ){ | |
1471 s->mb_type[i] &= ~MB_TYPE_INTER4V; | |
1472 s->mb_type[i] |= MB_TYPE_INTRA; | |
1473 } | |
1474 } | |
1475 } | |
502 | 1476 xy+=2; |
1477 i++; | |
324 | 1478 } |
1479 } | |
1480 } | |
1481 } | |
1482 | |
1483 void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type) | |
1484 { | |
1485 int y; | |
1486 UINT8 * fcode_tab= s->fcode_tab; | |
1487 | |
1488 /* clip / convert to intra 16x16 type MVs */ | |
1489 for(y=0; y<s->mb_height; y++){ | |
1490 int x; | |
1491 int xy= (y+1)* (s->mb_width+2)+1; | |
1492 int i= y*s->mb_width; | |
1493 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
|
1494 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
|
1495 || 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
|
1496 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
|
1497 else mv_table[xy][0]= -(16<<f_code); |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1498 } |
199b324b2693
fixing variance scaling for b frames (messed adaptive quants up)
michaelni
parents:
690
diff
changeset
|
1499 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
|
1500 || 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
|
1501 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
|
1502 else mv_table[xy][1]= -(16<<f_code); |
324 | 1503 } |
1504 xy++; | |
1505 i++; | |
1506 } | |
1507 } | |
1508 } |