Mercurial > libavcodec.hg
annotate motion_est.c @ 1950:a3c60fa850dc libavcodec
motion estimation cleanup
replace ugly macros by always_inline functions, that way its much more readable and flexible as always_inline can simply be removed while the macros couldnt be
about 0.5 % speedup with default parameters
author | michael |
---|---|
date | Thu, 22 Apr 2004 03:31:29 +0000 |
parents | e2501e6e7ff7 |
children | 5dafb10e0252 |
rev | line source |
---|---|
0 | 1 /* |
2 * Motion estimation | |
429 | 3 * Copyright (c) 2000,2001 Fabrice Bellard. |
1739
07a484280a82
copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents:
1729
diff
changeset
|
4 * Copyright (c) 2002-2004 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 */ |
1106 | 23 |
24 /** | |
25 * @file motion_est.c | |
26 * Motion estimation. | |
27 */ | |
28 | |
0 | 29 #include <stdlib.h> |
30 #include <stdio.h> | |
1426 | 31 #include <limits.h> |
0 | 32 #include "avcodec.h" |
33 #include "dsputil.h" | |
34 #include "mpegvideo.h" | |
35 | |
1950 | 36 #undef NDEBUG |
37 #include <assert.h> | |
936 | 38 |
455 | 39 #define SQ(a) ((a)*(a)) |
284 | 40 |
455 | 41 #define P_LEFT P[1] |
42 #define P_TOP P[2] | |
43 #define P_TOPRIGHT P[3] | |
44 #define P_MEDIAN P[4] | |
45 #define P_MV1 P[9] | |
46 | |
936 | 47 static inline int sad_hpel_motion_search(MpegEncContext * s, |
48 int *mx_ptr, int *my_ptr, int dmin, | |
1950 | 49 int src_index, int ref_index, |
50 int size, int h); | |
0 | 51 |
936 | 52 static inline int update_map_generation(MpegEncContext * s) |
53 { | |
54 s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2); | |
55 if(s->me.map_generation==0){ | |
56 s->me.map_generation= 1<<(ME_MAP_MV_BITS*2); | |
57 memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE); | |
58 } | |
59 return s->me.map_generation; | |
60 } | |
61 | |
948 | 62 /* shape adaptive search stuff */ |
63 typedef struct Minima{ | |
64 int height; | |
65 int x, y; | |
66 int checked; | |
67 }Minima; | |
936 | 68 |
948 | 69 static int minima_cmp(const void *a, const void *b){ |
1057 | 70 const Minima *da = (const Minima *) a; |
71 const Minima *db = (const Minima *) b; | |
948 | 72 |
73 return da->height - db->height; | |
74 } | |
1950 | 75 |
76 #define FLAG_QPEL 1 //must be 1 | |
77 #define FLAG_CHROMA 2 | |
78 #define FLAG_DIRECT 4 | |
936 | 79 |
1950 | 80 static inline void init_ref(MpegEncContext *s, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){ |
81 MotionEstContext * const c= &s->me; | |
82 const int offset[3]= { | |
83 y*c-> stride + x, | |
84 ((y*c->uvstride + x)>>1), | |
85 ((y*c->uvstride + x)>>1), | |
86 }; | |
87 int i; | |
88 for(i=0; i<3; i++){ | |
89 c->src[0][i]= src [i] + offset[i]; | |
90 c->ref[0][i]= ref [i] + offset[i]; | |
91 } | |
92 if(ref_index){ | |
93 for(i=0; i<3; i++){ | |
94 c->ref[ref_index][i]= ref2[i] + offset[i]; | |
95 } | |
96 } | |
936 | 97 } |
98 | |
1950 | 99 static int get_flags(MpegEncContext *s, int direct, int chroma){ |
100 return ((s->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0) | |
101 + (direct ? FLAG_DIRECT : 0) | |
102 + (chroma ? FLAG_CHROMA : 0); | |
103 } | |
1708 | 104 |
1950 | 105 static inline void init_mc(MpegEncContext *s, int size, int flags){ |
106 MotionEstContext * const c= &s->me; | |
107 | |
108 /*FIXME s->no_rounding b_type*/ | |
109 if(flags & FLAG_CHROMA){ | |
110 if(s->no_rounding) c->chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1]; | |
111 else c->chroma_hpel_put= &s->dsp.put_pixels_tab[size+1]; | |
112 } | |
113 | |
114 if(flags & FLAG_QPEL){ | |
115 c->qpel_avg= &s->dsp.avg_qpel_pixels_tab[size]; | |
116 if(s->no_rounding) c->qpel_put= &s->dsp.put_no_rnd_qpel_pixels_tab[size]; | |
117 else c->qpel_put= &s->dsp.put_qpel_pixels_tab[size]; | |
118 }else{ | |
119 c->hpel_avg= &s->dsp.avg_pixels_tab[size]; | |
120 if(s->no_rounding) c->hpel_put= &s->dsp.put_no_rnd_pixels_tab[size]; | |
121 else c->hpel_put= &s->dsp.put_pixels_tab[size]; | |
122 } | |
123 c->temp= c->scratchpad; | |
936 | 124 } |
125 | |
1950 | 126 static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
127 const int size, const int h, int ref_index, int src_index, | |
128 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ | |
129 MotionEstContext * const c= &s->me; | |
130 const int stride= c->stride; | |
131 const int uvstride= c->uvstride; | |
132 const int qpel= flags&FLAG_QPEL; | |
133 const int chroma= flags&FLAG_CHROMA; | |
134 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel? | |
135 const int hx= subx + (x<<(1+qpel)); | |
136 const int hy= suby + (y<<(1+qpel)); | |
137 uint8_t * const * const ref= c->ref[ref_index]; | |
138 uint8_t * const * const src= c->src[src_index]; | |
139 int d; | |
140 //FIXME check chroma 4mv, (no crashes ...) | |
141 if(flags&FLAG_DIRECT){ | |
142 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){ | |
143 const int time_pp= s->pp_time; | |
144 const int time_pb= s->pb_time; | |
145 const int mask= 2*qpel+1; | |
146 if(s->mv_type==MV_TYPE_8X8){ | |
147 int i; | |
148 for(i=0; i<4; i++){ | |
149 int fx = c->direct_basis_mv[i][0] + hx; | |
150 int fy = c->direct_basis_mv[i][1] + hy; | |
151 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4)); | |
152 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4)); | |
153 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1)); | |
154 int bxy= (bx&mask) + ((by&mask)<<(qpel+1)); | |
155 | |
156 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1); | |
157 if(qpel){ | |
158 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride); | |
159 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride); | |
160 }else{ | |
161 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8); | |
162 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8); | |
163 } | |
164 } | |
165 }else{ | |
166 int fx = c->direct_basis_mv[0][0] + hx; | |
167 int fy = c->direct_basis_mv[0][1] + hy; | |
168 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp); | |
169 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp); | |
170 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1)); | |
171 int bxy= (bx&mask) + ((by&mask)<<(qpel+1)); | |
172 | |
173 if(qpel){ | |
174 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride); | |
175 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride); | |
176 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride); | |
177 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride); | |
178 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride); | |
179 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride); | |
180 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride); | |
181 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride); | |
182 }else{ | |
183 assert((fx>>1) + 16*s->mb_x >= -16); | |
184 assert((fy>>1) + 16*s->mb_y >= -16); | |
185 assert((fx>>1) + 16*s->mb_x <= s->width); | |
186 assert((fy>>1) + 16*s->mb_y <= s->height); | |
187 assert((bx>>1) + 16*s->mb_x >= -16); | |
188 assert((by>>1) + 16*s->mb_y >= -16); | |
189 assert((bx>>1) + 16*s->mb_x <= s->width); | |
190 assert((by>>1) + 16*s->mb_y <= s->height); | |
936 | 191 |
1950 | 192 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16); |
193 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16); | |
194 } | |
195 } | |
196 d = cmp_func(s, c->temp, src[0], stride, 16); | |
197 }else | |
198 d= 256*256*256*32; | |
199 }else{ | |
200 int uvdxy; | |
201 if(dxy){ | |
202 if(qpel){ | |
203 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h) | |
204 if(chroma){ | |
205 int cx= hx/2; | |
206 int cy= hy/2; | |
207 cx= (cx>>1)|(cx&1); | |
208 cy= (cy>>1)|(cy&1); | |
209 uvdxy= (cx&1) + 2*(cy&1); | |
210 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264 | |
211 } | |
212 }else{ | |
213 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h); | |
214 if(chroma) | |
215 uvdxy= dxy | (x&1) | (2*(y&1)); | |
216 } | |
217 d = cmp_func(s, c->temp, src[0], stride, h); | |
218 }else{ | |
219 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h); | |
220 if(chroma) | |
221 uvdxy= (x&1) + 2*(y&1); | |
222 } | |
223 if(chroma){ | |
224 uint8_t * const uvtemp= c->temp + 16*stride; | |
225 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1); | |
226 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1); | |
227 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1); | |
228 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1); | |
229 } | |
230 } | |
231 #if 0 | |
232 if(full_pel){ | |
233 const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1); | |
234 score_map[index]= d; | |
235 } | |
936 | 236 |
1950 | 237 d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor; |
238 #endif | |
239 return d; | |
936 | 240 } |
241 | |
242 #include "motion_est_template.c" | |
243 | |
244 static inline int get_penalty_factor(MpegEncContext *s, int type){ | |
1013 | 245 switch(type&0xFF){ |
936 | 246 default: |
247 case FF_CMP_SAD: | |
1013 | 248 return s->qscale*2; |
936 | 249 case FF_CMP_DCT: |
1013 | 250 return s->qscale*3; |
936 | 251 case FF_CMP_SATD: |
1013 | 252 return s->qscale*6; |
1007 | 253 case FF_CMP_SSE: |
1013 | 254 return s->qscale*s->qscale*2; |
1007 | 255 case FF_CMP_BIT: |
256 return 1; | |
257 case FF_CMP_RD: | |
1013 | 258 case FF_CMP_PSNR: |
259 return (s->qscale*s->qscale*185 + 64)>>7; | |
936 | 260 } |
261 } | |
262 | |
263 void ff_init_me(MpegEncContext *s){ | |
1729 | 264 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp); |
265 ff_set_cmp(&s->dsp, s->dsp.me_cmp, s->avctx->me_cmp); | |
266 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp); | |
267 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, s->avctx->mb_cmp); | |
1950 | 268 |
269 s->me.flags = get_flags(s, 0, s->avctx->me_cmp &FF_CMP_CHROMA); | |
270 s->me.sub_flags= get_flags(s, 0, s->avctx->me_sub_cmp&FF_CMP_CHROMA); | |
271 s->me.mb_flags = get_flags(s, 0, s->avctx->mb_cmp &FF_CMP_CHROMA); | |
936 | 272 |
273 if(s->flags&CODEC_FLAG_QPEL){ | |
1950 | 274 s->me.sub_motion_search= qpel_motion_search; |
936 | 275 }else{ |
276 if(s->avctx->me_sub_cmp&FF_CMP_CHROMA) | |
1950 | 277 s->me.sub_motion_search= hpel_motion_search; |
1013 | 278 else if( s->avctx->me_sub_cmp == FF_CMP_SAD |
279 && s->avctx-> me_cmp == FF_CMP_SAD | |
280 && s->avctx-> mb_cmp == FF_CMP_SAD) | |
1708 | 281 s->me.sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles |
936 | 282 else |
1950 | 283 s->me.sub_motion_search= hpel_motion_search; |
936 | 284 } |
1950 | 285 if(s->linesize){ |
286 s->me.stride = s->linesize; | |
287 s->me.uvstride= s->uvlinesize; | |
954 | 288 }else{ |
1950 | 289 s->me.stride = 16*s->mb_width + 32; |
290 s->me.uvstride= 8*s->mb_width + 16; | |
1013 | 291 } |
936 | 292 } |
293 | |
1419 | 294 #if 0 |
1064 | 295 static int pix_dev(uint8_t * pix, int line_size, int mean) |
284 | 296 { |
297 int s, i, j; | |
298 | |
299 s = 0; | |
300 for (i = 0; i < 16; i++) { | |
301 for (j = 0; j < 16; j += 8) { | |
302 s += ABS(pix[0]-mean); | |
303 s += ABS(pix[1]-mean); | |
304 s += ABS(pix[2]-mean); | |
305 s += ABS(pix[3]-mean); | |
306 s += ABS(pix[4]-mean); | |
307 s += ABS(pix[5]-mean); | |
308 s += ABS(pix[6]-mean); | |
309 s += ABS(pix[7]-mean); | |
310 pix += 8; | |
311 } | |
312 pix += line_size - 16; | |
313 } | |
314 return s; | |
315 } | |
1419 | 316 #endif |
284 | 317 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
318 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
|
319 int *mx_ptr, int *my_ptr) |
0 | 320 { |
321 *mx_ptr = 16 * s->mb_x; | |
322 *my_ptr = 16 * s->mb_y; | |
323 } | |
324 | |
325 static int full_motion_search(MpegEncContext * s, | |
326 int *mx_ptr, int *my_ptr, int range, | |
324 | 327 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 328 { |
329 int x1, y1, x2, y2, xx, yy, x, y; | |
330 int mx, my, dmin, d; | |
1064 | 331 uint8_t *pix; |
0 | 332 |
333 xx = 16 * s->mb_x; | |
334 yy = 16 * s->mb_y; | |
335 x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */ | |
336 if (x1 < xmin) | |
337 x1 = xmin; | |
338 x2 = xx + range - 1; | |
339 if (x2 > xmax) | |
340 x2 = xmax; | |
341 y1 = yy - range + 1; | |
342 if (y1 < ymin) | |
343 y1 = ymin; | |
344 y2 = yy + range - 1; | |
345 if (y2 > ymax) | |
346 y2 = ymax; | |
903 | 347 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 348 dmin = 0x7fffffff; |
349 mx = 0; | |
350 my = 0; | |
351 for (y = y1; y <= y2; y++) { | |
352 for (x = x1; x <= x2; x++) { | |
1708 | 353 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, |
354 s->linesize, 16); | |
0 | 355 if (d < dmin || |
356 (d == dmin && | |
357 (abs(x - xx) + abs(y - yy)) < | |
358 (abs(mx - xx) + abs(my - yy)))) { | |
359 dmin = d; | |
360 mx = x; | |
361 my = y; | |
362 } | |
363 } | |
364 } | |
365 | |
366 *mx_ptr = mx; | |
367 *my_ptr = my; | |
368 | |
369 #if 0 | |
370 if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) || | |
371 *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) { | |
372 fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr); | |
373 } | |
374 #endif | |
375 return dmin; | |
376 } | |
377 | |
378 | |
379 static int log_motion_search(MpegEncContext * s, | |
380 int *mx_ptr, int *my_ptr, int range, | |
324 | 381 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 382 { |
383 int x1, y1, x2, y2, xx, yy, x, y; | |
384 int mx, my, dmin, d; | |
1064 | 385 uint8_t *pix; |
0 | 386 |
387 xx = s->mb_x << 4; | |
388 yy = s->mb_y << 4; | |
389 | |
390 /* Left limit */ | |
391 x1 = xx - range; | |
392 if (x1 < xmin) | |
393 x1 = xmin; | |
394 | |
395 /* Right limit */ | |
396 x2 = xx + range; | |
397 if (x2 > xmax) | |
398 x2 = xmax; | |
399 | |
400 /* Upper limit */ | |
401 y1 = yy - range; | |
402 if (y1 < ymin) | |
403 y1 = ymin; | |
404 | |
405 /* Lower limit */ | |
406 y2 = yy + range; | |
407 if (y2 > ymax) | |
408 y2 = ymax; | |
409 | |
903 | 410 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 411 dmin = 0x7fffffff; |
412 mx = 0; | |
413 my = 0; | |
414 | |
415 do { | |
416 for (y = y1; y <= y2; y += range) { | |
417 for (x = x1; x <= x2; x += range) { | |
1708 | 418 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16); |
0 | 419 if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
420 dmin = d; | |
421 mx = x; | |
422 my = y; | |
423 } | |
424 } | |
425 } | |
426 | |
427 range = range >> 1; | |
428 | |
429 x1 = mx - range; | |
430 if (x1 < xmin) | |
431 x1 = xmin; | |
432 | |
433 x2 = mx + range; | |
434 if (x2 > xmax) | |
435 x2 = xmax; | |
436 | |
437 y1 = my - range; | |
438 if (y1 < ymin) | |
439 y1 = ymin; | |
440 | |
441 y2 = my + range; | |
442 if (y2 > ymax) | |
443 y2 = ymax; | |
444 | |
445 } while (range >= 1); | |
446 | |
447 #ifdef DEBUG | |
448 fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my); | |
449 #endif | |
450 *mx_ptr = mx; | |
451 *my_ptr = my; | |
452 return dmin; | |
453 } | |
454 | |
455 static int phods_motion_search(MpegEncContext * s, | |
456 int *mx_ptr, int *my_ptr, int range, | |
324 | 457 int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture) |
0 | 458 { |
459 int x1, y1, x2, y2, xx, yy, x, y, lastx, d; | |
460 int mx, my, dminx, dminy; | |
1064 | 461 uint8_t *pix; |
0 | 462 |
463 xx = s->mb_x << 4; | |
464 yy = s->mb_y << 4; | |
465 | |
466 /* Left limit */ | |
467 x1 = xx - range; | |
468 if (x1 < xmin) | |
469 x1 = xmin; | |
470 | |
471 /* Right limit */ | |
472 x2 = xx + range; | |
473 if (x2 > xmax) | |
474 x2 = xmax; | |
475 | |
476 /* Upper limit */ | |
477 y1 = yy - range; | |
478 if (y1 < ymin) | |
479 y1 = ymin; | |
480 | |
481 /* Lower limit */ | |
482 y2 = yy + range; | |
483 if (y2 > ymax) | |
484 y2 = ymax; | |
485 | |
903 | 486 pix = s->new_picture.data[0] + (yy * s->linesize) + xx; |
0 | 487 mx = 0; |
488 my = 0; | |
489 | |
490 x = xx; | |
491 y = yy; | |
492 do { | |
493 dminx = 0x7fffffff; | |
494 dminy = 0x7fffffff; | |
495 | |
496 lastx = x; | |
497 for (x = x1; x <= x2; x += range) { | |
1708 | 498 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16); |
0 | 499 if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
500 dminx = d; | |
501 mx = x; | |
502 } | |
503 } | |
504 | |
505 x = lastx; | |
506 for (y = y1; y <= y2; y += range) { | |
1708 | 507 d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16); |
0 | 508 if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) { |
509 dminy = d; | |
510 my = y; | |
511 } | |
512 } | |
513 | |
514 range = range >> 1; | |
515 | |
516 x = mx; | |
517 y = my; | |
518 x1 = mx - range; | |
519 if (x1 < xmin) | |
520 x1 = xmin; | |
521 | |
522 x2 = mx + range; | |
523 if (x2 > xmax) | |
524 x2 = xmax; | |
525 | |
526 y1 = my - range; | |
527 if (y1 < ymin) | |
528 y1 = ymin; | |
529 | |
530 y2 = my + range; | |
531 if (y2 > ymax) | |
532 y2 = ymax; | |
533 | |
534 } while (range >= 1); | |
535 | |
536 #ifdef DEBUG | |
537 fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my); | |
538 #endif | |
539 | |
540 /* half pixel search */ | |
541 *mx_ptr = mx; | |
542 *my_ptr = my; | |
543 return dminy; | |
544 } | |
545 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
546 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
547 #define Z_THRESHOLD 256 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
548 |
936 | 549 #define CHECK_SAD_HALF_MV(suffix, x, y) \ |
455 | 550 {\ |
1708 | 551 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\ |
936 | 552 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\ |
455 | 553 COPY3_IF_LT(dminh, d, dx, x, dy, y)\ |
554 } | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
555 |
936 | 556 static inline int sad_hpel_motion_search(MpegEncContext * s, |
0 | 557 int *mx_ptr, int *my_ptr, int dmin, |
1950 | 558 int src_index, int ref_index, |
559 int size, int h) | |
0 | 560 { |
936 | 561 const int penalty_factor= s->me.sub_penalty_factor; |
1708 | 562 int mx, my, dminh; |
1064 | 563 uint8_t *pix, *ptr; |
1950 | 564 int stride= s->me.stride; |
565 const int flags= s->me.sub_flags; | |
566 LOAD_COMMON | |
567 | |
568 assert(flags == 0); | |
455 | 569 |
936 | 570 if(s->me.skip){ |
455 | 571 // printf("S"); |
572 *mx_ptr = 0; | |
573 *my_ptr = 0; | |
574 return dmin; | |
575 } | |
576 // printf("N"); | |
577 | |
1950 | 578 pix = s->me.src[src_index][0]; |
455 | 579 |
294 | 580 mx = *mx_ptr; |
581 my = *my_ptr; | |
1950 | 582 ptr = s->me.ref[ref_index][0] + (my * stride) + mx; |
455 | 583 |
294 | 584 dminh = dmin; |
585 | |
586 if (mx > xmin && mx < xmax && | |
587 my > ymin && my < ymax) { | |
455 | 588 int dx=0, dy=0; |
589 int d, pen_x, pen_y; | |
590 const int index= (my<<ME_MAP_SHIFT) + mx; | |
591 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
592 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; | |
593 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; | |
594 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; | |
595 mx<<=1; | |
596 my<<=1; | |
294 | 597 |
598 | |
599 pen_x= pred_x + mx; | |
600 pen_y= pred_y + my; | |
601 | |
1708 | 602 ptr-= stride; |
455 | 603 if(t<=b){ |
936 | 604 CHECK_SAD_HALF_MV(y2 , 0, -1) |
455 | 605 if(l<=r){ |
936 | 606 CHECK_SAD_HALF_MV(xy2, -1, -1) |
455 | 607 if(t+r<=b+l){ |
936 | 608 CHECK_SAD_HALF_MV(xy2, +1, -1) |
1708 | 609 ptr+= stride; |
455 | 610 }else{ |
1708 | 611 ptr+= stride; |
936 | 612 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 613 } |
936 | 614 CHECK_SAD_HALF_MV(x2 , -1, 0) |
455 | 615 }else{ |
936 | 616 CHECK_SAD_HALF_MV(xy2, +1, -1) |
455 | 617 if(t+l<=b+r){ |
936 | 618 CHECK_SAD_HALF_MV(xy2, -1, -1) |
1708 | 619 ptr+= stride; |
455 | 620 }else{ |
1708 | 621 ptr+= stride; |
936 | 622 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 623 } |
936 | 624 CHECK_SAD_HALF_MV(x2 , +1, 0) |
455 | 625 } |
626 }else{ | |
627 if(l<=r){ | |
628 if(t+l<=b+r){ | |
936 | 629 CHECK_SAD_HALF_MV(xy2, -1, -1) |
1708 | 630 ptr+= stride; |
455 | 631 }else{ |
1708 | 632 ptr+= stride; |
936 | 633 CHECK_SAD_HALF_MV(xy2, +1, +1) |
455 | 634 } |
936 | 635 CHECK_SAD_HALF_MV(x2 , -1, 0) |
636 CHECK_SAD_HALF_MV(xy2, -1, +1) | |
455 | 637 }else{ |
638 if(t+r<=b+l){ | |
936 | 639 CHECK_SAD_HALF_MV(xy2, +1, -1) |
1708 | 640 ptr+= stride; |
455 | 641 }else{ |
1708 | 642 ptr+= stride; |
936 | 643 CHECK_SAD_HALF_MV(xy2, -1, +1) |
455 | 644 } |
936 | 645 CHECK_SAD_HALF_MV(x2 , +1, 0) |
646 CHECK_SAD_HALF_MV(xy2, +1, +1) | |
455 | 647 } |
936 | 648 CHECK_SAD_HALF_MV(y2 , 0, +1) |
455 | 649 } |
650 mx+=dx; | |
651 my+=dy; | |
294 | 652 |
653 }else{ | |
455 | 654 mx<<=1; |
655 my<<=1; | |
294 | 656 } |
657 | |
658 *mx_ptr = mx; | |
659 *my_ptr = my; | |
455 | 660 return dminh; |
294 | 661 } |
662 | |
455 | 663 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4) |
294 | 664 { |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
665 const int xy= s->mb_x + s->mb_y*s->mb_stride; |
294 | 666 |
324 | 667 s->p_mv_table[xy][0] = mx; |
668 s->p_mv_table[xy][1] = my; | |
294 | 669 |
670 /* has allready been set to the 4 MV if 4MV is done */ | |
455 | 671 if(mv4){ |
294 | 672 int mot_xy= s->block_index[0]; |
673 | |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
674 s->current_picture.motion_val[0][mot_xy ][0]= mx; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
675 s->current_picture.motion_val[0][mot_xy ][1]= my; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
676 s->current_picture.motion_val[0][mot_xy+1][0]= mx; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
677 s->current_picture.motion_val[0][mot_xy+1][1]= my; |
294 | 678 |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
679 mot_xy += s->b8_stride; |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
680 s->current_picture.motion_val[0][mot_xy ][0]= mx; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
681 s->current_picture.motion_val[0][mot_xy ][1]= my; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
682 s->current_picture.motion_val[0][mot_xy+1][0]= mx; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
683 s->current_picture.motion_val[0][mot_xy+1][1]= my; |
294 | 684 } |
685 } | |
686 | |
1086 | 687 /** |
688 * get fullpel ME search limits. | |
689 */ | |
1708 | 690 static inline void get_limits(MpegEncContext *s, int x, int y) |
324 | 691 { |
1708 | 692 /* |
693 if(s->avctx->me_range) s->me.range= s->avctx->me_range >> 1; | |
694 else s->me.range= 16; | |
695 */ | |
324 | 696 if (s->unrestricted_mv) { |
1708 | 697 s->me.xmin = - x - 16; |
698 s->me.ymin = - y - 16; | |
699 s->me.xmax = - x + s->mb_width *16; | |
700 s->me.ymax = - y + s->mb_height*16; | |
324 | 701 } else { |
1708 | 702 s->me.xmin = - x; |
703 s->me.ymin = - y; | |
704 s->me.xmax = - x + s->mb_width *16 - 16; | |
705 s->me.ymax = - y + s->mb_height*16 - 16; | |
324 | 706 } |
707 } | |
708 | |
1708 | 709 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift) |
455 | 710 { |
1950 | 711 MotionEstContext * const c= &s->me; |
1708 | 712 const int size= 1; |
713 const int h=8; | |
455 | 714 int block; |
715 int P[10][2]; | |
1013 | 716 int dmin_sum=0, mx4_sum=0, my4_sum=0; |
1494
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
717 int same=1; |
1708 | 718 const int stride= s->linesize; |
719 const int uvstride= s->uvlinesize; | |
1950 | 720 uint8_t *mv_penalty= s->me.current_mv_penalty; |
455 | 721 |
1950 | 722 c->ref[1][0] = c->ref[0][0] + 8; |
723 c->ref[2][0] = c->ref[0][0] + 8*stride; | |
724 c->ref[3][0] = c->ref[2][0] + 8; | |
725 c->src[1][0] = c->src[0][0] + 8; | |
726 c->src[2][0] = c->src[0][0] + 8*stride; | |
727 c->src[3][0] = c->src[2][0] + 8; | |
728 init_mc(s, size, c->flags); | |
729 | |
455 | 730 for(block=0; block<4; block++){ |
731 int mx4, my4; | |
732 int pred_x4, pred_y4; | |
733 int dmin4; | |
734 static const int off[4]= {2, 1, 1, -1}; | |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
735 const int mot_stride = s->b8_stride; |
455 | 736 const int mot_xy = s->block_index[block]; |
1708 | 737 |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
738 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
739 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1]; |
455 | 740 |
1708 | 741 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift); |
455 | 742 |
743 /* special case for first line */ | |
1799 | 744 if (s->first_slice_line && block<2) { |
1950 | 745 s->me.pred_x= pred_x4= P_LEFT[0]; |
746 s->me.pred_y= pred_y4= P_LEFT[1]; | |
455 | 747 } else { |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
748 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
749 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
750 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
751 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1]; |
1708 | 752 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift); |
753 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift); | |
754 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift); | |
755 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift); | |
455 | 756 |
757 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
758 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
759 | |
1950 | 760 s->me.pred_x= pred_x4 = P_MEDIAN[0]; |
761 s->me.pred_y= pred_y4 = P_MEDIAN[1]; | |
455 | 762 } |
763 P_MV1[0]= mx; | |
764 P_MV1[1]= my; | |
765 | |
1950 | 766 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift); |
455 | 767 |
1950 | 768 dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h); |
1038 | 769 |
1950 | 770 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){ |
1013 | 771 int dxy; |
1708 | 772 const int offset= ((block&1) + (block>>1)*stride)*8; |
1013 | 773 uint8_t *dest_y = s->me.scratchpad + offset; |
774 if(s->quarter_sample){ | |
1950 | 775 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride; |
1013 | 776 dxy = ((my4 & 3) << 2) | (mx4 & 3); |
777 | |
778 if(s->no_rounding) | |
1799 | 779 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride); |
1013 | 780 else |
1708 | 781 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride); |
1013 | 782 }else{ |
1950 | 783 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride; |
1013 | 784 dxy = ((my4 & 1) << 1) | (mx4 & 1); |
785 | |
786 if(s->no_rounding) | |
1708 | 787 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h); |
1013 | 788 else |
1708 | 789 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h); |
1013 | 790 } |
791 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor; | |
792 }else | |
793 dmin_sum+= dmin4; | |
794 | |
795 if(s->quarter_sample){ | |
796 mx4_sum+= mx4/2; | |
797 my4_sum+= my4/2; | |
798 }else{ | |
799 mx4_sum+= mx4; | |
800 my4_sum+= my4; | |
801 } | |
802 | |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
803 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
804 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4; |
1494
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
805 |
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
806 if(mx4 != mx || my4 != my) same=0; |
1013 | 807 } |
808 | |
1494
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
809 if(same) |
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
810 return INT_MAX; |
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
811 |
1038 | 812 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){ |
1708 | 813 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, s->me.scratchpad, stride, 16); |
455 | 814 } |
1013 | 815 |
816 if(s->avctx->mb_cmp&FF_CMP_CHROMA){ | |
817 int dxy; | |
818 int mx, my; | |
819 int offset; | |
820 | |
821 mx= ff_h263_round_chroma(mx4_sum); | |
822 my= ff_h263_round_chroma(my4_sum); | |
823 dxy = ((my & 1) << 1) | (mx & 1); | |
824 | |
825 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize; | |
826 | |
827 if(s->no_rounding){ | |
828 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8); | |
829 s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8); | |
830 }else{ | |
831 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8); | |
832 s->dsp.put_pixels_tab [1][dxy](s->me.scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8); | |
833 } | |
834 | |
1708 | 835 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad , s->uvlinesize, 8); |
836 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad+8, s->uvlinesize, 8); | |
1013 | 837 } |
1950 | 838 |
839 s->me.pred_x= mx; | |
840 s->me.pred_y= my; | |
1013 | 841 |
842 switch(s->avctx->mb_cmp&0xFF){ | |
843 /*case FF_CMP_SSE: | |
844 return dmin_sum+ 32*s->qscale*s->qscale;*/ | |
845 case FF_CMP_RD: | |
846 return dmin_sum; | |
847 default: | |
848 return dmin_sum+ 11*s->me.mb_penalty_factor; | |
849 } | |
455 | 850 } |
851 | |
1950 | 852 static int interlaced_search(MpegEncContext *s, int ref_index, |
853 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my) | |
1708 | 854 { |
1950 | 855 MotionEstContext * const c= &s->me; |
1708 | 856 const int size=0; |
857 const int h=8; | |
858 int block; | |
859 int P[10][2]; | |
1950 | 860 uint8_t * const mv_penalty= c->current_mv_penalty; |
1708 | 861 int same=1; |
862 const int stride= 2*s->linesize; | |
863 const int uvstride= 2*s->uvlinesize; | |
864 int dmin_sum= 0; | |
865 const int mot_stride= s->mb_stride; | |
866 const int xy= s->mb_x + s->mb_y*mot_stride; | |
867 | |
1950 | 868 c->ymin>>=1; |
869 c->ymax>>=1; | |
870 c->stride<<=1; | |
871 c->uvstride<<=1; | |
872 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize; | |
873 c->src[1][0] = c->src[0][0] + s->linesize; | |
874 if(c->flags & FLAG_CHROMA){ | |
875 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize; | |
876 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize; | |
877 c->src[1][1] = c->src[0][1] + s->uvlinesize; | |
878 c->src[1][2] = c->src[0][2] + s->uvlinesize; | |
879 } | |
880 init_mc(s, size, c->flags); | |
1708 | 881 |
882 for(block=0; block<2; block++){ | |
883 int field_select; | |
884 int best_dmin= INT_MAX; | |
885 int best_field= -1; | |
886 | |
887 for(field_select=0; field_select<2; field_select++){ | |
1950 | 888 int dmin, mx_i, my_i; |
1708 | 889 int16_t (*mv_table)[2]= mv_tables[block][field_select]; |
890 | |
891 P_LEFT[0] = mv_table[xy - 1][0]; | |
892 P_LEFT[1] = mv_table[xy - 1][1]; | |
1950 | 893 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1); |
1708 | 894 |
1950 | 895 s->me.pred_x= P_LEFT[0]; |
896 s->me.pred_y= P_LEFT[1]; | |
1708 | 897 |
1799 | 898 if(!s->first_slice_line){ |
1708 | 899 P_TOP[0] = mv_table[xy - mot_stride][0]; |
900 P_TOP[1] = mv_table[xy - mot_stride][1]; | |
901 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0]; | |
902 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1]; | |
1950 | 903 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1); |
904 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1); | |
905 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1); | |
906 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1); | |
1708 | 907 |
908 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
909 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
910 } | |
911 P_MV1[0]= mx; //FIXME not correct if block != field_select | |
912 P_MV1[1]= my / 2; | |
913 | |
1950 | 914 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1); |
1708 | 915 |
1950 | 916 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h); |
1708 | 917 |
918 mv_table[xy][0]= mx_i; | |
919 mv_table[xy][1]= my_i; | |
920 | |
1950 | 921 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){ |
1708 | 922 int dxy; |
923 | |
924 //FIXME chroma ME | |
1950 | 925 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride; |
1708 | 926 dxy = ((my_i & 1) << 1) | (mx_i & 1); |
927 | |
928 if(s->no_rounding){ | |
1950 | 929 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h); |
1708 | 930 }else{ |
1950 | 931 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h); |
1708 | 932 } |
1950 | 933 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h); |
934 dmin+= (mv_penalty[mx_i-s->me.pred_x] + mv_penalty[my_i-s->me.pred_y] + 1)*c->mb_penalty_factor; | |
1708 | 935 }else |
1950 | 936 dmin+= c->mb_penalty_factor; //field_select bits |
1708 | 937 |
938 dmin += field_select != block; //slightly prefer same field | |
939 | |
940 if(dmin < best_dmin){ | |
941 best_dmin= dmin; | |
942 best_field= field_select; | |
943 } | |
944 } | |
945 { | |
946 int16_t (*mv_table)[2]= mv_tables[block][best_field]; | |
947 | |
948 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all | |
949 if(mv_table[xy][1]&1) same=0; | |
950 if(mv_table[xy][1]*2 != my) same=0; | |
951 if(best_field != block) same=0; | |
952 } | |
953 | |
954 field_select_tables[block][xy]= best_field; | |
955 dmin_sum += best_dmin; | |
956 } | |
957 | |
1950 | 958 c->ymin<<=1; |
959 c->ymax<<=1; | |
960 c->stride>>=1; | |
961 c->uvstride>>=1; | |
1708 | 962 |
963 if(same) | |
964 return INT_MAX; | |
965 | |
966 switch(s->avctx->mb_cmp&0xFF){ | |
967 /*case FF_CMP_SSE: | |
968 return dmin_sum+ 32*s->qscale*s->qscale;*/ | |
969 case FF_CMP_RD: | |
970 return dmin_sum; | |
971 default: | |
1950 | 972 return dmin_sum+ 11*c->mb_penalty_factor; |
1708 | 973 } |
974 } | |
975 | |
324 | 976 void ff_estimate_p_frame_motion(MpegEncContext * s, |
977 int mb_x, int mb_y) | |
0 | 978 { |
1950 | 979 MotionEstContext * const c= &s->me; |
1064 | 980 uint8_t *pix, *ppix; |
1708 | 981 int sum, varc, vard, mx, my, dmin, xx, yy; |
455 | 982 int P[10][2]; |
280 | 983 const int shift= 1+s->quarter_sample; |
294 | 984 int mb_type=0; |
903 | 985 Picture * const pic= &s->current_picture; |
1950 | 986 |
987 init_ref(s, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0); | |
988 init_mc(s, 0, c->flags); | |
1708 | 989 |
936 | 990 assert(s->quarter_sample==0 || s->quarter_sample==1); |
1950 | 991 assert(s->linesize == s->me.stride); |
992 assert(s->uvlinesize == s->me.uvstride); | |
936 | 993 |
994 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
995 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
1013 | 996 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp); |
1950 | 997 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
0 | 998 |
1708 | 999 get_limits(s, 16*mb_x, 16*mb_y); |
936 | 1000 s->me.skip=0; |
324 | 1001 |
320
cda7d0857baf
- ME setting moved to AVCodecContext/MpegEncContext, no longer a global.
pulento
parents:
304
diff
changeset
|
1002 switch(s->me_method) { |
0 | 1003 case ME_ZERO: |
1004 default: | |
1005 no_motion_search(s, &mx, &my); | |
455 | 1006 mx-= mb_x*16; |
1007 my-= mb_y*16; | |
0 | 1008 dmin = 0; |
1009 break; | |
1708 | 1010 #if 0 |
0 | 1011 case ME_FULL: |
1708 | 1012 dmin = full_motion_search(s, &mx, &my, range, ref_picture); |
455 | 1013 mx-= mb_x*16; |
1014 my-= mb_y*16; | |
0 | 1015 break; |
1016 case ME_LOG: | |
1708 | 1017 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture); |
455 | 1018 mx-= mb_x*16; |
1019 my-= mb_y*16; | |
0 | 1020 break; |
1021 case ME_PHODS: | |
1708 | 1022 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture); |
455 | 1023 mx-= mb_x*16; |
1024 my-= mb_y*16; | |
0 | 1025 break; |
1708 | 1026 #endif |
288 | 1027 case ME_X1: |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
1028 case ME_EPZS: |
288 | 1029 { |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
1030 const int mot_stride = s->b8_stride; |
294 | 1031 const int mot_xy = s->block_index[0]; |
288 | 1032 |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1033 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1034 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1]; |
288 | 1035 |
1708 | 1036 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift); |
280 | 1037 |
1799 | 1038 if(!s->first_slice_line) { |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1039 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1040 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1041 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1042 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1]; |
1708 | 1043 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1] = (s->me.ymax<<shift); |
1044 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift); | |
1045 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift); | |
280 | 1046 |
455 | 1047 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1048 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1049 | |
1050 if(s->out_format == FMT_H263){ | |
1950 | 1051 c->pred_x = P_MEDIAN[0]; |
1052 c->pred_y = P_MEDIAN[1]; | |
455 | 1053 }else { /* mpeg1 at least */ |
1950 | 1054 c->pred_x= P_LEFT[0]; |
1055 c->pred_y= P_LEFT[1]; | |
455 | 1056 } |
952 | 1057 }else{ |
1950 | 1058 c->pred_x= P_LEFT[0]; |
1059 c->pred_y= P_LEFT[1]; | |
288 | 1060 } |
952 | 1061 |
280 | 1062 } |
1950 | 1063 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift); |
1064 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
275
diff
changeset
|
1065 break; |
0 | 1066 } |
1067 | |
1068 /* intra / predictive decision */ | |
1069 xx = mb_x * 16; | |
1070 yy = mb_y * 16; | |
1071 | |
1950 | 1072 pix = c->src[0][0]; |
455 | 1073 /* At this point (mx,my) are full-pell and the relative displacement */ |
1950 | 1074 ppix = c->ref[0][0] + (my * s->linesize) + mx; |
289
648e9245546d
seems the old intra/inter decission is slightly better with a threshold, than the new one
michaelni
parents:
288
diff
changeset
|
1075 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1076 sum = s->dsp.pix_sum(pix, s->linesize); |
455 | 1077 |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1078 varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8; |
1708 | 1079 vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16)+128)>>8; |
608 | 1080 |
455 | 1081 //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); |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1082 pic->mb_var [s->mb_stride * mb_y + mb_x] = varc; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1083 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1084 pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1085 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; |
1799 | 1086 s->mb_var_sum_temp += varc; |
1087 s->mc_mb_var_sum_temp += vard; | |
455 | 1088 //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
|
1089 |
0 | 1090 #if 0 |
233
3f5b72726118
- More work on preliminary bit rate control, just to be able to get an
pulento
parents:
232
diff
changeset
|
1091 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
|
1092 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy); |
0 | 1093 #endif |
1389 | 1094 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){ |
608 | 1095 if (vard <= 64 || vard < varc) |
1096 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); | |
1097 else | |
913 | 1098 s->scene_change_score+= s->qscale; |
608 | 1099 |
294 | 1100 if (vard*2 + 200 > varc) |
1708 | 1101 mb_type|= CANDIDATE_MB_TYPE_INTRA; |
294 | 1102 if (varc*2 + 200 > vard){ |
1708 | 1103 mb_type|= CANDIDATE_MB_TYPE_INTER; |
1950 | 1104 s->me.sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
1494
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
1105 if(s->flags&CODEC_FLAG_MV0) |
3ee63c12ea30
optionally try to encode each MB with MV=<0,0> and choose the one with better RD
michaelni
parents:
1426
diff
changeset
|
1106 if(mx || my) |
1708 | 1107 mb_type |= CANDIDATE_MB_TYPE_SKIPED; //FIXME check difference |
304 | 1108 }else{ |
936 | 1109 mx <<=shift; |
1110 my <<=shift; | |
0 | 1111 } |
455 | 1112 if((s->flags&CODEC_FLAG_4MV) |
936 | 1113 && !s->me.skip && varc>50 && vard>10){ |
1708 | 1114 if(h263_mv4_search(s, mx, my, shift) < INT_MAX) |
1115 mb_type|=CANDIDATE_MB_TYPE_INTER4V; | |
455 | 1116 |
1117 set_p_mv_tables(s, mx, my, 0); | |
1118 }else | |
1119 set_p_mv_tables(s, mx, my, 1); | |
1708 | 1120 if((s->flags&CODEC_FLAG_INTERLACED_ME) |
1121 && !s->me.skip){ //FIXME varc/d checks | |
1950 | 1122 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my) < INT_MAX) |
1708 | 1123 mb_type |= CANDIDATE_MB_TYPE_INTER_I; |
1124 } | |
294 | 1125 }else{ |
1633 | 1126 int intra_score, i; |
1708 | 1127 mb_type= CANDIDATE_MB_TYPE_INTER; |
1013 | 1128 |
1950 | 1129 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
1013 | 1130 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip) |
1950 | 1131 dmin= get_mb_score(s, mx, my, 0, 0); |
1013 | 1132 |
1133 if((s->flags&CODEC_FLAG_4MV) | |
1134 && !s->me.skip && varc>50 && vard>10){ | |
1708 | 1135 int dmin4= h263_mv4_search(s, mx, my, shift); |
1013 | 1136 if(dmin4 < dmin){ |
1708 | 1137 mb_type= CANDIDATE_MB_TYPE_INTER4V; |
1013 | 1138 dmin=dmin4; |
1139 } | |
1140 } | |
1708 | 1141 if((s->flags&CODEC_FLAG_INTERLACED_ME) |
1142 && !s->me.skip){ //FIXME varc/d checks | |
1950 | 1143 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my); |
1708 | 1144 if(dmin_i < dmin){ |
1145 mb_type = CANDIDATE_MB_TYPE_INTER_I; | |
1146 dmin= dmin_i; | |
1147 } | |
1148 } | |
1633 | 1149 |
1150 // pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; | |
1708 | 1151 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V); |
1633 | 1152 |
1153 /* get intra luma score */ | |
1154 if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){ | |
1155 intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it | |
1156 }else{ | |
1157 int mean= (sum+128)>>8; | |
1158 mean*= 0x01010101; | |
1159 | |
1160 for(i=0; i<16; i++){ | |
1161 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean; | |
1162 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean; | |
1163 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean; | |
1164 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean; | |
1165 } | |
1166 | |
1708 | 1167 intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize, 16); |
1633 | 1168 } |
1169 #if 0 //FIXME | |
1170 /* get chroma score */ | |
1171 if(s->avctx->mb_cmp&FF_CMP_CHROMA){ | |
1172 for(i=1; i<3; i++){ | |
1173 uint8_t *dest_c; | |
1174 int mean; | |
1175 | |
1176 if(s->out_format == FMT_H263){ | |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
1177 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;) |
1633 | 1178 }else{ |
1179 mean= (s->last_dc[i] + 4)>>3; | |
1180 } | |
1181 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8; | |
1182 | |
1183 mean*= 0x01010101; | |
1184 for(i=0; i<8; i++){ | |
1185 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean; | |
1186 *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean; | |
1187 } | |
1188 | |
1189 intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize); | |
1190 } | |
1191 } | |
1192 #endif | |
1193 intra_score += s->me.mb_penalty_factor*16; | |
1013 | 1194 |
1633 | 1195 if(intra_score < dmin){ |
1708 | 1196 mb_type= CANDIDATE_MB_TYPE_INTRA; |
1197 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup | |
1633 | 1198 }else |
1199 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0; | |
1200 | |
1201 if (vard <= 64 || vard < varc) { //FIXME | |
608 | 1202 s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc); |
294 | 1203 }else{ |
1011 | 1204 s->scene_change_score+= s->qscale; |
294 | 1205 } |
1206 } | |
284 | 1207 |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1208 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type; |
0 | 1209 } |
1210 | |
951 | 1211 int ff_pre_estimate_p_frame_motion(MpegEncContext * s, |
1212 int mb_x, int mb_y) | |
1213 { | |
1950 | 1214 MotionEstContext * const c= &s->me; |
1708 | 1215 int mx, my, dmin; |
951 | 1216 int P[10][2]; |
1217 const int shift= 1+s->quarter_sample; | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1218 const int xy= mb_x + mb_y*s->mb_stride; |
1950 | 1219 init_ref(s, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0); |
1220 init_mc(s, 0, c->flags); | |
951 | 1221 |
1222 assert(s->quarter_sample==0 || s->quarter_sample==1); | |
1223 | |
954 | 1224 s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp); |
1950 | 1225 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
951 | 1226 |
1708 | 1227 get_limits(s, 16*mb_x, 16*mb_y); |
951 | 1228 s->me.skip=0; |
1229 | |
1230 P_LEFT[0] = s->p_mv_table[xy + 1][0]; | |
1231 P_LEFT[1] = s->p_mv_table[xy + 1][1]; | |
1232 | |
1708 | 1233 if(P_LEFT[0] < (s->me.xmin<<shift)) P_LEFT[0] = (s->me.xmin<<shift); |
951 | 1234 |
1235 /* special case for first line */ | |
1799 | 1236 if (s->first_slice_line) { |
1950 | 1237 c->pred_x= P_LEFT[0]; |
1238 c->pred_y= P_LEFT[1]; | |
952 | 1239 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]= |
1240 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME | |
951 | 1241 } else { |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1242 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0]; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1243 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1]; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1244 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0]; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1245 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1]; |
1708 | 1246 if(P_TOP[1] < (s->me.ymin<<shift)) P_TOP[1] = (s->me.ymin<<shift); |
1247 if(P_TOPRIGHT[0] > (s->me.xmax<<shift)) P_TOPRIGHT[0]= (s->me.xmax<<shift); | |
1248 if(P_TOPRIGHT[1] < (s->me.ymin<<shift)) P_TOPRIGHT[1]= (s->me.ymin<<shift); | |
951 | 1249 |
1250 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1251 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1252 | |
1950 | 1253 c->pred_x = P_MEDIAN[0]; |
1254 c->pred_y = P_MEDIAN[1]; | |
951 | 1255 } |
1950 | 1256 |
1257 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift); | |
952 | 1258 |
951 | 1259 s->p_mv_table[xy][0] = mx<<shift; |
1260 s->p_mv_table[xy][1] = my<<shift; | |
1261 | |
1262 return dmin; | |
1263 } | |
1264 | |
1057 | 1265 static int ff_estimate_motion_b(MpegEncContext * s, |
1950 | 1266 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code) |
0 | 1267 { |
1708 | 1268 int mx, my, dmin; |
455 | 1269 int P[10][2]; |
324 | 1270 const int shift= 1+s->quarter_sample; |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1271 const int mot_stride = s->mb_stride; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1272 const int mot_xy = mb_y*mot_stride + mb_x; |
1162 | 1273 uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV; |
948 | 1274 int mv_scale; |
936 | 1275 |
1276 s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp); | |
1277 s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp); | |
1013 | 1278 s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp); |
1950 | 1279 s->me.current_mv_penalty= mv_penalty; |
936 | 1280 |
1708 | 1281 get_limits(s, 16*mb_x, 16*mb_y); |
324 | 1282 |
1283 switch(s->me_method) { | |
1284 case ME_ZERO: | |
1285 default: | |
1286 no_motion_search(s, &mx, &my); | |
1287 dmin = 0; | |
455 | 1288 mx-= mb_x*16; |
1289 my-= mb_y*16; | |
324 | 1290 break; |
1708 | 1291 #if 0 |
324 | 1292 case ME_FULL: |
1708 | 1293 dmin = full_motion_search(s, &mx, &my, range, ref_picture); |
455 | 1294 mx-= mb_x*16; |
1295 my-= mb_y*16; | |
324 | 1296 break; |
1297 case ME_LOG: | |
1708 | 1298 dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture); |
455 | 1299 mx-= mb_x*16; |
1300 my-= mb_y*16; | |
324 | 1301 break; |
1302 case ME_PHODS: | |
1708 | 1303 dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture); |
455 | 1304 mx-= mb_x*16; |
1305 my-= mb_y*16; | |
324 | 1306 break; |
1708 | 1307 #endif |
324 | 1308 case ME_X1: |
1309 case ME_EPZS: | |
1310 { | |
455 | 1311 P_LEFT[0] = mv_table[mot_xy - 1][0]; |
1312 P_LEFT[1] = mv_table[mot_xy - 1][1]; | |
324 | 1313 |
1708 | 1314 if(P_LEFT[0] > (s->me.xmax<<shift)) P_LEFT[0] = (s->me.xmax<<shift); |
324 | 1315 |
1316 /* special case for first line */ | |
1799 | 1317 if (!s->first_slice_line) { |
455 | 1318 P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
1319 P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; | |
1320 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0]; | |
1321 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1]; | |
1708 | 1322 if(P_TOP[1] > (s->me.ymax<<shift)) P_TOP[1]= (s->me.ymax<<shift); |
1323 if(P_TOPRIGHT[0] < (s->me.xmin<<shift)) P_TOPRIGHT[0]= (s->me.xmin<<shift); | |
1324 if(P_TOPRIGHT[1] > (s->me.ymax<<shift)) P_TOPRIGHT[1]= (s->me.ymax<<shift); | |
324 | 1325 |
455 | 1326 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
1327 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
324 | 1328 } |
1950 | 1329 s->me.pred_x= P_LEFT[0]; |
1330 s->me.pred_y= P_LEFT[1]; | |
324 | 1331 } |
948 | 1332 |
1333 if(mv_table == s->b_forw_mv_table){ | |
1334 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift); | |
1335 }else{ | |
1336 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift); | |
1337 } | |
1338 | |
1950 | 1339 dmin = epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale); |
324 | 1340 |
1341 break; | |
1342 } | |
1343 | |
1950 | 1344 dmin= s->me.sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16); |
1013 | 1345 |
1346 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip) | |
1950 | 1347 dmin= get_mb_score(s, mx, my, 0, ref_index); |
1013 | 1348 |
455 | 1349 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my); |
324 | 1350 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
1351 mv_table[mot_xy][0]= mx; | |
1352 mv_table[mot_xy][1]= my; | |
936 | 1353 |
327 | 1354 return dmin; |
324 | 1355 } |
1356 | |
1950 | 1357 static inline int check_bidir_mv(MpegEncContext * s, |
327 | 1358 int motion_fx, int motion_fy, |
1359 int motion_bx, int motion_by, | |
1360 int pred_fx, int pred_fy, | |
1708 | 1361 int pred_bx, int pred_by, |
1362 int size, int h) | |
327 | 1363 { |
1364 //FIXME optimize? | |
936 | 1365 //FIXME better f_code prediction (max mv & distance) |
1708 | 1366 //FIXME pointers |
1950 | 1367 MotionEstContext * const c= &s->me; |
1162 | 1368 uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame |
1950 | 1369 int stride= s->me.stride; |
1370 int uvstride= s->me.uvstride; | |
936 | 1371 uint8_t *dest_y = s->me.scratchpad; |
327 | 1372 uint8_t *ptr; |
1373 int dxy; | |
1374 int src_x, src_y; | |
1375 int fbmin; | |
1950 | 1376 uint8_t **src_data= c->src[0]; |
1377 uint8_t **ref_data= c->ref[0]; | |
1378 uint8_t **ref2_data= c->ref[2]; | |
327 | 1379 |
936 | 1380 if(s->quarter_sample){ |
1381 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3); | |
1708 | 1382 src_x = motion_fx >> 2; |
1383 src_y = motion_fy >> 2; | |
327 | 1384 |
1708 | 1385 ptr = ref_data[0] + (src_y * stride) + src_x; |
1386 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride); | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1387 |
936 | 1388 dxy = ((motion_by & 3) << 2) | (motion_bx & 3); |
1708 | 1389 src_x = motion_bx >> 2; |
1390 src_y = motion_by >> 2; | |
936 | 1391 |
1950 | 1392 ptr = ref2_data[0] + (src_y * stride) + src_x; |
1708 | 1393 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride); |
936 | 1394 }else{ |
1395 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1); | |
1708 | 1396 src_x = motion_fx >> 1; |
1397 src_y = motion_fy >> 1; | |
327 | 1398 |
1708 | 1399 ptr = ref_data[0] + (src_y * stride) + src_x; |
1400 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h); | |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1401 |
936 | 1402 dxy = ((motion_by & 1) << 1) | (motion_bx & 1); |
1708 | 1403 src_x = motion_bx >> 1; |
1404 src_y = motion_by >> 1; | |
936 | 1405 |
1950 | 1406 ptr = ref2_data[0] + (src_y * stride) + src_x; |
1708 | 1407 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h); |
936 | 1408 } |
853
eacc2dd8fd9d
* using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents:
847
diff
changeset
|
1409 |
1013 | 1410 fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor |
1411 +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor | |
1708 | 1412 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic |
1013 | 1413 |
1414 if(s->avctx->mb_cmp&FF_CMP_CHROMA){ | |
1415 } | |
1416 //FIXME CHROMA !!! | |
1417 | |
327 | 1418 return fbmin; |
1419 } | |
1420 | |
1421 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/ | |
1950 | 1422 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y) |
327 | 1423 { |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1424 const int mot_stride = s->mb_stride; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1425 const int xy = mb_y *mot_stride + mb_x; |
327 | 1426 int fbmin; |
1427 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0]; | |
1428 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1]; | |
1429 int pred_bx= s->b_bidir_back_mv_table[xy-1][0]; | |
1430 int pred_by= s->b_bidir_back_mv_table[xy-1][1]; | |
1431 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0]; | |
1432 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1]; | |
1433 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0]; | |
1434 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1]; | |
1435 | |
1436 //FIXME do refinement and add flag | |
1437 | |
1950 | 1438 fbmin= check_bidir_mv(s, motion_fx, motion_fy, |
327 | 1439 motion_bx, motion_by, |
1440 pred_fx, pred_fy, | |
1708 | 1441 pred_bx, pred_by, |
1442 0, 16); | |
327 | 1443 |
1444 return fbmin; | |
1445 } | |
1446 | |
1950 | 1447 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y) |
324 | 1448 { |
455 | 1449 int P[10][2]; |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1450 const int mot_stride = s->mb_stride; |
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1451 const int mot_xy = mb_y*mot_stride + mb_x; |
936 | 1452 const int shift= 1+s->quarter_sample; |
1453 int dmin, i; | |
327 | 1454 const int time_pp= s->pp_time; |
664 | 1455 const int time_pb= s->pb_time; |
936 | 1456 int mx, my, xmin, xmax, ymin, ymax; |
327 | 1457 int16_t (*mv_table)[2]= s->b_direct_mv_table; |
936 | 1458 |
1950 | 1459 s->me.current_mv_penalty= s->me.mv_penalty[1] + MAX_MV; |
936 | 1460 ymin= xmin=(-32)>>shift; |
1461 ymax= xmax= 31>>shift; | |
1462 | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1463 if(IS_8X8(s->next_picture.mb_type[mot_xy])){ |
936 | 1464 s->mv_type= MV_TYPE_8X8; |
1465 }else{ | |
1466 s->mv_type= MV_TYPE_16X16; | |
327 | 1467 } |
936 | 1468 |
1469 for(i=0; i<4; i++){ | |
1470 int index= s->block_index[i]; | |
1471 int min, max; | |
1472 | |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1473 s->me.co_located_mv[i][0]= s->next_picture.motion_val[0][index][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1474 s->me.co_located_mv[i][1]= s->next_picture.motion_val[0][index][1]; |
936 | 1475 s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3)); |
1476 s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3)); | |
1477 // 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); | |
1478 // 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); | |
1479 | |
1480 max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1481 min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift; | |
1764 | 1482 max+= 16*mb_x + 1; // +-1 is for the simpler rounding |
1483 min+= 16*mb_x - 1; | |
960 | 1484 xmax= FFMIN(xmax, s->width - max); |
1485 xmin= FFMAX(xmin, - 16 - min); | |
936 | 1486 |
1487 max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1488 min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift; | |
1764 | 1489 max+= 16*mb_y + 1; // +-1 is for the simpler rounding |
1490 min+= 16*mb_y - 1; | |
960 | 1491 ymax= FFMIN(ymax, s->height - max); |
1492 ymin= FFMAX(ymin, - 16 - min); | |
936 | 1493 |
1494 if(s->mv_type == MV_TYPE_16X16) break; | |
327 | 1495 } |
936 | 1496 |
1497 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16); | |
1498 | |
1499 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){ | |
1500 s->b_direct_mv_table[mot_xy][0]= 0; | |
1501 s->b_direct_mv_table[mot_xy][1]= 0; | |
1502 | |
1503 return 256*256*256*64; | |
1504 } | |
1708 | 1505 |
1506 s->me.xmin= xmin; | |
1507 s->me.ymin= ymin; | |
1508 s->me.xmax= xmax; | |
1509 s->me.ymax= ymax; | |
1950 | 1510 s->me.flags |= FLAG_DIRECT; |
1511 s->me.sub_flags |= FLAG_DIRECT; | |
1512 s->me.pred_x=0; | |
1513 s->me.pred_y=0; | |
936 | 1514 |
950 | 1515 P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift); |
1516 P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift); | |
1517 | |
1518 /* special case for first line */ | |
1799 | 1519 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped |
950 | 1520 P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift); |
1521 P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift); | |
1522 P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift); | |
1523 P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift); | |
1524 | |
1525 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); | |
1526 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); | |
1527 } | |
1013 | 1528 |
1950 | 1529 dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift)); |
1530 if(s->me.sub_flags&FLAG_QPEL) | |
1531 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); | |
1532 else | |
1533 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); | |
1534 | |
1535 if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip) | |
1536 dmin= get_mb_score(s, mx, my, 0, 0); | |
1708 | 1537 |
1538 get_limits(s, 16*mb_x, 16*mb_y); //restore s->me.?min/max, maybe not needed | |
327 | 1539 |
1540 s->b_direct_mv_table[mot_xy][0]= mx; | |
1541 s->b_direct_mv_table[mot_xy][1]= my; | |
1950 | 1542 s->me.flags &= ~FLAG_DIRECT; |
1543 s->me.sub_flags &= ~FLAG_DIRECT; | |
1544 | |
327 | 1545 return dmin; |
324 | 1546 } |
1547 | |
1548 void ff_estimate_b_frame_motion(MpegEncContext * s, | |
1549 int mb_x, int mb_y) | |
1550 { | |
1013 | 1551 const int penalty_factor= s->me.mb_penalty_factor; |
1708 | 1552 int fmin, bmin, dmin, fbmin, bimin, fimin; |
327 | 1553 int type=0; |
1950 | 1554 init_ref(s, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2); |
1555 init_mc(s, 0, s->me.flags); | |
327 | 1556 |
1426 | 1557 s->me.skip=0; |
1558 if (s->codec_id == CODEC_ID_MPEG4) | |
1950 | 1559 dmin= direct_search(s, mb_x, mb_y); |
1426 | 1560 else |
1561 dmin= INT_MAX; | |
1708 | 1562 //FIXME penalty stuff for non mpeg4 |
1426 | 1563 s->me.skip=0; |
1950 | 1564 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor; |
1426 | 1565 |
1566 s->me.skip=0; | |
1950 | 1567 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor; |
324 | 1568 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
327 | 1569 |
1426 | 1570 s->me.skip=0; |
1950 | 1571 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor; |
1013 | 1572 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin); |
1708 | 1573 |
1574 if(s->flags & CODEC_FLAG_INTERLACED_ME){ | |
1575 const int xy = mb_y*s->mb_stride + mb_x; | |
1576 | |
1577 //FIXME mb type penalty | |
1578 s->me.skip=0; | |
1950 | 1579 s->me.current_mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; |
1580 fimin= interlaced_search(s, 0, | |
1581 s->b_field_mv_table[0], s->b_field_select_table[0], | |
1708 | 1582 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
1950 | 1583 s->me.current_mv_penalty= s->me.mv_penalty[s->b_code] + MAX_MV; |
1584 bimin= interlaced_search(s, 2, | |
1585 s->b_field_mv_table[1], s->b_field_select_table[1], | |
1708 | 1586 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1]); |
1587 }else | |
1588 fimin= bimin= INT_MAX; | |
1589 | |
612 | 1590 { |
1051
e5a9dbf597d4
mpeg1 bframe encoding patch by (Rapha¸«³l LEGRAND) with some modifications by me
michaelni
parents:
1050
diff
changeset
|
1591 int score= fmin; |
1708 | 1592 type = CANDIDATE_MB_TYPE_FORWARD; |
327 | 1593 |
1426 | 1594 if (dmin <= score){ |
1051
e5a9dbf597d4
mpeg1 bframe encoding patch by (Rapha¸«³l LEGRAND) with some modifications by me
michaelni
parents:
1050
diff
changeset
|
1595 score = dmin; |
1708 | 1596 type = CANDIDATE_MB_TYPE_DIRECT; |
327 | 1597 } |
1598 if(bmin<score){ | |
1599 score=bmin; | |
1708 | 1600 type= CANDIDATE_MB_TYPE_BACKWARD; |
327 | 1601 } |
1602 if(fbmin<score){ | |
1603 score=fbmin; | |
1708 | 1604 type= CANDIDATE_MB_TYPE_BIDIR; |
1605 } | |
1606 if(fimin<score){ | |
1607 score=fimin; | |
1608 type= CANDIDATE_MB_TYPE_FORWARD_I; | |
1609 } | |
1610 if(bimin<score){ | |
1611 score=bimin; | |
1612 type= CANDIDATE_MB_TYPE_BACKWARD_I; | |
327 | 1613 } |
1013 | 1614 |
807
0e1d375c537f
fixing q>0.0 assert failure caused by overflow of variance for b frames
michaelni
parents:
765
diff
changeset
|
1615 score= ((unsigned)(score*score + 128*256))>>16; |
1799 | 1616 s->mc_mb_var_sum_temp += score; |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1617 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE |
327 | 1618 } |
612 | 1619 |
1389 | 1620 if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){ |
1708 | 1621 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT; |
1622 if(fimin < INT_MAX) | |
1623 type |= CANDIDATE_MB_TYPE_FORWARD_I; | |
1624 if(bimin < INT_MAX) | |
1625 type |= CANDIDATE_MB_TYPE_BACKWARD_I; | |
1626 if(fimin < INT_MAX && bimin < INT_MAX){ | |
1627 type |= CANDIDATE_MB_TYPE_BIDIR_I; | |
1628 } | |
1629 //FIXME something smarter | |
1630 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB | |
1729 | 1631 #if 0 |
1632 if(s->out_format == FMT_MPEG1) | |
1633 type |= CANDIDATE_MB_TYPE_INTRA; | |
1634 #endif | |
612 | 1635 } |
1636 | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1637 s->mb_type[mb_y*s->mb_stride + mb_x]= type; |
324 | 1638 } |
0 | 1639 |
324 | 1640 /* find best f_code for ME which do unlimited searches */ |
1641 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type) | |
1642 { | |
1643 if(s->me_method>=ME_EPZS){ | |
455 | 1644 int score[8]; |
324 | 1645 int i, y; |
1064 | 1646 uint8_t * fcode_tab= s->fcode_tab; |
455 | 1647 int best_fcode=-1; |
1648 int best_score=-10000000; | |
324 | 1649 |
936 | 1650 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); |
324 | 1651 |
1652 for(y=0; y<s->mb_height; y++){ | |
1653 int x; | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1654 int xy= y*s->mb_stride; |
324 | 1655 for(x=0; x<s->mb_width; x++){ |
1634 | 1656 if(s->mb_type[xy] & type){ |
847 | 1657 int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV], |
1658 fcode_tab[mv_table[xy][1] + MAX_MV]); | |
455 | 1659 int j; |
1660 | |
1661 for(j=0; j<fcode && j<8; j++){ | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1662 if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy]) |
455 | 1663 score[j]-= 170; |
1664 } | |
324 | 1665 } |
1666 xy++; | |
1667 } | |
1668 } | |
455 | 1669 |
1670 for(i=1; i<8; i++){ | |
1671 if(score[i] > best_score){ | |
1672 best_score= score[i]; | |
1673 best_fcode= i; | |
1674 } | |
1675 // printf("%d %d\n", i, score[i]); | |
1676 } | |
327 | 1677 |
324 | 1678 // printf("fcode: %d type: %d\n", i, s->pict_type); |
455 | 1679 return best_fcode; |
324 | 1680 /* for(i=0; i<=MAX_FCODE; i++){ |
1681 printf("%d ", mv_num[i]); | |
1682 } | |
1683 printf("\n");*/ | |
1684 }else{ | |
1685 return 1; | |
0 | 1686 } |
1687 } | |
1688 | |
324 | 1689 void ff_fix_long_p_mvs(MpegEncContext * s) |
1690 { | |
1691 const int f_code= s->f_code; | |
1086 | 1692 int y, range; |
1421 | 1693 assert(s->pict_type==P_TYPE); |
1086 | 1694 |
1421 | 1695 range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code); |
1086 | 1696 |
1697 if(s->msmpeg4_version) range= 16; | |
1698 | |
1699 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range; | |
1700 | |
455 | 1701 //printf("%d no:%d %d//\n", clip, noclip, f_code); |
324 | 1702 if(s->flags&CODEC_FLAG_4MV){ |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
1703 const int wrap= s->b8_stride; |
324 | 1704 |
1705 /* clip / convert to intra 8x8 type MVs */ | |
1706 for(y=0; y<s->mb_height; y++){ | |
1938
e2501e6e7ff7
unify table indexing (motion_val,dc_val,ac_val,coded_block changed)
michael
parents:
1904
diff
changeset
|
1707 int xy= y*2*wrap; |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1708 int i= y*s->mb_stride; |
324 | 1709 int x; |
1710 | |
1711 for(x=0; x<s->mb_width; x++){ | |
1708 | 1712 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){ |
324 | 1713 int block; |
1714 for(block=0; block<4; block++){ | |
1715 int off= (block& 1) + (block>>1)*wrap; | |
1668
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1716 int mx= s->current_picture.motion_val[0][ xy + off ][0]; |
30746f429df6
move motion_val & mb_type to AVFrame patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
michael
parents:
1634
diff
changeset
|
1717 int my= s->current_picture.motion_val[0][ xy + off ][1]; |
324 | 1718 |
1086 | 1719 if( mx >=range || mx <-range |
1720 || my >=range || my <-range){ | |
1708 | 1721 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V; |
1722 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA; | |
1723 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA; | |
324 | 1724 } |
1725 } | |
1726 } | |
502 | 1727 xy+=2; |
1728 i++; | |
324 | 1729 } |
1730 } | |
1731 } | |
1732 } | |
1733 | |
1708 | 1734 /** |
1735 * | |
1736 * @param truncate 1 for truncation, 0 for using intra | |
1737 */ | |
1738 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select, | |
1739 int16_t (*mv_table)[2], int f_code, int type, int truncate) | |
324 | 1740 { |
1708 | 1741 int y, h_range, v_range; |
324 | 1742 |
1051
e5a9dbf597d4
mpeg1 bframe encoding patch by (Rapha¸«³l LEGRAND) with some modifications by me
michaelni
parents:
1050
diff
changeset
|
1743 // RAL: 8 in MPEG-1, 16 in MPEG-4 |
1421 | 1744 int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code); |
1708 | 1745 |
1746 if(s->msmpeg4_version) range= 16; | |
1086 | 1747 if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range; |
1051
e5a9dbf597d4
mpeg1 bframe encoding patch by (Rapha¸«³l LEGRAND) with some modifications by me
michaelni
parents:
1050
diff
changeset
|
1748 |
1708 | 1749 h_range= range; |
1750 v_range= field_select_table ? range>>1 : range; | |
1751 | |
324 | 1752 /* clip / convert to intra 16x16 type MVs */ |
1753 for(y=0; y<s->mb_height; y++){ | |
1754 int x; | |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1755 int xy= y*s->mb_stride; |
1086 | 1756 for(x=0; x<s->mb_width; x++){ |
1177
fea03d2c4946
simplified adressing of most mb based arrays (mb_x + mb_y*s->mb_stride) now instead of mb_x + mb_y*mb_width and 1+mb_x + (1+mb_y)*(mb_width+2) and ... mixture
michaelni
parents:
1162
diff
changeset
|
1757 if (s->mb_type[xy] & type){ // RAL: "type" test added... |
1708 | 1758 if(field_select_table==NULL || field_select_table[xy] == field_select){ |
1759 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range | |
1760 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){ | |
1086 | 1761 |
1708 | 1762 if(truncate){ |
1763 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1; | |
1764 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range; | |
1765 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1; | |
1766 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range; | |
1767 }else{ | |
1768 s->mb_type[xy] &= ~type; | |
1769 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA; | |
1770 mv_table[xy][0]= | |
1771 mv_table[xy][1]= 0; | |
1772 } | |
1051
e5a9dbf597d4
mpeg1 bframe encoding patch by (Rapha¸«³l LEGRAND) with some modifications by me
michaelni
parents:
1050
diff
changeset
|
1773 } |
1086 | 1774 } |
324 | 1775 } |
1776 xy++; | |
1777 } | |
1778 } | |
1779 } |