[svn] / branches / dev-api-4 / xvidcore / src / motion / motion_est.h Repository:
ViewVC logotype

Annotation of /branches/dev-api-4/xvidcore/src/motion/motion_est.h

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1084 - (view) (download)

1 : edgomez 1054 /*****************************************************************************
2 : edgomez 851 *
3 :     * XVID MPEG-4 VIDEO CODEC
4 : edgomez 1054 * - Motion Estimation related header -
5 : edgomez 851 *
6 : edgomez 1054 * Copyright(C) 2002 Christoph Lampert <gruel@web.de>
7 :     * 2002 Michael Militzer <michael@xvid.org>
8 :     * 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net>
9 : edgomez 851 *
10 : edgomez 1054 * This program is free software ; you can redistribute it and/or modify
11 : edgomez 851 * it under the terms of the GNU General Public License as published by
12 : edgomez 1054 * the Free Software Foundation ; either version 2 of the License, or
13 : edgomez 851 * (at your option) any later version.
14 :     *
15 :     * This program is distributed in the hope that it will be useful,
16 : edgomez 1054 * but WITHOUT ANY WARRANTY ; without even the implied warranty of
17 : edgomez 851 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 :     * GNU General Public License for more details.
19 :     *
20 :     * You should have received a copy of the GNU General Public License
21 : edgomez 1054 * along with this program ; if not, write to the Free Software
22 : edgomez 851 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 :     *
24 : syskin 1084 * $Id: motion_est.h,v 1.3.2.13 2003-07-13 09:57:28 syskin Exp $
25 : edgomez 851 *
26 : edgomez 1054 ****************************************************************************/
27 : edgomez 851
28 :     #ifndef _MOTION_EST_H_
29 :     #define _MOTION_EST_H_
30 :    
31 :     #include "../portab.h"
32 :     #include "../global.h"
33 :     #include "../image/reduced.h"
34 :    
35 :     /* hard coded motion search parameters for motion_est and smp_motion_est */
36 :    
37 : edgomez 1053 /* very large value */
38 : edgomez 851 #define MV_MAX_ERROR (4096 * 256)
39 :    
40 :     /* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */
41 :     #define MV16_INTER_BIAS 512
42 :    
43 :     /* vector map (vlc delta size) smoother parameters ! float !*/
44 :     #define NEIGH_TEND_16X16 10.5
45 :     #define NEIGH_TEND_8X8 40.0
46 :     #define NEIGH_8X8_BIAS 30
47 :    
48 : edgomez 1023 #define BITS_MULT 16
49 :    
50 : edgomez 851 /* Parameters which control inter/inter4v decision */
51 :     #define IMV16X16 2
52 :    
53 :     static const int lambda_vec16[32] =
54 :     { 0 ,(int)(1.00235 * NEIGH_TEND_16X16 + 0.5),
55 :     (int)(1.15582*NEIGH_TEND_16X16 + 0.5), (int)(1.31976*NEIGH_TEND_16X16 + 0.5),
56 :     (int)(1.49591*NEIGH_TEND_16X16 + 0.5), (int)(1.68601*NEIGH_TEND_16X16 + 0.5),
57 :     (int)(1.89187*NEIGH_TEND_16X16 + 0.5), (int)(2.11542*NEIGH_TEND_16X16 + 0.5),
58 :     (int)(2.35878*NEIGH_TEND_16X16 + 0.5), (int)(2.62429*NEIGH_TEND_16X16 + 0.5),
59 :     (int)(2.91455*NEIGH_TEND_16X16 + 0.5), (int)(3.23253*NEIGH_TEND_16X16 + 0.5),
60 :     (int)(3.58158*NEIGH_TEND_16X16 + 0.5), (int)(3.96555*NEIGH_TEND_16X16 + 0.5),
61 :     (int)(4.38887*NEIGH_TEND_16X16 + 0.5), (int)(4.85673*NEIGH_TEND_16X16 + 0.5),
62 :     (int)(5.37519*NEIGH_TEND_16X16 + 0.5), (int)(5.95144*NEIGH_TEND_16X16 + 0.5),
63 :     (int)(6.59408*NEIGH_TEND_16X16 + 0.5), (int)(7.31349*NEIGH_TEND_16X16 + 0.5),
64 :     (int)(8.12242*NEIGH_TEND_16X16 + 0.5), (int)(9.03669*NEIGH_TEND_16X16 + 0.5),
65 :     (int)(10.0763*NEIGH_TEND_16X16 + 0.5), (int)(11.2669*NEIGH_TEND_16X16 + 0.5),
66 :     (int)(12.6426*NEIGH_TEND_16X16 + 0.5), (int)(14.2493*NEIGH_TEND_16X16 + 0.5),
67 :     (int)(16.1512*NEIGH_TEND_16X16 + 0.5), (int)(18.442*NEIGH_TEND_16X16 + 0.5),
68 :     (int)(21.2656*NEIGH_TEND_16X16 + 0.5), (int)(24.8580*NEIGH_TEND_16X16 + 0.5),
69 :     (int)(29.6436*NEIGH_TEND_16X16 + 0.5), (int)(36.4949*NEIGH_TEND_16X16 + 0.5) };
70 :    
71 :     static const int lambda_vec8[32] =
72 :     { 0 ,(int)(1.00235 * NEIGH_TEND_8X8 + 0.5),
73 :     (int)(1.15582 + NEIGH_TEND_8X8 + 0.5), (int)(1.31976*NEIGH_TEND_8X8 + 0.5),
74 :     (int)(1.49591*NEIGH_TEND_8X8 + 0.5), (int)(1.68601*NEIGH_TEND_8X8 + 0.5),
75 :     (int)(1.89187*NEIGH_TEND_8X8 + 0.5), (int)(2.11542*NEIGH_TEND_8X8 + 0.5),
76 :     (int)(2.35878*NEIGH_TEND_8X8 + 0.5), (int)(2.62429*NEIGH_TEND_8X8 + 0.5),
77 :     (int)(2.91455*NEIGH_TEND_8X8 + 0.5), (int)(3.23253*NEIGH_TEND_8X8 + 0.5),
78 :     (int)(3.58158*NEIGH_TEND_8X8 + 0.5), (int)(3.96555*NEIGH_TEND_8X8 + 0.5),
79 :     (int)(4.38887*NEIGH_TEND_8X8 + 0.5), (int)(4.85673*NEIGH_TEND_8X8 + 0.5),
80 :     (int)(5.37519*NEIGH_TEND_8X8 + 0.5), (int)(5.95144*NEIGH_TEND_8X8 + 0.5),
81 :     (int)(6.59408*NEIGH_TEND_8X8 + 0.5), (int)(7.31349*NEIGH_TEND_8X8 + 0.5),
82 :     (int)(8.12242*NEIGH_TEND_8X8 + 0.5), (int)(9.03669*NEIGH_TEND_8X8 + 0.5),
83 :     (int)(10.0763*NEIGH_TEND_8X8 + 0.5), (int)(11.2669*NEIGH_TEND_8X8 + 0.5),
84 :     (int)(12.6426*NEIGH_TEND_8X8 + 0.5), (int)(14.2493*NEIGH_TEND_8X8 + 0.5),
85 :     (int)(16.1512*NEIGH_TEND_8X8 + 0.5), (int)(18.442*NEIGH_TEND_8X8 + 0.5),
86 :     (int)(21.2656*NEIGH_TEND_8X8 + 0.5), (int)(24.8580*NEIGH_TEND_8X8 + 0.5),
87 :     (int)(29.6436*NEIGH_TEND_8X8 + 0.5), (int)(36.4949*NEIGH_TEND_8X8 + 0.5) };
88 :    
89 : edgomez 1053 /* mv.length table */
90 : edgomez 974 static const int mvtab[64] = {
91 :     1, 2, 3, 4, 6, 7, 7, 7,
92 :     9, 9, 9, 10, 10, 10, 10, 10,
93 :     10, 10, 10, 10, 10, 10, 10, 10,
94 :     10, 11, 11, 11, 11, 11, 11, 12,
95 :     12, 12, 12, 12, 12, 12, 12, 12,
96 :     12, 12, 12, 12, 12, 12, 12, 12,
97 :     12, 12, 12, 12, 12, 12, 12, 12, 12 };
98 : edgomez 851
99 :     static const int DQtab[4] = {
100 :     -1, -2, 1, 2
101 :     };
102 :    
103 :     #define RRV_MV_SCALEDOWN(a) ( (a)>=0 ? (a+1)/2 : (a-1)/2 )
104 :    
105 :     typedef struct
106 :     {
107 : edgomez 1053 /* general fields */
108 : edgomez 851 int max_dx, min_dx, max_dy, min_dy;
109 :     uint32_t rounding;
110 :     VECTOR predMV;
111 :     VECTOR * currentMV;
112 :     VECTOR * currentQMV;
113 :     int32_t * iMinSAD;
114 : edgomez 1053 const uint8_t * RefP[6]; /* N, V, H, HV, cU, cV */
115 : edgomez 851 const uint8_t * CurU;
116 :     const uint8_t * CurV;
117 :     uint8_t * RefQ;
118 :     const uint8_t * Cur;
119 :     uint32_t lambda16;
120 :     uint32_t lambda8;
121 :     uint32_t iEdgedWidth;
122 :     uint32_t iFcode;
123 :     int * temp;
124 :     int qpel, qpel_precision;
125 :     int chroma;
126 :     int rrv;
127 : edgomez 1053
128 :     /* fields for interpolate and direct modes */
129 :     const uint8_t * b_RefP[6]; /* N, V, H, HV, cU, cV */
130 : edgomez 851 VECTOR bpredMV;
131 :     uint32_t bFcode;
132 : edgomez 1053
133 :     /* fields for direct mode */
134 : edgomez 851 VECTOR directmvF[4];
135 :     VECTOR directmvB[4];
136 :     const VECTOR * referencemv;
137 : edgomez 1053
138 :     /* BITS/R-D stuff */
139 : edgomez 959 int16_t * dctSpace;
140 : edgomez 1022 uint32_t iQuant;
141 :     uint32_t quant_type;
142 : syskin 1084 int * cbp;
143 : edgomez 851
144 :     } SearchData;
145 :    
146 :    
147 :     typedef void(CheckFunc)(const int x, const int y,
148 :     const int Direction, int * const dir,
149 :     const SearchData * const Data);
150 :     CheckFunc *CheckCandidate;
151 :    
152 :     /*
153 :     * Calculate the min/max range
154 :     * relative to the _MACROBLOCK_ position
155 :     */
156 :     static void __inline
157 :     get_range(int32_t * const min_dx,
158 :     int32_t * const max_dx,
159 :     int32_t * const min_dy,
160 :     int32_t * const max_dy,
161 :     const uint32_t x,
162 :     const uint32_t y,
163 :     uint32_t block_sz, /* block dimension, 8 or 16 */
164 :     const uint32_t width,
165 :     const uint32_t height,
166 :     const uint32_t fcode,
167 :     const int qpel, /* 1 if the resulting range should be in qpel precision; otherwise 0 */
168 :     const int rrv)
169 :     {
170 :     int k, m = qpel ? 4 : 2;
171 :     const int search_range = 32 << (fcode - 1);
172 :     int high = search_range - 1;
173 :     int low = -search_range;
174 :    
175 :     if (rrv) {
176 :     high = RRV_MV_SCALEUP(high);
177 :     low = RRV_MV_SCALEUP(low);
178 :     block_sz *= 2;
179 :     }
180 :    
181 :     k = m * (int)(width - x * block_sz);
182 :     *max_dx = MIN(high, k);
183 :     k = m * (int)(height - y * block_sz);
184 :     *max_dy = MIN(high, k);
185 :    
186 :     k = -m * (int)((x+1) * block_sz);
187 :     *min_dx = MAX(low, k);
188 :     k = -m * (int)((y+1) * block_sz);
189 :     *min_dy = MAX(low, k);
190 :     }
191 :    
192 :     typedef void MainSearchFunc(int x, int y, const SearchData * const Data, int bDirection);
193 :    
194 :     static MainSearchFunc DiamondSearch, AdvDiamondSearch, SquareSearch;
195 :    
196 :     static void Search8(const SearchData * const OldData,
197 :     const int x, const int y,
198 :     const uint32_t MotionFlags,
199 :     const MBParam * const pParam,
200 :     MACROBLOCK * const pMB,
201 :     const MACROBLOCK * const pMBs,
202 :     const int block,
203 :     SearchData * const Data);
204 :    
205 :     bool
206 :     MotionEstimation(MBParam * const pParam,
207 : syskin 884 FRAMEINFO * const current,
208 :     FRAMEINFO * const reference,
209 :     const IMAGE * const pRefH,
210 :     const IMAGE * const pRefV,
211 :     const IMAGE * const pRefHV,
212 : chl 1077 const IMAGE * const pGMC,
213 : syskin 884 const uint32_t iLimit);
214 : edgomez 851
215 :     static void
216 :     SearchP(const IMAGE * const pRef,
217 :     const uint8_t * const pRefH,
218 :     const uint8_t * const pRefV,
219 :     const uint8_t * const pRefHV,
220 :     const IMAGE * const pCur,
221 :     const int x,
222 :     const int y,
223 :     const uint32_t MotionFlags,
224 : edgomez 1022 const uint32_t VopFlags,
225 :     const uint32_t VolFlags,
226 : edgomez 851 SearchData * const Data,
227 :     const MBParam * const pParam,
228 :     const MACROBLOCK * const pMBs,
229 :     const MACROBLOCK * const prevMBs,
230 :     MACROBLOCK * const pMB);
231 :    
232 : chl 1077 static __inline void
233 :     GMEanalyzeMB (const uint8_t * const pCur,
234 :     const uint8_t * const pRef,
235 :     const uint8_t * const pRefH,
236 :     const uint8_t * const pRefV,
237 :     const uint8_t * const pRefHV,
238 :     const int x,
239 :     const int y,
240 :     const MBParam * const pParam,
241 :     MACROBLOCK * const pMBs,
242 :     SearchData * const Data);
243 :    
244 :     void
245 :     GMEanalysis(const MBParam * const pParam,
246 :     const FRAMEINFO * const current,
247 :     const FRAMEINFO * const reference,
248 :     const IMAGE * const pRefH,
249 :     const IMAGE * const pRefV,
250 :     const IMAGE * const pRefHV);
251 :    
252 :    
253 :    
254 :     WARPPOINTS
255 :     GlobalMotionEst(MACROBLOCK * const pMBs,
256 : edgomez 851 const MBParam * const pParam,
257 :     const FRAMEINFO * const current,
258 :     const FRAMEINFO * const reference,
259 :     const IMAGE * const pRefH,
260 :     const IMAGE * const pRefV,
261 :     const IMAGE * const pRefHV );
262 :    
263 : chl 1077
264 :     int
265 :     GlobalMotionEstRefine(WARPPOINTS *const startwp,
266 :     MACROBLOCK * const pMBs,
267 :     const MBParam * const pParam,
268 :     const FRAMEINFO * const current,
269 :     const FRAMEINFO * const reference,
270 :     const IMAGE * const pCurr,
271 :     const IMAGE * const pRef,
272 :     const IMAGE * const pRefH,
273 :     const IMAGE * const pRefV,
274 :     const IMAGE * const pRefHV);
275 :    
276 :    
277 :    
278 :    
279 :     int
280 :     globalSAD(const WARPPOINTS *const wp,
281 :     const MBParam * const pParam,
282 :     const MACROBLOCK * const pMBs,
283 :     const FRAMEINFO * const current,
284 :     const IMAGE * const pRef,
285 :     const IMAGE * const pCurr,
286 :     uint8_t *const GMCblock);
287 :    
288 : edgomez 851 #define iDiamondSize 2
289 :    
290 :     static __inline uint32_t
291 : suxen_drol 890 MakeGoodMotionFlags(const uint32_t MotionFlags, const uint32_t VopFlags, const uint32_t VolFlags)
292 : edgomez 851 {
293 :     uint32_t Flags = MotionFlags;
294 :    
295 : edgomez 949 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS))
296 :     Flags &= ~(XVID_ME_QUARTERPELREFINE16_BITS+XVID_ME_QUARTERPELREFINE8_BITS+XVID_ME_HALFPELREFINE16_BITS+XVID_ME_HALFPELREFINE8_BITS+XVID_ME_EXTSEARCH_BITS);
297 : edgomez 851
298 : edgomez 949 if (Flags & XVID_ME_EXTSEARCH_BITS)
299 :     Flags |= XVID_ME_HALFPELREFINE16_BITS;
300 : edgomez 851
301 : edgomez 949 if (Flags & XVID_ME_EXTSEARCH_BITS && MotionFlags & XVID_ME_EXTSEARCH8)
302 :     Flags |= XVID_ME_HALFPELREFINE8_BITS;
303 : edgomez 851
304 : edgomez 949 if (Flags & XVID_ME_HALFPELREFINE16_BITS)
305 :     Flags |= XVID_ME_QUARTERPELREFINE16_BITS;
306 : edgomez 851
307 : edgomez 949 if (Flags & XVID_ME_HALFPELREFINE8_BITS) {
308 :     Flags |= XVID_ME_QUARTERPELREFINE8_BITS;
309 :     Flags &= ~XVID_ME_HALFPELREFINE8;
310 : edgomez 851 }
311 :    
312 : edgomez 949 if (Flags & XVID_ME_QUARTERPELREFINE8_BITS)
313 :     Flags &= ~XVID_ME_QUARTERPELREFINE8;
314 : edgomez 851
315 : edgomez 949 if (!(VolFlags & XVID_VOL_QUARTERPEL))
316 :     Flags &= ~(XVID_ME_QUARTERPELREFINE16+XVID_ME_QUARTERPELREFINE8+XVID_ME_QUARTERPELREFINE16_BITS+XVID_ME_QUARTERPELREFINE8_BITS);
317 : edgomez 851
318 : edgomez 949 if (!(VopFlags & XVID_VOP_HALFPEL))
319 :     Flags &= ~(XVID_ME_EXTSEARCH16+XVID_ME_HALFPELREFINE16+XVID_ME_HALFPELREFINE8+XVID_ME_HALFPELREFINE16_BITS+XVID_ME_HALFPELREFINE8_BITS);
320 : edgomez 851
321 : edgomez 949 if ((VopFlags & XVID_VOP_GREYSCALE) || (VopFlags & XVID_VOP_REDUCED))
322 :     Flags &= ~(XVID_ME_CHROMA16 + XVID_ME_CHROMA8);
323 : edgomez 851
324 :     return Flags;
325 :     }
326 :    
327 :     /* BITS mode decision and search */
328 :    
329 :     #include "../bitstream/zigzag.h"
330 :     #include "../quant/quant_mpeg4.h"
331 :     #include "../quant/quant_h263.h"
332 :     #include "../bitstream/vlc_codes.h"
333 : edgomez 1022 #include "../dct/fdct.h"
334 : edgomez 851
335 :     static int
336 :     CountMBBitsInter(SearchData * const Data,
337 :     const MACROBLOCK * const pMBs, const int x, const int y,
338 :     const MBParam * const pParam,
339 :     const uint32_t MotionFlags);
340 :    
341 :     static int
342 :     CountMBBitsInter4v(const SearchData * const Data,
343 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
344 :     const int x, const int y,
345 :     const MBParam * const pParam, const uint32_t MotionFlags,
346 :     const VECTOR * const backup);
347 :    
348 :     static int
349 :     CountMBBitsIntra(const SearchData * const Data);
350 :    
351 : syskin 1084 static int
352 :     CountMBBitsGMC(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y);
353 :    
354 : edgomez 851 int CodeCoeffIntra_CalcBits(const int16_t qcoeff[64], const uint16_t * zigzag);
355 :     int CodeCoeffInter_CalcBits(const int16_t qcoeff[64], const uint16_t * zigzag);
356 : edgomez 1023
357 : syskin 1075 /* one over lambda for R-D mode decision and motion search */
358 :     #define LAMBDA ( (int)(BITS_MULT/1.0) )
359 : edgomez 1023
360 : edgomez 1022 static __inline unsigned int
361 : edgomez 1023 Block_CalcBits( int16_t * const coeff,
362 :     int16_t * const data,
363 : syskin 1075 int16_t * const dqcoeff,
364 : edgomez 1023 const uint32_t quant, const int quant_type,
365 :     uint32_t * cbp,
366 :     const int block)
367 : edgomez 1022 {
368 :     int sum;
369 : edgomez 1023 int bits;
370 :     int distortion = 0;
371 :     int i;
372 : edgomez 851
373 : edgomez 1022 fdct(data);
374 :    
375 : syskin 1084 if (quant_type) sum = quant_inter(coeff, data, quant);
376 : edgomez 1022 else sum = quant4_inter(coeff, data, quant);
377 :    
378 :     if (sum > 0) {
379 :     *cbp |= 1 << (5 - block);
380 : edgomez 1023 bits = BITS_MULT * CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
381 :    
382 : syskin 1084 if (quant_type) dequant_inter(dqcoeff, coeff, quant);
383 :     else dequant4_inter(dqcoeff, coeff, quant);
384 : edgomez 1023
385 : syskin 1084 for (i = 0; i < 64; i++)
386 :     distortion += (data[i] - dqcoeff[i])*(data[i] - dqcoeff[i]);
387 :    
388 :     } else {
389 :     bits = 0;
390 :     for (i = 0; i < 64; i++)
391 :     distortion += data[i]*data[i];
392 : edgomez 1023 }
393 :    
394 : syskin 1084
395 : syskin 1075 bits += (LAMBDA*distortion)/(quant*quant);
396 : edgomez 1023
397 :     return bits;
398 : edgomez 1022 }
399 :    
400 : edgomez 1023 static __inline unsigned int
401 :     Block_CalcBitsIntra(int16_t * const coeff,
402 :     int16_t * const data,
403 : syskin 1075 int16_t * const dqcoeff,
404 : edgomez 1023 const uint32_t quant, const int quant_type,
405 :     uint32_t * cbp,
406 :     const int block,
407 :     int * dcpred)
408 :     {
409 :     int bits, i;
410 :     int distortion = 0;
411 : syskin 1071 uint32_t iDcScaler = get_dc_scaler(quant, block < 4);
412 : edgomez 1023 int b_dc;
413 :    
414 :     fdct(data);
415 :     data[0] -= 1024;
416 :    
417 : syskin 1084 if (quant_type) quant_intra(coeff, data, quant, iDcScaler);
418 : syskin 1075 else quant4_intra(coeff, data, quant, iDcScaler);
419 : edgomez 1023
420 :     b_dc = coeff[0];
421 :     if (block < 4) {
422 :     coeff[0] -= *dcpred;
423 :     *dcpred = b_dc;
424 :     }
425 :    
426 :     bits = BITS_MULT*CodeCoeffIntra_CalcBits(coeff, scan_tables[0]);
427 :     if (bits != 0) *cbp |= 1 << (5 - block);
428 :    
429 : syskin 1071 if (block < 4) bits += BITS_MULT*dcy_tab[coeff[0] + 255].len;
430 :     else bits += BITS_MULT*dcc_tab[coeff[0] + 255].len;
431 :    
432 : edgomez 1023 coeff[0] = b_dc;
433 : syskin 1084 if (quant_type) dequant_intra(dqcoeff, coeff, quant, iDcScaler);
434 : syskin 1075 else dequant4_intra(dqcoeff, coeff, quant, iDcScaler);
435 : edgomez 1023
436 :     for (i = 0; i < 64; i++) {
437 : syskin 1075 distortion += (data[i] - dqcoeff[i])*(data[i] - dqcoeff[i]);
438 : edgomez 1023 }
439 :    
440 : syskin 1075 bits += (LAMBDA*distortion)/(quant*quant);
441 : edgomez 1023
442 :     return bits;
443 :     }
444 :    
445 : edgomez 851 #endif /* _MOTION_EST_H_ */

No admin address has been configured
ViewVC Help
Powered by ViewVC 1.0.4