1 /*****************************************************************************
2 *
3 * XVID MPEG-4 VIDEO CODEC
4 * - Motion Estimation for P- and S- VOPs -
5 *
6 * Copyright(C) 2002 Christoph Lampert <gruel@web.de>
7 * 2002-2010 Michael Militzer <michael@xvid.org>
8 * 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net>
9 *
10 * This program is free software ; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation ; either version 2 of the License, or
13 * (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY ; without even the implied warranty of
17 * 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 * along with this program ; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 *
24 * $Id: estimation_pvop.c 1985 2011-05-18 09:02:35Z Isibaar $
25 *
26 ****************************************************************************/
27
28 #include <assert.h>
29 #include <stdio.h>
30 #include <stdlib.h>
31 #include <string.h> /* memcpy */
32
33 #include "../encoder.h"
34 #include "../prediction/mbprediction.h"
35 #include "../global.h"
36 #include "../utils/timer.h"
37 #include "../image/interpolate8x8.h"
38 #include "estimation.h"
39 #include "motion.h"
40 #include "sad.h"
41 #include "motion_inlines.h"
42 #include "motion_smp.h"
43
44
45 static const int xvid_me_lambda_vec8[32] =
46 { 0 ,(int)(1.0 * NEIGH_TEND_8X8 + 0.5),
47 (int)(2.0*NEIGH_TEND_8X8 + 0.5), (int)(3.0*NEIGH_TEND_8X8 + 0.5),
48 (int)(4.0*NEIGH_TEND_8X8 + 0.5), (int)(5.0*NEIGH_TEND_8X8 + 0.5),
49 (int)(6.0*NEIGH_TEND_8X8 + 0.5), (int)(7.0*NEIGH_TEND_8X8 + 0.5),
50 (int)(8.0*NEIGH_TEND_8X8 + 0.5), (int)(9.0*NEIGH_TEND_8X8 + 0.5),
51 (int)(10.0*NEIGH_TEND_8X8 + 0.5), (int)(11.0*NEIGH_TEND_8X8 + 0.5),
52 (int)(12.0*NEIGH_TEND_8X8 + 0.5), (int)(13.0*NEIGH_TEND_8X8 + 0.5),
53 (int)(14.0*NEIGH_TEND_8X8 + 0.5), (int)(15.0*NEIGH_TEND_8X8 + 0.5),
54 (int)(16.0*NEIGH_TEND_8X8 + 0.5), (int)(17.0*NEIGH_TEND_8X8 + 0.5),
55 (int)(18.0*NEIGH_TEND_8X8 + 0.5), (int)(19.0*NEIGH_TEND_8X8 + 0.5),
56 (int)(20.0*NEIGH_TEND_8X8 + 0.5), (int)(21.0*NEIGH_TEND_8X8 + 0.5),
57 (int)(22.0*NEIGH_TEND_8X8 + 0.5), (int)(23.0*NEIGH_TEND_8X8 + 0.5),
58 (int)(24.0*NEIGH_TEND_8X8 + 0.5), (int)(25.0*NEIGH_TEND_8X8 + 0.5),
59 (int)(26.0*NEIGH_TEND_8X8 + 0.5), (int)(27.0*NEIGH_TEND_8X8 + 0.5),
60 (int)(28.0*NEIGH_TEND_8X8 + 0.5), (int)(29.0*NEIGH_TEND_8X8 + 0.5),
61 (int)(30.0*NEIGH_TEND_8X8 + 0.5), (int)(31.0*NEIGH_TEND_8X8 + 0.5)
62 };
63
64 static void
CheckCandidate16(const int x,const int y,SearchData * const data,const unsigned int Direction)65 CheckCandidate16(const int x, const int y, SearchData * const data, const unsigned int Direction)
66 {
67 const uint8_t * Reference;
68 int32_t sad, xc, yc; uint32_t t;
69 VECTOR * current;
70
71 if ( (x > data->max_dx) || (x < data->min_dx)
72 || (y > data->max_dy) || (y < data->min_dy) ) return;
73
74 if (data->qpel_precision) { /* x and y are in 1/4 precision */
75 Reference = xvid_me_interpolate16x16qpel(x, y, 0, data);
76 current = data->currentQMV;
77 xc = x/2; yc = y/2;
78 } else {
79 Reference = GetReference(x, y, data);
80 current = data->currentMV;
81 xc = x; yc = y;
82 }
83
84 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp);
85 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);
86
87 sad += (data->lambda16 * t);
88 data->temp[0] += (data->lambda8 * t);
89
90 if (data->chroma) {
91 if (sad >= data->iMinSAD[0]) goto no16;
92 sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
93 (yc >> 1) + roundtab_79[yc & 0x3], data);
94 }
95
96 if (sad < data->iMinSAD[0]) {
97 data->iMinSAD[0] = sad;
98 current[0].x = x; current[0].y = y;
99 data->dir = Direction;
100 }
101
102 no16:
103 if (data->temp[0] < data->iMinSAD[1]) {
104 data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; }
105 if (data->temp[1] < data->iMinSAD[2]) {
106 data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
107 if (data->temp[2] < data->iMinSAD[3]) {
108 data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
109 if (data->temp[3] < data->iMinSAD[4]) {
110 data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
111 }
112
113 static void
CheckCandidate8(const int x,const int y,SearchData * const data,const unsigned int Direction)114 CheckCandidate8(const int x, const int y, SearchData * const data, const unsigned int Direction)
115 {
116 int32_t sad; uint32_t t;
117 const uint8_t * Reference;
118 VECTOR * current;
119
120 if ( (x > data->max_dx) || (x < data->min_dx)
121 || (y > data->max_dy) || (y < data->min_dy) ) return;
122
123 if (!data->qpel_precision) {
124 Reference = GetReference(x, y, data);
125 current = data->currentMV;
126 } else { /* x and y are in 1/4 precision */
127 Reference = xvid_me_interpolate8x8qpel(x, y, 0, 0, data);
128 current = data->currentQMV;
129 }
130
131 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
132 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);
133
134 sad += (data->lambda8 * t);
135
136 if (sad < *(data->iMinSAD)) {
137 *(data->iMinSAD) = sad;
138 current->x = x; current->y = y;
139 data->dir = Direction;
140 }
141 }
142
143 int
xvid_me_SkipDecisionP(const IMAGE * current,const IMAGE * reference,const int x,const int y,const uint32_t stride,const uint32_t iQuant)144 xvid_me_SkipDecisionP(const IMAGE * current, const IMAGE * reference,
145 const int x, const int y,
146 const uint32_t stride, const uint32_t iQuant)
147 {
148 int offset = (x + y*stride)*8;
149 uint32_t sadC = sad8(current->u + offset,
150 reference->u + offset, stride);
151 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
152 sadC += sad8(current->v + offset,
153 reference->v + offset, stride);
154 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
155 return 1;
156 }
157
158 /*
159 * pmv are filled with:
160 * [0]: Median (or whatever is correct in a special case)
161 * [1]: left neighbour
162 * [2]: top neighbour
163 * [3]: topright neighbour
164 * psad are filled with:
165 * [0]: minimum of [1] to [3]
166 * [1]: left neighbour's SAD (NB:[1] to [3] are actually not needed)
167 * [2]: top neighbour's SAD
168 * [3]: topright neighbour's SAD
169 */
170
171 static __inline void
get_pmvdata2(const MACROBLOCK * const mbs,const int mb_width,const int bound,const int x,const int y,VECTOR * const pmv,int32_t * const psad)172 get_pmvdata2(const MACROBLOCK * const mbs,
173 const int mb_width,
174 const int bound,
175 const int x,
176 const int y,
177 VECTOR * const pmv,
178 int32_t * const psad)
179 {
180 int lx, ly, lz; /* left */
181 int tx, ty, tz; /* top */
182 int rx, ry, rz; /* top-right */
183 int lpos, tpos, rpos;
184 int num_cand = 0, last_cand = 1;
185
186 lx = x - 1; ly = y; lz = 1;
187 tx = x; ty = y - 1; tz = 2;
188 rx = x + 1; ry = y - 1; rz = 2;
189
190 lpos = lx + ly * mb_width;
191 rpos = rx + ry * mb_width;
192 tpos = tx + ty * mb_width;
193
194 if (lpos >= bound && lx >= 0) {
195 num_cand++;
196 last_cand = 1;
197 pmv[1] = mbs[lpos].mvs[lz];
198 psad[1] = mbs[lpos].sad8[lz];
199 } else {
200 pmv[1] = zeroMV;
201 psad[1] = MV_MAX_ERROR;
202 }
203
204 if (tpos >= bound) {
205 num_cand++;
206 last_cand = 2;
207 pmv[2]= mbs[tpos].mvs[tz];
208 psad[2] = mbs[tpos].sad8[tz];
209 } else {
210 pmv[2] = zeroMV;
211 psad[2] = MV_MAX_ERROR;
212 }
213
214 if (rpos >= bound && rx < mb_width) {
215 num_cand++;
216 last_cand = 3;
217 pmv[3] = mbs[rpos].mvs[rz];
218 psad[3] = mbs[rpos].sad8[rz];
219 } else {
220 pmv[3] = zeroMV;
221 psad[3] = MV_MAX_ERROR;
222 }
223
224 /* original pmvdata() compatibility hack */
225 if (x == 0 && y == 0) {
226 pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV;
227 psad[0] = 0;
228 psad[1] = psad[2] = psad[3] = MV_MAX_ERROR;
229 return;
230 }
231
232 /* if only one valid candidate preictor, the invalid candiates are set to the canidate */
233 if (num_cand == 1) {
234 pmv[0] = pmv[last_cand];
235 psad[0] = psad[last_cand];
236 return;
237 }
238
239 if ((MVequal(pmv[1], pmv[2])) && (MVequal(pmv[1], pmv[3]))) {
240 pmv[0] = pmv[1];
241 psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]);
242 return;
243 }
244
245 /* set median, minimum */
246
247 pmv[0].x =
248 MIN(MAX(pmv[1].x, pmv[2].x),
249 MIN(MAX(pmv[2].x, pmv[3].x), MAX(pmv[1].x, pmv[3].x)));
250 pmv[0].y =
251 MIN(MAX(pmv[1].y, pmv[2].y),
252 MIN(MAX(pmv[2].y, pmv[3].y), MAX(pmv[1].y, pmv[3].y)));
253
254 psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]);
255
256 }
257
258
259 static void
ModeDecision_SAD(SearchData * const Data,MACROBLOCK * const pMB,const MACROBLOCK * const pMBs,const int x,const int y,const MBParam * const pParam,const uint32_t MotionFlags,const uint32_t VopFlags,const uint32_t VolFlags,const IMAGE * const pCurrent,const IMAGE * const pRef,const IMAGE * const vGMC,const int coding_type,const int skip_sad)260 ModeDecision_SAD(SearchData * const Data,
261 MACROBLOCK * const pMB,
262 const MACROBLOCK * const pMBs,
263 const int x, const int y,
264 const MBParam * const pParam,
265 const uint32_t MotionFlags,
266 const uint32_t VopFlags,
267 const uint32_t VolFlags,
268 const IMAGE * const pCurrent,
269 const IMAGE * const pRef,
270 const IMAGE * const vGMC,
271 const int coding_type,
272 const int skip_sad)
273 {
274 int mode = MODE_INTER;
275 int mcsel = 0;
276 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
277 const uint32_t iQuant = pMB->quant;
278
279 const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);
280
281 int sad;
282 int InterBias = MV16_INTER_BIAS;
283
284 pMB->mcsel = 0;
285
286 if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
287 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
288 mode = MODE_INTER;
289 sad = Data->iMinSAD[0];
290 } else {
291 mode = MODE_INTER4V;
292 sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
293 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
294 Data->iMinSAD[0] = sad;
295 }
296
297 /* final skip decision, a.k.a. "the vector you found, really that good?" */
298 if (skip_possible && (skip_sad < (int)iQuant * MAX_SAD00_FOR_SKIP))
299 if ( (100*skip_sad)/(pMB->sad16+1) < FINAL_SKIP_THRESH)
300 if (Data->chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant)) {
301 mode = MODE_NOT_CODED;
302 sad = 0;
303 }
304
305 /* mcsel */
306 if (coding_type == S_VOP) {
307
308 int32_t iSAD = sad16(Data->Cur,
309 vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
310
311 if (Data->chroma) {
312 iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
313 iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
314 }
315
316 if (iSAD <= sad) { /* mode decision GMC */
317 mode = MODE_INTER;
318 mcsel = 1;
319 sad = iSAD;
320 }
321 }
322
323 /* intra decision */
324
325 if (iQuant > 10) InterBias += 60 * (iQuant - 10); /* to make high quants work */
326 if (y != 0)
327 if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
328 if (x != 0)
329 if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
330
331 if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */
332
333 if (InterBias < sad) {
334 int32_t deviation = dev16(Data->Cur, Data->iEdgedWidth);
335 if (deviation < (sad - InterBias)) mode = MODE_INTRA;
336 }
337
338 pMB->cbp = 63;
339 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
340
341 if (mode == MODE_INTER && mcsel == 0) {
342 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
343
344 if(Data->qpel) {
345 pMB->qmvs[0] = pMB->qmvs[1]
346 = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
347 pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
348 pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
349 } else {
350 pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
351 pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
352 }
353
354 } else if (mode == MODE_INTER ) { /* but mcsel == 1 */
355
356 pMB->mcsel = 1;
357 if (Data->qpel) {
358 pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
359 pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
360 pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
361 } else
362 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
363
364 } else
365 if (mode == MODE_INTER4V) ; /* anything here? */
366 else /* INTRA, NOT_CODED */
367 ZeroMacroblockP(pMB, 0);
368
369 pMB->mode = mode;
370 }
371
372 static __inline void
PreparePredictionsP(VECTOR * const pmv,int x,int y,int iWcount,int iHcount,const MACROBLOCK * const prevMB)373 PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
374 int iHcount, const MACROBLOCK * const prevMB)
375 {
376
377 if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
378 pmv[5].x = EVEN(pmv[3].x);
379 pmv[5].y = EVEN(pmv[3].y);
380 } else pmv[5].x = pmv[5].y = 0;
381
382 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
383 else pmv[3].x = pmv[3].y = 0;
384
385 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
386 else pmv[4].x = pmv[4].y = 0;
387
388 /* [1] median prediction */
389 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
390
391 pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */
392
393 pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
394 pmv[2].y = EVEN(prevMB->mvs[0].y);
395
396 if ((x < iWcount-1) && (y < iHcount-1)) {
397 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
398 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
399 } else pmv[6].x = pmv[6].y = 0;
400 }
401
402 static void
Search8(SearchData * const OldData,const int x,const int y,const uint32_t MotionFlags,const MBParam * const pParam,MACROBLOCK * const pMB,const MACROBLOCK * const pMBs,const int block,SearchData * const Data,const int bound)403 Search8(SearchData * const OldData,
404 const int x, const int y,
405 const uint32_t MotionFlags,
406 const MBParam * const pParam,
407 MACROBLOCK * const pMB,
408 const MACROBLOCK * const pMBs,
409 const int block,
410 SearchData * const Data,
411 const int bound)
412 {
413 int i = 0;
414 VECTOR vbest_q; int32_t sbest_q;
415 *Data->iMinSAD = *(OldData->iMinSAD + 1 + block);
416 *Data->currentMV = *(OldData->currentMV + 1 + block);
417 *Data->currentQMV = *(OldData->currentQMV + 1 + block);
418
419 if(Data->qpel) {
420 Data->predMV = get_qpmv2(pMBs, pParam->mb_width, bound, x/2, y/2, block);
421 if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
422 Data->predMV, Data->iFcode, 0);
423 } else {
424 Data->predMV = get_pmv2(pMBs, pParam->mb_width, bound, x/2, y/2, block);
425 if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
426 Data->predMV, Data->iFcode, 0);
427 }
428
429 *(Data->iMinSAD) += (Data->lambda8 * i);
430
431 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
432
433 vbest_q = Data->currentQMV[0];
434 sbest_q = Data->iMinSAD[0];
435
436
437 Data->RefP[0] = OldData->RefP[0] + 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
438 Data->RefP[1] = OldData->RefP[1] + 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
439 Data->RefP[2] = OldData->RefP[2] + 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
440 Data->RefP[3] = OldData->RefP[3] + 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
441
442 Data->Cur = OldData->Cur + 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
443 Data->qpel_precision = 0;
444
445 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
446 pParam->width, pParam->height, Data->iFcode - Data->qpel, 1);
447
448 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) {
449
450 MainSearchFunc *MainSearchPtr;
451 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = xvid_me_SquareSearch;
452 else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = xvid_me_AdvDiamondSearch;
453 else MainSearchPtr = xvid_me_DiamondSearch;
454
455 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate8);
456 }
457
458 if(!Data->qpel) {
459 /* halfpel mode */
460 if (MotionFlags & XVID_ME_HALFPELREFINE8)
461 /* perform halfpel refine of current best vector */
462 xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate8, 0);
463 } else {
464 /* qpel mode */
465 Data->currentQMV->x = 2*Data->currentMV->x;
466 Data->currentQMV->y = 2*Data->currentMV->y;
467
468 if(MotionFlags & XVID_ME_FASTREFINE8) {
469 /* fast */
470 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
471 pParam->width, pParam->height, Data->iFcode, 2);
472 FullRefine_Fast(Data, CheckCandidate8, 0);
473 } else if(MotionFlags & XVID_ME_QUARTERPELREFINE8) {
474 /* full */
475 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
476 xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate8, 0); /* hpel part */
477 Data->currentQMV->x = 2*Data->currentMV->x;
478 Data->currentQMV->y = 2*Data->currentMV->y;
479 }
480
481 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
482 pParam->width, pParam->height, Data->iFcode, 2);
483 Data->qpel_precision = 1;
484
485 xvid_me_SubpelRefine(Data->currentQMV[0], Data, CheckCandidate8, 0); /* qpel part */
486 }
487 }
488
489 if (sbest_q <= Data->iMinSAD[0]) /* we have not found a better match */
490 Data->currentQMV[0] = vbest_q;
491
492 }
493
494 if(Data->qpel) {
495 pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
496 pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
497 pMB->qmvs[block] = *Data->currentQMV;
498 } else {
499 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
500 pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
501 }
502
503 *(OldData->iMinSAD + 1 + block) = *Data->iMinSAD;
504 *(OldData->currentMV + 1 + block) = *Data->currentMV;
505 *(OldData->currentQMV + 1 + block) = *Data->currentQMV;
506
507 pMB->mvs[block] = *Data->currentMV;
508 pMB->sad8[block] = 4 * *Data->iMinSAD;
509 }
510
511
512
513 static void
SearchP(const IMAGE * const pRef,const uint8_t * const pRefH,const uint8_t * const pRefV,const uint8_t * const pRefHV,const IMAGE * const pCur,const int x,const int y,const uint32_t MotionFlags,const uint32_t VopFlags,SearchData * const Data,const MBParam * const pParam,const MACROBLOCK * const pMBs,const MACROBLOCK * const prevMBs,MACROBLOCK * const pMB,const int bound)514 SearchP(const IMAGE * const pRef,
515 const uint8_t * const pRefH,
516 const uint8_t * const pRefV,
517 const uint8_t * const pRefHV,
518 const IMAGE * const pCur,
519 const int x,
520 const int y,
521 const uint32_t MotionFlags,
522 const uint32_t VopFlags,
523 SearchData * const Data,
524 const MBParam * const pParam,
525 const MACROBLOCK * const pMBs,
526 const MACROBLOCK * const prevMBs,
527 MACROBLOCK * const pMB,
528 const int bound)
529 {
530
531 int i, threshA;
532 VECTOR pmv[7];
533 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
534 CheckFunc * CheckCandidate;
535
536 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
537 pParam->width, pParam->height, Data->iFcode - Data->qpel, 1);
538
539 get_pmvdata2(pMBs, pParam->mb_width, bound, x, y, pmv, Data->temp);
540
541 Data->chromaX = Data->chromaY = 0; /* chroma-sad cache */
542 Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16;
543 Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8;
544 Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8;
545
546 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
547 Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
548 Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
549 Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
550 Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
551 Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
552
553 Data->lambda16 = xvid_me_lambda_vec16[pMB->quant];
554 Data->lambda8 = xvid_me_lambda_vec8[pMB->quant];
555 Data->qpel_precision = 0;
556 Data->dir = 0;
557
558 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
559
560 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, bound, x, y, 0);
561 else Data->predMV = pmv[0];
562
563 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0);
564 Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i);
565 Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i);
566 Data->iMinSAD[2] = pMB->sad8[1];
567 Data->iMinSAD[3] = pMB->sad8[2];
568 Data->iMinSAD[4] = pMB->sad8[3];
569
570 if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) {
571 threshA = Data->temp[0]; /* that's where we keep this SAD atm */
572 if (threshA < 512) threshA = 512;
573 else if (threshA > 1024) threshA = 1024;
574 } else
575 threshA = 512;
576
577 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
578 prevMBs + x + y * pParam->mb_width);
579
580 if (inter4v) CheckCandidate = CheckCandidate16;
581 else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
582
583 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
584
585 for (i = 1; i < 7; i++)
586 if (!vector_repeats(pmv, i)) {
587 CheckCandidate(pmv[i].x, pmv[i].y, Data, i);
588 if (Data->iMinSAD[0] <= threshA) { i++; break; }
589 }
590
591 if ((Data->iMinSAD[0] <= threshA) ||
592 (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
593 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
594 inter4v = 0;
595 else {
596
597 MainSearchFunc * MainSearchPtr;
598 int mask = make_mask(pmv, i, Data->dir); /* all vectors pmv[0..i-1] have been checked */
599
600 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch;
601 else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch;
602 else MainSearchPtr = xvid_me_DiamondSearch;
603
604 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate);
605
606 /* extended search, diamond starting in 0,0 and in prediction.
607 note that this search is/might be done in halfpel positions,
608 which makes it more different than the diamond above */
609
610 if (MotionFlags & XVID_ME_EXTSEARCH16) {
611 int32_t bSAD;
612 VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
613 if (Data->qpel) {
614 startMV.x /= 2;
615 startMV.y /= 2;
616 }
617 if (!(MVequal(startMV, backupMV))) {
618 bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
619
620 CheckCandidate(startMV.x, startMV.y, Data, 255);
621 xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate);
622 if (bSAD < Data->iMinSAD[0]) {
623 Data->currentMV[0] = backupMV;
624 Data->iMinSAD[0] = bSAD;
625 }
626 }
627
628 backupMV = Data->currentMV[0];
629 startMV.x = startMV.y = 1;
630 if (!(MVequal(startMV, backupMV))) {
631 bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
632
633 CheckCandidate(startMV.x, startMV.y, Data, 255);
634 xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate);
635 if (bSAD < Data->iMinSAD[0]) {
636 Data->currentMV[0] = backupMV;
637 Data->iMinSAD[0] = bSAD;
638 }
639 }
640 }
641 }
642
643
644 if(!Data->qpel) {
645 /* halfpel mode */
646 if (MotionFlags & XVID_ME_HALFPELREFINE16)
647 xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate, 0);
648 } else {
649 /* qpel mode */
650
651 for(i = 0; i < 5; i++) {
652 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
653 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
654 }
655 if(MotionFlags & XVID_ME_FASTREFINE16 && MotionFlags & XVID_ME_QUARTERPELREFINE16) {
656 /* fast */
657 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
658 pParam->width, pParam->height, Data->iFcode, 2);
659 FullRefine_Fast(Data, CheckCandidate, 0);
660 } else {
661 if(MotionFlags & (XVID_ME_QUARTERPELREFINE16 | XVID_ME_QUARTERPELREFINE16_RD)) {
662 /* full */
663 if (MotionFlags & XVID_ME_HALFPELREFINE16) {
664 xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate, 0); /* hpel part */
665 for(i = 0; i < 5; i++) {
666 Data->currentQMV[i].x = 2 * Data->currentMV[i].x;
667 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
668 }
669 }
670 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
671 pParam->width, pParam->height, Data->iFcode, 2);
672 Data->qpel_precision = 1;
673 if(MotionFlags & XVID_ME_QUARTERPELREFINE16)
674 xvid_me_SubpelRefine(Data->currentQMV[0], Data, CheckCandidate, 0); /* qpel part */
675 }
676 }
677 }
678
679 if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30 * ((MotionFlags & XVID_ME_FASTREFINE16) ? 8 : 1))
680 inter4v = 0;
681
682 if (inter4v) {
683 SearchData Data8;
684 memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
685
686 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8, bound);
687 Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8, bound);
688 Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8, bound);
689 Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8, bound);
690
691 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) {
692 /* chroma is only used for comparison to INTER. if the comparison will be done in RD domain, it will not be used */
693 int sumx = 0, sumy = 0;
694
695 if (Data->qpel)
696 for (i = 1; i < 5; i++) {
697 sumx += Data->currentQMV[i].x/2;
698 sumy += Data->currentQMV[i].y/2;
699 }
700 else
701 for (i = 1; i < 5; i++) {
702 sumx += Data->currentMV[i].x;
703 sumy += Data->currentMV[i].y;
704 }
705
706 Data->iMinSAD[1] += xvid_me_ChromaSAD((sumx >> 3) + roundtab_76[sumx & 0xf],
707 (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
708 }
709 } else Data->iMinSAD[1] = 4096*256;
710 }
711
712 static int
InitialSkipDecisionP(int sad00,const MBParam * pParam,const FRAMEINFO * current,MACROBLOCK * pMB,const MACROBLOCK * prevMB,int x,int y,const SearchData * Data,const IMAGE * const pGMC,const IMAGE * const pCurrent,const IMAGE * const pRef,const uint32_t MotionFlags,const int bound)713 InitialSkipDecisionP(int sad00,
714 const MBParam * pParam,
715 const FRAMEINFO * current,
716 MACROBLOCK * pMB,
717 const MACROBLOCK * prevMB,
718 int x, int y,
719 const SearchData * Data,
720 const IMAGE * const pGMC,
721 const IMAGE * const pCurrent,
722 const IMAGE * const pRef,
723 const uint32_t MotionFlags,
724 const int bound)
725 {
726 const unsigned int iEdgedWidth = pParam->edged_width;
727
728 int skip_thresh = INITIAL_SKIP_THRESH * \
729 (current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1);
730 int stat_thresh = 0;
731
732 /* initial skip decision */
733 if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */
734 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
735 if (Data->chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant)) {
736 ZeroMacroblockP(pMB, sad00);
737 pMB->mode = MODE_NOT_CODED;
738 return 1;
739 }
740 }
741
742 if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) {
743 VECTOR *cmpMV;
744 VECTOR staticMV = { 0, 0 };
745 const MACROBLOCK * pMBs = current->mbs;
746
747 if (current->coding_type == S_VOP)
748 cmpMV = &pMB->amv;
749 else
750 cmpMV = &staticMV;
751
752 if(x > 0 && y > 0 && x < (int) pParam->mb_width) {
753 if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], *cmpMV) &&
754 MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], *cmpMV) &&
755 MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], *cmpMV) &&
756 MVequal(prevMB->mvs[0], *cmpMV)) {
757 stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
758 MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
759 MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
760 prevMB->sad16)));
761 } else {
762 stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
763 MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
764 MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
765 prevMB->sad16)));
766 }
767 }
768 }
769
770 /* favorize (0,0) or global vector for cartoons */
771 if (current->vop_flags & XVID_VOP_CARTOON) {
772 if (current->coding_type == S_VOP) {
773 int32_t iSAD = sad16(pCurrent->y + (x + y * iEdgedWidth) * 16,
774 pGMC->y + 16*y*iEdgedWidth + 16*x, iEdgedWidth, 65536);
775
776 if (Data->chroma) {
777 iSAD += sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, pGMC->u + 8*y*(iEdgedWidth/2) + 8*x, iEdgedWidth/2);
778 iSAD += sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, pGMC->v + 8*y*(iEdgedWidth/2) + 8*x, iEdgedWidth/2);
779 }
780
781 if (iSAD <= stat_thresh) { /* mode decision GMC */
782 pMB->mode = MODE_INTER;
783 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = iSAD;
784 pMB->mcsel = 1;
785 if (Data->qpel) {
786 pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
787 pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
788 pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
789 } else
790 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
791
792 return 1;
793 }
794 }
795 else if (sad00 < stat_thresh) {
796 VECTOR predMV;
797 if (Data->qpel)
798 predMV = get_qpmv2(current->mbs, pParam->mb_width, bound, x, y, 0);
799 else
800 predMV = get_pmv2(current->mbs, pParam->mb_width, bound, x, y, 0);
801
802 ZeroMacroblockP(pMB, sad00);
803 pMB->cbp = 0x3f;
804 pMB->pmvs[0].x = - predMV.x;
805 pMB->pmvs[0].y = - predMV.y;
806 return 1;
807 }
808 }
809
810 return 0;
811 }
812
813 static __inline uint32_t
MakeGoodMotionFlags(const uint32_t MotionFlags,const uint32_t VopFlags,const uint32_t VolFlags)814 MakeGoodMotionFlags(const uint32_t MotionFlags, const uint32_t VopFlags, const uint32_t VolFlags)
815 {
816 uint32_t Flags = MotionFlags;
817
818 if (!(VopFlags & XVID_VOP_MODEDECISION_RD))
819 Flags &= ~(XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD+XVID_ME_EXTSEARCH_RD);
820
821 if (Flags & XVID_ME_EXTSEARCH_RD)
822 Flags |= XVID_ME_HALFPELREFINE16_RD;
823
824 if (Flags & XVID_ME_EXTSEARCH_RD && MotionFlags & XVID_ME_EXTSEARCH8)
825 Flags |= XVID_ME_HALFPELREFINE8_RD;
826
827 if (Flags & XVID_ME_HALFPELREFINE16_RD)
828 Flags |= XVID_ME_QUARTERPELREFINE16_RD;
829
830 if (Flags & XVID_ME_HALFPELREFINE8_RD) {
831 Flags |= XVID_ME_QUARTERPELREFINE8_RD;
832 Flags &= ~XVID_ME_HALFPELREFINE8;
833 }
834
835 if (Flags & XVID_ME_QUARTERPELREFINE8_RD)
836 Flags &= ~XVID_ME_QUARTERPELREFINE8;
837
838 if (Flags & XVID_ME_QUARTERPELREFINE16_RD)
839 Flags &= ~XVID_ME_QUARTERPELREFINE16;
840
841 if (!(VolFlags & XVID_VOL_QUARTERPEL))
842 Flags &= ~(XVID_ME_QUARTERPELREFINE16+XVID_ME_QUARTERPELREFINE8+XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD);
843
844 if (!(VopFlags & XVID_VOP_HALFPEL))
845 Flags &= ~(XVID_ME_EXTSEARCH16+XVID_ME_HALFPELREFINE16+XVID_ME_HALFPELREFINE8+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD);
846
847 if (VopFlags & XVID_VOP_GREYSCALE)
848 Flags &= ~(XVID_ME_CHROMA_PVOP + XVID_ME_CHROMA_BVOP);
849
850 if (Flags & XVID_ME_FASTREFINE8)
851 Flags &= ~XVID_ME_HALFPELREFINE8_RD;
852
853 if (Flags & XVID_ME_FASTREFINE16)
854 Flags &= ~XVID_ME_HALFPELREFINE16_RD;
855
856 return Flags;
857 }
858
859 static __inline void
motionStatsPVOP(int * const MVmax,int * const mvCount,int * const mvSum,const MACROBLOCK * const pMB,const int qpel)860 motionStatsPVOP(int * const MVmax, int * const mvCount, int * const mvSum,
861 const MACROBLOCK * const pMB, const int qpel)
862 {
863 const VECTOR * const mv = qpel ? pMB->qmvs : pMB->mvs;
864 int i;
865 int max = *MVmax;
866
867 switch (pMB->mode) {
868 case MODE_INTER4V:
869 *mvCount += 3;
870 for(i = 3; i; i--) {
871 if (mv[i].x > max) max = mv[i].x;
872 else if (-mv[i].x - 1 > max) max = -mv[i].x - 1;
873 *mvSum += mv[i].x * mv[i].x;
874 if (mv[i].y > max) max = mv[i].y;
875 else if (-mv[i].y - 1 > max) max = -mv[i].y - 1;
876 *mvSum += mv[i].y * mv[i].y;
877 }
878 case MODE_INTER:
879 (*mvCount)++;
880 *mvSum += mv[0].x * mv[0].x;
881 *mvSum += mv[0].y * mv[0].y;
882 if (mv[0].x > max) max = mv[0].x;
883 else if (-mv[0].x - 1 > max) max = -mv[0].x - 1;
884 if (mv[0].y > max) max = mv[0].y;
885 else if (-mv[0].y - 1 > max) max = -mv[0].y - 1;
886 *MVmax = max;
887 default:
888 break;
889 }
890 }
891
892 void
MotionEstimation(MBParam * const pParam,FRAMEINFO * const current,FRAMEINFO * const reference,const IMAGE * const pRefH,const IMAGE * const pRefV,const IMAGE * const pRefHV,const IMAGE * const pGMC,const uint32_t iLimit,const int num_slices)893 MotionEstimation(MBParam * const pParam,
894 FRAMEINFO * const current,
895 FRAMEINFO * const reference,
896 const IMAGE * const pRefH,
897 const IMAGE * const pRefV,
898 const IMAGE * const pRefHV,
899 const IMAGE * const pGMC,
900 const uint32_t iLimit,
901 const int num_slices)
902 {
903 MACROBLOCK *const pMBs = current->mbs;
904 const IMAGE *const pCurrent = ¤t->image;
905 const IMAGE *const pRef = &reference->image;
906
907 const uint32_t mb_width = pParam->mb_width;
908 const uint32_t mb_height = pParam->mb_height;
909 const uint32_t iEdgedWidth = pParam->edged_width;
910 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
911 int bound = 0;
912 int MVmax = 0, mvSum = 0, mvCount = 0;
913
914 uint32_t x, y;
915 int sad00;
916 int block = 0;
917
918 /* some pre-initialized thingies for SearchP */
919 DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
920 SearchData Data;
921 memset(&Data, 0, sizeof(SearchData));
922 Data.iEdgedWidth = iEdgedWidth;
923 Data.iFcode = current->fcode;
924 Data.rounding = pParam->m_rounding_type;
925 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
926 Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
927 Data.dctSpace = dct_space;
928 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
929 Data.mpeg_quant_matrices = pParam->mpeg_quant_matrices;
930
931 Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
932 if (sadInit) (*sadInit) ();
933
934 for (y = 0; y < mb_height; y++) {
935
936 bound = mb_width * ((((y*num_slices) / mb_height) * mb_height + (num_slices-1)) / num_slices);
937
938 for (x = 0; x < mb_width; x++) {
939 MACROBLOCK *pMB = &pMBs[block];
940 MACROBLOCK *prevMB = &reference->mbs[block];
941 int skip;
942 block++;
943
944 pMB->sad16 =
945 sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
946 pRef->y + (x + y * iEdgedWidth) * 16,
947 pParam->edged_width, pMB->sad8);
948
949 sad00 = 4*MAX(MAX(pMB->sad8[0], pMB->sad8[1]), MAX(pMB->sad8[2], pMB->sad8[3]));
950
951 if (Data.chroma) {
952 Data.chromaSAD = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
953 pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
954 + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
955 pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
956 pMB->sad16 += Data.chromaSAD;
957 sad00 += Data.chromaSAD;
958 }
959
960 skip = InitialSkipDecisionP(sad00, pParam, current, pMB, prevMB, x, y, &Data, pGMC,
961 pCurrent, pRef, MotionFlags, bound);
962 if (skip) continue;
963
964 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
965 y, MotionFlags, current->vop_flags,
966 &Data, pParam, pMBs, reference->mbs, pMB, bound);
967
968 if (current->vop_flags & XVID_VOP_MODEDECISION_RD)
969 xvid_me_ModeDecision_RD(&Data, pMB, pMBs, x, y, pParam,
970 MotionFlags, current->vop_flags, current->vol_flags,
971 pCurrent, pRef, pGMC, current->coding_type, bound);
972
973 else if (current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD)
974 xvid_me_ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam,
975 MotionFlags, current->vop_flags, current->vol_flags,
976 pCurrent, pRef, pGMC, current->coding_type, bound);
977 else
978 ModeDecision_SAD(&Data, pMB, pMBs, x, y, pParam,
979 MotionFlags, current->vop_flags, current->vol_flags,
980 pCurrent, pRef, pGMC, current->coding_type, sad00);
981
982
983 motionStatsPVOP(&MVmax, &mvCount, &mvSum, pMB, Data.qpel);
984 }
985 }
986
987 current->fcode = getMinFcode(MVmax);
988 current->sStat.iMvSum = mvSum;
989 current->sStat.iMvCount = mvCount;
990 }
991
992 void
MotionEstimateSMP(SMPData * h)993 MotionEstimateSMP(SMPData * h)
994 {
995 Encoder *pEnc = (Encoder *) h->pEnc;
996
997 const MBParam * const pParam = &pEnc->mbParam;
998 const FRAMEINFO * const current = pEnc->current;
999 const FRAMEINFO * const reference = pEnc->reference;
1000 const IMAGE * const pRefH = &pEnc->vInterH;
1001 const IMAGE * const pRefV = &pEnc->vInterV;
1002 const IMAGE * const pRefHV = &pEnc->vInterHV;
1003 const IMAGE * const pGMC = &pEnc->vGMC;
1004 uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags,
1005 current->vop_flags,
1006 current->vol_flags);
1007
1008 MACROBLOCK *const pMBs = current->mbs;
1009 const IMAGE *const pCurrent = ¤t->image;
1010 const IMAGE *const pRef = &reference->image;
1011
1012 const int mb_width = pParam->mb_width;
1013 const int mb_height = pParam->mb_height;
1014 const uint32_t iEdgedWidth = pParam->edged_width;
1015 int bound = 0;
1016 int num_slices = pEnc->num_slices;
1017 int y_step = h->y_step;
1018 int y_row = h->y_row;
1019 int start_y = h->start_y;
1020 int stop_y = h->stop_y;
1021 int MVmax = 0, mvSum = 0, mvCount = 0;
1022
1023 int x, y;
1024 int sad00;
1025 int block = (start_y+y_row)*mb_width;
1026 int * complete_count_self = h->complete_count_self;
1027 const volatile int * complete_count_above = h->complete_count_above;
1028 int max_mbs;
1029 int current_mb = 0;
1030
1031 /* some pre-initialized thingies for SearchP */
1032 DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1033 SearchData Data;
1034 memset(&Data, 0, sizeof(SearchData));
1035 Data.iEdgedWidth = iEdgedWidth;
1036 Data.iFcode = current->fcode;
1037 Data.rounding = pParam->m_rounding_type;
1038 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1039 Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
1040 Data.dctSpace = dct_space;
1041 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1042 Data.mpeg_quant_matrices = pParam->mpeg_quant_matrices;
1043
1044 /* todo: sort out temp memory space */
1045 Data.RefQ = h->RefQ;
1046 if (sadInit) (*sadInit) ();
1047
1048 max_mbs = 0;
1049
1050 for (y = (start_y + y_row); y < stop_y; y += y_step) {
1051 bound = mb_width * ((((y*num_slices) / mb_height) * mb_height + (num_slices-1)) / num_slices);
1052
1053 if (y == start_y) max_mbs = mb_width; /* we can process all blocks of the first row */
1054
1055 for (x = 0; x < mb_width; x++) {
1056
1057 MACROBLOCK *pMB, *prevMB;
1058 int skip;
1059
1060 if (current_mb >= max_mbs) {
1061 /* we ME-ed all macroblocks we safely could. grab next portion */
1062 int above_count = *complete_count_above; /* sync point */
1063 if (above_count == mb_width) {
1064 /* full line above is ready */
1065 above_count = mb_width+1;
1066 if (y < (stop_y-y_step)) {
1067 /* this is not last line, grab a portion of MBs from the next line too */
1068 above_count += MAX(0, complete_count_above[1] - 1);
1069 }
1070 }
1071
1072 max_mbs = current_mb + above_count - x - 1;
1073
1074 if (current_mb >= max_mbs) {
1075 /* current workload is zero */
1076 x--;
1077 sched_yield();
1078 continue;
1079 }
1080 }
1081
1082 pMB = &pMBs[block];
1083 prevMB = &reference->mbs[block];
1084
1085 pMB->sad16 =
1086 sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1087 pRef->y + (x + y * iEdgedWidth) * 16,
1088 pParam->edged_width, pMB->sad8);
1089
1090 sad00 = 4*MAX(MAX(pMB->sad8[0], pMB->sad8[1]), MAX(pMB->sad8[2], pMB->sad8[3]));
1091
1092 if (Data.chroma) {
1093 Data.chromaSAD = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1094 pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1095 + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1096 pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1097 pMB->sad16 += Data.chromaSAD;
1098 sad00 += Data.chromaSAD;
1099 }
1100
1101 skip = InitialSkipDecisionP(sad00, pParam, current, pMB, prevMB, x, y, &Data, pGMC,
1102 pCurrent, pRef, MotionFlags, bound);
1103
1104 if (skip) {
1105 current_mb++;
1106 block++;
1107 *complete_count_self = x+1;
1108 continue;
1109 }
1110
1111 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1112 y, MotionFlags, current->vop_flags,
1113 &Data, pParam, pMBs, reference->mbs, pMB, bound);
1114
1115 if (current->vop_flags & XVID_VOP_MODEDECISION_RD)
1116 xvid_me_ModeDecision_RD(&Data, pMB, pMBs, x, y, pParam,
1117 MotionFlags, current->vop_flags, current->vol_flags,
1118 pCurrent, pRef, pGMC, current->coding_type, bound);
1119
1120 else if (current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD)
1121 xvid_me_ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam,
1122 MotionFlags, current->vop_flags, current->vol_flags,
1123 pCurrent, pRef, pGMC, current->coding_type, bound);
1124 else
1125 ModeDecision_SAD(&Data, pMB, pMBs, x, y, pParam,
1126 MotionFlags, current->vop_flags, current->vol_flags,
1127 pCurrent, pRef, pGMC, current->coding_type, sad00);
1128
1129 *complete_count_self = x+1;
1130
1131 current_mb++;
1132 block++;
1133
1134 motionStatsPVOP(&MVmax, &mvCount, &mvSum, pMB, Data.qpel);
1135
1136 }
1137
1138 block += (y_step-1)*pParam->mb_width;
1139
1140 complete_count_self++;
1141 complete_count_above++;
1142 }
1143
1144 h->minfcode = getMinFcode(MVmax);
1145
1146 h->MVmax = MVmax;
1147 h->mvSum = mvSum;
1148 h->mvCount = mvCount;
1149 }
1150
1151
1152