| 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 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 |
|
* $Id$ |
| 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 |
|
|
| 43 |
|
static const int xvid_me_lambda_vec8[32] = |
| 44 |
|
{ 0 ,(int)(1.00235 * NEIGH_TEND_8X8 + 0.5), |
| 45 |
|
(int)(1.15582 + NEIGH_TEND_8X8 + 0.5), (int)(1.31976*NEIGH_TEND_8X8 + 0.5), |
| 46 |
|
(int)(1.49591*NEIGH_TEND_8X8 + 0.5), (int)(1.68601*NEIGH_TEND_8X8 + 0.5), |
| 47 |
|
(int)(1.89187*NEIGH_TEND_8X8 + 0.5), (int)(2.11542*NEIGH_TEND_8X8 + 0.5), |
| 48 |
|
(int)(2.35878*NEIGH_TEND_8X8 + 0.5), (int)(2.62429*NEIGH_TEND_8X8 + 0.5), |
| 49 |
|
(int)(2.91455*NEIGH_TEND_8X8 + 0.5), (int)(3.23253*NEIGH_TEND_8X8 + 0.5), |
| 50 |
|
(int)(3.58158*NEIGH_TEND_8X8 + 0.5), (int)(3.96555*NEIGH_TEND_8X8 + 0.5), |
| 51 |
|
(int)(4.38887*NEIGH_TEND_8X8 + 0.5), (int)(4.85673*NEIGH_TEND_8X8 + 0.5), |
| 52 |
|
(int)(5.37519*NEIGH_TEND_8X8 + 0.5), (int)(5.95144*NEIGH_TEND_8X8 + 0.5), |
| 53 |
|
(int)(6.59408*NEIGH_TEND_8X8 + 0.5), (int)(7.31349*NEIGH_TEND_8X8 + 0.5), |
| 54 |
|
(int)(8.12242*NEIGH_TEND_8X8 + 0.5), (int)(9.03669*NEIGH_TEND_8X8 + 0.5), |
| 55 |
|
(int)(10.0763*NEIGH_TEND_8X8 + 0.5), (int)(11.2669*NEIGH_TEND_8X8 + 0.5), |
| 56 |
|
(int)(12.6426*NEIGH_TEND_8X8 + 0.5), (int)(14.2493*NEIGH_TEND_8X8 + 0.5), |
| 57 |
|
(int)(16.1512*NEIGH_TEND_8X8 + 0.5), (int)(18.442*NEIGH_TEND_8X8 + 0.5), |
| 58 |
|
(int)(21.2656*NEIGH_TEND_8X8 + 0.5), (int)(24.8580*NEIGH_TEND_8X8 + 0.5), |
| 59 |
|
(int)(29.6436*NEIGH_TEND_8X8 + 0.5), (int)(36.4949*NEIGH_TEND_8X8 + 0.5) |
| 60 |
|
}; |
| 61 |
|
|
| 62 |
|
static void |
| 63 |
|
CheckCandidate16(const int x, const int y, SearchData * const data, const unsigned int Direction) |
| 64 |
|
{ |
| 65 |
|
const uint8_t * Reference; |
| 66 |
|
int32_t sad; uint32_t t; |
| 67 |
|
|
| 68 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
| 69 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
| 70 |
|
|
| 71 |
|
Reference = GetReference(x, y, data); |
| 72 |
|
|
| 73 |
|
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
| 74 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel, 0); |
| 75 |
|
|
| 76 |
|
sad += (data->lambda16 * t * sad)>>10; |
| 77 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
| 78 |
|
|
| 79 |
|
if (data->chroma) { |
| 80 |
|
if (sad >= data->iMinSAD[0]) goto no16; |
| 81 |
|
sad += xvid_me_ChromaSAD((x >> 1) + roundtab_79[x & 0x3], |
| 82 |
|
(y >> 1) + roundtab_79[y & 0x3], data); |
| 83 |
|
} |
| 84 |
|
|
| 85 |
|
if (sad < data->iMinSAD[0]) { |
| 86 |
|
data->iMinSAD[0] = sad; |
| 87 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
| 88 |
|
data->dir = Direction; |
| 89 |
|
} |
| 90 |
|
|
| 91 |
|
no16: |
| 92 |
|
if (data->temp[0] < data->iMinSAD[1]) { |
| 93 |
|
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
| 94 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
| 95 |
|
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
| 96 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
| 97 |
|
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
| 98 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
| 99 |
|
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
| 100 |
|
} |
| 101 |
|
|
| 102 |
|
static void |
| 103 |
|
CheckCandidate16_qpel(const int x, const int y, SearchData * const data, const unsigned int Direction) |
| 104 |
|
{ |
| 105 |
|
const uint8_t *Reference; |
| 106 |
|
int32_t sad; uint32_t t; |
| 107 |
|
|
| 108 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
| 109 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
| 110 |
|
|
| 111 |
|
Reference = xvid_me_interpolate16x16qpel(x, y, 0, data); |
| 112 |
|
|
| 113 |
|
sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); |
| 114 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); |
| 115 |
|
|
| 116 |
|
sad += (data->lambda16 * t * sad)>>10; |
| 117 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
| 118 |
|
|
| 119 |
|
if (data->chroma && (sad < data->iMinSAD[0] || sad < data->iMinSAD2) ) |
| 120 |
|
sad += xvid_me_ChromaSAD(((x/2) >> 1) + roundtab_79[(x/2) & 0x3], |
| 121 |
|
((y/2) >> 1) + roundtab_79[(y/2) & 0x3], data); |
| 122 |
|
|
| 123 |
|
if (data->temp[0] < data->iMinSAD[1]) { |
| 124 |
|
data->iMinSAD[1] = data->temp[0]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } |
| 125 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
| 126 |
|
data->iMinSAD[2] = data->temp[1]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } |
| 127 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
| 128 |
|
data->iMinSAD[3] = data->temp[2]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } |
| 129 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
| 130 |
|
data->iMinSAD[4] = data->temp[3]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } |
| 131 |
|
|
| 132 |
|
if (sad < data->iMinSAD[0]) { |
| 133 |
|
data->iMinSAD2 = *(data->iMinSAD); |
| 134 |
|
data->currentQMV2.x = data->currentQMV->x; |
| 135 |
|
data->currentQMV2.y = data->currentQMV->y; |
| 136 |
|
|
| 137 |
|
data->iMinSAD[0] = sad; |
| 138 |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
| 139 |
|
} else if (sad < data->iMinSAD2) { |
| 140 |
|
data->iMinSAD2 = sad; |
| 141 |
|
data->currentQMV2.x = x; data->currentQMV2.y = y; |
| 142 |
|
} |
| 143 |
|
} |
| 144 |
|
|
| 145 |
|
static void |
| 146 |
|
CheckCandidate8(const int x, const int y, SearchData * const data, const unsigned int Direction) |
| 147 |
|
{ |
| 148 |
|
int32_t sad; uint32_t t; |
| 149 |
|
const uint8_t * Reference; |
| 150 |
|
VECTOR * current; |
| 151 |
|
|
| 152 |
|
if ( (x > data->max_dx) || (x < data->min_dx) |
| 153 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
| 154 |
|
|
| 155 |
|
if (!data->qpel_precision) { |
| 156 |
|
Reference = GetReference(x, y, data); |
| 157 |
|
current = data->currentMV; |
| 158 |
|
} else { /* x and y are in 1/4 precision */ |
| 159 |
|
Reference = xvid_me_interpolate8x8qpel(x, y, 0, 0, data); |
| 160 |
|
current = data->currentQMV; |
| 161 |
|
} |
| 162 |
|
|
| 163 |
|
sad = sad8(data->Cur, Reference, data->iEdgedWidth); |
| 164 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
| 165 |
|
|
| 166 |
|
sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; |
| 167 |
|
|
| 168 |
|
if (sad < *(data->iMinSAD)) { |
| 169 |
|
*(data->iMinSAD) = sad; |
| 170 |
|
current->x = x; current->y = y; |
| 171 |
|
data->dir = Direction; |
| 172 |
|
} |
| 173 |
|
} |
| 174 |
|
|
| 175 |
|
static void |
| 176 |
|
CheckCandidate32(const int x, const int y, SearchData * const data, const unsigned int Direction) |
| 177 |
|
{ |
| 178 |
|
uint32_t t; |
| 179 |
|
const uint8_t * Reference; |
| 180 |
|
int sad; |
| 181 |
|
|
| 182 |
|
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
| 183 |
|
(x > data->max_dx) || (x < data->min_dx) |
| 184 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
| 185 |
|
|
| 186 |
|
Reference = GetReference(x, y, data); |
| 187 |
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); |
| 188 |
|
|
| 189 |
|
sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp); |
| 190 |
|
|
| 191 |
|
sad += (data->lambda16 * t * sad) >> 10; |
| 192 |
|
data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; |
| 193 |
|
|
| 194 |
|
if (sad < data->iMinSAD[0]) { |
| 195 |
|
data->iMinSAD[0] = sad; |
| 196 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
| 197 |
|
data->dir = Direction; |
| 198 |
|
} |
| 199 |
|
|
| 200 |
|
if (data->temp[0] < data->iMinSAD[1]) { |
| 201 |
|
data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } |
| 202 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
| 203 |
|
data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } |
| 204 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
| 205 |
|
data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } |
| 206 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
| 207 |
|
data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } |
| 208 |
|
} |
| 209 |
|
|
| 210 |
|
static void |
| 211 |
|
SubpelRefine_Fast(SearchData * data, CheckFunc * CheckCandidate) |
| 212 |
|
{ |
| 213 |
|
/* Do a fast q-pel refinement */ |
| 214 |
|
VECTOR centerMV; |
| 215 |
|
VECTOR second_best; |
| 216 |
|
int best_sad = *data->iMinSAD; |
| 217 |
|
int xo, yo, xo2, yo2; |
| 218 |
|
int size = 2; |
| 219 |
|
data->iMinSAD2 = 0; |
| 220 |
|
|
| 221 |
|
/* check all halfpixel positions near our best halfpel position */ |
| 222 |
|
centerMV = *data->currentQMV; |
| 223 |
|
*data->iMinSAD = 256 * 4096; |
| 224 |
|
|
| 225 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - size, 0); |
| 226 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y - size, 0); |
| 227 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y, 0); |
| 228 |
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y + size, 0); |
| 229 |
|
|
| 230 |
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + size, 0); |
| 231 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y + size, 0); |
| 232 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y, 0); |
| 233 |
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y - size, 0); |
| 234 |
|
|
| 235 |
|
second_best = *data->currentQMV; |
| 236 |
|
|
| 237 |
|
/* after second_best has been found, go back to the vector we began with */ |
| 238 |
|
|
| 239 |
|
data->currentQMV[0] = centerMV; |
| 240 |
|
*data->iMinSAD = best_sad; |
| 241 |
|
|
| 242 |
|
xo = centerMV.x; |
| 243 |
|
yo = centerMV.y; |
| 244 |
|
xo2 = second_best.x; |
| 245 |
|
yo2 = second_best.y; |
| 246 |
|
|
| 247 |
|
data->iMinSAD2 = 256 * 4096; |
| 248 |
|
|
| 249 |
|
if (yo == yo2) { |
| 250 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); |
| 251 |
|
CHECK_CANDIDATE(xo, yo-1, 0); |
| 252 |
|
CHECK_CANDIDATE(xo, yo+1, 0); |
| 253 |
|
|
| 254 |
|
if(best_sad <= data->iMinSAD2) return; |
| 255 |
|
|
| 256 |
|
if(data->currentQMV[0].x == data->currentQMV2.x) { |
| 257 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo-1, 0); |
| 258 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo+1, 0); |
| 259 |
|
} else { |
| 260 |
|
CHECK_CANDIDATE((xo+xo2)>>1, |
| 261 |
|
(data->currentQMV[0].x == xo) ? data->currentQMV[0].y : data->currentQMV2.y, 0); |
| 262 |
|
} |
| 263 |
|
return; |
| 264 |
|
} |
| 265 |
|
|
| 266 |
|
if (xo == xo2) { |
| 267 |
|
CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); |
| 268 |
|
CHECK_CANDIDATE(xo-1, yo, 0); |
| 269 |
|
CHECK_CANDIDATE(xo+1, yo, 0); |
| 270 |
|
|
| 271 |
|
if(best_sad < data->iMinSAD2) return; |
| 272 |
|
|
| 273 |
|
if(data->currentQMV[0].y == data->currentQMV2.y) { |
| 274 |
|
CHECK_CANDIDATE(xo-1, (yo+yo2)>>1, 0); |
| 275 |
|
CHECK_CANDIDATE(xo+1, (yo+yo2)>>1, 0); |
| 276 |
|
} else { |
| 277 |
|
CHECK_CANDIDATE((data->currentQMV[0].y == yo) ? data->currentQMV[0].x : data->currentQMV2.x, (yo+yo2)>>1, 0); |
| 278 |
|
} |
| 279 |
|
return; |
| 280 |
|
} |
| 281 |
|
|
| 282 |
|
CHECK_CANDIDATE(xo, (yo+yo2)>>1, 0); |
| 283 |
|
CHECK_CANDIDATE((xo+xo2)>>1, yo, 0); |
| 284 |
|
|
| 285 |
|
if(best_sad <= data->iMinSAD2) return; |
| 286 |
|
|
| 287 |
|
CHECK_CANDIDATE((xo+xo2)>>1, (yo+yo2)>>1, 0); |
| 288 |
|
} |
| 289 |
|
|
| 290 |
|
int |
| 291 |
|
xvid_me_SkipDecisionP(const IMAGE * current, const IMAGE * reference, |
| 292 |
|
const int x, const int y, |
| 293 |
|
const uint32_t stride, const uint32_t iQuant, int rrv) |
| 294 |
|
{ |
| 295 |
|
int offset = (x + y*stride)*8; |
| 296 |
|
if(!rrv) { |
| 297 |
|
uint32_t sadC = sad8(current->u + offset, |
| 298 |
|
reference->u + offset, stride); |
| 299 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
| 300 |
|
sadC += sad8(current->v + offset, |
| 301 |
|
reference->v + offset, stride); |
| 302 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; |
| 303 |
|
return 1; |
| 304 |
|
|
| 305 |
|
} else { |
| 306 |
|
uint32_t sadC = sad16(current->u + 2*offset, |
| 307 |
|
reference->u + 2*offset, stride, 256*4096); |
| 308 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
| 309 |
|
sadC += sad16(current->v + 2*offset, |
| 310 |
|
reference->v + 2*offset, stride, 256*4096); |
| 311 |
|
if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; |
| 312 |
|
return 1; |
| 313 |
|
} |
| 314 |
|
} |
| 315 |
|
|
| 316 |
|
/* |
| 317 |
|
* pmv are filled with: |
| 318 |
|
* [0]: Median (or whatever is correct in a special case) |
| 319 |
|
* [1]: left neighbour |
| 320 |
|
* [2]: top neighbour |
| 321 |
|
* [3]: topright neighbour |
| 322 |
|
* psad are filled with: |
| 323 |
|
* [0]: minimum of [1] to [3] |
| 324 |
|
* [1]: left neighbour's SAD (NB:[1] to [3] are actually not needed) |
| 325 |
|
* [2]: top neighbour's SAD |
| 326 |
|
* [3]: topright neighbour's SAD |
| 327 |
|
*/ |
| 328 |
|
|
| 329 |
|
static __inline void |
| 330 |
|
get_pmvdata2(const MACROBLOCK * const mbs, |
| 331 |
|
const int mb_width, |
| 332 |
|
const int bound, |
| 333 |
|
const int x, |
| 334 |
|
const int y, |
| 335 |
|
VECTOR * const pmv, |
| 336 |
|
int32_t * const psad) |
| 337 |
|
{ |
| 338 |
|
int lx, ly, lz; /* left */ |
| 339 |
|
int tx, ty, tz; /* top */ |
| 340 |
|
int rx, ry, rz; /* top-right */ |
| 341 |
|
int lpos, tpos, rpos; |
| 342 |
|
int num_cand = 0, last_cand = 1; |
| 343 |
|
|
| 344 |
|
lx = x - 1; ly = y; lz = 1; |
| 345 |
|
tx = x; ty = y - 1; tz = 2; |
| 346 |
|
rx = x + 1; ry = y - 1; rz = 2; |
| 347 |
|
|
| 348 |
|
lpos = lx + ly * mb_width; |
| 349 |
|
rpos = rx + ry * mb_width; |
| 350 |
|
tpos = tx + ty * mb_width; |
| 351 |
|
|
| 352 |
|
if (lpos >= bound && lx >= 0) { |
| 353 |
|
num_cand++; |
| 354 |
|
last_cand = 1; |
| 355 |
|
pmv[1] = mbs[lpos].mvs[lz]; |
| 356 |
|
psad[1] = mbs[lpos].sad8[lz]; |
| 357 |
|
} else { |
| 358 |
|
pmv[1] = zeroMV; |
| 359 |
|
psad[1] = MV_MAX_ERROR; |
| 360 |
|
} |
| 361 |
|
|
| 362 |
|
if (tpos >= bound) { |
| 363 |
|
num_cand++; |
| 364 |
|
last_cand = 2; |
| 365 |
|
pmv[2]= mbs[tpos].mvs[tz]; |
| 366 |
|
psad[2] = mbs[tpos].sad8[tz]; |
| 367 |
|
} else { |
| 368 |
|
pmv[2] = zeroMV; |
| 369 |
|
psad[2] = MV_MAX_ERROR; |
| 370 |
|
} |
| 371 |
|
|
| 372 |
|
if (rpos >= bound && rx < mb_width) { |
| 373 |
|
num_cand++; |
| 374 |
|
last_cand = 3; |
| 375 |
|
pmv[3] = mbs[rpos].mvs[rz]; |
| 376 |
|
psad[3] = mbs[rpos].sad8[rz]; |
| 377 |
|
} else { |
| 378 |
|
pmv[3] = zeroMV; |
| 379 |
|
psad[3] = MV_MAX_ERROR; |
| 380 |
|
} |
| 381 |
|
|
| 382 |
|
/* original pmvdata() compatibility hack */ |
| 383 |
|
if (x == 0 && y == 0) { |
| 384 |
|
pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
| 385 |
|
psad[0] = 0; |
| 386 |
|
psad[1] = psad[2] = psad[3] = MV_MAX_ERROR; |
| 387 |
|
return; |
| 388 |
|
} |
| 389 |
|
|
| 390 |
|
/* if only one valid candidate preictor, the invalid candiates are set to the canidate */ |
| 391 |
|
if (num_cand == 1) { |
| 392 |
|
pmv[0] = pmv[last_cand]; |
| 393 |
|
psad[0] = psad[last_cand]; |
| 394 |
|
return; |
| 395 |
|
} |
| 396 |
|
|
| 397 |
|
if ((MVequal(pmv[1], pmv[2])) && (MVequal(pmv[1], pmv[3]))) { |
| 398 |
|
pmv[0] = pmv[1]; |
| 399 |
|
psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]); |
| 400 |
|
return; |
| 401 |
|
} |
| 402 |
|
|
| 403 |
|
/* set median, minimum */ |
| 404 |
|
|
| 405 |
|
pmv[0].x = |
| 406 |
|
MIN(MAX(pmv[1].x, pmv[2].x), |
| 407 |
|
MIN(MAX(pmv[2].x, pmv[3].x), MAX(pmv[1].x, pmv[3].x))); |
| 408 |
|
pmv[0].y = |
| 409 |
|
MIN(MAX(pmv[1].y, pmv[2].y), |
| 410 |
|
MIN(MAX(pmv[2].y, pmv[3].y), MAX(pmv[1].y, pmv[3].y))); |
| 411 |
|
|
| 412 |
|
psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]); |
| 413 |
|
|
| 414 |
|
} |
| 415 |
|
|
| 416 |
|
|
| 417 |
|
static void |
| 418 |
|
ModeDecision_SAD(SearchData * const Data, |
| 419 |
|
MACROBLOCK * const pMB, |
| 420 |
|
const MACROBLOCK * const pMBs, |
| 421 |
|
const int x, const int y, |
| 422 |
|
const MBParam * const pParam, |
| 423 |
|
const uint32_t MotionFlags, |
| 424 |
|
const uint32_t VopFlags, |
| 425 |
|
const uint32_t VolFlags, |
| 426 |
|
const IMAGE * const pCurrent, |
| 427 |
|
const IMAGE * const pRef, |
| 428 |
|
const IMAGE * const vGMC, |
| 429 |
|
const int coding_type) |
| 430 |
|
{ |
| 431 |
|
int mode = MODE_INTER; |
| 432 |
|
int mcsel = 0; |
| 433 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
| 434 |
|
const uint32_t iQuant = pMB->quant; |
| 435 |
|
|
| 436 |
|
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
| 437 |
|
|
| 438 |
|
int sad; |
| 439 |
|
int InterBias = MV16_INTER_BIAS; |
| 440 |
|
|
| 441 |
|
pMB->mcsel = 0; |
| 442 |
|
|
| 443 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
| 444 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
| 445 |
|
mode = MODE_INTER; |
| 446 |
|
sad = Data->iMinSAD[0]; |
| 447 |
|
} else { |
| 448 |
|
mode = MODE_INTER4V; |
| 449 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
| 450 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
| 451 |
|
Data->iMinSAD[0] = sad; |
| 452 |
|
} |
| 453 |
|
|
| 454 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
| 455 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
| 456 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
| 457 |
|
if (Data->chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
| 458 |
|
mode = MODE_NOT_CODED; |
| 459 |
|
sad = 0; |
| 460 |
|
} |
| 461 |
|
|
| 462 |
|
/* mcsel */ |
| 463 |
|
if (coding_type == S_VOP) { |
| 464 |
|
|
| 465 |
|
int32_t iSAD = sad16(Data->Cur, |
| 466 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
| 467 |
|
|
| 468 |
|
if (Data->chroma) { |
| 469 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
| 470 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
| 471 |
|
} |
| 472 |
|
|
| 473 |
|
if (iSAD <= sad) { /* mode decision GMC */ |
| 474 |
|
mode = MODE_INTER; |
| 475 |
|
mcsel = 1; |
| 476 |
|
sad = iSAD; |
| 477 |
|
} |
| 478 |
|
} |
| 479 |
|
|
| 480 |
|
/* intra decision */ |
| 481 |
|
|
| 482 |
|
if (iQuant > 10) InterBias += 60 * (iQuant - 10); /* to make high quants work */ |
| 483 |
|
if (y != 0) |
| 484 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
| 485 |
|
if (x != 0) |
| 486 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
| 487 |
|
|
| 488 |
|
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
| 489 |
|
if (Data->rrv) InterBias *= 4; |
| 490 |
|
|
| 491 |
|
if (InterBias < sad) { |
| 492 |
|
int32_t deviation; |
| 493 |
|
if (!Data->rrv) |
| 494 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
| 495 |
|
else |
| 496 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
| 497 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
| 498 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
| 499 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
| 500 |
|
|
| 501 |
|
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
| 502 |
|
} |
| 503 |
|
|
| 504 |
|
pMB->cbp = 63; |
| 505 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
| 506 |
|
|
| 507 |
|
if (Data->rrv) { |
| 508 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
| 509 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
| 510 |
|
} |
| 511 |
|
|
| 512 |
|
if (mode == MODE_INTER && mcsel == 0) { |
| 513 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
| 514 |
|
|
| 515 |
|
if(Data->qpel) { |
| 516 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
| 517 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
| 518 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
| 519 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
| 520 |
|
} else { |
| 521 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
| 522 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
| 523 |
|
} |
| 524 |
|
|
| 525 |
|
} else if (mode == MODE_INTER ) { /* but mcsel == 1 */ |
| 526 |
|
|
| 527 |
|
pMB->mcsel = 1; |
| 528 |
|
if (Data->qpel) { |
| 529 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
| 530 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
| 531 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
| 532 |
|
} else |
| 533 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
| 534 |
|
|
| 535 |
|
} else |
| 536 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
| 537 |
|
else /* INTRA, NOT_CODED */ |
| 538 |
|
ZeroMacroblockP(pMB, 0); |
| 539 |
|
|
| 540 |
|
pMB->mode = mode; |
| 541 |
|
} |
| 542 |
|
|
| 543 |
|
static __inline void |
| 544 |
|
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
| 545 |
|
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
| 546 |
|
{ |
| 547 |
|
/* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ |
| 548 |
|
if (rrv) { iWcount /= 2; iHcount /= 2; } |
| 549 |
|
|
| 550 |
|
if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
| 551 |
|
pmv[5].x = EVEN(pmv[3].x); |
| 552 |
|
pmv[5].y = EVEN(pmv[3].y); |
| 553 |
|
} else pmv[5].x = pmv[5].y = 0; |
| 554 |
|
|
| 555 |
|
if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */ |
| 556 |
|
else pmv[3].x = pmv[3].y = 0; |
| 557 |
|
|
| 558 |
|
if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */ |
| 559 |
|
else pmv[4].x = pmv[4].y = 0; |
| 560 |
|
|
| 561 |
|
/* [1] median prediction */ |
| 562 |
|
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
| 563 |
|
|
| 564 |
|
pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ |
| 565 |
|
|
| 566 |
|
pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
| 567 |
|
pmv[2].y = EVEN(prevMB->mvs[0].y); |
| 568 |
|
|
| 569 |
|
if ((x < iWcount-1) && (y < iHcount-1)) { |
| 570 |
|
pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ |
| 571 |
|
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
| 572 |
|
} else pmv[6].x = pmv[6].y = 0; |
| 573 |
|
|
| 574 |
|
if (rrv) { |
| 575 |
|
int i; |
| 576 |
|
for (i = 0; i < 7; i++) { |
| 577 |
|
pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); |
| 578 |
|
pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); |
| 579 |
|
} |
| 580 |
|
} |
| 581 |
|
} |
| 582 |
|
|
| 583 |
|
static void |
| 584 |
|
Search8(SearchData * const OldData, |
| 585 |
|
const int x, const int y, |
| 586 |
|
const uint32_t MotionFlags, |
| 587 |
|
const MBParam * const pParam, |
| 588 |
|
MACROBLOCK * const pMB, |
| 589 |
|
const MACROBLOCK * const pMBs, |
| 590 |
|
const int block, |
| 591 |
|
SearchData * const Data) |
| 592 |
|
{ |
| 593 |
|
int i = 0; |
| 594 |
|
CheckFunc * CheckCandidate; |
| 595 |
|
*Data->iMinSAD = *(OldData->iMinSAD + 1 + block); |
| 596 |
|
*Data->currentMV = *(OldData->currentMV + 1 + block); |
| 597 |
|
*Data->currentQMV = *(OldData->currentQMV + 1 + block); |
| 598 |
|
|
| 599 |
|
if(Data->qpel) { |
| 600 |
|
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
| 601 |
|
if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, |
| 602 |
|
Data->predMV, Data->iFcode, 0, 0); |
| 603 |
|
} else { |
| 604 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
| 605 |
|
if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, |
| 606 |
|
Data->predMV, Data->iFcode, 0, Data->rrv); |
| 607 |
|
} |
| 608 |
|
|
| 609 |
|
*(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; |
| 610 |
|
|
| 611 |
|
if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { |
| 612 |
|
|
| 613 |
|
if (Data->rrv) i = 16; else i = 8; |
| 614 |
|
|
| 615 |
|
Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
| 616 |
|
Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
| 617 |
|
Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
| 618 |
|
Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
| 619 |
|
|
| 620 |
|
Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); |
| 621 |
|
Data->qpel_precision = 0; |
| 622 |
|
|
| 623 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
| 624 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
| 625 |
|
|
| 626 |
|
if (!Data->rrv) CheckCandidate = CheckCandidate8; |
| 627 |
|
else CheckCandidate = CheckCandidate16no4v; |
| 628 |
|
|
| 629 |
|
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { |
| 630 |
|
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
| 631 |
|
|
| 632 |
|
MainSearchFunc *MainSearchPtr; |
| 633 |
|
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = xvid_me_SquareSearch; |
| 634 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = xvid_me_AdvDiamondSearch; |
| 635 |
|
else MainSearchPtr = xvid_me_DiamondSearch; |
| 636 |
|
|
| 637 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate); |
| 638 |
|
|
| 639 |
|
if(*(Data->iMinSAD) < temp_sad) { |
| 640 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
| 641 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
| 642 |
|
} |
| 643 |
|
} |
| 644 |
|
|
| 645 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
| 646 |
|
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
| 647 |
|
|
| 648 |
|
xvid_me_SubpelRefine(Data, CheckCandidate); /* perform halfpel refine of current best vector */ |
| 649 |
|
|
| 650 |
|
if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
| 651 |
|
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
| 652 |
|
Data->currentQMV->y = 2 * Data->currentMV->y; |
| 653 |
|
} |
| 654 |
|
} |
| 655 |
|
|
| 656 |
|
if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) { |
| 657 |
|
Data->qpel_precision = 1; |
| 658 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, |
| 659 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
| 660 |
|
xvid_me_SubpelRefine(Data, CheckCandidate); |
| 661 |
|
} |
| 662 |
|
} |
| 663 |
|
|
| 664 |
|
if (Data->rrv) { |
| 665 |
|
Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); |
| 666 |
|
Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); |
| 667 |
|
} |
| 668 |
|
|
| 669 |
|
if(Data->qpel) { |
| 670 |
|
pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; |
| 671 |
|
pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; |
| 672 |
|
pMB->qmvs[block] = *Data->currentQMV; |
| 673 |
|
} else { |
| 674 |
|
pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; |
| 675 |
|
pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; |
| 676 |
|
} |
| 677 |
|
|
| 678 |
|
*(OldData->iMinSAD + 1 + block) = *Data->iMinSAD; |
| 679 |
|
*(OldData->currentMV + 1 + block) = *Data->currentMV; |
| 680 |
|
*(OldData->currentQMV + 1 + block) = *Data->currentQMV; |
| 681 |
|
|
| 682 |
|
pMB->mvs[block] = *Data->currentMV; |
| 683 |
|
pMB->sad8[block] = 4 * *Data->iMinSAD; |
| 684 |
|
} |
| 685 |
|
|
| 686 |
|
|
| 687 |
|
|
| 688 |
|
static void |
| 689 |
|
SearchP(const IMAGE * const pRef, |
| 690 |
|
const uint8_t * const pRefH, |
| 691 |
|
const uint8_t * const pRefV, |
| 692 |
|
const uint8_t * const pRefHV, |
| 693 |
|
const IMAGE * const pCur, |
| 694 |
|
const int x, |
| 695 |
|
const int y, |
| 696 |
|
const uint32_t MotionFlags, |
| 697 |
|
const uint32_t VopFlags, |
| 698 |
|
SearchData * const Data, |
| 699 |
|
const MBParam * const pParam, |
| 700 |
|
const MACROBLOCK * const pMBs, |
| 701 |
|
const MACROBLOCK * const prevMBs, |
| 702 |
|
MACROBLOCK * const pMB) |
| 703 |
|
{ |
| 704 |
|
|
| 705 |
|
int i, threshA; |
| 706 |
|
VECTOR pmv[7]; |
| 707 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
| 708 |
|
CheckFunc * CheckCandidate; |
| 709 |
|
|
| 710 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
| 711 |
|
pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); |
| 712 |
|
|
| 713 |
|
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, pmv, Data->temp); |
| 714 |
|
|
| 715 |
|
Data->chromaX = Data->chromaY = 0; /* chroma-sad cache */ |
| 716 |
|
i = Data->rrv ? 2 : 1; |
| 717 |
|
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
| 718 |
|
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 719 |
|
Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 720 |
|
|
| 721 |
|
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; |
| 722 |
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; |
| 723 |
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; |
| 724 |
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; |
| 725 |
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 726 |
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
| 727 |
|
|
| 728 |
|
Data->lambda16 = xvid_me_lambda_vec16[pMB->quant]; |
| 729 |
|
Data->lambda8 = xvid_me_lambda_vec8[pMB->quant]; |
| 730 |
|
Data->qpel_precision = 0; |
| 731 |
|
Data->dir = 0; |
| 732 |
|
|
| 733 |
|
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
| 734 |
|
|
| 735 |
|
if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
| 736 |
|
else Data->predMV = pmv[0]; |
| 737 |
|
|
| 738 |
|
i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); |
| 739 |
|
Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); |
| 740 |
|
Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); |
| 741 |
|
Data->iMinSAD[2] = pMB->sad8[1]; |
| 742 |
|
Data->iMinSAD[3] = pMB->sad8[2]; |
| 743 |
|
Data->iMinSAD[4] = pMB->sad8[3]; |
| 744 |
|
|
| 745 |
|
if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { |
| 746 |
|
threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
| 747 |
|
if (threshA < 512) threshA = 512; |
| 748 |
|
else if (threshA > 1024) threshA = 1024; |
| 749 |
|
} else |
| 750 |
|
threshA = 512; |
| 751 |
|
|
| 752 |
|
PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, |
| 753 |
|
prevMBs + x + y * pParam->mb_width, Data->rrv); |
| 754 |
|
|
| 755 |
|
if (!Data->rrv) { |
| 756 |
|
if (inter4v) CheckCandidate = CheckCandidate16; |
| 757 |
|
else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
| 758 |
|
} else CheckCandidate = CheckCandidate32; |
| 759 |
|
|
| 760 |
|
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
| 761 |
|
|
| 762 |
|
for (i = 1; i < 7; i++) |
| 763 |
|
if (!vector_repeats(pmv, i)) { |
| 764 |
|
CheckCandidate(pmv[i].x, pmv[i].y, Data, i); |
| 765 |
|
if (Data->iMinSAD[0] <= threshA) { i++; break; } |
| 766 |
|
} |
| 767 |
|
|
| 768 |
|
if ((Data->iMinSAD[0] <= threshA) || |
| 769 |
|
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
| 770 |
|
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
| 771 |
|
inter4v = 0; |
| 772 |
|
else { |
| 773 |
|
|
| 774 |
|
MainSearchFunc * MainSearchPtr; |
| 775 |
|
int mask = make_mask(pmv, i, Data->dir); /* all vectors pmv[0..i-1] have been checked */ |
| 776 |
|
|
| 777 |
|
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; |
| 778 |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; |
| 779 |
|
else MainSearchPtr = xvid_me_DiamondSearch; |
| 780 |
|
|
| 781 |
|
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate); |
| 782 |
|
|
| 783 |
|
/* extended search, diamond starting in 0,0 and in prediction. |
| 784 |
|
note that this search is/might be done in halfpel positions, |
| 785 |
|
which makes it more different than the diamond above */ |
| 786 |
|
|
| 787 |
|
if (MotionFlags & XVID_ME_EXTSEARCH16) { |
| 788 |
|
int32_t bSAD; |
| 789 |
|
VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; |
| 790 |
|
if (Data->qpel) { |
| 791 |
|
startMV.x /= 2; |
| 792 |
|
startMV.y /= 2; |
| 793 |
|
} else if (Data->rrv) { |
| 794 |
|
startMV.x = RRV_MV_SCALEUP(startMV.x); |
| 795 |
|
startMV.y = RRV_MV_SCALEUP(startMV.y); |
| 796 |
|
} |
| 797 |
|
if (!(MVequal(startMV, backupMV))) { |
| 798 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
| 799 |
|
|
| 800 |
|
CheckCandidate(startMV.x, startMV.y, Data, 255); |
| 801 |
|
xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate); |
| 802 |
|
if (bSAD < Data->iMinSAD[0]) { |
| 803 |
|
Data->currentMV[0] = backupMV; |
| 804 |
|
Data->iMinSAD[0] = bSAD; } |
| 805 |
|
} |
| 806 |
|
|
| 807 |
|
backupMV = Data->currentMV[0]; |
| 808 |
|
startMV.x = startMV.y = 1; |
| 809 |
|
if (!(MVequal(startMV, backupMV))) { |
| 810 |
|
bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; |
| 811 |
|
|
| 812 |
|
CheckCandidate(startMV.x, startMV.y, Data, 255); |
| 813 |
|
xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate); |
| 814 |
|
if (bSAD < Data->iMinSAD[0]) { |
| 815 |
|
Data->currentMV[0] = backupMV; |
| 816 |
|
Data->iMinSAD[0] = bSAD; |
| 817 |
|
} |
| 818 |
|
} |
| 819 |
|
} |
| 820 |
|
} |
| 821 |
|
|
| 822 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
| 823 |
|
xvid_me_SubpelRefine(Data, CheckCandidate); |
| 824 |
|
|
| 825 |
|
for(i = 0; i < 5; i++) { |
| 826 |
|
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
| 827 |
|
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
| 828 |
|
} |
| 829 |
|
|
| 830 |
|
if (Data->qpel) { |
| 831 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
| 832 |
|
pParam->width, pParam->height, Data->iFcode, 2, 0); |
| 833 |
|
Data->qpel_precision = 1; |
| 834 |
|
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
| 835 |
|
if(MotionFlags & XVID_ME_FASTREFINE16) |
| 836 |
|
SubpelRefine_Fast(Data, CheckCandidate16_qpel); |
| 837 |
|
else |
| 838 |
|
xvid_me_SubpelRefine(Data, CheckCandidate16_qpel); |
| 839 |
|
} |
| 840 |
|
} |
| 841 |
|
|
| 842 |
|
if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
| 843 |
|
inter4v = 0; |
| 844 |
|
|
| 845 |
|
if (inter4v) { |
| 846 |
|
SearchData Data8; |
| 847 |
|
memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
| 848 |
|
|
| 849 |
|
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
| 850 |
|
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
| 851 |
|
Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); |
| 852 |
|
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
| 853 |
|
|
| 854 |
|
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { |
| 855 |
|
/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ |
| 856 |
|
int sumx = 0, sumy = 0; |
| 857 |
|
|
| 858 |
|
if (Data->qpel) |
| 859 |
|
for (i = 1; i < 5; i++) { |
| 860 |
|
sumx += Data->currentQMV[i].x/2; |
| 861 |
|
sumy += Data->currentQMV[i].y/2; |
| 862 |
|
} |
| 863 |
|
else |
| 864 |
|
for (i = 1; i < 5; i++) { |
| 865 |
|
sumx += Data->currentMV[i].x; |
| 866 |
|
sumy += Data->currentMV[i].y; |
| 867 |
|
} |
| 868 |
|
|
| 869 |
|
Data->iMinSAD[1] += xvid_me_ChromaSAD((sumx >> 3) + roundtab_76[sumx & 0xf], |
| 870 |
|
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
| 871 |
|
} |
| 872 |
|
} else Data->iMinSAD[1] = 4096*256; |
| 873 |
|
} |
| 874 |
|
|
| 875 |
|
static __inline uint32_t |
| 876 |
|
MakeGoodMotionFlags(const uint32_t MotionFlags, const uint32_t VopFlags, const uint32_t VolFlags) |
| 877 |
|
{ |
| 878 |
|
uint32_t Flags = MotionFlags; |
| 879 |
|
|
| 880 |
|
if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) |
| 881 |
|
Flags &= ~(XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD+XVID_ME_EXTSEARCH_RD); |
| 882 |
|
|
| 883 |
|
if (Flags & XVID_ME_EXTSEARCH_RD) |
| 884 |
|
Flags |= XVID_ME_HALFPELREFINE16_RD; |
| 885 |
|
|
| 886 |
|
if (Flags & XVID_ME_EXTSEARCH_RD && MotionFlags & XVID_ME_EXTSEARCH8) |
| 887 |
|
Flags |= XVID_ME_HALFPELREFINE8_RD; |
| 888 |
|
|
| 889 |
|
if (Flags & XVID_ME_HALFPELREFINE16_RD) |
| 890 |
|
Flags |= XVID_ME_QUARTERPELREFINE16_RD; |
| 891 |
|
|
| 892 |
|
if (Flags & XVID_ME_HALFPELREFINE8_RD) { |
| 893 |
|
Flags |= XVID_ME_QUARTERPELREFINE8_RD; |
| 894 |
|
Flags &= ~XVID_ME_HALFPELREFINE8; |
| 895 |
|
} |
| 896 |
|
|
| 897 |
|
if (Flags & XVID_ME_QUARTERPELREFINE8_RD) |
| 898 |
|
Flags &= ~XVID_ME_QUARTERPELREFINE8; |
| 899 |
|
|
| 900 |
|
if (!(VolFlags & XVID_VOL_QUARTERPEL)) |
| 901 |
|
Flags &= ~(XVID_ME_QUARTERPELREFINE16+XVID_ME_QUARTERPELREFINE8+XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD); |
| 902 |
|
|
| 903 |
|
if (!(VopFlags & XVID_VOP_HALFPEL)) |
| 904 |
|
Flags &= ~(XVID_ME_EXTSEARCH16+XVID_ME_HALFPELREFINE16+XVID_ME_HALFPELREFINE8+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD); |
| 905 |
|
|
| 906 |
|
if ((VopFlags & XVID_VOP_GREYSCALE) || (VopFlags & XVID_VOP_REDUCED)) |
| 907 |
|
Flags &= ~(XVID_ME_CHROMA_PVOP + XVID_ME_CHROMA_BVOP); |
| 908 |
|
|
| 909 |
|
return Flags; |
| 910 |
|
} |
| 911 |
|
|
| 912 |
|
bool |
| 913 |
|
MotionEstimation(MBParam * const pParam, |
| 914 |
|
FRAMEINFO * const current, |
| 915 |
|
FRAMEINFO * const reference, |
| 916 |
|
const IMAGE * const pRefH, |
| 917 |
|
const IMAGE * const pRefV, |
| 918 |
|
const IMAGE * const pRefHV, |
| 919 |
|
const IMAGE * const pGMC, |
| 920 |
|
const uint32_t iLimit) |
| 921 |
|
{ |
| 922 |
|
MACROBLOCK *const pMBs = current->mbs; |
| 923 |
|
const IMAGE *const pCurrent = ¤t->image; |
| 924 |
|
const IMAGE *const pRef = &reference->image; |
| 925 |
|
|
| 926 |
|
uint32_t mb_width = pParam->mb_width; |
| 927 |
|
uint32_t mb_height = pParam->mb_height; |
| 928 |
|
const uint32_t iEdgedWidth = pParam->edged_width; |
| 929 |
|
const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags); |
| 930 |
|
int stat_thresh = 0; |
| 931 |
|
|
| 932 |
|
uint32_t x, y; |
| 933 |
|
uint32_t iIntra = 0; |
| 934 |
|
int32_t sad00; |
| 935 |
|
int skip_thresh = INITIAL_SKIP_THRESH * \ |
| 936 |
|
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
| 937 |
|
(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); |
| 938 |
|
|
| 939 |
|
/* some pre-initialized thingies for SearchP */ |
| 940 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
| 941 |
|
SearchData Data; |
| 942 |
|
memset(&Data, 0, sizeof(SearchData)); |
| 943 |
|
Data.iEdgedWidth = iEdgedWidth; |
| 944 |
|
Data.iFcode = current->fcode; |
| 945 |
|
Data.rounding = pParam->m_rounding_type; |
| 946 |
|
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
| 947 |
|
Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; |
| 948 |
|
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
| 949 |
|
Data.dctSpace = dct_space; |
| 950 |
|
Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
| 951 |
|
Data.mpeg_quant_matrices = pParam->mpeg_quant_matrices; |
| 952 |
|
Data.iMinSAD2 = 0; |
| 953 |
|
|
| 954 |
|
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
| 955 |
|
mb_width = (pParam->width + 31) / 32; |
| 956 |
|
mb_height = (pParam->height + 31) / 32; |
| 957 |
|
Data.qpel = 0; |
| 958 |
|
} |
| 959 |
|
|
| 960 |
|
Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ |
| 961 |
|
if (sadInit) (*sadInit) (); |
| 962 |
|
|
| 963 |
|
for (y = 0; y < mb_height; y++) { |
| 964 |
|
for (x = 0; x < mb_width; x++) { |
| 965 |
|
MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; |
| 966 |
|
MACROBLOCK *prevMB = &reference->mbs[x + y * pParam->mb_width]; |
| 967 |
|
|
| 968 |
|
if (!Data.rrv) pMB->sad16 = |
| 969 |
|
sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, |
| 970 |
|
pRef->y + (x + y * iEdgedWidth) * 16, |
| 971 |
|
pParam->edged_width, pMB->sad8 ); |
| 972 |
|
|
| 973 |
|
else pMB->sad16 = |
| 974 |
|
sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, |
| 975 |
|
pRef->y + (x + y * iEdgedWidth) * 32, |
| 976 |
|
pParam->edged_width, pMB->sad8 ); |
| 977 |
|
|
| 978 |
|
if (Data.chroma) { |
| 979 |
|
Data.chromaSAD = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, |
| 980 |
|
pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) |
| 981 |
|
+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, |
| 982 |
|
pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); |
| 983 |
|
pMB->sad16 += Data.chromaSAD; |
| 984 |
|
} |
| 985 |
|
|
| 986 |
|
sad00 = pMB->sad16; |
| 987 |
|
|
| 988 |
|
/* initial skip decision */ |
| 989 |
|
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
| 990 |
|
if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
| 991 |
|
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
| 992 |
|
if (Data.chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
| 993 |
|
ZeroMacroblockP(pMB, sad00); |
| 994 |
|
pMB->mode = MODE_NOT_CODED; |
| 995 |
|
continue; |
| 996 |
|
} |
| 997 |
|
} |
| 998 |
|
|
| 999 |
|
if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) { |
| 1000 |
|
if(x > 0 && y > 0 && x < pParam->mb_width) { |
| 1001 |
|
if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], zeroMV) && |
| 1002 |
|
MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && |
| 1003 |
|
MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], zeroMV) && |
| 1004 |
|
MVequal(prevMB->mvs[0], zeroMV)) { |
| 1005 |
|
stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16, |
| 1006 |
|
MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16, |
| 1007 |
|
MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, |
| 1008 |
|
prevMB->sad16))); |
| 1009 |
|
} else { |
| 1010 |
|
stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16, |
| 1011 |
|
MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16, |
| 1012 |
|
MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16, |
| 1013 |
|
prevMB->sad16))); |
| 1014 |
|
} |
| 1015 |
|
} |
| 1016 |
|
} |
| 1017 |
|
|
| 1018 |
|
/* favorize (0,0) vector for cartoons */ |
| 1019 |
|
if ((current->vop_flags & XVID_VOP_CARTOON) && |
| 1020 |
|
((sad00 < pMB->quant * 4 * skip_thresh) || (sad00 < stat_thresh))) { |
| 1021 |
|
ZeroMacroblockP(pMB, sad00); |
| 1022 |
|
continue; |
| 1023 |
|
} |
| 1024 |
|
|
| 1025 |
|
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
| 1026 |
|
y, MotionFlags, current->vop_flags, |
| 1027 |
|
&Data, pParam, pMBs, reference->mbs, pMB); |
| 1028 |
|
|
| 1029 |
|
if (current->vop_flags & XVID_VOP_MODEDECISION_RD) |
| 1030 |
|
xvid_me_ModeDecision_RD(&Data, pMB, pMBs, x, y, pParam, |
| 1031 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
| 1032 |
|
pCurrent, pRef, pGMC, current->coding_type); |
| 1033 |
|
|
| 1034 |
|
else if (current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD) |
| 1035 |
|
xvid_me_ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam, |
| 1036 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
| 1037 |
|
pCurrent, pRef, pGMC, current->coding_type); |
| 1038 |
|
else |
| 1039 |
|
ModeDecision_SAD(&Data, pMB, pMBs, x, y, pParam, |
| 1040 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
| 1041 |
|
pCurrent, pRef, pGMC, current->coding_type); |
| 1042 |
|
|
| 1043 |
|
|
| 1044 |
|
if (pMB->mode == MODE_INTRA) |
| 1045 |
|
if (++iIntra > iLimit) return 1; |
| 1046 |
|
} |
| 1047 |
|
} |
| 1048 |
|
return 0; |
| 1049 |
|
} |
| 1050 |
|
|
| 1051 |
|
|
Parent Directory
| 
Revision Log
| 
Patch