| 4 |
* - Motion Estimation for B-VOPs - |
* - Motion Estimation for B-VOPs - |
| 5 |
* |
* |
| 6 |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
| 7 |
* 2002 Michael Militzer <michael@xvid.org> |
* 2002-2010 Michael Militzer <michael@xvid.org> |
| 8 |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
| 9 |
* |
* |
| 10 |
* This program is free software ; you can redistribute it and/or modify |
* This program is free software ; you can redistribute it and/or modify |
| 46 |
int sad; |
int sad; |
| 47 |
const uint32_t stride = data->iEdgedWidth/2; |
const uint32_t stride = data->iEdgedWidth/2; |
| 48 |
uint8_t *f_refu, *f_refv, *b_refu, *b_refv; |
uint8_t *f_refu, *f_refv, *b_refu, *b_refv; |
| 49 |
|
int offset, filter; |
| 50 |
|
|
| 51 |
const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { |
const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { |
| 52 |
NULL, |
NULL, |
| 55 |
interpolate8x8_halfpel_hv |
interpolate8x8_halfpel_hv |
| 56 |
}; |
}; |
| 57 |
|
|
| 58 |
int offset = (fx>>1) + (fy>>1)*stride; |
if (data->chromaX == fx && data->chromaY == fy && |
| 59 |
int filter = ((fx & 1) << 1) | (fy & 1); |
data->b_chromaX == bx && data->b_chromaY == by) |
| 60 |
|
return data->chromaSAD; |
| 61 |
|
|
| 62 |
|
offset = (fx>>1) + (fy>>1)*stride; |
| 63 |
|
filter = ((fx & 1) << 1) | (fy & 1); |
| 64 |
|
|
| 65 |
if (filter != 0) { |
if (filter != 0) { |
| 66 |
f_refu = data->RefQ; |
f_refu = data->RefQ + 64; |
| 67 |
f_refv = data->RefQ + 8; |
f_refv = data->RefQ + 64 + 8; |
| 68 |
|
if (data->chromaX != fx || data->chromaY != fy) { |
| 69 |
interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); |
interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); |
| 70 |
interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); |
interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); |
| 71 |
|
} |
| 72 |
} else { |
} else { |
| 73 |
f_refu = (uint8_t*)data->RefP[4] + offset; |
f_refu = (uint8_t*)data->RefP[4] + offset; |
| 74 |
f_refv = (uint8_t*)data->RefP[5] + offset; |
f_refv = (uint8_t*)data->RefP[5] + offset; |
| 75 |
} |
} |
| 76 |
|
data->chromaX = fx; data->chromaY = fy; |
| 77 |
|
|
| 78 |
offset = (bx>>1) + (by>>1)*stride; |
offset = (bx>>1) + (by>>1)*stride; |
| 79 |
filter = ((bx & 1) << 1) | (by & 1); |
filter = ((bx & 1) << 1) | (by & 1); |
| 80 |
|
|
| 81 |
if (filter != 0) { |
if (filter != 0) { |
| 82 |
b_refu = data->RefQ + 16; |
b_refu = data->RefQ + 64 + 16; |
| 83 |
b_refv = data->RefQ + 24; |
b_refv = data->RefQ + 64 + 24; |
| 84 |
|
if (data->b_chromaX != bx || data->b_chromaY != by) { |
| 85 |
interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); |
| 86 |
interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); |
| 87 |
|
} |
| 88 |
} else { |
} else { |
| 89 |
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
b_refu = (uint8_t*)data->b_RefP[4] + offset; |
| 90 |
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
b_refv = (uint8_t*)data->b_RefP[5] + offset; |
| 91 |
} |
} |
| 92 |
|
data->b_chromaX = bx; data->b_chromaY = by; |
| 93 |
|
|
| 94 |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
sad = sad8bi(data->CurU, b_refu, f_refu, stride); |
| 95 |
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
sad += sad8bi(data->CurV, b_refv, f_refv, stride); |
| 96 |
|
|
| 97 |
|
data->chromaSAD = sad; |
| 98 |
return sad; |
return sad; |
| 99 |
} |
} |
| 100 |
|
|
| 137 |
xcb = xb/2; ycb = yb/2; |
xcb = xb/2; ycb = yb/2; |
| 138 |
} |
} |
| 139 |
|
|
| 140 |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision) |
| 141 |
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision); |
| 142 |
|
|
| 143 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
| 144 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t); |
| 145 |
|
|
| 146 |
if (data->chroma && sad < *data->iMinSAD) |
if (data->chroma && sad < *data->iMinSAD) |
| 147 |
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
| 157 |
} |
} |
| 158 |
|
|
| 159 |
static void |
static void |
|
CheckCandidateInt_qpel(const int x, const int y, SearchData * const data, const unsigned int Direction) |
|
|
{ |
|
|
int32_t sad, xf, yf, xb, yb, xcf, ycf, xcb, ycb; |
|
|
uint32_t t; |
|
|
|
|
|
const uint8_t *ReferenceF, *ReferenceB; |
|
|
VECTOR *current; |
|
|
|
|
|
if ((x > data->max_dx) || (x < data->min_dx) || |
|
|
(y > data->max_dy) || (y < data->min_dy)) |
|
|
return; |
|
|
|
|
|
if (Direction == 1) { /* x and y mean forward vector */ |
|
|
VECTOR backward = data->qpel_precision ? data->currentQMV[1] : data->currentMV[1]; |
|
|
xb = backward.x; |
|
|
yb = backward.y; |
|
|
xf = x; yf = y; |
|
|
} else { /* x and y mean backward vector */ |
|
|
VECTOR forward = data->qpel_precision ? data->currentQMV[0] : data->currentMV[0]; |
|
|
xf = forward.x; |
|
|
yf = forward.y; |
|
|
xb = x; yb = y; |
|
|
} |
|
|
|
|
|
ReferenceF = xvid_me_interpolate16x16qpel(xf, yf, 0, data); |
|
|
current = data->currentQMV + Direction - 1; |
|
|
ReferenceB = xvid_me_interpolate16x16qpel(xb, yb, 1, data); |
|
|
xcf = xf/2; ycf = yf/2; |
|
|
xcb = xb/2; ycb = yb/2; |
|
|
|
|
|
t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) |
|
|
+ d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
|
|
|
|
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
|
|
sad += (data->lambda16 * t * sad)>>10; |
|
|
|
|
|
if (data->chroma && sad < *data->iMinSAD) |
|
|
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
|
|
(ycf >> 1) + roundtab_79[ycf & 0x3], |
|
|
(xcb >> 1) + roundtab_79[xcb & 0x3], |
|
|
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
*data->iMinSAD = sad; |
|
|
current->x = x; current->y = y; |
|
|
data->dir = Direction; |
|
|
} |
|
|
|
|
|
if (sad < *(data->iMinSAD)) { |
|
|
data->iMinSAD2 = *(data->iMinSAD); |
|
|
data->currentQMV2.x = current->x; |
|
|
data->currentQMV2.y = current->y; |
|
|
|
|
|
*data->iMinSAD = sad; |
|
|
current->x = x; current->y = y; |
|
|
} else if (sad < data->iMinSAD2) { |
|
|
data->iMinSAD2 = sad; |
|
|
data->currentQMV2.x = x; data->currentQMV2.y = y; |
|
|
} |
|
|
} |
|
|
|
|
|
static void |
|
| 160 |
CheckCandidateDirect(const int x, const int y, SearchData * const data, const unsigned int Direction) |
CheckCandidateDirect(const int x, const int y, SearchData * const data, const unsigned int Direction) |
| 161 |
{ |
{ |
| 162 |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; |
| 164 |
const uint8_t *ReferenceF; |
const uint8_t *ReferenceF; |
| 165 |
const uint8_t *ReferenceB; |
const uint8_t *ReferenceB; |
| 166 |
VECTOR mvs, b_mvs; |
VECTOR mvs, b_mvs; |
| 167 |
|
const int blocks[4] = {0, 8, 8*data->iEdgedWidth, 8*data->iEdgedWidth+8}; |
| 168 |
|
|
| 169 |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; |
| 170 |
|
|
| 188 |
if (data->qpel) { |
if (data->qpel) { |
| 189 |
xcf += mvs.x/2; ycf += mvs.y/2; |
xcf += mvs.x/2; ycf += mvs.y/2; |
| 190 |
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
xcb += b_mvs.x/2; ycb += b_mvs.y/2; |
| 191 |
|
if (data->qpel_precision) { |
| 192 |
|
ReferenceF = xvid_me_interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
| 193 |
|
ReferenceB = xvid_me_interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
| 194 |
|
goto done; |
| 195 |
|
} |
| 196 |
|
mvs.x >>=1; mvs.y >>=1; b_mvs.x >>=1; b_mvs.y >>=1; // qpel->hpel |
| 197 |
} else { |
} else { |
| 198 |
xcf += mvs.x; ycf += mvs.y; |
xcf += mvs.x; ycf += mvs.y; |
| 199 |
xcb += b_mvs.x; ycb += b_mvs.y; |
xcb += b_mvs.x; ycb += b_mvs.y; |
|
mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
|
|
b_mvs.x *= 2; b_mvs.y *= 2; |
|
| 200 |
} |
} |
| 201 |
|
ReferenceF = GetReference(mvs.x, mvs.y, data) + blocks[k]; |
| 202 |
ReferenceF = xvid_me_interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data) + blocks[k]; |
| 203 |
ReferenceB = xvid_me_interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); |
done: |
| 204 |
|
sad += data->iMinSAD[k+1] = |
| 205 |
sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), |
sad8bi(data->Cur + blocks[k], |
| 206 |
ReferenceF, ReferenceB, data->iEdgedWidth); |
ReferenceF, ReferenceB, data->iEdgedWidth); |
| 207 |
if (sad > *(data->iMinSAD)) return; |
if (sad > *(data->iMinSAD)) return; |
| 208 |
} |
} |
| 209 |
|
|
| 210 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0)); |
| 211 |
|
|
| 212 |
if (data->chroma && sad < *data->iMinSAD) |
if (data->chroma && sad < *data->iMinSAD) |
| 213 |
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
| 250 |
if (data->qpel) { |
if (data->qpel) { |
| 251 |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); |
| 252 |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); |
| 253 |
|
if (data->qpel_precision) { |
| 254 |
ReferenceF = xvid_me_interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
ReferenceF = xvid_me_interpolate16x16qpel(mvs.x, mvs.y, 0, data); |
| 255 |
ReferenceB = xvid_me_interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
ReferenceB = xvid_me_interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); |
| 256 |
|
goto done; |
| 257 |
|
} |
| 258 |
|
mvs.x >>=1; mvs.y >>=1; b_mvs.x >>=1; b_mvs.y >>=1; // qpel->hpel |
| 259 |
} else { |
} else { |
| 260 |
xcf = 4*mvs.x; ycf = 4*mvs.y; |
xcf = 4*mvs.x; ycf = 4*mvs.y; |
| 261 |
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; |
| 262 |
|
} |
| 263 |
ReferenceF = GetReference(mvs.x, mvs.y, data); |
ReferenceF = GetReference(mvs.x, mvs.y, data); |
| 264 |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); |
|
} |
|
| 265 |
|
|
| 266 |
|
done: |
| 267 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
| 268 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0)); |
| 269 |
|
|
| 270 |
if (data->chroma && sad < *data->iMinSAD) |
if (data->chroma && sad < *data->iMinSAD) |
| 271 |
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
| 291 |
if ( (x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) || ( x < data->min_dx) |
| 292 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
| 293 |
|
|
|
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
|
|
|
|
| 294 |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
| 295 |
Reference = xvid_me_interpolate16x16qpel(x, y, 0, data); |
Reference = xvid_me_interpolate16x16qpel(x, y, 0, data); |
| 296 |
current = data->currentQMV; |
current = data->currentQMV; |
| 301 |
xc = x; yc = y; |
xc = x; yc = y; |
| 302 |
} |
} |
| 303 |
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
| 304 |
data->qpel^data->qpel_precision, data->rrv); |
data->qpel^data->qpel_precision); |
| 305 |
|
|
| 306 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
| 307 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t); |
| 308 |
|
|
| 309 |
if (data->chroma && sad < *data->iMinSAD) |
if (data->chroma && sad < *data->iMinSAD) |
| 310 |
sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
| 317 |
} |
} |
| 318 |
} |
} |
| 319 |
|
|
|
void |
|
|
CheckCandidate16no4v_qpel(const int x, const int y, SearchData * const data, const unsigned int Direction) |
|
|
{ |
|
|
int32_t sad, xc, yc; |
|
|
const uint8_t * Reference; |
|
|
uint32_t t; |
|
|
|
|
|
if ( (x > data->max_dx) || ( x < data->min_dx) |
|
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|
|
|
|
|
if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */ |
|
|
|
|
|
Reference = xvid_me_interpolate16x16qpel(x, y, 0, data); |
|
|
|
|
|
xc = x/2; yc = y/2; |
|
|
t = d_mv_bits(x, y, data->predMV, data->iFcode, |
|
|
data->qpel^data->qpel_precision, data->rrv); |
|
| 320 |
|
|
| 321 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
static void |
| 322 |
sad += (data->lambda16 * t * sad)>>10; |
initialize_searchData(SearchData * Data_d, |
| 323 |
|
SearchData * Data_f, |
| 324 |
|
SearchData * Data_b, |
| 325 |
|
SearchData * Data_i, |
| 326 |
|
int x, int y, |
| 327 |
|
const IMAGE * const f_Ref, |
| 328 |
|
const uint8_t * const f_RefH, |
| 329 |
|
const uint8_t * const f_RefV, |
| 330 |
|
const uint8_t * const f_RefHV, |
| 331 |
|
const IMAGE * const b_Ref, |
| 332 |
|
const uint8_t * const b_RefH, |
| 333 |
|
const uint8_t * const b_RefV, |
| 334 |
|
const uint8_t * const b_RefHV, |
| 335 |
|
const IMAGE * const pCur, |
| 336 |
|
const MACROBLOCK * const b_mb) |
| 337 |
|
{ |
| 338 |
|
|
| 339 |
if (data->chroma && sad < *data->iMinSAD) |
/* per-macroblock SearchData initialization - too many things would be repeated 4 times */ |
| 340 |
sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
const uint8_t * RefP[6], * b_RefP[6], * Cur[3]; |
| 341 |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
const uint32_t iEdgedWidth = Data_d->iEdgedWidth; |
| 342 |
|
unsigned int lambda; |
| 343 |
|
int i; |
| 344 |
|
|
| 345 |
if (sad < *(data->iMinSAD)) { |
/* luma */ |
| 346 |
data->iMinSAD2 = *(data->iMinSAD); |
int offset = (x + iEdgedWidth*y) * 16; |
| 347 |
data->currentQMV2.x = data->currentQMV->x; |
RefP[0] = f_Ref->y + offset; |
| 348 |
data->currentQMV2.y = data->currentQMV->y; |
RefP[2] = f_RefH + offset; |
| 349 |
|
RefP[1] = f_RefV + offset; |
| 350 |
|
RefP[3] = f_RefHV + offset; |
| 351 |
|
b_RefP[0] = b_Ref->y + offset; |
| 352 |
|
b_RefP[2] = b_RefH + offset; |
| 353 |
|
b_RefP[1] = b_RefV + offset; |
| 354 |
|
b_RefP[3] = b_RefHV + offset; |
| 355 |
|
Cur[0] = pCur->y + offset; |
| 356 |
|
|
| 357 |
|
/* chroma */ |
| 358 |
|
offset = (x + (iEdgedWidth/2)*y) * 8; |
| 359 |
|
RefP[4] = f_Ref->u + offset; |
| 360 |
|
RefP[5] = f_Ref->v + offset; |
| 361 |
|
b_RefP[4] = b_Ref->u + offset; |
| 362 |
|
b_RefP[5] = b_Ref->v + offset; |
| 363 |
|
Cur[1] = pCur->u + offset; |
| 364 |
|
Cur[2] = pCur->v + offset; |
| 365 |
|
|
| 366 |
|
lambda = xvid_me_lambda_vec16[b_mb->quant]; |
| 367 |
|
|
| 368 |
|
for (i = 0; i < 6; i++) { |
| 369 |
|
Data_d->RefP[i] = Data_f->RefP[i] = Data_i->RefP[i] = RefP[i]; |
| 370 |
|
Data_d->b_RefP[i] = Data_b->RefP[i] = Data_i->b_RefP[i] = b_RefP[i]; |
| 371 |
|
} |
| 372 |
|
Data_d->Cur = Data_f->Cur = Data_b->Cur = Data_i->Cur = Cur[0]; |
| 373 |
|
Data_d->CurU = Data_f->CurU = Data_b->CurU = Data_i->CurU = Cur[1]; |
| 374 |
|
Data_d->CurV = Data_f->CurV = Data_b->CurV = Data_i->CurV = Cur[2]; |
| 375 |
|
|
| 376 |
|
Data_d->lambda16 = Data_f->lambda16 = Data_b->lambda16 = Data_i->lambda16 = lambda; |
| 377 |
|
|
| 378 |
|
/* reset chroma-sad cache */ |
| 379 |
|
Data_d->b_chromaX = Data_d->b_chromaY = Data_d->chromaX = Data_d->chromaY = Data_d->chromaSAD = 256*4096; |
| 380 |
|
Data_i->b_chromaX = Data_i->b_chromaY = Data_i->chromaX = Data_i->chromaY = Data_i->chromaSAD = 256*4096; |
| 381 |
|
Data_f->chromaX = Data_f->chromaY = Data_f->chromaSAD = 256*4096; |
| 382 |
|
Data_b->chromaX = Data_b->chromaY = Data_b->chromaSAD = 256*4096; |
| 383 |
|
|
| 384 |
data->iMinSAD[0] = sad; |
*Data_d->iMinSAD = *Data_b->iMinSAD = *Data_f->iMinSAD = *Data_i->iMinSAD = 4096*256; |
|
data->currentQMV[0].x = x; data->currentQMV[0].y = y; |
|
|
} else if (sad < data->iMinSAD2) { |
|
|
data->iMinSAD2 = sad; |
|
|
data->currentQMV2.x = x; data->currentQMV2.y = y; |
|
|
} |
|
| 385 |
} |
} |
| 386 |
|
|
| 387 |
static __inline VECTOR |
static __inline VECTOR |
| 395 |
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, |
| 396 |
const uint32_t iWcount, |
const uint32_t iWcount, |
| 397 |
const MACROBLOCK * const pMB, |
const MACROBLOCK * const pMB, |
| 398 |
const uint32_t mode_curr) |
const uint32_t mode_curr, |
| 399 |
|
const VECTOR hint, const int bound) |
| 400 |
{ |
{ |
| 401 |
|
int lx, ly; /* left */ |
| 402 |
|
int tx, ty; /* top */ |
| 403 |
|
int rtx, rty; /* top-right */ |
| 404 |
|
int ltx, lty; /* top-left */ |
| 405 |
|
int lpos, tpos, rtpos, ltpos; |
| 406 |
|
|
| 407 |
|
lx = x - 1; ly = y; |
| 408 |
|
tx = x; ty = y - 1; |
| 409 |
|
rtx = x + 1; rty = y - 1; |
| 410 |
|
ltx = x - 1; lty = y - 1; |
| 411 |
|
|
| 412 |
|
lpos = lx + ly * iWcount; |
| 413 |
|
rtpos = rtx + rty * iWcount; |
| 414 |
|
tpos = tx + ty * iWcount; |
| 415 |
|
ltpos = ltx + lty * iWcount; |
| 416 |
|
|
|
/* [0] is prediction */ |
|
|
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
|
| 417 |
|
|
| 418 |
pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
/* [0] is prediction */ |
| 419 |
|
/* [1] is zero */ |
| 420 |
|
pmv[1].x = pmv[1].y = 0; |
| 421 |
|
|
| 422 |
pmv[2] = ChoosePred(pMB, mode_curr); |
pmv[2].x = hint.x; pmv[2].y = hint.y; |
|
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
|
| 423 |
|
|
| 424 |
if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
if (rtpos >= bound && rtx < (int)iWcount) { /* [3] top-right neighbour */ |
| 425 |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
|
pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); |
|
| 426 |
} else pmv[3].x = pmv[3].y = 0; |
} else pmv[3].x = pmv[3].y = 0; |
| 427 |
|
|
| 428 |
if (y != 0) { |
if (tpos >= bound) { |
| 429 |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); |
pmv[4] = ChoosePred(pMB-iWcount, mode_curr); /* [4] top */ |
|
pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); |
|
| 430 |
} else pmv[4].x = pmv[4].y = 0; |
} else pmv[4].x = pmv[4].y = 0; |
| 431 |
|
|
| 432 |
if (x != 0) { |
if (lpos >= bound && lx >= 0) { |
| 433 |
pmv[5] = ChoosePred(pMB-1, mode_curr); |
pmv[5] = ChoosePred(pMB-1, mode_curr); /* [5] left */ |
|
pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); |
|
| 434 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
| 435 |
|
|
| 436 |
if (x != 0 && y != 0) { |
if (ltpos >= bound && ltx >= 0) { |
| 437 |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); |
pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); /* [6] top-left */ |
|
pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); |
|
| 438 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
| 439 |
} |
} |
| 440 |
|
|
|
|
|
| 441 |
/* search backward or forward */ |
/* search backward or forward */ |
| 442 |
static void |
static void |
| 443 |
SearchBF( const IMAGE * const pRef, |
SearchBF_initial(const int x, const int y, |
|
const uint8_t * const pRefH, |
|
|
const uint8_t * const pRefV, |
|
|
const uint8_t * const pRefHV, |
|
|
const int x, const int y, |
|
| 444 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 445 |
const uint32_t iFcode, |
const uint32_t iFcode, |
| 446 |
const MBParam * const pParam, |
const MBParam * const pParam, |
| 448 |
const VECTOR * const predMV, |
const VECTOR * const predMV, |
| 449 |
int32_t * const best_sad, |
int32_t * const best_sad, |
| 450 |
const int32_t mode_current, |
const int32_t mode_current, |
| 451 |
SearchData * const Data) |
SearchData * const Data, |
| 452 |
|
VECTOR hint, const int bound) |
| 453 |
{ |
{ |
| 454 |
|
|
| 455 |
int i; |
int i; |
| 456 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
|
int threshA = (MotionFlags & XVID_ME_FASTREFINE16) ? 150 : 300; |
|
| 457 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
|
Data->iFcode = iFcode; |
|
| 458 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
|
Data->chromaX = Data->chromaY = Data->chromaSAD = 256*4096; /* reset chroma-sad cache */ |
|
|
|
|
|
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; |
|
|
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
|
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; |
|
| 459 |
|
|
| 460 |
Data->predMV = *predMV; |
Data->predMV = *predMV; |
| 461 |
|
|
| 462 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
| 463 |
pParam->width, pParam->height, iFcode - Data->qpel, 1, 0); |
pParam->width, pParam->height, iFcode - Data->qpel, 1); |
| 464 |
|
|
| 465 |
pmv[0] = Data->predMV; |
pmv[0] = Data->predMV; |
| 466 |
if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } |
if (Data->qpel) { |
| 467 |
|
pmv[0].x /= 2; pmv[0].y /= 2; |
| 468 |
|
hint.x /= 2; hint.y /= 2; |
| 469 |
|
} |
| 470 |
|
|
| 471 |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); |
PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current, hint, bound); |
| 472 |
|
|
| 473 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
| 474 |
|
|
| 488 |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); |
| 489 |
} |
} |
| 490 |
|
|
| 491 |
xvid_me_SubpelRefine(Data, CheckCandidate16no4v); |
if (Data->iMinSAD[0] < *best_sad) *best_sad = Data->iMinSAD[0]; |
| 492 |
|
} |
| 493 |
|
|
| 494 |
if (Data->qpel && (*Data->iMinSAD < *best_sad + threshA)) { |
static void |
| 495 |
Data->currentQMV->x = 2*Data->currentMV->x; |
SearchBF_final(const int x, const int y, |
| 496 |
Data->currentQMV->y = 2*Data->currentMV->y; |
const uint32_t MotionFlags, |
| 497 |
Data->qpel_precision = 1; |
const MBParam * const pParam, |
| 498 |
|
int32_t * const best_sad, |
| 499 |
|
SearchData * const Data) |
| 500 |
|
{ |
| 501 |
|
if(!Data->qpel) { |
| 502 |
|
/* halfpel mode */ |
| 503 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
| 504 |
|
xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate16no4v, 0); |
| 505 |
|
} else { |
| 506 |
|
/* qpel mode */ |
| 507 |
|
if(MotionFlags & XVID_ME_FASTREFINE16) { |
| 508 |
|
/* fast */ |
| 509 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
| 510 |
pParam->width, pParam->height, iFcode, 2, 0); |
pParam->width, pParam->height, Data->iFcode, 2); |
| 511 |
|
FullRefine_Fast(Data, CheckCandidate16no4v, 0); |
| 512 |
|
|
| 513 |
|
} else { |
| 514 |
|
|
| 515 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
| 516 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
| 517 |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
| 518 |
if (MotionFlags & XVID_ME_FASTREFINE16) |
/* full */ |
| 519 |
SubpelRefine_Fast(Data, CheckCandidate16no4v_qpel); |
if (MotionFlags & XVID_ME_HALFPELREFINE16) { |
| 520 |
else |
xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate16no4v, 0); /* hpel part */ |
| 521 |
xvid_me_SubpelRefine(Data, CheckCandidate16no4v); |
Data->currentQMV->x = 2*Data->currentMV->x; |
| 522 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
| 523 |
} |
} |
| 524 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, |
| 525 |
|
pParam->width, pParam->height, Data->iFcode, 2); |
| 526 |
|
Data->qpel_precision = 1; |
| 527 |
|
xvid_me_SubpelRefine(Data->currentQMV[0], Data, CheckCandidate16no4v, 0); /* qpel part */ |
| 528 |
} |
} |
|
|
|
|
/* three bits are needed to code backward mode. four for forward */ |
|
|
|
|
|
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
|
|
else *Data->iMinSAD += 3 * Data->lambda16; |
|
|
|
|
|
if (*Data->iMinSAD < *best_sad) { |
|
|
*best_sad = *Data->iMinSAD; |
|
|
pMB->mode = mode_current; |
|
|
if (Data->qpel) { |
|
|
pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; |
|
|
pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; |
|
|
if (mode_current == MODE_FORWARD) |
|
|
pMB->qmvs[0] = *Data->currentQMV; |
|
|
else |
|
|
pMB->b_qmvs[0] = *Data->currentQMV; |
|
|
} else { |
|
|
pMB->pmvs[0].x = Data->currentMV->x - predMV->x; |
|
|
pMB->pmvs[0].y = Data->currentMV->y - predMV->y; |
|
| 529 |
} |
} |
|
if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; |
|
|
else pMB->b_mvs[0] = *Data->currentMV; |
|
| 530 |
} |
} |
| 531 |
|
if (Data->iMinSAD[0] < *best_sad) *best_sad = Data->iMinSAD[0]; |
| 532 |
|
|
|
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
|
|
else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
|
| 533 |
} |
} |
| 534 |
|
|
| 535 |
static void |
static void |
| 536 |
SkipDecisionB(const IMAGE * const pCur, |
SkipDecisionB(MACROBLOCK * const pMB, const SearchData * const Data) |
|
const IMAGE * const f_Ref, |
|
|
const IMAGE * const b_Ref, |
|
|
MACROBLOCK * const pMB, |
|
|
const uint32_t x, const uint32_t y, |
|
|
const SearchData * const Data) |
|
| 537 |
{ |
{ |
| 538 |
int k; |
int k; |
| 539 |
|
|
| 555 |
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; |
| 556 |
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; |
| 557 |
|
|
| 558 |
sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, |
sum = sad8bi(Data->CurU, |
| 559 |
f_Ref->u + ((int)y*8 + dy/2) * (int)stride + (int)x*8 + dx/2, |
Data->RefP[4] + (dy/2) * (int)stride + dx/2, |
| 560 |
b_Ref->u + ((int)y*8 + b_dy/2) * (int)stride + (int)x*8 + b_dx/2, |
Data->b_RefP[4] + (b_dy/2) * (int)stride + b_dx/2, |
| 561 |
stride); |
stride); |
| 562 |
|
|
| 563 |
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ |
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ |
| 564 |
|
|
| 565 |
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
sum += sad8bi(Data->CurV, |
| 566 |
f_Ref->v + ((int)y*8 + dy/2) * (int)stride + (int)x*8 + dx/2, |
Data->RefP[5] + (dy/2) * (int)stride + dx/2, |
| 567 |
b_Ref->v + ((int)y*8 + b_dy/2) * (int)stride + (int)x*8 + b_dx/2, |
Data->b_RefP[5] + (b_dy/2) * (int)stride + b_dx/2, |
| 568 |
stride); |
stride); |
| 569 |
|
|
| 570 |
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ |
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ |
| 571 |
|
} else { |
| 572 |
|
int sum = Data->chromaSAD; /* chroma-sad SAD caching keeps it there */ |
| 573 |
|
|
| 574 |
|
if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ |
| 575 |
} |
} |
| 576 |
|
|
| 577 |
/* skip */ |
/* skip */ |
| 579 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
| 580 |
pMB->qmvs[k] = pMB->mvs[k] = Data->directmvF[k]; |
pMB->qmvs[k] = pMB->mvs[k] = Data->directmvF[k]; |
| 581 |
pMB->b_qmvs[k] = pMB->b_mvs[k] = Data->directmvB[k]; |
pMB->b_qmvs[k] = pMB->b_mvs[k] = Data->directmvB[k]; |
| 582 |
|
if (Data->qpel) { |
| 583 |
|
pMB->mvs[k].x /= 2; pMB->mvs[k].y /= 2; /* it's a hint for future searches */ |
| 584 |
|
pMB->b_mvs[k].x /= 2; pMB->b_mvs[k].y /= 2; |
| 585 |
|
} |
| 586 |
} |
} |
| 587 |
} |
} |
| 588 |
|
|
| 589 |
static uint32_t |
static uint32_t |
| 590 |
SearchDirect(const IMAGE * const f_Ref, |
SearchDirect_initial(const int x, const int y, |
|
const uint8_t * const f_RefH, |
|
|
const uint8_t * const f_RefV, |
|
|
const uint8_t * const f_RefHV, |
|
|
const IMAGE * const b_Ref, |
|
|
const uint8_t * const b_RefH, |
|
|
const uint8_t * const b_RefV, |
|
|
const uint8_t * const b_RefHV, |
|
|
const IMAGE * const pCur, |
|
|
const int x, const int y, |
|
| 591 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 592 |
const int32_t TRB, const int32_t TRD, |
const int32_t TRB, const int32_t TRD, |
| 593 |
const MBParam * const pParam, |
const MBParam * const pParam, |
| 599 |
{ |
{ |
| 600 |
int32_t skip_sad; |
int32_t skip_sad; |
| 601 |
int k = (x + Data->iEdgedWidth*y) * 16; |
int k = (x + Data->iEdgedWidth*y) * 16; |
|
MainSearchFunc *MainSearchPtr; |
|
|
CheckFunc * CheckCandidate; |
|
|
|
|
|
*Data->iMinSAD = 256*4096; |
|
|
Data->RefP[0] = f_Ref->y + k; |
|
|
Data->RefP[2] = f_RefH + k; |
|
|
Data->RefP[1] = f_RefV + k; |
|
|
Data->RefP[3] = f_RefHV + k; |
|
|
Data->b_RefP[0] = b_Ref->y + k; |
|
|
Data->b_RefP[2] = b_RefH + k; |
|
|
Data->b_RefP[1] = b_RefV + k; |
|
|
Data->b_RefP[3] = b_RefHV + k; |
|
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
| 602 |
|
|
| 603 |
k = Data->qpel ? 4 : 2; |
k = Data->qpel ? 4 : 2; |
| 604 |
Data->max_dx = k * (pParam->width - x * 16); |
Data->max_dx = k * (pParam->width - x * 16); |
| 607 |
Data->min_dy = -k * (16 + y * 16); |
Data->min_dy = -k * (16 + y * 16); |
| 608 |
|
|
| 609 |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; |
|
Data->qpel_precision = 0; |
|
| 610 |
|
|
| 611 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
| 612 |
pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); |
| 613 |
pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; |
| 614 |
pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); |
| 615 |
pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; |
| 616 |
|
|
| 617 |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
if ( (Data->directmvB[k].x > Data->max_dx) | (Data->directmvB[k].x < Data->min_dx) |
| 618 |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
| (Data->directmvB[k].y > Data->max_dy) | (Data->directmvB[k].y < Data->min_dy) ) { |
| 619 |
|
|
| 620 |
*best_sad = 256*4096; /* in that case, we won't use direct mode */ |
Data->iMinSAD[0] = *best_sad = 256*4096; /* in that case, we won't use direct mode */ |
|
pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ |
|
|
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
|
| 621 |
return 256*4096; |
return 256*4096; |
| 622 |
} |
} |
| 623 |
if (b_mb->mode != MODE_INTER4V) { |
if (b_mb->mode != MODE_INTER4V) { |
|
pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; |
|
|
pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; |
|
| 624 |
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0]; |
| 625 |
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0]; |
| 626 |
break; |
break; |
| 627 |
} |
} |
| 628 |
} |
} |
| 629 |
|
Data->qpel_precision = Data->qpel; /* this initial check is done with full precision, to find real |
| 630 |
|
SKIP sad */ |
| 631 |
|
|
| 632 |
CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; |
CheckCandidateDirect(0, 0, Data, 255); /* will also fill iMinSAD[1..4] with 8x8 SADs */ |
|
|
|
|
CheckCandidate(0, 0, Data, 255); |
|
| 633 |
|
|
| 634 |
/* initial (fast) skip decision */ |
/* initial (fast) skip decision */ |
| 635 |
if (*Data->iMinSAD < (int)Data->iQuant * INITIAL_SKIP_THRESH) { |
if (Data->iMinSAD[1] < (int)Data->iQuant * INITIAL_SKIP_THRESH |
| 636 |
|
&& Data->iMinSAD[2] < (int)Data->iQuant * INITIAL_SKIP_THRESH |
| 637 |
|
&& Data->iMinSAD[3] < (int)Data->iQuant * INITIAL_SKIP_THRESH |
| 638 |
|
&& Data->iMinSAD[4] < (int)Data->iQuant * INITIAL_SKIP_THRESH) { |
| 639 |
/* possible skip */ |
/* possible skip */ |
| 640 |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
SkipDecisionB(pMB, Data); |
| 641 |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skipped */ |
if (pMB->mode == MODE_DIRECT_NONE_MV) |
| 642 |
|
return *Data->iMinSAD; /* skipped */ |
| 643 |
} |
} |
| 644 |
|
|
| 645 |
skip_sad = *Data->iMinSAD; |
if (Data->chroma && Data->chromaSAD >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) /* chroma doesn't allow skip */ |
| 646 |
|
skip_sad = 256*4096; |
| 647 |
|
else |
| 648 |
|
skip_sad = 4*MAX(MAX(Data->iMinSAD[1],Data->iMinSAD[2]), MAX(Data->iMinSAD[3],Data->iMinSAD[4])); |
| 649 |
|
|
| 650 |
if (!(MotionFlags & XVID_ME_SKIP_DELTASEARCH)) { |
Data->currentMV[1].x = Data->directmvF[0].x + Data->currentMV->x; /* hints for forward and backward searches */ |
| 651 |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; |
Data->currentMV[1].y = Data->directmvF[0].y + Data->currentMV->y; |
|
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; |
|
|
else MainSearchPtr = xvid_me_DiamondSearch; |
|
| 652 |
|
|
| 653 |
MainSearchPtr(0, 0, Data, 255, CheckCandidate); |
Data->currentMV[2].x = ((Data->currentMV->x == 0) ? |
| 654 |
|
Data->directmvB[0].x |
| 655 |
|
: Data->currentMV[1].x - Data->referencemv[0].x); |
| 656 |
|
|
| 657 |
|
Data->currentMV[2].y = ((Data->currentMV->y == 0) ? |
| 658 |
|
Data->directmvB[0].y |
| 659 |
|
: Data->currentMV[1].y - Data->referencemv[0].y); |
| 660 |
|
|
| 661 |
xvid_me_SubpelRefine(Data, CheckCandidate); |
*best_sad = Data->iMinSAD[0]; |
| 662 |
|
|
| 663 |
|
return skip_sad; |
| 664 |
} |
} |
| 665 |
|
|
| 666 |
*Data->iMinSAD += Data->lambda16; |
static void |
| 667 |
*best_sad = *Data->iMinSAD; |
SearchDirect_final( const uint32_t MotionFlags, |
| 668 |
|
const MACROBLOCK * const b_mb, |
| 669 |
|
int32_t * const best_sad, |
| 670 |
|
SearchData * const Data) |
| 671 |
|
|
| 672 |
|
{ |
| 673 |
|
CheckFunc * CheckCandidate = b_mb->mode == MODE_INTER4V ? |
| 674 |
|
CheckCandidateDirect : CheckCandidateDirectno4v; |
| 675 |
|
MainSearchFunc *MainSearchPtr; |
| 676 |
|
|
| 677 |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; |
| 678 |
else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; |
| 679 |
|
else MainSearchPtr = xvid_me_DiamondSearch; |
| 680 |
|
|
| 681 |
pMB->pmvs[3] = *Data->currentMV; |
Data->qpel_precision = 0; |
| 682 |
|
MainSearchPtr(0, 0, Data, 255, CheckCandidate); |
| 683 |
|
|
| 684 |
for (k = 0; k < 4; k++) { |
Data->qpel_precision = Data->qpel; |
|
pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; |
|
|
pMB->b_mvs[k].x = ( (Data->currentMV->x == 0) |
|
|
? Data->directmvB[k].x |
|
|
:pMB->mvs[k].x - Data->referencemv[k].x); |
|
|
pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); |
|
|
pMB->b_mvs[k].y = ((Data->currentMV->y == 0) |
|
|
? Data->directmvB[k].y |
|
|
: pMB->mvs[k].y - Data->referencemv[k].y); |
|
| 685 |
if (Data->qpel) { |
if (Data->qpel) { |
| 686 |
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
*Data->iMinSAD = 256*4096; /* this old SAD was not real, it was in hpel precision */ |
| 687 |
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
CheckCandidate(Data->currentMV->x, Data->currentMV->y, Data, 255); |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
|
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
|
| 688 |
} |
} |
| 689 |
|
|
| 690 |
if (b_mb->mode != MODE_INTER4V) { |
xvid_me_SubpelRefine(Data->currentMV[0], Data, CheckCandidate, 0); |
| 691 |
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
|
| 692 |
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
if (Data->iMinSAD[0] < *best_sad) { |
| 693 |
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
*best_sad = Data->iMinSAD[0]; |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
|
|
break; |
|
|
} |
|
| 694 |
} |
} |
| 695 |
return skip_sad; |
|
| 696 |
} |
} |
| 697 |
|
|
| 698 |
|
|
| 699 |
static void set_range(int * range, SearchData * Data) |
static __inline void |
| 700 |
|
set_range(int * range, SearchData * Data) |
| 701 |
{ |
{ |
| 702 |
Data->min_dx = range[0]; |
Data->min_dx = range[0]; |
| 703 |
Data->max_dx = range[1]; |
Data->max_dx = range[1]; |
| 706 |
} |
} |
| 707 |
|
|
| 708 |
static void |
static void |
| 709 |
SubpelRefine_dir(SearchData * const data, CheckFunc * const CheckCandidate, const int dir) |
SearchInterpolate_initial( |
|
{ |
|
|
/* Do a half-pel or q-pel refinement */ |
|
|
const VECTOR centerMV = data->qpel_precision ? |
|
|
data->currentQMV[dir-1] : data->currentMV[dir-1]; |
|
|
|
|
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, dir); |
|
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, dir); |
|
|
} |
|
|
|
|
|
/* Pretty much redundant code, just as SubpelRefine_dir above too |
|
|
* |
|
|
* TODO: Get rid off all the redundancy (SubpelRefine_Fast_dir, |
|
|
* CheckCandidate16no4v_qpel etc.) */ |
|
|
|
|
|
void |
|
|
SubpelRefine_Fast_dir(SearchData * data, CheckFunc * CheckCandidate, const int dir) |
|
|
{ |
|
|
/* Do a fast q-pel refinement */ |
|
|
VECTOR centerMV; |
|
|
VECTOR second_best; |
|
|
int best_sad = *data->iMinSAD; |
|
|
int xo, yo, xo2, yo2; |
|
|
int size = 2; |
|
|
data->iMinSAD2 = 0; |
|
|
|
|
|
/* check all halfpixel positions near our best halfpel position */ |
|
|
centerMV = data->currentQMV[dir-1]; |
|
|
*data->iMinSAD = 256 * 4096; |
|
|
|
|
|
CHECK_CANDIDATE(centerMV.x, centerMV.y - size, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y - size, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y, dir); |
|
|
CHECK_CANDIDATE(centerMV.x + size, centerMV.y + size, dir); |
|
|
|
|
|
CHECK_CANDIDATE(centerMV.x, centerMV.y + size, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y + size, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y, dir); |
|
|
CHECK_CANDIDATE(centerMV.x - size, centerMV.y - size, dir); |
|
|
|
|
|
second_best = data->currentQMV[dir-1]; |
|
|
|
|
|
/* after second_best has been found, go back to the vector we began with */ |
|
|
|
|
|
data->currentQMV[dir-1] = centerMV; |
|
|
*data->iMinSAD = best_sad; |
|
|
|
|
|
xo = centerMV.x; |
|
|
yo = centerMV.y; |
|
|
xo2 = second_best.x; |
|
|
yo2 = second_best.y; |
|
|
|
|
|
data->iMinSAD2 = 256 * 4096; |
|
|
|
|
|
if (yo == yo2) { |
|
|
CHECK_CANDIDATE((xo+xo2)>>1, yo, dir); |
|
|
CHECK_CANDIDATE(xo, yo-1, dir); |
|
|
CHECK_CANDIDATE(xo, yo+1, dir); |
|
|
|
|
|
if(best_sad <= data->iMinSAD2) return; |
|
|
|
|
|
if(data->currentQMV[dir-1].x == data->currentQMV2.x) { |
|
|
CHECK_CANDIDATE((xo+xo2)>>1, yo-1, dir); |
|
|
CHECK_CANDIDATE((xo+xo2)>>1, yo+1, dir); |
|
|
} else { |
|
|
CHECK_CANDIDATE((xo+xo2)>>1, |
|
|
(data->currentQMV[dir-1].x == xo) ? data->currentQMV[dir-1].y : data->currentQMV2.y, dir); |
|
|
} |
|
|
return; |
|
|
} |
|
|
|
|
|
if (xo == xo2) { |
|
|
CHECK_CANDIDATE(xo, (yo+yo2)>>1, dir); |
|
|
CHECK_CANDIDATE(xo-1, yo, dir); |
|
|
CHECK_CANDIDATE(xo+1, yo, dir); |
|
|
|
|
|
if(best_sad < data->iMinSAD2) return; |
|
|
|
|
|
if(data->currentQMV[dir-1].y == data->currentQMV2.y) { |
|
|
CHECK_CANDIDATE(xo-1, (yo+yo2)>>1, dir); |
|
|
CHECK_CANDIDATE(xo+1, (yo+yo2)>>1, dir); |
|
|
} else { |
|
|
CHECK_CANDIDATE((data->currentQMV[dir-1].y == yo) ? data->currentQMV[dir-1].x : data->currentQMV2.x, (yo+yo2)>>1, dir); |
|
|
} |
|
|
return; |
|
|
} |
|
|
|
|
|
CHECK_CANDIDATE(xo, (yo+yo2)>>1, dir); |
|
|
CHECK_CANDIDATE((xo+xo2)>>1, yo, dir); |
|
|
|
|
|
if(best_sad <= data->iMinSAD2) return; |
|
|
|
|
|
CHECK_CANDIDATE((xo+xo2)>>1, (yo+yo2)>>1, dir); |
|
|
} |
|
|
|
|
|
static void |
|
|
SearchInterpolate(const IMAGE * const f_Ref, |
|
|
const uint8_t * const f_RefH, |
|
|
const uint8_t * const f_RefV, |
|
|
const uint8_t * const f_RefHV, |
|
|
const IMAGE * const b_Ref, |
|
|
const uint8_t * const b_RefH, |
|
|
const uint8_t * const b_RefV, |
|
|
const uint8_t * const b_RefHV, |
|
| 710 |
const int x, const int y, |
const int x, const int y, |
|
const uint32_t fcode, |
|
|
const uint32_t bcode, |
|
| 711 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
| 712 |
const MBParam * const pParam, |
const MBParam * const pParam, |
| 713 |
const VECTOR * const f_predMV, |
const VECTOR * const f_predMV, |
| 714 |
const VECTOR * const b_predMV, |
const VECTOR * const b_predMV, |
|
MACROBLOCK * const pMB, |
|
| 715 |
int32_t * const best_sad, |
int32_t * const best_sad, |
| 716 |
SearchData * const Data) |
SearchData * const Data, |
| 717 |
|
const VECTOR startF, |
| 718 |
|
const VECTOR startB) |
| 719 |
|
|
| 720 |
{ |
{ |
|
int i, j; |
|
| 721 |
int b_range[4], f_range[4]; |
int b_range[4], f_range[4]; |
|
int threshA = (MotionFlags & XVID_ME_FAST_MODEINTERPOLATE) ? 250 : 500; |
|
|
int threshB = (MotionFlags & XVID_ME_FAST_MODEINTERPOLATE) ? 150 : 300; |
|
| 722 |
|
|
| 723 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
|
*Data->iMinSAD = 4096*256; |
|
|
Data->iFcode = fcode; Data->bFcode = bcode; |
|
|
|
|
|
i = (x + y * Data->iEdgedWidth) * 16; |
|
|
|
|
|
Data->RefP[0] = f_Ref->y + i; |
|
|
Data->RefP[2] = f_RefH + i; |
|
|
Data->RefP[1] = f_RefV + i; |
|
|
Data->RefP[3] = f_RefHV + i; |
|
|
Data->b_RefP[0] = b_Ref->y + i; |
|
|
Data->b_RefP[2] = b_RefH + i; |
|
|
Data->b_RefP[1] = b_RefV + i; |
|
|
Data->b_RefP[3] = b_RefHV + i; |
|
|
Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
|
Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; |
|
| 724 |
|
|
| 725 |
Data->predMV = *f_predMV; |
Data->predMV = *f_predMV; |
| 726 |
Data->bpredMV = *b_predMV; |
Data->bpredMV = *b_predMV; |
| 727 |
|
|
| 728 |
Data->currentMV[0] = Data->currentMV[2]; /* forward search left its vector here */ |
Data->currentMV[0] = startF; |
| 729 |
|
Data->currentMV[1] = startB; |
| 730 |
|
|
| 731 |
get_range(f_range, f_range+1, f_range+2, f_range+3, x, y, 4, pParam->width, pParam->height, fcode - Data->qpel, 1, 0); |
get_range(f_range, f_range+1, f_range+2, f_range+3, x, y, 4, pParam->width, pParam->height, Data->iFcode - Data->qpel, 1); |
| 732 |
get_range(b_range, b_range+1, b_range+2, b_range+3, x, y, 4, pParam->width, pParam->height, bcode - Data->qpel, 1, 0); |
get_range(b_range, b_range+1, b_range+2, b_range+3, x, y, 4, pParam->width, pParam->height, Data->bFcode - Data->qpel, 1); |
| 733 |
|
|
| 734 |
if (Data->currentMV[0].x > f_range[1]) Data->currentMV[0].x = f_range[1]; |
if (Data->currentMV[0].x > f_range[1]) Data->currentMV[0].x = f_range[1]; |
| 735 |
if (Data->currentMV[0].x < f_range[0]) Data->currentMV[0].x = f_range[0]; |
if (Data->currentMV[0].x < f_range[0]) Data->currentMV[0].x = f_range[0]; |
| 745 |
|
|
| 746 |
CheckCandidateInt(Data->currentMV[0].x, Data->currentMV[0].y, Data, 1); |
CheckCandidateInt(Data->currentMV[0].x, Data->currentMV[0].y, Data, 1); |
| 747 |
|
|
| 748 |
|
if (Data->iMinSAD[0] < *best_sad) *best_sad = Data->iMinSAD[0]; |
| 749 |
|
} |
| 750 |
|
|
| 751 |
|
static void |
| 752 |
|
SearchInterpolate_final(const int x, const int y, |
| 753 |
|
const uint32_t MotionFlags, |
| 754 |
|
const MBParam * const pParam, |
| 755 |
|
int32_t * const best_sad, |
| 756 |
|
SearchData * const Data) |
| 757 |
|
{ |
| 758 |
|
int i, j; |
| 759 |
|
int b_range[4], f_range[4]; |
| 760 |
|
|
| 761 |
|
get_range(f_range, f_range+1, f_range+2, f_range+3, x, y, 4, pParam->width, pParam->height, Data->iFcode - Data->qpel, 1); |
| 762 |
|
get_range(b_range, b_range+1, b_range+2, b_range+3, x, y, 4, pParam->width, pParam->height, Data->bFcode - Data->qpel, 1); |
| 763 |
|
|
| 764 |
/* diamond */ |
/* diamond */ |
| 765 |
do { |
do { |
| 766 |
Data->dir = 0; |
Data->dir = 0; |
| 787 |
|
|
| 788 |
/* qpel refinement */ |
/* qpel refinement */ |
| 789 |
if (Data->qpel) { |
if (Data->qpel) { |
|
if (*Data->iMinSAD > *best_sad + threshA) return; |
|
| 790 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
| 791 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, |
| 792 |
|
x, y, 4, pParam->width, pParam->height, Data->iFcode, 2); |
| 793 |
|
|
| 794 |
Data->currentQMV[0].x = 2 * Data->currentMV[0].x; |
Data->currentQMV[0].x = 2 * Data->currentMV[0].x; |
| 795 |
Data->currentQMV[0].y = 2 * Data->currentMV[0].y; |
Data->currentQMV[0].y = 2 * Data->currentMV[0].y; |
| 797 |
Data->currentQMV[1].y = 2 * Data->currentMV[1].y; |
Data->currentQMV[1].y = 2 * Data->currentMV[1].y; |
| 798 |
|
|
| 799 |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
| 800 |
if (MotionFlags & XVID_ME_FASTREFINE16) |
xvid_me_SubpelRefine(Data->currentQMV[0], Data, CheckCandidateInt, 1); |
| 801 |
SubpelRefine_Fast_dir(Data, CheckCandidateInt_qpel, 1); |
|
| 802 |
else |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, |
| 803 |
SubpelRefine_dir(Data, CheckCandidateInt, 1); |
x, y, 4, pParam->width, pParam->height, Data->bFcode, 2); |
| 804 |
|
|
| 805 |
|
xvid_me_SubpelRefine(Data->currentQMV[1], Data, CheckCandidateInt, 2); |
| 806 |
|
} |
| 807 |
} |
} |
| 808 |
|
|
| 809 |
if (*Data->iMinSAD > *best_sad + threshB) return; |
if (Data->iMinSAD[0] < *best_sad) *best_sad = Data->iMinSAD[0]; |
| 810 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0); |
} |
| 811 |
|
|
| 812 |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
static void |
| 813 |
if (MotionFlags & XVID_ME_FASTREFINE16) |
ModeDecision_BVOP_SAD(const SearchData * const Data_d, |
| 814 |
SubpelRefine_Fast_dir(Data, CheckCandidateInt_qpel, 2); |
const SearchData * const Data_b, |
| 815 |
else |
const SearchData * const Data_f, |
| 816 |
SubpelRefine_dir(Data, CheckCandidateInt, 2); |
const SearchData * const Data_i, |
| 817 |
|
MACROBLOCK * const pMB, |
| 818 |
|
const MACROBLOCK * const b_mb, |
| 819 |
|
VECTOR * f_predMV, |
| 820 |
|
VECTOR * b_predMV, |
| 821 |
|
int force_direct) |
| 822 |
|
{ |
| 823 |
|
int mode = MODE_DIRECT, k; |
| 824 |
|
int best_sad, f_sad, b_sad, i_sad; |
| 825 |
|
const int qpel = Data_d->qpel; |
| 826 |
|
|
| 827 |
|
/* evaluate cost of all modes - quite simple in SAD */ |
| 828 |
|
best_sad = Data_d->iMinSAD[0] + 1*Data_d->lambda16; |
| 829 |
|
b_sad = Data_b->iMinSAD[0] + 3*Data_d->lambda16; |
| 830 |
|
f_sad = Data_f->iMinSAD[0] + 4*Data_d->lambda16; |
| 831 |
|
i_sad = Data_i->iMinSAD[0] + 2*Data_d->lambda16; |
| 832 |
|
|
| 833 |
|
if (force_direct) |
| 834 |
|
goto set_mode; /* bypass checks for non-direct modes */ |
| 835 |
|
|
| 836 |
|
if (b_sad < best_sad) { |
| 837 |
|
mode = MODE_BACKWARD; |
| 838 |
|
best_sad = b_sad; |
| 839 |
} |
} |
| 840 |
|
|
| 841 |
|
if (f_sad < best_sad) { |
| 842 |
|
mode = MODE_FORWARD; |
| 843 |
|
best_sad = f_sad; |
| 844 |
} |
} |
| 845 |
|
|
| 846 |
*Data->iMinSAD += 2 * Data->lambda16; /* two bits are needed to code interpolate mode. */ |
if (i_sad < best_sad) { |
| 847 |
|
mode = MODE_INTERPOLATE; |
| 848 |
|
best_sad = i_sad; |
| 849 |
|
} |
| 850 |
|
|
| 851 |
if (*Data->iMinSAD < *best_sad) { |
set_mode: |
| 852 |
*best_sad = *Data->iMinSAD; |
pMB->sad16 = best_sad; |
| 853 |
pMB->mvs[0] = Data->currentMV[0]; |
pMB->mode = mode; |
| 854 |
pMB->b_mvs[0] = Data->currentMV[1]; |
pMB->cbp = 63; |
| 855 |
pMB->mode = MODE_INTERPOLATE; |
|
| 856 |
if (Data->qpel) { |
switch (mode) { |
| 857 |
pMB->qmvs[0] = Data->currentQMV[0]; |
|
| 858 |
pMB->b_qmvs[0] = Data->currentQMV[1]; |
case MODE_DIRECT: |
| 859 |
|
if (!qpel && b_mb->mode != MODE_INTER4V) pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
| 860 |
|
|
| 861 |
|
pMB->pmvs[3] = Data_d->currentMV[0]; |
| 862 |
|
|
| 863 |
|
for (k = 0; k < 4; k++) { |
| 864 |
|
pMB->mvs[k].x = Data_d->directmvF[k].x + Data_d->currentMV->x; |
| 865 |
|
pMB->b_mvs[k].x = ( (Data_d->currentMV->x == 0) |
| 866 |
|
? Data_d->directmvB[k].x |
| 867 |
|
:pMB->mvs[k].x - Data_d->referencemv[k].x); |
| 868 |
|
pMB->mvs[k].y = (Data_d->directmvF[k].y + Data_d->currentMV->y); |
| 869 |
|
pMB->b_mvs[k].y = ((Data_d->currentMV->y == 0) |
| 870 |
|
? Data_d->directmvB[k].y |
| 871 |
|
: pMB->mvs[k].y - Data_d->referencemv[k].y); |
| 872 |
|
if (qpel) { |
| 873 |
|
pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; |
| 874 |
|
pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; |
| 875 |
|
pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; |
| 876 |
|
pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; |
| 877 |
|
} |
| 878 |
|
|
| 879 |
|
if (b_mb->mode != MODE_INTER4V) { |
| 880 |
|
pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; |
| 881 |
|
pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; |
| 882 |
|
pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; |
| 883 |
|
pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; |
| 884 |
|
break; |
| 885 |
|
} |
| 886 |
|
} |
| 887 |
|
break; |
| 888 |
|
|
| 889 |
|
case MODE_FORWARD: |
| 890 |
|
if (qpel) { |
| 891 |
|
pMB->pmvs[0].x = Data_f->currentQMV->x - f_predMV->x; |
| 892 |
|
pMB->pmvs[0].y = Data_f->currentQMV->y - f_predMV->y; |
| 893 |
|
pMB->qmvs[0] = *Data_f->currentQMV; |
| 894 |
|
*f_predMV = Data_f->currentQMV[0]; |
| 895 |
|
} else { |
| 896 |
|
pMB->pmvs[0].x = Data_f->currentMV->x - f_predMV->x; |
| 897 |
|
pMB->pmvs[0].y = Data_f->currentMV->y - f_predMV->y; |
| 898 |
|
*f_predMV = Data_f->currentMV[0]; |
| 899 |
|
} |
| 900 |
|
pMB->mvs[0] = *Data_f->currentMV; |
| 901 |
|
pMB->b_mvs[0] = *Data_b->currentMV; /* hint for future searches */ |
| 902 |
|
break; |
| 903 |
|
|
| 904 |
|
case MODE_BACKWARD: |
| 905 |
|
if (qpel) { |
| 906 |
|
pMB->pmvs[0].x = Data_b->currentQMV->x - b_predMV->x; |
| 907 |
|
pMB->pmvs[0].y = Data_b->currentQMV->y - b_predMV->y; |
| 908 |
|
pMB->b_qmvs[0] = *Data_b->currentQMV; |
| 909 |
|
*b_predMV = Data_b->currentQMV[0]; |
| 910 |
|
} else { |
| 911 |
|
pMB->pmvs[0].x = Data_b->currentMV->x - b_predMV->x; |
| 912 |
|
pMB->pmvs[0].y = Data_b->currentMV->y - b_predMV->y; |
| 913 |
|
*b_predMV = Data_b->currentMV[0]; |
| 914 |
|
} |
| 915 |
|
pMB->b_mvs[0] = *Data_b->currentMV; |
| 916 |
|
pMB->mvs[0] = *Data_f->currentMV; /* hint for future searches */ |
| 917 |
|
break; |
| 918 |
|
|
| 919 |
|
|
| 920 |
|
case MODE_INTERPOLATE: |
| 921 |
|
pMB->mvs[0] = Data_i->currentMV[0]; |
| 922 |
|
pMB->b_mvs[0] = Data_i->currentMV[1]; |
| 923 |
|
if (qpel) { |
| 924 |
|
pMB->qmvs[0] = Data_i->currentQMV[0]; |
| 925 |
|
pMB->b_qmvs[0] = Data_i->currentQMV[1]; |
| 926 |
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; |
| 927 |
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; |
| 928 |
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; |
| 929 |
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; |
| 930 |
|
*f_predMV = Data_i->currentQMV[0]; |
| 931 |
|
*b_predMV = Data_i->currentQMV[1]; |
| 932 |
} else { |
} else { |
| 933 |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; |
| 934 |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; |
| 935 |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; |
| 936 |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; |
| 937 |
|
*f_predMV = Data_i->currentMV[0]; |
| 938 |
|
*b_predMV = Data_i->currentMV[1]; |
| 939 |
} |
} |
| 940 |
|
break; |
| 941 |
} |
} |
| 942 |
} |
} |
| 943 |
|
|
| 944 |
|
static __inline void |
| 945 |
|
maxMotionBVOP(int * const MVmaxF, int * const MVmaxB, const MACROBLOCK * const pMB, const int qpel) |
| 946 |
|
{ |
| 947 |
|
if (pMB->mode == MODE_FORWARD || pMB->mode == MODE_INTERPOLATE) { |
| 948 |
|
const VECTOR * const mv = qpel ? pMB->qmvs : pMB->mvs; |
| 949 |
|
int max = *MVmaxF; |
| 950 |
|
if (mv[0].x > max) max = mv[0].x; |
| 951 |
|
else if (-mv[0].x - 1 > max) max = -mv[0].x - 1; |
| 952 |
|
if (mv[0].y > max) max = mv[0].y; |
| 953 |
|
else if (-mv[0].y - 1 > max) max = -mv[0].y - 1; |
| 954 |
|
|
| 955 |
|
*MVmaxF = max; |
| 956 |
|
} |
| 957 |
|
|
| 958 |
|
if (pMB->mode == MODE_BACKWARD || pMB->mode == MODE_INTERPOLATE) { |
| 959 |
|
const VECTOR * const mv = qpel ? pMB->b_qmvs : pMB->b_mvs; |
| 960 |
|
int max = *MVmaxB; |
| 961 |
|
if (mv[0].x > max) max = mv[0].x; |
| 962 |
|
else if (-mv[0].x - 1 > max) max = -mv[0].x - 1; |
| 963 |
|
if (mv[0].y > max) max = mv[0].y; |
| 964 |
|
else if (-mv[0].y - 1 > max) max = -mv[0].y - 1; |
| 965 |
|
*MVmaxB = max; |
| 966 |
|
} |
| 967 |
|
} |
| 968 |
|
|
| 969 |
|
|
| 970 |
void |
void |
| 971 |
MotionEstimationBVOP(MBParam * const pParam, |
MotionEstimationBVOP(MBParam * const pParam, |
| 972 |
FRAMEINFO * const frame, |
FRAMEINFO * const frame, |
| 983 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
| 984 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
| 985 |
const IMAGE * const b_refV, |
const IMAGE * const b_refV, |
| 986 |
const IMAGE * const b_refHV) |
const IMAGE * const b_refHV, |
| 987 |
|
const int num_slices) |
| 988 |
{ |
{ |
| 989 |
uint32_t i, j; |
uint32_t i, j; |
| 990 |
int32_t best_sad; |
int32_t best_sad = 256*4096; |
| 991 |
uint32_t skip_sad; |
uint32_t skip_sad; |
| 992 |
|
int fb_thresh; |
| 993 |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
const MACROBLOCK * const b_mbs = b_reference->mbs; |
|
MACROBLOCK *const pMBs = frame->mbs; |
|
| 994 |
|
|
| 995 |
VECTOR f_predMV, b_predMV; |
VECTOR f_predMV, b_predMV; |
| 996 |
|
|
| 997 |
|
int mb_width = pParam->mb_width; |
| 998 |
|
int mb_height = pParam->mb_height; |
| 999 |
|
int MVmaxF = 0, MVmaxB = 0; |
| 1000 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
| 1001 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
| 1002 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
| 1003 |
|
|
| 1004 |
/* some pre-inintialized data for the rest of the search */ |
/* some pre-inintialized data for the rest of the search */ |
| 1005 |
|
SearchData Data_d, Data_f, Data_b, Data_i; |
| 1006 |
|
memset(&Data_d, 0, sizeof(SearchData)); |
| 1007 |
|
|
| 1008 |
SearchData Data; |
Data_d.iEdgedWidth = pParam->edged_width; |
| 1009 |
memset(&Data, 0, sizeof(SearchData)); |
Data_d.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
| 1010 |
|
Data_d.rounding = 0; |
| 1011 |
Data.iEdgedWidth = pParam->edged_width; |
Data_d.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
| 1012 |
Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
Data_d.iQuant = frame->quant; |
| 1013 |
Data.rounding = 0; |
Data_d.quant_sq = frame->quant*frame->quant; |
| 1014 |
Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
Data_d.dctSpace = dct_space; |
| 1015 |
Data.iQuant = frame->quant; |
Data_d.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
| 1016 |
|
Data_d.mpeg_quant_matrices = pParam->mpeg_quant_matrices; |
| 1017 |
|
|
| 1018 |
|
Data_d.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
| 1019 |
|
|
| 1020 |
|
memcpy(&Data_f, &Data_d, sizeof(SearchData)); |
| 1021 |
|
memcpy(&Data_b, &Data_d, sizeof(SearchData)); |
| 1022 |
|
memcpy(&Data_i, &Data_d, sizeof(SearchData)); |
| 1023 |
|
|
| 1024 |
Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
Data_f.iFcode = Data_i.iFcode = frame->fcode = b_reference->fcode; |
| 1025 |
|
Data_b.iFcode = Data_i.bFcode = frame->bcode = b_reference->fcode; |
| 1026 |
|
|
|
/* note: i==horizontal, j==vertical */ |
|
| 1027 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
| 1028 |
|
int new_bound = mb_width * ((((j*num_slices) / mb_height) * mb_height + (num_slices-1)) / num_slices); |
| 1029 |
|
|
| 1030 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
| 1031 |
|
|
| 1032 |
for (i = 0; i < pParam->mb_width; i++) { |
for (i = 0; i < pParam->mb_width; i++) { |
| 1033 |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
| 1034 |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
| 1035 |
int interpol_search = 0; |
int force_direct = (((j*mb_width+i)==new_bound) && (j > 0)) ? 1 : 0; /* MTK decoder chipsets do NOT reset predMVs upon resync marker in BVOPs. We workaround this problem |
| 1036 |
int bf_search = 0; |
by placing the slice border on second MB in a row and then force the first MB to be direct mode */ |
| 1037 |
int bf_thresh = 0; |
|
| 1038 |
|
pMB->mode = -1; |
| 1039 |
|
|
| 1040 |
|
initialize_searchData(&Data_d, &Data_f, &Data_b, &Data_i, |
| 1041 |
|
i, j, f_ref, f_refH->y, f_refV->y, f_refHV->y, |
| 1042 |
|
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
| 1043 |
|
&frame->image, b_mb); |
| 1044 |
|
|
| 1045 |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
| 1046 |
if (b_reference->coding_type != S_VOP) |
if (b_reference->coding_type != S_VOP) |
| 1051 |
continue; |
continue; |
| 1052 |
} |
} |
| 1053 |
|
|
|
Data.lambda16 = xvid_me_lambda_vec16[b_mb->quant]; |
|
|
|
|
|
Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; |
|
|
Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; |
|
|
Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; |
|
|
|
|
| 1054 |
/* direct search comes first, because it (1) checks for SKIP-mode |
/* direct search comes first, because it (1) checks for SKIP-mode |
| 1055 |
and (2) sets very good predictions for forward and backward search */ |
and (2) sets very good predictions for forward and backward search */ |
| 1056 |
skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
skip_sad = SearchDirect_initial(i, j, frame->motion_flags, TRB, TRD, pParam, pMB, |
| 1057 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_mb, &best_sad, &Data_d); |
|
&frame->image, |
|
|
i, j, |
|
|
frame->motion_flags, |
|
|
TRB, TRD, |
|
|
pParam, |
|
|
pMB, b_mb, |
|
|
&best_sad, |
|
|
&Data); |
|
| 1058 |
|
|
| 1059 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { |
if (pMB->mode == MODE_DIRECT_NONE_MV) { |
| 1060 |
pMB->sad16 = best_sad; |
pMB->sad16 = best_sad; |
| 1061 |
|
pMB->cbp = 0; |
| 1062 |
continue; |
continue; |
| 1063 |
} |
} |
| 1064 |
|
|
| 1065 |
if (frame->motion_flags & XVID_ME_BFRAME_EARLYSTOP) { |
SearchBF_initial(i, j, frame->motion_flags, frame->fcode, pParam, pMB, |
| 1066 |
if(i > 0 && j > 0 && i < pParam->mb_width) { |
&f_predMV, &best_sad, MODE_FORWARD, &Data_f, Data_d.currentMV[1], new_bound); |
|
bf_thresh = MIN((&pMBs[(i-1) + j * pParam->mb_width])->sad16, |
|
|
MIN((&pMBs[i + (j-1) * pParam->mb_width])->sad16, |
|
|
(&pMBs[(i+1) + (j-1) * pParam->mb_width])->sad16)); |
|
|
|
|
|
if (((&pMBs[(i-1) + j * pParam->mb_width])->mode != MODE_FORWARD) && |
|
|
((&pMBs[(i-1) + j * pParam->mb_width])->mode != MODE_BACKWARD) && |
|
|
((&pMBs[(i-1) + j * pParam->mb_width])->mode != MODE_INTERPOLATE)) |
|
|
bf_search++; |
|
|
|
|
|
if (((&pMBs[i + (j - 1) * pParam->mb_width])->mode != MODE_FORWARD) && |
|
|
((&pMBs[i + (j - 1) * pParam->mb_width])->mode != MODE_BACKWARD) && |
|
|
((&pMBs[i + (j - 1) * pParam->mb_width])->mode != MODE_INTERPOLATE)) |
|
|
bf_search++; |
|
|
|
|
|
if (((&pMBs[(i + 1) + (j - 1) * pParam->mb_width])->mode != MODE_FORWARD) && |
|
|
((&pMBs[(i + 1) + (j - 1) * pParam->mb_width])->mode != MODE_BACKWARD) && |
|
|
((&pMBs[(i + 1) + (j - 1) * pParam->mb_width])->mode != MODE_INTERPOLATE)) |
|
|
bf_search++; |
|
|
} |
|
| 1067 |
|
|
| 1068 |
if ((best_sad < bf_thresh) && (bf_search == 3)) |
SearchBF_initial(i, j, frame->motion_flags, frame->bcode, pParam, pMB, |
| 1069 |
|
&b_predMV, &best_sad, MODE_BACKWARD, &Data_b, Data_d.currentMV[2], new_bound); |
| 1070 |
|
|
| 1071 |
|
if (frame->motion_flags&XVID_ME_BFRAME_EARLYSTOP) |
| 1072 |
|
fb_thresh = best_sad; |
| 1073 |
|
else |
| 1074 |
|
fb_thresh = best_sad + (best_sad>>1); |
| 1075 |
|
|
| 1076 |
|
if (Data_f.iMinSAD[0] <= fb_thresh) |
| 1077 |
|
SearchBF_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_f); |
| 1078 |
|
|
| 1079 |
|
if (Data_b.iMinSAD[0] <= fb_thresh) |
| 1080 |
|
SearchBF_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_b); |
| 1081 |
|
|
| 1082 |
|
SearchInterpolate_initial(i, j, frame->motion_flags, pParam, &f_predMV, &b_predMV, &best_sad, |
| 1083 |
|
&Data_i, Data_f.currentMV[0], Data_b.currentMV[0]); |
| 1084 |
|
|
| 1085 |
|
if (((Data_i.iMinSAD[0] < best_sad +(best_sad>>3)) && !(frame->motion_flags&XVID_ME_FAST_MODEINTERPOLATE)) |
| 1086 |
|
|| Data_i.iMinSAD[0] <= best_sad) |
| 1087 |
|
|
| 1088 |
|
SearchInterpolate_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_i); |
| 1089 |
|
|
| 1090 |
|
if (Data_d.iMinSAD[0] <= 2*best_sad) |
| 1091 |
|
if ((!(frame->motion_flags&XVID_ME_SKIP_DELTASEARCH) && (best_sad > 750)) |
| 1092 |
|
|| (best_sad > 1000)) |
| 1093 |
|
|
| 1094 |
|
SearchDirect_final(frame->motion_flags, b_mb, &best_sad, &Data_d); |
| 1095 |
|
|
| 1096 |
|
/* final skip decision */ |
| 1097 |
|
if ( (skip_sad < 2 * Data_d.iQuant * MAX_SAD00_FOR_SKIP ) |
| 1098 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) { |
| 1099 |
|
|
| 1100 |
|
Data_d.chromaSAD = 0; /* green light for chroma check */ |
| 1101 |
|
|
| 1102 |
|
SkipDecisionB(pMB, &Data_d); |
| 1103 |
|
|
| 1104 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { /* skipped? */ |
| 1105 |
|
pMB->sad16 = skip_sad; |
| 1106 |
|
pMB->cbp = 0; |
| 1107 |
continue; |
continue; |
| 1108 |
} |
} |
|
|
|
|
/* forward search */ |
|
|
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, |
|
|
frame->motion_flags, |
|
|
frame->fcode, pParam, |
|
|
pMB, &f_predMV, &best_sad, |
|
|
MODE_FORWARD, &Data); |
|
|
|
|
|
/* backward search */ |
|
|
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
frame->motion_flags, |
|
|
frame->bcode, pParam, |
|
|
pMB, &b_predMV, &best_sad, |
|
|
MODE_BACKWARD, &Data); |
|
|
|
|
|
/* interpolate search comes last, because it uses data from forward and backward as prediction */ |
|
|
if (frame->motion_flags & XVID_ME_FAST_MODEINTERPOLATE) { |
|
|
|
|
|
if(i > 0 && j > 0 && i < pParam->mb_width) { |
|
|
if ((&pMBs[(i-1) + j * pParam->mb_width])->mode == MODE_INTERPOLATE) |
|
|
interpol_search++; |
|
|
if ((&pMBs[i + (j - 1) * pParam->mb_width])->mode == MODE_INTERPOLATE) |
|
|
interpol_search++; |
|
|
if ((&pMBs[(i + 1) + (j - 1) * pParam->mb_width])->mode == MODE_INTERPOLATE) |
|
|
interpol_search++; |
|
| 1109 |
} |
} |
| 1110 |
|
|
| 1111 |
|
if (frame->vop_flags & XVID_VOP_RD_BVOP) |
| 1112 |
|
ModeDecision_BVOP_RD(&Data_d, &Data_b, &Data_f, &Data_i, |
| 1113 |
|
pMB, b_mb, &f_predMV, &b_predMV, frame->motion_flags, frame->vop_flags, pParam, i, j, best_sad, force_direct); |
| 1114 |
else |
else |
| 1115 |
interpol_search = 1; |
ModeDecision_BVOP_SAD(&Data_d, &Data_b, &Data_f, &Data_i, pMB, b_mb, &f_predMV, &b_predMV, force_direct); |
| 1116 |
|
|
| 1117 |
|
maxMotionBVOP(&MVmaxF, &MVmaxB, pMB, Data_d.qpel); |
| 1118 |
|
|
|
interpol_search |= !(best_sad < 3 * Data.iQuant * MAX_SAD00_FOR_SKIP * (Data.chroma ? 3:2)); |
|
| 1119 |
} |
} |
| 1120 |
else |
} |
| 1121 |
interpol_search = 1; |
|
| 1122 |
|
frame->fcode = getMinFcode(MVmaxF); |
| 1123 |
|
frame->bcode = getMinFcode(MVmaxB); |
| 1124 |
|
} |
| 1125 |
|
|
| 1126 |
|
|
| 1127 |
|
|
| 1128 |
|
void |
| 1129 |
|
SMPMotionEstimationBVOP(SMPData * h) |
| 1130 |
|
{ |
| 1131 |
|
Encoder *pEnc = (Encoder *) h->pEnc; |
| 1132 |
|
|
| 1133 |
if (interpol_search) { |
const MBParam * const pParam = &pEnc->mbParam; |
| 1134 |
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
const FRAMEINFO * const frame = h->current; |
| 1135 |
|
const int32_t time_bp = (int32_t)(pEnc->current->stamp - frame->stamp); |
| 1136 |
|
const int32_t time_pp = (int32_t)(pEnc->current->stamp - pEnc->reference->stamp); |
| 1137 |
|
/* forward (past) reference */ |
| 1138 |
|
const IMAGE * const f_ref = &pEnc->reference->image; |
| 1139 |
|
const IMAGE * const f_refH = &pEnc->f_refh; |
| 1140 |
|
const IMAGE * const f_refV = &pEnc->f_refv; |
| 1141 |
|
const IMAGE * const f_refHV = &pEnc->f_refhv; |
| 1142 |
|
/* backward (future) reference */ |
| 1143 |
|
const FRAMEINFO * const b_reference = pEnc->current; |
| 1144 |
|
const IMAGE * const b_ref = &pEnc->current->image; |
| 1145 |
|
const IMAGE * const b_refH = &pEnc->vInterH; |
| 1146 |
|
const IMAGE * const b_refV = &pEnc->vInterV; |
| 1147 |
|
const IMAGE * const b_refHV = &pEnc->vInterHV; |
| 1148 |
|
|
| 1149 |
|
int mb_width = pParam->mb_width; |
| 1150 |
|
int mb_height = pParam->mb_height; |
| 1151 |
|
int num_slices = pEnc->num_slices; |
| 1152 |
|
int y_row = h->y_row; |
| 1153 |
|
int y_step = h->y_step; |
| 1154 |
|
int start_y = h->start_y; |
| 1155 |
|
int stop_y = h->stop_y; |
| 1156 |
|
int * complete_count_self = h->complete_count_self; |
| 1157 |
|
const int * complete_count_above = h->complete_count_above; |
| 1158 |
|
int max_mbs; |
| 1159 |
|
int current_mb = 0; |
| 1160 |
|
|
| 1161 |
|
int32_t i, j; |
| 1162 |
|
int32_t best_sad = 256*4096; |
| 1163 |
|
uint32_t skip_sad; |
| 1164 |
|
int fb_thresh; |
| 1165 |
|
const MACROBLOCK * const b_mbs = b_reference->mbs; |
| 1166 |
|
|
| 1167 |
|
VECTOR f_predMV, b_predMV; |
| 1168 |
|
|
| 1169 |
|
int MVmaxF = 0, MVmaxB = 0; |
| 1170 |
|
const int32_t TRB = time_pp - time_bp; |
| 1171 |
|
const int32_t TRD = time_pp; |
| 1172 |
|
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
| 1173 |
|
|
| 1174 |
|
/* some pre-inintialized data for the rest of the search */ |
| 1175 |
|
SearchData Data_d, Data_f, Data_b, Data_i; |
| 1176 |
|
memset(&Data_d, 0, sizeof(SearchData)); |
| 1177 |
|
|
| 1178 |
|
Data_d.iEdgedWidth = pParam->edged_width; |
| 1179 |
|
Data_d.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; |
| 1180 |
|
Data_d.rounding = 0; |
| 1181 |
|
Data_d.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; |
| 1182 |
|
Data_d.iQuant = frame->quant; |
| 1183 |
|
Data_d.quant_sq = frame->quant*frame->quant; |
| 1184 |
|
Data_d.dctSpace = dct_space; |
| 1185 |
|
Data_d.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
| 1186 |
|
Data_d.mpeg_quant_matrices = pParam->mpeg_quant_matrices; |
| 1187 |
|
|
| 1188 |
|
Data_d.RefQ = h->RefQ; |
| 1189 |
|
|
| 1190 |
|
memcpy(&Data_f, &Data_d, sizeof(SearchData)); |
| 1191 |
|
memcpy(&Data_b, &Data_d, sizeof(SearchData)); |
| 1192 |
|
memcpy(&Data_i, &Data_d, sizeof(SearchData)); |
| 1193 |
|
|
| 1194 |
|
Data_f.iFcode = Data_i.iFcode = frame->fcode; |
| 1195 |
|
Data_b.iFcode = Data_i.bFcode = frame->bcode; |
| 1196 |
|
|
| 1197 |
|
max_mbs = 0; |
| 1198 |
|
|
| 1199 |
|
for (j = (start_y+y_row); j < stop_y; j += y_step) { |
| 1200 |
|
int new_bound = mb_width * ((((j*num_slices) / mb_height) * mb_height + (num_slices-1)) / num_slices); |
| 1201 |
|
|
| 1202 |
|
if (j == start_y) max_mbs = pParam->mb_width; /* we can process all blocks of the first row */ |
| 1203 |
|
|
| 1204 |
|
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
| 1205 |
|
|
| 1206 |
|
for (i = 0; i < (int) pParam->mb_width; i++) { |
| 1207 |
|
MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; |
| 1208 |
|
const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; |
| 1209 |
|
int force_direct = (((j*mb_width+i)==new_bound) && (j > 0)) ? 1 : 0; /* MTK decoder chipsets do NOT reset predMVs upon resync marker in BVOPs. We workaround this problem |
| 1210 |
|
by placing the slice border on second MB in a row and then force the first MB to be direct mode */ |
| 1211 |
|
pMB->mode = -1; |
| 1212 |
|
|
| 1213 |
|
initialize_searchData(&Data_d, &Data_f, &Data_b, &Data_i, |
| 1214 |
|
i, j, f_ref, f_refH->y, f_refV->y, f_refHV->y, |
| 1215 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
| 1216 |
i, j, |
&frame->image, b_mb); |
| 1217 |
frame->fcode, frame->bcode, |
|
| 1218 |
frame->motion_flags, |
if (current_mb >= max_mbs) { |
| 1219 |
pParam, |
/* we ME-ed all macroblocks we safely could. grab next portion */ |
| 1220 |
&f_predMV, &b_predMV, |
int above_count = *complete_count_above; /* sync point */ |
| 1221 |
pMB, &best_sad, |
if (above_count == pParam->mb_width) { |
| 1222 |
&Data); |
/* full line above is ready */ |
| 1223 |
|
above_count = pParam->mb_width+1; |
| 1224 |
|
if (j < stop_y-y_step) { |
| 1225 |
|
/* this is not last line, grab a portion of MBs from the next line too */ |
| 1226 |
|
above_count += MAX(0, complete_count_above[1] - 1); |
| 1227 |
|
} |
| 1228 |
} |
} |
| 1229 |
|
|
| 1230 |
/* final skip decision */ |
max_mbs = current_mb + above_count - i - 1; |
| 1231 |
if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * (Data.chroma ? 3:2) ) |
|
| 1232 |
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
if (current_mb >= max_mbs) { |
| 1233 |
|
/* current workload is zero */ |
| 1234 |
|
i--; |
| 1235 |
|
sched_yield(); |
| 1236 |
|
continue; |
| 1237 |
|
} |
| 1238 |
|
} |
| 1239 |
|
|
| 1240 |
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ |
| 1241 |
|
if (b_reference->coding_type != S_VOP) |
| 1242 |
|
if (b_mb->mode == MODE_NOT_CODED) { |
| 1243 |
|
pMB->mode = MODE_NOT_CODED; |
| 1244 |
|
pMB->mvs[0] = pMB->b_mvs[0] = zeroMV; |
| 1245 |
|
pMB->sad16 = 0; |
| 1246 |
|
*complete_count_self = i+1; |
| 1247 |
|
current_mb++; |
| 1248 |
|
continue; |
| 1249 |
|
} |
| 1250 |
|
|
| 1251 |
switch (pMB->mode) { |
/* direct search comes first, because it (1) checks for SKIP-mode |
| 1252 |
case MODE_FORWARD: |
and (2) sets very good predictions for forward and backward search */ |
| 1253 |
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
skip_sad = SearchDirect_initial(i, j, frame->motion_flags, TRB, TRD, pParam, pMB, |
| 1254 |
pMB->sad16 = best_sad; |
b_mb, &best_sad, &Data_d); |
| 1255 |
break; |
|
| 1256 |
case MODE_BACKWARD: |
if (pMB->mode == MODE_DIRECT_NONE_MV) { |
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
|
pMB->sad16 = best_sad; |
|
|
break; |
|
|
case MODE_INTERPOLATE: |
|
|
f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; |
|
|
b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; |
|
|
pMB->sad16 = best_sad; |
|
|
break; |
|
|
default: |
|
| 1257 |
pMB->sad16 = best_sad; |
pMB->sad16 = best_sad; |
| 1258 |
break; |
pMB->cbp = 0; |
| 1259 |
|
*complete_count_self = i+1; |
| 1260 |
|
current_mb++; |
| 1261 |
|
continue; |
| 1262 |
|
} |
| 1263 |
|
|
| 1264 |
|
SearchBF_initial(i, j, frame->motion_flags, frame->fcode, pParam, pMB, |
| 1265 |
|
&f_predMV, &best_sad, MODE_FORWARD, &Data_f, Data_d.currentMV[1], new_bound); |
| 1266 |
|
|
| 1267 |
|
SearchBF_initial(i, j, frame->motion_flags, frame->bcode, pParam, pMB, |
| 1268 |
|
&b_predMV, &best_sad, MODE_BACKWARD, &Data_b, Data_d.currentMV[2], new_bound); |
| 1269 |
|
|
| 1270 |
|
if (frame->motion_flags&XVID_ME_BFRAME_EARLYSTOP) |
| 1271 |
|
fb_thresh = best_sad; |
| 1272 |
|
else |
| 1273 |
|
fb_thresh = best_sad + (best_sad>>1); |
| 1274 |
|
|
| 1275 |
|
if (Data_f.iMinSAD[0] <= fb_thresh) |
| 1276 |
|
SearchBF_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_f); |
| 1277 |
|
|
| 1278 |
|
if (Data_b.iMinSAD[0] <= fb_thresh) |
| 1279 |
|
SearchBF_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_b); |
| 1280 |
|
|
| 1281 |
|
SearchInterpolate_initial(i, j, frame->motion_flags, pParam, &f_predMV, &b_predMV, &best_sad, |
| 1282 |
|
&Data_i, Data_f.currentMV[0], Data_b.currentMV[0]); |
| 1283 |
|
|
| 1284 |
|
if (((Data_i.iMinSAD[0] < best_sad +(best_sad>>3)) && !(frame->motion_flags&XVID_ME_FAST_MODEINTERPOLATE)) |
| 1285 |
|
|| Data_i.iMinSAD[0] <= best_sad) |
| 1286 |
|
|
| 1287 |
|
SearchInterpolate_final(i, j, frame->motion_flags, pParam, &best_sad, &Data_i); |
| 1288 |
|
|
| 1289 |
|
if (Data_d.iMinSAD[0] <= 2*best_sad) |
| 1290 |
|
if ((!(frame->motion_flags&XVID_ME_SKIP_DELTASEARCH) && (best_sad > 750)) |
| 1291 |
|
|| (best_sad > 1000)) |
| 1292 |
|
|
| 1293 |
|
SearchDirect_final(frame->motion_flags, b_mb, &best_sad, &Data_d); |
| 1294 |
|
|
| 1295 |
|
/* final skip decision */ |
| 1296 |
|
if ( (skip_sad < 2 * Data_d.iQuant * MAX_SAD00_FOR_SKIP ) |
| 1297 |
|
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) { |
| 1298 |
|
|
| 1299 |
|
Data_d.chromaSAD = 0; /* green light for chroma check */ |
| 1300 |
|
|
| 1301 |
|
SkipDecisionB(pMB, &Data_d); |
| 1302 |
|
|
| 1303 |
|
if (pMB->mode == MODE_DIRECT_NONE_MV) { /* skipped? */ |
| 1304 |
|
pMB->sad16 = skip_sad; |
| 1305 |
|
pMB->cbp = 0; |
| 1306 |
|
*complete_count_self = i+1; |
| 1307 |
|
current_mb++; |
| 1308 |
|
continue; |
| 1309 |
} |
} |
| 1310 |
} |
} |
| 1311 |
|
|
| 1312 |
|
if (frame->vop_flags & XVID_VOP_RD_BVOP) |
| 1313 |
|
ModeDecision_BVOP_RD(&Data_d, &Data_b, &Data_f, &Data_i, |
| 1314 |
|
pMB, b_mb, &f_predMV, &b_predMV, frame->motion_flags, frame->vop_flags, pParam, i, j, best_sad, force_direct); |
| 1315 |
|
else |
| 1316 |
|
ModeDecision_BVOP_SAD(&Data_d, &Data_b, &Data_f, &Data_i, pMB, b_mb, &f_predMV, &b_predMV, force_direct); |
| 1317 |
|
|
| 1318 |
|
*complete_count_self = i+1; |
| 1319 |
|
current_mb++; |
| 1320 |
|
maxMotionBVOP(&MVmaxF, &MVmaxB, pMB, Data_d.qpel); |
| 1321 |
} |
} |
| 1322 |
|
|
| 1323 |
|
complete_count_self++; |
| 1324 |
|
complete_count_above++; |
| 1325 |
|
} |
| 1326 |
|
|
| 1327 |
|
h->minfcode = getMinFcode(MVmaxF); |
| 1328 |
|
h->minbcode = getMinFcode(MVmaxB); |
| 1329 |
} |
} |
Parent Directory
| 
Revision Log
| 
Patch