--- branches/dev-api-4/xvidcore/src/motion/estimation_bvop.c 2003/11/19 12:24:25 1215 +++ branches/dev-api-4/xvidcore/src/motion/estimation_bvop.c 2003/12/18 14:47:44 1276 @@ -21,7 +21,7 @@ * along with this program ; if not, write to the Free Software * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * - * $Id: estimation_bvop.c,v 1.1.2.5 2003-11-19 12:24:25 syskin Exp $ + * $Id: estimation_bvop.c,v 1.1.2.11 2003-12-18 14:47:44 edgomez Exp $ * ****************************************************************************/ @@ -39,7 +39,6 @@ #include "sad.h" #include "motion_inlines.h" - static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, SearchData * const data) @@ -298,6 +297,44 @@ } } +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); + + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); + sad += (data->lambda16 * t * sad)>>10; + + if (data->chroma && sad < *data->iMinSAD) + sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); + + if (sad < *(data->iMinSAD)) { + data->iMinSAD2 = *(data->iMinSAD); + data->currentQMV2.x = data->currentQMV->x; + data->currentQMV2.y = data->currentQMV->y; + + data->iMinSAD[0] = sad; + 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; + } +} + static __inline VECTOR ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) { @@ -361,11 +398,10 @@ int i; VECTOR pmv[7]; - MainSearchFunc *MainSearchPtr; *Data->iMinSAD = MV_MAX_ERROR; Data->iFcode = iFcode; Data->qpel_precision = 0; - Data->chromaX = Data->chromaX = Data->chromaSAD = 256*4096; /* reset chroma-sad cache */ + 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; @@ -391,12 +427,14 @@ if (!vector_repeats(pmv, i) ) CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i); - if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; - else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; - else MainSearchPtr = xvid_me_DiamondSearch; - if (*Data->iMinSAD > 512) { unsigned int mask = make_mask(pmv, 7, Data->dir); + + MainSearchFunc *MainSearchPtr; + if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; + else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; + else MainSearchPtr = xvid_me_DiamondSearch; + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); } @@ -408,7 +446,13 @@ Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, pParam->width, pParam->height, iFcode, 2, 0); - xvid_me_SubpelRefine(Data, CheckCandidate16no4v); + + if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { + if(MotionFlags & XVID_ME_FASTREFINE16) + SubpelRefine_Fast(Data, CheckCandidate16no4v_qpel); + else + xvid_me_SubpelRefine(Data, CheckCandidate16no4v); + } } /* three bits are needed to code backward mode. four for forward */ @@ -446,42 +490,46 @@ const uint32_t x, const uint32_t y, const SearchData * const Data) { - int dx = 0, dy = 0, b_dx = 0, b_dy = 0; - int32_t sum; int k; - const uint32_t stride = Data->iEdgedWidth/2; - /* this is not full chroma compensation, only it's fullpel approximation. should work though */ - for (k = 0; k < 4; k++) { - dy += Data->directmvF[k].y >> Data->qpel; - dx += Data->directmvF[k].x >> Data->qpel; - b_dy += Data->directmvB[k].y >> Data->qpel; - b_dx += Data->directmvB[k].x >> Data->qpel; - } - - dy = (dy >> 3) + roundtab_76[dy & 0xf]; - dx = (dx >> 3) + roundtab_76[dx & 0xf]; - b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; - b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; - - sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, - f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, - b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, - stride); - - if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ - - sum += sad8bi(pCur->v + 8*x + 8 * y * stride, - f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, - b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, - stride); + if (!Data->chroma) { + int dx = 0, dy = 0, b_dx = 0, b_dy = 0; + int32_t sum; + const uint32_t stride = Data->iEdgedWidth/2; + /* this is not full chroma compensation, only it's fullpel approximation. should work though */ - if (sum < MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) { - pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ for (k = 0; k < 4; k++) { - pMB->qmvs[k] = pMB->mvs[k] = Data->directmvF[k]; - pMB->b_qmvs[k] = pMB->b_mvs[k] = Data->directmvB[k]; + dy += Data->directmvF[k].y >> Data->qpel; + dx += Data->directmvF[k].x >> Data->qpel; + b_dy += Data->directmvB[k].y >> Data->qpel; + b_dx += Data->directmvB[k].x >> Data->qpel; } + + dy = (dy >> 3) + roundtab_76[dy & 0xf]; + dx = (dx >> 3) + roundtab_76[dx & 0xf]; + b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; + b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; + + sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, + f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); + + if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ + + sum += sad8bi(pCur->v + 8*x + 8 * y * stride, + f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); + + if (sum >= MAX_CHROMA_SAD_FOR_SKIP * (int)Data->iQuant) return; /* no skip */ + } + + /* skip */ + pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ + for (k = 0; k < 4; k++) { + pMB->qmvs[k] = pMB->mvs[k] = Data->directmvF[k]; + pMB->b_qmvs[k] = pMB->b_mvs[k] = Data->directmvB[k]; } } @@ -563,30 +611,22 @@ /* initial (fast) skip decision */ if (*Data->iMinSAD < (int)Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { /* possible skip */ - if (Data->chroma) { - pMB->mode = MODE_DIRECT_NONE_MV; - return *Data->iMinSAD; /* skip. */ - } else { - SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); - if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ - } + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skipped */ } *Data->iMinSAD += Data->lambda16; skip_sad = *Data->iMinSAD; - /* - * DIRECT MODE DELTA VECTOR SEARCH. - * This has to be made more effective, but at the moment I'm happy it's running at all - */ + if (!(MotionFlags & XVID_ME_SKIP_DELTASEARCH)) { + if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; + else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; + else MainSearchPtr = xvid_me_DiamondSearch; - if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch; - else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch; - else MainSearchPtr = xvid_me_DiamondSearch; - - MainSearchPtr(0, 0, Data, 255, CheckCandidate); + MainSearchPtr(0, 0, Data, 255, CheckCandidate); - xvid_me_SubpelRefine(Data, CheckCandidate); + xvid_me_SubpelRefine(Data, CheckCandidate); + } *best_sad = *Data->iMinSAD; @@ -671,6 +711,8 @@ { int i, j; int b_range[4], f_range[4]; + int threshA = (MotionFlags & XVID_ME_FAST_MODEINTERPOLATE) ? 0 : 500; + int threshB = (MotionFlags & XVID_ME_FAST_MODEINTERPOLATE) ? 0 : 300; Data->qpel_precision = 0; *Data->iMinSAD = 4096*256; @@ -739,7 +781,7 @@ /* qpel refinement */ if (Data->qpel) { - if (*Data->iMinSAD > *best_sad + 500) return; + if (*Data->iMinSAD > *best_sad + threshA) return; Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); @@ -748,12 +790,12 @@ Data->currentQMV[1].x = 2 * Data->currentMV[1].x; Data->currentQMV[1].y = 2 * Data->currentMV[1].y; SubpelRefine_dir(Data, CheckCandidateInt, 1); - if (*Data->iMinSAD > *best_sad + 300) return; + if (*Data->iMinSAD > *best_sad + threshB) return; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0); SubpelRefine_dir(Data, CheckCandidateInt, 2); } - *Data->iMinSAD += (2+3) * Data->lambda16; /* two bits are needed to code interpolate mode. */ + *Data->iMinSAD += 2 * Data->lambda16; /* two bits are needed to code interpolate mode. */ if (*Data->iMinSAD < *best_sad) { *best_sad = *Data->iMinSAD; @@ -799,6 +841,7 @@ uint32_t skip_sad; const MACROBLOCK * const b_mbs = b_reference->mbs; + MACROBLOCK *const pMBs = frame->mbs; VECTOR f_predMV, b_predMV; @@ -811,7 +854,6 @@ memset(&Data, 0, sizeof(SearchData)); Data.iEdgedWidth = pParam->edged_width; - Data.lambda16 = xvid_me_lambda_vec16[MAX(frame->quant-2, 2)]; Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; Data.rounding = 0; Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; @@ -827,14 +869,18 @@ for (i = 0; i < pParam->mb_width; i++) { MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; + int interpol_search; /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ if (b_reference->coding_type != S_VOP) if (b_mb->mode == MODE_NOT_CODED) { pMB->mode = MODE_NOT_CODED; pMB->mvs[0] = pMB->b_mvs[0] = zeroMV; + pMB->sad16 = 0; continue; } + + 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; @@ -853,7 +899,39 @@ &best_sad, &Data); - if (pMB->mode == MODE_DIRECT_NONE_MV) continue; + if (pMB->mode == MODE_DIRECT_NONE_MV) { + pMB->sad16 = best_sad; + continue; + } + + if (frame->motion_flags & XVID_ME_BFRAME_EARLYSTOP) { + int bf_search = 0; + int bf_thresh = 0; + + if(i > 0 && j > 0 && i < pParam->mb_width) { + bf_thresh = ((&pMBs[(i-1) + j * pParam->mb_width])->sad16 + + (&pMBs[i + (j-1) * pParam->mb_width])->sad16 + + (&pMBs[(i+1) + (j-1) * pParam->mb_width])->sad16) / 3; + + 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++; + } + + if ((best_sad < bf_thresh) && (bf_search == 3)) + continue; + } /* forward search */ SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, @@ -872,33 +950,45 @@ MODE_BACKWARD, &Data); /* interpolate search comes last, because it uses data from forward and backward as prediction */ - SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, - b_ref, b_refH->y, b_refV->y, b_refHV->y, - i, j, - frame->fcode, frame->bcode, - frame->motion_flags, - pParam, - &f_predMV, &b_predMV, - pMB, &best_sad, - &Data); + if (frame->motion_flags & XVID_ME_FAST_MODEINTERPOLATE) + interpol_search = (best_sad > Data.iQuant * 3 * MAX_SAD00_FOR_SKIP * (Data.chroma ? 3:2)); + else + interpol_search = 1; + + if (interpol_search) { + SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, + b_ref, b_refH->y, b_refV->y, b_refHV->y, + i, j, + frame->fcode, frame->bcode, + frame->motion_flags, + pParam, + &f_predMV, &b_predMV, + pMB, &best_sad, + &Data); + } /* final skip decision */ - if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2) - && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) + if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * (Data.chroma ? 3:2) ) + && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) + SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); switch (pMB->mode) { case MODE_FORWARD: f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; + pMB->sad16 = best_sad; break; case MODE_BACKWARD: 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: + pMB->sad16 = best_sad; break; } }