--- branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/10/17 13:50:23 601 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/11/02 15:52:31 619 @@ -131,6 +131,45 @@ } static void +CheckCandidate16Q(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + //this function expects X and Y in halfpel positions, but assumes that qpel positions are our goal + int t; + const uint8_t * Reference; + + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; + + switch ( ((x&1)<<1) + (y&1) ) { + case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; + case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; + case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; + default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; + } + + data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + + t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); + data->temp[0] += data->lambda16 * t; + data->temp[1] += data->lambda8 * t; + + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } + + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } + +} + +static void CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; @@ -174,6 +213,7 @@ if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases switch( ((x&1)<<1) + (y&1) ) { case 0: // pure halfpel position - shouldn't happen during a refinement step @@ -181,9 +221,7 @@ break; case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); @@ -191,9 +229,7 @@ break; case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); @@ -202,7 +238,6 @@ default: // x and y in qpel resolution - the "corners" (top left/right and // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); @@ -236,75 +271,6 @@ } static void -CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) - -// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution -// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) -// around currentMV! -{ - int32_t sad; - uint8_t * Reference = (uint8_t *) data->RefQ; - const uint8_t *ref1, *ref2, *ref3, *ref4; - VECTOR halfpelMV = *(data->currentMV); - - int32_t iEdgedWidth = data->iEdgedWidth; - uint32_t rounding = data->rounding; - - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; - - switch( ((x&1)<<1) + (y&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); - - interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); - interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); - interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); - interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; - } - - sad = data->lambda16 * - d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); - sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentQMV[0].x = x; data->currentQMV[0].y = y; -// *dir = Direction; - } -} - -static void CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; @@ -491,6 +457,32 @@ } static void +CheckCandidate8Q(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + const uint8_t * Reference; + + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; + + switch ( ((x&1)<<1) + (y&1) ) + { + case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; + case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; + case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; + default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; + } + + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + sad += data->lambda8 * d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} + +static void CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) // CheckCandidate16no4v variant which expects x and y in quarter pixel resolution // Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) @@ -508,6 +500,7 @@ if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); switch( ((x&1)<<1) + (y&1) ) { case 0: // pure halfpel position - shouldn't happen during a refinement step @@ -515,14 +508,12 @@ break; case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); break; case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); @@ -530,7 +521,6 @@ default: // x and y in qpel resolution - the "corners" (top left/right and // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); @@ -764,14 +754,13 @@ pMB->mode = MODE_NOT_CODED; pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; - pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; } -bool +bool MotionEstimation(MBParam * const pParam, FRAMEINFO * const current, FRAMEINFO * const reference, @@ -833,13 +822,16 @@ pMB->quant = quant; } -//initial skip decision +/* initial skip decision */ +/* no early skip for GMC (global vector = skip vector is unknown!) */ - if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) - if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { - SkipMacroblockP(pMB, sad00); - continue; - } + if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ + if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) + if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, y, current->motion_flags, pMB->quant, @@ -847,13 +839,15 @@ current->global_flags & XVID_INTER4V, pMB); /* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP + if (current->coding_type == P_VOP) { + if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) ) - if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { - SkipMacroblockP(pMB, sad00); - continue; - } - + if (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } + /* finally, intra decision */ InterBias = MV16_INTER_BIAS; @@ -882,7 +876,13 @@ } } free(qimage); - return 0; + + if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ + current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); + else + current->GMC_MV = zeroMV; + + return 0; } @@ -960,8 +960,6 @@ int i, iDirection = 255, mask, threshA; VECTOR pmv[7]; - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); - get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); @@ -976,6 +974,7 @@ Data->RefQ = pRefQ; Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -985,13 +984,14 @@ if (pMB->dquant != NO_CHANGE) inter4v = 0; - if (inter4v) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; - for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; - i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); + if (pParam->m_quarterpel) { + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); + i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); + } else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); + Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; Data->iMinSAD[2] = pMB->sad8[1]; @@ -1007,9 +1007,10 @@ PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, prevMBs + x + y * pParam->mb_width); - if (inter4v) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; - + if (pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; + else + if (inter4v) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; /* main loop. checking all predictions */ @@ -1044,7 +1045,7 @@ if (!(MVequal(startMV, backupMV))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); + (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; @@ -1057,7 +1058,7 @@ if (!(MVequal(startMV, backupMV))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); + (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; @@ -1075,22 +1076,8 @@ if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - if(inter4v) - CheckCandidate = CheckCandidate16_qpel; - else - CheckCandidate = CheckCandidate16no4v_qpel; - - Data->iMinSAD[0] -= lambda_vec16[iQuant] * - d_mv_bits(Data->predMV.x - Data->currentMV[0].x, Data->predMV.y - Data->currentMV[0].y, Data->iFcode); - Data->iMinSAD[1] -= lambda_vec8[iQuant] * - d_mv_bits(Data->predMV.x - Data->currentMV[1].x, Data->predMV.y - Data->currentMV[1].y, Data->iFcode); - - Data->iMinSAD[0] += lambda_vec16[iQuant] * - d_mv_bits(Data->predQMV.x - Data->currentQMV[0].x, Data->predMV.y - Data->currentQMV[0].y, Data->iFcode); - Data->iMinSAD[1] += lambda_vec8[iQuant] * - d_mv_bits(Data->predQMV.x - Data->currentQMV[1].x, Data->predMV.y - Data->currentQMV[1].y, Data->iFcode); - - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + CheckCandidate = CheckCandidate16_qpel; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, pParam->width, pParam->height, Data->iFcode, 0); QuarterpelRefine(Data); @@ -1099,8 +1086,9 @@ if (inter4v) { SearchData Data8; Data8.iFcode = Data->iFcode; - Data8.lambda8 = lambda_vec8[iQuant]; + Data8.lambda8 = Data->lambda8; Data8.iEdgedWidth = Data->iEdgedWidth; + Data8.RefQ = Data->RefQ; Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); @@ -1147,30 +1135,25 @@ const int block, SearchData * const Data) { - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); Data->iMinSAD = OldData->iMinSAD + 1 + block; Data->currentMV = OldData->currentMV + 1 + block; Data->currentQMV = OldData->currentQMV + 1 + block; if(pParam->m_quarterpel) { - //it is qpel. substract d_mv_bits[qpel] from 0, add d_mv_bits[hpel] everywhere - if (block == 0) - *(Data->iMinSAD) -= Data->lambda8 * + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + if (block != 0) *(Data->iMinSAD) += Data->lambda8 * d_mv_bits( Data->currentQMV->x - Data->predQMV.x, Data->currentQMV->y - Data->predQMV.y, Data->iFcode); - - *(Data->iMinSAD) += Data->lambda8 * - d_mv_bits( Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, - Data->iFcode); - } else //it is not qpel. add d_mv_bits[hpel] everywhere but not in 0 (it's already there) + CheckCandidate = CheckCandidate8Q; + } else { + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); if (block != 0) *(Data->iMinSAD) += Data->lambda8 * d_mv_bits( Data->currentMV->x - Data->predMV.x, Data->currentMV->y - Data->predMV.y, Data->iFcode); - + CheckCandidate = CheckCandidate8; + } if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { @@ -1178,15 +1161,12 @@ Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); - Data->RefQ = OldData->RefQ; Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); - CheckCandidate = CheckCandidate8; - if (MotionFlags & PMV_EXTSEARCH8) { int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD @@ -1217,15 +1197,10 @@ if(pParam->m_quarterpel) { if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && (MotionFlags & PMV_QUARTERPELREFINE8)) { - CheckCandidate = CheckCandidate8_qpel; - Data->iMinSAD[0] -= Data->lambda8 * - d_mv_bits(Data->predMV.x - Data->currentMV[0].x, Data->predMV.y - Data->currentMV[0].y, Data->iFcode); - - Data->iMinSAD[0] += Data->lambda8 * - d_mv_bits(Data->predQMV.x - Data->currentQMV[0].x, Data->predQMV.y - Data->currentQMV[0].y, Data->iFcode); - - QuarterpelRefine(Data); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, + pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); + QuarterpelRefine(Data); } } } @@ -1601,7 +1576,8 @@ } while (!(iDirection)); // two bits are needed to code interpolate mode. we treat the bits just like they were vector's - *fData->iMinSAD += 2 * fData->lambda16 * 2; + *fData->iMinSAD += 2 * fData->lambda16 * 2; + if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; pMB->mvs[0] = fData->currentMV[0]; @@ -1664,8 +1640,8 @@ MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; -/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ - if (b_mb->mode == MODE_NOT_CODED) { +/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ + if ( (b_mb->mode == MODE_NOT_CODED) ) { pMB->mode = MODE_NOT_CODED; continue; } @@ -1763,44 +1739,36 @@ { int32_t temp_sad; MainSearchFunc *MainSearchPtr; - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); Data->iMinSAD = OldData->iMinSAD + 1 + block; Data->currentMV = OldData->currentMV + 1 + block; Data->currentQMV = OldData->currentQMV + 1 + block; + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); if(pParam->m_quarterpel) { - //it is qpel. substract d_mv_bits[qpel] from 0, add d_mv_bits[hpel] everywhere - if (block == 0) - *(Data->iMinSAD) -= Data->lambda8 * + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + if (block != 0) *(Data->iMinSAD) += Data->lambda8 * d_mv_bits( Data->currentQMV->x - Data->predQMV.x, Data->currentQMV->y - Data->predQMV.y, Data->iFcode); - - *(Data->iMinSAD) += Data->lambda8 * - d_mv_bits( Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, - Data->iFcode); - } else //it is not qpel. add d_mv_bits[hpel] everywhere but not in 0 (it's already there) + CheckCandidate = CheckCandidate8Q; + } else { if (block != 0) *(Data->iMinSAD) += Data->lambda8 * d_mv_bits( Data->currentMV->x - Data->predMV.x, Data->currentMV->y - Data->predMV.y, Data->iFcode); - + CheckCandidate = CheckCandidate8; + } Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); - Data->RefQ = OldData->RefQ; Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); - CheckCandidate = CheckCandidate8; - temp_sad = *(Data->iMinSAD); // store current MinSAD if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; @@ -1828,15 +1796,10 @@ if(pParam->m_quarterpel) { if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && (MotionFlags & PMV_QUARTERPELREFINE8)) { - + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, + pParam->width, pParam->height, Data->iFcode, 0); CheckCandidate = CheckCandidate8_qpel; - Data->iMinSAD[0] -= Data->lambda8 * - d_mv_bits(Data->predMV.x - Data->currentMV[0].x, Data->predMV.y - Data->currentMV[0].y, Data->iFcode); - - Data->iMinSAD[0] += Data->lambda8 * - d_mv_bits(Data->predQMV.x - Data->currentQMV[0].x, Data->predQMV.y - Data->currentQMV[0].y, Data->iFcode); - - QuarterpelRefine(Data); + QuarterpelRefine(Data); } pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; @@ -1851,7 +1814,6 @@ pMB->sad8[block] = 4 * (*Data->iMinSAD); } - static void SearchPhinted ( const uint8_t * const pRef, const uint8_t * const pRefH, @@ -1884,6 +1846,7 @@ Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -1896,10 +1859,10 @@ if (pMB->dquant != NO_CHANGE) inter4v = 0; - if (inter4v) - CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; - + if(pParam->m_quarterpel) CheckCandidate = CheckCandidate16Q; + else + if (inter4v) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; pMB->mvs[0].x = EVEN(pMB->mvs[0].x); pMB->mvs[0].y = EVEN(pMB->mvs[0].y); @@ -1934,21 +1897,18 @@ } if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - - if(inter4v) - CheckCandidate = CheckCandidate16_qpel; - else - CheckCandidate = CheckCandidate16no4v_qpel; - + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode, 0); + CheckCandidate = CheckCandidate16_qpel; QuarterpelRefine(Data); } - if (inter4v) { SearchData Data8; Data8.iFcode = Data->iFcode; - Data8.lambda8 = lambda_vec8[pMB->quant]; + Data8.lambda8 = Data->lambda8; Data8.iEdgedWidth = Data->iEdgedWidth; + Data8.RefQ = Data->RefQ; Search8hinted(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); Search8hinted(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); Search8hinted(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); @@ -1959,16 +1919,24 @@ (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { // INTER MODE - pMB->mode = MODE_INTER; pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->qmvs[0] = pMB->qmvs[1] + = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; - pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; - pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; + if(pParam->m_quarterpel) { + pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; + pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; + } + else { + pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; + } } else { // INTER4V MODE; all other things are already set in Search8hinted pMB->mode = MODE_INTER4V; @@ -1994,10 +1962,11 @@ uint8_t * qimage; int32_t temp[5], quant = current->quant; int32_t iMinSAD[5]; - VECTOR currentMV[5]; + VECTOR currentMV[5], currentQMV[5]; SearchData Data; Data.iEdgedWidth = pParam->edged_width; Data.currentMV = currentMV; + Data.currentQMV = currentQMV; Data.iMinSAD = iMinSAD; Data.temp = temp; Data.iFcode = current->fcode; @@ -2024,13 +1993,14 @@ if (!(current->global_flags & XVID_LUMIMASKING)) { pMB->dquant = NO_CHANGE; pMB->quant = current->quant; } - else + else { if (pMB->dquant != NO_CHANGE) { quant += DQtab[pMB->dquant]; if (quant > 31) quant = 31; else if (quant < 1) quant = 1; - pMB->quant = quant; } + pMB->quant = quant; + } SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, y, current->motion_flags, pMB->quant, @@ -2053,7 +2023,7 @@ SearchData * const Data) { - int i, mask; + int i = 255, mask; VECTOR pmv[3]; *(Data->iMinSAD) = MV_MAX_ERROR; @@ -2078,10 +2048,9 @@ if (!(mask = make_mask(pmv, 2))) CheckCandidate16no4vI(0, 0, mask, &i, Data); - DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask); + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); - pMB->mvs[0] = pMB->mvs[1] - = pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() + pMB->mvs[0] = *Data->currentMV; return *(Data->iMinSAD); } @@ -2157,7 +2126,62 @@ min = -min; max += 1; if (min > max) max = min; + if (pParam->m_quarterpel) max *= 2; for (i = 1; (max > 32 << (i - 1)); i++); return i; } + + +/* Global Motion Estimation, (C) 2002 by sysKin */ +/* calculate global translation from local (halfpel) motion field */ + +static VECTOR +GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) +{ + + int count, bestcount = 0; + int x, y; + VECTOR gmc; + int step, min_x, max_x, min_y, max_y; + uint32_t mx, my; + + min_x = min_y = -32<= 2; step /= 2) { + bestcount = 0; + for (y = min_y; y <= max_y; y += step) + for (x = min_x ; x <= max_x; x += step) { + count = 0; + //for all macroblocks + for (my = 1; my < pParam->mb_height-1; my++) + for (mx = 1; mx < pParam->mb_width-1; mx++) { + const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; + VECTOR mv; + + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) + continue; + + mv = pMB->mvs[0]; + if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ + count++; + } + if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } + } + min_x = gmc.x - step; + max_x = gmc.x + step; + min_y = gmc.y - step; + max_y = gmc.y + step; + + } + if (pParam->m_quarterpel) { + gmc.x *= 2; + gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ + } + + if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) + gmc.x = gmc.y = 0; //no camara pan, no GMC + + return gmc; +}