--- 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 2002/11/12 14:46:03 644 @@ -89,9 +89,47 @@ return xb + yb; } +static int32_t +ChromaSAD(int dx, int dy, const SearchData * const data) +{ + int sad; + dx = (dx >> 1) + roundtab_79[dx & 0x3]; + dy = (dy >> 1) + roundtab_79[dy & 0x3]; + + switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) + case 0: + sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + break; + case 1: + dx = dx / 2; dy = (dy - 1) / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + break; + case 2: + dx = (dx - 1) / 2; dy = dy / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + break; + default: + dx = (dx - 1) / 2; dy = (dy - 1) / 2; + interpolate8x8_halfpel_hv(data->RefQ, + data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, + data->rounding); + sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); + interpolate8x8_halfpel_hv(data->RefQ, + data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, + data->rounding); + sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); + break; + } + return sad; +} + /* CHECK_CANDIATE FUNCTIONS START */ + static void CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { @@ -110,9 +148,13 @@ data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - data->temp[0] += data->lambda16 * t; - data->temp[1] += data->lambda8 * t; + if (data->qpel) t = d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); + else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + + data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; + + if (data->chroma) data->temp[0] += ChromaSAD(x, y, data); if (data->temp[0] < data->iMinSAD[0]) { data->iMinSAD[0] = data->temp[0]; @@ -147,8 +189,8 @@ default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; } - sad = data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); + sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); + sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -174,6 +216,7 @@ 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 +224,7 @@ 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 +232,7 @@ 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 +241,6 @@ 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); @@ -217,8 +255,11 @@ data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); - data->temp[0] += data->lambda16 * t; - data->temp[1] += data->lambda8 * t; + data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; + + if (data->chroma) + data->temp[0] += ChromaSAD(x/2, y/2, data); if (data->temp[0] < data->iMinSAD[0]) { data->iMinSAD[0] = data->temp[0]; @@ -236,75 +277,6 @@ } 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; @@ -347,11 +319,11 @@ default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; } - sad = data->lambda16 * - ( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + - d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * + ( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + + d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -362,7 +334,7 @@ static void CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; + int32_t sad = 0; int k; const uint8_t *ReferenceF; const uint8_t *ReferenceB; @@ -370,8 +342,6 @@ if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - sad = data->lambda16 * d_mv_bits(x, y, 1); - for (k = 0; k < 4; k++) { mvs.x = data->directmvF[k].x + x; b_mvs.x = ((x == 0) ? @@ -409,6 +379,8 @@ if (sad > *(data->iMinSAD)) return; } + sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; + if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV->x = x; data->currentMV->y = y; @@ -424,8 +396,6 @@ VECTOR mvs, b_mvs; if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - - sad = data->lambda16 * d_mv_bits(x, y, 1); mvs.x = data->directmvF[0].x + x; b_mvs.x = ((x == 0) ? @@ -456,7 +426,8 @@ default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; } - sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -467,7 +438,7 @@ static void CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; + int32_t sad; int t; const uint8_t * Reference; if (( x > data->max_dx) || ( x < data->min_dx) @@ -482,7 +453,10 @@ } sad = sad8(data->Cur, Reference, data->iEdgedWidth); - sad += data->lambda8 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); + else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + + sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -508,6 +482,7 @@ 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 +490,12 @@ 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 +503,6 @@ 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); @@ -540,7 +512,7 @@ } sad = sad8(data->Cur, Reference, data->iEdgedWidth); - sad += data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); + sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -788,7 +760,7 @@ uint32_t x, y; uint32_t iIntra = 0; - int32_t InterBias, quant = current->quant; + int32_t InterBias, quant = current->quant, sad00; uint8_t *qimage; // some pre-initialized thingies for SearchP @@ -804,23 +776,35 @@ Data.temp = temp; Data.iFcode = current->fcode; Data.rounding = pParam->m_rounding_type; + Data.qpel = pParam->m_quarterpel; + Data.chroma = current->global_flags & XVID_ME_COLOUR; if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) return 1; // allocate some mem for qpel interpolated blocks // somehow this is dirty since I think we shouldn't use malloc outside // encoder_create() - so please fix me! - + Data.RefQ = qimage; if (sadInit) (*sadInit) (); for (y = 0; y < pParam->mb_height; y++) { for (x = 0; x < pParam->mb_width; x++) { - MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; - int32_t sad00 = pMB->sad16 + + pMB->sad16 = sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, pRef->y + (x + y * pParam->edged_width) * 16, pParam->edged_width, pMB->sad8 ); + if (Data.chroma) { + pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, + pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); + + pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, + pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); + } + + sad00 = pMB->sad16; //if no gmc; else sad00 = (..) + if (!(current->global_flags & XVID_LUMIMASKING)) { pMB->dquant = NO_CHANGE; pMB->quant = current->quant; @@ -834,34 +818,40 @@ } //initial skip decision - - 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, +/* no early skip for GMC (global vector = skip vector is unknown!) */ + 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 (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } + + SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, y, current->motion_flags, pMB->quant, &Data, pParam, pMBs, reference->mbs, 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 (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } /* finally, intra decision */ InterBias = MV16_INTER_BIAS; - if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work + if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work if (y != 0) - if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; + if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; if (x != 0) - if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; + if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; + + if (Data.chroma) InterBias += 50; // to compensate bigger SAD if (InterBias < pMB->sad16) { const int32_t deviation = @@ -882,6 +872,12 @@ } } free(qimage); + + 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; } @@ -939,11 +935,10 @@ } static void -SearchP(const uint8_t * const pRef, +SearchP(const IMAGE * const pRef, const uint8_t * const pRefH, const uint8_t * const pRefV, const uint8_t * const pRefHV, - const uint8_t * const pRefQ, const IMAGE * const pCur, const int x, const int y, @@ -960,8 +955,6 @@ 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); @@ -969,13 +962,18 @@ Data->predMV = pmv[0]; Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; - Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; + Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; + + Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; - Data->RefQ = pRefQ; + Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -985,15 +983,16 @@ 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); - Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; - Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; + 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 + (Data->lambda16 * i * pMB->sad16)/1000; + Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; Data->iMinSAD[2] = pMB->sad8[1]; Data->iMinSAD[3] = pMB->sad8[2]; Data->iMinSAD[4] = pMB->sad8[3]; @@ -1007,10 +1006,9 @@ PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, prevMBs + x + y * pParam->mb_width); - if (inter4v) CheckCandidate = CheckCandidate16; + if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; else CheckCandidate = CheckCandidate16no4v; - /* main loop. checking all predictions */ for (i = 1; i < 7; i++) { @@ -1044,7 +1042,7 @@ 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 +1055,7 @@ 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; @@ -1075,36 +1073,41 @@ 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); } + if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; 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; + Data8.qpel = Data->qpel; 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); Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + + if (Data->chroma) { + int sum, dx, dy; + + if(pParam->m_quarterpel) { + sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; + } else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; + dy = (sum >> 3) + roundtab_76[sum & 0xf]; + + if(pParam->m_quarterpel) { + sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; + } else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; + dx = (sum >> 3) + roundtab_76[sum & 0xf]; + + Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); + } } if (!(inter4v) || @@ -1124,8 +1127,7 @@ 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 { + } else { pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; } @@ -1147,30 +1149,23 @@ 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) - if (block != 0) *(Data->iMinSAD) += Data->lambda8 * + Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + } 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); - + Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + } if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { @@ -1178,13 +1173,11 @@ 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) { @@ -1217,15 +1210,10 @@ 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); } } } @@ -1355,8 +1343,8 @@ // three bits are needed to code backward mode. four for forward // we treat the bits just like they were vector's - if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16 * 2; - else *Data->iMinSAD += 3 * Data->lambda16 * 2; + 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; @@ -1441,7 +1429,7 @@ (*CheckCandidate)(0, 0, 255, &k, Data); // skip decision - if (*Data->iMinSAD - 2 * Data->lambda16 < (uint32_t)pMB->quant * SKIP_THRESH_B) { + if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { //possible skip - checking chroma. everything copied from MC //this is not full chroma compensation, only it's fullpel approximation. should work though int sum, dx, dy, b_dx, b_dy; @@ -1486,7 +1474,7 @@ HalfpelRefine(Data); - *Data->iMinSAD += 1 * Data->lambda16 * 2; // one bit is needed to code direct mode + *Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode *best_sad = *Data->iMinSAD; if (b_mb->mode == MODE_INTER4V) @@ -1600,8 +1588,8 @@ } 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; // two bits are needed to code interpolate mode. + if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; pMB->mvs[0] = fData->currentMV[0]; @@ -1615,7 +1603,6 @@ } } - void MotionEstimationBVOP(MBParam * const pParam, FRAMEINFO * const frame, @@ -1628,7 +1615,7 @@ const IMAGE * const f_refV, const IMAGE * const f_refHV, // backward (future) reference - const MACROBLOCK * const b_mbs, + const FRAMEINFO * const b_reference, const IMAGE * const b_ref, const IMAGE * const b_refH, const IMAGE * const b_refV, @@ -1638,6 +1625,7 @@ int32_t best_sad, skip_sad; int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; static const VECTOR zeroMV={0,0}; + const MACROBLOCK * const b_mbs = b_reference->mbs; VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ @@ -1664,17 +1652,18 @@ 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) { - pMB->mode = MODE_NOT_CODED; - continue; - } +/* 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; + continue; + } Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; pMB->quant = frame->quant; + /* direct search comes first, because it (1) checks for SKIP-mode and (2) sets very good predictions for forward and backward search */ - skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, b_ref, b_refH->y, b_refV->y, b_refHV->y, &frame->image, @@ -1688,9 +1677,6 @@ if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } -// best_sad = 256*4096; //uncomment to disable Directsearch. -// To disable any other mode, just comment the function call - // forward search SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, &frame->image, i, j, @@ -1743,117 +1729,12 @@ } } } - -// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", -// f_count,b_count,i_count,d_count,n_count); - } /* Hinted ME starts here */ static void -Search8hinted(const SearchData * const OldData, - const int x, const int y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMB, - const MACROBLOCK * const pMBs, - const int block, - SearchData * const Data) -{ - 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; - - 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 * - 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) - if (block != 0) *(Data->iMinSAD) += Data->lambda8 * - d_mv_bits( Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, - Data->iFcode); - - - 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; - else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; - else MainSearchPtr = DiamondSearch; - - (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); - - if(*(Data->iMinSAD) < temp_sad) { - Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector - Data->currentQMV->y = 2 * Data->currentMV->y; - } - - if (MotionFlags & PMV_HALFPELREFINE8) { - temp_sad = *(Data->iMinSAD); // store current MinSAD - - HalfpelRefine(Data); // perform halfpel refine of current best vector - - if(*(Data->iMinSAD) < temp_sad) { // we have found a better match - Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector - Data->currentQMV->y = 2 * Data->currentMV->y; - } - } - - 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); - } - pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; - pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; - } else { - pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; - pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; - } - - pMB->mvs[block] = *(Data->currentMV); - pMB->qmvs[block] = *(Data->currentQMV); - - pMB->sad8[block] = 4 * (*Data->iMinSAD); -} - - -static void -SearchPhinted ( const uint8_t * const pRef, +SearchPhinted ( const IMAGE * const pRef, const uint8_t * const pRefH, const uint8_t * const pRefV, const uint8_t * const pRefHV, @@ -1869,21 +1750,24 @@ SearchData * const Data) { - const int32_t iEdgedWidth = pParam->edged_width; - int i, t; MainSearchFunc * MainSearchPtr; Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); 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); - Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; - Data->Ref = pRef + (x + iEdgedWidth*y)*16; - Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; - Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; - Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; - Data->lambda16 = lambda_vec16[iQuant]; + Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; + Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; + + Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; + Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; + Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; + Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; + Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -1896,11 +1780,9 @@ if (pMB->dquant != NO_CHANGE) inter4v = 0; - if (inter4v) - CheckCandidate = CheckCandidate16; + if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; else CheckCandidate = CheckCandidate16no4v; - pMB->mvs[0].x = EVEN(pMB->mvs[0].x); pMB->mvs[0].y = EVEN(pMB->mvs[0].y); if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed @@ -1934,43 +1816,65 @@ } 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; - 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); - Search8hinted(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + Data8.RefQ = Data->RefQ; + Data8.qpel = Data->qpel; + 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); + Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + + if (Data->chroma) { + int sum, dx, dy; + + if(pParam->m_quarterpel) + sum = (pMB->qmvs[0].y + pMB->qmvs[1].y + pMB->qmvs[2].y + pMB->qmvs[3].y)/2; + else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; + dy = (sum ? SIGN(sum) * + (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); + + if(pParam->m_quarterpel) + sum = (pMB->qmvs[0].x + pMB->qmvs[1].x + pMB->qmvs[2].x + pMB->qmvs[3].x)/2; + else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; + dx = (sum ? SIGN(sum) * + (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); + Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); + } } if (!(inter4v) || (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 +// INTER4V MODE; all other things are already set in Search8 pMB->mode = MODE_INTER4V; pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; @@ -1994,14 +1898,17 @@ 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; Data.rounding = pParam->m_rounding_type; + Data.qpel = pParam->m_quarterpel; + Data.chroma = current->global_flags & XVID_ME_COLOUR; if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) return; // allocate some mem for qpel interpolated blocks @@ -2020,19 +1927,19 @@ //intra mode is copied from the first pass. At least for the time being if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; - 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, + SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, y, current->motion_flags, pMB->quant, pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, &Data); @@ -2053,51 +1960,70 @@ SearchData * const Data) { - int i, mask; + int i = 255, mask; VECTOR pmv[3]; - *(Data->iMinSAD) = MV_MAX_ERROR; - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + + //median is only used as prediction. it doesn't have to be real + if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; + else + if (x == 1) //left macroblock does not have any vector now + Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median + else if (y == 1) // top macroblock don't have it's vector + Data->predMV = (pMB - 1)->mvs[0]; // left instead of median + else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median + 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); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->Ref = pRef + (x + y * pParam->edged_width) * 16; - CheckCandidate = CheckCandidate16no4vI; - pmv[1].x = EVEN(pMB->mvs[0].x); pmv[1].y = EVEN(pMB->mvs[0].y); - pmv[0].x = EVEN(Data->predMV.x); - pmv[0].y = EVEN(Data->predMV.y); - pmv[2].x = pmv[2].y = 0; + pmv[2].x = EVEN(Data->predMV.x); + pmv[2].y = EVEN(Data->predMV.y); + pmv[0].x = pmv[0].y = 0; + + (*CheckCandidate)(0, 0, 255, &i, Data); + +//early skip for 0,0 + if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->mode = MODE_NOT_CODED; + return 0; + } - CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); if (!(mask = make_mask(pmv, 1))) - CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); + (*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); if (!(mask = make_mask(pmv, 2))) - CheckCandidate16no4vI(0, 0, mask, &i, Data); - - DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask); + (*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); - pMB->mvs[0] = pMB->mvs[1] - = pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() + if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->mode = MODE_INTER; return *(Data->iMinSAD); } #define INTRA_THRESH 1350 #define INTER_THRESH 900 + int MEanalysis( const IMAGE * const pRef, - const IMAGE * const pCurrent, + FRAMEINFO * const Current, MBParam * const pParam, - MACROBLOCK * const pMBs, - const uint32_t iFcode) + int maxIntra, //maximum number if non-I frames + int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame + int bCount) // number if B frames in a row { uint32_t x, y, intra = 0; int sSAD = 0; + MACROBLOCK * const pMBs = Current->mbs; + const IMAGE * const pCurrent = &Current->image; + int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; VECTOR currentMV; int32_t iMinSAD; @@ -2105,7 +2031,18 @@ Data.iEdgedWidth = pParam->edged_width; Data.currentMV = ¤tMV; Data.iMinSAD = &iMinSAD; - Data.iFcode = iFcode; + Data.iFcode = Current->fcode; + CheckCandidate = CheckCandidate16no4vI; + + if (intraCount < 12) // we're right after an I frame + IntraThresh += 4 * (intraCount - 12) * (intraCount - 12); + else + if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec + IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; + + + InterThresh += 300 * (1 - bCount); + if (InterThresh < 200) InterThresh = 200; if (sadInit) (*sadInit) (); @@ -2117,17 +2054,19 @@ sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, pMB, &Data); - if (sad > INTRA_THRESH) { + if (sad > IntraThresh) { dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, pParam->edged_width); - if (dev + INTRA_THRESH < sad) intra++; - if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame + if (dev + IntraThresh < sad) { + pMB->mode = MODE_INTRA; + if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame } + } sSAD += sad; } } sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); - if (sSAD > INTER_THRESH ) return 1; //P frame + if (sSAD > InterThresh ) return 1; //P frame emms(); return 0; // B frame @@ -2157,7 +2096,100 @@ 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; } + +static void +CheckGMC(int x, int y, const int dir, int * iDirection, + const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, + const MBParam * const pParam) +{ + uint32_t mx, my, a, count = 0; + + for (my = 1; my < pParam->mb_height-1; my++) + for (mx = 1; mx < pParam->mb_width-1; mx++) { + VECTOR mv; + const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; + mv = pMB->mvs[0]; + a = ABS(mv.x - x) + ABS(mv.y - y); + if (a < 6) count += 6 - a; + } + + if (count > *bestcount) { + *bestcount = count; + *iDirection = dir; + GMC->x = x; GMC->y = y; + } +} + + +static VECTOR +GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) +{ + + uint32_t count, bestcount = 0; + int x, y; + VECTOR gmc = {0,0}; + int step, min_x, max_x, min_y, max_y; + uint32_t mx, my; + int iDirection, bDirection; + + 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 (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) + gmc.x = gmc.y = 0; //no camara pan, no GMC + +// step2: let's refine camera panning using gradiend-descent approach. +// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) + bestcount = 0; + CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); + do { + x = gmc.x; y = gmc.y; + bDirection = iDirection; iDirection = 0; + if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + + } while (iDirection); + + if (pParam->m_quarterpel) { + gmc.x *= 2; + gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ + } + + return gmc; +}