--- branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/10/17 13:27:22 600 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/10/21 13:07:33 606 @@ -55,10 +55,10 @@ #define GET_REFERENCE(X, Y, REF) { \ switch ( (((X)&1)<<1) + ((Y)&1) ) \ { \ - case 0 : REF = data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - case 1 : REF = data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ - case 2 : REF = data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - default : REF = data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ + case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ + case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ + case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ + default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ } \ } @@ -110,13 +110,48 @@ data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - if(data->quarterpel) - 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); + 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->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 +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); - data->temp[0] += lambda_vec16[data->iQuant] * t; - data->temp[1] += lambda_vec8[data->iQuant] * t; + 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]; @@ -151,13 +186,7 @@ default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; } - if(data->quarterpel) - sad = lambda_vec16[data->iQuant] * - d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); - else - sad = lambda_vec16[data->iQuant] * - d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - + 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); if (sad < *(data->iMinSAD)) { @@ -174,7 +203,7 @@ // around currentMV! { int t; - uint8_t * Reference = (uint8_t *) data->RefQ; + uint8_t * Reference = (uint8_t *)data->RefQ; const uint8_t *ref1, *ref2, *ref3, *ref4; VECTOR halfpelMV = *(data->currentMV); @@ -184,16 +213,15 @@ 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 - GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); + 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); @@ -201,9 +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); @@ -212,7 +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); @@ -227,8 +252,8 @@ 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] += lambda_vec16[data->iQuant] * t; - data->temp[1] += lambda_vec8[data->iQuant] * t; + 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]; @@ -246,75 +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, (const uint8_t *) 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 = lambda_vec16[data->iQuant] * - 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; @@ -322,10 +278,7 @@ if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - sad = lambda_vec16[data->iQuant] * - d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - - sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), data->iEdgedWidth, 256*4096); if (sad < *(data->iMinSAD)) { @@ -360,7 +313,7 @@ default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; } - sad = lambda_vec16[data->iQuant] * + 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) ); @@ -383,7 +336,7 @@ if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); + sad = data->lambda16 * d_mv_bits(x, y, 1); for (k = 0; k < 4; k++) { mvs.x = data->directmvF[k].x + x; @@ -438,7 +391,7 @@ if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); + sad = data->lambda16 * d_mv_bits(x, y, 1); mvs.x = data->directmvF[0].x + x; b_mvs.x = ((x == 0) ? @@ -495,11 +448,33 @@ } sad = sad8(data->Cur, Reference, data->iEdgedWidth); - - if(data->quarterpel) - sad += lambda_vec8[data->iQuant] * d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); - else - sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + sad += data->lambda8 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} + +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; @@ -525,21 +500,20 @@ 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 - GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); + 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); 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); @@ -547,7 +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); @@ -557,7 +530,7 @@ } sad = sad8(data->Cur, Reference, data->iEdgedWidth); - sad += lambda_vec8[data->iQuant] * 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); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -821,7 +794,6 @@ Data.temp = temp; Data.iFcode = current->fcode; Data.rounding = pParam->m_rounding_type; - Data.quarterpel = pParam->m_quarterpel; if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) return 1; // allocate some mem for qpel interpolated blocks @@ -841,25 +813,23 @@ if (!(current->global_flags & XVID_LUMIMASKING)) { pMB->dquant = NO_CHANGE; - pMB->quant = current->quant; } - else + pMB->quant = current->quant; + } 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; + } //initial skip decision - if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) - && (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { - if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { - SkipMacroblockP(pMB, sad00); - continue; - sad00 = 256 * 4096; + 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; } - } else sad00 = 256*4096; // skip not allowed - for final skip decision SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, y, current->motion_flags, pMB->quant, @@ -867,9 +837,12 @@ current->global_flags & XVID_INTER4V, pMB); /* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) - if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) - { SkipMacroblockP(pMB, sad00); continue; } + 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; + } /* finally, intra decision */ @@ -977,8 +950,6 @@ 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); @@ -992,7 +963,8 @@ Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; Data->RefQ = pRefQ; - Data->iQuant = iQuant; + Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -1002,17 +974,14 @@ 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; - if(Data->quarterpel) + 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); - + } 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]; @@ -1028,9 +997,10 @@ 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 */ @@ -1065,7 +1035,7 @@ 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; @@ -1078,7 +1048,7 @@ 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; @@ -1096,13 +1066,9 @@ if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - if(inter4v) - CheckCandidate = CheckCandidate16_qpel; - else - CheckCandidate = CheckCandidate16no4v_qpel; - + 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); // get real range + pParam->width, pParam->height, Data->iFcode, 0); QuarterpelRefine(Data); } @@ -1110,8 +1076,9 @@ if (inter4v) { SearchData Data8; Data8.iFcode = Data->iFcode; - Data8.iQuant = Data->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); @@ -1158,27 +1125,25 @@ 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; - Data->quarterpel = OldData->quarterpel; - if(Data->quarterpel) // add d_mv_bits[qpel] everywhere but not in 0 (it's already there) - { - if (block != 0) - *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * - d_mv_bits(Data->currentQMV->x - Data->predQMV.x, - Data->currentQMV->y - Data->predQMV.y, - Data->iFcode); - - } else // add d_mv_bits[hpel] everywhere but not in 0 (it's already there) - if (block != 0) - *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * - d_mv_bits(Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, - Data->iFcode); + if(pParam->m_quarterpel) { + 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); + 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)) { @@ -1186,15 +1151,12 @@ 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 @@ -1205,7 +1167,7 @@ (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); - if(*(Data->iMinSAD) < temp_sad) { //found a better match? + if(*(Data->iMinSAD) < temp_sad) { Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector Data->currentQMV->y = 2 * Data->currentMV->y; } @@ -1222,15 +1184,14 @@ } } - if((Data->quarterpel) && (!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && - (MotionFlags & PMV_QUARTERPELREFINE8)) { - + if(pParam->m_quarterpel) { + if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && + (MotionFlags & PMV_QUARTERPELREFINE8)) { CheckCandidate = CheckCandidate8_qpel; - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, - pParam->width, pParam->height, OldData->iFcode, 0); // get real range - + pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); QuarterpelRefine(Data); + } } } @@ -1246,7 +1207,7 @@ pMB->mvs[block] = *(Data->currentMV); pMB->qmvs[block] = *(Data->currentQMV); - pMB->sad8[block] = 4 * (*Data->iMinSAD); // Isibaar: why? + pMB->sad8[block] = 4 * (*Data->iMinSAD); } /* B-frames code starts here */ @@ -1359,17 +1320,16 @@ // 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 * lambda_vec16[Data->iQuant]; - else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; - + if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16 * 2; + else *Data->iMinSAD += 3 * Data->lambda16 * 2; if (*Data->iMinSAD < *best_sad) { *best_sad = *Data->iMinSAD; pMB->mode = mode_current; pMB->pmvs[0].x = Data->currentMV->x - predMV->x; pMB->pmvs[0].y = Data->currentMV->y - predMV->y; - if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; - else pMB->b_mvs[0] = *Data->currentMV; + if (mode_current == MODE_FORWARD) pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; + else pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search } } @@ -1446,7 +1406,7 @@ (*CheckCandidate)(0, 0, 255, &k, Data); // skip decision - if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { + if (*Data->iMinSAD - 2 * Data->lambda16 < (uint32_t)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; @@ -1472,7 +1432,7 @@ b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, Data->iEdgedWidth/2); - if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { + if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { pMB->mode = MODE_DIRECT_NONE_MV; return *Data->iMinSAD; } @@ -1491,7 +1451,7 @@ HalfpelRefine(Data); - *Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's + *Data->iMinSAD += 1 * Data->lambda16 * 2; // one bit is needed to code direct mode *best_sad = *Data->iMinSAD; if (b_mb->mode == MODE_INTER4V) @@ -1547,12 +1507,11 @@ int iDirection, i, j; SearchData bData; - bData.iMinSAD = fData->iMinSAD; - *bData.iMinSAD = 4096*256; + *(bData.iMinSAD = fData->iMinSAD) = 4096*256; bData.Cur = fData->Cur; fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; bData.currentMV = fData->currentMV + 1; - bData.iQuant = fData->iQuant; + bData.lambda16 = fData->lambda16; fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; @@ -1567,8 +1526,7 @@ bData.bpredMV = fData->predMV = *f_predMV; fData->bpredMV = bData.predMV = *b_predMV; - fData->currentMV[0] = pMB->mvs[0]; - fData->currentMV[1] = pMB->b_mvs[0]; + fData->currentMV[0] = fData->currentMV[3]; //forward search stored it's vector here. backward stored it in the place it's needed get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); @@ -1608,7 +1566,8 @@ } while (!(iDirection)); // two bits are needed to code interpolate mode. we treat the bits just like they were vector's - *fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; + *fData->iMinSAD += 2 * fData->lambda16 * 2; + if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; pMB->mvs[0] = fData->currentMV[0]; @@ -1659,7 +1618,7 @@ Data.iEdgedWidth = pParam->edged_width; Data.currentMV = currentMV; Data.iMinSAD = &iMinSAD; - Data.iQuant = frame->quant; + Data.lambda16 = lambda_vec16[frame->quant]; // note: i==horizontal, j==vertical @@ -1678,6 +1637,7 @@ } 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 */ @@ -1759,52 +1719,47 @@ 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) + 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; - Data->quarterpel = OldData->quarterpel; + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - if (block != 0) { - if(pParam->m_quarterpel) { - *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * + if(pParam->m_quarterpel) { + 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); - } - else { - *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * + 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 + temp_sad = *(Data->iMinSAD); // store current MinSAD if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; @@ -1828,22 +1783,17 @@ } } - if((Data->quarterpel) && (!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && - (MotionFlags & PMV_QUARTERPELREFINE8)) { - - CheckCandidate = CheckCandidate8_qpel; - - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, - pParam->width, pParam->height, OldData->iFcode, 0); // get real range - - QuarterpelRefine(Data); - } - if(pParam->m_quarterpel) { - pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; - pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; - } - else { + 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; + 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; } @@ -1854,13 +1804,11 @@ pMB->sad8[block] = 4 * (*Data->iMinSAD); } - static void SearchPhinted ( const uint8_t * 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, @@ -1878,7 +1826,6 @@ int i, t; MainSearchFunc * MainSearchPtr; - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); Data->predMV = get_pmv2(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); @@ -1888,9 +1835,8 @@ Data->RefH = pRefH + (x + iEdgedWidth*y) * 16; Data->RefV = pRefV + (x + iEdgedWidth*y) * 16; Data->RefHV = pRefHV + (x + iEdgedWidth*y) * 16; - Data->RefQ = pRefQ; - - Data->iQuant = iQuant; + Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; if (!(MotionFlags & PMV_HALFPEL16)) { Data->min_dx = EVEN(Data->min_dx); @@ -1903,10 +1849,10 @@ 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); @@ -1941,22 +1887,18 @@ } 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); // get real range - + pParam->width, pParam->height, Data->iFcode, 0); + CheckCandidate = CheckCandidate16_qpel; QuarterpelRefine(Data); } if (inter4v) { SearchData Data8; Data8.iFcode = Data->iFcode; - Data8.iQuant = Data->iQuant; + 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); @@ -1967,16 +1909,24 @@ (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; @@ -1999,11 +1949,10 @@ const IMAGE *const pRef = &reference->image; uint32_t x, y; - uint8_t *qimage; + uint8_t * qimage; int32_t temp[5], quant = current->quant; int32_t iMinSAD[5]; - VECTOR currentMV[5]; - VECTOR currentQMV[5]; + VECTOR currentMV[5], currentQMV[5]; SearchData Data; Data.iEdgedWidth = pParam->edged_width; Data.currentMV = currentMV; @@ -2018,6 +1967,8 @@ // 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++) { @@ -2032,15 +1983,16 @@ 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, qimage, pCurrent, x, + SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, y, current->motion_flags, pMB->quant, pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, &Data); @@ -2061,7 +2013,7 @@ SearchData * const Data) { - int i, mask; + int i = 255, mask; VECTOR pmv[3]; *(Data->iMinSAD) = MV_MAX_ERROR; @@ -2088,8 +2040,7 @@ 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); } @@ -2114,28 +2065,24 @@ Data.currentMV = ¤tMV; Data.iMinSAD = &iMinSAD; Data.iFcode = iFcode; - Data.iQuant = 2; if (sadInit) (*sadInit) (); - for (y = 0; y < pParam->mb_height-1; y++) { - for (x = 0; x < pParam->mb_width; x++) { + for (y = 1; y < pParam->mb_height-1; y++) { + for (x = 1; x < pParam->mb_width-1; x++) { int sad, dev; MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, pMB, &Data); - if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work - if (sad > INTRA_THRESH) { - 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 (sad > INTRA_THRESH) { + 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 } - sSAD += sad; - } - + sSAD += sad; } } sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); @@ -2169,6 +2116,7 @@ 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;