--- branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/12/29 11:22:20 748 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2003/01/08 14:37:52 767 @@ -84,8 +84,6 @@ { int sad; const uint32_t stride = data->iEdgedWidth/2; - dx = (dx >> 1) + roundtab_79[dx & 0x3]; - dy = (dy >> 1) + roundtab_79[dy & 0x3]; if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently data->temp[5] = dx; data->temp[6] = dy; // backup @@ -264,7 +262,8 @@ 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(xc, yc, data); + if (data->chroma) data->temp[0] += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); if (data->temp[0] < data->iMinSAD[0]) { data->iMinSAD[0] = data->temp[0]; @@ -754,7 +753,6 @@ uint32_t x, y; uint32_t iIntra = 0; int32_t InterBias, quant = current->quant, sad00; - uint8_t *qimage; // some pre-initialized thingies for SearchP int32_t temp[8]; @@ -780,11 +778,7 @@ Data.qpel = Data.chroma = 0; } - 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; + Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) if (sadInit) (*sadInit) (); for (y = 0; y < mb_height; y++) { @@ -876,7 +870,7 @@ pParam->edged_width); if (deviation < (pMB->sad16 - InterBias)) { - if (++iIntra >= iLimit) { free(qimage); return 1; } + if (++iIntra >= iLimit) return 1; pMB->mode = MODE_INTRA; pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; @@ -888,7 +882,6 @@ } } } - free(qimage); if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); @@ -939,10 +932,7 @@ else pmv[4].x = pmv[4].y = 0; // [1] median prediction - if (rrv) { //median is in halfzero-precision - pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); - pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); - } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } + pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask @@ -957,8 +947,8 @@ if (rrv) { int i; for (i = 0; i < 7; i++) { - pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); - pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick + pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // halfzero->halfpel + pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); } } } @@ -1124,19 +1114,18 @@ Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); if (Data->chroma) { - int sumx, sumy, dx, dy; + int sumx, sumy; if(pParam->m_quarterpel) { - sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; + sumx = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; } else { sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; } - dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; - dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; - Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); + Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], + (sumy >> 3) + roundtab_76[sumy & 0xf], Data); } } @@ -1318,7 +1307,7 @@ if ((x != 0)&&(y != 0)) { pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); - pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); + pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); } else pmv[6].x = pmv[6].y = 0; // more? @@ -1423,44 +1412,26 @@ static void SkipDecisionB(const IMAGE * const pCur, - const IMAGE * const f_Ref, - const IMAGE * const b_Ref, - MACROBLOCK * const pMB, - const uint32_t quant, - const uint32_t x, const uint32_t y, - const SearchData * const Data) + const IMAGE * const f_Ref, + const IMAGE * const b_Ref, + MACROBLOCK * const pMB, + const uint32_t x, const uint32_t y, + const SearchData * const Data) { - int dx, dy, b_dx, b_dy; + int dx = 0, dy = 0, b_dx = 0, b_dy = 0; uint32_t sum; + const int div = 1 + Data->qpel; + int k; + const uint32_t quant = pMB->quant; //this is not full chroma compensation, only it's fullpel approximation. should work though - if (Data->qpel) { - dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + - Data->directmvF[2].y/2 + Data->directmvF[3].y/2; - - dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + - Data->directmvF[2].x/2 + Data->directmvF[3].x/2; - b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + - Data->directmvB[2].y/2 + Data->directmvB[3].y/2; - - b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + - Data->directmvB[2].x/2 + Data->directmvB[3].x/2; - - } else { - dy = Data->directmvF[0].y + Data->directmvF[1].y + - Data->directmvF[2].y + Data->directmvF[3].y; - - dx = Data->directmvF[0].x + Data->directmvF[1].x + - Data->directmvF[2].x + Data->directmvF[3].x; - - b_dy = Data->directmvB[0].y + Data->directmvB[1].y + - Data->directmvB[2].y + Data->directmvB[3].y; - - b_dx = Data->directmvB[0].x + Data->directmvB[1].x + - Data->directmvB[2].x + Data->directmvB[3].x; + for (k = 0; k < 4; k++) { + dy += Data->directmvF[k].y / div; + dx += Data->directmvF[0].x / div; + b_dy += Data->directmvB[0].y / div; + b_dx += Data->directmvB[0].x / div; } - dy = (dy >> 3) + roundtab_76[dy & 0xf]; dx = (dx >> 3) + roundtab_76[dx & 0xf]; b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; @@ -1470,12 +1441,15 @@ f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, Data->iEdgedWidth/2); + + if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * quant) return; //no skip + sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, Data->iEdgedWidth/2); - if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped + if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped } @@ -1559,7 +1533,7 @@ // initial (fast) skip decision if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { - SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); //possible skip - checking chroma if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. } @@ -1762,7 +1736,6 @@ const int32_t TRB = time_pp - time_bp; const int32_t TRD = time_pp; - uint8_t * qimage; // some pre-inintialized data for the rest of the search @@ -1779,12 +1752,7 @@ Data.qpel = pParam->m_quarterpel; Data.rounding = 0; - if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) - return; // 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; - + Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) // note: i==horizontal, j==vertical for (j = 0; j < pParam->mb_height; j++) { @@ -1850,7 +1818,7 @@ // final skip decision if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) - SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); + SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); switch (pMB->mode) { case MODE_FORWARD: @@ -1881,7 +1849,6 @@ } } } - free(qimage); } static __inline void