EncCu.cpp

#if JVET_M0428_ENC_DB_OPT
    if ( m_bestModeUpdated && bestCS->cost != MAX_DOUBLE )
    {
      xCalDebCost( *bestCS, partitioner );
    }
#endif
  }
  tempCS->initStructData(encTestMode.qp, encTestMode.lossless);

  if (cu.lwidth() != 64)
  {
    isPerfectMatch = false;
  }
  m_modeCtrl->setIsHashPerfectMatch(isPerfectMatch);
}
#endif

void EncCu::xCheckRDCostMerge2Nx2N( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner, const EncTestMode& encTestMode )
{
  const Slice &slice = *tempCS->slice;

  CHECK( slice.getSliceType() == I_SLICE, "Merge modes not available for I-slices" );

  tempCS->initStructData( encTestMode.qp, encTestMode.lossless );

  MergeCtx mergeCtx;
  const SPS &sps = *tempCS->sps;

  if( sps.getSBTMVPEnabledFlag() )
  {
    Size bufSize = g_miScaling.scale( tempCS->area.lumaSize() );
    mergeCtx.subPuMvpMiBuf    = MotionBuf( m_SubPuMiBuf,    bufSize );
  }

#if JVET_M0147_DMVR
  Mv   refinedMvdL0[MAX_NUM_PARTS_IN_CTU][MRG_MAX_NUM_CANDS];
#endif
  setMergeBestSATDCost( MAX_DOUBLE );

  {
    // first get merge candidates
    CodingUnit cu( tempCS->area );
    cu.cs       = tempCS;
    cu.predMode = MODE_INTER;
    cu.slice    = tempCS->slice;
#if HEVC_TILES_WPP
    cu.tileIdx  = tempCS->picture->tileMap->getTileIdxMap(tempCS->area.lumaPos());
#endif

    PredictionUnit pu( tempCS->area );
    pu.cu = &cu;
    pu.cs = tempCS;
#if JVET_M0170_MRG_SHARELIST
    pu.shareParentPos = tempCS->sharedBndPos;
    pu.shareParentSize = tempCS->sharedBndSize;
#endif
    PU::getInterMergeCandidates(pu, mergeCtx
      , 0
    );
#if !JVET_M0068_M0171_MMVD_CLEANUP
    PU::restrictBiPredMergeCands(pu, mergeCtx);
#endif
    PU::getInterMMVDMergeCandidates(pu, mergeCtx);
  }
  bool candHasNoResidual[MRG_MAX_NUM_CANDS + MMVD_ADD_NUM];
  for (uint32_t ui = 0; ui < MRG_MAX_NUM_CANDS + MMVD_ADD_NUM; ui++)
  {
    candHasNoResidual[ui] = false;
  }

  bool                                        bestIsSkip = false;
  bool                                        bestIsMMVDSkip = true;
  PelUnitBuf                                  acMergeBuffer[MRG_MAX_NUM_CANDS];
  PelUnitBuf                                  acMergeRealBuffer[MMVD_MRG_MAX_RD_BUF_NUM];
  PelUnitBuf *                                acMergeTempBuffer[MMVD_MRG_MAX_RD_NUM];
  PelUnitBuf *                                singleMergeTempBuffer;
  int                                         insertPos;
  unsigned                                    uiNumMrgSATDCand = mergeCtx.numValidMergeCand + MMVD_ADD_NUM;

  static_vector<unsigned, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM>  RdModeList;
  bool                                        mrgTempBufSet = false;

  for (unsigned i = 0; i < MRG_MAX_NUM_CANDS + MMVD_ADD_NUM; i++)
  {
    RdModeList.push_back(i);
  }

  const UnitArea localUnitArea(tempCS->area.chromaFormat, Area(0, 0, tempCS->area.Y().width, tempCS->area.Y().height));
  for (unsigned i = 0; i < MMVD_MRG_MAX_RD_BUF_NUM; i++)
  {
    acMergeRealBuffer[i] = m_acMergeBuffer[i].getBuf(localUnitArea);
    if (i < MMVD_MRG_MAX_RD_NUM)
    {
      acMergeTempBuffer[i] = acMergeRealBuffer + i;
    }
    else
    {
      singleMergeTempBuffer = acMergeRealBuffer + i;
    }
  }

  static_vector<unsigned, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM>  RdModeList2; // store the Intra mode for Intrainter
  RdModeList2.clear();
  bool isIntrainterEnabled = sps.getUseMHIntra();
  if (bestCS->area.lwidth() * bestCS->area.lheight() < 64 || bestCS->area.lwidth() >= MAX_CU_SIZE || bestCS->area.lheight() >= MAX_CU_SIZE)
  {
    isIntrainterEnabled = false;
  }
  bool isTestSkipMerge[MRG_MAX_NUM_CANDS]; // record if the merge candidate has tried skip mode
  for (uint32_t idx = 0; idx < MRG_MAX_NUM_CANDS; idx++)
  {
    isTestSkipMerge[idx] = false;
  }
  if( m_pcEncCfg->getUseFastMerge() || isIntrainterEnabled)
  {
    uiNumMrgSATDCand = NUM_MRG_SATD_CAND;
    if (isIntrainterEnabled)
    {
      uiNumMrgSATDCand += 1;
    }
    bestIsSkip       = false;

    if( auto blkCache = dynamic_cast< CacheBlkInfoCtrl* >( m_modeCtrl ) )
    {
#if JVET_M0483_IBC
      if (slice.getSPS()->getIBCFlag())
#else
      if (slice.getSPS()->getIBCMode())
#endif
      {
        ComprCUCtx cuECtx = m_modeCtrl->getComprCUCtx();
        bestIsSkip = blkCache->isSkip(tempCS->area) && cuECtx.bestCU;
      }
      else
      bestIsSkip = blkCache->isSkip( tempCS->area );
      bestIsMMVDSkip = blkCache->isMMVDSkip(tempCS->area);
    }

    if (isIntrainterEnabled) // always perform low complexity check
    {
      bestIsSkip = false;
    }

    static_vector<double, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> candCostList;

    // 1. Pass: get SATD-cost for selected candidates and reduce their count
    if( !bestIsSkip )
    {
      RdModeList.clear();
      mrgTempBufSet       = true;
      const double sqrtLambdaForFirstPass = m_pcRdCost->getMotionLambda( encTestMode.lossless );

      CodingUnit &cu      = tempCS->addCU( tempCS->area, partitioner.chType );
      const double sqrtLambdaForFirstPassIntra = m_pcRdCost->getMotionLambda(cu.transQuantBypass) / double(1 << SCALE_BITS);

      partitioner.setCUData( cu );
      cu.slice            = tempCS->slice;
#if HEVC_TILES_WPP
      cu.tileIdx          = tempCS->picture->tileMap->getTileIdxMap( tempCS->area.lumaPos() );
#endif
      cu.skip             = false;
      cu.mmvdSkip = false;
      cu.triangle         = false;
    //cu.affine
      cu.predMode         = MODE_INTER;
    //cu.LICFlag
      cu.transQuantBypass = encTestMode.lossless;
      cu.chromaQpAdj      = cu.transQuantBypass ? 0 : m_cuChromaQpOffsetIdxPlus1;
      cu.qp               = encTestMode.qp;
    //cu.emtFlag  is set below

      PredictionUnit &pu  = tempCS->addPU( cu, partitioner.chType );

      DistParam distParam;
      const bool bUseHadamard= !encTestMode.lossless;
      m_pcRdCost->setDistParam (distParam, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth (CHANNEL_TYPE_LUMA), COMPONENT_Y, bUseHadamard);

      const UnitArea localUnitArea( tempCS->area.chromaFormat, Area( 0, 0, tempCS->area.Y().width, tempCS->area.Y().height) );
#if JVET_M0483_IBC==0
      uint32_t ibcCand = 0;
      uint32_t numValidMv = mergeCtx.numValidMergeCand;
#endif
      for( uint32_t uiMergeCand = 0; uiMergeCand < mergeCtx.numValidMergeCand; uiMergeCand++ )
      {
#if JVET_M0483_IBC==0
        if ((mergeCtx.interDirNeighbours[uiMergeCand] == 1 || mergeCtx.interDirNeighbours[uiMergeCand] == 3) && tempCS->slice->getRefPic(REF_PIC_LIST_0, mergeCtx.mvFieldNeighbours[uiMergeCand << 1].refIdx)->getPOC() == tempCS->slice->getPOC())
        {
          ibcCand++;
          numValidMv--;
          continue;
        }
#endif
        mergeCtx.setMergeInfo( pu, uiMergeCand );

        PU::spanMotionInfo( pu, mergeCtx );
#if JVET_M0147_DMVR
        pu.mvRefine = true;
#endif
        distParam.cur = singleMergeTempBuffer->Y();
        m_pcInterSearch->motionCompensation(pu, *singleMergeTempBuffer);
        acMergeBuffer[uiMergeCand] = m_acRealMergeBuffer[uiMergeCand].getBuf(localUnitArea);
        acMergeBuffer[uiMergeCand].copyFrom(*singleMergeTempBuffer);
#if JVET_M0147_DMVR
        pu.mvRefine = false;
#endif
        if( mergeCtx.interDirNeighbours[uiMergeCand] == 3 && mergeCtx.mrgTypeNeighbours[uiMergeCand] == MRG_TYPE_DEFAULT_N )
        {
          mergeCtx.mvFieldNeighbours[2*uiMergeCand].mv   = pu.mv[0];
          mergeCtx.mvFieldNeighbours[2*uiMergeCand+1].mv = pu.mv[1];
#if JVET_M0147_DMVR
          {
            int dx, dy, i, j, num = 0;
            dy = std::min<int>(pu.lumaSize().height, DMVR_SUBCU_HEIGHT);
            dx = std::min<int>(pu.lumaSize().width, DMVR_SUBCU_WIDTH);
            if (PU::checkDMVRCondition(pu))
            {
              for (i = 0; i < (pu.lumaSize().height); i += dy)
              {
                for (j = 0; j < (pu.lumaSize().width); j += dx)
                {
                  refinedMvdL0[num][uiMergeCand] = pu.mvdL0SubPu[num];
                  num++;
                }
              }
            }
          }
#endif
        }

        Distortion uiSad = distParam.distFunc(distParam);
        uint32_t uiBitsCand = uiMergeCand + 1;
        if( uiMergeCand == tempCS->slice->getMaxNumMergeCand() - 1 )
        {
          uiBitsCand--;
        }
        uiBitsCand++; // for mmvd_flag
        double cost     = (double)uiSad + (double)uiBitsCand * sqrtLambdaForFirstPass;
        insertPos = -1;
        updateDoubleCandList(uiMergeCand, cost, RdModeList, candCostList, RdModeList2, (uint32_t)NUM_LUMA_MODE, uiNumMrgSATDCand, &insertPos);
        if (insertPos != -1)
        {
          if (insertPos == RdModeList.size() - 1)
          {
            swap(singleMergeTempBuffer, acMergeTempBuffer[insertPos]);
          }
          else
          {
            for (uint32_t i = uint32_t(RdModeList.size()) - 1; i > insertPos; i--)
            {
              swap(acMergeTempBuffer[i - 1], acMergeTempBuffer[i]);
            }
            swap(singleMergeTempBuffer, acMergeTempBuffer[insertPos]);
          }
        }
#if JVET_M0483_IBC==0
        CHECK(std::min(uiMergeCand + 1 - ibcCand, uiNumMrgSATDCand) != RdModeList.size(), "");
#else
        CHECK(std::min(uiMergeCand + 1, uiNumMrgSATDCand) != RdModeList.size(), "");
#endif
      }
#if JVET_M0483_IBC==0
      if (numValidMv < uiNumMrgSATDCand)
        uiNumMrgSATDCand = numValidMv;
      if (numValidMv == 0)
        return;
#endif

      if (isIntrainterEnabled)
      {
        int numTestIntraMode = 4;
        // prepare for Intra bits calculation
        const TempCtx ctxStart(m_CtxCache, m_CABACEstimator->getCtx());
        const TempCtx ctxStartIntraMode(m_CtxCache, SubCtx(Ctx::MHIntraPredMode, m_CABACEstimator->getCtx()));

        // for Intrainter fast, recored the best intra mode during the first round for mrege 0
        int bestMHIntraMode = -1;
        double bestMHIntraCost = MAX_DOUBLE;

        pu.mhIntraFlag = true;

        // save the to-be-tested merge candidates
        uint32_t MHIntraMergeCand[NUM_MRG_SATD_CAND];
#if JVET_M0483_IBC==0
        for (uint32_t mergeCnt = 0; mergeCnt < std::min(NUM_MRG_SATD_CAND, (const int) uiNumMrgSATDCand); mergeCnt++)
#else
        for (uint32_t mergeCnt = 0; mergeCnt < NUM_MRG_SATD_CAND; mergeCnt++)
#endif
        {
          MHIntraMergeCand[mergeCnt] = RdModeList[mergeCnt];
        }
#if JVET_M0483_IBC==0
        for (uint32_t mergeCnt = 0; mergeCnt < std::min( std::min(NUM_MRG_SATD_CAND, (const int)uiNumMrgSATDCand), 4); mergeCnt++)
#else
        for (uint32_t mergeCnt = 0; mergeCnt < std::min(NUM_MRG_SATD_CAND, 4); mergeCnt++)
#endif
        {
          uint32_t mergeCand = MHIntraMergeCand[mergeCnt];
          acMergeBuffer[mergeCand] = m_acRealMergeBuffer[mergeCand].getBuf(localUnitArea);

          // estimate merge bits
          uint32_t bitsCand = mergeCand + 1;
          if (mergeCand == pu.cs->slice->getMaxNumMergeCand() - 1)
          {
            bitsCand--;
          }

          // first round
          for (uint32_t intraCnt = 0; intraCnt < numTestIntraMode; intraCnt++)
          {
            pu.intraDir[0] = (intraCnt < 2) ? intraCnt : ((intraCnt == 2) ? HOR_IDX : VER_IDX);

            // fast 2
            if (mergeCnt > 0 && bestMHIntraMode != pu.intraDir[0])
            {
              continue;
            }
            int narrowCase = PU::getNarrowShape(pu.lwidth(), pu.lheight());
            if (narrowCase == 1 && pu.intraDir[0] == HOR_IDX)
            {
              continue;
            }
            if (narrowCase == 2 && pu.intraDir[0] == VER_IDX)
            {
              continue;
            }
            // generate intrainter Y prediction
            if (mergeCnt == 0)
            {
              bool isUseFilter = IntraPrediction::useFilteredIntraRefSamples(COMPONENT_Y, pu, true, pu);
              m_pcIntraSearch->initIntraPatternChType(*pu.cu, pu.Y(), isUseFilter);
              m_pcIntraSearch->predIntraAng(COMPONENT_Y, pu.cs->getPredBuf(pu).Y(), pu, isUseFilter);
              m_pcIntraSearch->switchBuffer(pu, COMPONENT_Y, pu.cs->getPredBuf(pu).Y(), m_pcIntraSearch->getPredictorPtr2(COMPONENT_Y, intraCnt));
            }
            pu.cs->getPredBuf(pu).copyFrom(acMergeBuffer[mergeCand]);
#if JVET_M0427_INLOOP_RESHAPER
            if (pu.cs->slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag())
            {
              pu.cs->getPredBuf(pu).Y().rspSignal(m_pcReshape->getFwdLUT());
            }
#endif
            m_pcIntraSearch->geneWeightedPred(COMPONENT_Y, pu.cs->getPredBuf(pu).Y(), pu, m_pcIntraSearch->getPredictorPtr2(COMPONENT_Y, intraCnt));

            // calculate cost
#if JVET_M0427_INLOOP_RESHAPER
            if (pu.cs->slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag())
            {
               pu.cs->getPredBuf(pu).Y().rspSignal(m_pcReshape->getInvLUT());
            }
#endif
            distParam.cur = pu.cs->getPredBuf(pu).Y();
            Distortion sadValue = distParam.distFunc(distParam);
#if JVET_M0427_INLOOP_RESHAPER
            if (pu.cs->slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag())
            {
              pu.cs->getPredBuf(pu).Y().rspSignal(m_pcReshape->getFwdLUT());
            }
#endif
            m_CABACEstimator->getCtx() = SubCtx(Ctx::MHIntraPredMode, ctxStartIntraMode);
            uint64_t fracModeBits = m_pcIntraSearch->xFracModeBitsIntra(pu, pu.intraDir[0], CHANNEL_TYPE_LUMA);
            double cost = (double)sadValue + (double)(bitsCand + 1) * sqrtLambdaForFirstPass + (double)fracModeBits * sqrtLambdaForFirstPassIntra;
            insertPos = -1;
            updateDoubleCandList(mergeCand + MRG_MAX_NUM_CANDS + MMVD_ADD_NUM, cost, RdModeList, candCostList, RdModeList2, pu.intraDir[0], uiNumMrgSATDCand, &insertPos);
            if (insertPos != -1)
            {
              for (int i = int(RdModeList.size()) - 1; i > insertPos; i--)
              {
                swap(acMergeTempBuffer[i - 1], acMergeTempBuffer[i]);
              }
              swap(singleMergeTempBuffer, acMergeTempBuffer[insertPos]);
            }
            // fast 2
            if (mergeCnt == 0 && cost < bestMHIntraCost)
            {
              bestMHIntraMode = pu.intraDir[0];
              bestMHIntraCost = cost;
            }
          }
        }
        pu.mhIntraFlag = false;
        m_CABACEstimator->getCtx() = ctxStart;
      }

      cu.mmvdSkip = true;
      int tempNum = 0;
      tempNum = MMVD_ADD_NUM;
#if !JVET_M0823_MMVD_ENCOPT
      bool allowDirection[4] = { true, true, true, true };
#endif
      for (uint32_t mergeCand = mergeCtx.numValidMergeCand; mergeCand < mergeCtx.numValidMergeCand + tempNum; mergeCand++)
      {
        const int mmvdMergeCand = mergeCand - mergeCtx.numValidMergeCand;
        int bitsBaseIdx = 0;
        int bitsRefineStep = 0;
        int bitsDirection = 2;
        int bitsCand = 0;
        int baseIdx;
        int refineStep;
#if !JVET_M0823_MMVD_ENCOPT
        int direction;
#endif
        baseIdx = mmvdMergeCand / MMVD_MAX_REFINE_NUM;
        refineStep = (mmvdMergeCand - (baseIdx * MMVD_MAX_REFINE_NUM)) / 4;
#if !JVET_M0823_MMVD_ENCOPT
        direction = (mmvdMergeCand - baseIdx * MMVD_MAX_REFINE_NUM - refineStep * 4) % 4;
        if (refineStep == 0)
        {
          allowDirection[direction] = true;
        }
        if (allowDirection[direction] == false)
        {
          continue;
        }
#endif
        bitsBaseIdx = baseIdx + 1;
        if (baseIdx == MMVD_BASE_MV_NUM - 1)
        {
          bitsBaseIdx--;
        }

        bitsRefineStep = refineStep + 1;
        if (refineStep == MMVD_REFINE_STEP - 1)
        {
          bitsRefineStep--;
        }

        bitsCand = bitsBaseIdx + bitsRefineStep + bitsDirection;
        bitsCand++; // for mmvd_flag

        mergeCtx.setMmvdMergeCandiInfo(pu, mmvdMergeCand);

        PU::spanMotionInfo(pu, mergeCtx);
#if JVET_M0147_DMVR
        pu.mvRefine = true;
#endif
        distParam.cur = singleMergeTempBuffer->Y();
#if JVET_M0823_MMVD_ENCOPT
        pu.mmvdEncOptMode = (refineStep > 2 ? 2 : 1);
        CHECK(!pu.mmvdMergeFlag, "MMVD merge should be set");
        // Don't do chroma MC here
        m_pcInterSearch->motionCompensation(pu, *singleMergeTempBuffer, REF_PIC_LIST_X, true, false);
        pu.mmvdEncOptMode = 0;
#else
        m_pcInterSearch->motionCompensation(pu, *singleMergeTempBuffer);
#endif
#if JVET_M0147_DMVR // store the refined MV
        pu.mvRefine = false;
#endif
        Distortion uiSad = distParam.distFunc(distParam);


        double cost = (double)uiSad + (double)bitsCand * sqrtLambdaForFirstPass;
#if !JVET_M0823_MMVD_ENCOPT
        allowDirection[direction] = cost >  1.3 * candCostList[0] ? 0 : 1;
#endif
        insertPos = -1;
        updateDoubleCandList(mergeCand, cost, RdModeList, candCostList, RdModeList2, (uint32_t)NUM_LUMA_MODE, uiNumMrgSATDCand, &insertPos);
        if (insertPos != -1)
        {
          for (int i = int(RdModeList.size()) - 1; i > insertPos; i--)
          {
            swap(acMergeTempBuffer[i - 1], acMergeTempBuffer[i]);
          }
          swap(singleMergeTempBuffer, acMergeTempBuffer[insertPos]);
        }
      }

      // Try to limit number of candidates using SATD-costs
      for( uint32_t i = 1; i < uiNumMrgSATDCand; i++ )
      {
        if( candCostList[i] > MRG_FAST_RATIO * candCostList[0] )
        {
          uiNumMrgSATDCand = i;
          break;
        }
      }

      setMergeBestSATDCost( candCostList[0] );

      if (isIntrainterEnabled)
      {
        pu.mhIntraFlag = true;
        for (uint32_t mergeCnt = 0; mergeCnt < uiNumMrgSATDCand; mergeCnt++)
        {
          if (RdModeList[mergeCnt] >= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM))
          {
            pu.intraDir[0] = RdModeList2[mergeCnt];
            pu.intraDir[1] = DM_CHROMA_IDX;
            uint32_t bufIdx = (pu.intraDir[0] > 1) ? (pu.intraDir[0] == HOR_IDX ? 2 : 3) : pu.intraDir[0];
            bool isUseFilter = IntraPrediction::useFilteredIntraRefSamples(COMPONENT_Cb, pu, true, pu);
            m_pcIntraSearch->initIntraPatternChType(*pu.cu, pu.Cb(), isUseFilter);
            m_pcIntraSearch->predIntraAng(COMPONENT_Cb, pu.cs->getPredBuf(pu).Cb(), pu, isUseFilter);
            m_pcIntraSearch->switchBuffer(pu, COMPONENT_Cb, pu.cs->getPredBuf(pu).Cb(), m_pcIntraSearch->getPredictorPtr2(COMPONENT_Cb, bufIdx));
            isUseFilter = IntraPrediction::useFilteredIntraRefSamples(COMPONENT_Cr, pu, true, pu);
            m_pcIntraSearch->initIntraPatternChType(*pu.cu, pu.Cr(), isUseFilter);
            m_pcIntraSearch->predIntraAng(COMPONENT_Cr, pu.cs->getPredBuf(pu).Cr(), pu, isUseFilter);
            m_pcIntraSearch->switchBuffer(pu, COMPONENT_Cr, pu.cs->getPredBuf(pu).Cr(), m_pcIntraSearch->getPredictorPtr2(COMPONENT_Cr, bufIdx));
          }
        }
        pu.mhIntraFlag = false;
      }

      tempCS->initStructData( encTestMode.qp, encTestMode.lossless );
    }
    else
    {
      if (bestIsMMVDSkip)
      {
        uiNumMrgSATDCand = mergeCtx.numValidMergeCand + MMVD_ADD_NUM;
      }
      else
      {
        uiNumMrgSATDCand = mergeCtx.numValidMergeCand;
      }
    }
  }
#if JVET_M0428_ENC_DB_OPT
  m_bestModeUpdated = tempCS->useDbCost = bestCS->useDbCost = false;
#endif
#if !JVET_M0253_HASH_ME
  const uint32_t iteration = encTestMode.lossless ? 1 : 2;

  // 2. Pass: check candidates using full RD test
  for( uint32_t uiNoResidualPass = 0; uiNoResidualPass < iteration; uiNoResidualPass++ )
#else
  uint32_t iteration;
  uint32_t iterationBegin = m_modeCtrl->getIsHashPerfectMatch() ? 1 : 0;
  if (encTestMode.lossless)
  {
    iteration = 1;
    iterationBegin = 0;
  }
  else
  {
    iteration = 2;
  }
  for (uint32_t uiNoResidualPass = iterationBegin; uiNoResidualPass < iteration; ++uiNoResidualPass)
#endif
  {
    for( uint32_t uiMrgHADIdx = 0; uiMrgHADIdx < uiNumMrgSATDCand; uiMrgHADIdx++ )
    {
      uint32_t uiMergeCand = RdModeList[uiMrgHADIdx];

#if JVET_M0483_IBC==0
      if(uiMergeCand < mergeCtx.numValidMergeCand)
        if ((mergeCtx.interDirNeighbours[uiMergeCand] == 1 || mergeCtx.interDirNeighbours[uiMergeCand] == 3) && tempCS->slice->getRefPic(REF_PIC_LIST_0, mergeCtx.mvFieldNeighbours[uiMergeCand << 1].refIdx)->getPOC() == tempCS->slice->getPOC())
        {
          continue;
        }
#endif

      if (uiNoResidualPass != 0 && uiMergeCand >= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM)) // intrainter does not support skip mode
      {
        uiMergeCand -= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM); // for skip, map back to normal merge candidate idx and try RDO
        if (isTestSkipMerge[uiMergeCand])
        {
          continue;
        }
      }

      if (((uiNoResidualPass != 0) && candHasNoResidual[uiMrgHADIdx])
       || ( (uiNoResidualPass == 0) && bestIsSkip ) )
      {
        continue;
      }

      // first get merge candidates
      CodingUnit &cu      = tempCS->addCU( tempCS->area, partitioner.chType );

      partitioner.setCUData( cu );
      cu.slice            = tempCS->slice;
#if HEVC_TILES_WPP
      cu.tileIdx          = tempCS->picture->tileMap->getTileIdxMap( tempCS->area.lumaPos() );
#endif
      cu.skip             = false;
      cu.mmvdSkip = false;
      cu.triangle         = false;
    //cu.affine
      cu.predMode         = MODE_INTER;
    //cu.LICFlag
      cu.transQuantBypass = encTestMode.lossless;
      cu.chromaQpAdj      = cu.transQuantBypass ? 0 : m_cuChromaQpOffsetIdxPlus1;
      cu.qp               = encTestMode.qp;
      PredictionUnit &pu  = tempCS->addPU( cu, partitioner.chType );

      if (uiNoResidualPass == 0 && uiMergeCand >= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM))
      {
        uiMergeCand -= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM);
        cu.mmvdSkip = false;
        mergeCtx.setMergeInfo(pu, uiMergeCand);
        pu.mhIntraFlag = true;
        pu.intraDir[0] = RdModeList2[uiMrgHADIdx];
        CHECK(pu.intraDir[0]<0 || pu.intraDir[0]>(NUM_LUMA_MODE - 1), "out of intra mode");
        pu.intraDir[1] = DM_CHROMA_IDX;
      }

      else if (uiMergeCand >= mergeCtx.numValidMergeCand && uiMergeCand < MRG_MAX_NUM_CANDS + MMVD_ADD_NUM)
      {
        cu.mmvdSkip = true;
        mergeCtx.setMmvdMergeCandiInfo(pu, uiMergeCand - mergeCtx.numValidMergeCand);
      }
      else
      {
        cu.mmvdSkip = false;
        mergeCtx.setMergeInfo(pu, uiMergeCand);
      }
      PU::spanMotionInfo( pu, mergeCtx );

      if( mrgTempBufSet )
      {
#if JVET_M0147_DMVR
        {
          int dx, dy, i, j, num = 0;
          dy = std::min<int>(pu.lumaSize().height, DMVR_SUBCU_HEIGHT);
          dx = std::min<int>(pu.lumaSize().width, DMVR_SUBCU_WIDTH);
          if (PU::checkDMVRCondition(pu))
          {
            for (i = 0; i < (pu.lumaSize().height); i += dy)
            {
              for (j = 0; j < (pu.lumaSize().width); j += dx)
              {
                pu.mvdL0SubPu[num] = refinedMvdL0[num][uiMergeCand];
                num++;
              }
            }
          }
        }
#endif
        if (pu.mhIntraFlag)
        {
          uint32_t bufIdx = (pu.intraDir[0] > 1) ? (pu.intraDir[0] == HOR_IDX ? 2 : 3) : pu.intraDir[0];
          PelBuf tmpBuf = tempCS->getPredBuf(pu).Y();
          tmpBuf.copyFrom(acMergeBuffer[uiMergeCand].Y());
#if JVET_M0427_INLOOP_RESHAPER
          if (pu.cs->slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag())
          {
            tmpBuf.rspSignal(m_pcReshape->getFwdLUT());
          }
#endif
          m_pcIntraSearch->geneWeightedPred(COMPONENT_Y, tmpBuf, pu, m_pcIntraSearch->getPredictorPtr2(COMPONENT_Y, bufIdx));
          tmpBuf = tempCS->getPredBuf(pu).Cb();
          tmpBuf.copyFrom(acMergeBuffer[uiMergeCand].Cb());
          m_pcIntraSearch->geneWeightedPred(COMPONENT_Cb, tmpBuf, pu, m_pcIntraSearch->getPredictorPtr2(COMPONENT_Cb, bufIdx));
          tmpBuf = tempCS->getPredBuf(pu).Cr();
          tmpBuf.copyFrom(acMergeBuffer[uiMergeCand].Cr());
          m_pcIntraSearch->geneWeightedPred(COMPONENT_Cr, tmpBuf, pu, m_pcIntraSearch->getPredictorPtr2(COMPONENT_Cr, bufIdx));
        }
        else
        {
#if JVET_M0823_MMVD_ENCOPT
          if (uiMergeCand >= mergeCtx.numValidMergeCand && uiMergeCand < MRG_MAX_NUM_CANDS + MMVD_ADD_NUM) {
            pu.mmvdEncOptMode = 0;
            m_pcInterSearch->motionCompensation(pu);
          }
          else
#endif
          if (uiNoResidualPass != 0 && uiMergeCand < mergeCtx.numValidMergeCand && RdModeList[uiMrgHADIdx] >= (MRG_MAX_NUM_CANDS + MMVD_ADD_NUM))
          {
            tempCS->getPredBuf().copyFrom(acMergeBuffer[uiMergeCand]);
          }
          else
          {
            tempCS->getPredBuf().copyFrom(*acMergeTempBuffer[uiMrgHADIdx]);
          }
        }
      }
      else
      {
#if JVET_M0147_DMVR
        pu.mvRefine = true;
#endif
        m_pcInterSearch->motionCompensation( pu );
#if JVET_M0147_DMVR
        pu.mvRefine = false;
#endif
      }
      if (!cu.mmvdSkip && !pu.mhIntraFlag && uiNoResidualPass != 0)
      {
        CHECK(uiMergeCand >= mergeCtx.numValidMergeCand, "out of normal merge");
        isTestSkipMerge[uiMergeCand] = true;
      }

#if JVET_M0464_UNI_MTS
      xEncodeInterResidual( tempCS, bestCS, partitioner, encTestMode, uiNoResidualPass, NULL, uiNoResidualPass == 0 ? &candHasNoResidual[uiMrgHADIdx] : NULL );
#else
      xEncodeInterResidual(tempCS, bestCS, partitioner, encTestMode, uiNoResidualPass
        , NULL
        , 1
        , uiNoResidualPass == 0 ? &candHasNoResidual[uiMrgHADIdx] : NULL);
#endif

      if( m_pcEncCfg->getUseFastDecisionForMerge() && !bestIsSkip && !pu.mhIntraFlag)
      {
        bestIsSkip = bestCS->getCU( partitioner.chType )->rootCbf == 0;
      }
      tempCS->initStructData( encTestMode.qp, encTestMode.lossless );
    }// end loop uiMrgHADIdx

    if( uiNoResidualPass == 0 && m_pcEncCfg->getUseEarlySkipDetection() )
    {
      const CodingUnit     &bestCU = *bestCS->getCU( partitioner.chType );
      const PredictionUnit &bestPU = *bestCS->getPU( partitioner.chType );

      if( bestCU.rootCbf == 0 )
      {
        if( bestPU.mergeFlag )
        {
          m_modeCtrl->setEarlySkipDetected();
        }
        else if( m_pcEncCfg->getMotionEstimationSearchMethod() != MESEARCH_SELECTIVE )
        {
          int absolute_MV = 0;

          for( uint32_t uiRefListIdx = 0; uiRefListIdx < 2; uiRefListIdx++ )
          {
            if( slice.getNumRefIdx( RefPicList( uiRefListIdx ) ) > 0 )
            {
              absolute_MV += bestPU.mvd[uiRefListIdx].getAbsHor() + bestPU.mvd[uiRefListIdx].getAbsVer();
            }
          }

          if( absolute_MV == 0 )
          {
            m_modeCtrl->setEarlySkipDetected();
          }
        }
      }
    }
  }
#if JVET_M0428_ENC_DB_OPT
  if ( m_bestModeUpdated && bestCS->cost != MAX_DOUBLE )
  {
    xCalDebCost( *bestCS, partitioner );
  }
#endif
}

void EncCu::xCheckRDCostMergeTriangle2Nx2N( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner, const EncTestMode& encTestMode )
{
  const Slice &slice = *tempCS->slice;
  const SPS &sps = *tempCS->sps;

  CHECK( slice.getSliceType() != B_SLICE, "Triangle mode is only applied to B-slices" );

  tempCS->initStructData( encTestMode.qp, encTestMode.lossless );

  bool trianglecandHasNoResidual[TRIANGLE_MAX_NUM_CANDS];
  for( int mergeCand = 0; mergeCand < TRIANGLE_MAX_NUM_CANDS; mergeCand++ )
  {
    trianglecandHasNoResidual[mergeCand] = false;
  }

  bool bestIsSkip;
  CodingUnit* cuTemp = bestCS->getCU(partitioner.chType);
  if (cuTemp)
    bestIsSkip = m_pcEncCfg->getUseFastDecisionForMerge() ? bestCS->getCU(partitioner.chType)->rootCbf == 0 : false;
  else
    bestIsSkip = false;
  uint8_t                                         numTriangleCandidate   = TRIANGLE_MAX_NUM_CANDS;
  uint8_t                                         triangleNumMrgSATDCand = TRIANGLE_MAX_NUM_SATD_CANDS;
  PelUnitBuf                                      triangleBuffer[TRIANGLE_MAX_NUM_UNI_CANDS];
  PelUnitBuf                                      triangleWeightedBuffer[TRIANGLE_MAX_NUM_CANDS];
  static_vector<uint8_t, TRIANGLE_MAX_NUM_CANDS> triangleRdModeList;
  static_vector<double,  TRIANGLE_MAX_NUM_CANDS> tianglecandCostList;

  if( auto blkCache = dynamic_cast< CacheBlkInfoCtrl* >( m_modeCtrl ) )
  {
    bestIsSkip |= blkCache->isSkip( tempCS->area );
  }

  DistParam distParam;
  const bool useHadamard = !encTestMode.lossless;
  m_pcRdCost->setDistParam( distParam, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth( CHANNEL_TYPE_LUMA ), COMPONENT_Y, useHadamard );

  const UnitArea localUnitArea( tempCS->area.chromaFormat, Area( 0, 0, tempCS->area.Y().width, tempCS->area.Y().height) );

  const double sqrtLambdaForFirstPass = m_pcRdCost->getMotionLambda(encTestMode.lossless);

  MergeCtx triangleMrgCtx;
  {
    CodingUnit cu( tempCS->area );
    cu.cs       = tempCS;
    cu.predMode = MODE_INTER;
    cu.slice    = tempCS->slice;
#if HEVC_TILES_WPP
    cu.tileIdx          = tempCS->picture->tileMap->getTileIdxMap( tempCS->area.lumaPos() );
#endif
    cu.triangle = true;
    cu.mmvdSkip = false;
    cu.GBiIdx   = GBI_DEFAULT;

    PredictionUnit pu( tempCS->area );
    pu.cu = &cu;
    pu.cs = tempCS;


    PU::getTriangleMergeCandidates( pu, triangleMrgCtx );
    for( uint8_t mergeCand = 0; mergeCand < TRIANGLE_MAX_NUM_UNI_CANDS; mergeCand++ )
    {
      triangleBuffer[mergeCand] = m_acMergeBuffer[mergeCand].getBuf(localUnitArea);
      triangleMrgCtx.setMergeInfo( pu, mergeCand );
      PU::spanMotionInfo( pu, triangleMrgCtx );

      m_pcInterSearch->motionCompensation( pu, triangleBuffer[mergeCand] );
    }
  }

  bool tempBufSet = bestIsSkip ? false : true;
  triangleNumMrgSATDCand = bestIsSkip ? TRIANGLE_MAX_NUM_CANDS : TRIANGLE_MAX_NUM_SATD_CANDS;
  if( bestIsSkip )
  {
    for( uint8_t i = 0; i < TRIANGLE_MAX_NUM_CANDS; i++ )
    {
      triangleRdModeList.push_back(i);
    }
  }
  else
  {
    CodingUnit &cu      = tempCS->addCU( tempCS->area, partitioner.chType );

    partitioner.setCUData( cu );
    cu.slice            = tempCS->slice;
#if HEVC_TILES_WPP
    cu.tileIdx          = tempCS->picture->tileMap->getTileIdxMap( tempCS->area.lumaPos() );
#endif
    cu.skip             = false;
    cu.predMode         = MODE_INTER;
    cu.transQuantBypass = encTestMode.lossless;
    cu.chromaQpAdj      = cu.transQuantBypass ? 0 : m_cuChromaQpOffsetIdxPlus1;
    cu.qp               = encTestMode.qp;
    cu.triangle         = true;
    cu.mmvdSkip         = false;
    cu.GBiIdx           = GBI_DEFAULT;

    PredictionUnit &pu  = tempCS->addPU( cu, partitioner.chType );

    if( abs(g_aucLog2[cu.lwidth()] - g_aucLog2[cu.lheight()]) >= 2 )
    {
      numTriangleCandidate = 30;
    }
    else
    {
      numTriangleCandidate = TRIANGLE_MAX_NUM_CANDS;
    }

    for( uint8_t mergeCand = 0; mergeCand < numTriangleCandidate; mergeCand++ )
    {
#if JVET_M0883_TRIANGLE_SIGNALING
      bool    splitDir = m_triangleModeTest[mergeCand].m_splitDir;
      uint8_t candIdx0 = m_triangleModeTest[mergeCand].m_candIdx0;
      uint8_t candIdx1 = m_triangleModeTest[mergeCand].m_candIdx1;
#else
      bool    splitDir = g_triangleCombination[mergeCand][0];
      uint8_t candIdx0 = g_triangleCombination[mergeCand][1];
      uint8_t candIdx1 = g_triangleCombination[mergeCand][2];
#endif

#if JVET_M0883_TRIANGLE_SIGNALING
      pu.triangleSplitDir = splitDir;
      pu.triangleMergeIdx0 = candIdx0;
      pu.triangleMergeIdx1 = candIdx1;
#else
      pu.mergeIdx  = mergeCand;
#endif
      pu.mergeFlag = true;
      triangleWeightedBuffer[mergeCand] = m_acTriangleWeightedBuffer[mergeCand].getBuf( localUnitArea );
      triangleBuffer[candIdx0] = m_acMergeBuffer[candIdx0].getBuf( localUnitArea );
      triangleBuffer[candIdx1] = m_acMergeBuffer[candIdx1].getBuf( localUnitArea );

#if JVET_M0328_KEEP_ONE_WEIGHT_GROUP
      m_pcInterSearch->weightedTriangleBlk( pu, splitDir, CHANNEL_TYPE_LUMA, triangleWeightedBuffer[mergeCand], triangleBuffer[candIdx0], triangleBuffer[candIdx1] );
#else
      m_pcInterSearch->weightedTriangleBlk( pu, PU::getTriangleWeights(pu, triangleMrgCtx, candIdx0, candIdx1), splitDir, CHANNEL_TYPE_LUMA, triangleWeightedBuffer[mergeCand], triangleBuffer[candIdx0], triangleBuffer[candIdx1] );
#endif
      distParam.cur = triangleWeightedBuffer[mergeCand].Y();

      Distortion uiSad = distParam.distFunc( distParam );

#if JVET_M0883_TRIANGLE_SIGNALING
      uint32_t uiBitsCand = m_triangleIdxBins[splitDir][candIdx0][candIdx1];
#else
      uint32_t uiBitsCand = g_triangleIdxBins[mergeCand];
#endif

      double cost = (double)uiSad + (double)uiBitsCand * sqrtLambdaForFirstPass;

      static_vector<int, TRIANGLE_MAX_NUM_CANDS> * nullList = nullptr;
      updateCandList( mergeCand, cost, triangleRdModeList, tianglecandCostList
        , *nullList, -1
        , triangleNumMrgSATDCand );
    }

    // limit number of candidates using SATD-costs
    for( uint8_t i = 0; i < triangleNumMrgSATDCand; i++ )
    {
      if( tianglecandCostList[i] > MRG_FAST_RATIO * tianglecandCostList[0] || tianglecandCostList[i] > getMergeBestSATDCost() )
      {
        triangleNumMrgSATDCand = i;
        break;
      }
    }

    // perform chroma weighting process
    for( uint8_t i = 0; i < triangleNumMrgSATDCand; i++ )
    {
      uint8_t  mergeCand = triangleRdModeList[i];
#if JVET_M0883_TRIANGLE_SIGNALING
      bool     splitDir  = m_triangleModeTest[mergeCand].m_splitDir;
      uint8_t  candIdx0  = m_triangleModeTest[mergeCand].m_candIdx0;
      uint8_t  candIdx1  = m_triangleModeTest[mergeCand].m_candIdx1;
#else
      bool     splitDir  = g_triangleCombination[mergeCand][0];
      uint8_t  candIdx0  = g_triangleCombination[mergeCand][1];
      uint8_t  candIdx1  = g_triangleCombination[mergeCand][2];
#endif

#if JVET_M0883_TRIANGLE_SIGNALING
      pu.triangleSplitDir = splitDir;
      pu.triangleMergeIdx0 = candIdx0;
      pu.triangleMergeIdx1 = candIdx1;
#else
      pu.mergeIdx  = mergeCand;
#endif
      pu.mergeFlag = true;

#if JVET_M0328_KEEP_ONE_WEIGHT_GROUP
      m_pcInterSearch->weightedTriangleBlk( pu, splitDir, CHANNEL_TYPE_CHROMA, triangleWeightedBuffer[mergeCand], triangleBuffer[candIdx0], triangleBuffer[candIdx1] );
#else
      m_pcInterSearch->weightedTriangleBlk( pu, PU::getTriangleWeights(pu, triangleMrgCtx, candIdx0, candIdx1), splitDir, CHANNEL_TYPE_CHROMA, triangleWeightedBuffer[mergeCand], triangleBuffer[candIdx0], triangleBuffer[candIdx1] );
#endif
    }

    tempCS->initStructData( encTestMode.qp, encTestMode.lossless );
  }
#if JVET_M0428_ENC_DB_OPT
  m_bestModeUpdated = tempCS->useDbCost = bestCS->useDbCost = false;
#endif
  {
#if !JVET_M0253_HASH_ME
    const uint8_t iteration = encTestMode.lossless ? 1 : 2;
    for( uint8_t noResidualPass = 0; noResidualPass < iteration; noResidualPass++ )
#else
    uint8_t iteration;
    uint8_t iterationBegin = m_modeCtrl->getIsHashPerfectMatch() ? 1 : 0;
    if (encTestMode.lossless)
    {
      iteration = 1;
      iterationBegin = 0;
    }
    else
    {
      iteration = 2;
    }
    for (uint8_t noResidualPass = iterationBegin; noResidualPass < iteration; ++noResidualPass)
#endif
    {
      for( uint8_t mrgHADIdx = 0; mrgHADIdx < triangleNumMrgSATDCand; mrgHADIdx++ )
      {
        uint8_t mergeCand = triangleRdModeList[mrgHADIdx];

        if ( ( (noResidualPass != 0) && trianglecandHasNoResidual[mergeCand] )
          || ( (noResidualPass == 0) && bestIsSkip ) )
        {
          continue;
        }

#if JVET_M0883_TRIANGLE_SIGNALING
        bool    splitDir = m_triangleModeTest[mergeCand].m_splitDir;
        uint8_t candIdx0 = m_triangleModeTest[mergeCand].m_candIdx0;
        uint8_t candIdx1 = m_triangleModeTest[mergeCand].m_candIdx1;
#else
        bool    splitDir = g_triangleCombination[mergeCand][0];
        uint8_t candIdx0 = g_triangleCombination[mergeCand][1];
        uint8_t candIdx1 = g_triangleCombination[mergeCand][2];
#endif

        CodingUnit &cu = tempCS->addCU(tempCS->area, partitioner.chType);

        partitioner.setCUData(cu);
        cu.slice = tempCS->slice;
#if HEVC_TILES_WPP
        cu.tileIdx          = tempCS->picture->tileMap->getTileIdxMap( tempCS->area.lumaPos() );
#endif
        cu.skip = false;
        cu.predMode = MODE_INTER;
        cu.transQuantBypass = encTestMode.lossless;
        cu.chromaQpAdj = cu.transQuantBypass ? 0 : m_cuChromaQpOffsetIdxPlus1;
        cu.qp = encTestMode.qp;
        cu.triangle = true;
        cu.mmvdSkip = false;
        cu.GBiIdx   = GBI_DEFAULT;
        PredictionUnit &pu = tempCS->addPU(cu, partitioner.chType);

#if JVET_M0883_TRIANGLE_SIGNALING
        pu.triangleSplitDir = splitDir;
        pu.triangleMergeIdx0 = candIdx0;
        pu.triangleMergeIdx1 = candIdx1;
#else
        pu.mergeIdx = mergeCand;
#endif