IntraSearch.cpp

// Intra search
// -------------------------------------------------------------------------------------------------------------------

void IntraSearch::xEncIntraHeader(CodingStructure &cs, Partitioner &partitioner, const bool &bLuma, const bool &bChroma)
{
  CodingUnit &cu = *cs.getCU( partitioner.chType );

  if (bLuma)
  {
    bool isFirst = partitioner.currArea().lumaPos() == cs.area.lumaPos();

    // CU header
    if( isFirst )
    {
#if JVET_M0483_IBC 
      if ((!cs.slice->isIntra() || cs.slice->getSPS()->getSpsNext().getIBCMode())
#else
      if( !cs.slice->isIntra() 
#endif
        && cu.Y().valid()
        )
      {
        if( cs.pps->getTransquantBypassEnabledFlag() )
        {
          m_CABACEstimator->cu_transquant_bypass_flag( cu );
        }
        m_CABACEstimator->cu_skip_flag( cu );
        m_CABACEstimator->pred_mode   ( cu );
      }
      m_CABACEstimator->extend_ref_line(cu);
      if( CU::isIntra(cu) )
      {
        m_CABACEstimator->pcm_data( cu, partitioner );
        if( cu.ipcm )
        {
          return;
        }
      }
    }

    PredictionUnit &pu = *cs.getPU(partitioner.currArea().lumaPos(), partitioner.chType);

    // luma prediction mode
    if (isFirst)
    {
      if ( !cu.Y().valid())
        m_CABACEstimator->pred_mode( cu );
      m_CABACEstimator->intra_luma_pred_mode( pu );
    }
  }

  if (bChroma)
  {
    bool isFirst = partitioner.currArea().Cb().valid() && partitioner.currArea().chromaPos() == cs.area.chromaPos();

    PredictionUnit &pu = *cs.getPU( partitioner.currArea().chromaPos(), CHANNEL_TYPE_CHROMA );

    if( isFirst )
    {
      m_CABACEstimator->intra_chroma_pred_mode( pu );
    }
  }
}

void IntraSearch::xEncSubdivCbfQT(CodingStructure &cs, Partitioner &partitioner, const bool &bLuma, const bool &bChroma)
{
  const UnitArea &currArea = partitioner.currArea();
  TransformUnit &currTU    = *cs.getTU( currArea.blocks[partitioner.chType], partitioner.chType );
#if !JVET_M0464_UNI_MTS
  CodingUnit &currCU       = *currTU.cu;
#endif
  uint32_t currDepth           = partitioner.currTrDepth;

  const bool subdiv        = currTU.depth > currDepth;

  if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
  {
    CHECK( !subdiv, "TU split implied" );
  }
  else
  {
    CHECK( subdiv, "No TU subdivision is allowed with QTBT" );
  }

  if (bChroma)
  {
    const uint32_t numberValidComponents = getNumberValidComponents(currArea.chromaFormat);

    for (uint32_t ch = COMPONENT_Cb; ch < numberValidComponents; ch++)
    {
      const ComponentID compID = ComponentID(ch);

      if( currDepth == 0 || TU::getCbfAtDepth( currTU, compID, currDepth - 1 ) )
      {
        const bool prevCbf = ( compID == COMPONENT_Cr ? TU::getCbfAtDepth( currTU, COMPONENT_Cb, currDepth ) : false );
        m_CABACEstimator->cbf_comp( cs, TU::getCbfAtDepth( currTU, compID, currDepth ), currArea.blocks[compID], currDepth, prevCbf );

      }
    }
  }

  if (subdiv)
  {
#if !JVET_M0464_UNI_MTS
    if( currDepth == 0 && bLuma ) m_CABACEstimator->emt_cu_flag( currCU );
#endif

    if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
    {
      partitioner.splitCurrArea( TU_MAX_TR_SPLIT, cs );
    }
    else
    THROW( "Cannot perform an implicit split!" );

    do
    {
      xEncSubdivCbfQT( cs, partitioner, bLuma, bChroma );
    } while( partitioner.nextPart( cs ) );

    partitioner.exitCurrSplit();
  }
  else
  {
#if !JVET_M0464_UNI_MTS
    if( currDepth == 0 && bLuma && TU::getCbfAtDepth( currTU, COMPONENT_Y, 0 ) ) m_CABACEstimator->emt_cu_flag( currCU );
#endif
    //===== Cbfs =====
    if (bLuma)
    {
      m_CABACEstimator->cbf_comp( cs, TU::getCbfAtDepth( currTU, COMPONENT_Y, currDepth ), currTU.Y(), currTU.depth );
    }
  }
}

void IntraSearch::xEncCoeffQT(CodingStructure &cs, Partitioner &partitioner, const ComponentID &compID)
{
  const UnitArea &currArea  = partitioner.currArea();
  TransformUnit &currTU     = *cs.getTU( currArea.blocks[partitioner.chType], partitioner.chType );
  uint32_t      currDepth       = partitioner.currTrDepth;
  const bool subdiv         = currTU.depth > currDepth;

  if (subdiv)
  {
    if (partitioner.canSplit(TU_MAX_TR_SPLIT, cs))
    {
      partitioner.splitCurrArea(TU_MAX_TR_SPLIT, cs);
    }
    else
      THROW("Implicit TU split not available!");

    do
    {
      xEncCoeffQT( cs, partitioner, compID );
    } while( partitioner.nextPart( cs ) );

    partitioner.exitCurrSplit();
  }
  else

  if( currArea.blocks[compID].valid() )
  {
    if( TU::hasCrossCompPredInfo( currTU, compID ) )
    {
      m_CABACEstimator->cross_comp_pred( currTU, compID );
    }
    if( TU::getCbf( currTU, compID ) )
    {
      m_CABACEstimator->residual_coding( currTU, compID );
    }
  }
}

uint64_t IntraSearch::xGetIntraFracBitsQT( CodingStructure &cs, Partitioner &partitioner, const bool &bLuma, const bool &bChroma )
{
  m_CABACEstimator->resetBits();

  xEncIntraHeader( cs, partitioner, bLuma, bChroma );
  xEncSubdivCbfQT( cs, partitioner, bLuma, bChroma );

  if( bLuma )
  {
    xEncCoeffQT( cs, partitioner, COMPONENT_Y );
  }
  if( bChroma )
  {
    xEncCoeffQT( cs, partitioner, COMPONENT_Cb );
    xEncCoeffQT( cs, partitioner, COMPONENT_Cr );
  }

  uint64_t fracBits = m_CABACEstimator->getEstFracBits();
  return fracBits;
}

uint64_t IntraSearch::xGetIntraFracBitsQTChroma(TransformUnit& currTU, const ComponentID &compID)
{
  m_CABACEstimator->resetBits();

  if( TU::hasCrossCompPredInfo( currTU, compID ) )
  {
    m_CABACEstimator->cross_comp_pred( currTU, compID );
  }
  if( TU::getCbf( currTU, compID ) )
  {
    m_CABACEstimator->residual_coding( currTU, compID );
  }

  uint64_t fracBits = m_CABACEstimator->getEstFracBits();
  return fracBits;
}

#if JVET_M0464_UNI_MTS
void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &compID, const bool &checkCrossCPrediction, Distortion& ruiDist, const int &default0Save1Load2, uint32_t* numSig, std::vector<TrMode>* trModes, const bool loadTr)
#else
void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &compID, const bool &checkCrossCPrediction, Distortion& ruiDist, const int &default0Save1Load2, uint32_t* numSig )
#endif
{
  if (!tu.blocks[compID].valid())
  {
    return;
  }

  CodingStructure &cs                       = *tu.cs;

  const CompArea      &area                 = tu.blocks[compID];
  const SPS           &sps                  = *cs.sps;
  const PPS           &pps                  = *cs.pps;

  const ChannelType    chType               = toChannelType(compID);
  const int            bitDepth             = sps.getBitDepth(chType);

  PelBuf         piOrg                      = cs.getOrgBuf    (area);
  PelBuf         piPred                     = cs.getPredBuf   (area);
  PelBuf         piResi                     = cs.getResiBuf   (area);
  PelBuf         piOrgResi                  = cs.getOrgResiBuf(area);
  PelBuf         piReco                     = cs.getRecoBuf   (area);

  const PredictionUnit &pu                  = *cs.getPU(area.pos(), chType);
  const uint32_t           uiChFinalMode        = PU::getFinalIntraMode(pu, chType);

  const bool           bUseCrossCPrediction = pps.getPpsRangeExtension().getCrossComponentPredictionEnabledFlag() && isChroma( compID ) && PU::isChromaIntraModeCrossCheckMode( pu ) && checkCrossCPrediction;
  const bool           ccUseRecoResi        = m_pcEncCfg->getUseReconBasedCrossCPredictionEstimate();


  //===== init availability pattern =====
  PelBuf sharedPredTS( m_pSharedPredTransformSkip[compID], area );
  if( default0Save1Load2 != 2 )
  {
    const bool bUseFilteredPredictions = IntraPrediction::useFilteredIntraRefSamples( compID, pu, true, tu );
    initIntraPatternChType( *tu.cu, area, bUseFilteredPredictions );

    //===== get prediction signal =====
    if( compID != COMPONENT_Y && PU::isLMCMode( uiChFinalMode ) )
    {
      {
        xGetLumaRecPixels( pu, area );
      }
      predIntraChromaLM( compID, piPred, pu, area, uiChFinalMode );
    }
    else
    {
      predIntraAng( compID, piPred, pu, bUseFilteredPredictions );
    }


    // save prediction
    if( default0Save1Load2 == 1 )
    {
      sharedPredTS.copyFrom( piPred );
    }
  }
  else
  {
    // load prediction
    piPred.copyFrom( sharedPredTS );
  }


  DTRACE( g_trace_ctx, D_PRED, "@(%4d,%4d) [%2dx%2d] IMode=%d\n", tu.lx(), tu.ly(), tu.lwidth(), tu.lheight(), uiChFinalMode );
  //DTRACE_PEL_BUF( D_PRED, piPred, tu, tu.cu->predMode, COMPONENT_Y );

  //===== get residual signal =====
  piResi.copyFrom( piOrg  );
  piResi.subtract( piPred );

  if (pps.getPpsRangeExtension().getCrossComponentPredictionEnabledFlag() && isLuma(compID))
  {
    piOrgResi.copyFrom (piResi);
  }

  if (bUseCrossCPrediction)
  {
    if (xCalcCrossComponentPredictionAlpha(tu, compID, ccUseRecoResi) == 0)
    {
      return;
    }
    CrossComponentPrediction::crossComponentPrediction(tu, compID, cs.getResiBuf(tu.Y()), piResi, piResi, false);
  }

  //===== transform and quantization =====
  //--- init rate estimation arrays for RDOQ ---
  //--- transform and quantization           ---
  TCoeff uiAbsSum = 0;

  const QpParam cQP(tu, compID);

#if RDOQ_CHROMA_LAMBDA
  m_pcTrQuant->selectLambda(compID);
#endif

#if JVET_M0464_UNI_MTS
  if( trModes )
  {
    m_pcTrQuant->transformNxN( tu, compID, cQP, trModes, CU::isIntra( *tu.cu ) ? m_pcEncCfg->getIntraMTSMaxCand() : m_pcEncCfg->getInterMTSMaxCand() );
    tu.mtsIdx = trModes->at(0).first;
  }
  m_pcTrQuant->transformNxN(tu, compID, cQP, uiAbsSum, m_CABACEstimator->getCtx(), loadTr);
#else
  m_pcTrQuant->transformNxN(tu, compID, cQP, uiAbsSum, m_CABACEstimator->getCtx());
#endif


  DTRACE( g_trace_ctx, D_TU_ABS_SUM, "%d: comp=%d, abssum=%d\n", DTRACE_GET_COUNTER( g_trace_ctx, D_TU_ABS_SUM ), compID, uiAbsSum );


  //--- inverse transform ---
  if (uiAbsSum > 0)
  {
    m_pcTrQuant->invTransformNxN(tu, compID, piResi, cQP);
  }
  else
  {
    piResi.fill(0);
  }

  //===== reconstruction =====
  if (bUseCrossCPrediction)
  {
    CrossComponentPrediction::crossComponentPrediction(tu, compID, cs.getResiBuf(tu.Y()), piResi, piResi, true);
  }

  piReco.reconstruct(piPred, piResi, cs.slice->clpRng( compID ));

  //===== update distortion =====
#if WCG_EXT
  if( m_pcEncCfg->getLumaLevelToDeltaQPMapping().isEnabled() )
  {
    const CPelBuf orgLuma = cs.getOrgBuf( cs.area.blocks[COMPONENT_Y] );
    ruiDist += m_pcRdCost->getDistPart( piOrg, piReco, bitDepth, compID, DF_SSE_WTD, &orgLuma );
  }
  else
#endif
  {
    ruiDist += m_pcRdCost->getDistPart( piOrg, piReco, bitDepth, compID, DF_SSE );
  }
}

void IntraSearch::xRecurIntraCodingLumaQT( CodingStructure &cs, Partitioner &partitioner )
{
  const UnitArea &currArea = partitioner.currArea();
#if !JVET_M0464_UNI_MTS
  const CodingUnit &cu     = *cs.getCU(currArea.lumaPos(), partitioner.chType);
#endif
  uint32_t currDepth       = partitioner.currTrDepth;
  const PPS &pps           = *cs.pps;
  const bool keepResi      = pps.getPpsRangeExtension().getCrossComponentPredictionEnabledFlag() || KEEP_PRED_AND_RESI_SIGNALS;
  bool bCheckFull          = true;
  bool bCheckSplit         = false;
  bCheckFull               = !partitioner.canSplit( TU_MAX_TR_SPLIT, cs );
  bCheckSplit              = partitioner.canSplit( TU_MAX_TR_SPLIT, cs );

  uint32_t    numSig           = 0;

#if JVET_M0464_UNI_MTS
  double     dSingleCost                        = MAX_DOUBLE;
  Distortion uiSingleDistLuma                   = 0;
  uint64_t   singleFracBits                     = 0;
  int        bestModeId[MAX_NUM_COMPONENT]      = { 0, 0, 0 };
#else
  bool    checkInitTrDepth = false, checkInitTrDepthTransformSkipWinner = false;

  double     dSingleCost                        = MAX_DOUBLE;
  Distortion uiSingleDistLuma                   = 0;
  uint64_t     singleFracBits                     = 0;
  bool       checkTransformSkip                 = pps.getUseTransformSkip();
  int        bestModeId[MAX_NUM_COMPONENT]      = {0, 0, 0};
  uint8_t      nNumTransformCands                 = cu.emtFlag ? 4 : 1; //4 is the number of transforms of emt
  bool       isAllIntra                         = m_pcEncCfg->getIntraPeriod() == 1;

  uint8_t numTransformIndexCands                  = nNumTransformCands;
#endif

  const TempCtx ctxStart  ( m_CtxCache, m_CABACEstimator->getCtx() );
  TempCtx       ctxBest   ( m_CtxCache );

  CodingStructure *csSplit = nullptr;
  CodingStructure *csFull  = nullptr;

  if( bCheckSplit )
  {
    csSplit = &cs;
  }
  else if( bCheckFull )
  {
    csFull = &cs;
  }

  if( bCheckFull )
  {
    csFull->cost = 0.0;

    TransformUnit &tu = csFull->addTU( CS::getArea( *csFull, currArea, partitioner.chType ), partitioner.chType );
    tu.depth = currDepth;

#if JVET_M0464_UNI_MTS
    const bool tsAllowed  = TU::isTSAllowed ( tu, COMPONENT_Y );
    const bool mtsAllowed = TU::isMTSAllowed( tu, COMPONENT_Y );
    uint8_t nNumTransformCands = 1 + ( tsAllowed ? 1 : 0 ) + ( mtsAllowed ? 4 : 0 ); // DCT + TS + 4 MTS = 6 tests
    std::vector<TrMode> trModes;
    trModes.push_back( TrMode( 0, true ) ); //DCT2
    if( tsAllowed )
    {
      trModes.push_back( TrMode( 1, true ) );
    }
    if( mtsAllowed )
    {
      for( int i = 2; i < 6; i++ )
      {
        trModes.push_back( TrMode( i, true) );
      }
    }

    CHECK( !tu.Y().valid(), "Invalid TU" );
#else
    checkTransformSkip &= TU::hasTransformSkipFlag( *tu.cs, tu.Y() );
    checkTransformSkip &= !cu.transQuantBypass;
    checkTransformSkip &= !cu.emtFlag;

    CHECK( !tu.Y().valid(), "Invalid TU" );

    //this prevents transformSkip from being checked because we already know it's not the best mode
    checkTransformSkip = ( checkInitTrDepth && !checkInitTrDepthTransformSkipWinner ) ? false : checkTransformSkip;


    CHECK( checkInitTrDepthTransformSkipWinner && !checkTransformSkip, "Transform Skip must be enabled if it was the winner in the previous call of xRecurIntraCodingLumaQT!" );
#endif

    CodingStructure &saveCS = *m_pSaveCS[0];

    TransformUnit *tmpTU = nullptr;

    Distortion singleDistTmpLuma = 0;
    uint64_t     singleTmpFracBits = 0;
    double     singleCostTmp     = 0;
    int        firstCheckId      = 0;

#if JVET_M0464_UNI_MTS
    int       lastCheckId        = trModes[nNumTransformCands-1].first;
    bool isNotOnlyOneMode        = nNumTransformCands != 1;
#else
    //we add the EMT candidates to the loop. TransformSkip will still be the last one to be checked (when modeId == lastCheckId) as long as checkTransformSkip is true
    int        lastCheckId       = numTransformIndexCands - ( firstCheckId + 1 ) + ( int ) checkTransformSkip;
    bool isNotOnlyOneMode        = lastCheckId != firstCheckId && !checkInitTrDepthTransformSkipWinner;
#endif

    if( isNotOnlyOneMode )
    {
      saveCS.pcv     = cs.pcv;
      saveCS.picture = cs.picture;
      saveCS.area.repositionTo(cs.area);
      saveCS.clearTUs();
      tmpTU = &saveCS.addTU(currArea, partitioner.chType);
    }

#if JVET_M0464_UNI_MTS
    bool    cbfDCT2  = true;
#else
    bool cbfBestMode = false;
#endif

#if JVET_M0464_UNI_MTS
    for( int modeId = firstCheckId; modeId < nNumTransformCands; modeId++ )
    {
      if( !cbfDCT2 || ( m_pcEncCfg->getUseTransformSkipFast() && bestModeId[COMPONENT_Y] == 1 ) )
      {
        break;
      }
      if( !trModes[modeId].second )
      {
        continue;
      }
      tu.mtsIdx = trModes[modeId].first;
#else
    for( int modeId = firstCheckId; modeId <= lastCheckId; modeId++ )
    {
      if( checkInitTrDepthTransformSkipWinner )
      {
        //If this is a full RQT call and the winner of the first call (checkFirst=true) was transformSkip, then we skip the first iteration of the loop, since transform skip always comes at the end
        if( modeId == firstCheckId )
        {
          continue;
        }
      }

      uint8_t transformIndex = modeId;


      if( ( transformIndex < lastCheckId ) || ( ( transformIndex == lastCheckId ) && !checkTransformSkip ) ) //we avoid this if the mode is transformSkip
      {
        // Skip checking other transform candidates if zero CBF is encountered and it is the best transform so far
        if( m_pcEncCfg->getFastIntraEMT() && isAllIntra && transformIndex && !cbfBestMode )
        {
          continue;
        }
      }
#endif

      if ((modeId != firstCheckId) && isNotOnlyOneMode)
      {
        m_CABACEstimator->getCtx() = ctxStart;
      }

      int default0Save1Load2 = 0;
      singleDistTmpLuma = 0;

#if JVET_M0464_UNI_MTS
      if( modeId == firstCheckId && nNumTransformCands > 1 )
#else
      if (modeId == firstCheckId && modeId != lastCheckId && !checkInitTrDepthTransformSkipWinner )
#endif
      {
        default0Save1Load2 = 1;
      }
      else if (modeId != firstCheckId)
      {
        default0Save1Load2 = 2;
      }
#if JVET_M0464_UNI_MTS
      if( nNumTransformCands > 1 )
      {
        xIntraCodingTUBlock( tu, COMPONENT_Y, false, singleDistTmpLuma, default0Save1Load2, &numSig, modeId == 0 ? &trModes : nullptr, true );
        if( modeId == 0 )
        {
          for( int i = 0; i < nNumTransformCands; i++ )
          {
            if( trModes[i].second )
            {
              lastCheckId = trModes[i].first;
            }
          }
        }
      }
      else
      {
        xIntraCodingTUBlock( tu, COMPONENT_Y, false, singleDistTmpLuma, default0Save1Load2, &numSig );
      }
#else
      if (cu.emtFlag)
      {
        tu.emtIdx = transformIndex;
      }
      if( !checkTransformSkip )
      {
        tu.transformSkip[COMPONENT_Y] = false;
      }
      else
      {
        tu.transformSkip[COMPONENT_Y] = modeId == lastCheckId;
      }

      xIntraCodingTUBlock( tu, COMPONENT_Y, false, singleDistTmpLuma, default0Save1Load2, &numSig );
#endif

      //----- determine rate and r-d cost -----
#if JVET_M0464_UNI_MTS
      if( ( trModes[modeId].first != 0 && !TU::getCbfAtDepth( tu, COMPONENT_Y, currDepth ) ) )
#else
      //the condition (transformIndex != DCT2_EMT) seems to be irrelevant, since DCT2_EMT=7 and the highest value of transformIndex is 4
      if( ( modeId == lastCheckId && checkTransformSkip && !TU::getCbfAtDepth( tu, COMPONENT_Y, currDepth ) ) )
#endif
      {
        //In order not to code TS flag when cbf is zero, the case for TS with cbf being zero is forbidden.
        singleCostTmp = MAX_DOUBLE;
      }
      else
      {
        singleTmpFracBits = xGetIntraFracBitsQT( *csFull, partitioner, true, false );
        singleCostTmp     = m_pcRdCost->calcRdCost( singleTmpFracBits, singleDistTmpLuma );
      }

      if (singleCostTmp < dSingleCost)
      {
        dSingleCost       = singleCostTmp;
        uiSingleDistLuma  = singleDistTmpLuma;
        singleFracBits    = singleTmpFracBits;

#if JVET_M0464_UNI_MTS
        bestModeId[COMPONENT_Y] = trModes[modeId].first;
        if( trModes[modeId].first == 0 )
        {
          cbfDCT2  = TU::getCbfAtDepth( tu, COMPONENT_Y, currDepth );
        }
#else
        bestModeId[COMPONENT_Y] = modeId;
        cbfBestMode       = TU::getCbfAtDepth( tu, COMPONENT_Y, currDepth );
#endif

        if( bestModeId[COMPONENT_Y] != lastCheckId )
        {
#if KEEP_PRED_AND_RESI_SIGNALS
          saveCS.getPredBuf( tu.Y() ).copyFrom( csFull->getPredBuf( tu.Y() ) );
#endif
          saveCS.getRecoBuf( tu.Y() ).copyFrom( csFull->getRecoBuf( tu.Y() ) );

          if( keepResi )
          {
            saveCS.getResiBuf   ( tu.Y() ).copyFrom( csFull->getResiBuf   ( tu.Y() ) );
            saveCS.getOrgResiBuf( tu.Y() ).copyFrom( csFull->getOrgResiBuf( tu.Y() ) );
          }

          tmpTU->copyComponentFrom( tu, COMPONENT_Y );

          ctxBest = m_CABACEstimator->getCtx();
        }
      }
    }

    if( bestModeId[COMPONENT_Y] != lastCheckId )
    {
#if KEEP_PRED_AND_RESI_SIGNALS
      csFull->getPredBuf( tu.Y() ).copyFrom( saveCS.getPredBuf( tu.Y() ) );
#endif
      csFull->getRecoBuf( tu.Y() ).copyFrom( saveCS.getRecoBuf( tu.Y() ) );

      if( keepResi )
      {
        csFull->getResiBuf   ( tu.Y() ).copyFrom( saveCS.getResiBuf   ( tu.Y() ) );
        csFull->getOrgResiBuf( tu.Y() ).copyFrom( saveCS.getOrgResiBuf( tu.Y() ) );
      }

      tu.copyComponentFrom( *tmpTU, COMPONENT_Y );

      if( !bCheckSplit )
      {
        m_CABACEstimator->getCtx() = ctxBest;
      }
    }
    else if( bCheckSplit )
    {
      ctxBest = m_CABACEstimator->getCtx();
    }

    csFull->cost     += dSingleCost;
    csFull->dist     += uiSingleDistLuma;
    csFull->fracBits += singleFracBits;
  }

  if( bCheckSplit )
  {
    //----- store full entropy coding status, load original entropy coding status -----
    if( bCheckFull )
    {
      m_CABACEstimator->getCtx() = ctxStart;
    }
    //----- code splitted block -----
    csSplit->cost = 0;

    bool uiSplitCbfLuma  = false;
    bool splitIsSelected = true;
    if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
    {
      partitioner.splitCurrArea( TU_MAX_TR_SPLIT, cs );
    }

    do
    {
      xRecurIntraCodingLumaQT( *csSplit, partitioner );

      csSplit->setDecomp( partitioner.currArea().Y() );

      uiSplitCbfLuma |= TU::getCbfAtDepth( *csSplit->getTU( partitioner.currArea().lumaPos(), partitioner.chType ), COMPONENT_Y, partitioner.currTrDepth );



    } while( partitioner.nextPart( *csSplit ) );

    partitioner.exitCurrSplit();

    if( splitIsSelected )
    {
      for( auto &ptu : csSplit->tus )
      {
        if( currArea.Y().contains( ptu->Y() ) )
        {
          TU::setCbfAtDepth( *ptu, COMPONENT_Y, currDepth, uiSplitCbfLuma ? 1 : 0 );
        }
      }

      //----- restore context states -----
      m_CABACEstimator->getCtx() = ctxStart;

      //----- determine rate and r-d cost -----
      csSplit->fracBits = xGetIntraFracBitsQT(*csSplit, partitioner, true, false);

      //--- update cost ---
      csSplit->cost     = m_pcRdCost->calcRdCost(csSplit->fracBits, csSplit->dist);
    }
  }

  if( csFull || csSplit )
  {
    {
      // otherwise this would've happened in useSubStructure
      cs.picture->getRecoBuf( currArea.Y() ).copyFrom( cs.getRecoBuf( currArea.Y() ) );
    }

    cs.cost = m_pcRdCost->calcRdCost( cs.fracBits, cs.dist );
  }
}

ChromaCbfs IntraSearch::xRecurIntraChromaCodingQT(CodingStructure &cs, Partitioner& partitioner)
{
  UnitArea currArea                   = partitioner.currArea();
  const bool keepResi                 = cs.sps->getSpsNext().getUseLMChroma() || KEEP_PRED_AND_RESI_SIGNALS;
  if( !currArea.Cb().valid() ) return ChromaCbfs( false );


  TransformUnit &currTU               = *cs.getTU( currArea.chromaPos(), CHANNEL_TYPE_CHROMA );
  const PredictionUnit &pu            = *cs.getPU( currArea.chromaPos(), CHANNEL_TYPE_CHROMA );
#if !JVET_M0464_UNI_MTS
  const TransformUnit &currTULuma     = CS::isDualITree( cs ) ? *cs.picture->cs->getTU( currArea.lumaPos(), CHANNEL_TYPE_LUMA ) : currTU;
#endif

  uint32_t     currDepth                  = partitioner.currTrDepth;
  const PPS &pps                      = *cs.pps;
  ChromaCbfs cbfs                     ( false );

  if (currDepth == currTU.depth)
  {
    if (!currArea.Cb().valid() || !currArea.Cr().valid())
    {
      return cbfs;
    }

#if !JVET_M0464_UNI_MTS
    bool checkTransformSkip = pps.getUseTransformSkip();
    checkTransformSkip &= TU::hasTransformSkipFlag( *currTU.cs, partitioner.currArea().Cb() );

    if( m_pcEncCfg->getUseTransformSkipFast() )
    {
      checkTransformSkip &= TU::hasTransformSkipFlag( *currTU.cs, partitioner.currArea().Y() );

      if( checkTransformSkip && cs.pcv->noChroma2x2 )
      {
        int nbLumaSkip = currTULuma.transformSkip[0] ? 1 : 0;

        {
          // the chroma blocks are co-located with the last luma block, so backwards references are needed
          nbLumaSkip += cs.getTU( currTULuma.Y().topLeft().offset( -1,  0 ), partitioner.chType )->transformSkip[0] ? 1 : 0;
          nbLumaSkip += cs.getTU( currTULuma.Y().topLeft().offset( -1, -1 ), partitioner.chType )->transformSkip[0] ? 1 : 0;
          nbLumaSkip += cs.getTU( currTULuma.Y().topLeft().offset(  0, -1 ), partitioner.chType )->transformSkip[0] ? 1 : 0;
        }

        checkTransformSkip &= ( nbLumaSkip > 0 );
      }
    }
#endif

    CodingStructure &saveCS = *m_pSaveCS[1];
    saveCS.pcv      = cs.pcv;
    saveCS.picture  = cs.picture;
    saveCS.area.repositionTo( cs.area );
    saveCS.initStructData( MAX_INT, false, true );

    TransformUnit &tmpTU = saveCS.addTU(currArea, partitioner.chType);


    cs.setDecomp(currArea.Cb(), true); // set in advance (required for Cb2/Cr2 in 4:2:2 video)

    const unsigned      numTBlocks  = ::getNumberValidTBlocks( *cs.pcv );

    for( uint32_t c = COMPONENT_Cb; c < numTBlocks; c++)
    {
      const ComponentID compID  = ComponentID(c);
      const CompArea&   area    = currTU.blocks[compID];

      double     dSingleCost    = MAX_DOUBLE;
      int        bestModeId     = 0;
      Distortion singleDistC    = 0;
      Distortion singleDistCTmp = 0;
      double     singleCostTmp  = 0;

      const bool checkCrossComponentPrediction = PU::isChromaIntraModeCrossCheckMode( pu ) && pps.getPpsRangeExtension().getCrossComponentPredictionEnabledFlag() && TU::getCbf( currTU, COMPONENT_Y );

      const int  crossCPredictionModesToTest = checkCrossComponentPrediction ? 2 : 1;
#if JVET_M0464_UNI_MTS
      const int  totalModesToTest            = crossCPredictionModesToTest;
#else
      const int  transformSkipModesToTest    = checkTransformSkip ? 2 : 1;
      const int  totalModesToTest            = crossCPredictionModesToTest * transformSkipModesToTest;
#endif
      const bool isOneMode                   = (totalModesToTest == 1);

      int currModeId = 0;
      int default0Save1Load2 = 0;

      TempCtx ctxStart  ( m_CtxCache );
      TempCtx ctxBest   ( m_CtxCache );

      if (!isOneMode)
      {
        ctxStart = m_CABACEstimator->getCtx();
      }

#if !JVET_M0464_UNI_MTS
      for (int transformSkipModeId = 0; transformSkipModeId < transformSkipModesToTest; transformSkipModeId++)
#endif
      {
        for (int crossCPredictionModeId = 0; crossCPredictionModeId < crossCPredictionModesToTest; crossCPredictionModeId++)
        {
          currTU.compAlpha    [compID] = 0;
#if !JVET_M0464_UNI_MTS
          currTU.transformSkip[compID] = transformSkipModeId;
#endif

          currModeId++;

          const bool isFirstMode = (currModeId == 1);
          const bool isLastMode  = (currModeId == totalModesToTest); // currModeId is indexed from 1

          if (isOneMode)
          {
            default0Save1Load2 = 0;
          }
#if JVET_M0464_UNI_MTS
          else if (!isOneMode && (crossCPredictionModeId == 0))
#else
          else if (!isOneMode && (transformSkipModeId == 0) && (crossCPredictionModeId == 0))
#endif
          {
            default0Save1Load2 = 1; //save prediction on first mode
          }
          else
          {
            default0Save1Load2 = 2; //load it on subsequent modes
          }

          if (!isFirstMode) // if not first mode to be tested
          {
            m_CABACEstimator->getCtx() = ctxStart;
          }

          singleDistCTmp = 0;

          xIntraCodingTUBlock( currTU, compID, crossCPredictionModeId != 0, singleDistCTmp, default0Save1Load2 );

#if JVET_M0464_UNI_MTS
          if( ( ( crossCPredictionModeId == 1 ) && ( currTU.compAlpha[compID] == 0 ) ) ) //In order not to code TS flag when cbf is zero, the case for TS with cbf being zero is forbidden.
#else
          if( ( ( crossCPredictionModeId == 1 ) && ( currTU.compAlpha[compID] == 0 ) ) || ( ( transformSkipModeId == 1 ) && !TU::getCbf( currTU, compID ) ) ) //In order not to code TS flag when cbf is zero, the case for TS with cbf being zero is forbidden.
#endif
          {
            singleCostTmp = MAX_DOUBLE;
          }
          else if( !isOneMode )
          {
            uint64_t fracBitsTmp = xGetIntraFracBitsQTChroma( currTU, compID );
            singleCostTmp = m_pcRdCost->calcRdCost( fracBitsTmp, singleDistCTmp );
          }

          if( singleCostTmp < dSingleCost )
          {
            dSingleCost = singleCostTmp;
            singleDistC = singleDistCTmp;
            bestModeId  = currModeId;

            if( !isLastMode )
            {
#if KEEP_PRED_AND_RESI_SIGNALS
              saveCS.getPredBuf   (area).copyFrom(cs.getPredBuf   (area));
              saveCS.getOrgResiBuf(area).copyFrom(cs.getOrgResiBuf(area));
#endif
              if( keepResi )
              {
                saveCS.getResiBuf (area).copyFrom(cs.getResiBuf   (area));
              }
              saveCS.getRecoBuf   (area).copyFrom(cs.getRecoBuf   (area));

              tmpTU.copyComponentFrom(currTU, compID);

              ctxBest = m_CABACEstimator->getCtx();
            }
          }
        }
      }

      if (bestModeId < totalModesToTest)
      {
#if KEEP_PRED_AND_RESI_SIGNALS
        cs.getPredBuf   (area).copyFrom(saveCS.getPredBuf   (area));
        cs.getOrgResiBuf(area).copyFrom(saveCS.getOrgResiBuf(area));
#endif
        if( keepResi )
        {
          cs.getResiBuf (area).copyFrom(saveCS.getResiBuf   (area));
        }
        cs.getRecoBuf   (area).copyFrom(saveCS.getRecoBuf   (area));

        currTU.copyComponentFrom(tmpTU, compID);

        m_CABACEstimator->getCtx() = ctxBest;
      }

      cs.picture->getRecoBuf(area).copyFrom(cs.getRecoBuf(area));

      cbfs.cbf(compID) = TU::getCbf(currTU, compID);

      cs.dist += singleDistC;
    }
  }
  else
  {
    unsigned    numValidTBlocks   = ::getNumberValidTBlocks( *cs.pcv );
    ChromaCbfs  SplitCbfs         ( false );

    if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
    {
      partitioner.splitCurrArea( TU_MAX_TR_SPLIT, cs );
    }
    else
      THROW( "Implicit TU split not available" );

    do
    {
      ChromaCbfs subCbfs = xRecurIntraChromaCodingQT( cs, partitioner );

      for( uint32_t ch = COMPONENT_Cb; ch < numValidTBlocks; ch++ )
      {
        const ComponentID compID = ComponentID( ch );
        SplitCbfs.cbf( compID ) |= subCbfs.cbf( compID );
      }
    } while( partitioner.nextPart( cs ) );

    partitioner.exitCurrSplit();

    {

      cbfs.Cb |= SplitCbfs.Cb;
      cbfs.Cr |= SplitCbfs.Cr;

      for( auto &ptu : cs.tus )
      {
        if( currArea.Cb().contains( ptu->Cb() ) || ( !ptu->Cb().valid() && currArea.Y().contains( ptu->Y() ) ) )
        {
          TU::setCbfAtDepth( *ptu, COMPONENT_Cb, currDepth, SplitCbfs.Cb );
          TU::setCbfAtDepth( *ptu, COMPONENT_Cr, currDepth, SplitCbfs.Cr );
        }
      }
    }
  }

  return cbfs;
}

uint64_t IntraSearch::xFracModeBitsIntra(PredictionUnit &pu, const uint32_t &uiMode, const ChannelType &chType)
{
  uint32_t orgMode = uiMode;

  if (!pu.mhIntraFlag)
  std::swap(orgMode, pu.intraDir[chType]);

  m_CABACEstimator->resetBits();

  if( isLuma( chType ) )
  {
    if ( pu.mhIntraFlag )
      m_CABACEstimator->MHIntra_luma_pred_modes(*pu.cu);
    else
    {
      m_CABACEstimator->extend_ref_line(pu);
      m_CABACEstimator->intra_luma_pred_mode(pu);
    }
  }
  else
  {
    m_CABACEstimator->intra_chroma_pred_mode( pu );
  }

  if ( !pu.mhIntraFlag )
  std::swap(orgMode, pu.intraDir[chType]);

  return m_CABACEstimator->getEstFracBits();
}



void IntraSearch::encPredIntraDPCM( const ComponentID &compID, PelBuf &pOrg, PelBuf &pDst, const uint32_t &uiDirMode )
{
  CHECK( pOrg.buf == 0, "Encoder DPCM called without original buffer" );

  const int srcStride = m_topRefLength + 1;