IntraSearch.cpp

  Pel  *runIndex = tu.getPLTIndex(compBegin);
  bool *runType  = tu.getRunTypes(compBegin);
  m_scanOrder = g_scanOrder[SCAN_UNGROUPED][pltScanMode ? SCAN_TRAV_VER : SCAN_TRAV_HOR][gp_sizeIdxInfo->idxFrom(width)][gp_sizeIdxInfo->idxFrom(height)];
// Trellis initialization
  for (int i = 0; i < 2; i++)
  {
    memset(m_prevRunTypeRDOQ[i], 0, sizeof(Pel)*NUM_TRELLIS_STATE);
    memset(m_prevRunPosRDOQ[i],  0, sizeof(int)*NUM_TRELLIS_STATE);
    memset(m_stateCostRDOQ[i],  0, sizeof (double)*NUM_TRELLIS_STATE);
  }
  for (int state = 0; state < NUM_TRELLIS_STATE; state++)
  {
    m_statePtRDOQ[state][0] = 0;
  }
// Context modeling
  const FracBitsAccess& fracBits = m_CABACEstimator->getCtx().getFracBitsAcess();
  BinFracBits fracBitsPltCopyFlagIndex[RUN_IDX_THRE + 1];
  for (int dist = 0; dist <= RUN_IDX_THRE; dist++)
  {
    const unsigned  ctxId = DeriveCtx::CtxPltCopyFlag(PLT_RUN_INDEX, dist);
    fracBitsPltCopyFlagIndex[dist] = fracBits.getFracBitsArray(Ctx::IdxRunModel( ctxId ) );
  }
  BinFracBits fracBitsPltCopyFlagAbove[RUN_IDX_THRE + 1];
  for (int dist = 0; dist <= RUN_IDX_THRE; dist++)
  {
    const unsigned  ctxId = DeriveCtx::CtxPltCopyFlag(PLT_RUN_COPY, dist);
    fracBitsPltCopyFlagAbove[dist] = fracBits.getFracBitsArray(Ctx::CopyRunModel( ctxId ) );
  }
  const BinFracBits fracBitsPltRunType = fracBits.getFracBitsArray( Ctx::RunTypeFlag() );

// Trellis RDO per CG
  bool contTrellisRD = true;
  for (int subSetId = 0; ( subSetId <= (total - 1) >> LOG2_PALETTE_CG_SIZE ) && contTrellisRD; subSetId++)
  {
    int minSubPos = subSetId << LOG2_PALETTE_CG_SIZE;
    int maxSubPos = minSubPos + (1 << LOG2_PALETTE_CG_SIZE);
    maxSubPos = (maxSubPos > total) ? total : maxSubPos; // if last position is out of the current CU size
    contTrellisRD = deriveSubblockIndexMap(cs, partitioner, compBegin, pltScanMode, minSubPos, maxSubPos, fracBitsPltRunType, fracBitsPltCopyFlagIndex, fracBitsPltCopyFlagAbove, dMinCost, (bool)pltScanMode);
  }
  if (!contTrellisRD)
  {
    return;
  }


// best state at the last scan position
  double  sumRdCost = MAX_DOUBLE;
  uint8_t bestState = 0;
  for (uint8_t state = 0; state < NUM_TRELLIS_STATE; state++)
  {
    if (m_stateCostRDOQ[0][state] < sumRdCost)
    {
      sumRdCost = m_stateCostRDOQ[0][state];
      bestState = state;
    }
  }

     bool checkRunTable  [MAX_CU_BLKSIZE_PLT*MAX_CU_BLKSIZE_PLT];
  uint8_t checkIndexTable[MAX_CU_BLKSIZE_PLT*MAX_CU_BLKSIZE_PLT];
  uint8_t bestStateTable [MAX_CU_BLKSIZE_PLT*MAX_CU_BLKSIZE_PLT];
  uint8_t nextState = bestState;
// best trellis path
  for (int i = (width*height - 1); i >= 0; i--)
  {
    bestStateTable[i] = nextState;
    int rasterPos = m_scanOrder[i].idx;
    nextState = m_statePtRDOQ[nextState][rasterPos];
  }
// reconstruct index and runs based on the state pointers
  for (int i = 0; i < (width*height); i++)
  {
    int rasterPos = m_scanOrder[i].idx;
    int  abovePos = (pltScanMode == PLT_SCAN_HORTRAV) ? m_scanOrder[i].idx - width : m_scanOrder[i].idx - 1;
        nextState = bestStateTable[i];
    if ( nextState == 0 ) // same as the previous
    {
      checkRunTable[rasterPos] = checkRunTable[ m_scanOrder[i - 1].idx ];
      if ( checkRunTable[rasterPos] == PLT_RUN_INDEX )
      {
        checkIndexTable[rasterPos] = checkIndexTable[m_scanOrder[i - 1].idx];
      }
      else
      {
        checkIndexTable[rasterPos] = checkIndexTable[ abovePos ];
      }
    }
    else if (nextState == 1) // CopyAbove mode
    {
      checkRunTable[rasterPos] = PLT_RUN_COPY;
      checkIndexTable[rasterPos] = checkIndexTable[abovePos];
    }
    else if (nextState == 2) // Index mode
    {
      checkRunTable[rasterPos] = PLT_RUN_INDEX;
      checkIndexTable[rasterPos] = m_minErrorIndexMap[rasterPos];
    }
  }

// Escape flag
  m_bestEscape = false;
  for (int pos = 0; pos < (width*height); pos++)
  {
    uint8_t index = checkIndexTable[pos];
    if (index == cu.curPLTSize[compBegin])
    {
      m_bestEscape = true;
      break;
    }
  }

// Horizontal scan v.s vertical scan
  if (sumRdCost < dMinCost)
  {
    cu.useEscape[compBegin] = m_bestEscape;
    m_bestScanRotationMode = pltScanMode;
    memset(idxExist, false, sizeof(bool) * (MAXPLTSIZE + 1));
    for (int pos = 0; pos < (width*height); pos++)
    {
      runIndex[pos] = checkIndexTable[pos];
      runType[pos] = checkRunTable[pos];
      idxExist[checkIndexTable[pos]] = true;
    }
    dMinCost = sumRdCost;
  }
}

bool IntraSearch::deriveSubblockIndexMap(
  CodingStructure& cs,
  Partitioner&  partitioner,
  ComponentID   compBegin,
  PLTScanMode   pltScanMode,
  int           minSubPos,
  int           maxSubPos,
  const BinFracBits& fracBitsPltRunType,
  const BinFracBits* fracBitsPltIndexINDEX,
  const BinFracBits* fracBitsPltIndexCOPY,
  const double minCost,
  bool         useRotate
)
{
  CodingUnit &cu    = *cs.getCU(partitioner.chType);
  uint32_t   height = cu.block(compBegin).height;
  uint32_t   width  = cu.block(compBegin).width;
  int indexMaxValue = cu.curPLTSize[compBegin];

  int refId = 0;
  int currRasterPos, currScanPos, prevScanPos, aboveScanPos, roffset;
  int log2Width = (pltScanMode == PLT_SCAN_HORTRAV) ? floorLog2(width): floorLog2(height);
  int buffersize = (pltScanMode == PLT_SCAN_HORTRAV) ? 2*width: 2*height;
  for (int curPos = minSubPos; curPos < maxSubPos; curPos++)
  {
    currRasterPos = m_scanOrder[curPos].idx;
    prevScanPos = (curPos == 0) ? 0 : (curPos - 1) % buffersize;
    roffset = (curPos >> log2Width) << log2Width;
    aboveScanPos = roffset - (curPos - roffset + 1);
    aboveScanPos %= buffersize;
    currScanPos = curPos % buffersize;
    if ((pltScanMode == PLT_SCAN_HORTRAV && curPos < width) || (pltScanMode == PLT_SCAN_VERTRAV && curPos < height))
    {
      aboveScanPos = -1; // first column/row: above row is not valid
    }

// Trellis stats:
// 1st state: same as previous scanned sample
// 2nd state: Copy_Above mode
// 3rd state: Index mode
// Loop of current state
    for ( int curState = 0; curState < NUM_TRELLIS_STATE; curState++ )
    {
      double    minRdCost          = MAX_DOUBLE;
      int       minState           = 0; // best prevState
      uint8_t   bestRunIndex       = 0;
      bool      bestRunType        = 0;
      bool      bestPrevCodedType  = 0;
      int       bestPrevCodedPos   = 0;
      if ( ( curState == 0 && curPos == 0 ) || ( curState == 1 && aboveScanPos < 0 ) ) // state not available
      {
        m_stateCostRDOQ[1 - refId][curState] = MAX_DOUBLE;
        continue;
      }

      bool    runType  = 0;
      uint8_t runIndex = 0;
      if ( curState == 1 ) // 2nd state: Copy_Above mode
      {
        runType = PLT_RUN_COPY;
      }
      else if ( curState == 2 ) // 3rd state: Index mode
      {
        runType = PLT_RUN_INDEX;
        runIndex = m_minErrorIndexMap[currRasterPos];
      }

// Loop of previous state
      for ( int stateID = 0; stateID < NUM_TRELLIS_STATE; stateID++ )
      {
        if ( m_stateCostRDOQ[refId][stateID] == MAX_DOUBLE )
        {
          continue;
        }
        if ( curState == 0 ) // 1st state: same as previous scanned sample
        {
          runType = m_runMapRDOQ[refId][stateID][prevScanPos];
          runIndex = ( runType == PLT_RUN_INDEX ) ? m_indexMapRDOQ[refId][stateID][ prevScanPos ] : m_indexMapRDOQ[refId][stateID][ aboveScanPos ];
        }
        else if ( curState == 1 ) // 2nd state: Copy_Above mode
        {
          runIndex = m_indexMapRDOQ[refId][stateID][aboveScanPos];
        }
        bool    prevRunType   = m_runMapRDOQ[refId][stateID][prevScanPos];
        uint8_t prevRunIndex  = m_indexMapRDOQ[refId][stateID][prevScanPos];
        uint8_t aboveRunIndex = (aboveScanPos >= 0) ? m_indexMapRDOQ[refId][stateID][aboveScanPos] : 0;
        int      dist = curPos - m_prevRunPosRDOQ[refId][stateID] - 1;
        double rdCost = m_stateCostRDOQ[refId][stateID];
        if ( rdCost >= minRdCost ) continue;

// Calculate Rd cost
        bool prevCodedRunType = m_prevRunTypeRDOQ[refId][stateID];
        int  prevCodedPos     = m_prevRunPosRDOQ [refId][stateID];
        const BinFracBits* fracBitsPt = (m_prevRunTypeRDOQ[refId][stateID] == PLT_RUN_INDEX) ? fracBitsPltIndexINDEX : fracBitsPltIndexCOPY;
        rdCost += rateDistOptPLT(runType, runIndex, prevRunType, prevRunIndex, aboveRunIndex, prevCodedRunType, prevCodedPos, curPos, (pltScanMode == PLT_SCAN_HORTRAV) ? width : height, dist, indexMaxValue, fracBitsPt, fracBitsPltRunType);
        if (rdCost < minRdCost) // update minState ( minRdCost )
        {
          minRdCost    = rdCost;
          minState     = stateID;
          bestRunType  = runType;
          bestRunIndex = runIndex;
          bestPrevCodedType = prevCodedRunType;
          bestPrevCodedPos  = prevCodedPos;
        }
      }
// Update trellis info of current state
      m_stateCostRDOQ  [1 - refId][curState]  = minRdCost;
      m_prevRunTypeRDOQ[1 - refId][curState]  = bestPrevCodedType;
      m_prevRunPosRDOQ [1 - refId][curState]  = bestPrevCodedPos;
      m_statePtRDOQ[curState][currRasterPos] = minState;
      int buffer2update = std::min(buffersize, curPos);
      memcpy(m_indexMapRDOQ[1 - refId][curState], m_indexMapRDOQ[refId][minState], sizeof(uint8_t)*buffer2update);
      memcpy(m_runMapRDOQ[1 - refId][curState], m_runMapRDOQ[refId][minState], sizeof(bool)*buffer2update);
      m_indexMapRDOQ[1 - refId][curState][currScanPos] = bestRunIndex;
      m_runMapRDOQ  [1 - refId][curState][currScanPos] = bestRunType;
    }

    if (useRotate) // early terminate: Rd cost >= min cost in horizontal scan
    {
      if ((m_stateCostRDOQ[1 - refId][0] >= minCost) &&
         (m_stateCostRDOQ[1 - refId][1] >= minCost) &&
         (m_stateCostRDOQ[1 - refId][2] >= minCost) )
      {
        return 0;
      }
    }
    refId = 1 - refId;
  }
  return 1;
}

double IntraSearch::rateDistOptPLT(
  bool      runType,
  uint8_t   runIndex,
  bool      prevRunType,
  uint8_t   prevRunIndex,
  uint8_t   aboveRunIndex,
  bool&     prevCodedRunType,
  int&      prevCodedPos,
  int       scanPos,
  uint32_t  width,
  int       dist,
  int       indexMaxValue,
  const BinFracBits* IndexfracBits,
  const BinFracBits& TypefracBits)
{
  double rdCost = 0.0;
  bool identityFlag = !( (runType != prevRunType) || ( (runType == PLT_RUN_INDEX) && (runIndex != prevRunIndex) ) );

  if ( ( !identityFlag && runType == PLT_RUN_INDEX ) || scanPos == 0 ) // encode index value
  {
    uint8_t refIndex = (prevRunType == PLT_RUN_INDEX) ? prevRunIndex : aboveRunIndex;
    refIndex = (scanPos == 0) ? ( indexMaxValue + 1) : refIndex;
    if ( runIndex == refIndex )
    {
      rdCost = MAX_DOUBLE;
      return rdCost;
    }
    rdCost += m_pcRdCost->getLambda()*(m_truncBinBits[(runIndex > refIndex) ? runIndex - 1 : runIndex][(scanPos == 0) ? (indexMaxValue + 1) : indexMaxValue] << SCALE_BITS);
  }
  rdCost += m_indexError[runIndex][m_scanOrder[scanPos].idx] * (1 << SCALE_BITS);
  if (scanPos > 0)
  {
    rdCost += m_pcRdCost->getLambda()*( identityFlag ? (IndexfracBits[(dist < RUN_IDX_THRE) ? dist : RUN_IDX_THRE].intBits[1]) : (IndexfracBits[(dist < RUN_IDX_THRE) ? dist : RUN_IDX_THRE].intBits[0] ) );
  }
  if ( !identityFlag && scanPos >= width && prevRunType != PLT_RUN_COPY )
  {
    rdCost += m_pcRdCost->getLambda()*TypefracBits.intBits[runType];
  }
  if (!identityFlag || scanPos == 0)
  {
    prevCodedRunType = runType;
    prevCodedPos = scanPos;
  }
  return rdCost;
}
uint32_t IntraSearch::getEpExGolombNumBins(uint32_t symbol, uint32_t count)
{
  uint32_t numBins = 0;
  while (symbol >= (uint32_t)(1 << count))
  {
    numBins++;
    symbol -= 1 << count;
    count++;
  }
  numBins++;
  numBins += count;
  assert(numBins <= 32);
  return numBins;
}

uint32_t IntraSearch::getTruncBinBits(uint32_t symbol, uint32_t maxSymbol)
{
  uint32_t idxCodeBit = 0;
  uint32_t thresh;
  if (maxSymbol > 256)
  {
    uint32_t threshVal = 1 << 8;
    thresh = 8;
    while (threshVal <= maxSymbol)
    {
      thresh++;
      threshVal <<= 1;
    }
    thresh--;
  }
  else
  {
    thresh = g_tbMax[maxSymbol];
  }
  uint32_t uiVal = 1 << thresh;
  assert(uiVal <= maxSymbol);
  assert((uiVal << 1) > maxSymbol);
  assert(symbol < maxSymbol);
  uint32_t b = maxSymbol - uiVal;
  assert(b < uiVal);
  if (symbol < uiVal - b)
  {
    idxCodeBit = thresh;
  }
  else
  {
    idxCodeBit = thresh + 1;
  }
  return idxCodeBit;
}

void IntraSearch::initTBCTable(int bitDepth)
{
  for (uint32_t i = 0; i < m_symbolSize; i++)
  {
    memset(m_truncBinBits[i], 0, sizeof(uint16_t)*(m_symbolSize + 1));
  }
  for (uint32_t i = 0; i < (m_symbolSize + 1); i++)
  {
    for (uint32_t j = 0; j < i; j++)
    {
      m_truncBinBits[j][i] = getTruncBinBits(j, i);
    }
  }
  memset(m_escapeNumBins, 0, sizeof(uint16_t)*m_symbolSize);
  for (uint32_t i = 0; i < m_symbolSize; i++)
  {
    m_escapeNumBins[i] = getEpExGolombNumBins(i, 5);
  }
}

void IntraSearch::calcPixelPred(CodingStructure& cs, Partitioner& partitioner, uint32_t yPos, uint32_t xPos, ComponentID compBegin, uint32_t numComp)
{
  CodingUnit    &cu = *cs.getCU(partitioner.chType);
  TransformUnit &tu = *cs.getTU(partitioner.chType);
  bool lossless = (m_pcEncCfg->getCostMode() == COST_LOSSLESS_CODING && cs.slice->isLossless());

  CPelBuf   orgBuf[3];
  for (int comp = compBegin; comp < (compBegin + numComp); comp++)
  {
    CompArea  area = cu.blocks[comp];
    if (m_pcEncCfg->getLmcs() && (cs.slice->getLmcsEnabledFlag() && m_pcReshape->getCTUFlag()))
    {
      orgBuf[comp] = cs.getPredBuf(area);
    }
    else
    {
      orgBuf[comp] = cs.getOrgBuf(area);
    }
  }

  int qp[3];
  int qpRem[3];
  int qpPer[3];
  int quantiserScale[3];
  int quantiserRightShift[3];
  int rightShiftOffset[3];
  int invquantiserRightShift[3];
  int add[3];
  if (!lossless)
  {
    for (uint32_t ch = compBegin; ch < (compBegin + numComp); ch++)
    {
      QpParam cQP(tu, ComponentID(ch));
      qp[ch]                     = cQP.Qp(true);
      qpRem[ch]                  = qp[ch] % 6;
      qpPer[ch]                  = qp[ch] / 6;
      quantiserScale[ch]         = g_quantScales[0][qpRem[ch]];
      quantiserRightShift[ch]    = QUANT_SHIFT + qpPer[ch];
      rightShiftOffset[ch]       = 1 << (quantiserRightShift[ch] - 1);
      invquantiserRightShift[ch] = IQUANT_SHIFT;
      add[ch]                    = 1 << (invquantiserRightShift[ch] - 1);
    }
  }

  uint32_t scaleX = getComponentScaleX(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
  uint32_t scaleY = getComponentScaleY(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
  for (uint32_t ch = compBegin; ch < (compBegin + numComp); ch++)
  {
    const int channelBitDepth = cu.cs->sps->getBitDepth(toChannelType((ComponentID)ch));
    CompArea  area = cu.blocks[ch];
    PelBuf    recBuf = cs.getRecoBuf(area);
    PLTescapeBuf escapeValue = tu.getescapeValue((ComponentID)ch);
    if (compBegin != COMPONENT_Y || ch == 0)
    {
      if (lossless)
      {
        escapeValue.at(xPos, yPos) = orgBuf[ch].at(xPos, yPos);
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT_VS
        recBuf.at(xPos, yPos)      = orgBuf[ch].at(xPos, yPos);
#else
        recBuf.at(xPos, yPos)      = escapeValue.at(xPos, yPos);
#endif
      }
      else
      {
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT_VS
      escapeValue.at(xPos, yPos) = std::max<TCoeff>(0, ((orgBuf[ch].at(xPos, yPos) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch]));
      assert(escapeValue.at(xPos, yPos) < (TCoeff(1) << (channelBitDepth + 1)));
      TCoeff value = (((escapeValue.at(xPos, yPos)*g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch]) >> invquantiserRightShift[ch];
      recBuf.at(xPos, yPos) = Pel(ClipBD<TCoeff>(value, channelBitDepth));//to be checked
#else
      escapeValue.at(xPos, yPos) = TCoeff(std::max<int>(0, ((orgBuf[ch].at(xPos, yPos) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch])));
      assert(escapeValue.at(xPos, yPos) < (1 << (channelBitDepth + 1)));
      recBuf.at(xPos, yPos) = (((escapeValue.at(xPos, yPos)*g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch]) >> invquantiserRightShift[ch];
      recBuf.at(xPos, yPos) = Pel(ClipBD<int>(recBuf.at(xPos, yPos), channelBitDepth));//to be checked
#endif
      }
    }
    else if (compBegin == COMPONENT_Y && ch > 0 && yPos % (1 << scaleY) == 0 && xPos % (1 << scaleX) == 0)
    {
      uint32_t yPosC = yPos >> scaleY;
      uint32_t xPosC = xPos >> scaleX;
      if (lossless)
      {
        escapeValue.at(xPosC, yPosC) = orgBuf[ch].at(xPosC, yPosC);
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT_VS
        recBuf.at(xPosC, yPosC)      = orgBuf[ch].at(xPosC, yPosC);
#else
        recBuf.at(xPosC, yPosC)      = escapeValue.at(xPosC, yPosC);
#endif
      }
      else
      {
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT_VS
        escapeValue.at(xPosC, yPosC) = std::max<TCoeff>(
          0, ((orgBuf[ch].at(xPosC, yPosC) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch]));
        assert(escapeValue.at(xPosC, yPosC) < (TCoeff(1) << (channelBitDepth + 1)));
        TCoeff value = (((escapeValue.at(xPosC, yPosC) * g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch])
                       >> invquantiserRightShift[ch];
        recBuf.at(xPosC, yPosC) = Pel(ClipBD<TCoeff>(value, channelBitDepth));   // to be checked
#else
        escapeValue.at(xPosC, yPosC) = TCoeff(std::max<int>(
          0, ((orgBuf[ch].at(xPosC, yPosC) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch])));
        assert(escapeValue.at(xPosC, yPosC) < (1 << (channelBitDepth + 1)));
        recBuf.at(xPosC, yPosC) =
          (((escapeValue.at(xPosC, yPosC) * g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch])
          >> invquantiserRightShift[ch];
        recBuf.at(xPosC, yPosC) = Pel(ClipBD<int>(recBuf.at(xPosC, yPosC), channelBitDepth));   // to be checked
#endif
      }
    }
  }
}

void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner, ComponentID compBegin, uint32_t numComp)
{
  CodingUnit &cu = *cs.getCU(partitioner.chType);
  const int channelBitDepth_L = cs.sps->getBitDepth(CHANNEL_TYPE_LUMA);
  const int channelBitDepth_C = cs.sps->getBitDepth(CHANNEL_TYPE_CHROMA);

  bool lossless        = (m_pcEncCfg->getCostMode() == COST_LOSSLESS_CODING && cs.slice->isLossless());
  int  pcmShiftRight_L = (channelBitDepth_L - PLT_ENCBITDEPTH);
  int  pcmShiftRight_C = (channelBitDepth_C - PLT_ENCBITDEPTH);
  if (lossless)
  {
    pcmShiftRight_L = 0;
    pcmShiftRight_C = 0;
  }
#if !INTRA_RM_SMALL_BLOCK_SIZE_CONSTRAINTS
  int maxPltSize = cu.isSepTree() ? MAXPLTSIZE_DUALTREE : MAXPLTSIZE;
#else
  int maxPltSize = CS::isDualITree(cs) ? MAXPLTSIZE_DUALTREE : MAXPLTSIZE;
#endif
  uint32_t height = cu.block(compBegin).height;
  uint32_t width = cu.block(compBegin).width;

  CPelBuf   orgBuf[3];
  for (int comp = compBegin; comp < (compBegin + numComp); comp++)
  {
    CompArea  area = cu.blocks[comp];
    if (m_pcEncCfg->getLmcs() && (cs.slice->getLmcsEnabledFlag() && m_pcReshape->getCTUFlag()))
    {
      orgBuf[comp] = cs.getPredBuf(area);
    }
    else
    {
      orgBuf[comp] = cs.getOrgBuf(area);
    }
  }

  TransformUnit &tu = *cs.getTU(partitioner.chType);
  QpParam cQP(tu, compBegin);
  int qp = cQP.Qp(true) - 12;
  qp = (qp < 0) ? 0 : ((qp > 56) ? 56 : qp);
  int errorLimit = g_paletteQuant[qp];
  if (lossless)
  {
    errorLimit = 0;
  }
  uint32_t totalSize = height*width;
  SortingElement *pelList = new SortingElement[totalSize];
  SortingElement  element;
  SortingElement *pelListSort = new SortingElement[MAXPLTSIZE + 1];
  uint32_t dictMaxSize = maxPltSize;
  uint32_t idx = 0;
  int last = -1;

  uint32_t scaleX = getComponentScaleX(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
  uint32_t scaleY = getComponentScaleY(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
  for (uint32_t y = 0; y < height; y++)
  {
    for (uint32_t x = 0; x < width; x++)
    {
      uint32_t org[3], pX, pY;
      for (int comp = compBegin; comp < (compBegin + numComp); comp++)
      {
        pX = (comp > 0 && compBegin == COMPONENT_Y) ? (x >> scaleX) : x;
        pY = (comp > 0 && compBegin == COMPONENT_Y) ? (y >> scaleY) : y;
        org[comp] = orgBuf[comp].at(pX, pY);
      }
      element.setAll(org, compBegin, numComp);

      ComponentID tmpCompBegin = compBegin;
      int tmpNumComp = numComp;
      if( cs.sps->getChromaFormatIdc() != CHROMA_444 &&
          numComp == 3 &&
         (x != ((x >> scaleX) << scaleX) || (y != ((y >> scaleY) << scaleY))) )
      {
        tmpCompBegin = COMPONENT_Y;
        tmpNumComp   = 1;
      }
      int besti = last, bestSAD = (last == -1) ? MAX_UINT : pelList[last].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
      if (lossless)
      {
        if (bestSAD)
        {
          for (int i = idx - 1; i >= 0; i--)
          {
            uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
            if (sad == 0)
            {
              bestSAD = sad;
              besti   = i;
              break;
            }
          }
        }
      }
      else
      {
        if (bestSAD)
        {
          for (int i = idx - 1; i >= 0; i--)
          {
            uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
            if (sad < bestSAD)
            {
              bestSAD = sad;
              besti   = i;
              if (!sad)
              {
                break;
              }
            }
          }
        }
      }
      if (besti >= 0 && pelList[besti].almostEqualData(element, errorLimit, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless))
      {
        pelList[besti].addElement(element, tmpCompBegin, tmpNumComp);
        last = besti;
      }
      else
      {
        pelList[idx].copyDataFrom(element, tmpCompBegin, tmpNumComp);
        for (int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++)
        {
          pelList[idx].setCnt(1, comp);
        }
        last = idx;
        idx++;
      }
    }
  }

  if( cs.sps->getChromaFormatIdc() != CHROMA_444 && numComp == 3 )
  {
    for( int i = 0; i < idx; i++ )
    {
      pelList[i].setCnt( pelList[i].getCnt(COMPONENT_Y) + (pelList[i].getCnt(COMPONENT_Cb) >> 2), MAX_NUM_COMPONENT);
    }
  }
  else
  {
    if( compBegin == 0 )
    {
      for( int i = 0; i < idx; i++ )
      {
        pelList[i].setCnt(pelList[i].getCnt(COMPONENT_Y), COMPONENT_Cb);
        pelList[i].setCnt(pelList[i].getCnt(COMPONENT_Y), COMPONENT_Cr);
        pelList[i].setCnt(pelList[i].getCnt(COMPONENT_Y), MAX_NUM_COMPONENT);
      }
    }
    else
    {
      for( int i = 0; i < idx; i++ )
      {
        pelList[i].setCnt(pelList[i].getCnt(COMPONENT_Cb), COMPONENT_Y);
        pelList[i].setCnt(pelList[i].getCnt(COMPONENT_Cb), MAX_NUM_COMPONENT);
      }
    }
  }

  for (int i = 0; i < dictMaxSize; i++)
  {
    pelListSort[i].setCnt(0, COMPONENT_Y);
    pelListSort[i].setCnt(0, COMPONENT_Cb);
    pelListSort[i].setCnt(0, COMPONENT_Cr);
    pelListSort[i].setCnt(0, MAX_NUM_COMPONENT);
    pelListSort[i].resetAll(compBegin, numComp);
  }

  //bubble sorting
  dictMaxSize = 1;
  for (int i = 0; i < idx; i++)
  {
    if( pelList[i].getCnt(MAX_NUM_COMPONENT) > pelListSort[dictMaxSize - 1].getCnt(MAX_NUM_COMPONENT) )
    {
      int j;
      for (j = dictMaxSize; j > 0; j--)
      {
        if (pelList[i].getCnt(MAX_NUM_COMPONENT) > pelListSort[j - 1].getCnt(MAX_NUM_COMPONENT))
        {
          pelListSort[j].copyAllFrom(pelListSort[j - 1], compBegin, numComp);
          dictMaxSize = std::min(dictMaxSize + 1, (uint32_t)maxPltSize);
        }
        else
        {
          break;
        }
      }
      pelListSort[j].copyAllFrom(pelList[i], compBegin, numComp);
    }
  }

  uint32_t paletteSize = 0;
  uint64_t numColorBits = 0;
  for (int comp = compBegin; comp < (compBegin + numComp); comp++)
  {
    numColorBits += (comp > 0) ? channelBitDepth_C : channelBitDepth_L;
  }
  const int plt_lambda_shift = (compBegin > 0) ? pcmShiftRight_C : pcmShiftRight_L;
  double    bitCost          = m_pcRdCost->getLambda() / (double) (1 << (2 * plt_lambda_shift)) * numColorBits;
  bool   reuseflag[MAXPLTPREDSIZE] = { false };
  int    run;
  double reuseflagCost;
  for (int i = 0; i < maxPltSize; i++)
  {
    if( pelListSort[i].getCnt(MAX_NUM_COMPONENT) )
    {
      ComponentID tmpCompBegin = compBegin;
      int tmpNumComp = numComp;
      if( cs.sps->getChromaFormatIdc() != CHROMA_444 && numComp == 3 && pelListSort[i].getCnt(COMPONENT_Cb) == 0 )
      {
        tmpCompBegin = COMPONENT_Y;
        tmpNumComp   = 1;
      }

      for( int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++ )
      {
        int half = pelListSort[i].getCnt(comp) >> 1;
        cu.curPLT[comp][paletteSize] = (pelListSort[i].getSumData(comp) + half) / pelListSort[i].getCnt(comp);
      }

      int best = -1;
      if( errorLimit )
      {
        double pal[MAX_NUM_COMPONENT], err = 0.0, bestCost = 0.0;
        for( int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++ )
        {
          pal[comp] = pelListSort[i].getSumData(comp) / (double)pelListSort[i].getCnt(comp);
          err = pal[comp] - cu.curPLT[comp][paletteSize];
          if( isChroma((ComponentID) comp) )
          {
            bestCost += (err * err * PLT_CHROMA_WEIGHTING) / (1 << (2 * pcmShiftRight_C)) * pelListSort[i].getCnt(comp);
          }
          else
          {
            bestCost += (err * err) / (1 << (2 * pcmShiftRight_L)) * pelListSort[i].getCnt(comp);
          }
        }
        bestCost += bitCost;

        for( int t = 0; t < cs.prevPLT.curPLTSize[compBegin]; t++ )
        {
          double cost = 0.0;
          for( int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++ )
          {
            err = pal[comp] - cs.prevPLT.curPLT[comp][t];
            if( isChroma((ComponentID) comp) )
            {
              cost += (err * err * PLT_CHROMA_WEIGHTING) / (1 << (2 * pcmShiftRight_C)) * pelListSort[i].getCnt(comp);
            }
            else
            {
              cost += (err * err) / (1 << (2 * pcmShiftRight_L)) * pelListSort[i].getCnt(comp);
            }
          }
          run = 0;
          for (int t2 = t; t2 >= 0; t2--)
          {
            if (!reuseflag[t2])
            {
              run++;
            }
            else
            {
              break;
            }
          }
          reuseflagCost = m_pcRdCost->getLambda() / (double)(1 << (2 * plt_lambda_shift)) * getEpExGolombNumBins(run ? run + 1 : run, 0);
          cost += reuseflagCost;

          if( cost < bestCost )
          {
            best = t;
            bestCost = cost;
          }
        }
        if( best != -1 )
        {
          for( int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++ )
          {
            cu.curPLT[comp][paletteSize] = cs.prevPLT.curPLT[comp][best];
          }
          reuseflag[best] = true;
        }
      }

      bool duplicate = false;
      if( pelListSort[i].getCnt(MAX_NUM_COMPONENT) == 1 && best == -1 )
      {
        duplicate = true;
      }
      else
      {
        for( int t = 0; t < paletteSize; t++ )
        {
          bool duplicateTmp = true;
          for( int comp = tmpCompBegin; comp < (tmpCompBegin + tmpNumComp); comp++ )
          {
            duplicateTmp = duplicateTmp && (cu.curPLT[comp][paletteSize] == cu.curPLT[comp][t]);
          }
          if( duplicateTmp )
          {
            duplicate = true;
            break;
          }
        }
      }
      if( !duplicate )
      {
        if( cs.sps->getChromaFormatIdc() != CHROMA_444 && numComp == 3 && pelListSort[i].getCnt(COMPONENT_Cb) == 0 )
        {
          if( best != -1 )
          {
            cu.curPLT[COMPONENT_Cb][paletteSize] = cs.prevPLT.curPLT[COMPONENT_Cb][best];
            cu.curPLT[COMPONENT_Cr][paletteSize] = cs.prevPLT.curPLT[COMPONENT_Cr][best];
          }
          else
          {
            cu.curPLT[COMPONENT_Cb][paletteSize] = 1 << (channelBitDepth_C - 1);
            cu.curPLT[COMPONENT_Cr][paletteSize] = 1 << (channelBitDepth_C - 1);
          }
        }
        paletteSize++;
      }
    }
    else
    {
      break;
    }
  }
  cu.curPLTSize[compBegin] = paletteSize;
#if !INTRA_RM_SMALL_BLOCK_SIZE_CONSTRAINTS
  if( cu.isLocalSepTree() )
  {
    cu.curPLTSize[COMPONENT_Y] = paletteSize;
  }
#endif
  delete[] pelList;
  delete[] pelListSort;
}
// -------------------------------------------------------------------------------------------------------------------
// Intra search
// -------------------------------------------------------------------------------------------------------------------

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

  if (bLuma)
  {
    bool isFirst = cu.ispMode ? subTuIdx == 0 : partitioner.currArea().lumaPos() == cs.area.lumaPos();

    // CU header
    if( isFirst )
    {
      if ((!cs.slice->isIntra() || cs.slice->getSPS()->getIBCFlag() || cs.slice->getSPS()->getPLTMode())
          && cu.Y().valid())
      {
        m_CABACEstimator->cu_skip_flag( cu );
        m_CABACEstimator->pred_mode   ( cu );
      }
#if ENABLE_DIMD
      m_CABACEstimator->cu_dimd_flag(cu);
#endif
      if (CU::isPLT(cu))
      {
        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->bdpcm_mode( cu, COMPONENT_Y );
      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->bdpcm_mode( cu, ComponentID(CHANNEL_TYPE_CHROMA) );
      m_CABACEstimator->intra_chroma_pred_mode( pu );
    }
  }
}

void IntraSearch::xEncSubdivCbfQT( CodingStructure &cs, Partitioner &partitioner, const bool &bLuma, const bool &bChroma, const int subTuIdx, const PartSplit ispType )
{
  const UnitArea &currArea = partitioner.currArea();
          int subTuCounter = subTuIdx;
  TransformUnit &currTU = *cs.getTU( currArea.blocks[partitioner.chType], partitioner.chType, subTuCounter );
  CodingUnit    &currCU = *currTU.cu;
  uint32_t currDepth           = partitioner.currTrDepth;

  const bool subdiv        = currTU.depth > currDepth;
  ComponentID compID = partitioner.chType == CHANNEL_TYPE_LUMA ? COMPONENT_Y : COMPONENT_Cb;

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

  if (bChroma)
  {
    const bool chromaCbfISP = currArea.blocks[COMPONENT_Cb].valid() && currCU.ispMode && !subdiv;
    if ( !currCU.ispMode || chromaCbfISP )
    {
      const uint32_t numberValidComponents = getNumberValidComponents(currArea.chromaFormat);
      const uint32_t cbfDepth              = (chromaCbfISP ? currDepth - 1 : currDepth);

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

        if (currDepth == 0 || TU::getCbfAtDepth(currTU, compID, currDepth - 1) || chromaCbfISP)
        {
          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],
                                     cbfDepth, prevCbf);
        }
      }
    }
  }

  if (subdiv)
  {
    if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
    {
      partitioner.splitCurrArea( TU_MAX_TR_SPLIT, cs );
    }
    else if( currCU.ispMode && isLuma( compID ) )
    {
      partitioner.splitCurrArea( ispType, cs );
    }
    else
    {
      THROW("Cannot perform an implicit split!");
    }

    do
    {
      xEncSubdivCbfQT( cs, partitioner, bLuma, bChroma, subTuCounter, ispType );
      subTuCounter += subTuCounter != -1 ? 1 : 0;
    } while( partitioner.nextPart( cs ) );

    partitioner.exitCurrSplit();
  }
  else
  {
    //===== Cbfs =====
    if (bLuma)
    {
      bool previousCbf       = false;
      bool lastCbfIsInferred = false;
      if( ispType != TU_NO_ISP )
      {
        bool rootCbfSoFar = false;
        uint32_t nTus = currCU.ispMode == HOR_INTRA_SUBPARTITIONS ? currCU.lheight() >> floorLog2(currTU.lheight()) : currCU.lwidth() >> floorLog2(currTU.lwidth());
        if( subTuCounter == nTus - 1 )
        {
          TransformUnit* tuPointer = currCU.firstTU;
          for( int tuIdx = 0; tuIdx < nTus - 1; tuIdx++ )
          {
            rootCbfSoFar |= TU::getCbfAtDepth( *tuPointer, COMPONENT_Y, currDepth );
            tuPointer = tuPointer->next;
          }
          if( !rootCbfSoFar )
          {
            lastCbfIsInferred = true;
          }
        }
        if( !lastCbfIsInferred )
        {
          previousCbf = TU::getPrevTuCbfAtDepth( currTU, COMPONENT_Y, partitioner.currTrDepth );
        }
      }
      if( !lastCbfIsInferred )
      {
        m_CABACEstimator->cbf_comp( cs, TU::getCbfAtDepth( currTU, COMPONENT_Y, currDepth ), currTU.Y(), currTU.depth, previousCbf, currCU.ispMode );
      }
    }
  }
}

void IntraSearch::xEncCoeffQT( CodingStructure &cs, Partitioner &partitioner, const ComponentID compID, const int subTuIdx, const PartSplit ispType, CUCtx* cuCtx )
{
  const UnitArea &currArea  = partitioner.currArea();

       int subTuCounter     = subTuIdx;
  TransformUnit &currTU     = *cs.getTU( currArea.blocks[partitioner.chType], partitioner.chType, subTuIdx );
  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);