H.266/VVC代码学习7：VTM4.0帧内色度预测代码（estIntraPredChromaQT）

关于帧内亮度预测的过程在之前H.266/VVC代码学习1：帧内预测框架中已简要讲述。重要的操作如下

0.初始化；H.266/VVC代码学习22：帧内预测的初始化（initIntraPatternChType）
1.获取色度预测列表getIntraChromaCandModes，并对pu和tu初始化；
2.其中把5种模式使用SATD运算进行色度预测列表的第一轮精简，去掉2个；（DC、Ver、Hor、LM_L、LM_T）
3.将剩下的6种模式进行遍历，都进行后续变换编码等操作，得到代价；见H.266/VVC代码学习8：xRecurIntraChromaCodingQT函数
4.比较代价，选出最优模式，更新代价

关于色度预测的模式，在VTM4.0中有角度预测、LM和DM三类共71种，如需查阅代码请点击：
H.266/VVC代码学习20：角度预测入口——predIntraAng
H.266/VVC代码学习15：LM模式——predIntraChromaLM
H.266/VVC代码学习14：DM模式——getFinalIntraMode

下面vtm4.0中代码部分的是详细内容，可结合注释理解：

#if JVET_M0102_INTRA_SUBPARTITIONS
void IntraSearch::estIntraPredChromaQT( CodingUnit &cu, Partitioner &partitioner, const double maxCostAllowed )
#else
void IntraSearch::estIntraPredChromaQT(CodingUnit &cu, Partitioner &partitioner)
#endif
{
	/******************************************** 初始化 ***************************************************/
  const ChromaFormat format   = cu.chromaFormat;
  const uint32_t    numberValidComponents = getNumberValidComponents(format);
  CodingStructure &cs = *cu.cs;
  const TempCtx ctxStart  ( m_CtxCache, m_CABACEstimator->getCtx() );

  cs.setDecomp( cs.area.Cb(), false );

#if JVET_M0102_INTRA_SUBPARTITIONS
  double    bestCostSoFar = maxCostAllowed;
  bool      lumaUsesISP   = !CS::isDualITree( *cu.cs ) && cu.ispMode;
  PartSplit ispType       = lumaUsesISP ? CU::getISPType( cu, COMPONENT_Y ) : TU_NO_ISP;
  CHECK( cu.ispMode && bestCostSoFar < 0, "bestCostSoFar must be positive!" );
#endif
  /******************************************* 对pu的处理 ***************************************************/
  auto &pu = *cu.firstPU;

  {
	  /************************* 处理列表 ********************/
    uint32_t       uiBestMode = 0;
    Distortion uiBestDist = 0;
    double     dBestCost = MAX_DOUBLE;

    //----- init mode list 初始化模式列表（列出空表） ----
    {
      uint32_t  uiMinMode = 0;
      uint32_t  uiMaxMode = NUM_CHROMA_MODE;

      //----- check chroma modes 获取列表的值（填满表）-----
      uint32_t chromaCandModes[ NUM_CHROMA_MODE ];
      PU::getIntraChromaCandModes( pu, chromaCandModes );

      // create a temporary CS 创建一个临时CS
      CodingStructure &saveCS = *m_pSaveCS[0];
      saveCS.pcv      = cs.pcv;
      saveCS.picture  = cs.picture;
      saveCS.area.repositionTo( cs.area );
      saveCS.clearTUs();

#if JVET_M0102_INTRA_SUBPARTITIONS
      if( !CS::isDualITree( cs ) && cu.ispMode )
      {
        saveCS.clearCUs();
        saveCS.clearPUs();
      }
#endif

	  /*********************************** 将TU保存 **********************************/
	  //如果是色度单独划分的话，将色度进行划分，并把TU存好
      if( CS::isDualITree( cs ) )
      {
        if( partitioner.canSplit( TU_MAX_TR_SPLIT, cs ) )
        {
          partitioner.splitCurrArea( TU_MAX_TR_SPLIT, cs );

          do
          {
            cs.addTU( CS::getArea( cs, partitioner.currArea(), partitioner.chType ), partitioner.chType ).depth = partitioner.currTrDepth;
          } while( partitioner.nextPart( cs ) );

          partitioner.exitCurrSplit();
        }
        else
        cs.addTU( CS::getArea( cs, partitioner.currArea(), partitioner.chType ), partitioner.chType );
      }

      std::vector<TransformUnit*> orgTUs;

#if JVET_M0102_INTRA_SUBPARTITIONS
      if( lumaUsesISP )
      {
        CodingUnit& auxCU = saveCS.addCU( cu, partitioner.chType );
        auxCU.ispMode = cu.ispMode;
        saveCS.sps = cu.cs->sps;
        saveCS.addPU( *cu.firstPU, partitioner.chType );
      }
#endif


      // create a store for the TUs 保存tu进orgTUs中以备后续处理
      for( const auto &ptu : cs.tus )
      {
        // for split TUs in HEVC, add the TUs without Chroma parts for correct setting of Cbfs
#if JVET_M0102_INTRA_SUBPARTITIONS
        if( lumaUsesISP || pu.contains( *ptu, CHANNEL_TYPE_CHROMA ) )
#else
        if( pu.contains( *ptu, CHANNEL_TYPE_CHROMA ) )
#endif
        {
          saveCS.addTU( *ptu, partitioner.chType );
          orgTUs.push_back( ptu );
        }
      }
#if JVET_M0102_INTRA_SUBPARTITIONS
      if( lumaUsesISP )
      {
        saveCS.clearCUs();
      }
#endif
	  /*************进行第一次选择，按DC、Ver、Hor、LM_L、LM_T进行预测，按SATD排序，去掉两个SATD最大的模式**************/
      // SATD pre-selecting. 
      int satdModeList[NUM_CHROMA_MODE];
      int64_t satdSortedCost[NUM_CHROMA_MODE];
      for (int i = 0; i < NUM_CHROMA_MODE; i++)
      {//对于未通过SATD预先选择的模式，默认情况下执行RDO，因此将初始值设置为0。
        satdSortedCost[i] = 0; // for the mode not pre-select by SATD, do RDO by default, so set the initial value 0.
        satdModeList[i] = 0;
      }
      bool modeIsEnable[NUM_INTRA_MODE + 1]; // use intra mode idx to check whether enable
      for (int i = 0; i < NUM_INTRA_MODE + 1; i++)
      {
        modeIsEnable[i] = 1;
      }

      DistParam distParam;
      const bool useHadamard = true;
      pu.intraDir[1] = MDLM_L_IDX; // temporary assigned, just to indicate this is a MDLM mode. for luma down-sampling operation.

      initIntraPatternChType(cu, pu.Cb());//参考像素滤波
      initIntraPatternChType(cu, pu.Cr());
      xGetLumaRecPixels(pu, pu.Cb());//获取亮度重建值

      for (int idx = uiMinMode; idx <= uiMaxMode - 1; idx++)
      {
        int mode = chromaCandModes[idx];
        satdModeList[idx] = mode;
        if (PU::isLMCMode(mode) && !PU::isLMCModeEnabled(pu, mode))
        {
          continue;
        }
        if ((mode == LM_CHROMA_IDX) || (mode == PLANAR_IDX) || (mode == DM_CHROMA_IDX)) // only pre-check regular modes and MDLM modes, not including DM ,Planar, and LM
        {
          continue;
        }
        pu.intraDir[1] = mode; // temporary assigned, for SATD checking.

        int64_t sad = 0;
        CodingStructure& cs = *(pu.cs);

        CompArea areaCb = pu.Cb();
        PelBuf orgCb = cs.getOrgBuf(areaCb);
        PelBuf predCb = cs.getPredBuf(areaCb);

        m_pcRdCost->setDistParam(distParam, orgCb, predCb, pu.cs->sps->getBitDepth(CHANNEL_TYPE_CHROMA), COMPONENT_Cb, useHadamard);
        distParam.applyWeight = false;

        if (PU::isLMCMode(mode))
        {
          predIntraChromaLM(COMPONENT_Cb, predCb, pu, areaCb, mode);
        }
        else
        {
          predIntraAng(COMPONENT_Cb, predCb, pu, false);
        }

        sad += distParam.distFunc(distParam);

        CompArea areaCr = pu.Cr();
        PelBuf orgCr = cs.getOrgBuf(areaCr);
        PelBuf predCr = cs.getPredBuf(areaCr);

        m_pcRdCost->setDistParam(distParam, orgCr, predCr, pu.cs->sps->getBitDepth(CHANNEL_TYPE_CHROMA), COMPONENT_Cr, useHadamard);
        distParam.applyWeight = false;

        if (PU::isLMCMode(mode))
        {
          predIntraChromaLM(COMPONENT_Cr, predCr, pu, areaCr, mode);
        }
        else
        {
          predIntraAng(COMPONENT_Cr, predCr, pu, false);
        }
        sad += distParam.distFunc(distParam);
        satdSortedCost[idx] = sad;
      }
      // sort the mode based on the cost from small to large. 根据SATD从小到大排序模式。
      int tempIdx = 0;
      int64_t tempCost = 0;
      for (int i = uiMinMode; i <= uiMaxMode - 1; i++)
      {
        for (int j = i + 1; j <= uiMaxMode - 1; j++)
        {
          if (satdSortedCost[j] < satdSortedCost[i])
          {
            tempIdx = satdModeList[i];
            satdModeList[i] = satdModeList[j];
            satdModeList[j] = tempIdx;

            tempCost = satdSortedCost[i];
            satdSortedCost[i] = satdSortedCost[j];
            satdSortedCost[j] = tempCost;

          }
        }
      }
      int reducedModeNumber = 2; // reduce the number of chroma modes 把SATD最大的两个排除出去
      for (int i = 0; i < reducedModeNumber; i++)
      {
        modeIsEnable[satdModeList[uiMaxMode - 1 - i]] = 0; // disable the last reducedModeNumber modes
      }

      // save the dist
      Distortion baseDist = cs.dist;

	  /***********进行第二次选择，把SATD小的三种和未比较的三种共六种模式，编码后计算RDcost *******************/
      for (uint32_t uiMode = uiMinMode; uiMode < uiMaxMode; uiMode++)
      {
        const int chromaIntraMode = chromaCandModes[uiMode];
        if( PU::isLMCMode( chromaIntraMode ) && ! PU::isLMCModeEnabled( pu, chromaIntraMode ) )
        {
          continue;
        }
        if (!modeIsEnable[chromaIntraMode] && PU::isLMCModeEnabled(pu, chromaIntraMode)) // when CCLM is disable, then MDLM is disable. not use satd checking
        {
          continue;
        }
        cs.setDecomp( pu.Cb(), false );
        cs.dist = baseDist;
        //----- restore context models -----
        m_CABACEstimator->getCtx() = ctxStart;
		
        //----- chroma coding -----
        pu.intraDir[1] = chromaIntraMode;

#if JVET_M0102_INTRA_SUBPARTITIONS
        xRecurIntraChromaCodingQT( cs, partitioner, bestCostSoFar, ispType );
        if( lumaUsesISP && cs.dist == MAX_UINT )
        {
          continue;
        }
#else
        xRecurIntraChromaCodingQT( cs, partitioner );
#endif

        if (cs.pps->getUseTransformSkip())
        {
          m_CABACEstimator->getCtx() = ctxStart;
        }

#if JVET_M0102_INTRA_SUBPARTITIONS
        uint64_t fracBits   = xGetIntraFracBitsQT( cs, partitioner, false, true, -1, ispType );


#else
        uint64_t fracBits   = xGetIntraFracBitsQT( cs, partitioner, false, true );
#endif
        Distortion uiDist = cs.dist;
        double    dCost   = m_pcRdCost->calcRdCost( fracBits, uiDist - baseDist );

		/********************************* 获取最佳色度模式 ***********************************/
        //----- compare -----
        if( dCost < dBestCost )
        {
#if JVET_M0102_INTRA_SUBPARTITIONS
          if( lumaUsesISP && dCost < bestCostSoFar )
          {
            bestCostSoFar = dCost;
          }
#endif
          for( uint32_t i = getFirstComponentOfChannel( CHANNEL_TYPE_CHROMA ); i < numberValidComponents; i++ )
          {
            const CompArea &area = pu.blocks[i];

            saveCS.getRecoBuf     ( area ).copyFrom( cs.getRecoBuf   ( area ) );
#if KEEP_PRED_AND_RESI_SIGNALS
            saveCS.getPredBuf     ( area ).copyFrom( cs.getPredBuf   ( area ) );
            saveCS.getResiBuf     ( area ).copyFrom( cs.getResiBuf   ( area ) );
#endif
#if JVET_M0427_INLOOP_RESHAPER
            saveCS.getPredBuf     ( area ).copyFrom( cs.getPredBuf   (area ) );
            cs.picture->getPredBuf( area ).copyFrom( cs.getPredBuf   (area ) );
#endif
            cs.picture->getRecoBuf( area ).copyFrom( cs.getRecoBuf( area ) );

            for( uint32_t j = 0; j < saveCS.tus.size(); j++ )
            {
              saveCS.tus[j]->copyComponentFrom( *orgTUs[j], area.compID );
            }
          }

          dBestCost  = dCost;
          uiBestDist = uiDist;
          uiBestMode = chromaIntraMode;
        }

      }

      for( uint32_t i = getFirstComponentOfChannel( CHANNEL_TYPE_CHROMA ); i < numberValidComponents; i++ )
      {
        const CompArea &area = pu.blocks[i];

        cs.getRecoBuf         ( area ).copyFrom( saveCS.getRecoBuf( area ) );
#if KEEP_PRED_AND_RESI_SIGNALS
        cs.getPredBuf         ( area ).copyFrom( saveCS.getPredBuf( area ) );
        cs.getResiBuf         ( area ).copyFrom( saveCS.getResiBuf( area ) );
#endif
#if JVET_M0427_INLOOP_RESHAPER
        cs.getPredBuf         ( area ).copyFrom( saveCS.getPredBuf( area ) );
        cs.picture->getPredBuf( area ).copyFrom( cs.getPredBuf    ( area ) );
#endif

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

        for( uint32_t j = 0; j < saveCS.tus.size(); j++ )
        {
          orgTUs[ j ]->copyComponentFrom( *saveCS.tus[ j ], area.compID );
        }
      }
    }

    pu.intraDir[1] = uiBestMode;
    cs.dist        = uiBestDist;
#if PRINT
	//printf("xGetIntraFracBitsQT (%d,%d) %d %d : %lld\n", cs.area.blocks[0].x, cs.area.blocks[0].y, cs.area.blocks[0].width, cs.area.blocks[0].height, cs.fracBits);
#endif 
  }

  /******************************** 重建上下文模型 ************************************/
  //----- restore context models -----
  m_CABACEstimator->getCtx() = ctxStart;
#if JVET_M0102_INTRA_SUBPARTITIONS
  if( lumaUsesISP && bestCostSoFar >= maxCostAllowed )
  {
    cu.ispMode = 0;
  }
#endif
}