【HEVC学习与研究】29、解码第一个Coding Quadtree结构（1）

在开始研究这部分代码之前，重新回顾一下理论部分。

众所周知，H.264及其之前的编码标准中，VCL层的核心结构称作“宏块”（MacroBlock, MB），大小为16×16像素。一个宏块里面包含着一个16×16分辨率的亮度采样矩阵和两个8×8分辨率的色度采样矩阵。相应的，在HEVC中，编码所采用的块结构为树形编码单元（Coding Tree Unit, CTU），每个CTU的大小由编码器决定，并且通常要大于宏块的16×16大小。比CTU更小的单位称为树形编码块（Coding Tree Block, CTB），一个CTU由一个亮度信号CTB及其相应的色度CTB，以及其他一些语法元素构成。一个正方形的亮度CTB的边长（以像素为单位）可能取16、32或者64个像素，CTB越大则压缩效率越高。HEVC标准支持将一个CTB按照四叉树的方法分割为多个更小的块结构。

采用了四叉树分割的方法后，这个树结构保存了亮度和色度编码单元（Coding Unit, CU）的位置和大小。这个四叉树结构的根与当前的这个CTU对应，也就是说，一个亮度CU最大可能的大小就是亮度CTB的大小。通常，一个CU的组成包括一个亮度编码块（Coding Block, CB），一个色度CB以及其他相应的语法元素。一个CTB可能包含一个CU，也可能被分割成多个CU，每一个CU进一步分割成预测单元（Prediction Unit, PU）和树结构的变换单元（Transform Unit, TU）。

判断采用帧内预测还是帧间预测在CU层完成。PU的分割结构就以CU为根，根据基本的预测模式判定，亮度和色度CB被进一步分割，由亮度和色度预测块(Prediction Block, PB)中获取预测值。每一个PB的大小可能由64×64到4×4不等。预测的残差信号由TU进行编码。同PU类似，TU树形结构的根节点也是CU，一个亮度CB的残差信号可能由一个完整的变换块（Transform Block, TB）表示，也可能树形分割成多个子TB。对4×4、8×8、16×16和32×32的正方形TB块，HEVC依然采用整数变换；当TB大小为4×4，且采用帧内编码的残差信号则采用了一种类离散正弦变换的方法。

标准文档的7.3.8.4节给出了Coding quadtree的语法结构：

该结构中的第一个语法元素split_cu_flag[x0][y0]表明这个Coding quadtree是否继续四等分，下标x0和y0表示当前拟作为cu的像素块左上角像素相对于该帧左上角的坐标。从该结构的定义来看，Coding quadtree采用了递归的结构，只要split_cu_flag[x0][y0]为1，则这个Coding quadtree就会继续分割下去，直到split_cu_flag[x0][y0]为0时，这个Coding quadtree进一步处理为一个coding unit。

回到代码中。实现解析Coding quadtree的函数由Void TDecCu::xDecodeCU( TComDataCU* pcCU, UInt uiAbsPartIdx, UInt uiDepth, UInt& ruiIsLast)实现：

Void TDecCu::xDecodeCU( TComDataCU* pcCU, UInt uiAbsPartIdx, UInt uiDepth, UInt& ruiIsLast)
{
	TComPic* pcPic = pcCU->getPic();
	UInt uiCurNumParts    = pcPic->getNumPartInCU() >> (uiDepth<<1);
	UInt uiQNumParts      = uiCurNumParts>>2;

	Bool bBoundary = false;
	UInt uiLPelX   = pcCU->getCUPelX() + g_auiRasterToPelX[ g_auiZscanToRaster[uiAbsPartIdx] ];
	UInt uiRPelX   = uiLPelX + (g_uiMaxCUWidth>>uiDepth)  - 1;
	UInt uiTPelY   = pcCU->getCUPelY() + g_auiRasterToPelY[ g_auiZscanToRaster[uiAbsPartIdx] ];
	UInt uiBPelY   = uiTPelY + (g_uiMaxCUHeight>>uiDepth) - 1;

	TComSlice * pcSlice = pcCU->getPic()->getSlice(pcCU->getPic()->getCurrSliceIdx());
	Bool bStartInCU = pcCU->getSCUAddr()+uiAbsPartIdx+uiCurNumParts>pcSlice->getSliceSegmentCurStartCUAddr()&&pcCU->getSCUAddr()+uiAbsPartIdx<pcSlice->getSliceSegmentCurStartCUAddr();
	if((!bStartInCU) && ( uiRPelX < pcSlice->getSPS()->getPicWidthInLumaSamples() ) && ( uiBPelY < pcSlice->getSPS()->getPicHeightInLumaSamples() ) )
	{
		m_pcEntropyDecoder->decodeSplitFlag( pcCU, uiAbsPartIdx, uiDepth );
	}
	else
	{
		bBoundary = true;
	}

	if( ( ( uiDepth < pcCU->getDepth( uiAbsPartIdx ) ) && ( uiDepth < g_uiMaxCUDepth - g_uiAddCUDepth ) ) || bBoundary )
	{
		UInt uiIdx = uiAbsPartIdx;
		if( (g_uiMaxCUWidth>>uiDepth) == pcCU->getSlice()->getPPS()->getMinCuDQPSize() && pcCU->getSlice()->getPPS()->getUseDQP())
		{
			setdQPFlag(true);
			pcCU->setQPSubParts( pcCU->getRefQP(uiAbsPartIdx), uiAbsPartIdx, uiDepth ); // set QP to default QP
		}

		for ( UInt uiPartUnitIdx = 0; uiPartUnitIdx < 4; uiPartUnitIdx++ )
		{
			uiLPelX   = pcCU->getCUPelX() + g_auiRasterToPelX[ g_auiZscanToRaster[uiIdx] ];
			uiTPelY   = pcCU->getCUPelY() + g_auiRasterToPelY[ g_auiZscanToRaster[uiIdx] ];

			Bool bSubInSlice = pcCU->getSCUAddr()+uiIdx+uiQNumParts>pcSlice->getSliceSegmentCurStartCUAddr();
			if ( bSubInSlice )
			{
				if ( !ruiIsLast && ( uiLPelX < pcCU->getSlice()->getSPS()->getPicWidthInLumaSamples() ) && ( uiTPelY < pcCU->getSlice()->getSPS()->getPicHeightInLumaSamples() ) )
				{
					xDecodeCU( pcCU, uiIdx, uiDepth+1, ruiIsLast );
				}
				else
				{
					pcCU->setOutsideCUPart( uiIdx, uiDepth+1 );
				}
			}

			uiIdx += uiQNumParts;
		}
		if( (g_uiMaxCUWidth>>uiDepth) == pcCU->getSlice()->getPPS()->getMinCuDQPSize() && pcCU->getSlice()->getPPS()->getUseDQP())
		{
			if ( getdQPFlag() )
			{
				UInt uiQPSrcPartIdx;
				if ( pcPic->getCU( pcCU->getAddr() )->getSliceSegmentStartCU(uiAbsPartIdx) != pcSlice->getSliceSegmentCurStartCUAddr() )
				{
					uiQPSrcPartIdx = pcSlice->getSliceSegmentCurStartCUAddr() % pcPic->getNumPartInCU();
				}
				else
				{
					uiQPSrcPartIdx = uiAbsPartIdx;
				}
				pcCU->setQPSubParts( pcCU->getRefQP( uiQPSrcPartIdx ), uiAbsPartIdx, uiDepth ); // set QP to default QP
			}
		}
		return;
	}

	if( (g_uiMaxCUWidth>>uiDepth) >= pcCU->getSlice()->getPPS()->getMinCuDQPSize() && pcCU->getSlice()->getPPS()->getUseDQP())
	{
		setdQPFlag(true);
		pcCU->setQPSubParts( pcCU->getRefQP(uiAbsPartIdx), uiAbsPartIdx, uiDepth ); // set QP to default QP
	}

	if (pcCU->getSlice()->getPPS()->getTransquantBypassEnableFlag())
	{
		m_pcEntropyDecoder->decodeCUTransquantBypassFlag( pcCU, uiAbsPartIdx, uiDepth );
	}

	// decode CU mode and the partition size
	if( !pcCU->getSlice()->isIntra())
	{
		m_pcEntropyDecoder->decodeSkipFlag( pcCU, uiAbsPartIdx, uiDepth );
	}

	if( pcCU->isSkipped(uiAbsPartIdx) )
	{
		m_ppcCU[uiDepth]->copyInterPredInfoFrom( pcCU, uiAbsPartIdx, REF_PIC_LIST_0 );
		m_ppcCU[uiDepth]->copyInterPredInfoFrom( pcCU, uiAbsPartIdx, REF_PIC_LIST_1 );
		TComMvField cMvFieldNeighbours[MRG_MAX_NUM_CANDS << 1]; // double length for mv of both lists
		UChar uhInterDirNeighbours[MRG_MAX_NUM_CANDS];
		Int numValidMergeCand = 0;
		for( UInt ui = 0; ui < m_ppcCU[uiDepth]->getSlice()->getMaxNumMergeCand(); ++ui )
		{
			uhInterDirNeighbours[ui] = 0;
		}
		m_pcEntropyDecoder->decodeMergeIndex( pcCU, 0, uiAbsPartIdx, uiDepth );
		UInt uiMergeIndex = pcCU->getMergeIndex(uiAbsPartIdx);
		m_ppcCU[uiDepth]->getInterMergeCandidates( 0, 0, cMvFieldNeighbours, uhInterDirNeighbours, numValidMergeCand, uiMergeIndex );
		pcCU->setInterDirSubParts( uhInterDirNeighbours[uiMergeIndex], uiAbsPartIdx, 0, uiDepth );

		TComMv cTmpMv( 0, 0 );
		for ( UInt uiRefListIdx = 0; uiRefListIdx < 2; uiRefListIdx++ )
		{        
			if ( pcCU->getSlice()->getNumRefIdx( RefPicList( uiRefListIdx ) ) > 0 )
			{
				pcCU->setMVPIdxSubParts( 0, RefPicList( uiRefListIdx ), uiAbsPartIdx, 0, uiDepth);
				pcCU->setMVPNumSubParts( 0, RefPicList( uiRefListIdx ), uiAbsPartIdx, 0, uiDepth);
				pcCU->getCUMvField( RefPicList( uiRefListIdx ) )->setAllMvd( cTmpMv, SIZE_2Nx2N, uiAbsPartIdx, uiDepth );
				pcCU->getCUMvField( RefPicList( uiRefListIdx ) )->setAllMvField( cMvFieldNeighbours[ 2*uiMergeIndex + uiRefListIdx ], SIZE_2Nx2N, uiAbsPartIdx, uiDepth );
			}
		}
		xFinishDecodeCU( pcCU, uiAbsPartIdx, uiDepth, ruiIsLast );
		return;
	}

	m_pcEntropyDecoder->decodePredMode( pcCU, uiAbsPartIdx, uiDepth );
	m_pcEntropyDecoder->decodePartSize( pcCU, uiAbsPartIdx, uiDepth );

	if (pcCU->isIntra( uiAbsPartIdx ) && pcCU->getPartitionSize( uiAbsPartIdx ) == SIZE_2Nx2N )
	{
		m_pcEntropyDecoder->decodeIPCMInfo( pcCU, uiAbsPartIdx, uiDepth );

		if(pcCU->getIPCMFlag(uiAbsPartIdx))
		{
			xFinishDecodeCU( pcCU, uiAbsPartIdx, uiDepth, ruiIsLast );
			return;
		}
	}

	UInt uiCurrWidth      = pcCU->getWidth ( uiAbsPartIdx );
	UInt uiCurrHeight     = pcCU->getHeight( uiAbsPartIdx );

	// prediction mode ( Intra : direction mode, Inter : Mv, reference idx )
	m_pcEntropyDecoder->decodePredInfo( pcCU, uiAbsPartIdx, uiDepth, m_ppcCU[uiDepth]);

	// Coefficient decoding
	Bool bCodeDQP = getdQPFlag();
	m_pcEntropyDecoder->decodeCoeff( pcCU, uiAbsPartIdx, uiDepth, uiCurrWidth, uiCurrHeight, bCodeDQP );
	setdQPFlag( bCodeDQP );
	xFinishDecodeCU( pcCU, uiAbsPartIdx, uiDepth, ruiIsLast );
}

TDecEntropy::decodeSplitFlag   ( TComDataCU* pcCU, UInt uiAbsPartIdx, UInt uiDepth )函数调用m_pcEntropyDecoderIf->parseSplitFlag( pcCU, uiAbsPartIdx, uiDepth )来解析 split_cu_flag[x0][y0]的值。在这个过程中，由于规定了最大的CUDepth为4，所以只进行了4次分割便返回。其后，将处理coding block数据。