解析完成条带头之后下一步的工作是解析条带数据slice_segment_data。slice_segment_data数据主要由一个个的Coding_Tree_Unit(CTU)组成。每一个CTU的结构如下所示:
coding_tree_unit( )
{
xCtb = ( CtbAddrInRs % PicWidthInCtbsY ) << CtbLog2SizeY
yCtb = ( CtbAddrInRs / PicWidthInCtbsY ) << CtbLog2SizeY
if( slice_sao_luma_flag | | slice_sao_chroma_flag )
sao( xCtb >> CtbLog2SizeY, yCtb >> CtbLog2SizeY )
coding_quadtree( xCtb, yCtb, CtbLog2SizeY, 0 )
}
可以看出,每一个CTU的coding_quadtree部分之前是语法结构sao。那么现在来研究一下如何从码流中解码出sao,这里着重讨论cabac而非sao的意义。
对sao的解析部分在TDecSlice::decompressSlice()中。在真正开始解码之前,TDecSlice::decompressSlice()调用了m_pcEntropyDecoderIf->resetEntropy(p)函数进行熵解码器的一些初始化操作,如为不同语法元素进行initBuffer。在最后,调用了m_pcTDecBinIf->start()函数。该函数读出了码流中的前两个字节组成一个UInt16的数字赋给m_uiValue。
TDecBinCABAC::start()
{
assert( m_pcTComBitstream->getNumBitsUntilByteAligned() == 0 );
m_uiRange = 510;
m_bitsNeeded = -8;
m_uiValue = (m_pcTComBitstream->readByte() << 8);
m_uiValue |= m_pcTComBitstream->readByte();
}
以下是TDecSlice::decompressSlice()中sao的解析部分:
if ( pcSlice->getSPS()->getUseSAO() && (pcSlice->getSaoEnabledFlag()||pcSlice->getSaoEnabledFlagChroma()) )
{
SAOParam *saoParam = rpcPic->getPicSym()->getSaoParam();
saoParam->bSaoFlag[0] = pcSlice->getSaoEnabledFlag();
if (iCUAddr == iStartCUAddr)
{
saoParam->bSaoFlag[1] = pcSlice->getSaoEnabledFlagChroma();
}
Int numCuInWidth = saoParam->numCuInWidth;
Int cuAddrInSlice = iCUAddr - rpcPic->getPicSym()->getCUOrderMap(pcSlice->getSliceCurStartCUAddr()/rpcPic->getNumPartInCU());
Int cuAddrUpInSlice = cuAddrInSlice - numCuInWidth;
Int rx = iCUAddr % numCuInWidth;
Int ry = iCUAddr / numCuInWidth;
Int allowMergeLeft = 1;
Int allowMergeUp = 1;
if (rx!=0)
{
if (rpcPic->getPicSym()->getTileIdxMap(iCUAddr-1) != rpcPic->getPicSym()->getTileIdxMap(iCUAddr))
{
allowMergeLeft = 0;
}
}
if (ry!=0)
{
if (rpcPic->getPicSym()->getTileIdxMap(iCUAddr-numCuInWidth) != rpcPic->getPicSym()->getTileIdxMap(iCUAddr))
{
allowMergeUp = 0;
}
}
pcSbacDecoder->parseSaoOneLcuInterleaving(rx, ry, saoParam,pcCU, cuAddrInSlice, cuAddrUpInSlice, allowMergeLeft, allowMergeUp);
}
else if ( pcSlice->getSPS()->getUseSAO() )
{
Int addr = pcCU->getAddr();
SAOParam *saoParam = rpcPic->getPicSym()->getSaoParam();
for (Int cIdx=0; cIdx<3; cIdx++)
{
SaoLcuParam *saoLcuParam = &(saoParam->saoLcuParam[cIdx][addr]);
if ( ((cIdx == 0) && !pcSlice->getSaoEnabledFlag()) || ((cIdx == 1 || cIdx == 2) && !pcSlice->getSaoEnabledFlagChroma()))
{
saoLcuParam->mergeUpFlag = 0;
saoLcuParam->mergeLeftFlag = 0;
saoLcuParam->subTypeIdx = 0;
saoLcuParam->typeIdx = -1;
saoLcuParam->offset[0] = 0;
saoLcuParam->offset[1] = 0;
saoLcuParam->offset[2] = 0;
saoLcuParam->offset[3] = 0;
}
}
}
具体的解析过程在pcSbacDecoder->parseSaoOneLcuInterleaving()中实现。实现如下:
Void TDecSbac::parseSaoOneLcuInterleaving(Int rx, Int ry, SAOParam* pSaoParam, TComDataCU* pcCU, Int iCUAddrInSlice, Int iCUAddrUpInSlice, Int allowMergeLeft, Int allowMergeUp)
{
Int iAddr = pcCU->getAddr();
UInt uiSymbol;
for (Int iCompIdx=0; iCompIdx<3; iCompIdx++)
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeUpFlag = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeLeftFlag = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].subTypeIdx = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].typeIdx = -1;
pSaoParam->saoLcuParam[iCompIdx][iAddr].offset[0] = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].offset[1] = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].offset[2] = 0;
pSaoParam->saoLcuParam[iCompIdx][iAddr].offset[3] = 0;
}
if (pSaoParam->bSaoFlag[0] || pSaoParam->bSaoFlag[1] )
{
if (rx>0 && iCUAddrInSlice!=0 && allowMergeLeft)
{
parseSaoMerge(uiSymbol);
pSaoParam->saoLcuParam[0][iAddr].mergeLeftFlag = (Bool)uiSymbol;
}
if (pSaoParam->saoLcuParam[0][iAddr].mergeLeftFlag==0)
{
if ((ry > 0) && (iCUAddrUpInSlice>=0) && allowMergeUp)
{
parseSaoMerge(uiSymbol);
pSaoParam->saoLcuParam[0][iAddr].mergeUpFlag = (Bool)uiSymbol;
}
}
}
for (Int iCompIdx=0; iCompIdx<3; iCompIdx++)
{
if ((iCompIdx == 0 && pSaoParam->bSaoFlag[0]) || (iCompIdx > 0 && pSaoParam->bSaoFlag[1]) )
{
if (rx>0 && iCUAddrInSlice!=0 && allowMergeLeft)
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeLeftFlag = pSaoParam->saoLcuParam[0][iAddr].mergeLeftFlag;
}
else
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeLeftFlag = 0;
}
if (pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeLeftFlag==0)
{
if ((ry > 0) && (iCUAddrUpInSlice>=0) && allowMergeUp)
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeUpFlag = pSaoParam->saoLcuParam[0][iAddr].mergeUpFlag;
}
else
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeUpFlag = 0;
}
if (!pSaoParam->saoLcuParam[iCompIdx][iAddr].mergeUpFlag)
{
pSaoParam->saoLcuParam[2][iAddr].typeIdx = pSaoParam->saoLcuParam[1][iAddr].typeIdx;
parseSaoOffset(&(pSaoParam->saoLcuParam[iCompIdx][iAddr]), iCompIdx);
}
else
{
copySaoOneLcuParam(&pSaoParam->saoLcuParam[iCompIdx][iAddr], &pSaoParam->saoLcuParam[iCompIdx][iAddr-pSaoParam->numCuInWidth]);
}
}
else
{
copySaoOneLcuParam(&pSaoParam->saoLcuParam[iCompIdx][iAddr], &pSaoParam->saoLcuParam[iCompIdx][iAddr-1]);
}
}
else
{
pSaoParam->saoLcuParam[iCompIdx][iAddr].typeIdx = -1;
pSaoParam->saoLcuParam[iCompIdx][iAddr].subTypeIdx = 0;
}
}
}
在调试过程中,对于第一个ctu来讲,rx和ry均为0,因此不会解析MergeUp和MergeLeft语法元素。实际上执行的部分是parseSaoOffset(&(pSaoParam->saoLcuParam[iCompIdx][iAddr]), iCompIdx);
Void TDecSbac::parseSaoOffset(SaoLcuParam* psSaoLcuParam, UInt compIdx)
{
UInt uiSymbol;
static Int iTypeLength[MAX_NUM_SAO_TYPE] =
{
SAO_EO_LEN,
SAO_EO_LEN,
SAO_EO_LEN,
SAO_EO_LEN,
SAO_BO_LEN
};
if (compIdx==2)
{
uiSymbol = (UInt)( psSaoLcuParam->typeIdx + 1);
}
else
{
parseSaoTypeIdx(uiSymbol);
}
psSaoLcuParam->typeIdx = (Int)uiSymbol - 1;
if (uiSymbol)
{
psSaoLcuParam->length = iTypeLength[psSaoLcuParam->typeIdx];
Int bitDepth = compIdx ? g_bitDepthC : g_bitDepthY;
Int offsetTh = 1 << min(bitDepth - 5,5);
if( psSaoLcuParam->typeIdx == SAO_BO )
{
for(Int i=0; i< psSaoLcuParam->length; i++)
{
parseSaoMaxUvlc(uiSymbol, offsetTh -1 );
psSaoLcuParam->offset[i] = uiSymbol;
}
for(Int i=0; i< psSaoLcuParam->length; i++)
{
if (psSaoLcuParam->offset[i] != 0)
{
m_pcTDecBinIf->decodeBinEP ( uiSymbol);
if (uiSymbol)
{
psSaoLcuParam->offset[i] = -psSaoLcuParam->offset[i] ;
}
}
}
parseSaoUflc(5, uiSymbol );
psSaoLcuParam->subTypeIdx = uiSymbol;
}
else if( psSaoLcuParam->typeIdx < 4 )
{
parseSaoMaxUvlc(uiSymbol, offsetTh -1 ); psSaoLcuParam->offset[0] = uiSymbol;
parseSaoMaxUvlc(uiSymbol, offsetTh -1 ); psSaoLcuParam->offset[1] = uiSymbol;
parseSaoMaxUvlc(uiSymbol, offsetTh -1 ); psSaoLcuParam->offset[2] = -(Int)uiSymbol;
parseSaoMaxUvlc(uiSymbol, offsetTh -1 ); psSaoLcuParam->offset[3] = -(Int)uiSymbol;
if (compIdx != 2)
{
parseSaoUflc(2, uiSymbol );
psSaoLcuParam->subTypeIdx = uiSymbol;
psSaoLcuParam->typeIdx += psSaoLcuParam->subTypeIdx;
}
}
}
else
{
psSaoLcuParam->length = 0;
}
}
在该函数中执行parseSaoTypeIdx(uiSymbol);
Void TDecSbac::parseSaoTypeIdx (UInt& ruiVal)
{
UInt uiCode;
m_pcTDecBinIf->decodeBin( uiCode, m_cSaoTypeIdxSCModel.get( 0, 0, 0 ) );
if (uiCode == 0)
{
ruiVal = 0;
}
else
{
m_pcTDecBinIf->decodeBinEP( uiCode );
if (uiCode == 0)
{
ruiVal = 5;
}
else
{
ruiVal = 1;
}
}
}
该函数调用了上篇文章中解释过的decodeBin函数实现解析。回顾上篇文章中对该函数的解释,可以看出该函数的输出值实际就是valMps或者1-valMps,组成一串二进制码流,也就是算数编码前的二值化过的语法元素。通过这个函数输出了一个二进制字符,但是却并没有从原始码流中读取数据,这种现象开始时相当让我感到费解,不过想了一下便明白了这是很正常的,因为这正是算数编码过程的逆过程。在编码的时候,编码器“吸收”了一个、两个甚至更多个字符,多次进行MPS/LPS判断和区间划分,直到满足一定条件后才会输出一个0或1。那么在解码的时候,也会可能会出现输出了多个字符才“消耗”了1bit码流的情况。从上述函数的实现中可以得知,parseSaoTypeIdx函数返回给uiSymbol的值为0,那么parseSaoOffset(SaoLcuParam* psSaoLcuParam, UInt compIdx)函数便不会做后续操作,将psSaoLcuParam->length置0后直接返回。
从parseSaoOneLcuInterleaving函数中可以看出,parseSaoOffset将会循环调用3次。在第二次调用中,我们的demo程序显示,parseSaoTypeIdx中uiCode返回值为1,因此,下面继续调用了decodeBinEP函数:
Void
TDecBinCABAC::decodeBinEP( UInt& ruiBin )
{
m_uiValue += m_uiValue;
if ( ++m_bitsNeeded >= 0 )
{
m_bitsNeeded = -8;
m_uiValue += m_pcTComBitstream->readByte();
}
ruiBin = 0;
UInt scaledRange = m_uiRange << 7;
if ( m_uiValue >= scaledRange )
{
ruiBin = 1;
m_uiValue -= scaledRange;
}
}
该函数向实参uiCode输出值为0,根据条件,parseSaoTypeIdx的输出值为5。并且将执行4次parseSaoMaxUvlc,将解码的结果赋予psSaoLcuParam->offset[i]。
Void TDecSbac::parseSaoMaxUvlc ( UInt& val, UInt maxSymbol )
{
if (maxSymbol == 0)
{
val = 0;
return;
}
UInt code;
Int i;
m_pcTDecBinIf->decodeBinEP( code );//①
if ( code == 0 )
{
val = 0;
return;
}
i=1;
while (1)
{
m_pcTDecBinIf->decodeBinEP( code );//②
if ( code == 0 )
{
break;
}
i++;
if (i == maxSymbol)
{
break;
}
}
val = i;
}
在执行第一次parseSaoMaxUvlc中,只执行了函数①,code等于0因此返回val=0;执行第二次parseSaoMaxUvlc,①返回code=1,因此将会执行下面的while循环,共执行三次,返回val=3;值得注意的是,在第三次循环中,decodeBinEP函数读取了码流中的第三个字符加到了m_uiValue上;执行第三次parseSaoMaxUvlc,①返回code=1,下面的while循环只会执行一次,返回val=1;第四次执行parseSaoMaxUvlc,①返回code=1,while循环执行一次,返回val=1。
接下来,对第2~4个offset值,将各调用一次decodeBinEP,为1则取其负值。
调用parseSaoUflc(5, uiSymbol ),该函数调用了decodeBinsEP,该函数可以认为一次处理多位。实现如下:
Void TDecBinCABAC::decodeBinsEP( UInt& ruiBin, Int numBins )
{
UInt bins = 0;
while ( numBins > 8 )
{
m_uiValue = ( m_uiValue << 8 ) + ( m_pcTComBitstream->readByte() << ( 8 + m_bitsNeeded ) );
UInt scaledRange = m_uiRange << 15;
for ( Int i = 0; i < 8; i++ )
{
bins += bins;
scaledRange >>= 1;
if ( m_uiValue >= scaledRange )
{
bins++;
m_uiValue -= scaledRange;
}
}
numBins -= 8;
}
m_bitsNeeded += numBins;
m_uiValue <<= numBins;
if ( m_bitsNeeded >= 0 )
{
m_uiValue += m_pcTComBitstream->readByte() << m_bitsNeeded;
m_bitsNeeded -= 8;
}
UInt scaledRange = m_uiRange << ( numBins + 7 );
for ( Int i = 0; i < numBins; i++ )
{
bins += bins;
scaledRange >>= 1;
if ( m_uiValue >= scaledRange )
{
bins++;
m_uiValue -= scaledRange;
}
}
ruiBin = bins;
}
在该函数中读取了码流中的第四个字节(107),在for循环中比较m_uiValue和scaledRange的值并决定bins加1或倍增,最终返回值为13。该值赋予psSaoLcuParam->subTypeIdx。
自此,parseSaoOffset的任务便已完成。而类似的工作将循环三次以完成pcSbacDecoder->parseSaoOneLcuInterleaving的功能。
PS:跟踪了这么久的代码,但是现在的问题依然是知其然不知其所以然……看来接下来需要在标准文档和其他一些资料上面下些功夫了。