xEncodeInterResidual function, the function function is just like the function name, it is the function entry of encoding the inter-frame prediction residual. This function mainly selects the best residual coding mode from skip residual coding, DCT-2, MTS and SBT modes.
Mainly divided into the following steps:
1. Pre-analysis residual
(1) Calculate the current SSE of the entire CU , calculate and derive the minimum distortion of several modes of SBT, and select at most four modes with the smallest distortion of SBT (SBT mode selection involves the fast algorithm of SBT) ( calcMinDistSbt function)
(2) Find the optimal transformation mode of the same SSE as the current CU in history
Two, coding mode selection
(1) Skip mode or MTS/DCT-2 situation: If the current mode is the skip residual coding mode or there is no conversion mode with the same SSE as the current CU in history, or the historically best conversion mode is MTS or DCT2, then Call encodeResAndCalcRdInterCU function for residual coding and calculate RD Cost
(2) SBT situation: select numSbtRdo SBT modes through a fast algorithm, traverse numSbtRdo SBT modes, call encodeResAndCalcRdInterCU function for residual coding and calculate RD Cost
Three, save the best transformation mode
Note: Since SBT has up to 8 modes, in order to reduce the complexity, many fast algorithms on SBT are involved
The function code and related comments are as follows:
void EncCu::xEncodeInterResidual( CodingStructure *&tempCS
, CodingStructure *&bestCS
, Partitioner &partitioner
, const EncTestMode& encTestMode
, int residualPass
, bool* bestHasNonResi
, double* equBcwCost
)
{
CodingUnit* cu = tempCS->getCU( partitioner.chType );
double bestCostInternal = MAX_DOUBLE;
double bestCost = bestCS->cost;
double bestCostBegin = bestCS->cost;
CodingUnit* prevBestCU = bestCS->getCU( partitioner.chType );
uint8_t prevBestSbt = ( prevBestCU == nullptr ) ? 0 : prevBestCU->sbtInfo;
bool swapped = false; // avoid unwanted data copy
bool reloadCU = false;
const PredictionUnit& pu = *cu->firstPU;
// clang-format off
const int affineShiftTab[3] =
{
MV_PRECISION_INTERNAL - MV_PRECISION_QUARTER,
MV_PRECISION_INTERNAL - MV_PRECISION_SIXTEENTH,
MV_PRECISION_INTERNAL - MV_PRECISION_INT
};
const int normalShiftTab[NUM_IMV_MODES] =
{
MV_PRECISION_INTERNAL - MV_PRECISION_QUARTER,
MV_PRECISION_INTERNAL - MV_PRECISION_INT,
MV_PRECISION_INTERNAL - MV_PRECISION_4PEL,
MV_PRECISION_INTERNAL - MV_PRECISION_HALF,
};
// clang-format on
int mvShift;
for (int refList = 0; refList < NUM_REF_PIC_LIST_01; refList++)
{
if (pu.refIdx[refList] >= 0)
{
if (!cu->affine)
{
mvShift = normalShiftTab[cu->imv];
Mv signaledmvd(pu.mvd[refList].getHor() >> mvShift, pu.mvd[refList].getVer() >> mvShift);
if (!((signaledmvd.getHor() >= MVD_MIN) && (signaledmvd.getHor() <= MVD_MAX)) || !((signaledmvd.getVer() >= MVD_MIN) && (signaledmvd.getVer() <= MVD_MAX)))
return;
}
else
{
for (int ctrlP = 1 + (cu->affineType == AFFINEMODEL_6PARAM); ctrlP >= 0; ctrlP--)
{
mvShift = affineShiftTab[cu->imv];
Mv signaledmvd(pu.mvdAffi[refList][ctrlP].getHor() >> mvShift, pu.mvdAffi[refList][ctrlP].getVer() >> mvShift);
if (!((signaledmvd.getHor() >= MVD_MIN) && (signaledmvd.getHor() <= MVD_MAX)) || !((signaledmvd.getVer() >= MVD_MIN) && (signaledmvd.getVer() <= MVD_MAX)))
return;
}
}
}
}
// avoid MV exceeding 18-bit dynamic range
// 避免MV超过18位动态范围
const int maxMv = 1 << 17;
if (!cu->affine && !pu.mergeFlag)
{
if ( (pu.refIdx[0] >= 0 && (pu.mv[0].getAbsHor() >= maxMv || pu.mv[0].getAbsVer() >= maxMv))
|| (pu.refIdx[1] >= 0 && (pu.mv[1].getAbsHor() >= maxMv || pu.mv[1].getAbsVer() >= maxMv)))
{
return;
}
}
if (cu->affine && !pu.mergeFlag)
{
for (int refList = 0; refList < NUM_REF_PIC_LIST_01; refList++)
{
if (pu.refIdx[refList] >= 0)
{
for (int ctrlP = 1 + (cu->affineType == AFFINEMODEL_6PARAM); ctrlP >= 0; ctrlP--)
{
if (pu.mvAffi[refList][ctrlP].getAbsHor() >= maxMv || pu.mvAffi[refList][ctrlP].getAbsVer() >= maxMv)
{
return;
}
}
}
}
}
const bool mtsAllowed = tempCS->sps->getUseInterMTS() && CU::isInter( *cu ) && partitioner.currArea().lwidth() <= MTS_INTER_MAX_CU_SIZE && partitioner.currArea().lheight() <= MTS_INTER_MAX_CU_SIZE;
uint8_t sbtAllowed = cu->checkAllowedSbt();//是否允许使用SBT，判断当前预测模式、CU尺寸是否符合要求
// SBT resolution-dependent fast algorithm: not try size-64 SBT in RDO for low-resolution sequences (now resolution below HD)
// SBT分辨率相关快速算法：对于低分辨率序列（现在分辨率低于HD），不要尝试RDO中的size-64sbt
if( tempCS->pps->getPicWidthInLumaSamples() < (uint32_t)m_pcEncCfg->getSBTFast64WidthTh() )
{
sbtAllowed = ((cu->lwidth() > 32 || cu->lheight() > 32)) ? 0 : sbtAllowed;
}
uint8_t numRDOTried = 0;
Distortion sbtOffDist = 0;
bool sbtOffRootCbf = 0;
double sbtOffCost = MAX_DOUBLE;
double currBestCost = MAX_DOUBLE;
bool doPreAnalyzeResi = ( sbtAllowed || mtsAllowed ) && residualPass == 0;
m_pcInterSearch->initTuAnalyzer();
if( doPreAnalyzeResi )
{
m_pcInterSearch->calcMinDistSbt( *tempCS, *cu, sbtAllowed );
}
auto slsSbt = dynamic_cast<SaveLoadEncInfoSbt*>( m_modeCtrl );
int slShift = 4 + std::min( (int)gp_sizeIdxInfo->idxFrom( cu->lwidth() ) + (int)gp_sizeIdxInfo->idxFrom( cu->lheight() ), 9 );
Distortion curPuSse = m_pcInterSearch->getEstDistSbt( NUMBER_SBT_MODE );//当前PU的SSE
uint8_t currBestSbt = 0;
uint8_t currBestTrs = MAX_UCHAR;
uint8_t histBestSbt = MAX_UCHAR;
uint8_t histBestTrs = MAX_UCHAR;
m_pcInterSearch->setHistBestTrs( MAX_UCHAR, MAX_UCHAR );
// 设置历史最佳SBT模式和MTS模式
if( doPreAnalyzeResi )
{
if( m_pcInterSearch->getSkipSbtAll() && !mtsAllowed ) //emt is off MTS关闭
{
histBestSbt = 0; //try DCT2
m_pcInterSearch->setHistBestTrs( histBestSbt, histBestTrs );
}
else
{
assert( curPuSse != std::numeric_limits<uint64_t>::max() );
//寻找历史上与当前PU相同SSE的最优SBT模式
uint16_t compositeSbtTrs = slsSbt->findBestSbt( cu->cs->area, (uint32_t)( curPuSse >> slShift ) );
histBestSbt = ( compositeSbtTrs >> 0 ) & 0xff;
histBestTrs = ( compositeSbtTrs >> 8 ) & 0xff;
// 特殊情况，加载SBT时跳过SBT
if( m_pcInterSearch->getSkipSbtAll() && CU::isSbtMode( histBestSbt ) ) //special case, skip SBT when loading SBT
{
histBestSbt = 0; //try DCT2
}
m_pcInterSearch->setHistBestTrs( histBestSbt, histBestTrs );
}
}
{// EMT loop
if( reloadCU )
{
if( bestCost == bestCS->cost ) //The first EMT pass didn't become the bestCS, so we clear the TUs generated
{
tempCS->clearTUs();
}
else if( false == swapped )
{
tempCS->initStructData( encTestMode.qp );
tempCS->copyStructure( *bestCS, partitioner.chType );
tempCS->getPredBuf().copyFrom( bestCS->getPredBuf() );
bestCost = bestCS->cost;
cu = tempCS->getCU( partitioner.chType );
swapped = true;
}
else
{
tempCS->clearTUs();
bestCost = bestCS->cost;
cu = tempCS->getCU( partitioner.chType );
}
//we need to restart the distortion for the new tempCS, the bit count and the cost
//我们需要为新的tempCS重新开始失真，比特数和成本
tempCS->dist = 0;
tempCS->fracBits = 0;
tempCS->cost = MAX_DOUBLE;
tempCS->costDbOffset = 0;
}
reloadCU = true; // enable cu reloading
cu->skip = false;
cu->sbtInfo = 0;
const bool skipResidual = residualPass == 1;
//跳过残差编码（Skip模式）或历史上没有和当前相同SSE的SBT模式或者历史上最佳模式是MTS或DCT-2
if( skipResidual || histBestSbt == MAX_UCHAR || !CU::isSbtMode( histBestSbt ) )
{
//编码残差并计算CU的RD Cost
m_pcInterSearch->encodeResAndCalcRdInterCU( *tempCS, partitioner, skipResidual );
if (tempCS->slice->getSPS()->getUseColorTrans())
{
bestCS->tmpColorSpaceCost = tempCS->tmpColorSpaceCost;
bestCS->firstColorSpaceSelected = tempCS->firstColorSpaceSelected;
}
numRDOTried += mtsAllowed ? 2 : 1;
xEncodeDontSplit( *tempCS, partitioner );
xCheckDQP( *tempCS, partitioner );
#if JVET_Q0267_RESET_CHROMA_QP_OFFSET
xCheckChromaQPOffset( *tempCS, partitioner );
#endif
if( NULL != bestHasNonResi && (bestCostInternal > tempCS->cost) )
{
bestCostInternal = tempCS->cost;
if (!(tempCS->getPU(partitioner.chType)->ciipFlag))
*bestHasNonResi = !cu->rootCbf;
}
if (cu->rootCbf == false)
{
if (tempCS->getPU(partitioner.chType)->ciipFlag)
{
tempCS->cost = MAX_DOUBLE;
tempCS->costDbOffset = 0;
return;
}
}
currBestCost = tempCS->cost;
sbtOffCost = tempCS->cost;
sbtOffDist = tempCS->dist;
sbtOffRootCbf = cu->rootCbf;
//设置当前最佳变换模式（MTS或者DCT-2）和变换类型
currBestSbt = CU::getSbtInfo(cu->firstTU->mtsIdx[COMPONENT_Y] > MTS_SKIP ? SBT_OFF_MTS : SBT_OFF_DCT, 0);
currBestTrs = cu->firstTU->mtsIdx[COMPONENT_Y];
#if WCG_EXT
DTRACE_MODE_COST( *tempCS, m_pcRdCost->getLambda( true ) );
#else
DTRACE_MODE_COST( *tempCS, m_pcRdCost->getLambda() );
#endif
xCheckBestMode( tempCS, bestCS, partitioner, encTestMode );
}
uint8_t numSbtRdo = CU::numSbtModeRdo( sbtAllowed );
//early termination if all SBT modes are not allowed 如果不允许所有SBT模式，则提前终止
//normative
if( !sbtAllowed || skipResidual )
{
numSbtRdo = 0;
}
//fast algorithm
//快速算法：存在历史最佳SBT模式但是最终选择的模式不是SBT模式
if( ( histBestSbt != MAX_UCHAR && !CU::isSbtMode( histBestSbt ) ) || m_pcInterSearch->getSkipSbtAll() )
{
numSbtRdo = 0;
}
//如果当前PU关闭SBT的Cost大于最佳Cost*th，则跳过SBT的RDO过程，将numSbtRdo设置为0
if( bestCost != MAX_DOUBLE && sbtOffCost != MAX_DOUBLE )
{
double th = 1.07;
if( !( prevBestSbt == 0 || m_sbtCostSave[0] == MAX_DOUBLE ) )
{
//m_sbtCostSave[0] = sbtOffCost;上一次关闭SBT的Cost
//m_sbtCostSave[1] = currBestCost;上次最佳的Cost
assert( m_sbtCostSave[1] <= m_sbtCostSave[0] );
th *= ( m_sbtCostSave[0] / m_sbtCostSave[1] );
}
if( sbtOffCost > bestCost * th )
{
numSbtRdo = 0;
}
}
//当SbtOffCost小于某一阈值的时候，跳过SBT的RDO过程
if( !sbtOffRootCbf && sbtOffCost != MAX_DOUBLE )
{
double th = Clip3( 0.05, 0.55, ( 27 - cu->qp ) * 0.02 + 0.35 );
if( sbtOffCost < m_pcRdCost->calcRdCost( ( cu->lwidth() * cu->lheight() ) << SCALE_BITS, 0 ) * th )
{
numSbtRdo = 0;
}
}
//当历史最佳SBT模式存在且可用时，将numSbtRdo置1，仅对这一种SBT模式进行RD Cost的计算
if( histBestSbt != MAX_UCHAR && numSbtRdo != 0 )
{
numSbtRdo = 1;
m_pcInterSearch->initSbtRdoOrder( CU::getSbtMode( CU::getSbtIdx( histBestSbt ), CU::getSbtPos( histBestSbt ) ) );
}
//遍历numSbtRdo种SBT模式
for( int sbtModeIdx = 0; sbtModeIdx < numSbtRdo; sbtModeIdx++ )
{
uint8_t sbtMode = m_pcInterSearch->getSbtRdoOrder( sbtModeIdx );
uint8_t sbtIdx = CU::getSbtIdxFromSbtMode( sbtMode );
uint8_t sbtPos = CU::getSbtPosFromSbtMode( sbtMode );
//fast algorithm (early skip, save & load)
//快速算法（提前跳过，保存和加载）
if( histBestSbt == MAX_UCHAR )
{
//快速算法，提前跳过SBT（返回0/1/2/3，表示跳过当前SBT模式）
uint8_t skipCode = m_pcInterSearch->skipSbtByRDCost( cu->lwidth(), cu->lheight(), cu->mtDepth, sbtIdx, sbtPos, bestCS->cost, sbtOffDist, sbtOffCost, sbtOffRootCbf );
if( skipCode != MAX_UCHAR )
{
continue;
}
//比较当前SBT模式与前一个相同划分方式的SBT模式的Est失真
if( sbtModeIdx > 0 )
{
uint8_t prevSbtMode = m_pcInterSearch->getSbtRdoOrder( sbtModeIdx - 1 );//获得前一个SBT模式
//make sure the prevSbtMode is the same size as the current SBT mode (otherwise the estimated dist may not be comparable)
//确保prevSbtMode与当前SBT模式划分后的尺寸相同（否则估计的dist可能无法比较）
if( CU::isSameSbtSize( prevSbtMode, sbtMode ) )
{
Distortion currEstDist = m_pcInterSearch->getEstDistSbt( sbtMode );
Distortion prevEstDist = m_pcInterSearch->getEstDistSbt( prevSbtMode );
if( currEstDist > prevEstDist * 1.15 )
{
continue;
}
}
}
}
//init tempCS and TU
//初始化tempCS和TU
if( bestCost == bestCS->cost ) //The first EMT pass didn't become the bestCS, so we clear the TUs generated
{ //第一次EMT过程并没有成为bestCS，所以我们清除了生成的TUs
tempCS->clearTUs();
}
else if( false == swapped )
{
tempCS->initStructData( encTestMode.qp );
tempCS->copyStructure( *bestCS, partitioner.chType );
tempCS->getPredBuf().copyFrom( bestCS->getPredBuf() );
bestCost = bestCS->cost;
cu = tempCS->getCU( partitioner.chType );
swapped = true;
}
else
{
tempCS->clearTUs();
bestCost = bestCS->cost;
cu = tempCS->getCU( partitioner.chType );
}
//we need to restart the distortion for the new tempCS, the bit count and the cost
//我们需要为新的tempCS重新开始失真，比特数和成本
tempCS->dist = 0;
tempCS->fracBits = 0;
tempCS->cost = MAX_DOUBLE;
cu->skip = false;
//set SBT info
//设置SBT信息
cu->setSbtIdx( sbtIdx );
cu->setSbtPos( sbtPos );
//try residual coding
//尝试残差编码
m_pcInterSearch->encodeResAndCalcRdInterCU( *tempCS, partitioner, skipResidual );
if (tempCS->slice->getSPS()->getUseColorTrans())
{
bestCS->tmpColorSpaceCost = tempCS->tmpColorSpaceCost;
bestCS->firstColorSpaceSelected = tempCS->firstColorSpaceSelected;
}
numRDOTried++;
xEncodeDontSplit( *tempCS, partitioner );
xCheckDQP( *tempCS, partitioner );
#if JVET_Q0267_RESET_CHROMA_QP_OFFSET
xCheckChromaQPOffset( *tempCS, partitioner );
#endif
if( NULL != bestHasNonResi && ( bestCostInternal > tempCS->cost ) )
{
bestCostInternal = tempCS->cost;
if( !( tempCS->getPU( partitioner.chType )->ciipFlag ) )
*bestHasNonResi = !cu->rootCbf;
}
//保存当前最佳Cost和SBT信息
if( tempCS->cost < currBestCost )
{
currBestSbt = cu->sbtInfo;
currBestTrs = tempCS->tus[cu->sbtInfo ? cu->getSbtPos() : 0]->mtsIdx[COMPONENT_Y];
assert( currBestTrs == 0 || currBestTrs == 1 );
currBestCost = tempCS->cost;
}
#if WCG_EXT
DTRACE_MODE_COST( *tempCS, m_pcRdCost->getLambda( true ) );
#else
DTRACE_MODE_COST( *tempCS, m_pcRdCost->getLambda() );
#endif
xCheckBestMode( tempCS, bestCS, partitioner, encTestMode );
}//end numSbtRdo loop
if( bestCostBegin != bestCS->cost )
{
m_sbtCostSave[0] = sbtOffCost;
m_sbtCostSave[1] = currBestCost;
}
} //end emt loop
if( histBestSbt == MAX_UCHAR && doPreAnalyzeResi && numRDOTried > 1 )
{
//保存最佳的变换模式
slsSbt->saveBestSbt( cu->cs->area, (uint32_t)( curPuSse >> slShift ), currBestSbt, currBestTrs );
}
tempCS->cost = currBestCost;
if( ETM_INTER_ME == encTestMode.type )
{
if( equBcwCost != NULL )
{
if( tempCS->cost < ( *equBcwCost ) && cu->BcwIdx == BCW_DEFAULT )
{
( *equBcwCost ) = tempCS->cost;
}
}
else
{
CHECK( equBcwCost == NULL, "equBcwCost == NULL" );
}
if( tempCS->slice->getCheckLDC() && !cu->imv && cu->BcwIdx != BCW_DEFAULT && tempCS->cost < m_bestBcwCost[1] )
{
if( tempCS->cost < m_bestBcwCost[0] )
{
m_bestBcwCost[1] = m_bestBcwCost[0];
m_bestBcwCost[0] = tempCS->cost;
m_bestBcwIdx[1] = m_bestBcwIdx[0];
m_bestBcwIdx[0] = cu->BcwIdx;
}
else
{
m_bestBcwCost[1] = tempCS->cost;
m_bestBcwIdx[1] = cu->BcwIdx;
}
}
}
}The CheckAllowSbt function is mainly used to check whether the current CU can use SBT technology:
const uint8_t CodingUnit::checkAllowedSbt() const
{
if( !slice->getSPS()->getUseSBT() )
{
// SPS层没有开启SBT
return 0;
}
//check on prediction mode
// 检查预测模式：帧内预测、IBC、PLT模式不使用SBT
if (predMode == MODE_INTRA || predMode == MODE_IBC || predMode == MODE_PLT ) //intra, palette or IBC
{
return 0;
}
if( firstPU->ciipFlag )
{
return 0;
}
#if !JVET_Q0806
if( triangle )
{
return 0;
}
#endif
uint8_t sbtAllowed = 0;
int cuWidth = lwidth();//Cu宽度
int cuHeight = lheight();//Cu高度
bool allow_type[NUMBER_SBT_IDX];
memset( allow_type, false, NUMBER_SBT_IDX * sizeof( bool ) );
//parameter
// 参数：最大和最小支持SBT的变换块尺寸
int maxSbtCUSize = cs->sps->getMaxTbSize();
int minSbtCUSize = 1 << ( MIN_CU_LOG2 + 1 );//8
//check on size
// 检查尺寸
if( cuWidth > maxSbtCUSize || cuHeight > maxSbtCUSize )
{
return 0;
}
allow_type[SBT_VER_HALF] = cuWidth >= minSbtCUSize;//是否允许2:2垂直划分
allow_type[SBT_HOR_HALF] = cuHeight >= minSbtCUSize;//是否允许2:2水平划分
allow_type[SBT_VER_QUAD] = cuWidth >= ( minSbtCUSize << 1 );//是否允许1:3/3:1垂直划分
allow_type[SBT_HOR_QUAD] = cuHeight >= ( minSbtCUSize << 1 );//是否允许1:3/3:1水平划分
for( int i = 0; i < NUMBER_SBT_IDX; i++ )
{
sbtAllowed += (uint8_t)allow_type[i] << i;
}
return sbtAllowed;
}The function of calcMinDistSbt is as follows:
(1) Calculate the SSE distortion between the predicted value and the original value of the entire block
(2) Calculate the weighted distortion of the SSE between the predicted value and the original value of the eight SBT modes that need to be coded and the sub-block that does not need to be coded: SSE_SBT = SSE (Code)/16 + SSE (Uncode) , and follow Sort from smallest to largest
void InterSearch::calcMinDistSbt( CodingStructure &cs, const CodingUnit& cu, const uint8_t sbtAllowed )
{
if( !sbtAllowed ) //不允许使用SBT时
{
m_estMinDistSbt[NUMBER_SBT_MODE] = 0;
for( int comp = 0; comp < getNumberValidTBlocks( *cs.pcv ); comp++ )
{
const ComponentID compID = ComponentID( comp );
CPelBuf pred = cs.getPredBuf( compID );
CPelBuf org = cs.getOrgBuf( compID );
m_estMinDistSbt[NUMBER_SBT_MODE] += m_pcRdCost->getDistPart( org, pred, cs.sps->getBitDepth( toChannelType( compID ) ), compID, DF_SSE );
}
return;
}
//SBT fast algorithm 2.1 : estimate a minimum RD cost of a SBT mode based on the luma distortion of uncoded part and coded part (assuming distorted can be reduced to 1/16);
// if this cost is larger than the best cost, no need to try a specific SBT mode
//SBT快速算法2.1：根据未编码部分和编码部分的luma失真估计SBT模式的最小RD Cost（假设失真可以减少到1/16）；
// 如果该成本大于最佳成本，则无需尝试特定的SBT模式
int cuWidth = cu.lwidth();
int cuHeight = cu.lheight();
int numPartX = cuWidth >= 16 ? 4 : ( cuWidth == 4 ? 1 : 2 );
int numPartY = cuHeight >= 16 ? 4 : ( cuHeight == 4 ? 1 : 2 );
Distortion dist[4][4];
memset( dist, 0, sizeof( Distortion ) * 16 );
for( uint32_t c = 0; c < getNumberValidTBlocks( *cs.pcv ); c++ )
{
const ComponentID compID = ComponentID( c );
const CompArea& compArea = cu.blocks[compID];
const CPelBuf orgPel = cs.getOrgBuf( compArea );
const CPelBuf predPel = cs.getPredBuf( compArea );
int lengthX = compArea.width / numPartX;
int lengthY = compArea.height / numPartY;
int strideOrg = orgPel.stride;
int stridePred = predPel.stride;
uint32_t uiShift = DISTORTION_PRECISION_ADJUSTMENT( ( *cs.sps.getBitDepth( toChannelType( compID ) ) - 8 ) << 1 );
Intermediate_Int iTemp;
//calc distY of 16 sub parts
//计算16个子部分的距离
for( int j = 0; j < numPartY; j++ )
{
for( int i = 0; i < numPartX; i++ )
{
int posX = i * lengthX;
int posY = j * lengthY;
const Pel* ptrOrg = orgPel.bufAt( posX, posY );
const Pel* ptrPred = predPel.bufAt( posX, posY );
Distortion uiSum = 0;
for( int n = 0; n < lengthY; n++ )
{
for( int m = 0; m < lengthX; m++ )
{
iTemp = ptrOrg[m] - ptrPred[m];
uiSum += Distortion( ( iTemp * iTemp ) >> uiShift );
}
ptrOrg += strideOrg;//加个CU的长度
ptrPred += stridePred;
}
if( isChroma( compID ) )
{
uiSum = (Distortion)( uiSum * m_pcRdCost->getChromaWeight() );
}
dist[j][i] += uiSum;
}
}
}
// SSE of a CU
// 计算出整个CU的SSE
m_estMinDistSbt[NUMBER_SBT_MODE] = 0;
for( int j = 0; j < numPartY; j++ )
{
for( int i = 0; i < numPartX; i++ )
{
m_estMinDistSbt[NUMBER_SBT_MODE] += dist[j][i];
}
}
//init per-mode dist
// 初始化每个模式的失真
for( int i = SBT_VER_H0; i < NUMBER_SBT_MODE; i++ )
{
m_estMinDistSbt[i] = std::numeric_limits<uint64_t>::max();
}
//SBT fast algorithm 1: not try SBT if the residual is too small to compensate bits for encoding residual info
//SBT快速算法1：如果残差太小，无法补偿编码残差信息的比特数，则不要尝试SBT
uint64_t minNonZeroResiFracBits = 12 << SCALE_BITS;
if( m_pcRdCost->calcRdCost( 0, m_estMinDistSbt[NUMBER_SBT_MODE] ) < m_pcRdCost->calcRdCost( minNonZeroResiFracBits, 0 ) )
{
m_skipSbtAll = true;
return;
}
//derive estimated minDist of SBT = zero-residual part distortion + non-zero residual part distortion / 16
//推导出SBT的估计值minDist=零残差部分失真+非零残差部分失真/16
int shift = 5;
Distortion distResiPart = 0, distNoResiPart = 0;
//如果允许垂直2:2划分
if( CU::targetSbtAllowed( SBT_VER_HALF, sbtAllowed ) )
{
int offsetResiPart = 0;
int offsetNoResiPart = numPartX / 2;
distResiPart = distNoResiPart = 0;
assert( numPartX >= 2 );
for( int j = 0; j < numPartY; j++ )
{
for( int i = 0; i < numPartX / 2; i++ )
{
distResiPart += dist[j][i + offsetResiPart];
distNoResiPart += dist[j][i + offsetNoResiPart];
}
}
m_estMinDistSbt[SBT_VER_H0] = ( distResiPart >> shift ) + distNoResiPart;
m_estMinDistSbt[SBT_VER_H1] = ( distNoResiPart >> shift ) + distResiPart;
}
if( CU::targetSbtAllowed( SBT_HOR_HALF, sbtAllowed ) )
{
int offsetResiPart = 0;
int offsetNoResiPart = numPartY / 2;
assert( numPartY >= 2 );
distResiPart = distNoResiPart = 0;
for( int j = 0; j < numPartY / 2; j++ )
{
for( int i = 0; i < numPartX; i++ )
{
distResiPart += dist[j + offsetResiPart][i];
distNoResiPart += dist[j + offsetNoResiPart][i];
}
}
m_estMinDistSbt[SBT_HOR_H0] = ( distResiPart >> shift ) + distNoResiPart;
m_estMinDistSbt[SBT_HOR_H1] = ( distNoResiPart >> shift ) + distResiPart;
}
if( CU::targetSbtAllowed( SBT_VER_QUAD, sbtAllowed ) )
{
assert( numPartX == 4 );
m_estMinDistSbt[SBT_VER_Q0] = m_estMinDistSbt[SBT_VER_Q1] = 0;
for( int j = 0; j < numPartY; j++ )
{
m_estMinDistSbt[SBT_VER_Q0] += dist[j][0] + ( ( dist[j][1] + dist[j][2] + dist[j][3] ) << shift );
m_estMinDistSbt[SBT_VER_Q1] += dist[j][3] + ( ( dist[j][0] + dist[j][1] + dist[j][2] ) << shift );
}
m_estMinDistSbt[SBT_VER_Q0] = m_estMinDistSbt[SBT_VER_Q0] >> shift;
m_estMinDistSbt[SBT_VER_Q1] = m_estMinDistSbt[SBT_VER_Q1] >> shift;
}
if( CU::targetSbtAllowed( SBT_HOR_QUAD, sbtAllowed ) )
{
assert( numPartY == 4 );
m_estMinDistSbt[SBT_HOR_Q0] = m_estMinDistSbt[SBT_HOR_Q1] = 0;
for( int i = 0; i < numPartX; i++ )
{
m_estMinDistSbt[SBT_HOR_Q0] += dist[0][i] + ( ( dist[1][i] + dist[2][i] + dist[3][i] ) << shift );
m_estMinDistSbt[SBT_HOR_Q1] += dist[3][i] + ( ( dist[0][i] + dist[1][i] + dist[2][i] ) << shift );
}
m_estMinDistSbt[SBT_HOR_Q0] = m_estMinDistSbt[SBT_HOR_Q0] >> shift;
m_estMinDistSbt[SBT_HOR_Q1] = m_estMinDistSbt[SBT_HOR_Q1] >> shift;
}
//SBT fast algorithm 5: try N SBT modes with the lowest distortion
//SBT快速算法5：尝试N个失真最低的SBT模式
Distortion temp[NUMBER_SBT_MODE];
memcpy( temp, m_estMinDistSbt, sizeof( Distortion ) * NUMBER_SBT_MODE );
memset( m_sbtRdoOrder, 255, NUMBER_SBT_MODE );
//从二分划分的4种模式中找出两种SSE最小的模式
int startIdx = 0, numRDO;
numRDO = CU::targetSbtAllowed( SBT_VER_HALF, sbtAllowed ) + CU::targetSbtAllowed( SBT_HOR_HALF, sbtAllowed );
numRDO = std::min( ( numRDO << 1 ), SBT_NUM_RDO );
for( int i = startIdx; i < startIdx + numRDO; i++ )
{
Distortion minDist = std::numeric_limits<uint64_t>::max();
for( int n = SBT_VER_H0; n <= SBT_HOR_H1; n++ )
{ //寻找四种二分划分模式中最小失真的模式
if( temp[n] < minDist )
{
minDist = temp[n];
m_sbtRdoOrder[i] = n;
}
}
//将找到的最小失真的模式的临时失真值修改为max值
temp[m_sbtRdoOrder[i]] = std::numeric_limits<uint64_t>::max();
}
//从四分划分的四种模式中选择两种SSE最小的模式
startIdx += numRDO;
numRDO = CU::targetSbtAllowed( SBT_VER_QUAD, sbtAllowed ) + CU::targetSbtAllowed( SBT_HOR_QUAD, sbtAllowed );
numRDO = std::min( ( numRDO << 1 ), SBT_NUM_RDO );
for( int i = startIdx; i < startIdx + numRDO; i++ )
{
Distortion minDist = std::numeric_limits<uint64_t>::max();
for( int n = SBT_VER_Q0; n <= SBT_HOR_Q1; n++ )
{
if( temp[n] < minDist )
{
minDist = temp[n];
m_sbtRdoOrder[i] = n;
}
}
temp[m_sbtRdoOrder[i]] = std::numeric_limits<uint64_t>::max();
}
}The skipSbtByRDCost function involves the fast algorithm of SBT:
(1) Estimate the minimum RD cost of the SBT mode based on the luma distortion of the uncoded part and the coded part (assuming that the distortion can be reduced to 1/16); if the Cost is greater than the optimal Cost, there is no need to try a specific SBT mode
(2) Skip SBT when the residual is too small
(3) Skip SBT when the estimated RD Cost is greater than the best Cost
uint8_t InterSearch::skipSbtByRDCost( int width, int height, int mtDepth, uint8_t sbtIdx, uint8_t sbtPos, double bestCost, Distortion distSbtOff, double costSbtOff, bool rootCbfSbtOff )
{
int sbtMode = CU::getSbtMode( sbtIdx, sbtPos );
//SBT fast algorithm 2.2 : estimate a minimum RD cost of a SBT mode based on the luma distortion of uncoded part and coded part (assuming distorted can be reduced to 1/16);
// if this cost is larger than the best cost, no need to try a specific SBT mode
//SBT快速算法2.2：根据未编码部分和编码部分的luma失真估计SBT模式的最小RD成本（假设失真可以减少到1/16）；
// 如果该Cost大于最佳Cost，则无需尝试特定的SBT模式
if( m_pcRdCost->calcRdCost( 11 << SCALE_BITS, m_estMinDistSbt[sbtMode] ) > bestCost )
{
return 0; //early skip type 0 提前退出
}
if( costSbtOff != MAX_DOUBLE )
{
if( !rootCbfSbtOff )
{
//SBT fast algorithm 3: skip SBT when the residual is too small (estCost is more accurate than fast algorithm 1, counting PU mode bits)
//SBT快速算法3：当残差太小时跳过SBT（estCost比快速算法1更精确，计算PU模式位）
uint64_t minNonZeroResiFracBits = 10 << SCALE_BITS;//最小非零残差编码的比特数
Distortion distResiPart;
if( sbtIdx == SBT_VER_HALF || sbtIdx == SBT_HOR_HALF )
{
distResiPart = (Distortion)( ( ( m_estMinDistSbt[NUMBER_SBT_MODE] - m_estMinDistSbt[sbtMode] ) * 9 ) >> 4 );
}
else
{
distResiPart = (Distortion)( ( ( m_estMinDistSbt[NUMBER_SBT_MODE] - m_estMinDistSbt[sbtMode] ) * 3 ) >> 3 );
}
double estCost = ( costSbtOff - m_pcRdCost->calcRdCost( 0 << SCALE_BITS, distSbtOff ) ) + m_pcRdCost->calcRdCost( minNonZeroResiFracBits, m_estMinDistSbt[sbtMode] + distResiPart );
if( estCost > costSbtOff )
{
return 1;
}
if( estCost > bestCost )
{
return 2;
}
}
else
{
//SBT fast algorithm 4: skip SBT when an estimated RD cost is larger than the bestCost
//SBT快速算法4：当估计的RD成本大于最佳成本时跳过SBT
double weight = sbtMode > SBT_HOR_H1 ? 0.4 : 0.6;
double estCost = ( ( costSbtOff - m_pcRdCost->calcRdCost( 0 << SCALE_BITS, distSbtOff ) ) * weight ) + m_pcRdCost->calcRdCost( 0 << SCALE_BITS, m_estMinDistSbt[sbtMode] );
if( estCost > bestCost )
{
return 3;
}
}
}
return MAX_UCHAR;
}