VTM中色度预测一共测试8种模式,其顺序为为PLANAR
,VER
, HOR
,DC
,LM_CHROMA
,MDLM_L
,MDLM_T
,DM_CHROMA
,其中LM_CHROMA
,MDLM_L
和MDLM_T
分别表示CCLM
,LM_L
和LM_A
跨分量预测模式,DM_CHROMA
为从亮度分量获取到的模式。其中如果DM_CHROMA
和前四个默认预测模式相同,则将对应默认模式替换为VDIA
模式。
在classic CCLM
之外,模型参数还有另外两种计算方式,即CCLM还有两位两种模式,称之为LM_A
和LM_L
模式。
LM_A
模式仅使用已重建亮度像素的上方元素作为参考像素计算线性模型参数,为了获取足够参考像素,上方亮度参考像素扩展为W+H个LM_L
模式仅使用已重建像素左侧元素计算线性模型参数,同样,左侧参考像素扩展为H+W个填充色度预测列表通过函数getIntraChromaCandModes()完成,其具体实现如下:
void PU::getIntraChromaCandModes( const PredictionUnit &pu, unsigned modeList[NUM_CHROMA_MODE] )
{
{
modeList[ 0 ] = PLANAR_IDX;
modeList[ 1 ] = VER_IDX;
modeList[ 2 ] = HOR_IDX;
modeList[ 3 ] = DC_IDX;
modeList[4] = LM_CHROMA_IDX;
modeList[5] = MDLM_L_IDX;
modeList[6] = MDLM_T_IDX;
modeList[7] = DM_CHROMA_IDX;
Position topLeftPos = pu.blocks[pu.chType].lumaPos();
Position refPos = topLeftPos.offset( pu.blocks[pu.chType].lumaSize().width >> 1, pu.blocks[pu.chType].lumaSize().height >> 1 );
const PredictionUnit *lumaPU = CS::isDualITree( *pu.cs ) ? pu.cs->picture->cs->getPU( refPos, CHANNEL_TYPE_LUMA ) : &pu;
const uint32_t lumaMode = lumaPU->intraDir[CHANNEL_TYPE_LUMA];
for( int i = 0; i < 4; i++ )
{
if( lumaMode == modeList[i] )
{
modeList[i] = VDIA_IDX;
break;
}
}
}
}
色度预测的主要流程如下:
chroma
预测模式候选列表void IntraSearch::estIntraPredChromaQT( CodingUnit &cu, Partitioner &partitioner, const double maxCostAllowed )
{
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 ); // 设置重建标志false
#if JVET_M0102_INTRA_SUBPARTITIONS
double bestCostSoFar = maxCostAllowed;
bool lumaUsesISP = !CS::isDualITree( *cu.cs ) && cu.ispMode; // 帧间亮度是否为ISP模式
PartSplit ispType = lumaUsesISP ? CU::getISPType( cu, COMPONENT_Y ) : TU_NO_ISP; // 亮度ISP类型
CHECK( cu.ispMode && bestCostSoFar < 0, "bestCostSoFar must be positive!" );
#endif
auto &pu = *cu.firstPU;
{
uint32_t uiBestMode = 0; // 记录最优模式
Distortion uiBestDist = 0; // 最优模式对应distortion
double dBestCost = MAX_DOUBLE; // 最优模式rdCost
//----- init mode list ----
{
uint32_t uiMinMode = 0;
uint32_t uiMaxMode = NUM_CHROMA_MODE;
//----- check chroma modes -----
uint32_t chromaCandModes[ NUM_CHROMA_MODE ];
PU::getIntraChromaCandModes( pu, chromaCandModes ); // 构建chroma模式列表
// create a temporary 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
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 ); // add tu
}
std::vector<TransformUnit*> orgTUs;
#if JVET_M0102_INTRA_SUBPARTITIONS
if( lumaUsesISP ) // 如果亮度是ISP,记录当前cu,pu信息
{
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
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
// SATD pre-selecting.
int satdModeList[NUM_CHROMA_MODE]; // chroma基于satd模式列表
int64_t satdSortedCost[NUM_CHROMA_MODE]; // 对应satd cost
for (int i = 0; i < NUM_CHROMA_MODE; i++)
{
// 默认跳过基于SATD预测的模式,直接进行fullRD代价计算
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;
// 这里将pu模式设置为LM_L模式,为了在下面对luma重建信号进行下采样
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()); // 获取Cb分量参考像素
initIntraPatternChType(cu, pu.Cr()); // 获取Cr分量参考像素
xGetLumaRecPixels(pu, pu.Cb()); // CCLM预测模式亮度重建参考像素获取
// *************** 基于SATD模式粗选择 ************* //
for (int idx = uiMinMode; idx <= uiMaxMode - 1; idx++) // 遍历所有候选
{
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)) // CCLM,LM_A,LM_L模式,进行跨分量预测
{
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); // chroma 代价为Cb和Cr的和
satdSortedCost[idx] = sad;
}
// sort the mode based on the cost from small to large.
// 基于satd cost 进行模式排序,冒泡
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
for (int i = 0; i < reducedModeNumber; i++) // 剔除两个SATD cost最大的模式
{
modeIsEnable[satdModeList[uiMaxMode - 1 - i]] = 0; // disable the last reducedModeNumber modes
}
// save the dist 记录当前cost
Distortion baseDist = cs.dist;
// *************** 进行RQT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; //重置上下文
//----- chroma coding -----
pu.intraDir[1] = chromaIntraMode;
#if JVET_M0102_INTRA_SUBPARTITIONS
xRecurIntraChromaCodingQT( cs, partitioner, bestCostSoFar, ispType ); // RQT计算变换量化,反变换反量化,计算重建之后distortion
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 ); // distortion = currentCost - storeCost
//----- compare -----
if( dCost < dBestCost ) // 更新best信息,记录预测、残差、重建信号
{
#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;
}
}
// 将最优信号,更新在picture buffer
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;
}
//----- restore context models ----- 更新上下文模型
m_CABACEstimator->getCtx() = ctxStart;
#if JVET_M0102_INTRA_SUBPARTITIONS
if( lumaUsesISP && bestCostSoFar >= maxCostAllowed )
{
cu.ispMode = 0;
}
#endif
}