From e621c4a43f6e5ccce5486c5957b1933f6fbe7460 Mon Sep 17 00:00:00 2001
From: deluxan <santiago.de.luxan@hhi.fraunhofer.de>
Date: Fri, 2 Aug 2019 20:02:06 +0200
Subject: [PATCH] Integration of JVET-0502
---
cfg/encoder_lowdelay_P_vtm.cfg | 2 +-
cfg/encoder_lowdelay_vtm.cfg | 2 +-
cfg/encoder_randomaccess_vtm.cfg | 2 +-
source/Lib/CommonLib/IntraPrediction.cpp | 145 ++++++-
source/Lib/CommonLib/IntraPrediction.h | 11 +
source/Lib/CommonLib/TrQuant.cpp | 14 +
source/Lib/CommonLib/TrQuant.h | 6 +-
source/Lib/CommonLib/TypeDef.h | 6 +
source/Lib/CommonLib/UnitTools.cpp | 27 ++
source/Lib/CommonLib/UnitTools.h | 5 +
source/Lib/DecoderLib/CABACReader.cpp | 4 +
source/Lib/DecoderLib/DecCu.cpp | 34 ++
source/Lib/EncoderLib/CABACWriter.cpp | 8 +
source/Lib/EncoderLib/EncCu.cpp | 34 ++
source/Lib/EncoderLib/EncCu.h | 4 +
source/Lib/EncoderLib/EncModeCtrl.h | 3 +
source/Lib/EncoderLib/IntraSearch.cpp | 523 ++++++++++++++++++++++-
source/Lib/EncoderLib/IntraSearch.h | 133 +++++-
18 files changed, 956 insertions(+), 7 deletions(-)
diff --git a/cfg/encoder_lowdelay_P_vtm.cfg b/cfg/encoder_lowdelay_P_vtm.cfg
index 2d24145ce..be04c02b9 100644
--- a/cfg/encoder_lowdelay_P_vtm.cfg
+++ b/cfg/encoder_lowdelay_P_vtm.cfg
@@ -133,7 +133,7 @@ MIP : 1
# Fast tools
PBIntraFast : 1
-ISPFast : 1
+ISPFast : 0
FastMrg : 1
AMaxBT : 1
FastMIP : 0
diff --git a/cfg/encoder_lowdelay_vtm.cfg b/cfg/encoder_lowdelay_vtm.cfg
index 148f03230..15864d6a3 100644
--- a/cfg/encoder_lowdelay_vtm.cfg
+++ b/cfg/encoder_lowdelay_vtm.cfg
@@ -138,7 +138,7 @@ PROF : 1
# Fast tools
PBIntraFast : 1
-ISPFast : 1
+ISPFast : 0
FastMrg : 1
AMaxBT : 1
FastMIP : 0
diff --git a/cfg/encoder_randomaccess_vtm.cfg b/cfg/encoder_randomaccess_vtm.cfg
index 2e4c7a033..58a9f20fd 100644
--- a/cfg/encoder_randomaccess_vtm.cfg
+++ b/cfg/encoder_randomaccess_vtm.cfg
@@ -156,7 +156,7 @@ PROF : 1
# Fast tools
PBIntraFast : 1
-ISPFast : 1
+ISPFast : 0
FastMrg : 1
AMaxBT : 1
FastMIP : 0
diff --git a/source/Lib/CommonLib/IntraPrediction.cpp b/source/Lib/CommonLib/IntraPrediction.cpp
index f9a5914b3..ebd48f18f 100644
--- a/source/Lib/CommonLib/IntraPrediction.cpp
+++ b/source/Lib/CommonLib/IntraPrediction.cpp
@@ -328,8 +328,12 @@ void IntraPrediction::predIntraAng( const ComponentID compId, PelBuf &piPred, co
const int srcStride = m_topRefLength + 1 + (whRatio + 1) * multiRefIdx;
const int srcHStride = m_leftRefLength + 1 + (hwRatio + 1) * multiRefIdx;
#endif
-
+
+#if JVET_O0502_ISP_CLEANUP
+ const CPelBuf& srcBuf = pu.cu->ispMode && isLuma(compID) ? getISPBuffer() : CPelBuf(getPredictorPtr(compID), srcStride, srcHStride);
+#else
const CPelBuf & srcBuf = CPelBuf(getPredictorPtr(compID), srcStride, srcHStride);
+#endif
const ClpRng& clpRng(pu.cu->cs->slice->clpRng(compID));
switch (uiDirMode)
@@ -519,7 +523,11 @@ void IntraPrediction::initPredIntraParams(const PredictionUnit & pu, const CompA
m_ipaParam.multiRefIndex = isLuma (chType) ? pu.multiRefIdx : 0 ;
m_ipaParam.refFilterFlag = false;
m_ipaParam.interpolationFlag = false;
+#if JVET_O0502_ISP_CLEANUP
+ m_ipaParam.applyPDPC = ((puSize.width >= MIN_TB_SIZEY && puSize.height >= MIN_TB_SIZEY) || !isLuma(compId)) && m_ipaParam.multiRefIndex == 0;
+#else
m_ipaParam.applyPDPC = !useISP && m_ipaParam.multiRefIndex == 0;
+#endif
const int intraPredAngleMode = (m_ipaParam.isModeVer) ? predMode - VER_IDX : -(predMode - HOR_IDX);
@@ -573,10 +581,12 @@ void IntraPrediction::initPredIntraParams(const PredictionUnit & pu, const CompA
|| DC_IDX == dirMode
)
{
+#if !JVET_O0502_ISP_CLEANUP
if (useISP)
{
m_ipaParam.interpolationFlag = (m_ipaParam.isModeVer ? puSize.width : puSize.height) > 8 ? true : false ;
}
+#endif
}
else if (isLuma( chType ) && pu.cu->bdpcmMode) // BDPCM
{
@@ -1065,6 +1075,132 @@ void IntraPrediction::initIntraPatternChType(const CodingUnit &cu, const CompAre
}
}
+#if JVET_O0502_ISP_CLEANUP
+void IntraPrediction::initIntraPatternChTypeISP(const CodingUnit& cu, const CompArea& area, PelBuf& recBuf, const bool forceRefFilterFlag)
+{
+ const CodingStructure& cs = *cu.cs;
+
+ if (!forceRefFilterFlag)
+ {
+ initPredIntraParams(*cu.firstPU, area, *cs.sps);
+ }
+
+ const Position posLT = area;
+ bool isLeftAvail = cs.isDecomp(posLT.offset(-1, 0), CHANNEL_TYPE_LUMA);
+ bool isAboveAvail = cs.isDecomp(posLT.offset(0, -1), CHANNEL_TYPE_LUMA);
+ // ----- Step 1: unfiltered reference samples -----
+#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
+ if (cu.blocks[area.compID].x == area.x && cu.blocks[area.compID].y == area.y)
+#else
+ if (CU::isISPFirst(cu, area, area.compID))
+#endif
+ {
+ Pel* refBufUnfiltered = m_piYuvExt[area.compID][PRED_BUF_UNFILTERED];
+ // With the first subpartition all the CU reference samples are fetched at once in a single call to xFillReferenceSamples
+ if (cu.ispMode == HOR_INTRA_SUBPARTITIONS)
+ {
+ m_leftRefLength = cu.Y().height << 1;
+ m_topRefLength = cu.Y().width + area.width;
+ }
+ else //if (cu.ispMode == VER_INTRA_SUBPARTITIONS)
+ {
+ m_leftRefLength = cu.Y().height + area.height;
+ m_topRefLength = cu.Y().width << 1;
+ }
+
+ const int multiRefIdx = m_ipaParam.multiRefIndex;
+ const int whRatio = m_ipaParam.whRatio;
+ const int hwRatio = m_ipaParam.hwRatio;
+ const int srcStride = m_topRefLength + 1 + (whRatio + 1) * multiRefIdx;
+ const int srcHStride = m_leftRefLength + 1 + (hwRatio + 1) * multiRefIdx;
+ m_pelBufISP[0] = m_pelBufISPBase[0] = PelBuf(m_piYuvExt[area.compID][PRED_BUF_UNFILTERED], srcStride, srcHStride);
+ m_pelBufISP[1] = m_pelBufISPBase[1] = PelBuf(m_piYuvExt[area.compID][PRED_BUF_FILTERED], srcStride, srcHStride);
+
+ xFillReferenceSamples(cs.picture->getRecoBuf(cu.Y()), refBufUnfiltered, cu.Y(), cu);
+
+ // After having retrieved all the CU reference samples, the number of reference samples is now adjusted for the current subpartition
+ m_topRefLength = cu.blocks[area.compID].width + area.width;
+ m_leftRefLength = cu.blocks[area.compID].height + area.height;
+ }
+ else
+ {
+ //Now we only need to fetch the newly available reconstructed samples from the previously coded TU
+ Position tuPos = area;
+ tuPos.relativeTo(cu.Y());
+ m_pelBufISP[0] = m_pelBufISPBase[0].subBuf(tuPos, area.size());
+ m_pelBufISP[1] = m_pelBufISPBase[1].subBuf(tuPos, area.size());
+
+ PelBuf& dstBuf = m_pelBufISP[0];
+
+ m_topRefLength = cu.blocks[area.compID].width + area.width;
+ m_leftRefLength = cu.blocks[area.compID].height + area.height;
+
+ const int predSizeHor = m_topRefLength;
+ const int predSizeVer = m_leftRefLength;
+ if (cu.ispMode == HOR_INTRA_SUBPARTITIONS)
+ {
+ Pel* src = recBuf.bufAt(0, -1);
+ Pel* dst = dstBuf.bufAt(1, 0);
+ for (int i = 0; i < area.width; i++)
+ {
+ dst[i] = src[i];
+ }
+ Pel sample = src[area.width - 1];
+ dst += area.width;
+ for (int i = 0; i < predSizeHor - area.width; i++)
+ {
+ dst[i] = sample;
+ }
+ if (!isLeftAvail) //if left is not avaible, then it is necessary to fetch these samples for each subpartition
+ {
+ Pel* dst = dstBuf.bufAt(0, 0);
+ Pel sample = src[0];
+ for (int i = 0; i < predSizeVer + 1; i++)
+ {
+ *dst = sample;
+ dst += dstBuf.stride;
+ }
+ }
+ }
+ else
+ {
+ Pel* src = recBuf.bufAt(-1, 0);
+ Pel* dst = dstBuf.bufAt(0, 1);
+ for (int i = 0; i < area.height; i++)
+ {
+ *dst = *src;
+ src += recBuf.stride;
+ dst += dstBuf.stride;
+ }
+ Pel sample = src[-recBuf.stride];
+ for (int i = 0; i < predSizeVer - area.height; i++)
+ {
+ *dst = sample;
+ dst += dstBuf.stride;
+ }
+
+ if (!isAboveAvail) //if above is not avaible, then it is necessary to fetch these samples for each subpartition
+ {
+ Pel* dst = dstBuf.bufAt(0, 0);
+ Pel sample = recBuf.at(-1, 0);
+ for (int i = 0; i < predSizeHor + 1; i++)
+ {
+ dst[i] = sample;
+ }
+ }
+ }
+ }
+ // ----- Step 2: filtered reference samples -----
+ if (m_ipaParam.refFilterFlag || forceRefFilterFlag)
+ {
+ Pel* refBufUnfiltered = m_pelBufISP[0].buf;
+ Pel* refBufFiltered = m_pelBufISP[1].buf;
+ xFilterReferenceSamples(refBufUnfiltered, refBufFiltered, area, *cs.sps, cu.firstPU->multiRefIdx, m_pelBufISP[0].stride);
+ }
+}
+#endif
+
+
void IntraPrediction::xFillReferenceSamples( const CPelBuf &recoBuf, Pel* refBufUnfiltered, const CompArea &area, const CodingUnit &cu )
{
const ChannelType chType = toChannelType( area.compID );
@@ -1322,6 +1458,9 @@ void IntraPrediction::xFillReferenceSamples( const CPelBuf &recoBuf, Pel* refBuf
void IntraPrediction::xFilterReferenceSamples( const Pel* refBufUnfiltered, Pel* refBufFiltered, const CompArea &area, const SPS &sps
, int multiRefIdx
+#if JVET_O0502_ISP_CLEANUP
+ , int stride
+#endif
)
{
if (area.compID != COMPONENT_Y)
@@ -1337,7 +1476,11 @@ void IntraPrediction::xFilterReferenceSamples( const Pel* refBufUnfiltered, Pel*
const int predSize = m_topRefLength + (whRatio + 1) * multiRefIdx;
const int predHSize = m_leftRefLength + (hwRatio + 1) * multiRefIdx;
#endif
+#if JVET_O0502_ISP_CLEANUP
+ const int predStride = stride == 0 ? predSize + 1 : stride;
+#else
const int predStride = predSize + 1;
+#endif
diff --git a/source/Lib/CommonLib/IntraPrediction.h b/source/Lib/CommonLib/IntraPrediction.h
index 901eff095..157283e03 100644
--- a/source/Lib/CommonLib/IntraPrediction.h
+++ b/source/Lib/CommonLib/IntraPrediction.h
@@ -68,6 +68,10 @@ class IntraPrediction
private:
Pel* m_piYuvExt[MAX_NUM_COMPONENT][NUM_PRED_BUF];
+#if JVET_O0502_ISP_CLEANUP
+ PelBuf m_pelBufISPBase[2];
+ PelBuf m_pelBufISP[2];
+#endif
int m_iYuvExtSize;
Pel* m_yuvExt2[MAX_NUM_COMPONENT][4];
@@ -129,6 +133,9 @@ protected:
void xFillReferenceSamples ( const CPelBuf &recoBuf, Pel* refBufUnfiltered, const CompArea &area, const CodingUnit &cu );
void xFilterReferenceSamples ( const Pel* refBufUnfiltered, Pel* refBufFiltered, const CompArea &area, const SPS &sps
, int multiRefIdx
+#if JVET_O0502_ISP_CLEANUP
+ , int predStride = 0
+#endif
);
static int getWideAngle ( int width, int height, int predMode );
@@ -152,6 +159,10 @@ public:
void xGetLumaRecPixels(const PredictionUnit &pu, CompArea chromaArea);
/// set parameters from CU data for accessing intra data
void initIntraPatternChType (const CodingUnit &cu, const CompArea &area, const bool forceRefFilterFlag = false); // use forceRefFilterFlag to get both filtered and unfiltered buffers
+#if JVET_O0502_ISP_CLEANUP
+ void initIntraPatternChTypeISP (const CodingUnit& cu, const CompArea& area, PelBuf& piReco, const bool forceRefFilterFlag = false); // use forceRefFilterFlag to get both filtered and unfiltered buffers
+ const PelBuf& getISPBuffer () { return m_pelBufISP[m_ipaParam.refFilterFlag ? PRED_BUF_FILTERED : PRED_BUF_UNFILTERED]; }
+#endif
// Matrix-based intra prediction
void initIntraMip (const PredictionUnit &pu);
diff --git a/source/Lib/CommonLib/TrQuant.cpp b/source/Lib/CommonLib/TrQuant.cpp
index d45791d5d..ddf049944 100644
--- a/source/Lib/CommonLib/TrQuant.cpp
+++ b/source/Lib/CommonLib/TrQuant.cpp
@@ -1056,7 +1056,11 @@ void TrQuant::xQuant(TransformUnit &tu, const ComponentID &compID, const CCoeffB
m_quant->quant( tu, compID, pSrc, uiAbsSum, cQP, ctx );
}
+#if JVET_O0502_ISP_CLEANUP
+void TrQuant::transformNxN( TransformUnit& tu, const ComponentID& compID, const QpParam& cQP, std::vector<TrMode>* trModes, const int maxCand )
+#else
void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, std::vector<TrMode>* trModes, const int maxCand, double* diagRatio, double* horVerRatio )
+#endif
{
CodingStructure &cs = *tu.cs;
const CompArea &rect = tu.blocks[compID];
@@ -1111,9 +1115,11 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const
it++;
}
+#if !JVET_O0502_ISP_CLEANUP
// it gets the distribution of the DCT-II coefficients energy, which will be useful to discard ISP tests
CoeffBuf coeffsDCT( m_mtsCoeffs[0], rect );
xGetCoeffEnergy( tu, compID, coeffsDCT, diagRatio, horVerRatio );
+#endif
int numTests = 0;
std::vector<TrCost>::iterator itC = trCosts.begin();
const double fac = facBB[g_aucLog2[std::max(width, height)]-2];
@@ -1128,7 +1134,11 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const
}
}
+#if JVET_O0502_ISP_CLEANUP
+void TrQuant::transformNxN( TransformUnit& tu, const ComponentID& compID, const QpParam& cQP, TCoeff& uiAbsSum, const Ctx& ctx, const bool loadTr )
+#else
void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, TCoeff &uiAbsSum, const Ctx &ctx, const bool loadTr, double* diagRatio, double* horVerRatio )
+#endif
{
CodingStructure &cs = *tu.cs;
const SPS &sps = *cs.sps;
@@ -1207,6 +1217,7 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const
}
}
+#if !JVET_O0502_ISP_CLEANUP
//we do this only with the DCT-II coefficients
if( isLuma(compID) &&
!loadTr && tu.mtsIdx == MTS_DCT2_DCT2
@@ -1215,6 +1226,7 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const
//it gets the distribution of the coefficients energy, which will be useful to discard ISP tests
xGetCoeffEnergy( tu, compID, tempCoeff, diagRatio, horVerRatio );
}
+#endif
if( sps.getUseLFNST() )
{
@@ -1233,6 +1245,7 @@ void TrQuant::transformNxN( TransformUnit &tu, const ComponentID &compID, const
TU::setCbfAtDepth (tu, compID, tu.depth, uiAbsSum > 0);
}
+#if !JVET_O0502_ISP_CLEANUP
void TrQuant::xGetCoeffEnergy( TransformUnit &tu, const ComponentID &compID, const CoeffBuf& coeffs, double* diagRatio, double* horVerRatio )
{
if( nullptr == diagRatio || nullptr == horVerRatio ) return;
@@ -1266,6 +1279,7 @@ void TrQuant::xGetCoeffEnergy( TransformUnit &tu, const ComponentID &compID, con
*diagRatio = 0 == wdtE && 0 == hgtE && 0 == diaE ? 1 : double( diaE ) / double( wdtE + hgtE );
}
}
+#endif
void TrQuant::applyForwardRDPCM(TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, TCoeff &uiAbsSum, const RDPCMMode &mode)
{
diff --git a/source/Lib/CommonLib/TrQuant.h b/source/Lib/CommonLib/TrQuant.h
index f762386ac..ba4e30823 100644
--- a/source/Lib/CommonLib/TrQuant.h
+++ b/source/Lib/CommonLib/TrQuant.h
@@ -94,9 +94,13 @@ protected:
public:
void invTransformNxN (TransformUnit &tu, const ComponentID &compID, PelBuf &pResi, const QpParam &cQPs);
-
+#if JVET_O0502_ISP_CLEANUP
+ void transformNxN ( TransformUnit& tu, const ComponentID& compID, const QpParam& cQP, std::vector<TrMode>* trModes, const int maxCand );
+ void transformNxN ( TransformUnit& tu, const ComponentID& compID, const QpParam& cQP, TCoeff& uiAbsSum, const Ctx& ctx, const bool loadTr = false );
+#else
void transformNxN ( TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, std::vector<TrMode>* trModes, const int maxCand, double* diagRatio = nullptr, double* horVerRatio = nullptr );
void transformNxN ( TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, TCoeff &uiAbsSum, const Ctx &ctx, const bool loadTr = false, double* diagRatio = nullptr, double* horVerRatio = nullptr );
+#endif
void rdpcmNxN (TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, TCoeff &uiAbsSum, RDPCMMode &rdpcmMode);
void applyForwardRDPCM(TransformUnit &tu, const ComponentID &compID, const QpParam &cQP, TCoeff &uiAbsSum, const RDPCMMode &rdpcmMode);
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 2e3d38e17..2990155da 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -50,6 +50,7 @@
#include <assert.h>
#include <cassert>
+#define JVET_O0502_ISP_CLEANUP 1 // JVET-O0502: Enable PDPC and all 67 intra modes and apply the cubic filter always (also included in JVET-O0341) for ISP
#define JVET_O0070_PROF 1 // JVET-O0070 method 4-2.1a: Prediction refinement with optical flow for affine mode
@@ -484,7 +485,12 @@ enum ISPType
NOT_INTRA_SUBPARTITIONS = 0,
HOR_INTRA_SUBPARTITIONS = 1,
VER_INTRA_SUBPARTITIONS = 2,
+#if JVET_O0502_ISP_CLEANUP
+ NUM_INTRA_SUBPARTITIONS_MODES = 3,
+ INTRA_SUBPARTITIONS_RESERVED = 4
+#else
NUM_INTRA_SUBPARTITIONS_MODES = 3
+#endif
};
enum SbtIdx
diff --git a/source/Lib/CommonLib/UnitTools.cpp b/source/Lib/CommonLib/UnitTools.cpp
index 413d6dacd..1bf6e46ab 100755
--- a/source/Lib/CommonLib/UnitTools.cpp
+++ b/source/Lib/CommonLib/UnitTools.cpp
@@ -282,6 +282,7 @@ bool CU::divideTuInRows( const CodingUnit &cu )
return cu.ispMode == HOR_INTRA_SUBPARTITIONS ? true : false;
}
+#if !JVET_O0502_ISP_CLEANUP
bool CU::firstTestISPHorSplit( const int width, const int height, const ComponentID compID, const CodingUnit *cuLeft, const CodingUnit *cuAbove )
{
//this function decides which split mode (horizontal or vertical) is tested first (encoder only)
@@ -351,6 +352,7 @@ bool CU::firstTestISPHorSplit( const int width, const int height, const Componen
return true;
}
}
+#endif
PartSplit CU::getISPType( const CodingUnit &cu, const ComponentID compID )
{
@@ -432,6 +434,31 @@ uint32_t CU::getISPSplitDim( const int width, const int height, const PartSplit
return partitionSize;
}
+#if JVET_O0502_ISP_CLEANUP
+bool CU::allLumaCBfsAreZero(const CodingUnit& cu)
+{
+ if (!cu.ispMode)
+ {
+ return TU::getCbf(*cu.firstTU, COMPONENT_Y) == false;
+ }
+ else
+ {
+ int numTotalTUs = cu.ispMode == HOR_INTRA_SUBPARTITIONS ? cu.lheight() >> g_aucLog2[cu.firstTU->lheight()] : cu.lwidth() >> g_aucLog2[cu.firstTU->lwidth()];
+ //bool allZeroCbfs = false;
+ TransformUnit* tuPtr = cu.firstTU;
+ for (int tuIdx = 0; tuIdx < numTotalTUs; tuIdx++)
+ {
+ if (TU::getCbf(*tuPtr, COMPONENT_Y) == true)
+ {
+ return false;
+ }
+ tuPtr = tuPtr->next;
+ }
+ return true;
+ }
+}
+#endif
+
PUTraverser CU::traversePUs( CodingUnit& cu )
{
diff --git a/source/Lib/CommonLib/UnitTools.h b/source/Lib/CommonLib/UnitTools.h
index 0e8f46b1a..05b03ea61 100644
--- a/source/Lib/CommonLib/UnitTools.h
+++ b/source/Lib/CommonLib/UnitTools.h
@@ -96,13 +96,18 @@ namespace CU
bool divideTuInRows ( const CodingUnit &cu );
+#if !JVET_O0502_ISP_CLEANUP
bool firstTestISPHorSplit ( const int width, const int height, const ComponentID compID, const CodingUnit *cuLeft = nullptr, const CodingUnit *cuAbove = nullptr );
+#endif
PartSplit getISPType ( const CodingUnit &cu, const ComponentID compID );
bool isISPLast ( const CodingUnit &cu, const CompArea &tuArea, const ComponentID compID );
bool isISPFirst ( const CodingUnit &cu, const CompArea &tuArea, const ComponentID compID );
bool canUseISP ( const CodingUnit &cu, const ComponentID compID );
bool canUseISP ( const int width, const int height, const int maxTrSize = MAX_TB_SIZEY );
uint32_t getISPSplitDim ( const int width, const int height, const PartSplit ispType );
+#if JVET_O0502_ISP_CLEANUP
+ bool allLumaCBfsAreZero ( const CodingUnit& cu );
+#endif
PUTraverser traversePUs ( CodingUnit& cu);
TUTraverser traverseTUs ( CodingUnit& cu);
diff --git a/source/Lib/DecoderLib/CABACReader.cpp b/source/Lib/DecoderLib/CABACReader.cpp
index 296ecdf77..64f2393ff 100755
--- a/source/Lib/DecoderLib/CABACReader.cpp
+++ b/source/Lib/DecoderLib/CABACReader.cpp
@@ -1269,7 +1269,11 @@ void CABACReader::intra_luma_pred_modes( CodingUnit &cu )
for( int k = 0; k < numBlocks; k++ )
{
CHECK(numBlocks != 1, "not supported yet");
+#if JVET_O0502_ISP_CLEANUP
+ if ( cu.firstPU->multiRefIdx )
+#else
if( cu.firstPU->multiRefIdx || ( cu.ispMode && isLuma( cu.chType ) ) )
+#endif
{
mpmFlag[0] = true;
}
diff --git a/source/Lib/DecoderLib/DecCu.cpp b/source/Lib/DecoderLib/DecCu.cpp
index ba1571a52..998eb3394 100644
--- a/source/Lib/DecoderLib/DecCu.cpp
+++ b/source/Lib/DecoderLib/DecCu.cpp
@@ -204,8 +204,39 @@ void DecCu::xIntraRecBlk( TransformUnit& tu, const ComponentID compID )
const PredictionUnit &pu = *tu.cs->getPU( area.pos(), chType );
const uint32_t uiChFinalMode = PU::getFinalIntraMode( pu, chType );
+#if JVET_O0502_ISP_CLEANUP
+ PelBuf pReco = cs.getRecoBuf(area);
+#endif
//===== init availability pattern =====
+#if JVET_O0502_ISP_CLEANUP
+#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
+ bool predRegDiffFromTB = CU::isPredRegDiffFromTB(*tu.cu, compID);
+ bool firstTBInPredReg = CU::isFirstTBInPredReg(*tu.cu, compID, area);
+ CompArea areaPredReg(COMPONENT_Y, tu.chromaFormat, area);
+#endif
+ if (tu.cu->ispMode && isLuma(compID))
+ {
+#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
+ if (predRegDiffFromTB)
+ {
+ if (firstTBInPredReg)
+ {
+ CU::adjustPredArea(areaPredReg);
+ m_pcIntraPred->initIntraPatternChTypeISP(*tu.cu, areaPredReg, pReco);
+ }
+ }
+ else
+#endif
+ {
+ m_pcIntraPred->initIntraPatternChTypeISP(*tu.cu, area, pReco);
+ }
+ }
+ else
+ {
+ m_pcIntraPred->initIntraPatternChType(*tu.cu, area);
+ }
+#else
#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
bool predRegDiffFromTB = CU::isPredRegDiffFromTB(*tu.cu, compID);
bool firstTBInPredReg = CU::isFirstTBInPredReg (*tu.cu, compID, area);
@@ -221,6 +252,7 @@ void DecCu::xIntraRecBlk( TransformUnit& tu, const ComponentID compID )
else
#endif
m_pcIntraPred->initIntraPatternChType(*tu.cu, area);
+#endif
//===== get prediction signal =====
if( compID != COMPONENT_Y && PU::isLMCMode( uiChFinalMode ) )
@@ -323,7 +355,9 @@ void DecCu::xIntraRecBlk( TransformUnit& tu, const ComponentID compID )
CrossComponentPrediction::crossComponentPrediction( tu, compID, cs.getResiBuf( tu.Y() ), piResi, piResi, true );
}
+#if !JVET_O0502_ISP_CLEANUP
PelBuf pReco = cs.getRecoBuf( area );
+#endif
if( !tu.cu->ispMode || !isLuma( compID ) )
{
diff --git a/source/Lib/EncoderLib/CABACWriter.cpp b/source/Lib/EncoderLib/CABACWriter.cpp
index 188e7d6c3..0b14b65cc 100755
--- a/source/Lib/EncoderLib/CABACWriter.cpp
+++ b/source/Lib/EncoderLib/CABACWriter.cpp
@@ -1054,7 +1054,11 @@ void CABACWriter::intra_luma_pred_modes( const CodingUnit& cu )
break;
}
}
+#if JVET_O0502_ISP_CLEANUP
+ if ( pu->multiRefIdx )
+#else
if( pu->multiRefIdx || ( cu.ispMode && isLuma( cu.chType ) ) )
+#endif
{
CHECK(mpm_idx >= numMPMs, "use of non-MPM");
}
@@ -1154,7 +1158,11 @@ void CABACWriter::intra_luma_pred_mode( const PredictionUnit& pu )
break;
}
}
+#if JVET_O0502_ISP_CLEANUP
+ if ( pu.multiRefIdx )
+#else
if( pu.multiRefIdx || ( pu.cu->ispMode && isLuma( pu.cu->chType ) ) )
+#endif
{
CHECK(mpm_idx >= numMPMs, "use of non-MPM");
}
diff --git a/source/Lib/EncoderLib/EncCu.cpp b/source/Lib/EncoderLib/EncCu.cpp
index fe5b1633d..140adf9fc 100644
--- a/source/Lib/EncoderLib/EncCu.cpp
+++ b/source/Lib/EncoderLib/EncCu.cpp
@@ -572,8 +572,15 @@ bool EncCu::xCheckBestMode( CodingStructure *&tempCS, CodingStructure *&bestCS,
}
+#if JVET_O0502_ISP_CLEANUP
+void EncCu::xCompressCU( CodingStructure*& tempCS, CodingStructure*& bestCS, Partitioner& partitioner, double maxCostAllowed )
+#else
void EncCu::xCompressCU( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner )
+#endif
{
+#if JVET_O0502_ISP_CLEANUP
+ CHECK(maxCostAllowed < 0, "Wrong value of maxCostAllowed!");
+#endif
if (m_shareState == NO_SHARE)
{
tempCS->sharedBndPos = tempCS->area.Y().lumaPos();
@@ -660,6 +667,9 @@ void EncCu::xCompressCU( CodingStructure *&tempCS, CodingStructure *&bestCS, Par
do
{
EncTestMode currTestMode = m_modeCtrl->currTestMode();
+#if JVET_O0502_ISP_CLEANUP
+ currTestMode.maxCostAllowed = maxCostAllowed;
+#endif
if (pps.getUseDQP() && CS::isDualITree(*tempCS) && isChroma(partitioner.chType))
{
@@ -1246,7 +1256,13 @@ void EncCu::xCheckModeSplit(CodingStructure *&tempCS, CodingStructure *&bestCS,
#if JVET_O0070_PROF
tempSubCS->bestParent = bestSubCS->bestParent = bestCS;
#endif
+#if JVET_O0502_ISP_CLEANUP
+ double newMaxCostAllowed = isLuma(partitioner.chType) ? std::min(encTestMode.maxCostAllowed, bestCS->cost - m_pcRdCost->calcRdCost(tempCS->fracBits, tempCS->dist)) : MAX_DOUBLE;
+ newMaxCostAllowed = std::max(0.0, newMaxCostAllowed);
+ xCompressCU(tempSubCS, bestSubCS, partitioner, newMaxCostAllowed);
+#else
xCompressCU( tempSubCS, bestSubCS, partitioner );
+#endif
#if JVET_O0070_PROF
tempSubCS->bestParent = bestSubCS->bestParent = nullptr;
#endif
@@ -1464,8 +1480,17 @@ void EncCu::xCheckRDCostIntra( CodingStructure *&tempCS, CodingStructure *&bestC
bool validCandRet = false;
if( isLuma( partitioner.chType ) )
{
+#if JVET_O0502_ISP_CLEANUP
+ //ISP uses the value of the best cost so far (luma if it is the fast version) to avoid test non-necessary subpartitions
+ double bestCostSoFar = CS::isDualITree(*tempCS) ? m_modeCtrl->getBestCostWithoutSplitFlags() : bestCU && bestCU->predMode == MODE_INTRA ? bestCS->lumaCost : bestCS->cost;
+ if (CS::isDualITree(*tempCS) && encTestMode.maxCostAllowed < bestCostSoFar)
+ {
+ bestCostSoFar = encTestMode.maxCostAllowed;
+ }
+#else
//the Intra SubPartitions mode uses the value of the best cost so far (luma if it is the fast version) to avoid test non-necessary lines
const double bestCostSoFar = CS::isDualITree( *tempCS ) ? m_modeCtrl->getBestCostWithoutSplitFlags() : bestCU && bestCU->predMode == MODE_INTRA ? bestCS->lumaCost : bestCS->cost;
+#endif
validCandRet = m_pcIntraSearch->estIntraPredLumaQT( cu, partitioner, bestCostSoFar, mtsFlag, startMTSIdx[ trGrpIdx ], endMTSIdx[ trGrpIdx ], ( trGrpIdx > 0 ) );
if( sps.getUseLFNST() && ( !validCandRet || ( cu.ispMode && cu.firstTU->cbf[ COMPONENT_Y ] == 0 ) ) )
{
@@ -1626,13 +1651,22 @@ void EncCu::xCheckRDCostIntra( CodingStructure *&tempCS, CodingStructure *&bestC
}
}
+#if JVET_O0502_ISP_CLEANUP
+ //we decide to skip the non-DCT-II transforms and LFNST according to the ISP results
+ if ((endMtsFlag > 0 || endLfnstIdx > 0) && cu.ispMode && !mtsFlag && !lfnstIdx && tempCS->slice->isIntra() && m_pcEncCfg->getUseFastISP())
+#else
//we decide to skip the second emt pass or not according to the ISP results
if( considerMtsSecondPass && cu.ispMode && !mtsFlag && tempCS->slice->isIntra() )
+#endif
{
double bestCostDct2NoIsp = m_modeCtrl->getMtsFirstPassNoIspCost();
CHECKD( bestCostDct2NoIsp <= bestIspCost, "wrong cost!" );
+#if JVET_O0502_ISP_CLEANUP
+ double threshold = 1.4;
+#else
double nSamples = ( double ) ( cu.lwidth() << g_aucLog2[ cu.lheight() ] );
double threshold = 1 + 1.4 / sqrt( nSamples );
+#endif
double lfnstThreshold = 1.01 * threshold;
if( bestCostDct2NoIsp > bestIspCost*lfnstThreshold )
diff --git a/source/Lib/EncoderLib/EncCu.h b/source/Lib/EncoderLib/EncCu.h
index 994992656..9ea4376c5 100644
--- a/source/Lib/EncoderLib/EncCu.h
+++ b/source/Lib/EncoderLib/EncCu.h
@@ -175,7 +175,11 @@ protected:
void xCalDebCost ( CodingStructure &cs, Partitioner &partitioner, bool calDist = false );
Distortion getDistortionDb ( CodingStructure &cs, CPelBuf org, CPelBuf reco, ComponentID compID, const CompArea& compArea, bool afterDb );
+#if JVET_O0502_ISP_CLEANUP
+ void xCompressCU ( CodingStructure*& tempCS, CodingStructure*& bestCS, Partitioner& pm, double maxCostAllowed = MAX_DOUBLE );
+#else
void xCompressCU ( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm );
+#endif
#if ENABLE_SPLIT_PARALLELISM
void xCompressCUParallel ( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm );
void copyState ( EncCu* other, Partitioner& pm, const UnitArea& currArea, const bool isDist );
diff --git a/source/Lib/EncoderLib/EncModeCtrl.h b/source/Lib/EncoderLib/EncModeCtrl.h
index d29d095c4..a690dc538 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.h
+++ b/source/Lib/EncoderLib/EncModeCtrl.h
@@ -112,6 +112,9 @@ struct EncTestMode
EncTestModeOpts opts;
int qp;
bool lossless;
+#if JVET_O0502_ISP_CLEANUP
+ double maxCostAllowed;
+#endif
};
diff --git a/source/Lib/EncoderLib/IntraSearch.cpp b/source/Lib/EncoderLib/IntraSearch.cpp
index 31956e602..e90c5e545 100644
--- a/source/Lib/EncoderLib/IntraSearch.cpp
+++ b/source/Lib/EncoderLib/IntraSearch.cpp
@@ -323,6 +323,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
#else
bool testISP = sps.getUseISP() && cu.mtsFlag == 0 && cu.lfnstIdx == 0 && CU::canUseISP( width, height, MAX_TB_SIZEY );
#endif
+#if !JVET_O0502_ISP_CLEANUP
bool ispHorIsFirstTest = testISP ? CU::firstTestISPHorSplit( width, height, COMPONENT_Y, nullptr, nullptr ) : true;
int ispOptions[] = { NOT_INTRA_SUBPARTITIONS, HOR_INTRA_SUBPARTITIONS, VER_INTRA_SUBPARTITIONS };
if ( !ispHorIsFirstTest )
@@ -330,8 +331,19 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
ispOptions[1] = VER_INTRA_SUBPARTITIONS;
ispOptions[2] = HOR_INTRA_SUBPARTITIONS;
}
+#endif
if( testISP )
{
+#if JVET_O0502_ISP_CLEANUP
+ //reset the variables used for the tests
+ m_ispCandListHor.clear();
+ m_ispCandListVer.clear();
+ m_regIntraRDListWithCosts.clear();
+ m_ispTestedModes.clear();
+ //save the number of subpartitions
+ m_ispTestedModes.numTotalParts[0] = (int)height >> g_aucLog2[CU::getISPSplitDim(width, height, TU_1D_HORZ_SPLIT)];
+ m_ispTestedModes.numTotalParts[1] = (int)width >> g_aucLog2[CU::getISPSplitDim(width, height, TU_1D_VERT_SPLIT)];
+#else
//variables for the full RD list without MRL modes
m_rdModeListWithoutMrl .clear();
m_rdModeListWithoutMrlHor .clear();
@@ -340,6 +352,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
m_intraModeDiagRatio .clear();
m_intraModeHorVerRatio .clear();
m_intraModeTestedNormalIntra.clear();
+#endif
}
#if JVET_O1136_TS_BDPCM_SIGNALLING
@@ -557,8 +570,13 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
if ( testISP )
{
+#if JVET_O0502_ISP_CLEANUP
+ // we save the regular intra modes list
+ m_ispCandListHor = uiRdModeList;
+#else
//we save the list with no mrl modes to keep only the Hadamard selected modes (no mpms)
m_rdModeListWithoutMrl = uiRdModeList;
+#endif
}
#if ENABLE_JVET_L0283_MRL
pu.multiRefIdx = 1;
@@ -671,6 +689,23 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
if ( testISP )
{
+#if JVET_O0502_ISP_CLEANUP
+ // we add the MPMs to list that contains only regular intra modes
+ for (int j = 0; j < numCand; j++)
+ {
+ bool mostProbableModeIncluded = false;
+ ModeInfo mostProbableMode(false, 0, NOT_INTRA_SUBPARTITIONS, uiPreds[j]);
+
+ for (int i = 0; i < m_ispCandListHor.size(); i++)
+ {
+ mostProbableModeIncluded |= (mostProbableMode == m_ispCandListHor[i]);
+ }
+ if (!mostProbableModeIncluded)
+ {
+ m_ispCandListHor.push_back(mostProbableMode);
+ }
+ }
+#else
//we add the ISP MPMs to the list without mrl modes
m_rdModeListWithoutMrlHor = m_rdModeListWithoutMrl;
m_rdModeListWithoutMrlVer = m_rdModeListWithoutMrl;
@@ -701,6 +736,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
}
cu.ispMode = NOT_INTRA_SUBPARTITIONS;
+#endif
}
}
//*** Add MPMs for MIP to candidate list
@@ -765,6 +801,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
}
+#if !JVET_O0502_ISP_CLEANUP
if( testISP ) // we remove the non-MPMs from the ISP lists
{
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> uiRdModeListCopyHor = m_rdModeListWithoutMrlHor;
@@ -800,6 +837,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
cu.ispMode = NOT_INTRA_SUBPARTITIONS;
}
+#endif
CHECK( numModesForFullRD != uiRdModeList.size(), "Inconsistent state!" );
@@ -821,8 +859,12 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
uiRdModeList.resize(std::min<size_t>(uiRdModeList.size(), maxSize));
if ( testISP )
{
+#if JVET_O0502_ISP_CLEANUP
+ m_ispCandListHor.resize(std::min<size_t>(m_ispCandListHor.size(), maxSize));
+#else
m_rdModeListWithoutMrlHor.resize(std::min<size_t>(m_rdModeListWithoutMrlHor.size(), maxSize));
m_rdModeListWithoutMrlVer.resize(std::min<size_t>(m_rdModeListWithoutMrlVer.size(), maxSize));
+#endif
}
}
if (maxSize == 0)
@@ -842,8 +884,18 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
}
+#if JVET_O0502_ISP_CLEANUP
+ int numNonISPModes = (int)uiRdModeList.size();
+#endif
+
if ( testISP )
{
+#if JVET_O0502_ISP_CLEANUP
+ // we reserve positions for ISP in the common full RD list
+ const int maxNumRDModesISP = 16;
+ for (int i = 0; i < maxNumRDModesISP; i++)
+ uiRdModeList.push_back(ModeInfo(false, 0, INTRA_SUBPARTITIONS_RESERVED, 0));
+#else
//we create a single full RD list that includes all intra modes using regular intra, MRL and ISP
auto* firstIspList = ispOptions[1] == HOR_INTRA_SUBPARTITIONS ? &m_rdModeListWithoutMrlHor : &m_rdModeListWithoutMrlVer;
auto* secondIspList = ispOptions[1] == HOR_INTRA_SUBPARTITIONS ? &m_rdModeListWithoutMrlVer : &m_rdModeListWithoutMrlHor;
@@ -877,6 +929,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
uiRdModeList.insert( uiRdModeList.end(), secondIspList->begin(), secondIspList->end() );
uiRdModeList.insert( uiRdModeList.end(), firstIspList->begin() , firstIspList->end() );
}
+#endif
}
//===== check modes (using r-d costs) =====
@@ -920,14 +973,18 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
}
// just to be sure
numModesForFullRD = ( int ) uiRdModeList.size();
+#if !JVET_O0502_ISP_CLEANUP
PartSplit intraSubPartitionsProcOrder = TU_NO_ISP;
int bestNormalIntraModeIndex = -1;
+#endif
TUIntraSubPartitioner subTuPartitioner( partitioner );
if( !cu.ispMode && !cu.mtsFlag )
{
m_modeCtrl->setMtsFirstPassNoIspCost( MAX_DOUBLE );
}
+#if !JVET_O0502_ISP_CLEANUP
bool ispHorAllZeroCbfs = false, ispVerAllZeroCbfs = false;
+#endif
for (int mode = -2 * int(testBDPCM); mode < (int)uiRdModeList.size(); mode++)
{
@@ -952,6 +1009,18 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
else
{
cu.bdpcmMode = 0;
+#if JVET_O0502_ISP_CLEANUP
+ if (uiRdModeList[mode].ispMod == INTRA_SUBPARTITIONS_RESERVED)
+ {
+ if (mode == numNonISPModes) // the list needs to be sorted only once
+ {
+ xSortISPCandList(bestCurrentCost, csBest->cost);
+ }
+ xGetNextISPMode(uiRdModeList[mode], (mode > 0 ? &uiRdModeList[mode - 1] : nullptr), Size(width, height));
+ if (uiRdModeList[mode].ispMod == INTRA_SUBPARTITIONS_RESERVED)
+ continue;
+ }
+#endif
uiOrgMode = uiRdModeList[mode];
cu.mipFlag = uiOrgMode.mipFlg;
cu.ispMode = uiOrgMode.ispMod;
@@ -962,6 +1031,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
CHECK(pu.multiRefIdx && (pu.intraDir[0] == PLANAR_IDX), "Error: combination of MRL and Planar mode not supported");
CHECK(cu.ispMode && cu.mipFlag, "Error: combination of ISP and MIP not supported");
CHECK(cu.ispMode && pu.multiRefIdx, "Error: combination of ISP and MRL not supported");
+#if !JVET_O0502_ISP_CLEANUP
if( cu.ispMode )
{
intraSubPartitionsProcOrder = CU::getISPType( cu, COMPONENT_Y );
@@ -979,6 +1049,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
continue;
}
}
+#endif
}
// set context models
@@ -990,8 +1061,23 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
bool tmpValidReturn = false;
if( cu.ispMode )
{
+#if JVET_O0502_ISP_CLEANUP
+ tmpValidReturn = xIntraCodingLumaISP(*csTemp, subTuPartitioner, bestCurrentCost);
+ if (csTemp->tus.size() == 0)
+ {
+ // no TUs were coded
+ csTemp->cost = MAX_DOUBLE;
+ continue;
+ }
+ if (!cu.mtsFlag && !cu.lfnstIdx)
+ {
+ // we save the data for future tests
+ m_ispTestedModes.setModeResults((ISPType)cu.ispMode, (int)uiOrgMode.modeId, (int)csTemp->tus.size(), csTemp->cus[0]->firstTU->cbf[COMPONENT_Y] ? csTemp->cost : MAX_DOUBLE, csBest->cost);
+ }
+#else
tmpValidReturn = xRecurIntraCodingLumaQT( *csTemp, subTuPartitioner, bestCurrentCost, 0, intraSubPartitionsProcOrder, false,
mtsCheckRangeFlag, mtsFirstCheckId, mtsLastCheckId, moreProbMTSIdxFirst );
+#endif
}
else
{
@@ -1003,8 +1089,16 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
mtsCheckRangeFlag, mtsFirstCheckId, mtsLastCheckId, moreProbMTSIdxFirst );
}
+#if JVET_O0502_ISP_CLEANUP
+ if (!cu.ispMode && !cu.mtsFlag && !cu.lfnstIdx && !cu.bdpcmMode && !pu.multiRefIdx && !cu.mipFlag && testISP)
+ {
+ m_regIntraRDListWithCosts.push_back(ModeInfoWithCost(cu.mipFlag, pu.multiRefIdx, cu.ispMode, uiOrgMode.modeId, csTemp->cost));
+ }
+#endif
+
if( cu.ispMode && !csTemp->cus[0]->firstTU->cbf[COMPONENT_Y] )
{
+#if !JVET_O0502_ISP_CLEANUP
if( !sps.getUseLFNST() )
{
if ( cu.ispMode == HOR_INTRA_SUBPARTITIONS )
@@ -1016,6 +1110,7 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
ispVerAllZeroCbfs |= ( m_pcEncCfg->getUseFastISP() && csTemp->tus[0]->lwidth() > 2 && csTemp->cost >= bestCurrentCost );
}
}
+#endif
csTemp->cost = MAX_DOUBLE;
csTemp->costDbOffset = 0;
tmpValidReturn = false;
@@ -1027,8 +1122,13 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
m_modeCostStore[ lfnstIdx ][ testMip ? rdModeIdxList[ mode ] : mode ] = tmpValidReturn ? csTemp->cost : ( MAX_DOUBLE / 2.0 ); //(MAX_DOUBLE / 2.0) ??
}
-
+#if JVET_O0502_ISP_CLEANUP
+ DTRACE(g_trace_ctx, D_INTRA_COST, "IntraCost T [x=%d,y=%d,w=%d,h=%d] %f (%d,%d,%d,%d,%d,%d) \n", cu.blocks[0].x,
+ cu.blocks[0].y, (int)width, (int)height, csTemp->cost, uiOrgMode.modeId, uiOrgMode.ispMod,
+ pu.multiRefIdx, cu.mipFlag, cu.lfnstIdx, cu.mtsFlag);
+#else
DTRACE( g_trace_ctx, D_INTRA_COST, "IntraCost T %f (%d) \n", csTemp->cost, uiOrgMode.modeId );
+#endif
if( tmpValidReturn )
@@ -1056,7 +1156,9 @@ bool IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
if( !cu.ispMode && !cu.bdpcmMode && csBest->cost < bestCostNonBDPCM )
{
bestCostNonBDPCM = csBest->cost;
+#if !JVET_O0502_ISP_CLEANUP
bestNormalIntraModeIndex = mode;
+#endif
}
}
@@ -1802,7 +1904,9 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
const bool bUseCrossCPrediction = pps.getPpsRangeExtension().getCrossComponentPredictionEnabledFlag() && isChroma( compID ) && PU::isChromaIntraModeCrossCheckMode( pu ) && checkCrossCPrediction;
const bool ccUseRecoResi = m_pcEncCfg->getUseReconBasedCrossCPredictionEstimate();
+#if !JVET_O0502_ISP_CLEANUP
const bool ispSplitIsAllowed = sps.getUseISP() && CU::canUseISP( *tu.cu, compID );
+#endif
//===== init availability pattern =====
@@ -1816,6 +1920,32 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
PelBuf sharedPredTS( m_pSharedPredTransformSkip[compID], area );
if( default0Save1Load2 != 2 )
{
+#if JVET_O0502_ISP_CLEANUP
+#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
+ bool predRegDiffFromTB = CU::isPredRegDiffFromTB(*tu.cu, compID);
+ bool firstTBInPredReg = CU::isFirstTBInPredReg(*tu.cu, compID, area);
+ CompArea areaPredReg(COMPONENT_Y, tu.chromaFormat, area);
+#endif
+ if (tu.cu->ispMode && isLuma(compID))
+ {
+#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
+ if (predRegDiffFromTB)
+ {
+ if (firstTBInPredReg)
+ {
+ CU::adjustPredArea(areaPredReg);
+ initIntraPatternChTypeISP(*tu.cu, areaPredReg, piReco);
+ }
+ }
+ else
+#endif
+ initIntraPatternChTypeISP(*tu.cu, area, piReco);
+ }
+ else
+ {
+ initIntraPatternChType(*tu.cu, area);
+ }
+#else
#if JVET_O0106_ISP_4xN_PREDREG_FOR_1xN_2xN
bool predRegDiffFromTB = CU::isPredRegDiffFromTB(*tu.cu, compID);
bool firstTBInPredReg = CU::isFirstTBInPredReg(*tu.cu, compID, area);
@@ -1831,6 +1961,7 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
else
#endif
initIntraPatternChType(*tu.cu, area);
+#endif
//===== get prediction signal =====
if( compID != COMPONENT_Y && PU::isLMCMode( uiChFinalMode ) )
@@ -2001,6 +2132,14 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
if( isLuma(compID) )
{
#endif
+#if JVET_O0502_ISP_CLEANUP
+ if (trModes)
+ {
+ m_pcTrQuant->transformNxN(tu, compID, cQP, trModes, CU::isIntra(*tu.cu) ? m_pcEncCfg->getIntraMTSMaxCand() : m_pcEncCfg->getInterMTSMaxCand());
+ tu.mtsIdx = trModes->at(0).first;
+ }
+ m_pcTrQuant->transformNxN(tu, compID, cQP, uiAbsSum, m_CABACEstimator->getCtx(), loadTr);
+#else
double diagRatio = 0, horVerRatio = 0;
if( trModes )
@@ -2015,10 +2154,20 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
m_intraModeHorVerRatio .push_back(horVerRatio);
m_intraModeTestedNormalIntra.push_back((int)uiChFinalMode);
}
+#endif
DTRACE( g_trace_ctx, D_TU_ABS_SUM, "%d: comp=%d, abssum=%d\n", DTRACE_GET_COUNTER( g_trace_ctx, D_TU_ABS_SUM ), compID, uiAbsSum );
+#if JVET_O0502_ISP_CLEANUP
+ if (tu.cu->ispMode && isLuma(compID) && CU::isISPLast(*tu.cu, area, area.compID) && CU::allLumaCBfsAreZero(*tu.cu))
+ {
+ // ISP has to have at least one non-zero CBF
+ ruiDist = MAX_INT;
+ return;
+ }
+#endif
+
//--- inverse transform ---
if (uiAbsSum > 0)
@@ -2189,6 +2338,118 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
}
}
+#if JVET_O0502_ISP_CLEANUP
+bool IntraSearch::xIntraCodingLumaISP(CodingStructure& cs, Partitioner& partitioner, const double bestCostSoFar)
+{
+ int subTuCounter = 0;
+ const CodingUnit& cu = *cs.getCU(partitioner.currArea().lumaPos(), partitioner.chType);
+ bool earlySkipISP = false;
+ bool uiSplitCbfLuma = false;
+ const PartSplit ispType = CU::getISPType(cu, COMPONENT_Y);
+
+ cs.cost = 0;
+
+ partitioner.splitCurrArea(ispType, cs);
+
+ do // subpartitions loop
+ {
+ uint32_t numSig = 0;
+ Distortion singleDistTmpLuma = 0;
+ uint64_t singleTmpFracBits = 0;
+ double singleCostTmp = 0;
+
+ TransformUnit& tu = cs.addTU(CS::getArea(cs, partitioner.currArea(), partitioner.chType), partitioner.chType);
+ tu.depth = partitioner.currTrDepth;
+
+ // Encode TU
+ xIntraCodingTUBlock(tu, COMPONENT_Y, false, singleDistTmpLuma, 0, &numSig);
+
+ if (singleDistTmpLuma == MAX_INT) // all zero CBF skip
+ {
+ earlySkipISP = true;
+ partitioner.exitCurrSplit();
+ cs.cost = MAX_DOUBLE;
+ return false;
+ }
+
+ {
+ if (m_pcRdCost->calcRdCost(cs.fracBits, cs.dist + singleDistTmpLuma) > bestCostSoFar)
+ {
+ // The accumulated cost + distortion is already larger than the best cost so far, so it is not necessary to calculate the rate
+ earlySkipISP = true;
+ }
+ else
+ {
+ singleTmpFracBits = xGetIntraFracBitsQT(cs, partitioner, true, false, subTuCounter, ispType);
+ }
+ singleCostTmp = m_pcRdCost->calcRdCost(singleTmpFracBits, singleDistTmpLuma);
+ }
+
+ cs.cost += singleCostTmp;
+ cs.dist += singleDistTmpLuma;
+ cs.fracBits += singleTmpFracBits;
+
+ subTuCounter++;
+
+ uiSplitCbfLuma |= TU::getCbfAtDepth(*cs.getTU(partitioner.currArea().lumaPos(), partitioner.chType, subTuCounter - 1), COMPONENT_Y, partitioner.currTrDepth);
+ int nSubPartitions = m_ispTestedModes.numTotalParts[cu.ispMode - 1];
+ if (subTuCounter < nSubPartitions)
+ {
+ // exit condition if the accumulated cost is already larger than the best cost so far (no impact in RD performance)
+ if (cs.cost > bestCostSoFar)
+ {
+ earlySkipISP = true;
+ break;
+ }
+ else if (subTuCounter < nSubPartitions)
+ {
+ // more restrictive exit condition
+ double threshold = nSubPartitions == 2 ? 0.95 : subTuCounter == 1 ? 0.83 : 0.91;
+ if (subTuCounter < nSubPartitions && cs.cost > bestCostSoFar * threshold)
+ {
+ earlySkipISP = true;
+ break;
+ }
+ }
+ }
+ } while (partitioner.nextPart(cs)); // subpartitions loop
+
+ partitioner.exitCurrSplit();
+ const UnitArea& currArea = partitioner.currArea();
+ const uint32_t currDepth = partitioner.currTrDepth;
+
+ if (earlySkipISP)
+ {
+ cs.cost = MAX_DOUBLE;
+ }
+ else
+ {
+ cs.cost = m_pcRdCost->calcRdCost(cs.fracBits, cs.dist);
+ // The cost check is necessary here again to avoid superfluous operations if the maximum number of coded subpartitions was reached and yet ISP did not win
+ if (cs.cost < bestCostSoFar)
+ {
+ cs.setDecomp(cu.Y());
+ cs.picture->getRecoBuf(currArea.Y()).copyFrom(cs.getRecoBuf(currArea.Y()));
+
+ for (auto& ptu : cs.tus)
+ {
+ if (currArea.Y().contains(ptu->Y()))
+ {
+ TU::setCbfAtDepth(*ptu, COMPONENT_Y, currDepth, uiSplitCbfLuma ? 1 : 0);
+ }
+ }
+ }
+ else
+ {
+ cs.cost = MAX_DOUBLE;
+ earlySkipISP = true;
+ }
+ }
+ return !earlySkipISP;
+}
+#endif
+
+
bool IntraSearch::xRecurIntraCodingLumaQT( CodingStructure &cs, Partitioner &partitioner, const double bestCostSoFar, const int subTuIdx, const PartSplit ispType, const bool ispIsCurrentWinner, bool mtsCheckRangeFlag, int mtsFirstCheckId, int mtsLastCheckId, bool moreProbMTSIdxFirst )
{
int subTuCounter = subTuIdx;
@@ -3276,3 +3537,263 @@ void IntraSearch::reduceHadCandList(static_vector<T, N>& candModeList, static_ve
candCostList = tempCandCostList;
numModesForFullRD = int(candModeList.size());
}
+
+#if JVET_O0502_ISP_CLEANUP
+// It decides which modes from the ISP lists can be full RD tested
+void IntraSearch::xGetNextISPMode(ModeInfo& modeInfo, const ModeInfo* lastMode, const Size cuSize)
+{
+ static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM>* rdModeLists[2] = { &m_ispCandListHor, &m_ispCandListVer };
+
+ ISPType nextISPcandSplitType;
+ if (!m_ispTestedModes.stopTestingHorSplit && !m_ispTestedModes.stopTestingVerSplit)
+ {
+ nextISPcandSplitType = !lastMode ? HOR_INTRA_SUBPARTITIONS : lastMode->ispMod == HOR_INTRA_SUBPARTITIONS ? VER_INTRA_SUBPARTITIONS : HOR_INTRA_SUBPARTITIONS;
+ }
+ else if (!m_ispTestedModes.stopTestingHorSplit && m_ispTestedModes.stopTestingVerSplit)
+ {
+ nextISPcandSplitType = HOR_INTRA_SUBPARTITIONS;
+ }
+ else if (m_ispTestedModes.stopTestingHorSplit && !m_ispTestedModes.stopTestingVerSplit)
+ {
+ nextISPcandSplitType = VER_INTRA_SUBPARTITIONS;
+ }
+ else
+ {
+ return; // no more modes will be tested
+ }
+
+ int maxNumSubPartitions = m_ispTestedModes.numTotalParts[nextISPcandSplitType - 1];
+
+ if (m_ispTestedModes.numTestedModes[nextISPcandSplitType - 1] >= 2)
+ {
+ // Split stop criteria after checking the performance of previously tested intra modes
+ const int thresholdSplit1 = maxNumSubPartitions;
+
+ int mode1 = m_ispTestedModes.getTestedIntraMode((ISPType)nextISPcandSplitType, 0);
+ mode1 = mode1 == DC_IDX ? -1 : mode1;
+ int numSubPartsBestMode1 = mode1 != -1 ? m_ispTestedModes.getNumCompletedSubParts((ISPType)nextISPcandSplitType, mode1) : -1;
+ int mode2 = m_ispTestedModes.getTestedIntraMode((ISPType)nextISPcandSplitType, 1);
+ mode2 = mode2 == DC_IDX ? -1 : mode2;
+ int numSubPartsBestMode2 = mode2 != -1 ? m_ispTestedModes.getNumCompletedSubParts((ISPType)nextISPcandSplitType, mode2) : -1;
+
+ // 1) The 2 most promising modes do not reach a certain number of sub-partitions
+ if (numSubPartsBestMode1 != -1 && numSubPartsBestMode2 != -1)
+ {
+ if (numSubPartsBestMode1 < thresholdSplit1 && numSubPartsBestMode2 < thresholdSplit1)
+ {
+ m_ispTestedModes.stopTestingVerSplit = nextISPcandSplitType == VER_INTRA_SUBPARTITIONS ? true : m_ispTestedModes.stopTestingVerSplit;
+ m_ispTestedModes.stopTestingHorSplit = nextISPcandSplitType == HOR_INTRA_SUBPARTITIONS ? true : m_ispTestedModes.stopTestingHorSplit;
+ return;
+ }
+ }
+
+ // 2) One split is better than the other after PLANAR and one angle have been tested
+ ISPType otherSplit = nextISPcandSplitType == HOR_INTRA_SUBPARTITIONS ? VER_INTRA_SUBPARTITIONS : HOR_INTRA_SUBPARTITIONS;
+ int numSubPartsBestAngleOtherSplit = mode2 != -1 ? m_ispTestedModes.getNumCompletedSubParts(otherSplit, mode2) : -1;
+ bool stopThisSplit = false;
+ if (numSubPartsBestAngleOtherSplit != -1 && numSubPartsBestMode2 != -1)
+ {
+ if (numSubPartsBestAngleOtherSplit > numSubPartsBestMode2)
+ {
+ stopThisSplit = true;
+ }
+ else if (numSubPartsBestAngleOtherSplit == numSubPartsBestMode2 && numSubPartsBestAngleOtherSplit == maxNumSubPartitions)
+ {
+ double rdCostBestAngleThisSplit = m_ispTestedModes.getRDCost(nextISPcandSplitType, mode2, maxNumSubPartitions);
+ double rdCostBestAngleOtherSplit = m_ispTestedModes.getRDCost(otherSplit, mode2, maxNumSubPartitions);
+
+ if (rdCostBestAngleThisSplit == MAX_DOUBLE || rdCostBestAngleOtherSplit < rdCostBestAngleThisSplit * 1.3)
+ {
+ stopThisSplit = true;
+ }
+ }
+ }
+ if (stopThisSplit)
+ {
+ m_ispTestedModes.stopTestingVerSplit = nextISPcandSplitType == VER_INTRA_SUBPARTITIONS ? true : m_ispTestedModes.stopTestingVerSplit;
+ m_ispTestedModes.stopTestingHorSplit = nextISPcandSplitType == HOR_INTRA_SUBPARTITIONS ? true : m_ispTestedModes.stopTestingHorSplit;
+ return;
+ }
+ }
+
+ // Now a new mode is retrieved from the list and it has to be decided whether it should be tested or not
+ if (m_ispTestedModes.candIndexInList[nextISPcandSplitType - 1] < rdModeLists[nextISPcandSplitType - 1]->size())
+ {
+ ModeInfo candidate = rdModeLists[nextISPcandSplitType - 1]->at(m_ispTestedModes.candIndexInList[nextISPcandSplitType - 1]);
+ m_ispTestedModes.candIndexInList[nextISPcandSplitType - 1]++;
+
+ // extra modes are only tested if ISP has won so far
+ if (m_ispTestedModes.candIndexInList[nextISPcandSplitType - 1] > m_ispTestedModes.numOrigModesToTest)
+ {
+ if (m_ispTestedModes.bestSplitSoFar != candidate.ispMod || m_ispTestedModes.bestModeSoFar == PLANAR_IDX)
+ {
+ return;
+ }
+ }
+
+ bool testCandidate = true;
+
+ // we look for a reference mode that has already been tested within the window and decide to test the new one according to the reference mode costs
+ if (candidate.modeId >= DC_IDX && maxNumSubPartitions > 2 && m_ispTestedModes.numTestedModes[nextISPcandSplitType - 1] >= 2)
+ {
+ const int angWindowSize = 5;
+ int numSubPartsLeftMode, numSubPartsRightMode, numSubPartsRefMode, leftIntraMode = -1, rightIntraMode = -1;
+ int windowSize = candidate.modeId > DC_IDX ? angWindowSize : 1;
+ int numSamples = cuSize.width << g_aucLog2[cuSize.height];
+ int numSubPartsLimit = numSamples >= 256 ? maxNumSubPartitions - 1 : 2;
+
+ xFindAlreadyTestedNearbyIntraModes((int)candidate.modeId, &leftIntraMode, &rightIntraMode, (ISPType)candidate.ispMod, windowSize);
+
+ numSubPartsLeftMode = leftIntraMode != -1 ? m_ispTestedModes.getNumCompletedSubParts((ISPType)candidate.ispMod, leftIntraMode) : -1;
+ numSubPartsRightMode = rightIntraMode != -1 ? m_ispTestedModes.getNumCompletedSubParts((ISPType)candidate.ispMod, rightIntraMode) : -1;
+
+ numSubPartsRefMode = std::max(numSubPartsLeftMode, numSubPartsRightMode);
+
+ if (numSubPartsRefMode > 0)
+ {
+ // The mode was found. Now we check the condition
+ testCandidate = numSubPartsRefMode > numSubPartsLimit;
+ }
+ }
+
+ if (testCandidate)
+ {
+ modeInfo = candidate;
+ }
+ }
+}
+
+void IntraSearch::xFindAlreadyTestedNearbyIntraModes(int currentIntraMode, int* leftIntraMode, int* rightIntraMode, ISPType ispOption, int windowSize)
+{
+ bool leftModeFound = false, rightModeFound = false;
+ *leftIntraMode = -1;
+ *rightIntraMode = -1;
+ const unsigned st = ispOption - 1;
+
+ for (int k = 1; k <= windowSize; k++)
+ {
+ int off = currentIntraMode - 2 - k;
+ int leftMode = (off < 0) ? NUM_LUMA_MODE + off : currentIntraMode - k;
+ int rightMode = currentIntraMode > DC_IDX ? (((int)currentIntraMode - 2 + k) % 65) + 2 : PLANAR_IDX;
+
+ leftModeFound = leftMode != (int)currentIntraMode ? m_ispTestedModes.modeHasBeenTested[leftMode][st] : false;
+ rightModeFound = rightMode != (int)currentIntraMode ? m_ispTestedModes.modeHasBeenTested[rightMode][st] : false;
+ if (leftModeFound || rightModeFound)
+ {
+ *leftIntraMode = leftModeFound ? leftMode : -1;
+ *rightIntraMode = rightModeFound ? rightMode : -1;
+ break;
+ }
+ }
+}
+
+void IntraSearch::xSortISPCandList(double bestCostSoFar, double bestNonISPCost)
+{
+ if (m_pcEncCfg->getUseFastISP())
+ {
+ double thSkipISP = 1.4;
+ if (bestNonISPCost > bestCostSoFar * thSkipISP)
+ {
+ m_ispTestedModes.stopTestingHorSplit = true;
+ m_ispTestedModes.stopTestingVerSplit = true;
+ return;
+ }
+ }
+
+ for (int k = 0; k < m_ispCandListHor.size(); k++)
+ {
+ m_ispCandListHor.at(k).ispMod = HOR_INTRA_SUBPARTITIONS; //we set the correct ISP split type value
+ }
+
+ auto origHadList = m_ispCandListHor; // save the original hadamard list of regular intra
+ bool modeIsInList[NUM_LUMA_MODE] = { false };
+
+ m_ispCandListHor.clear();
+ m_ispCandListVer.clear();
+
+ // we sort the normal intra modes according to their full RD costs
+ std::sort(m_regIntraRDListWithCosts.begin(), m_regIntraRDListWithCosts.end(), ModeInfoWithCost::compareModeInfoWithCost);
+
+ // we get the best angle from the regular intra list
+ int bestNormalIntraAngle = -1;
+ for (int modeIdx = 0; modeIdx < m_regIntraRDListWithCosts.size(); modeIdx++)
+ {
+ if (bestNormalIntraAngle == -1 && m_regIntraRDListWithCosts.at(modeIdx).modeId > DC_IDX)
+ {
+ bestNormalIntraAngle = m_regIntraRDListWithCosts.at(modeIdx).modeId;
+ break;
+ }
+ }
+
+ int mode1 = PLANAR_IDX;
+ int mode2 = bestNormalIntraAngle;
+
+ ModeInfo refMode = origHadList.at(0);
+ auto* destListPtr = &m_ispCandListHor;
+ // 1) Planar
+ destListPtr->push_back(ModeInfo(refMode.mipFlg, refMode.mRefId, refMode.ispMod, mode1));
+ modeIsInList[mode1] = true;
+ // 2) Best angle in regular intra
+ if (mode2 != -1)
+ {
+ destListPtr->push_back(ModeInfo(refMode.mipFlg, refMode.mRefId, refMode.ispMod, mode2));
+ modeIsInList[mode2] = true;
+ }
+ // 3) Remaining regular intra modes that were full RD tested (except DC, which is added after the angles from regular intra)
+ int dcModeIndex = -1;
+ for (int remModeIdx = 0; remModeIdx < m_regIntraRDListWithCosts.size(); remModeIdx++)
+ {
+ int currentMode = m_regIntraRDListWithCosts.at(remModeIdx).modeId;
+ if (currentMode != mode1 && currentMode != mode2)
+ {
+ if (currentMode > DC_IDX)
+ {
+ destListPtr->push_back(ModeInfo(refMode.mipFlg, refMode.mRefId, refMode.ispMod, currentMode));
+ modeIsInList[currentMode] = true;
+ }
+ else if (currentMode == DC_IDX)
+ {
+ dcModeIndex = remModeIdx;
+ }
+ }
+ }
+ // 4) DC is added after the angles from regular intra
+ if (dcModeIndex != -1)
+ {
+ destListPtr->push_back(ModeInfo(refMode.mipFlg, refMode.mRefId, refMode.ispMod, DC_IDX));
+ modeIsInList[DC_IDX] = true;
+ }
+
+ // 5) We add extra candidates to the list that will only be tested is likely to win
+ m_ispTestedModes.numOrigModesToTest = (int)destListPtr->size();
+ const int addedModesFromHadList = 3;
+ int newModesAdded = 0;
+
+ for (int k = 0; k < origHadList.size(); k++)
+ {
+ if (newModesAdded == addedModesFromHadList)
+ {
+ break;
+ }
+ if (!modeIsInList[origHadList.at(k).modeId])
+ {
+ destListPtr->push_back(ModeInfo(refMode.mipFlg, refMode.mRefId, refMode.ispMod, origHadList.at(k).modeId));
+ newModesAdded++;
+ }
+ }
+
+ // Copy modes to other split-type list
+ m_ispCandListVer = m_ispCandListHor;
+ for (int i = 0; i < m_ispCandListVer.size(); i++)
+ {
+ m_ispCandListVer[i].ispMod = VER_INTRA_SUBPARTITIONS;
+ }
+
+ // Reset the tested modes information to 0
+ for (int i = 0; i < m_ispCandListHor.size(); i++)
+ {
+ m_ispTestedModes.clearISPModeInfo(m_ispCandListHor[i].modeId);
+ }
+}
+
+#endif
diff --git a/source/Lib/EncoderLib/IntraSearch.h b/source/Lib/EncoderLib/IntraSearch.h
index fd005a552..d038d7310 100644
--- a/source/Lib/EncoderLib/IntraSearch.h
+++ b/source/Lib/EncoderLib/IntraSearch.h
@@ -87,10 +87,131 @@ private:
ModeInfo(const bool mipf, const int mrid, const uint8_t ispm, const uint32_t mode) : mipFlg(mipf), mRefId(mrid), ispMod(ispm), modeId(mode) {}
bool operator==(const ModeInfo cmp) const { return (mipFlg == cmp.mipFlg && mRefId == cmp.mRefId && ispMod == cmp.ispMod && modeId == cmp.modeId); }
};
+#if JVET_O0502_ISP_CLEANUP
+ struct ModeInfoWithCost : public ModeInfo
+ {
+ double rdCost;
+ ModeInfoWithCost() : ModeInfo(), rdCost(MAX_DOUBLE) {}
+ ModeInfoWithCost(const bool mipf, const int mrid, const uint8_t ispm, const uint32_t mode, double cost) : ModeInfo(mipf, mrid, ispm, mode), rdCost(cost) {}
+ bool operator==(const ModeInfoWithCost cmp) const { return (mipFlg == cmp.mipFlg && mRefId == cmp.mRefId && ispMod == cmp.ispMod && modeId == cmp.modeId && rdCost == cmp.rdCost); }
+ static bool compareModeInfoWithCost(ModeInfoWithCost a, ModeInfoWithCost b) { return a.rdCost < b.rdCost; }
+ };
+
+ struct ISPTestedModeInfo
+ {
+ int numCompSubParts;
+ double rdCost;
+
+ ISPTestedModeInfo() {}
+
+ void setMode(int numParts, double cost)
+ {
+ numCompSubParts = numParts;
+ rdCost = cost;
+ }
+ void clear()
+ {
+ numCompSubParts = -1;
+ rdCost = MAX_DOUBLE;
+ }
+ };
+ struct ISPTestedModesInfo
+ {
+ ISPTestedModeInfo intraMode[NUM_LUMA_MODE][2];
+ bool modeHasBeenTested[NUM_LUMA_MODE][2];
+ int numTotalParts[2];
+ static_vector<int, FAST_UDI_MAX_RDMODE_NUM> testedModes[2];
+ int bestModeSoFar;
+ ISPType bestSplitSoFar;
+ int bestMode[2];
+ double bestCost[2];
+ int numTestedModes[2];
+ int candIndexInList[2];
+ bool stopTestingHorSplit;
+ bool stopTestingVerSplit;
+ int numOrigModesToTest;
+
+ // set a tested mode results
+ void setModeResults(ISPType splitType, int iModeIdx, int numCompletedParts, double rdCost, double currentBestCost)
+ {
+ const unsigned st = splitType - 1;
+ CHECKD(st > 1, "The split type is invalid!");
+ const int maxNumParts = numTotalParts[st];
+ intraMode[iModeIdx][st].setMode(numCompletedParts, numCompletedParts == maxNumParts ? rdCost : MAX_DOUBLE);
+ testedModes[st].push_back(iModeIdx);
+ numTestedModes[st]++;
+ modeHasBeenTested[iModeIdx][st] = true;
+ if (numCompletedParts == maxNumParts && rdCost < bestCost[st]) // best mode update
+ {
+ bestMode[st] = iModeIdx;
+ bestCost[st] = rdCost;
+ }
+ if (numCompletedParts == maxNumParts && rdCost < currentBestCost) // best mode update
+ {
+ bestModeSoFar = iModeIdx;
+ bestSplitSoFar = splitType;
+ }
+ }
+
+ int getNumCompletedSubParts(ISPType splitType, int iModeIdx)
+ {
+ const unsigned st = splitType - 1;
+ CHECK(st < 0 || st > 1, "The split type is invalid!");
+ CHECK(iModeIdx < 0 || iModeIdx >(NUM_LUMA_MODE - 1), "The modeIdx is invalid");
+ return modeHasBeenTested[iModeIdx][st] ? intraMode[iModeIdx][st].numCompSubParts : -1;
+ }
+
+ double getRDCost(ISPType splitType, int iModeIdx, int maxNumSubParts)
+ {
+ const unsigned st = splitType - 1;
+ CHECKD(st > 1, "The split type is invalid!");
+ return modeHasBeenTested[iModeIdx][st] && intraMode[iModeIdx][st].numCompSubParts == maxNumSubParts
+ ? intraMode[iModeIdx][st].rdCost
+ : -1;
+ }
+
+ // get a tested intra mode index
+ int getTestedIntraMode(ISPType splitType, int pos)
+ {
+ const unsigned st = splitType - 1;
+ CHECKD(st > 1, "The split type is invalid!");
+ return pos < testedModes[st].size() ? testedModes[st].at(pos) : -1;
+ }
+
+ // set everything to default values
+ void clear()
+ {
+ numTestedModes[0] = numTestedModes[1] = 0;
+ candIndexInList[0] = candIndexInList[1] = 0;
+ stopTestingHorSplit = false;
+ stopTestingVerSplit = false;
+ testedModes[0].clear();
+ testedModes[1].clear();
+ bestCost[0] = MAX_DOUBLE;
+ bestCost[1] = MAX_DOUBLE;
+ bestMode[0] = -1;
+ bestMode[1] = -1;
+ bestModeSoFar = -1;
+ bestSplitSoFar = NOT_INTRA_SUBPARTITIONS;
+ numOrigModesToTest = -1;
+ memset(modeHasBeenTested, 0, sizeof(modeHasBeenTested));
+ }
+ void clearISPModeInfo(int idx)
+ {
+ intraMode[idx][0].clear();
+ intraMode[idx][1].clear();
+ }
+ };
+
+ static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_ispCandListHor, m_ispCandListVer;
+ static_vector<ModeInfoWithCost, FAST_UDI_MAX_RDMODE_NUM> m_regIntraRDListWithCosts;
+ ISPTestedModesInfo m_ispTestedModes;
+#else
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_rdModeListWithoutMrl;
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_rdModeListWithoutMrlHor;
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_rdModeListWithoutMrlVer;
+#endif
//cost variables for the EMT algorithm and new modes list
double m_bestModeCostStore[ NUM_LFNST_NUM_PER_SET ]; // RD cost of the best mode for each PU using DCT2
@@ -98,9 +219,11 @@ private:
ModeInfo m_savedRdModeList[ NUM_LFNST_NUM_PER_SET ][ NUM_LUMA_MODE ];
int32_t m_savedNumRdModes[ NUM_LFNST_NUM_PER_SET ];
+#if !JVET_O0502_ISP_CLEANUP
static_vector<double, FAST_UDI_MAX_RDMODE_NUM> m_intraModeDiagRatio;
static_vector<double, FAST_UDI_MAX_RDMODE_NUM> m_intraModeHorVerRatio;
static_vector<int, FAST_UDI_MAX_RDMODE_NUM> m_intraModeTestedNormalIntra;
+#endif
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_uiSavedRdModeListLFNST;
static_vector<ModeInfo, FAST_UDI_MAX_RDMODE_NUM> m_uiSavedHadModeListLFNST;
@@ -180,7 +303,9 @@ protected:
ChromaCbfs xRecurIntraChromaCodingQT( CodingStructure &cs, Partitioner& pm, const double bestCostSoFar = MAX_DOUBLE, const PartSplit ispType = TU_NO_ISP );
bool xRecurIntraCodingLumaQT ( CodingStructure &cs, Partitioner& pm, const double bestCostSoFar = MAX_DOUBLE, const int subTuIdx = -1, const PartSplit ispType = TU_NO_ISP, const bool ispIsCurrentWinner = false, bool mtsCheckRangeFlag = false, int mtsFirstCheckId = 0, int mtsLastCheckId = 0, bool moreProbMTSIdxFirst = false );
-
+#if JVET_O0502_ISP_CLEANUP
+ bool xIntraCodingLumaISP ( CodingStructure& cs, Partitioner& pm, const double bestCostSoFar = MAX_DOUBLE );
+#endif
void encPredIntraDPCM( const ComponentID &compID, PelBuf &pOrg, PelBuf &pDst, const uint32_t &uiDirMode );
static bool useDPCMForFirstPassIntraEstimation( const PredictionUnit &pu, const uint32_t &uiDirMode );
@@ -189,6 +314,12 @@ protected:
void reduceHadCandList(static_vector<T, N>& candModeList, static_vector<double, N>& candCostList, int& numModesForFullRD, const double thresholdHadCost, const double thresholdHadCostConv);
double m_bestCostNonMip;
+
+#if JVET_O0502_ISP_CLEANUP
+ void xGetNextISPMode ( ModeInfo& modeInfo, const ModeInfo* lastMode, const Size cuSize );
+ void xFindAlreadyTestedNearbyIntraModes ( int currentIntraMode, int* leftIntraMode, int* rightIntraMode, ISPType ispOption, int windowSize );
+ void xSortISPCandList ( double bestCostSoFar, double bestNonISPCost );
+#endif
};// END CLASS DEFINITION EncSearch
//! \}
--
GitLab