From 5e7c70d2d25e029ecb35cb59c2f42fd8b35bcbc0 Mon Sep 17 00:00:00 2001
From: Alexey Filippov <alexey.filippov@huawei.com>
Date: Sun, 7 Apr 2019 21:04:24 +0300
Subject: [PATCH] cShift-related modifications and macro-related clean-up in
UnitTools.cpp
---
source/Lib/CommonLib/RdCost.cpp | 77 +++++++++++++++++++++------
source/Lib/CommonLib/RdCost.h | 15 +++++-
source/Lib/CommonLib/TypeDef.h | 8 +--
source/Lib/CommonLib/UnitTools.cpp | 6 +--
source/Lib/EncoderLib/EncCu.cpp | 6 +++
source/Lib/EncoderLib/InterSearch.cpp | 8 +++
source/Lib/EncoderLib/IntraSearch.cpp | 3 ++
7 files changed, 99 insertions(+), 24 deletions(-)
diff --git a/source/Lib/CommonLib/RdCost.cpp b/source/Lib/CommonLib/RdCost.cpp
index fe4b62f78..202f55456 100644
--- a/source/Lib/CommonLib/RdCost.cpp
+++ b/source/Lib/CommonLib/RdCost.cpp
@@ -380,6 +380,10 @@ Distortion RdCost::getDistPart( const CPelBuf &org, const CPelBuf &cur, int bitD
#if WCG_EXT
if( orgLuma )
{
+#if JVET_N0671_RDCOST_FIX
+ cDtParam.cShiftX = getComponentScaleX(compID, m_cf);
+ cDtParam.cShiftY = getComponentScaleY(compID, m_cf);
+#endif
if( isChroma(compID) )
{
cDtParam.orgLuma = *orgLuma;
@@ -3025,9 +3029,10 @@ Distortion RdCost::xGetSSE_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const int iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const int cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const int cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
@@ -3040,7 +3045,12 @@ Distortion RdCost::xGetSSE_WTD( const DistParam &rcDtParam )
}
piOrg += iStrideOrg;
piCur += iStrideCur;
+
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3061,9 +3071,10 @@ Distortion RdCost::xGetSSE2_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
@@ -3074,7 +3085,11 @@ Distortion RdCost::xGetSSE2_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[1 ], piCur[1 ], uiShift, piOrgLuma[size_t(1)<<cShift]); // piOrg[1] - piCur[1]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3095,9 +3110,10 @@ Distortion RdCost::xGetSSE4_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
@@ -3110,7 +3126,11 @@ Distortion RdCost::xGetSSE4_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[3 ], piCur[3 ], uiShift, piOrgLuma[size_t(3)<<cShift] ); // piOrg[3] - piCur[3]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3131,9 +3151,10 @@ Distortion RdCost::xGetSSE8_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
@@ -3150,7 +3171,11 @@ Distortion RdCost::xGetSSE8_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[7 ], piCur[7 ], uiShift, piOrgLuma[size_t(7)<<cShift ]); // piOrg[7] - piCur[7]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3170,9 +3195,10 @@ Distortion RdCost::xGetSSE16_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
uint32_t uiShift = DISTORTION_PRECISION_ADJUSTMENT(rcDtParam.bitDepth) << 1;
@@ -3196,7 +3222,12 @@ Distortion RdCost::xGetSSE16_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[15 ], piCur[15 ], uiShift, piOrgLuma[size_t(15)<<cShift ]); //piOrg[15] - piCur[15]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3216,9 +3247,10 @@ Distortion RdCost::xGetSSE16N_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
uint32_t uiShift = DISTORTION_PRECISION_ADJUSTMENT(rcDtParam.bitDepth) << 1;
@@ -3245,7 +3277,11 @@ Distortion RdCost::xGetSSE16N_WTD( const DistParam &rcDtParam )
}
piOrg += iStrideOrg;
piCur += iStrideCur;
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3265,9 +3301,10 @@ Distortion RdCost::xGetSSE32_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
- const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
+ const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
Distortion uiSum = 0;
@@ -3308,7 +3345,11 @@ Distortion RdCost::xGetSSE32_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[31], piCur[31], uiShift, piOrgLuma[size_t(31)<<cShift ]); // iTemp = piOrg[31] - piCur[31]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
@@ -3328,7 +3369,8 @@ Distortion RdCost::xGetSSE64_WTD( const DistParam &rcDtParam )
const Pel* piOrgLuma = rcDtParam.orgLuma.buf;
const size_t iStrideOrgLuma = rcDtParam.orgLuma.stride;
#if JVET_N0671_RDCOST_FIX
- const size_t cShift = 0;
+ const size_t cShift = rcDtParam.cShiftX;
+ const size_t cShiftY = rcDtParam.cShiftY;
#else
const size_t cShift = (rcDtParam.compID==COMPONENT_Y) ? 0 : 1; // assume 420, could use getComponentScaleX, getComponentScaleY
#endif
@@ -3403,7 +3445,12 @@ Distortion RdCost::xGetSSE64_WTD( const DistParam &rcDtParam )
uiSum += getWeightedMSE(rcDtParam.compID, piOrg[63], piCur[63], uiShift, piOrgLuma[size_t(63)<<cShift]); // iTemp = piOrg[63] - piCur[63]; uiSum += Distortion(( iTemp * iTemp ) >> uiShift);
piOrg += iStrideOrg;
piCur += iStrideCur;
+
+#if JVET_N0671_RDCOST_FIX
+ piOrgLuma += iStrideOrgLuma<<cShiftY;
+#else
piOrgLuma += iStrideOrgLuma<<cShift;
+#endif
}
return ( uiSum );
}
diff --git a/source/Lib/CommonLib/RdCost.h b/source/Lib/CommonLib/RdCost.h
index 22e5b1b68..5523d62c9 100644
--- a/source/Lib/CommonLib/RdCost.h
+++ b/source/Lib/CommonLib/RdCost.h
@@ -87,10 +87,15 @@ public:
// (vertical) subsampling shift (for reducing complexity)
// - 0 = no subsampling, 1 = even rows, 2 = every 4th, etc.
int subShift;
-
+#if JVET_N0671_RDCOST_FIX
+ int cShiftX;
+ int cShiftY;
+#endif
DistParam() :
org(), cur(), step( 1 ), bitDepth( 0 ), useMR( false ), applyWeight( false ), isBiPred( false ), wpCur( nullptr ), compID( MAX_NUM_COMPONENT ), maximumDistortionForEarlyExit( std::numeric_limits<Distortion>::max() ), subShift( 0 )
-
+#if JVET_N0671_RDCOST_FIX
+ , cShiftX(-1), cShiftY(-1)
+#endif
{ }
};
@@ -112,6 +117,9 @@ private:
static uint32_t m_signalType;
static double m_chromaWeight;
static int m_lumaBD;
+#if JVET_N0671_RDCOST_FIX
+ ChromaFormat m_cf;
+#endif // JVET_N0671_RDCOST_FIX
#endif
double m_DistScale;
double m_dLambdaMotionSAD[2 /* 0=standard, 1=for transquant bypass when mixed-lossless cost evaluation enabled*/];
@@ -128,6 +136,9 @@ public:
virtual ~RdCost();
#if WCG_EXT
+#if JVET_N0671_RDCOST_FIX
+ void setChromaFormat ( const ChromaFormat & _cf) { m_cf = _cf; }
+#endif
double calcRdCost ( uint64_t fracBits, Distortion distortion, bool useUnadjustedLambda = true );
#else
double calcRdCost ( uint64_t fracBits, Distortion distortion );
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 2ce1e1f73..848259148 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -50,6 +50,10 @@
#include <assert.h>
#include <cassert>
+#define JVET_N0470_SMVD_FIX 1 // remove mvd_l1_zero_flag condition, align to spec text.
+
+#define JVET_N0235_SMVD_SPS 1
+
#define JVET_N0671 1
#if JVET_N0671
@@ -65,10 +69,6 @@
#endif //JVET_N0671
-#define JVET_N0470_SMVD_FIX 1 // remove mvd_l1_zero_flag condition, align to spec text.
-
-#define JVET_N0235_SMVD_SPS 1
-
#define JVET_N0843_BVP_SIMPLIFICATION 1
#define JVET_N0448_N0380 1 // When MaxNumMergeCand is 1, MMVD_BASE_MV_NUM is inferred to be 1.
diff --git a/source/Lib/CommonLib/UnitTools.cpp b/source/Lib/CommonLib/UnitTools.cpp
index ec5d36eca..d0050c240 100644
--- a/source/Lib/CommonLib/UnitTools.cpp
+++ b/source/Lib/CommonLib/UnitTools.cpp
@@ -987,11 +987,11 @@ uint32_t PU::getFinalIntraMode( const PredictionUnit &pu, const ChannelType &chT
uiIntraMode = lumaPU.intraDir[0];
}
- if( pu.chromaFormat == CHROMA_422 && !isLuma( chType )
#if JVET_N0671_CHROMA_FORMAT_422
- && uiIntraMode < NUM_LUMA_MODE
+ if( pu.chromaFormat == CHROMA_422 && !isLuma( chType ) && uiIntraMode < NUM_LUMA_MODE ) // map directional, planar and dc
+#else
+ if( pu.chromaFormat == CHROMA_422 && !isLuma( chType ) )
#endif //JVET_N0671_CHROMA_FORMAT_422
- ) // map directional, planar and dc
{
uiIntraMode = g_chroma422IntraAngleMappingTable[uiIntraMode];
}
diff --git a/source/Lib/EncoderLib/EncCu.cpp b/source/Lib/EncoderLib/EncCu.cpp
index c5e361b80..3d0f61a4d 100644
--- a/source/Lib/EncoderLib/EncCu.cpp
+++ b/source/Lib/EncoderLib/EncCu.cpp
@@ -3707,6 +3707,9 @@ Distortion EncCu::getDistortionDb( CodingStructure &cs, CPelBuf org, CPelBuf rec
{
Distortion dist = 0;
#if WCG_EXT
+#if JVET_N0671_RDCOST_FIX
+ m_pcRdCost->setChromaFormat(cs.sps->getChromaFormatIdc());
+#endif
CPelBuf orgLuma = cs.picture->getOrigBuf( cs.area.blocks[COMPONENT_Y] );
if ( m_pcEncCfg->getLumaLevelToDeltaQPMapping().isEnabled() || (
m_pcEncCfg->getReshaper() && ( cs.slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag() ) ) )
@@ -4140,6 +4143,9 @@ void EncCu::xEncodeDontSplit( CodingStructure &cs, Partitioner &partitioner )
#if REUSE_CU_RESULTS
void EncCu::xReuseCachedResult( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner )
{
+#if JVET_N0671_RDCOST_FIX
+ m_pcRdCost->setChromaFormat(tempCS->sps->getChromaFormatIdc());
+#endif
BestEncInfoCache* bestEncCache = dynamic_cast<BestEncInfoCache*>( m_modeCtrl );
CHECK( !bestEncCache, "If this mode is chosen, mode controller has to implement the mode caching capabilities" );
EncTestMode cachedMode;
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index 51079bb74..af871b34e 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -6287,6 +6287,10 @@ void InterSearch::xEstimateInterResidualQT(CodingStructure &cs, Partitioner &par
{
const UnitArea& currArea = partitioner.currArea();
const SPS &sps = *cs.sps;
+#if JVET_N0671_RDCOST_FIX
+ m_pcRdCost->setChromaFormat(sps.getChromaFormatIdc());
+#endif
+
const uint32_t numValidComp = getNumberValidComponents( sps.getChromaFormatIdc() );
const uint32_t numTBlocks = getNumberValidTBlocks ( *cs.pcv );
const CodingUnit &cu = *cs.getCU(partitioner.chType);
@@ -6921,6 +6925,10 @@ void InterSearch::encodeResAndCalcRdInterCU(CodingStructure &cs, Partitioner &pa
, const bool luma, const bool chroma
)
{
+#if JVET_N0671_RDCOST_FIX
+ m_pcRdCost->setChromaFormat(cs.sps->getChromaFormatIdc());
+#endif
+
CodingUnit &cu = *cs.getCU( partitioner.chType );
const ChromaFormat format = cs.area.chromaFormat;;
diff --git a/source/Lib/EncoderLib/IntraSearch.cpp b/source/Lib/EncoderLib/IntraSearch.cpp
index 141da8b2c..258b886a8 100644
--- a/source/Lib/EncoderLib/IntraSearch.cpp
+++ b/source/Lib/EncoderLib/IntraSearch.cpp
@@ -1581,6 +1581,9 @@ void IntraSearch::xIntraCodingTUBlock(TransformUnit &tu, const ComponentID &comp
}
CodingStructure &cs = *tu.cs;
+#if JVET_N0671_RDCOST_FIX
+ m_pcRdCost->setChromaFormat(cs.sps->getChromaFormatIdc());
+#endif
const CompArea &area = tu.blocks[compID];
const SPS &sps = *cs.sps;
--
GitLab