diff --git a/cfg/encoder_intra_vtm.cfg b/cfg/encoder_intra_vtm.cfg
index e3cc0a9ad25e16106211814f797f69a40551b340..041ade27fedf5422edd41dd3a888a607f05f737b 100644
--- a/cfg/encoder_intra_vtm.cfg
+++ b/cfg/encoder_intra_vtm.cfg
@@ -120,6 +120,9 @@ PBIntraFast : 1
FastMrg : 1
AMaxBT : 1
+# Encoder optimization tools
+AffineAmvrEncOpt : 0
+
### DO NOT ADD ANYTHING BELOW THIS LINE ###
### DO NOT DELETE THE EMPTY LINE BELOW ###
diff --git a/cfg/encoder_lowdelay_P_vtm.cfg b/cfg/encoder_lowdelay_P_vtm.cfg
index 9c2f8ece3e201df90b0200240d366f634222ae48..e9aa837f2171481534ac8592c269c6dc83b6a682 100644
--- a/cfg/encoder_lowdelay_P_vtm.cfg
+++ b/cfg/encoder_lowdelay_P_vtm.cfg
@@ -137,6 +137,9 @@ PBIntraFast : 1
FastMrg : 1
AMaxBT : 1
+# Encoder optimization tools
+AffineAmvrEncOpt : 0
+
### DO NOT ADD ANYTHING BELOW THIS LINE ###
### DO NOT DELETE THE EMPTY LINE BELOW ###
diff --git a/cfg/encoder_lowdelay_vtm.cfg b/cfg/encoder_lowdelay_vtm.cfg
index 324320970afffe7c5a042636337590cdb55b0c94..0922ffc13b62dab779cd616704db7945e19d3feb 100644
--- a/cfg/encoder_lowdelay_vtm.cfg
+++ b/cfg/encoder_lowdelay_vtm.cfg
@@ -140,6 +140,9 @@ PBIntraFast : 1
FastMrg : 1
AMaxBT : 1
+# Encoder optimization tools
+AffineAmvrEncOpt : 0
+
### DO NOT ADD ANYTHING BELOW THIS LINE ###
### DO NOT DELETE THE EMPTY LINE BELOW ###
diff --git a/cfg/encoder_randomaccess_vtm.cfg b/cfg/encoder_randomaccess_vtm.cfg
index ab7c99c0cb22fc888ec9da541acde853d45ecc60..7ca088122f6b1439b67870bb4c2adbcc64c56038 100644
--- a/cfg/encoder_randomaccess_vtm.cfg
+++ b/cfg/encoder_randomaccess_vtm.cfg
@@ -155,6 +155,9 @@ PBIntraFast : 1
FastMrg : 1
AMaxBT : 1
+# Encoder optimization tools
+AffineAmvrEncOpt : 1
+
### DO NOT ADD ANYTHING BELOW THIS LINE ###
### DO NOT DELETE THE EMPTY LINE BELOW ###
diff --git a/source/App/EncoderApp/EncApp.cpp b/source/App/EncoderApp/EncApp.cpp
index 611f6f82f6b5ca37a2b62ac3108be0f98d9380bc..54cfe178e2cfab8cf953d86c3f2a5ab06c59bc3b 100644
--- a/source/App/EncoderApp/EncApp.cpp
+++ b/source/App/EncoderApp/EncApp.cpp
@@ -262,6 +262,9 @@ void EncApp::xInitLibCfg()
#endif
#if JVET_M0246_AFFINE_AMVR
m_cEncLib.setUseAffineAmvr ( m_AffineAmvr );
+#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ m_cEncLib.setUseAffineAmvrEncOpt ( m_AffineAmvrEncOpt );
#endif
m_cEncLib.setIBCMode ( m_IBCMode );
m_cEncLib.setIBCLocalSearchRangeX ( m_IBCLocalSearchRangeX );
diff --git a/source/App/EncoderApp/EncAppCfg.cpp b/source/App/EncoderApp/EncAppCfg.cpp
index ea91679e6aa8ba27564bce3f45a20baadae298ba..f31fad8636275253c227ef20c31d5251b929fa5e 100644
--- a/source/App/EncoderApp/EncAppCfg.cpp
+++ b/source/App/EncoderApp/EncAppCfg.cpp
@@ -869,6 +869,9 @@ bool EncAppCfg::parseCfg( int argc, char* argv[] )
#endif
#if JVET_M0246_AFFINE_AMVR
("AffineAmvr", m_AffineAmvr, false, "Eanble AMVR for affine inter mode")
+#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ ("AffineAmvrEncOpt", m_AffineAmvrEncOpt, false, "Enable encoder optimization of affine AMVR")
#endif
( "IBC", m_IBCMode, 0u, "IBCMode (0x1:enabled, 0x0:disabled) [default: disabled]")
( "IBCLocalSearchRangeX", m_IBCLocalSearchRangeX, 128u, "Search range of IBC local search in x direction")
@@ -3155,6 +3158,10 @@ void EncAppCfg::xPrintParameter()
#endif
#if JVET_M0246_AFFINE_AMVR
msg( VERBOSE, "AffineAmvr:%d ", m_AffineAmvr );
+#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ m_AffineAmvrEncOpt = m_AffineAmvr ? m_AffineAmvrEncOpt : false;
+ msg( VERBOSE, "AffineAmvrEncOpt:%d ", m_AffineAmvrEncOpt );
#endif
}
msg(VERBOSE, "IBC:%d ", m_IBCMode);
diff --git a/source/App/EncoderApp/EncAppCfg.h b/source/App/EncoderApp/EncAppCfg.h
index 9d7b9a62ca4abd2a412fdf2c4546995ad7303a04..c91c7ec580402ba1b0a7941bb7222fb0e4404b7f 100644
--- a/source/App/EncoderApp/EncAppCfg.h
+++ b/source/App/EncoderApp/EncAppCfg.h
@@ -243,6 +243,9 @@ protected:
#if JVET_M0246_AFFINE_AMVR
bool m_AffineAmvr;
#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ bool m_AffineAmvrEncOpt;
+#endif
unsigned m_IBCMode;
unsigned m_IBCLocalSearchRangeX;
diff --git a/source/Lib/CommonLib/MotionInfo.h b/source/Lib/CommonLib/MotionInfo.h
index 43acd88f1ca37ff7be0488afd9133ee92bf765fe..c193228a79b0c9523180c969f7199c7e18a2fb8e 100644
--- a/source/Lib/CommonLib/MotionInfo.h
+++ b/source/Lib/CommonLib/MotionInfo.h
@@ -161,6 +161,9 @@ class GBiMotionParam
bool m_readOnlyAffine[2][2][33];
Mv m_mvAffine[2][2][33][3];
Distortion m_distAffine[2][2][33];
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ int m_mvpIdx[2][2][33];
+#endif
public:
@@ -182,6 +185,9 @@ public:
memset(m_dist, -1, 2 * 33 * sizeof(Distortion));
memset(m_readOnlyAffine, false, 2 * 2 * 33 * sizeof(bool));
memset(m_distAffine, -1, 2 * 2 * 33 * sizeof(Distortion));
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ memset( m_mvpIdx, 0, 2 * 2 * 33 * sizeof( int ) );
+#endif
}
void setReadMode(bool b, uint32_t uiRefList, uint32_t uiRefIdx) { m_readOnly[uiRefList][uiRefIdx] = b; }
@@ -206,16 +212,30 @@ public:
Mv& getAffineMv(uint32_t uiRefList, uint32_t uiRefIdx, uint32_t uiAffineMvIdx, int bP4) { return m_mvAffine[bP4][uiRefList][uiRefIdx][uiAffineMvIdx]; }
- void copyAffineMvFrom(Mv(&racAffineMvs)[3], Distortion uiDist, uint32_t uiRefList, uint32_t uiRefIdx, int bP4)
+ void copyAffineMvFrom(Mv(&racAffineMvs)[3], Distortion uiDist, uint32_t uiRefList, uint32_t uiRefIdx, int bP4
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , const int mvpIdx
+#endif
+ )
{
memcpy(m_mvAffine[bP4][uiRefList][uiRefIdx], racAffineMvs, 3 * sizeof(Mv));
m_distAffine[bP4][uiRefList][uiRefIdx] = uiDist;
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ m_mvpIdx[bP4][uiRefList][uiRefIdx] = mvpIdx;
+#endif
}
- void copyAffineMvTo(Mv acAffineMvs[3], Distortion& ruiDist, uint32_t uiRefList, uint32_t uiRefIdx, int bP4)
+ void copyAffineMvTo(Mv acAffineMvs[3], Distortion& ruiDist, uint32_t uiRefList, uint32_t uiRefIdx, int bP4
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , int& mvpIdx
+#endif
+ )
{
memcpy(acAffineMvs, m_mvAffine[bP4][uiRefList][uiRefIdx], 3 * sizeof(Mv));
ruiDist = m_distAffine[bP4][uiRefList][uiRefIdx];
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ mvpIdx = m_mvpIdx[bP4][uiRefList][uiRefIdx];
+#endif
}
};
struct LutMotionCand
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index e0dfe4779d6d787859714e1b12467df7b4b07818..cb1dc56721259efdd479c56284673317f95b2bfc 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -115,6 +115,9 @@ typedef std::pair<int, int> TrCost;
#endif
#define JVET_M0246_AFFINE_AMVR 1
+#if JVET_M0246_AFFINE_AMVR
+#define JVET_M0247_AFFINE_AMVR_ENCOPT 1
+#endif
#define JVET_M0421_SPLIT_SIG 1
#define JVET_M0173_MOVE_GT2_TO_FIRST_PASS 1 // Moving the gtr2 flag to the first coding pass
diff --git a/source/Lib/EncoderLib/EncCfg.h b/source/Lib/EncoderLib/EncCfg.h
index a34fdac1aa7579b157ad0bf5ce8ff5338ff848f0..bbfd74821a63dea93ca14b8dec8a7450abc75930 100644
--- a/source/Lib/EncoderLib/EncCfg.h
+++ b/source/Lib/EncoderLib/EncCfg.h
@@ -243,6 +243,9 @@ protected:
#endif
#if JVET_M0246_AFFINE_AMVR
bool m_AffineAmvr;
+#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ bool m_AffineAmvrEncOpt;
#endif
unsigned m_IBCMode;
unsigned m_IBCLocalSearchRangeX;
@@ -767,6 +770,10 @@ public:
void setUseAffineAmvr ( bool b ) { m_AffineAmvr = b; }
bool getUseAffineAmvr () const { return m_AffineAmvr; }
#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ void setUseAffineAmvrEncOpt ( bool b ) { m_AffineAmvrEncOpt = b; }
+ bool getUseAffineAmvrEncOpt () const { return m_AffineAmvrEncOpt; }
+#endif
void setIBCMode (unsigned n) { m_IBCMode = n; }
unsigned getIBCMode () const { return m_IBCMode; }
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index 1b5559dff1c59e6f5298d1f710ecffc744f132ad..65d35faf49c6580fbe3c340ae2ce1b958e9ae45e 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -3950,20 +3950,35 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
}
else
{
- xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp );
+ xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , aaiMvpIdx[iRefList][iRefIdxTemp], affiAMVPInfoTemp[eRefPicList]
+#endif
+ );
}
}
else
{
- xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp );
+ xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , aaiMvpIdx[iRefList][iRefIdxTemp], affiAMVPInfoTemp[eRefPicList]
+#endif
+ );
}
if(pu.cu->cs->sps->getSpsNext().getUseGBi() && pu.cu->GBiIdx == GBI_DEFAULT && pu.cu->slice->isInterB())
{
m_uniMotions.setReadModeAffine(true, (uint8_t)iRefList, (uint8_t)iRefIdxTemp, pu.cu->affineType);
- m_uniMotions.copyAffineMvFrom(cMvTemp[iRefList][iRefIdxTemp], uiCostTemp - m_pcRdCost->getCost(uiBitsTemp), (uint8_t)iRefList, (uint8_t)iRefIdxTemp, pu.cu->affineType);
+ m_uniMotions.copyAffineMvFrom(cMvTemp[iRefList][iRefIdxTemp], uiCostTemp - m_pcRdCost->getCost(uiBitsTemp), (uint8_t)iRefList, (uint8_t)iRefIdxTemp, pu.cu->affineType
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , aaiMvpIdx[iRefList][iRefIdxTemp]
+#endif
+ );
}
// Set best AMVP Index
xCopyAffineAMVPInfo( affiAMVPInfoTemp[eRefPicList], aacAffineAMVPInfo[iRefList][iRefIdxTemp] );
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( pu.cu->imv != 2 || !m_pcEncCfg->getUseAffineAmvrEncOpt() )
+#endif
xCheckBestAffineMVP( pu, affiAMVPInfoTemp[eRefPicList], eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPred[iRefList][iRefIdxTemp], aaiMvpIdx[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp );
if ( iRefList == 0 )
@@ -4185,8 +4200,15 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
uiBitsTemp += m_auiMVPIdxCost[aaiMvpIdxBi[iRefList][iRefIdxTemp]][AMVP_MAX_NUM_CANDS];
// call Affine ME
- xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPredBi[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp, true );
+ xAffineMotionEstimation( pu, origBuf, eRefPicList, cMvPredBi[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp,
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ aaiMvpIdxBi[iRefList][iRefIdxTemp], aacAffineAMVPInfo[iRefList][iRefIdxTemp],
+#endif
+ true );
xCopyAffineAMVPInfo( aacAffineAMVPInfo[iRefList][iRefIdxTemp], affiAMVPInfoTemp[eRefPicList] );
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( pu.cu->imv != 2 || !m_pcEncCfg->getUseAffineAmvrEncOpt() )
+#endif
xCheckBestAffineMVP( pu, affiAMVPInfoTemp[eRefPicList], eRefPicList, cMvTemp[iRefList][iRefIdxTemp], cMvPredBi[iRefList][iRefIdxTemp], aaiMvpIdxBi[iRefList][iRefIdxTemp], uiBitsTemp, uiCostTemp );
if ( uiCostTemp < uiCostBi )
@@ -4532,13 +4554,25 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
Mv acMv[3],
uint32_t& ruiBits,
Distortion& ruiCost,
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ int& mvpIdx,
+ const AffineAMVPInfo& aamvpi,
+#endif
bool bBi)
{
- if( pu.cu->cs->sps->getSpsNext().getUseGBi() && pu.cu->GBiIdx != GBI_DEFAULT && !bBi && xReadBufferedAffineUniMv(pu, eRefPicList, iRefIdxPred, acMvPred, acMv, ruiBits, ruiCost) )
+ if( pu.cu->cs->sps->getSpsNext().getUseGBi() && pu.cu->GBiIdx != GBI_DEFAULT && !bBi && xReadBufferedAffineUniMv(pu, eRefPicList, iRefIdxPred, acMvPred, acMv, ruiBits, ruiCost
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , mvpIdx, aamvpi
+#endif
+ ) )
{
return;
}
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ uint32_t dirBits = ruiBits - m_auiMVPIdxCost[mvpIdx][aamvpi.numCand];
+ int bestMvpIdx = mvpIdx;
+#endif
const int width = pu.Y().width;
const int height = pu.Y().height;
@@ -4634,10 +4668,24 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
// get cost with mv
m_pcRdCost->setCostScale(0);
uiBitsBest = ruiBits;
- DTRACE( g_trace_ctx, D_COMMON, " (%d) xx uiBitsBest=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiBitsBest );
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( pu.cu->imv == 2 && m_pcEncCfg->getUseAffineAmvrEncOpt() )
+ {
+ uiBitsBest = dirBits + xDetermineBestMvp( pu, acMvTemp, mvpIdx, aamvpi );
+ acMvPred[0] = aamvpi.mvCandLT[mvpIdx];
+ acMvPred[1] = aamvpi.mvCandRT[mvpIdx];
+ acMvPred[2] = aamvpi.mvCandLB[mvpIdx];
+ }
+ else
+ {
+#endif
+ DTRACE( g_trace_ctx, D_COMMON, " (%d) xx uiBitsBest=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiBitsBest );
#if JVET_M0246_AFFINE_AMVR
- uiBitsBest += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
- DTRACE( g_trace_ctx, D_COMMON, " (%d) yy uiBitsBest=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiBitsBest );
+ uiBitsBest += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
+ DTRACE( g_trace_ctx, D_COMMON, " (%d) yy uiBitsBest=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiBitsBest );
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ }
+#endif
#else
for ( int i = 0; i < mvNum; i++ )
{
@@ -4666,7 +4714,9 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
const int bufStride = pBuf->Y().stride;
const int predBufStride = predBuf.Y().stride;
-
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ Mv prevIterMv[7][3];
+#endif
int iIterTime;
if ( pu.cu->affineType == AFFINEMODEL_6PARAM )
{
@@ -4683,6 +4733,9 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
}
for ( int iter=0; iter<iIterTime; iter++ ) // iterate loop
{
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ memcpy( prevIterMv[iter], acMvTemp, sizeof( Mv ) * 3 );
+#endif
/*********************************************************************************
* use gradient to update mv
*********************************************************************************/
@@ -4774,29 +4827,35 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
);
}
#endif
- bool bAllZero = false;
- for ( int i = 0; i < mvNum; i++ )
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( !m_pcEncCfg->getUseAffineAmvrEncOpt() )
{
-#if JVET_M0246_AFFINE_AMVR
- Mv deltaMv = acDeltaMv[i];
- if ( pu.cu->imv == 2 )
+#endif
+ bool bAllZero = false;
+ for ( int i = 0; i < mvNum; i++ )
{
- deltaMv.roundToPrecision( MV_PRECISION_INTERNAL, MV_PRECISION_HALF );
- }
- if ( deltaMv.getHor() != 0 || deltaMv.getVer() != 0 )
+#if JVET_M0246_AFFINE_AMVR
+ Mv deltaMv = acDeltaMv[i];
+ if ( pu.cu->imv == 2 )
+ {
+ deltaMv.roundToPrecision( MV_PRECISION_INTERNAL, MV_PRECISION_HALF );
+ }
+ if ( deltaMv.getHor() != 0 || deltaMv.getVer() != 0 )
#else
- if ( acDeltaMv[i].getHor() != 0 || acDeltaMv[i].getVer() != 0 )
+ if ( acDeltaMv[i].getHor() != 0 || acDeltaMv[i].getVer() != 0 )
#endif
- {
- bAllZero = false;
- break;
+ {
+ bAllZero = false;
+ break;
+ }
+ bAllZero = true;
}
- bAllZero = true;
- }
-
- if ( bAllZero )
- break;
+ if ( bAllZero )
+ break;
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ }
+#endif
// do motion compensation with updated mv
for ( int i = 0; i < mvNum; i++ )
{
@@ -4824,6 +4883,29 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
pu.cu->lumaSize(),
*pu.cs->sps);
}
+
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( m_pcEncCfg->getUseAffineAmvrEncOpt() )
+ {
+ bool identical = false;
+ for ( int k = iter; k >= 0; k-- )
+ {
+ if ( acMvTemp[0] == prevIterMv[k][0] && acMvTemp[1] == prevIterMv[k][1] )
+ {
+ identical = pu.cu->affineType ? acMvTemp[2] == prevIterMv[k][2] : true;
+ if ( identical )
+ {
+ break;
+ }
+ }
+ }
+ if ( identical )
+ {
+ break;
+ }
+ }
+#endif
+
xPredAffineBlk( COMPONENT_Y, pu, refPic, acMvTemp, predBuf, false, pu.cu->slice->clpRng( COMPONENT_Y ) );
// get error
@@ -4834,7 +4916,21 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
m_pcRdCost->setCostScale(0);
uint32_t uiBitsTemp = ruiBits;
#if JVET_M0246_AFFINE_AMVR
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ if ( pu.cu->imv == 2 && m_pcEncCfg->getUseAffineAmvrEncOpt() )
+ {
+ uiBitsTemp = dirBits + xDetermineBestMvp( pu, acMvTemp, bestMvpIdx, aamvpi );
+ acMvPred[0] = aamvpi.mvCandLT[bestMvpIdx];
+ acMvPred[1] = aamvpi.mvCandRT[bestMvpIdx];
+ acMvPred[2] = aamvpi.mvCandLB[bestMvpIdx];
+ }
+ else
+ {
+ uiBitsTemp += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
+ }
+#else
uiBitsTemp += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
+#endif
#else
for ( int i = 0; i < mvNum; i++ )
{
@@ -4860,6 +4956,9 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
uiCostBest = uiCostTemp;
uiBitsBest = uiBitsTemp;
memcpy( acMv, acMvTemp, sizeof(Mv) * 3 );
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ mvpIdx = bestMvpIdx;
+#endif
}
}
@@ -4902,6 +5001,56 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
if (uiCostBest <= AFFINE_ME_LIST_MVP_TH*m_hevcCost)
{
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ //search 8 nearest neighbors; integer distance
+ int testPos[8][2] = { { -1, 0 },{ 0, -1 },{ 0, 1 },{ 1, 0 },{ -1, -1 },{ -1, 1 },{ 1, 1 },{ 1, -1 } };
+ const uint32_t mvShift = pu.cu->imv == 1 ? 0 : ( pu.cu->imv == 2 ? ( MV_FRACTIONAL_BITS_DIFF << 1 ) : MV_FRACTIONAL_BITS_DIFF );
+ const int maxSearchRound = 3;
+
+ if ( m_pcEncCfg->getUseAffineAmvrEncOpt() && m_pcEncCfg->getIntraPeriod() != ( uint32_t ) -1 && pu.cu->imv )
+ {
+ for ( int rnd = 0; rnd < ( pu.cu->slice->getTLayer() <= 2 ? maxSearchRound : maxSearchRound - 1 ); rnd++ )
+ {
+ bool modelChange = false;
+ //search the model parameters with finear granularity;
+ for ( int j = 0; j < mvNum; j++ )
+ {
+ for ( int iter = 0; iter < 2; iter++ )
+ {
+ Mv centerMv[3];
+ memcpy( centerMv, acMv, sizeof( Mv ) * 3 );
+ memcpy( acMvTemp, acMv, sizeof( Mv ) * 3 );
+ for ( int i = ( iter ? 0: 4 ); i < ( iter ? 4 : 8 ); i++ )
+ {
+ acMvTemp[j].set( centerMv[j].getHor() + ( testPos[i][0] << mvShift ), centerMv[j].getVer() + ( testPos[i][1] << mvShift ) );
+
+ clipMv( acMvTemp[j], pu.cu->lumaPos(), pu.cu->lumaSize(), *pu.cs->sps );
+ xPredAffineBlk( COMPONENT_Y, pu, refPic, acMvTemp, predBuf, false, pu.cu->slice->clpRng( COMPONENT_Y ) );
+
+ Distortion costTemp = m_pcRdCost->getDistPart( predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth( CHANNEL_TYPE_LUMA ), COMPONENT_Y, DF_HAD );
+ uint32_t bitsTemp = ruiBits;
+ bitsTemp += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
+ costTemp = ( Distortion ) ( floor( fWeight * ( double ) costTemp ) + ( double ) m_pcRdCost->getCost( bitsTemp ) );
+
+ if ( costTemp < uiCostBest )
+ {
+ uiCostBest = costTemp;
+ uiBitsBest = bitsTemp;
+ ::memcpy( acMv, acMvTemp, sizeof( Mv ) * 3 );
+ modelChange = true;
+ }
+ }
+ }
+ }
+
+ if ( !modelChange )
+ {
+ break;
+ }
+ }
+ }
+#endif
+
Mv mvPredTmp[3] = { acMvPred[0], acMvPred[1], acMvPred[2] };
#if JVET_M0246_AFFINE_AMVR
if ( pu.cu->imv != 1 )
@@ -4988,6 +5137,11 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
#if JVET_M0246_AFFINE_AMVR
}
#endif
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ acMvPred[0] = aamvpi.mvCandLT[mvpIdx];
+ acMvPred[1] = aamvpi.mvCandRT[mvpIdx];
+ acMvPred[2] = aamvpi.mvCandLB[mvpIdx];
+#endif
// free buffer
for (int i = 0; i<iParaNum; i++)
@@ -6207,12 +6361,25 @@ bool InterSearch::xReadBufferedUniMv(PredictionUnit& pu, RefPicList eRefPicList,
return false;
}
-bool InterSearch::xReadBufferedAffineUniMv(PredictionUnit& pu, RefPicList eRefPicList, int32_t iRefIdx, Mv acMvPred[3], Mv acMv[3], uint32_t& ruiBits, Distortion& ruiCost)
+bool InterSearch::xReadBufferedAffineUniMv(PredictionUnit& pu, RefPicList eRefPicList, int32_t iRefIdx, Mv acMvPred[3], Mv acMv[3], uint32_t& ruiBits, Distortion& ruiCost
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , int& mvpIdx, const AffineAMVPInfo& aamvpi
+#endif
+)
{
if (m_uniMotions.isReadModeAffine((uint32_t)eRefPicList, (uint32_t)iRefIdx, pu.cu->affineType))
{
- m_uniMotions.copyAffineMvTo(acMv, ruiCost, (uint32_t)eRefPicList, (uint32_t)iRefIdx, pu.cu->affineType);
+ m_uniMotions.copyAffineMvTo(acMv, ruiCost, (uint32_t)eRefPicList, (uint32_t)iRefIdx, pu.cu->affineType
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , mvpIdx
+#endif
+ );
m_pcRdCost->setCostScale(0);
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ acMvPred[0] = aamvpi.mvCandLT[mvpIdx];
+ acMvPred[1] = aamvpi.mvCandRT[mvpIdx];
+ acMvPred[2] = aamvpi.mvCandLB[mvpIdx];
+#endif
uint32_t uiMvBits = 0;
for (int iVerIdx = 0; iVerIdx<(pu.cu->affineType ? 3 : 2); iVerIdx++)
@@ -6264,6 +6431,29 @@ void InterSearch::xClipMv( Mv& rcMv, const Position& pos, const struct Size& siz
rcMv.setVer( std::min( verMax, std::max( verMin, rcMv.getVer() ) ) );
}
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+uint32_t InterSearch::xDetermineBestMvp( PredictionUnit& pu, Mv acMvTemp[3], int& mvpIdx, const AffineAMVPInfo& aamvpi )
+{
+ bool mvpUpdated = false;
+ uint32_t minBits = std::numeric_limits<uint32_t>::max();
+ for ( int i = 0; i < aamvpi.numCand; i++ )
+ {
+ Mv mvPred[3] = { aamvpi.mvCandLT[i], aamvpi.mvCandRT[i], aamvpi.mvCandLB[i] };
+ uint32_t candBits = m_auiMVPIdxCost[i][aamvpi.numCand];
+ candBits += xCalcAffineMVBits( pu, acMvTemp, mvPred, pu.cu->imv != 1 );
+
+ if ( candBits < minBits )
+ {
+ minBits = candBits;
+ mvpIdx = i;
+ mvpUpdated = true;
+ }
+ }
+ CHECK( !mvpUpdated, "xDetermineBestMvp() error" );
+ return minBits;
+}
+#endif
+
#if JVET_M0444_SMVD
void InterSearch::symmvdCheckBestMvp(
PredictionUnit& pu,
diff --git a/source/Lib/EncoderLib/InterSearch.h b/source/Lib/EncoderLib/InterSearch.h
index 27030ab31ec24eddb4085f35b4a0f7ba54efb931..da9abdf1bea40095271e8b8f1c0eede5da81fddc 100644
--- a/source/Lib/EncoderLib/InterSearch.h
+++ b/source/Lib/EncoderLib/InterSearch.h
@@ -404,6 +404,10 @@ protected:
Mv acMv[3],
uint32_t& ruiBits,
Distortion& ruiCost,
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ int& mvpIdx,
+ const AffineAMVPInfo& aamvpi,
+#endif
bool bBi = false
);
@@ -430,7 +434,11 @@ protected:
void xSymmetricMotionEstimation( PredictionUnit& pu, PelUnitBuf& origBuf, Mv& rcMvCurPred, Mv& rcMvTarPred, RefPicList eRefPicList, MvField& rCurMvField, MvField& rTarMvField, Distortion& ruiCost, int gbiIdx );
#endif
- bool xReadBufferedAffineUniMv ( PredictionUnit& pu, RefPicList eRefPicList, int32_t iRefIdx, Mv acMvPred[3], Mv acMv[3], uint32_t& ruiBits, Distortion& ruiCost);
+ bool xReadBufferedAffineUniMv ( PredictionUnit& pu, RefPicList eRefPicList, int32_t iRefIdx, Mv acMvPred[3], Mv acMv[3], uint32_t& ruiBits, Distortion& ruiCost
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ , int& mvpIdx, const AffineAMVPInfo& aamvpi
+#endif
+ );
double xGetMEDistortionWeight ( uint8_t gbiIdx, RefPicList eRefPicList);
bool xReadBufferedUniMv ( PredictionUnit& pu, RefPicList eRefPicList, int32_t iRefIdx, Mv& pcMvPred, Mv& rcMv, uint32_t& ruiBits, Distortion& ruiCost);
@@ -458,7 +466,9 @@ protected:
void xExtDIFUpSamplingH ( CPelBuf* pcPattern );
void xExtDIFUpSamplingQ ( CPelBuf* pcPatternKey, Mv halfPelRef );
-
+#if JVET_M0247_AFFINE_AMVR_ENCOPT
+ uint32_t xDetermineBestMvp ( PredictionUnit& pu, Mv acMvTemp[3], int& mvpIdx, const AffineAMVPInfo& aamvpi );
+#endif
// -------------------------------------------------------------------------------------------------------------------
// compute symbol bits
// -------------------------------------------------------------------------------------------------------------------