From e8c2ead982b18869653d82d285c6d5b5f1276fa2 Mon Sep 17 00:00:00 2001
From: Xiaoyu Xiu <jamesxxiu@gmail.com>
Date: Wed, 21 Jul 2021 04:33:54 +0000
Subject: [PATCH] JVET-W0097: GPM-MMVD and GPM-TM
---
source/Lib/CommonLib/Buffer.cpp | 95 +-
source/Lib/CommonLib/Buffer.h | 6 +-
source/Lib/CommonLib/CommonDef.h | 16 +-
source/Lib/CommonLib/ContextModelling.cpp | 112 +-
source/Lib/CommonLib/ContextModelling.h | 6 +-
source/Lib/CommonLib/Contexts.cpp | 40 +
source/Lib/CommonLib/Contexts.h | 4 +
source/Lib/CommonLib/InterPrediction.cpp | 84 +-
source/Lib/CommonLib/InterPrediction.h | 4 +
source/Lib/CommonLib/Reshape.cpp | 1 -
source/Lib/CommonLib/Rom.cpp | 19 +
source/Lib/CommonLib/Rom.h | 3 +
source/Lib/CommonLib/Slice.cpp | 6 +
source/Lib/CommonLib/Slice.h | 7 +
source/Lib/CommonLib/TypeDef.h | 12 +-
source/Lib/CommonLib/Unit.cpp | 30 +
source/Lib/CommonLib/Unit.h | 12 +-
source/Lib/CommonLib/UnitTools.cpp | 620 +++++++-
source/Lib/CommonLib/UnitTools.h | 140 +-
source/Lib/CommonLib/x86/BufferX86.h | 208 +++
source/Lib/DecoderLib/CABACReader.cpp | 140 ++
source/Lib/DecoderLib/CABACReader.h | 5 +
source/Lib/DecoderLib/DecCu.cpp | 54 +
source/Lib/DecoderLib/DecCu.h | 3 +
source/Lib/DecoderLib/VLCReader.cpp | 12 +
source/Lib/EncoderLib/CABACWriter.cpp | 223 ++-
source/Lib/EncoderLib/CABACWriter.h | 13 +
source/Lib/EncoderLib/EncCu.cpp | 1561 +++++++++++++++++++--
source/Lib/EncoderLib/EncCu.h | 78 +
source/Lib/EncoderLib/EncLib.cpp | 24 +
source/Lib/EncoderLib/EncModeCtrl.cpp | 40 +-
source/Lib/EncoderLib/EncModeCtrl.h | 41 +-
source/Lib/EncoderLib/EncReshape.cpp | 1 -
source/Lib/EncoderLib/InterSearch.cpp | 9 +
source/Lib/EncoderLib/VLCWriter.cpp | 6 +
35 files changed, 3495 insertions(+), 140 deletions(-)
diff --git a/source/Lib/CommonLib/Buffer.cpp b/source/Lib/CommonLib/Buffer.cpp
index 1c62b65a6..79852a83f 100644
--- a/source/Lib/CommonLib/Buffer.cpp
+++ b/source/Lib/CommonLib/Buffer.cpp
@@ -90,7 +90,71 @@ int64_t getSumOfDifferenceCore(const Pel* src0, int src0Stride, const Pel* src1,
return sum;
}
#endif
+#if JVET_W0097_GPM_MMVD_TM
+void roundBDCore(const Pel* srcp, const int srcStride, Pel* dest, const int destStride, int width, int height, const ClpRng& clpRng)
+{
+ const int32_t clipbd = clpRng.bd;
+#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+ const int32_t shiftDefault = IF_INTERNAL_FRAC_BITS(clipbd);
+#else
+ const int32_t shiftDefault = std::max<int>(2, (IF_INTERNAL_PREC - clipbd));
+#endif
+ const int32_t offsetDefault = (1 << (shiftDefault - 1)) + IF_INTERNAL_OFFS;
+
+ if (width == 1)
+ {
+ THROW("Blocks of width = 1 not supported");
+ }
+ else
+ {
+#define RND_OP( ADDR ) dest[ADDR] = ClipPel( rightShift( srcp[ADDR] + offsetDefault, shiftDefault), clpRng )
+#define RND_INC \
+ srcp += srcStride; \
+ dest += destStride; \
+
+ SIZE_AWARE_PER_EL_OP(RND_OP, RND_INC);
+
+#undef RND_OP
+#undef RND_INC
+ }
+}
+
+void weightedAvgCore(const Pel* src0, const unsigned src0Stride, const Pel* src1, const unsigned src1Stride, Pel* dest, const unsigned destStride, const int8_t w0, const int8_t w1, int width, int height, const ClpRng& clpRng)
+{
+ const int8_t log2WeightBase = g_BcwLog2WeightBase;
+ const int clipbd = clpRng.bd;
+#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+ const int shiftNum = IF_INTERNAL_FRAC_BITS(clipbd) + log2WeightBase;
+#else
+ const int shiftNum = std::max<int>(2, (IF_INTERNAL_PREC - clipbd)) + log2WeightBase;
+#endif
+ const int offset = (1 << (shiftNum - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
+
+#define ADD_AVG_OP( ADDR ) dest[ADDR] = ClipPel( rightShift( ( src0[ADDR]*w0 + src1[ADDR]*w1 + offset ), shiftNum ), clpRng )
+#define ADD_AVG_INC \
+ src0 += src0Stride; \
+ src1 += src1Stride; \
+ dest += destStride; \
+
+ SIZE_AWARE_PER_EL_OP(ADD_AVG_OP, ADD_AVG_INC);
+
+#undef ADD_AVG_OP
+#undef ADD_AVG_INC
+}
+void copyClipCore(const Pel* srcp, const unsigned srcStride, Pel* dest, const unsigned destStride, int width, int height, const ClpRng& clpRng)
+{
+#define RECO_OP( ADDR ) dest[ADDR] = ClipPel( srcp[ADDR], clpRng )
+#define RECO_INC \
+ srcp += srcStride; \
+ dest += destStride; \
+
+ SIZE_AWARE_PER_EL_OP(RECO_OP, RECO_INC);
+
+#undef RECO_OP
+#undef RECO_INC
+}
+#endif
template< typename T >
void addAvgCore( const T* src1, int src1Stride, const T* src2, int src2Stride, T* dest, int dstStride, int width, int height, int rshift, int offset, const ClpRng& clpRng )
{
@@ -521,6 +585,11 @@ void linTfCore( const T* src, int srcStride, Pel *dst, int dstStride, int width,
PelBufferOps::PelBufferOps()
{
+#if JVET_W0097_GPM_MMVD_TM
+ roundBD = roundBDCore;
+ weightedAvg = weightedAvgCore;
+ copyClip = copyClipCore;
+#endif
addAvg4 = addAvgCore<Pel>;
addAvg8 = addAvgCore<Pel>;
@@ -625,6 +694,19 @@ void AreaBuf<Pel>::addHypothesisAndClip(const AreaBuf<const Pel> &other, const i
template<>
void AreaBuf<Pel>::addWeightedAvg(const AreaBuf<const Pel> &other1, const AreaBuf<const Pel> &other2, const ClpRng& clpRng, const int8_t bcwIdx)
{
+#if JVET_W0097_GPM_MMVD_TM
+ const int8_t w0 = getBcwWeight(bcwIdx, REF_PIC_LIST_0);
+ const int8_t w1 = getBcwWeight(bcwIdx, REF_PIC_LIST_1);
+
+ const Pel* src0 = other1.buf;
+ const Pel* src1 = other2.buf;
+ Pel* dest = buf;
+ const unsigned src0Stride = other1.stride;
+ const unsigned src1Stride = other2.stride;
+ const unsigned destStride = stride;
+
+ g_pelBufOP.weightedAvg(src0, src0Stride, src1, src1Stride, dest, destStride, w0, w1, width, height, clpRng);
+#else
const int8_t w0 = getBcwWeight(bcwIdx, REF_PIC_LIST_0);
const int8_t w1 = getBcwWeight(bcwIdx, REF_PIC_LIST_1);
const int8_t log2WeightBase = g_BcwLog2WeightBase;
@@ -654,6 +736,7 @@ void AreaBuf<Pel>::addWeightedAvg(const AreaBuf<const Pel> &other1, const AreaBu
#undef ADD_AVG_OP
#undef ADD_AVG_INC
+#endif
}
template<>
@@ -919,6 +1002,9 @@ void AreaBuf<Pel>::copyClip( const AreaBuf<const Pel> &src, const ClpRng& clpRng
}
else
{
+#if JVET_W0097_GPM_MMVD_TM
+ g_pelBufOP.copyClip(srcp, srcStride, dest, destStride, width, height, clpRng);
+#else
#define RECO_OP( ADDR ) dest[ADDR] = ClipPel( srcp[ADDR], clpRng )
#define RECO_INC \
srcp += srcStride; \
@@ -928,6 +1014,7 @@ void AreaBuf<Pel>::copyClip( const AreaBuf<const Pel> &src, const ClpRng& clpRng
#undef RECO_OP
#undef RECO_INC
+#endif
}
}
@@ -938,7 +1025,7 @@ void AreaBuf<Pel>::roundToOutputBitdepth( const AreaBuf<const Pel> &src, const C
Pel* dest = buf;
const unsigned srcStride = src.stride;
const unsigned destStride = stride;
-
+#if !JVET_W0097_GPM_MMVD_TM
const int32_t clipbd = clpRng.bd;
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
const int32_t shiftDefault = IF_INTERNAL_FRAC_BITS(clipbd);
@@ -946,13 +1033,16 @@ void AreaBuf<Pel>::roundToOutputBitdepth( const AreaBuf<const Pel> &src, const C
const int32_t shiftDefault = std::max<int>(2, (IF_INTERNAL_PREC - clipbd));
#endif
const int32_t offsetDefault = (1<<(shiftDefault-1)) + IF_INTERNAL_OFFS;
-
+#endif
if( width == 1 )
{
THROW( "Blocks of width = 1 not supported" );
}
else
{
+#if JVET_W0097_GPM_MMVD_TM
+ g_pelBufOP.roundBD(srcp, srcStride, dest, destStride, width, height, clpRng);
+#else
#define RND_OP( ADDR ) dest[ADDR] = ClipPel( rightShift( srcp[ADDR] + offsetDefault, shiftDefault), clpRng )
#define RND_INC \
srcp += srcStride; \
@@ -962,6 +1052,7 @@ void AreaBuf<Pel>::roundToOutputBitdepth( const AreaBuf<const Pel> &src, const C
#undef RND_OP
#undef RND_INC
+#endif
}
}
diff --git a/source/Lib/CommonLib/Buffer.h b/source/Lib/CommonLib/Buffer.h
index 36f47f694..873dfbaf0 100644
--- a/source/Lib/CommonLib/Buffer.h
+++ b/source/Lib/CommonLib/Buffer.h
@@ -60,7 +60,11 @@ struct PelBufferOps
template<X86_VEXT vext>
void _initPelBufOpsX86();
#endif
-
+#if JVET_W0097_GPM_MMVD_TM
+ void(*roundBD) (const Pel* srcp, const int srcStride, Pel* dest, const int destStride, int width, int height, const ClpRng& clpRng);
+ void(*weightedAvg) (const Pel* src0, const unsigned src0Stride, const Pel* src1, const unsigned src1Stride, Pel* dest, const unsigned destStride, const int8_t w0, const int8_t w1, int width, int height, const ClpRng& clpRng);
+ void(*copyClip) (const Pel* srcp, const unsigned srcStride, Pel* dest, const unsigned destStride, int width, int height, const ClpRng& clpRng);
+#endif
void ( *addAvg4 ) ( const Pel* src0, int src0Stride, const Pel* src1, int src1Stride, Pel *dst, int dstStride, int width, int height, int shift, int offset, const ClpRng& clpRng );
void ( *addAvg8 ) ( const Pel* src0, int src0Stride, const Pel* src1, int src1Stride, Pel *dst, int dstStride, int width, int height, int shift, int offset, const ClpRng& clpRng );
void ( *reco4 ) ( const Pel* src0, int src0Stride, const Pel* src1, int src1Stride, Pel *dst, int dstStride, int width, int height, const ClpRng& clpRng );
diff --git a/source/Lib/CommonLib/CommonDef.h b/source/Lib/CommonLib/CommonDef.h
index 510b0f039..afc3b0e7d 100644
--- a/source/Lib/CommonLib/CommonDef.h
+++ b/source/Lib/CommonLib/CommonDef.h
@@ -456,7 +456,14 @@ static const int MMVD_MRG_MAX_RD_NUM = 20;
static const int MMVD_MRG_MAX_RD_NUM = MRG_MAX_NUM_CANDS;
#endif
static const int MMVD_MRG_MAX_RD_BUF_NUM = (MMVD_MRG_MAX_RD_NUM + 1);///< increase buffer size by 1
-
+#if JVET_W0097_GPM_MMVD_TM
+static const int GPM_MMVD_REFINE_STEP = 8;
+static const int GPM_MMVD_REFINE_DIRECTION = 4;
+static const int GPM_MMVD_MAX_REFINE_NUM = (GPM_MMVD_REFINE_STEP * GPM_MMVD_REFINE_DIRECTION);
+static const int GPM_EXT_MMVD_REFINE_STEP = 9;
+static const int GPM_EXT_MMVD_REFINE_DIRECTION = 8;
+static const int GPM_EXT_MMVD_MAX_REFINE_NUM = (GPM_EXT_MMVD_REFINE_STEP * GPM_EXT_MMVD_REFINE_DIRECTION);
+#endif
static const int MAX_TU_LEVEL_CTX_CODED_BIN_CONSTRAINT_LUMA = 28;
static const int MAX_TU_LEVEL_CTX_CODED_BIN_CONSTRAINT_CHROMA = 28;
@@ -618,7 +625,14 @@ static const int GEO_NUM_DISTANCES = 4;
static const int GEO_NUM_PRESTORED_MASK = 6;
static const int GEO_WEIGHT_MASK_SIZE = 3 * (GEO_MAX_CU_SIZE >> 3) * 2 + GEO_MAX_CU_SIZE;
static const int GEO_MV_MASK_SIZE = GEO_WEIGHT_MASK_SIZE >> 2;
+#if JVET_W0097_GPM_MMVD_TM
+static const int GEO_MAX_TRY_WEIGHTED_SAD = 70;
+#if TM_MRG
+static const int GEO_TM_MAX_NUM_CANDS = GEO_MAX_NUM_UNI_CANDS * (GEO_NUM_TM_MV_CAND - 1);
+#endif
+#else
static const int GEO_MAX_TRY_WEIGHTED_SAD = 60;
+#endif
static const int GEO_MAX_TRY_WEIGHTED_SATD = 8;
#if ENABLE_OBMC
diff --git a/source/Lib/CommonLib/ContextModelling.cpp b/source/Lib/CommonLib/ContextModelling.cpp
index e6db8469e..a86b14b13 100644
--- a/source/Lib/CommonLib/ContextModelling.cpp
+++ b/source/Lib/CommonLib/ContextModelling.cpp
@@ -513,7 +513,7 @@ void MergeCtx::setMergeInfo( PredictionUnit& pu, int candIdx )
#endif
}
-#if NON_ADJACENT_MRG_CAND || TM_MRG || MULTI_PASS_DMVR
+#if NON_ADJACENT_MRG_CAND || TM_MRG || MULTI_PASS_DMVR || JVET_W0097_GPM_MMVD_TM
bool MergeCtx::xCheckSimilarMotion(int mergeCandIndex, uint32_t mvdSimilarityThresh) const
{
if (mvFieldNeighbours[(mergeCandIndex << 1)].refIdx < 0 && mvFieldNeighbours[(mergeCandIndex << 1) + 1].refIdx < 0)
@@ -606,7 +606,117 @@ bool MergeCtx::xCheckSimilarMotion(int mergeCandIndex, uint32_t mvdSimilarityThr
return false;
}
#endif
+#if JVET_W0097_GPM_MMVD_TM
+void MergeCtx::setGeoMmvdMergeInfo(PredictionUnit& pu, int mergeIdx, int mmvdIdx)
+{
+ bool extMMVD = pu.cs->picHeader->getGPMMMVDTableFlag();
+ CHECK(mergeIdx >= numValidMergeCand, "Merge candidate does not exist");
+ CHECK(mmvdIdx >= (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM), "GPM MMVD index is invalid");
+ CHECK(!pu.cu->geoFlag || CU::isIBC(*pu.cu), "incorrect GPM setting")
+ pu.regularMergeFlag = !(pu.ciipFlag || pu.cu->geoFlag);
+ pu.mergeFlag = true;
+ pu.mmvdMergeFlag = false;
+ pu.interDir = interDirNeighbours[mergeIdx];
+ pu.cu->imv = 0;
+ pu.mergeIdx = mergeIdx;
+ pu.mergeType = MRG_TYPE_DEFAULT_N;
+#if MULTI_PASS_DMVR
+ pu.bdmvrRefine = false;
+#endif
+
+ const int mvShift = MV_FRACTIONAL_BITS_DIFF;
+ const int refMvdCands[8] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 16 << mvShift , 32 << mvShift, 64 << mvShift , 128 << mvShift };
+ const int refExtMvdCands[9] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 12 << mvShift , 16 << mvShift, 24 << mvShift, 32 << mvShift, 64 << mvShift };
+ int fPosStep = (extMMVD ? (mmvdIdx >> 3) : (mmvdIdx >> 2));
+ int fPosPosition = (extMMVD ? (mmvdIdx - (fPosStep << 3)) : (mmvdIdx - (fPosStep << 2)));
+ int offset = (extMMVD ? refExtMvdCands[fPosStep] : refMvdCands[fPosStep]);
+ Mv mvOffset;
+
+ if (fPosPosition == 0)
+ {
+ mvOffset = Mv(offset, 0);
+ }
+ else if (fPosPosition == 1)
+ {
+ mvOffset = Mv(-offset, 0);
+ }
+ else if (fPosPosition == 2)
+ {
+ mvOffset = Mv(0, offset);
+ }
+ else if (fPosPosition == 3)
+ {
+ mvOffset = Mv(0, -offset);
+ }
+ else if (fPosPosition == 4)
+ {
+ mvOffset = Mv(offset, offset);
+ }
+ else if (fPosPosition == 5)
+ {
+ mvOffset = Mv(offset, -offset);
+ }
+ else if (fPosPosition == 6)
+ {
+ mvOffset = Mv(-offset, offset);
+ }
+ else if (fPosPosition == 7)
+ {
+ mvOffset = Mv(-offset, -offset);
+ }
+
+ pu.refIdx[REF_PIC_LIST_0] = mvFieldNeighbours[(mergeIdx << 1) + 0].refIdx;
+ pu.refIdx[REF_PIC_LIST_1] = mvFieldNeighbours[(mergeIdx << 1) + 1].refIdx;
+ if (pu.refIdx[REF_PIC_LIST_0] >= 0)
+ {
+ pu.mv[REF_PIC_LIST_0] = mvFieldNeighbours[(mergeIdx << 1) + 0].mv + mvOffset;
+ }
+ else
+ {
+ pu.mv[REF_PIC_LIST_0] = Mv();
+ }
+
+ if (pu.refIdx[REF_PIC_LIST_1] >= 0)
+ {
+ pu.mv[REF_PIC_LIST_1] = mvFieldNeighbours[(mergeIdx << 1) + 1].mv + mvOffset;
+ }
+ else
+ {
+ pu.mv[REF_PIC_LIST_1] = Mv();
+ }
+ pu.mvd[REF_PIC_LIST_0] = Mv();
+ pu.mvd[REF_PIC_LIST_1] = Mv();
+ pu.mvpIdx[REF_PIC_LIST_0] = NOT_VALID;
+ pu.mvpIdx[REF_PIC_LIST_1] = NOT_VALID;
+ pu.mvpNum[REF_PIC_LIST_0] = NOT_VALID;
+ pu.mvpNum[REF_PIC_LIST_1] = NOT_VALID;
+ pu.cu->BcwIdx = (interDirNeighbours[mergeIdx] == 3) ? BcwIdx[mergeIdx] : BCW_DEFAULT;
+
+#if MULTI_HYP_PRED
+ pu.addHypData.clear();
+ pu.numMergedAddHyps = 0;
+#endif
+
+#if !INTER_RM_SIZE_CONSTRAINTS
+ PU::restrictBiPredMergeCandsOne(pu);
+#endif
+ pu.mmvdEncOptMode = 0;
+
+#if INTER_LIC
+ pu.cu->LICFlag = pu.cs->slice->getUseLIC() ? LICFlags[mergeIdx] : false;
+ if (pu.interDir == 3)
+ {
+ CHECK(pu.cu->LICFlag, "LIC is not used with bi-prediction in merge");
+ }
+#endif
+}
+void MergeCtx::copyMergeCtx(MergeCtx & orgMergeCtx)
+{
+ memcpy(interDirNeighbours, orgMergeCtx.interDirNeighbours, MRG_MAX_NUM_CANDS * sizeof(unsigned char));
+ memcpy(mvFieldNeighbours, orgMergeCtx.mvFieldNeighbours, (MRG_MAX_NUM_CANDS << 1) * sizeof(MvField));
+}
+#endif
void MergeCtx::setMmvdMergeCandiInfo(PredictionUnit& pu, int candIdx)
{
const Slice &slice = *pu.cs->slice;
diff --git a/source/Lib/CommonLib/ContextModelling.h b/source/Lib/CommonLib/ContextModelling.h
index dd33025f3..eda11d04e 100644
--- a/source/Lib/CommonLib/ContextModelling.h
+++ b/source/Lib/CommonLib/ContextModelling.h
@@ -568,13 +568,17 @@ public:
bool mmvdUseAltHpelIf [ MMVD_BASE_MV_NUM ];
bool useAltHpelIf [ MRG_MAX_NUM_CANDS ];
void setMergeInfo( PredictionUnit& pu, int candIdx );
-#if NON_ADJACENT_MRG_CAND || TM_MRG || MULTI_PASS_DMVR
+#if NON_ADJACENT_MRG_CAND || TM_MRG || MULTI_PASS_DMVR || JVET_W0097_GPM_MMVD_TM
bool xCheckSimilarMotion(int mergeCandIndex, uint32_t mvdSimilarityThresh = 1) const;
#endif
#if TM_MRG
void copyRegularMergeCand( int dstCandIdx, MergeCtx& srcCtx, int srcCandIdx );
void convertRegularMergeCandToBi(int candIdx);
#endif
+#if JVET_W0097_GPM_MMVD_TM
+ void setGeoMmvdMergeInfo(PredictionUnit& pu, int mergeIdx, int mmvdIdx);
+ void copyMergeCtx(MergeCtx &orgMergeCtx);
+#endif
};
class AffineMergeCtx
diff --git a/source/Lib/CommonLib/Contexts.cpp b/source/Lib/CommonLib/Contexts.cpp
index 1bd41905e..b5548efd0 100644
--- a/source/Lib/CommonLib/Contexts.cpp
+++ b/source/Lib/CommonLib/Contexts.cpp
@@ -853,6 +853,28 @@ const CtxSet ContextSetCfg::MmvdStepMvpIdx = ContextSetCfg::addCtxSet
{ 0 }
});
+#if JVET_W0097_GPM_MMVD_TM
+const CtxSet ContextSetCfg::GeoMmvdFlag = ContextSetCfg::addCtxSet
+({
+ { 25 },
+ { 33 },
+ { 35 },
+ { 5 },
+ { 4 },
+ { 4 }
+ });
+
+const CtxSet ContextSetCfg::GeoMmvdStepMvpIdx = ContextSetCfg::addCtxSet
+({
+ { 59 },
+ { 60 },
+ { 35 },
+ { 0 },
+ { 0 },
+ { 0 }
+ });
+#endif
+
#if AFFINE_MMVD
const CtxSet ContextSetCfg::AfMmvdFlag = ContextSetCfg::addCtxSet
({
@@ -2002,6 +2024,24 @@ const CtxSet ContextSetCfg::MmvdStepMvpIdx = ContextSetCfg::addCtxSet
{ 0, },
});
+#if JVET_W0097_GPM_MMVD_TM
+const CtxSet ContextSetCfg::GeoMmvdFlag = ContextSetCfg::addCtxSet
+({
+ { 25, },
+ { 26, },
+ { CNU, },
+ { 4, },
+ });
+
+const CtxSet ContextSetCfg::GeoMmvdStepMvpIdx = ContextSetCfg::addCtxSet
+({
+ { 59, },
+ { 60, },
+ { CNU, },
+ { 0, },
+ });
+#endif
+
#if AFFINE_MMVD
const CtxSet ContextSetCfg::AfMmvdFlag = ContextSetCfg::addCtxSet
({
diff --git a/source/Lib/CommonLib/Contexts.h b/source/Lib/CommonLib/Contexts.h
index a49689489..4f49ae343 100644
--- a/source/Lib/CommonLib/Contexts.h
+++ b/source/Lib/CommonLib/Contexts.h
@@ -257,6 +257,10 @@ public:
static const CtxSet MmvdFlag;
static const CtxSet MmvdMergeIdx;
static const CtxSet MmvdStepMvpIdx;
+#if JVET_W0097_GPM_MMVD_TM
+ static const CtxSet GeoMmvdFlag;
+ static const CtxSet GeoMmvdStepMvpIdx;
+#endif
static const CtxSet SubblockMergeFlag;
static const CtxSet AffineFlag;
static const CtxSet AffineType;
diff --git a/source/Lib/CommonLib/InterPrediction.cpp b/source/Lib/CommonLib/InterPrediction.cpp
index be9096990..3c5a701c4 100644
--- a/source/Lib/CommonLib/InterPrediction.cpp
+++ b/source/Lib/CommonLib/InterPrediction.cpp
@@ -626,10 +626,21 @@ void InterPrediction::xSubPuBio(PredictionUnit& pu, PelUnitBuf& predBuf, const R
int dx = secStep;
int dy = fstStep;
+#if !JVET_W0097_GPM_MMVD_TM
const MotionInfo &curMi = pu.getMotionInfo(Position{ x, y });
+#endif
subPu.UnitArea::operator=(UnitArea(pu.chromaFormat, Area(x, y, dx, dy)));
+#if JVET_W0097_GPM_MMVD_TM
+ subPu.interDir = pu.interDir;
+ for (uint32_t i = 0; i < NUM_REF_PIC_LIST_01; i++)
+ {
+ subPu.refIdx[i] = pu.refIdx[i];
+ subPu.mv[i] = pu.mv[i];
+ }
+#else
subPu = curMi;
+#endif
PelUnitBuf subPredBuf = predBuf.subBuf(UnitAreaRelative(pu, subPu));
if (yuvDstTmp)
@@ -2810,7 +2821,11 @@ void InterPrediction::xWeightedAverage(
{
if( pu.cu->geoFlag )
{
+#if JVET_W0097_GPM_MMVD_TM
+ pcYuvDst.copyFrom(pcYuvSrc0, lumaOnly, chromaOnly);
+#else
pcYuvDst.copyFrom( pcYuvSrc0 );
+#endif
}
else
{
@@ -2825,7 +2840,11 @@ void InterPrediction::xWeightedAverage(
{
if( pu.cu->geoFlag )
{
+#if JVET_W0097_GPM_MMVD_TM
+ pcYuvDst.copyFrom(pcYuvSrc1, lumaOnly, chromaOnly);
+#else
pcYuvDst.copyFrom( pcYuvSrc1 );
+#endif
}
else
{
@@ -3489,18 +3508,45 @@ int InterPrediction::rightShiftMSB(int numer, int denom)
return numer >> floorLog2(denom);
}
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+void InterPrediction::motionCompensationGeo(CodingUnit &cu, MergeCtx &geoMrgCtx, MergeCtx &geoTmMrgCtx0, MergeCtx &geoTmMrgCtx1)
+#else
void InterPrediction::motionCompensationGeo( CodingUnit &cu, MergeCtx &geoMrgCtx )
+#endif
{
const uint8_t splitDir = cu.firstPU->geoSplitDir;
const uint8_t candIdx0 = cu.firstPU->geoMergeIdx0;
const uint8_t candIdx1 = cu.firstPU->geoMergeIdx1;
+#if JVET_W0097_GPM_MMVD_TM
+ const bool geoMMVDFlag0 = cu.firstPU->geoMMVDFlag0;
+ const uint8_t geoMMVDIdx0 = cu.firstPU->geoMMVDIdx0;
+ const bool geoMMVDFlag1 = cu.firstPU->geoMMVDFlag1;
+ const uint8_t geoMMVDIdx1 = cu.firstPU->geoMMVDIdx1;
+#if TM_MRG
+ const bool geoTmFlag0 = cu.firstPU->geoTmFlag0;
+ const bool geoTmFlag1 = cu.firstPU->geoTmFlag1;
+#endif
+#endif
for( auto &pu : CU::traversePUs( cu ) )
{
const UnitArea localUnitArea( cu.cs->area.chromaFormat, Area( 0, 0, pu.lwidth(), pu.lheight() ) );
PelUnitBuf tmpGeoBuf0 = m_geoPartBuf[0].getBuf( localUnitArea );
PelUnitBuf tmpGeoBuf1 = m_geoPartBuf[1].getBuf( localUnitArea );
PelUnitBuf predBuf = cu.cs->getPredBuf( pu );
-
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ if (geoTmFlag0)
+ {
+ geoTmMrgCtx0.setMergeInfo(pu, candIdx0);
+ }
+ else
+#endif
+ if (geoMMVDFlag0)
+ {
+ geoMrgCtx.setGeoMmvdMergeInfo(pu, candIdx0, geoMMVDIdx0);
+ }
+ else
+#endif
geoMrgCtx.setMergeInfo( pu, candIdx0 );
motionCompensation(pu, tmpGeoBuf0, REF_PIC_LIST_X, true, isChromaEnabled(pu.chromaFormat)); // TODO: check 4:0:0 interaction with weighted prediction.
@@ -3508,7 +3554,20 @@ void InterPrediction::motionCompensationGeo( CodingUnit &cu, MergeCtx &geoMrgCtx
{
printf( "DECODER_GEO_PU: pu motion vector across tile boundaries (%d,%d,%d,%d)\n", pu.lx(), pu.ly(), pu.lwidth(), pu.lheight() );
}
-
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ if (geoTmFlag1)
+ {
+ geoTmMrgCtx1.setMergeInfo(pu, candIdx1);
+ }
+ else
+#endif
+ if (geoMMVDFlag1)
+ {
+ geoMrgCtx.setGeoMmvdMergeInfo(pu, candIdx1, geoMMVDIdx1);
+ }
+ else
+#endif
geoMrgCtx.setMergeInfo( pu, candIdx1 );
motionCompensation(pu, tmpGeoBuf1, REF_PIC_LIST_X, true, isChromaEnabled(pu.chromaFormat)); // TODO: check 4:0:0 interaction with weighted prediction.
@@ -4937,7 +4996,26 @@ bool TplMatchingCtrl::xFillCurTemplate(Pel* tpl)
const CPelBuf recBuf = currPic.getRecoBuf(m_cu.cs->picture->blocks[m_compID]);
std::vector<Pel>& invLUT = m_interRes.m_pcReshape->getInvLUT();
const bool useLUT = isLuma(m_compID) && m_cu.cs->picHeader->getLmcsEnabledFlag() && m_interRes.m_pcReshape->getCTUFlag();
-
+#if JVET_W0097_GPM_MMVD_TM & TM_MRG
+ if (m_cu.geoFlag)
+ {
+ CHECK(m_pu.geoTmType == GEO_TM_OFF, "invalid geo template type value");
+ if (m_pu.geoTmType == GEO_TM_SHAPE_A)
+ {
+ if (TrueA_FalseL == 0)
+ {
+ return false;
+ }
+ }
+ if (m_pu.geoTmType == GEO_TM_SHAPE_L)
+ {
+ if (TrueA_FalseL == 1)
+ {
+ return false;
+ }
+ }
+ }
+#endif
const Size dstSize = (TrueA_FalseL ? Size(m_pu.lwidth(), tplSize) : Size(tplSize, m_pu.lheight()));
for (int h = 0; h < (int)dstSize.height; h++)
{
diff --git a/source/Lib/CommonLib/InterPrediction.h b/source/Lib/CommonLib/InterPrediction.h
index 0333d8448..bfd8f5282 100644
--- a/source/Lib/CommonLib/InterPrediction.h
+++ b/source/Lib/CommonLib/InterPrediction.h
@@ -280,7 +280,11 @@ public:
#if ENABLE_OBMC
void subBlockOBMC(PredictionUnit &pu, PelUnitBuf *pDst = nullptr);
#endif
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+ void motionCompensationGeo(CodingUnit &cu, MergeCtx &geoMrgCtx, MergeCtx &geoTmMrgCtx0, MergeCtx &geoTmMrgCtx1);
+#else
void motionCompensationGeo(CodingUnit &cu, MergeCtx &GeoMrgCtx);
+#endif
void weightedGeoBlk(PredictionUnit &pu, const uint8_t splitDir, int32_t channel, PelUnitBuf& predDst, PelUnitBuf& predSrc0, PelUnitBuf& predSrc1);
void xPrefetch(PredictionUnit& pu, PelUnitBuf &pcPad, RefPicList refId, bool forLuma);
void xPad(PredictionUnit& pu, PelUnitBuf &pcPad, RefPicList refId);
diff --git a/source/Lib/CommonLib/Reshape.cpp b/source/Lib/CommonLib/Reshape.cpp
index bfe772ed6..24f11ab00 100644
--- a/source/Lib/CommonLib/Reshape.cpp
+++ b/source/Lib/CommonLib/Reshape.cpp
@@ -90,7 +90,6 @@ void Reshape::destroy()
}
-
/** compute chroma residuce scale for TU
* \param average luma pred of TU
* \return chroma residue scale
diff --git a/source/Lib/CommonLib/Rom.cpp b/source/Lib/CommonLib/Rom.cpp
index fb4317e06..6f327a8c6 100644
--- a/source/Lib/CommonLib/Rom.cpp
+++ b/source/Lib/CommonLib/Rom.cpp
@@ -4540,4 +4540,23 @@ int8_t g_angle2mirror[GEO_NUM_ANGLES] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1,
const int g_addHypWeight[MULTI_HYP_PRED_NUM_WEIGHTS] = { 2, -1 };
static_assert(g_BcwLog2WeightBase == MULTI_HYP_PRED_WEIGHT_BITS, "number of bits for gbi and multi-hyp weights do not match");
#endif
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+uint8_t g_geoTmShape[2][GEO_NUM_ANGLES] = {
+ { GEO_TM_SHAPE_A, 0, GEO_TM_SHAPE_A, GEO_TM_SHAPE_A,
+ GEO_TM_SHAPE_A, GEO_TM_SHAPE_AL, 0, 0,
+ GEO_TM_SHAPE_AL, 0, 0, GEO_TM_SHAPE_AL,
+ GEO_TM_SHAPE_AL, GEO_TM_SHAPE_A, GEO_TM_SHAPE_A, 0,
+ GEO_TM_SHAPE_A, 0, GEO_TM_SHAPE_A, GEO_TM_SHAPE_A,
+ GEO_TM_SHAPE_A, GEO_TM_SHAPE_AL, 0, 0,
+ GEO_TM_SHAPE_AL, 0, 0, GEO_TM_SHAPE_AL,
+ GEO_TM_SHAPE_AL, GEO_TM_SHAPE_A, GEO_TM_SHAPE_A, 0, },
+ { GEO_TM_SHAPE_AL, 0, GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL,
+ GEO_TM_SHAPE_L, GEO_TM_SHAPE_L, 0, 0,
+ GEO_TM_SHAPE_L, 0, 0, GEO_TM_SHAPE_L,
+ GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL, 0,
+ GEO_TM_SHAPE_AL, 0, GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL,
+ GEO_TM_SHAPE_L, GEO_TM_SHAPE_L, 0, 0,
+ GEO_TM_SHAPE_L, 0, 0, GEO_TM_SHAPE_L,
+ GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL, GEO_TM_SHAPE_AL, 0, } };
+#endif
//! \}
diff --git a/source/Lib/CommonLib/Rom.h b/source/Lib/CommonLib/Rom.h
index b41f205ba..7158a30c9 100644
--- a/source/Lib/CommonLib/Rom.h
+++ b/source/Lib/CommonLib/Rom.h
@@ -304,6 +304,9 @@ extern int16_t g_weightOffset [GEO_NUM_PARTITION_MODE][GEO_NUM_CU_SIZE][
extern int8_t g_angle2mask [GEO_NUM_ANGLES];
extern int8_t g_Dis[GEO_NUM_ANGLES];
extern int8_t g_angle2mirror[GEO_NUM_ANGLES];
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+extern uint8_t g_geoTmShape[2][GEO_NUM_ANGLES];
+#endif
#if MULTI_HYP_PRED
extern const int g_addHypWeight[MULTI_HYP_PRED_NUM_WEIGHTS];
#endif
diff --git a/source/Lib/CommonLib/Slice.cpp b/source/Lib/CommonLib/Slice.cpp
index d82e0bd29..77d4e6ba4 100644
--- a/source/Lib/CommonLib/Slice.cpp
+++ b/source/Lib/CommonLib/Slice.cpp
@@ -2819,6 +2819,9 @@ PicHeader::PicHeader()
, m_disDmvrFlag ( 0 )
, m_disProfFlag ( 0 )
, m_jointCbCrSignFlag ( 0 )
+#if JVET_W0097_GPM_MMVD_TM
+, m_gpmMMVDTableFlag (false)
+#endif
, m_qpDelta ( 0 )
, m_numAlfAps ( 0 )
, m_alfApsId ( 0 )
@@ -2911,6 +2914,9 @@ void PicHeader::initPicHeader()
m_disDmvrFlag = 0;
m_disProfFlag = 0;
m_jointCbCrSignFlag = 0;
+#if JVET_W0097_GPM_MMVD_TM
+ m_gpmMMVDTableFlag = 0;
+#endif
m_qpDelta = 0;
m_numAlfAps = 0;
m_alfChromaApsId = 0;
diff --git a/source/Lib/CommonLib/Slice.h b/source/Lib/CommonLib/Slice.h
index 70d7b8bc6..d2abdaa22 100644
--- a/source/Lib/CommonLib/Slice.h
+++ b/source/Lib/CommonLib/Slice.h
@@ -2658,6 +2658,9 @@ private:
bool m_disDmvrFlag; //!< picture level DMVR disable flag
bool m_disProfFlag; //!< picture level PROF disable flag
bool m_jointCbCrSignFlag; //!< joint Cb/Cr residual sign flag
+#if JVET_W0097_GPM_MMVD_TM
+ bool m_gpmMMVDTableFlag;
+#endif
int m_qpDelta; //!< value of Qp delta
bool m_saoEnabledFlag[MAX_NUM_CHANNEL_TYPE]; //!< sao enabled flags for each channel
#if ALF_IMPROVEMENT
@@ -2786,6 +2789,10 @@ public:
bool getDisProfFlag() const { return m_disProfFlag; }
void setJointCbCrSignFlag( bool b ) { m_jointCbCrSignFlag = b; }
bool getJointCbCrSignFlag() const { return m_jointCbCrSignFlag; }
+#if JVET_W0097_GPM_MMVD_TM
+ void setGPMMMVDTableFlag(bool b) { m_gpmMMVDTableFlag = b; }
+ bool getGPMMMVDTableFlag() const { return m_gpmMMVDTableFlag; }
+#endif
void setQpDelta(int b) { m_qpDelta = b; }
int getQpDelta() const { return m_qpDelta; }
void setSaoEnabledFlag(ChannelType chType, bool b) { m_saoEnabledFlag[chType] = b; }
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 63d251c26..740e8ddfa 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -126,6 +126,7 @@
#define TM_AMVP 1 // Add template matching to non-subblock inter to refine regular AMVP candidates
#define TM_MRG 1 // Add template matching to non-subblock inter to refine regular merge candidates
#define ENABLE_OBMC 1 // Enable Overlapped Block Motion Compensation
+#define JVET_W0097_GPM_MMVD_TM 1 // JVET-W0097: GPM-MMVD and GPM-TM
// Transform and coefficient coding
#define TCQ_8STATES 1
@@ -715,7 +716,16 @@ enum RefPicList
#define L0 REF_PIC_LIST_0
#define L1 REF_PIC_LIST_1
-
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+enum GeoTmMvCand
+{
+ GEO_TM_OFF = 0,
+ GEO_TM_SHAPE_AL,
+ GEO_TM_SHAPE_A,
+ GEO_TM_SHAPE_L,
+ GEO_NUM_TM_MV_CAND
+};
+#endif
/// distortion function index
enum DFunc
{
diff --git a/source/Lib/CommonLib/Unit.cpp b/source/Lib/CommonLib/Unit.cpp
index 4e28e52a2..9cfac21f2 100644
--- a/source/Lib/CommonLib/Unit.cpp
+++ b/source/Lib/CommonLib/Unit.cpp
@@ -637,6 +637,16 @@ void PredictionUnit::initData()
geoSplitDir = MAX_UCHAR;
geoMergeIdx0 = MAX_UCHAR;
geoMergeIdx1 = MAX_UCHAR;
+#if JVET_W0097_GPM_MMVD_TM
+ geoMMVDFlag0 = false;
+ geoMMVDIdx0 = MAX_UCHAR;
+ geoMMVDFlag1 = false;
+ geoMMVDIdx1 = MAX_UCHAR;
+#if TM_MRG
+ geoTmFlag0 = false;
+ geoTmFlag1 = false;
+#endif
+#endif
mmvdMergeFlag = false;
mmvdMergeIdx = MAX_UCHAR;
#if AFFINE_MMVD
@@ -717,6 +727,16 @@ PredictionUnit& PredictionUnit::operator=(const InterPredictionData& predData)
geoSplitDir = predData.geoSplitDir;
geoMergeIdx0 = predData.geoMergeIdx0;
geoMergeIdx1 = predData.geoMergeIdx1;
+#if JVET_W0097_GPM_MMVD_TM
+ geoMMVDFlag0 = predData.geoMMVDFlag0;
+ geoMMVDIdx0 = predData.geoMMVDIdx0;
+ geoMMVDFlag1 = predData.geoMMVDFlag1;
+ geoMMVDIdx1 = predData.geoMMVDIdx1;
+#if TM_MRG
+ geoTmFlag0 = predData.geoTmFlag0;
+ geoTmFlag1 = predData.geoTmFlag1;
+#endif
+#endif
mmvdMergeFlag = predData.mmvdMergeFlag;
mmvdMergeIdx = predData.mmvdMergeIdx;
#if AFFINE_MMVD
@@ -793,6 +813,16 @@ PredictionUnit& PredictionUnit::operator=( const PredictionUnit& other )
geoSplitDir = other.geoSplitDir;
geoMergeIdx0 = other.geoMergeIdx0;
geoMergeIdx1 = other.geoMergeIdx1;
+#if JVET_W0097_GPM_MMVD_TM
+ geoMMVDFlag0 = other.geoMMVDFlag0;
+ geoMMVDIdx0 = other.geoMMVDIdx0;
+ geoMMVDFlag1 = other.geoMMVDFlag1;
+ geoMMVDIdx1 = other.geoMMVDIdx1;
+#if TM_MRG
+ geoTmFlag0 = other.geoTmFlag0;
+ geoTmFlag1 = other.geoTmFlag1;
+#endif
+#endif
mmvdMergeFlag = other.mmvdMergeFlag;
mmvdMergeIdx = other.mmvdMergeIdx;
#if AFFINE_MMVD
diff --git a/source/Lib/CommonLib/Unit.h b/source/Lib/CommonLib/Unit.h
index 1bd52574d..4e1273375 100644
--- a/source/Lib/CommonLib/Unit.h
+++ b/source/Lib/CommonLib/Unit.h
@@ -431,6 +431,17 @@ struct InterPredictionData
uint8_t geoSplitDir;
uint8_t geoMergeIdx0;
uint8_t geoMergeIdx1;
+#if JVET_W0097_GPM_MMVD_TM
+ bool geoMMVDFlag0;
+ uint8_t geoMMVDIdx0;
+ bool geoMMVDFlag1;
+ uint8_t geoMMVDIdx1;
+#if TM_MRG
+ bool geoTmFlag0;
+ bool geoTmFlag1;
+ uint8_t geoTmType;
+#endif
+#endif
bool mmvdMergeFlag;
uint8_t mmvdMergeIdx;
#if AFFINE_MMVD
@@ -527,7 +538,6 @@ struct TransformUnit : public UnitArea
CodingStructure *cs;
ChannelType chType;
int m_chromaResScaleInv;
-
uint8_t depth;
uint8_t mtsIdx [ MAX_NUM_TBLOCKS ];
bool noResidual;
diff --git a/source/Lib/CommonLib/UnitTools.cpp b/source/Lib/CommonLib/UnitTools.cpp
index d8528c0f7..1e46844cf 100644
--- a/source/Lib/CommonLib/UnitTools.cpp
+++ b/source/Lib/CommonLib/UnitTools.cpp
@@ -1441,7 +1441,7 @@ void PU::getIBCMergeCandidates(const PredictionUnit &pu, MergeCtx& mrgCtx, const
mrgCtx.numValidMergeCand = cnt;
}
-#if MULTI_PASS_DMVR
+#if MULTI_PASS_DMVR || JVET_W0097_GPM_MMVD_TM
uint32_t PU::getBDMVRMvdThreshold(const PredictionUnit &pu)
{
uint32_t numPixels = pu.lwidth() * pu.lheight();
@@ -4777,7 +4777,15 @@ void PU::restrictBiPredMergeCandsOne(PredictionUnit &pu)
}
}
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+void PU::getGeoMergeCandidates(PredictionUnit &pu, MergeCtx& geoMrgCtx, MergeCtx* mergeCtx)
+#else
+void PU::getGeoMergeCandidates(const PredictionUnit &pu, MergeCtx& geoMrgCtx, MergeCtx* mergeCtx)
+#endif
+#else
void PU::getGeoMergeCandidates( const PredictionUnit &pu, MergeCtx& geoMrgCtx )
+#endif
{
MergeCtx tmpMergeCtx;
@@ -4797,9 +4805,26 @@ void PU::getGeoMergeCandidates( const PredictionUnit &pu, MergeCtx& geoMrgCtx )
geoMrgCtx.LICFlags[i] = false;
#endif
}
-
+#if JVET_W0097_GPM_MMVD_TM
+ if (mergeCtx == NULL)
+ {
+#if TM_MRG
+ const bool tmMergeFlag = pu.tmMergeFlag;
+ pu.tmMergeFlag = false;
+#endif
+#endif
PU::getInterMergeCandidates(pu, tmpMergeCtx, 0);
-
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ pu.tmMergeFlag = tmMergeFlag;
+#endif
+ }
+ else
+ {
+ memcpy(tmpMergeCtx.interDirNeighbours, mergeCtx->interDirNeighbours, maxNumMergeCand * sizeof(unsigned char));
+ memcpy(tmpMergeCtx.mvFieldNeighbours, mergeCtx->mvFieldNeighbours, (maxNumMergeCand << 1) * sizeof(MvField));
+ }
+#endif
for (int32_t i = 0; i < maxNumMergeCand; i++)
{
int parity = i & 1;
@@ -4810,6 +4835,12 @@ void PU::getGeoMergeCandidates( const PredictionUnit &pu, MergeCtx& geoMrgCtx )
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].mv = tmpMergeCtx.mvFieldNeighbours[(i << 1) + parity].mv;
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].refIdx = -1;
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].refIdx = tmpMergeCtx.mvFieldNeighbours[(i << 1) + parity].refIdx;
+#if JVET_W0097_GPM_MMVD_TM
+ if (geoMrgCtx.xCheckSimilarMotion(geoMrgCtx.numValidMergeCand, PU::getBDMVRMvdThreshold(pu)))
+ {
+ continue;
+ }
+#endif
geoMrgCtx.numValidMergeCand++;
if (geoMrgCtx.numValidMergeCand == GEO_MAX_NUM_UNI_CANDS)
{
@@ -4825,6 +4856,12 @@ void PU::getGeoMergeCandidates( const PredictionUnit &pu, MergeCtx& geoMrgCtx )
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].mv = Mv(0, 0);
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].refIdx = tmpMergeCtx.mvFieldNeighbours[(i << 1) + !parity].refIdx;
geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].refIdx = -1;
+#if JVET_W0097_GPM_MMVD_TM
+ if (geoMrgCtx.xCheckSimilarMotion(geoMrgCtx.numValidMergeCand, PU::getBDMVRMvdThreshold(pu)))
+ {
+ continue;
+ }
+#endif
geoMrgCtx.numValidMergeCand++;
if (geoMrgCtx.numValidMergeCand == GEO_MAX_NUM_UNI_CANDS)
{
@@ -4832,6 +4869,148 @@ void PU::getGeoMergeCandidates( const PredictionUnit &pu, MergeCtx& geoMrgCtx )
}
}
}
+#if JVET_W0097_GPM_MMVD_TM
+ // add more parity based geo candidates, in an opposite parity rule
+ if (geoMrgCtx.numValidMergeCand < pu.cs->sps->getMaxNumGeoCand())
+ {
+ for (int32_t i = 0; i < maxNumMergeCand; i++)
+ {
+ int parity = i & 1;
+ if (tmpMergeCtx.interDirNeighbours[i] == 3)
+ {
+ geoMrgCtx.interDirNeighbours[geoMrgCtx.numValidMergeCand] = 2 - parity;
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].mv = tmpMergeCtx.mvFieldNeighbours[(i << 1) + !parity].mv;
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].mv = Mv(0, 0);
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].refIdx = tmpMergeCtx.mvFieldNeighbours[(i << 1) + !parity].refIdx;
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].refIdx = -1;
+ if (geoMrgCtx.xCheckSimilarMotion(geoMrgCtx.numValidMergeCand, PU::getBDMVRMvdThreshold(pu)))
+ {
+ continue;
+ }
+ geoMrgCtx.numValidMergeCand++;
+ if (geoMrgCtx.numValidMergeCand == pu.cs->sps->getMaxNumGeoCand())
+ {
+ return;
+ }
+ }
+ }
+ }
+
+ // add at most two average based geo candidates
+ if (geoMrgCtx.numValidMergeCand < pu.cs->sps->getMaxNumGeoCand())
+ {
+ // add one L0 cand by averaging the first two available L0 candidates
+ int cnt = 0;
+ int firstAvailRefIdx = -1;
+ Mv avgMv;
+ avgMv.setZero();
+ for (int i = 0; i < geoMrgCtx.numValidMergeCand; i++)
+ {
+ if (cnt == 2)
+ {
+ break;
+ }
+ if (geoMrgCtx.interDirNeighbours[i] == 1)
+ {
+ avgMv += geoMrgCtx.mvFieldNeighbours[i * 2].mv;
+ if (firstAvailRefIdx == -1)
+ {
+ firstAvailRefIdx = geoMrgCtx.mvFieldNeighbours[i * 2].refIdx;
+ }
+ cnt++;
+ }
+ }
+ if (cnt == 2)
+ {
+ roundAffineMv(avgMv.hor, avgMv.ver, 1);
+ geoMrgCtx.interDirNeighbours[geoMrgCtx.numValidMergeCand] = 1;
+ geoMrgCtx.mvFieldNeighbours[geoMrgCtx.numValidMergeCand * 2].setMvField(avgMv, firstAvailRefIdx);
+ geoMrgCtx.mvFieldNeighbours[geoMrgCtx.numValidMergeCand * 2 + 1].setMvField(Mv(0, 0), NOT_VALID);
+
+ if (!geoMrgCtx.xCheckSimilarMotion(geoMrgCtx.numValidMergeCand, PU::getBDMVRMvdThreshold(pu)))
+ {
+ geoMrgCtx.numValidMergeCand++;
+ }
+ if (geoMrgCtx.numValidMergeCand == pu.cs->sps->getMaxNumGeoCand())
+ {
+ return;
+ }
+ }
+
+ // add one L1 cand by averaging the first two available L1 candidates
+ cnt = 0;
+ firstAvailRefIdx = -1;
+ avgMv.setZero();
+ for (int i = 0; i < geoMrgCtx.numValidMergeCand; i++)
+ {
+ if (cnt == 2)
+ {
+ break;
+ }
+ if (geoMrgCtx.interDirNeighbours[i] == 2)
+ {
+ avgMv += geoMrgCtx.mvFieldNeighbours[i * 2 + 1].mv;
+ if (firstAvailRefIdx == -1)
+ {
+ firstAvailRefIdx = geoMrgCtx.mvFieldNeighbours[i * 2 + 1].refIdx;
+ }
+ cnt++;
+ }
+ }
+ if (cnt == 2)
+ {
+ roundAffineMv(avgMv.hor, avgMv.ver, 1);
+ geoMrgCtx.interDirNeighbours[geoMrgCtx.numValidMergeCand] = 2;
+ geoMrgCtx.mvFieldNeighbours[geoMrgCtx.numValidMergeCand * 2 + 1].setMvField(avgMv, firstAvailRefIdx);
+ geoMrgCtx.mvFieldNeighbours[geoMrgCtx.numValidMergeCand * 2].setMvField(Mv(0, 0), NOT_VALID);
+ if (!geoMrgCtx.xCheckSimilarMotion(geoMrgCtx.numValidMergeCand, PU::getBDMVRMvdThreshold(pu)))
+ {
+ geoMrgCtx.numValidMergeCand++;
+ }
+ if (geoMrgCtx.numValidMergeCand == pu.cs->sps->getMaxNumGeoCand())
+ {
+ return;
+ }
+ }
+ }
+ if (geoMrgCtx.numValidMergeCand < pu.cs->sps->getMaxNumGeoCand())
+ {
+ const Slice &slice = *pu.cs->slice;
+ int iNumRefIdx = std::min(slice.getNumRefIdx(REF_PIC_LIST_0), slice.getNumRefIdx(REF_PIC_LIST_1));
+
+ int r = 0;
+ int refcnt = 0;
+
+ for (int32_t i = geoMrgCtx.numValidMergeCand; i < pu.cs->sps->getMaxNumGeoCand(); i++)
+ {
+ int parity = i & 1;
+ if (0x01 + parity)
+ {
+ geoMrgCtx.interDirNeighbours[geoMrgCtx.numValidMergeCand] = 1 + parity;
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].mv = Mv(0, 0);
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].mv = Mv(0, 0);
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + !parity].refIdx = -1;
+ geoMrgCtx.mvFieldNeighbours[(geoMrgCtx.numValidMergeCand << 1) + parity].refIdx = r;
+
+ if (refcnt == iNumRefIdx - 1)
+ {
+ r = 0;
+ }
+ else
+ {
+ ++r;
+ ++refcnt;
+ }
+
+ geoMrgCtx.numValidMergeCand++;
+ if (geoMrgCtx.numValidMergeCand == pu.cs->sps->getMaxNumGeoCand())
+ {
+ return;
+ }
+ }
+ }
+ }
+#endif
}
void PU::spanGeoMotionInfo( PredictionUnit &pu, MergeCtx &geoMrgCtx, const uint8_t splitDir, const uint8_t candIdx0, const uint8_t candIdx1)
@@ -4950,6 +5129,441 @@ void PU::spanGeoMotionInfo( PredictionUnit &pu, MergeCtx &geoMrgCtx, const uint8
#endif
}
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+void PU::spanGeoMMVDMotionInfo(PredictionUnit &pu, MergeCtx &geoMrgCtx, MergeCtx &geoTmMrgCtx0, MergeCtx &geoTmMrgCtx1, const uint8_t splitDir, const uint8_t mergeIdx0, const uint8_t mergeIdx1, const bool tmFlag0, const bool mmvdFlag0, const uint8_t mmvdIdx0, const bool tmFlag1, const bool mmvdFlag1, const uint8_t mmvdIdx1)
+{
+ pu.geoSplitDir = splitDir;
+ pu.geoMergeIdx0 = mergeIdx0;
+ pu.geoMergeIdx1 = mergeIdx1;
+ pu.geoTmFlag0 = tmFlag0;
+ pu.geoMMVDFlag0 = mmvdFlag0;
+ pu.geoMMVDIdx0 = mmvdIdx0;
+ pu.geoTmFlag1 = tmFlag1;
+ pu.geoMMVDFlag1 = mmvdFlag1;
+ pu.geoMMVDIdx1 = mmvdIdx1;
+
+ MergeCtx *mergeCtx0 = (tmFlag0 ? &geoTmMrgCtx0 : &geoMrgCtx);
+ MergeCtx *mergeCtx1 = (tmFlag1 ? &geoTmMrgCtx1 : &geoMrgCtx);
+
+ const int mvShift = MV_FRACTIONAL_BITS_DIFF;
+ const bool extMMVD = pu.cs->picHeader->getGPMMMVDTableFlag();
+ const int MmvdCands[8] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 16 << mvShift , 32 << mvShift, 64 << mvShift , 128 << mvShift };
+ const int ExtMmvdCands[9] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 12 << mvShift , 16 << mvShift, 24 << mvShift, 32 << mvShift, 64 << mvShift };
+ const int* refMvdCands = (extMMVD ? ExtMmvdCands : MmvdCands);
+ Mv mvOffset0[2], mvOffset1[2], deltaMv;
+
+ if (mmvdFlag0)
+ {
+ int fPosStep = (extMMVD ? (mmvdIdx0 >> 3) : (mmvdIdx0 >> 2));
+ int fPosPosition = (extMMVD ? (mmvdIdx0 - (fPosStep << 3)) : (mmvdIdx0 - (fPosStep << 2)));
+
+ if (fPosPosition == 0)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 1)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 2)
+ {
+ deltaMv = Mv(0, refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 3)
+ {
+ deltaMv = Mv(0, -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 4)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 5)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 6)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 7)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+
+ if (mergeCtx0->interDirNeighbours[mergeIdx0] == 1)
+ {
+ mvOffset0[0] = deltaMv;
+ }
+ else
+ {
+ mvOffset0[1] = deltaMv;
+ }
+ }
+
+ if (mmvdFlag1)
+ {
+ int fPosStep = (extMMVD ? (mmvdIdx1 >> 3) : (mmvdIdx1 >> 2));
+ int fPosPosition = (extMMVD ? (mmvdIdx1 - (fPosStep << 3)) : (mmvdIdx1 - (fPosStep << 2)));
+
+ if (fPosPosition == 0)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 1)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 2)
+ {
+ deltaMv = Mv(0, refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 3)
+ {
+ deltaMv = Mv(0, -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 4)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 5)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 6)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 7)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+
+ if (mergeCtx1->interDirNeighbours[mergeIdx1] == 1)
+ {
+ mvOffset1[0] = deltaMv;
+ }
+ else
+ {
+ mvOffset1[1] = deltaMv;
+ }
+ }
+
+ MotionBuf mb = pu.getMotionBuf();
+
+ MotionInfo biMv;
+ biMv.isInter = true;
+ biMv.sliceIdx = pu.cs->slice->getIndependentSliceIdx();
+
+ if (mergeCtx0->interDirNeighbours[mergeIdx0] == 1 && mergeCtx1->interDirNeighbours[mergeIdx1] == 2)
+ {
+ biMv.interDir = 3;
+ biMv.mv[0] = mergeCtx0->mvFieldNeighbours[mergeIdx0 << 1].mv + mvOffset0[0];
+ biMv.mv[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ biMv.refIdx[0] = mergeCtx0->mvFieldNeighbours[mergeIdx0 << 1].refIdx;
+ biMv.refIdx[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ }
+ else if (mergeCtx0->interDirNeighbours[mergeIdx0] == 2 && mergeCtx1->interDirNeighbours[mergeIdx1] == 1)
+ {
+ biMv.interDir = 3;
+ biMv.mv[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ biMv.mv[1] = mergeCtx0->mvFieldNeighbours[(mergeIdx0 << 1) + 1].mv + mvOffset0[1];
+ biMv.refIdx[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ biMv.refIdx[1] = mergeCtx0->mvFieldNeighbours[(mergeIdx0 << 1) + 1].refIdx;
+ }
+ else if (mergeCtx0->interDirNeighbours[mergeIdx0] == 1 && mergeCtx1->interDirNeighbours[mergeIdx1] == 1)
+ {
+ biMv.interDir = 1;
+ biMv.mv[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ biMv.mv[1] = Mv(0, 0);
+ biMv.refIdx[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ biMv.refIdx[1] = -1;
+ }
+ else if (mergeCtx0->interDirNeighbours[mergeIdx0] == 2 && mergeCtx1->interDirNeighbours[mergeIdx1] == 2)
+ {
+ biMv.interDir = 2;
+ biMv.mv[0] = Mv(0, 0);
+ biMv.mv[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ biMv.refIdx[0] = -1;
+ biMv.refIdx[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ }
+
+ int16_t angle = g_GeoParams[splitDir][0];
+ int tpmMask = 0;
+ int lookUpY = 0, motionIdx = 0;
+ bool isFlip = angle >= 13 && angle <= 27;
+ int distanceIdx = g_GeoParams[splitDir][1];
+ int distanceX = angle;
+ int distanceY = (distanceX + (GEO_NUM_ANGLES >> 2)) % GEO_NUM_ANGLES;
+ int offsetX = (-(int)pu.lwidth()) >> 1;
+ int offsetY = (-(int)pu.lheight()) >> 1;
+ if (distanceIdx > 0)
+ {
+ if (angle % 16 == 8 || (angle % 16 != 0 && pu.lheight() >= pu.lwidth()))
+ {
+ offsetY += angle < 16 ? ((distanceIdx * pu.lheight()) >> 3) : -(int)((distanceIdx * pu.lheight()) >> 3);
+ }
+ else
+ {
+ offsetX += angle < 16 ? ((distanceIdx * pu.lwidth()) >> 3) : -(int)((distanceIdx * pu.lwidth()) >> 3);
+ }
+ }
+ for (int y = 0; y < mb.height; y++)
+ {
+ lookUpY = (((4 * y + offsetY) << 1) + 5) * g_Dis[distanceY];
+ for (int x = 0; x < mb.width; x++)
+ {
+ motionIdx = (((4 * x + offsetX) << 1) + 5) * g_Dis[distanceX] + lookUpY;
+ tpmMask = abs(motionIdx) < 32 ? 2 : (motionIdx <= 0 ? (1 - isFlip) : isFlip);
+ if (tpmMask == 2)
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = biMv.interDir;
+ mb.at(x, y).refIdx[0] = biMv.refIdx[0];
+ mb.at(x, y).refIdx[1] = biMv.refIdx[1];
+ mb.at(x, y).mv[0] = biMv.mv[0];
+ mb.at(x, y).mv[1] = biMv.mv[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ else if (tpmMask == 0)
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = mergeCtx0->interDirNeighbours[mergeIdx0];
+ mb.at(x, y).refIdx[0] = mergeCtx0->mvFieldNeighbours[mergeIdx0 << 1].refIdx;
+ mb.at(x, y).refIdx[1] = mergeCtx0->mvFieldNeighbours[(mergeIdx0 << 1) + 1].refIdx;
+ mb.at(x, y).mv[0] = mergeCtx0->mvFieldNeighbours[mergeIdx0 << 1].mv + mvOffset0[0];
+ mb.at(x, y).mv[1] = mergeCtx0->mvFieldNeighbours[(mergeIdx0 << 1) + 1].mv + mvOffset0[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ else
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = mergeCtx1->interDirNeighbours[mergeIdx1];
+ mb.at(x, y).refIdx[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ mb.at(x, y).refIdx[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ mb.at(x, y).mv[0] = mergeCtx1->mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ mb.at(x, y).mv[1] = mergeCtx1->mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ }
+ }
+}
+#else
+void PU::spanGeoMMVDMotionInfo(PredictionUnit &pu, MergeCtx &geoMrgCtx, const uint8_t splitDir, const uint8_t mergeIdx0, const uint8_t mergeIdx1, const bool mmvdFlag0, const uint8_t mmvdIdx0, const bool mmvdFlag1, const uint8_t mmvdIdx1)
+{
+ pu.geoSplitDir = splitDir;
+ pu.geoMergeIdx0 = mergeIdx0;
+ pu.geoMergeIdx1 = mergeIdx1;
+ pu.geoMMVDFlag0 = mmvdFlag0;
+ pu.geoMMVDIdx0 = mmvdIdx0;
+ pu.geoMMVDFlag1 = mmvdFlag1;
+ pu.geoMMVDIdx1 = mmvdIdx1;
+
+ const int mvShift = MV_FRACTIONAL_BITS_DIFF;
+ const bool extMMVD = pu.cs->picHeader->getGPMMMVDTableFlag();
+ const int MmvdCands[8] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 16 << mvShift , 32 << mvShift, 64 << mvShift , 128 << mvShift };
+ const int ExtMmvdCands[9] = { 1 << mvShift , 2 << mvShift , 4 << mvShift , 8 << mvShift , 12 << mvShift , 16 << mvShift, 24 << mvShift, 32 << mvShift, 64 << mvShift };
+ const int* refMvdCands = (extMMVD ? ExtMmvdCands : MmvdCands);
+ Mv mvOffset0[2], mvOffset1[2], deltaMv;
+
+ if (mmvdFlag0)
+ {
+ int fPosStep = (extMMVD ? (mmvdIdx0 >> 3) : (mmvdIdx0 >> 2));
+ int fPosPosition = (extMMVD ? (mmvdIdx0 - (fPosStep << 3)) : (mmvdIdx0 - (fPosStep << 2)));
+
+ if (fPosPosition == 0)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 1)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 2)
+ {
+ deltaMv = Mv(0, refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 3)
+ {
+ deltaMv = Mv(0, -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 4)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 5)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 6)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 7)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+
+ if (geoMrgCtx.interDirNeighbours[mergeIdx0] == 1)
+ {
+ mvOffset0[0] = deltaMv;
+ }
+ else
+ {
+ mvOffset0[1] = deltaMv;
+ }
+ }
+
+ if (mmvdFlag1)
+ {
+ int fPosStep = (extMMVD ? (mmvdIdx1 >> 3) : (mmvdIdx1 >> 2));
+ int fPosPosition = (extMMVD ? (mmvdIdx1 - (fPosStep << 3)) : (mmvdIdx1 - (fPosStep << 2)));
+
+ if (fPosPosition == 0)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 1)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], 0);
+ }
+ else if (fPosPosition == 2)
+ {
+ deltaMv = Mv(0, refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 3)
+ {
+ deltaMv = Mv(0, -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 4)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 5)
+ {
+ deltaMv = Mv(refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 6)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], refMvdCands[fPosStep]);
+ }
+ else if (fPosPosition == 7)
+ {
+ deltaMv = Mv(-refMvdCands[fPosStep], -refMvdCands[fPosStep]);
+ }
+
+ if (geoMrgCtx.interDirNeighbours[mergeIdx1] == 1)
+ {
+ mvOffset1[0] = deltaMv;
+ }
+ else
+ {
+ mvOffset1[1] = deltaMv;
+ }
+ }
+
+ MotionBuf mb = pu.getMotionBuf();
+
+ MotionInfo biMv;
+ biMv.isInter = true;
+ biMv.sliceIdx = pu.cs->slice->getIndependentSliceIdx();
+
+ if (geoMrgCtx.interDirNeighbours[mergeIdx0] == 1 && geoMrgCtx.interDirNeighbours[mergeIdx1] == 2)
+ {
+ biMv.interDir = 3;
+ biMv.mv[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx0 << 1].mv + mvOffset0[0];
+ biMv.mv[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ biMv.refIdx[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx0 << 1].refIdx;
+ biMv.refIdx[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ }
+ else if (geoMrgCtx.interDirNeighbours[mergeIdx0] == 2 && geoMrgCtx.interDirNeighbours[mergeIdx1] == 1)
+ {
+ biMv.interDir = 3;
+ biMv.mv[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ biMv.mv[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx0 << 1) + 1].mv + mvOffset0[1];
+ biMv.refIdx[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ biMv.refIdx[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx0 << 1) + 1].refIdx;
+ }
+ else if (geoMrgCtx.interDirNeighbours[mergeIdx0] == 1 && geoMrgCtx.interDirNeighbours[mergeIdx1] == 1)
+ {
+ biMv.interDir = 1;
+ biMv.mv[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ biMv.mv[1] = Mv(0, 0);
+ biMv.refIdx[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ biMv.refIdx[1] = -1;
+ }
+ else if (geoMrgCtx.interDirNeighbours[mergeIdx0] == 2 && geoMrgCtx.interDirNeighbours[mergeIdx1] == 2)
+ {
+ biMv.interDir = 2;
+ biMv.mv[0] = Mv(0, 0);
+ biMv.mv[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ biMv.refIdx[0] = -1;
+ biMv.refIdx[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ }
+
+ int16_t angle = g_GeoParams[splitDir][0];
+ int tpmMask = 0;
+ int lookUpY = 0, motionIdx = 0;
+ bool isFlip = angle >= 13 && angle <= 27;
+ int distanceIdx = g_GeoParams[splitDir][1];
+ int distanceX = angle;
+ int distanceY = (distanceX + (GEO_NUM_ANGLES >> 2)) % GEO_NUM_ANGLES;
+ int offsetX = (-(int)pu.lwidth()) >> 1;
+ int offsetY = (-(int)pu.lheight()) >> 1;
+ if (distanceIdx > 0)
+ {
+ if (angle % 16 == 8 || (angle % 16 != 0 && pu.lheight() >= pu.lwidth()))
+ {
+ offsetY += angle < 16 ? ((distanceIdx * pu.lheight()) >> 3) : -(int)((distanceIdx * pu.lheight()) >> 3);
+ }
+ else
+ {
+ offsetX += angle < 16 ? ((distanceIdx * pu.lwidth()) >> 3) : -(int)((distanceIdx * pu.lwidth()) >> 3);
+ }
+ }
+ for (int y = 0; y < mb.height; y++)
+ {
+ lookUpY = (((4 * y + offsetY) << 1) + 5) * g_Dis[distanceY];
+ for (int x = 0; x < mb.width; x++)
+ {
+ motionIdx = (((4 * x + offsetX) << 1) + 5) * g_Dis[distanceX] + lookUpY;
+ tpmMask = abs(motionIdx) < 32 ? 2 : (motionIdx <= 0 ? (1 - isFlip) : isFlip);
+ if (tpmMask == 2)
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = biMv.interDir;
+ mb.at(x, y).refIdx[0] = biMv.refIdx[0];
+ mb.at(x, y).refIdx[1] = biMv.refIdx[1];
+ mb.at(x, y).mv[0] = biMv.mv[0];
+ mb.at(x, y).mv[1] = biMv.mv[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ else if (tpmMask == 0)
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = geoMrgCtx.interDirNeighbours[mergeIdx0];
+ mb.at(x, y).refIdx[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx0 << 1].refIdx;
+ mb.at(x, y).refIdx[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx0 << 1) + 1].refIdx;
+ mb.at(x, y).mv[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx0 << 1].mv + mvOffset0[0];
+ mb.at(x, y).mv[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx0 << 1) + 1].mv + mvOffset0[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ else
+ {
+ mb.at(x, y).isInter = true;
+ mb.at(x, y).interDir = geoMrgCtx.interDirNeighbours[mergeIdx1];
+ mb.at(x, y).refIdx[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].refIdx;
+ mb.at(x, y).refIdx[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].refIdx;
+ mb.at(x, y).mv[0] = geoMrgCtx.mvFieldNeighbours[mergeIdx1 << 1].mv + mvOffset1[0];
+ mb.at(x, y).mv[1] = geoMrgCtx.mvFieldNeighbours[(mergeIdx1 << 1) + 1].mv + mvOffset1[1];
+ mb.at(x, y).sliceIdx = biMv.sliceIdx;
+ }
+ }
+ }
+}
+#endif
+#endif
+
bool CU::hasSubCUNonZeroMVd( const CodingUnit& cu )
{
bool bNonZeroMvd = false;
diff --git a/source/Lib/CommonLib/UnitTools.h b/source/Lib/CommonLib/UnitTools.h
index 0f1fed6c9..7f781108b 100644
--- a/source/Lib/CommonLib/UnitTools.h
+++ b/source/Lib/CommonLib/UnitTools.h
@@ -155,7 +155,7 @@ namespace PU
uint32_t getFinalIntraMode (const PredictionUnit &pu, const ChannelType &chType);
uint32_t getCoLocatedIntraLumaMode (const PredictionUnit &pu);
int getWideAngle ( const TransformUnit &tu, const uint32_t dirMode, const ComponentID compID );
-#if MULTI_PASS_DMVR
+#if MULTI_PASS_DMVR || JVET_W0097_GPM_MMVD_TM
uint32_t getBDMVRMvdThreshold (const PredictionUnit &pu);
#endif
#if TM_MRG
@@ -236,8 +236,23 @@ namespace PU
bool isLMCModeEnabled (const PredictionUnit &pu, unsigned mode);
bool isChromaIntraModeCrossCheckMode(const PredictionUnit &pu);
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ void getGeoMergeCandidates(PredictionUnit &pu, MergeCtx &GeoMrgCtx, MergeCtx* mergeCtx = NULL);
+#else
+ void getGeoMergeCandidates(const PredictionUnit &pu, MergeCtx &GeoMrgCtx, MergeCtx* mergeCtx = NULL);
+#endif
+#else
void getGeoMergeCandidates (const PredictionUnit &pu, MergeCtx &GeoMrgCtx);
+#endif
void spanGeoMotionInfo ( PredictionUnit &pu, MergeCtx &GeoMrgCtx, const uint8_t splitDir, const uint8_t candIdx0, const uint8_t candIdx1);
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ void spanGeoMMVDMotionInfo(PredictionUnit &pu, MergeCtx &geoMrgCtx, MergeCtx &geoTmMrgCtx0, MergeCtx &geoTmMrgCtx1, const uint8_t splitDir, const uint8_t mergeIdx0, const uint8_t mergeIdx1, const bool tmFlag0, const bool mmvdFlag0, const uint8_t mmvdIdx0, const bool tmFlag1, const bool mmvdFlag1, const uint8_t mmvdIdx1);
+#else
+ void spanGeoMMVDMotionInfo(PredictionUnit &pu, MergeCtx &GeoMrgCtx, const uint8_t splitDir, const uint8_t mergeIdx0, const uint8_t mergeIdx1, const bool mmvdFlag0, const uint8_t mmvdIdx0, const bool mmvdFlag1, const uint8_t mmvdIdx1);
+#endif
+#endif
bool isAddNeighborMv (const Mv& currMv, Mv* neighborMvs, int numNeighborMv);
void getIbcMVPsEncOnly(PredictionUnit &pu, Mv* mvPred, int& nbPred);
bool getDerivedBV(PredictionUnit &pu, const Mv& currentMv, Mv& derivedMv);
@@ -337,5 +352,128 @@ uint32_t updateCandList(T uiMode, double uiCost, static_vector<T, N>& candModeLi
}
return 0;
}
+#if JVET_W0097_GPM_MMVD_TM
+template<size_t N>
+void orderCandList(uint8_t uiMode, bool bNonMMVDListCand, int splitDir, double uiCost, static_vector<uint8_t, N>& candModeList, static_vector<bool, N>& isNonMMVDListIdx, static_vector<int, N>& candSplitDirList, static_vector<double, N>& candCostList, size_t uiFastCandNum = N)
+{
+ CHECK(std::min(uiFastCandNum, candModeList.size()) != std::min(uiFastCandNum, candCostList.size()), "Sizes do not match!");
+ CHECK(uiFastCandNum > candModeList.capacity(), "The vector is to small to hold all the candidates!");
+
+ size_t i;
+ size_t shift = 0;
+ size_t currSize = std::min(uiFastCandNum, candCostList.size());
+
+ while (shift < uiFastCandNum && shift < currSize && uiCost < candCostList[currSize - 1 - shift])
+ {
+ shift++;
+ }
+
+ if (candModeList.size() >= uiFastCandNum && shift != 0)
+ {
+ for (i = 1; i < shift; i++)
+ {
+ candModeList[currSize - i] = candModeList[currSize - 1 - i];
+ isNonMMVDListIdx[currSize - i] = isNonMMVDListIdx[currSize - 1 - i];
+ candSplitDirList[currSize - i] = candSplitDirList[currSize - 1 - i];
+ candCostList[currSize - i] = candCostList[currSize - 1 - i];
+ }
+ candModeList[currSize - shift] = uiMode;
+ isNonMMVDListIdx[currSize - shift] = bNonMMVDListCand;
+ candSplitDirList[currSize - shift] = splitDir;
+ candCostList[currSize - shift] = uiCost;
+ return;
+ }
+ else if (currSize < uiFastCandNum)
+ {
+ candModeList.insert(candModeList.end() - shift, uiMode);
+ isNonMMVDListIdx.insert(isNonMMVDListIdx.end() - shift, bNonMMVDListCand);
+ candSplitDirList.insert(candSplitDirList.end() - shift, splitDir);
+ candCostList.insert(candCostList.end() - shift, uiCost);
+ return;
+ }
+ return;
+}
+
+template<size_t N>
+uint32_t updateGeoMMVDCandList(double uiCost, int splitDir, int mergeCand0, int mergeCand1, int mmvdCand0, int mmvdCand1,
+ static_vector<double, N>& candCostList, static_vector<int, N>& geoSplitDirList, static_vector<int, N>& geoMergeCand0, static_vector<int, N>& geoMergeCand1, static_vector<int, N>& geoMmvdCand0, static_vector<int, N>& geoMmvdCand1,
+ size_t uiFastCandNum)
+{
+ CHECK(std::min(uiFastCandNum, geoSplitDirList.size()) != std::min(uiFastCandNum, candCostList.size()), "Sizes do not match!");
+ CHECK(uiFastCandNum > candCostList.capacity(), "The vector is to small to hold all the candidates!");
+
+ size_t i;
+ size_t shift = 0;
+ size_t currSize = std::min(uiFastCandNum, candCostList.size());
+
+ while (shift < uiFastCandNum && shift < currSize && uiCost < candCostList[currSize - 1 - shift])
+ {
+ shift++;
+ }
+
+ if (candCostList.size() >= uiFastCandNum && shift != 0)
+ {
+ for (i = 1; i < shift; i++)
+ {
+ geoSplitDirList[currSize - i] = geoSplitDirList[currSize - 1 - i];
+ geoMergeCand0[currSize - i] = geoMergeCand0[currSize - 1 - i];
+ geoMergeCand1[currSize - i] = geoMergeCand1[currSize - 1 - i];
+ geoMmvdCand0[currSize - i] = geoMmvdCand0[currSize - 1 - i];
+ geoMmvdCand1[currSize - i] = geoMmvdCand1[currSize - 1 - i];
+ candCostList[currSize - i] = candCostList[currSize - 1 - i];
+ }
+ geoSplitDirList[currSize - shift] = splitDir;
+ geoMergeCand0[currSize - shift] = mergeCand0;
+ geoMergeCand1[currSize - shift] = mergeCand1;
+ geoMmvdCand0[currSize - shift] = mmvdCand0;
+ geoMmvdCand1[currSize - shift] = mmvdCand1;
+ candCostList[currSize - shift] = uiCost;
+ return 1;
+ }
+ else if (currSize < uiFastCandNum)
+ {
+ geoSplitDirList.insert(geoSplitDirList.end() - shift, splitDir);
+ geoMergeCand0.insert(geoMergeCand0.end() - shift, mergeCand0);
+ geoMergeCand1.insert(geoMergeCand1.end() - shift, mergeCand1);
+ geoMmvdCand0.insert(geoMmvdCand0.end() - shift, mmvdCand0);
+ geoMmvdCand1.insert(geoMmvdCand1.end() - shift, mmvdCand1);
+ candCostList.insert(candCostList.end() - shift, uiCost);
+ return 1;
+ }
+ return 0;
+}
+template<size_t N>
+void sortCandList(double uiCost, int mergeCand, int mmvdCand, static_vector<double, N>& candCostList, static_vector<int, N>& mergeCandList, static_vector<int, N>& mmvdCandList, int fastCandNum)
+{
+ size_t i;
+ size_t shift = 0;
+ size_t currSize = candCostList.size();
+ CHECK(currSize > fastCandNum, "list overflow!");
+
+ while (shift < currSize && uiCost < candCostList[currSize - 1 - shift])
+ {
+ shift++;
+ }
+
+ if (currSize == fastCandNum && shift != 0)
+ {
+ for (i = 1; i < shift; i++)
+ {
+ mergeCandList[currSize - i] = mergeCandList[currSize - 1 - i];
+ mmvdCandList[currSize - i] = mmvdCandList[currSize - 1 - i];
+ candCostList[currSize - i] = candCostList[currSize - 1 - i];
+ }
+ mergeCandList[currSize - shift] = mergeCand;
+ mmvdCandList[currSize - shift] = mmvdCand;
+ candCostList[currSize - shift] = uiCost;
+ }
+ else if (currSize < fastCandNum)
+ {
+ mergeCandList.insert(mergeCandList.end() - shift, mergeCand);
+ mmvdCandList.insert(mmvdCandList.end() - shift, mmvdCand);
+ candCostList.insert(candCostList.end() - shift, uiCost);
+ }
+}
+#endif
#endif
diff --git a/source/Lib/CommonLib/x86/BufferX86.h b/source/Lib/CommonLib/x86/BufferX86.h
index 0d538ec62..7f69bb735 100644
--- a/source/Lib/CommonLib/x86/BufferX86.h
+++ b/source/Lib/CommonLib/x86/BufferX86.h
@@ -47,6 +47,209 @@
#if ENABLE_SIMD_OPT_BUFFER
#ifdef TARGET_SIMD_X86
+#if JVET_W0097_GPM_MMVD_TM
+template< X86_VEXT vext >
+void roundBD_SSE(const Pel* srcp, const int srcStride, Pel* dest, const int destStride, int width, int height, const ClpRng& clpRng)
+{
+ const int32_t clipbd = clpRng.bd;
+#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+ const int32_t shiftDefault = IF_INTERNAL_FRAC_BITS(clipbd);
+#else
+ const int32_t shiftDefault = std::max<int>(2, (IF_INTERNAL_PREC - clipbd));
+#endif
+ const int32_t offsetDefault = (1 << (shiftDefault - 1)) + IF_INTERNAL_OFFS;
+
+#if USE_AVX2
+ if (vext >= AVX2 && (width & 0x0f) == 0)
+ {
+ __m256i voffset = _mm256_set1_epi16((short)offsetDefault);
+ __m256i vibdmin = _mm256_set1_epi16((short)clpRng.min);
+ __m256i vibdmax = _mm256_set1_epi16((short)clpRng.max);
+ __m256i vsrc;
+ for (int row = 0; row < height; row++)
+ {
+ for (int col = 0; col < width; col += 16)
+ {
+ vsrc = _mm256_lddqu_si256((__m256i *)&srcp[col]);
+ vsrc = _mm256_adds_epi16(vsrc, voffset);
+ vsrc = _mm256_srai_epi16(vsrc, shiftDefault);
+ vsrc = _mm256_min_epi16(vibdmax, _mm256_max_epi16(vibdmin, vsrc));
+ _mm256_storeu_si256((__m256i *)&dest[col], vsrc);
+ }
+ srcp += srcStride;
+ dest += destStride;
+ }
+ }
+ else
+ {
+#endif
+ __m128i voffset = _mm_set1_epi16((short)offsetDefault);
+ __m128i vibdmin = _mm_set1_epi16((short)clpRng.min);
+ __m128i vibdmax = _mm_set1_epi16((short)clpRng.max);
+ __m128i vsrc;
+ for (int row = 0; row < height; row++)
+ {
+ int col = 0;
+ for (; col < ((width >> 3) << 3); col += 8)
+ {
+ vsrc = _mm_lddqu_si128((__m128i *)&srcp[col]);
+ vsrc = _mm_adds_epi16(vsrc, voffset);
+ vsrc = _mm_srai_epi16(vsrc, shiftDefault);
+ vsrc = _mm_min_epi16(vibdmax, _mm_max_epi16(vibdmin, vsrc));
+ _mm_storeu_si128((__m128i *)&dest[col], vsrc);
+ }
+ for (; col < ((width >> 2) << 2); col += 4)
+ {
+ vsrc = _mm_loadl_epi64((__m128i *)&srcp[col]);
+ vsrc = _mm_adds_epi16(vsrc, voffset);
+ vsrc = _mm_srai_epi16(vsrc, shiftDefault);
+ vsrc = _mm_min_epi16(vibdmax, _mm_max_epi16(vibdmin, vsrc));
+ _mm_storel_epi64((__m128i *)&dest[col], vsrc);
+ }
+ for (; col < width; col++)
+ {
+ dest[col] = ClipPel(rightShift(srcp[col] + offsetDefault, shiftDefault), clpRng);
+ }
+ srcp += srcStride;
+ dest += destStride;
+ }
+#if USE_AVX2
+ }
+#endif
+}
+
+template< X86_VEXT vext >
+void weightedAvg_SSE(const Pel* src0, const unsigned src0Stride, const Pel* src1, const unsigned src1Stride, Pel* dest, const unsigned destStride, const int8_t w0, const int8_t w1, int width, int height, const ClpRng& clpRng)
+{
+ const int8_t log2WeightBase = g_BcwLog2WeightBase;
+ const int clipbd = clpRng.bd;
+#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+ const int shiftNum = IF_INTERNAL_FRAC_BITS(clipbd) + log2WeightBase;
+#else
+ const int shiftNum = std::max<int>(2, (IF_INTERNAL_PREC - clipbd)) + log2WeightBase;
+#endif
+ const int offset = (1 << (shiftNum - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
+
+#if USE_AVX2
+ if ((vext >= AVX2) && (width & 0x7) == 0)
+ {
+ __m256i mw = _mm256_unpacklo_epi16(_mm256_set1_epi16(w0), _mm256_set1_epi16(w1));
+ __m256i voffset = _mm256_set1_epi32(offset);
+ __m256i vibdmin = _mm256_set1_epi16((short)clpRng.min);
+ __m256i vibdmax = _mm256_set1_epi16((short)clpRng.max);
+ __m256i msrc0, msrc1, msum0, msum1;
+
+ for (int row = 0; row < height; row++)
+ {
+ for (int col = 0; col < width; col += 8)
+ {
+ msrc0 = _mm256_castsi128_si256(_mm_lddqu_si128((__m128i*)(&src0[col])));
+ msrc1 = _mm256_castsi128_si256(_mm_lddqu_si128((__m128i*)(&src1[col])));
+ msum0 = _mm256_unpacklo_epi16(msrc0, msrc1);
+ msum1 = _mm256_unpackhi_epi16(msrc0, msrc1);
+ msum0 = _mm256_madd_epi16(msum0, mw);
+ msum1 = _mm256_madd_epi16(msum1, mw);
+ msum0 = _mm256_add_epi32(msum0, voffset);
+ msum1 = _mm256_add_epi32(msum1, voffset);
+ msum0 = _mm256_srai_epi32(msum0, shiftNum);
+ msum1 = _mm256_srai_epi32(msum1, shiftNum);
+ msum0 = _mm256_packs_epi32(msum0, msum1);
+ msum0 = _mm256_min_epi16(vibdmax, _mm256_max_epi16(vibdmin, msum0));
+ _mm_storeu_si128((__m128i *)&dest[col], _mm256_castsi256_si128(msum0));
+ }
+ src0 += src0Stride;
+ src1 += src1Stride;
+ dest += destStride;
+ }
+ }
+ else
+ {
+#endif
+ __m128i mw = _mm_unpacklo_epi16(_mm_set1_epi16(w0), _mm_set1_epi16(w1));
+ __m128i voffset = _mm_set1_epi32(offset);
+ __m128i vibdmin = _mm_set1_epi16((short)clpRng.min);
+ __m128i vibdmax = _mm_set1_epi16((short)clpRng.max);
+
+ for (int row = 0; row < height; row++)
+ {
+ int col = 0;
+ for (; col < ((width >> 2) << 2); col += 4)
+ {
+ __m128i msrc = _mm_unpacklo_epi16(_mm_loadl_epi64((__m128i *)&src0[col]), _mm_loadl_epi64((__m128i *)&src1[col]));
+ msrc = _mm_madd_epi16(msrc, mw);
+ msrc = _mm_add_epi32(msrc, voffset);
+ msrc = _mm_srai_epi32(msrc, shiftNum);
+ msrc = _mm_packs_epi32(msrc, msrc);
+ msrc = _mm_min_epi16(vibdmax, _mm_max_epi16(vibdmin, msrc));
+ _mm_storel_epi64((__m128i *)&dest[col], msrc);
+ }
+ for (; col < width; col++)
+ {
+ dest[col] = ClipPel(rightShift(src0[col] * w0 + src1[col] * w1 + offset, shiftNum), clpRng);
+ }
+ src0 += src0Stride;
+ src1 += src1Stride;
+ dest += destStride;
+ }
+#if USE_AVX2
+ }
+#endif
+}
+
+template< X86_VEXT vext >
+void copyClip_SSE(const Pel* srcp, const unsigned srcStride, Pel* dest, const unsigned destStride, int width, int height, const ClpRng& clpRng)
+{
+#if USE_AVX2
+ if (vext >= AVX2 && (width & 0x0f) == 0)
+ {
+ __m256i vibdmin = _mm256_set1_epi16((short)clpRng.min);
+ __m256i vibdmax = _mm256_set1_epi16((short)clpRng.max);
+ __m256i vsrc;
+ for (int row = 0; row < height; row++)
+ {
+ for (int col = 0; col < width; col += 16)
+ {
+ vsrc = _mm256_lddqu_si256((__m256i *)&srcp[col]);
+ vsrc = _mm256_min_epi16(vibdmax, _mm256_max_epi16(vibdmin, vsrc));
+ _mm256_storeu_si256((__m256i *)&dest[col], vsrc);
+ }
+ srcp += srcStride;
+ dest += destStride;
+ }
+ }
+ else
+ {
+#endif
+ __m128i vibdmin = _mm_set1_epi16((short)clpRng.min);
+ __m128i vibdmax = _mm_set1_epi16((short)clpRng.max);
+ __m128i vsrc;
+ for (int row = 0; row < height; row++)
+ {
+ int col = 0;
+ for (; col < ((width >> 3) << 3); col += 8)
+ {
+ vsrc = _mm_lddqu_si128((__m128i *)&srcp[col]);
+ vsrc = _mm_min_epi16(vibdmax, _mm_max_epi16(vibdmin, vsrc));
+ _mm_storeu_si128((__m128i *)&dest[col], vsrc);
+ }
+ for (; col < ((width >> 2) << 2); col += 4)
+ {
+ vsrc = _mm_loadl_epi64((__m128i *)&srcp[col]);
+ vsrc = _mm_min_epi16(vibdmax, _mm_max_epi16(vibdmin, vsrc));
+ _mm_storel_epi64((__m128i *)&dest[col], vsrc);
+ }
+ for (; col < width; col++)
+ {
+ dest[col] = ClipPel(srcp[col], clpRng);
+ }
+ srcp += srcStride;
+ dest += destStride;
+ }
+#if USE_AVX2
+ }
+#endif
+}
+#endif
template< X86_VEXT vext, int W >
void addAvg_SSE( const int16_t* src0, int src0Stride, const int16_t* src1, int src1Stride, int16_t *dst, int dstStride, int width, int height, int shift, int offset, const ClpRng& clpRng )
@@ -1767,6 +1970,11 @@ int64_t getSumOfDifference_SSE(const Pel* src0, int src0Stride, const Pel* src1,
template<X86_VEXT vext>
void PelBufferOps::_initPelBufOpsX86()
{
+#if JVET_W0097_GPM_MMVD_TM
+ roundBD = roundBD_SSE<vext>;
+ weightedAvg = weightedAvg_SSE<vext>;
+ copyClip = copyClip_SSE<vext>;
+#endif
addAvg8 = addAvg_SSE<vext, 8>;
addAvg4 = addAvg_SSE<vext, 4>;
diff --git a/source/Lib/DecoderLib/CABACReader.cpp b/source/Lib/DecoderLib/CABACReader.cpp
index a0ced2ca8..ad78f9274 100644
--- a/source/Lib/DecoderLib/CABACReader.cpp
+++ b/source/Lib/DecoderLib/CABACReader.cpp
@@ -2958,6 +2958,58 @@ void CABACReader::merge_idx( PredictionUnit& pu )
if (pu.cu->geoFlag)
{
RExt__DECODER_DEBUG_BIT_STATISTICS_CREATE_SET(STATS__CABAC_BITS__GEO_INDEX);
+#if JVET_W0097_GPM_MMVD_TM
+ pu.geoMMVDFlag0 = m_BinDecoder.decodeBin(Ctx::GeoMmvdFlag());
+ if (pu.geoMMVDFlag0)
+ {
+ geo_mmvd_idx(pu, REF_PIC_LIST_0);
+ }
+
+ pu.geoMMVDFlag1 = m_BinDecoder.decodeBin(Ctx::GeoMmvdFlag());
+ if (pu.geoMMVDFlag1)
+ {
+ geo_mmvd_idx(pu, REF_PIC_LIST_1);
+ }
+
+#if TM_MRG
+ if (!pu.geoMMVDFlag0 && !pu.geoMMVDFlag1)
+ {
+ tm_merge_flag(pu);
+ if (pu.tmMergeFlag)
+ {
+ pu.geoTmFlag0 = true;
+ pu.geoTmFlag1 = true;
+ geo_merge_idx(pu);
+ }
+ else
+ {
+ pu.geoTmFlag0 = false;
+ pu.geoTmFlag1 = false;
+ geo_merge_idx(pu);
+ }
+ }
+#else
+ if (!pu.geoMMVDFlag0 && !pu.geoMMVDFlag1)
+ {
+ geo_merge_idx(pu);
+ }
+#endif
+ else if (pu.geoMMVDFlag0 && pu.geoMMVDFlag1)
+ {
+ if (pu.geoMMVDIdx0 == pu.geoMMVDIdx1)
+ {
+ geo_merge_idx(pu);
+ }
+ else
+ {
+ geo_merge_idx1(pu);
+ }
+ }
+ else
+ {
+ geo_merge_idx1(pu);
+ }
+#else
uint32_t splitDir = 0;
xReadTruncBinCode(splitDir, GEO_NUM_PARTITION_MODE);
pu.geoSplitDir = splitDir;
@@ -2985,6 +3037,7 @@ void CABACReader::merge_idx( PredictionUnit& pu )
DTRACE(g_trace_ctx, D_SYNTAX, "merge_idx() geo_split_dir=%d\n", splitDir);
DTRACE(g_trace_ctx, D_SYNTAX, "merge_idx() geo_idx0=%d\n", mergeCand0);
DTRACE(g_trace_ctx, D_SYNTAX, "merge_idx() geo_idx1=%d\n", mergeCand1);
+#endif
return;
}
@@ -3038,7 +3091,94 @@ void CABACReader::merge_idx( PredictionUnit& pu )
DTRACE( g_trace_ctx, D_SYNTAX, "merge_idx() merge_idx=%d\n", pu.mergeIdx );
}
}
+#if JVET_W0097_GPM_MMVD_TM
+void CABACReader::geo_mmvd_idx(PredictionUnit& pu, RefPicList eRefPicList)
+{
+ bool extMMVD = pu.cs->picHeader->getGPMMMVDTableFlag();
+ int numCandminus1_step = (extMMVD ? GPM_EXT_MMVD_REFINE_STEP : GPM_MMVD_REFINE_STEP) - 1;
+ int step = 0;
+ if (m_BinDecoder.decodeBin(Ctx::GeoMmvdStepMvpIdx()))
+ {
+ step++;
+ for (; step < numCandminus1_step; step++)
+ {
+ if (!m_BinDecoder.decodeBinEP())
+ {
+ break;
+ }
+ }
+ }
+
+ int idxToMMVDStep[GPM_EXT_MMVD_REFINE_STEP] = { 1, 2, 3, 4, 5, 0, 6, 7, 8 };
+ step = idxToMMVDStep[step];
+ int direction = 0;
+ int maxMMVDDir = (extMMVD ? GPM_EXT_MMVD_REFINE_DIRECTION : GPM_MMVD_REFINE_DIRECTION);
+ direction = m_BinDecoder.decodeBinsEP(maxMMVDDir > 4 ? 3 : 2);
+ int mvpIdx = (step * maxMMVDDir + direction);
+ if (eRefPicList == REF_PIC_LIST_0)
+ {
+ pu.geoMMVDIdx0 = mvpIdx;
+ }
+ else
+ {
+ pu.geoMMVDIdx1 = mvpIdx;
+ }
+}
+
+
+void CABACReader::geo_merge_idx(PredictionUnit& pu)
+{
+ uint32_t splitDir = 0;
+ xReadTruncBinCode(splitDir, GEO_NUM_PARTITION_MODE);
+ pu.geoSplitDir = splitDir;
+ const int maxNumGeoCand = pu.cs->sps->getMaxNumGeoCand();
+ CHECK(maxNumGeoCand < 2, "Incorrect max number of geo candidates");
+ CHECK(pu.cu->lheight() > 64 || pu.cu->lwidth() > 64, "Incorrect block size of geo flag");
+ int numCandminus2 = maxNumGeoCand - 2;
+ pu.mergeIdx = 0;
+ int mergeCand0 = 0;
+ int mergeCand1 = 0;
+ if (m_BinDecoder.decodeBin(Ctx::MergeIdx()))
+ {
+ mergeCand0 += unary_max_eqprob(numCandminus2) + 1;
+ }
+ if (numCandminus2 > 0)
+ {
+ if (m_BinDecoder.decodeBin(Ctx::MergeIdx()))
+ {
+ mergeCand1 += unary_max_eqprob(numCandminus2 - 1) + 1;
+ }
+ }
+ mergeCand1 += mergeCand1 >= mergeCand0 ? 1 : 0;
+ pu.geoMergeIdx0 = mergeCand0;
+ pu.geoMergeIdx1 = mergeCand1;
+}
+
+void CABACReader::geo_merge_idx1(PredictionUnit& pu)
+{
+ uint32_t splitDir = 0;
+ xReadTruncBinCode(splitDir, GEO_NUM_PARTITION_MODE);
+ pu.geoSplitDir = splitDir;
+ const int maxNumGeoCand = pu.cs->sps->getMaxNumGeoCand();
+ CHECK(maxNumGeoCand < 2, "Incorrect max number of geo candidates");
+ CHECK(pu.cu->lheight() > 64 || pu.cu->lwidth() > 64, "Incorrect block size of geo flag");
+ int numCandminus2 = maxNumGeoCand - 2;
+ pu.mergeIdx = 0;
+ int mergeCand0 = 0;
+ int mergeCand1 = 0;
+ if (m_BinDecoder.decodeBin(Ctx::MergeIdx()))
+ {
+ mergeCand0 += unary_max_eqprob(numCandminus2) + 1;
+ }
+ if (m_BinDecoder.decodeBin(Ctx::MergeIdx()))
+ {
+ mergeCand1 += unary_max_eqprob(numCandminus2) + 1;
+ }
+ pu.geoMergeIdx0 = mergeCand0;
+ pu.geoMergeIdx1 = mergeCand1;
+}
+#endif
void CABACReader::mmvd_merge_idx(PredictionUnit& pu)
{
RExt__DECODER_DEBUG_BIT_STATISTICS_CREATE_SET(STATS__CABAC_BITS__MERGE_INDEX);
diff --git a/source/Lib/DecoderLib/CABACReader.h b/source/Lib/DecoderLib/CABACReader.h
index 4a8b362b9..5d0e3329a 100644
--- a/source/Lib/DecoderLib/CABACReader.h
+++ b/source/Lib/DecoderLib/CABACReader.h
@@ -131,6 +131,11 @@ public:
#endif
#if TM_MRG
void tm_merge_flag ( PredictionUnit& pu );
+#endif
+#if JVET_W0097_GPM_MMVD_TM
+ void geo_mmvd_idx(PredictionUnit& pu, RefPicList eRefPicList);
+ void geo_merge_idx(PredictionUnit& pu);
+ void geo_merge_idx1(PredictionUnit& pu);
#endif
void imv_mode ( CodingUnit& cu, MergeCtx& mrgCtx );
void affine_amvr_mode ( CodingUnit& cu, MergeCtx& mrgCtx );
diff --git a/source/Lib/DecoderLib/DecCu.cpp b/source/Lib/DecoderLib/DecCu.cpp
index f3c5ec18b..1e24c5d7e 100644
--- a/source/Lib/DecoderLib/DecCu.cpp
+++ b/source/Lib/DecoderLib/DecCu.cpp
@@ -908,8 +908,18 @@ void DecCu::xReconInter(CodingUnit &cu)
{
if( cu.geoFlag )
{
+#if JVET_W0097_GPM_MMVD_TM
+#if TM_MRG
+ m_pcInterPred->motionCompensationGeo(cu, m_geoMrgCtx, m_geoTmMrgCtx0, m_geoTmMrgCtx1);
+ PU::spanGeoMMVDMotionInfo(*cu.firstPU, m_geoMrgCtx, m_geoTmMrgCtx0, m_geoTmMrgCtx1, cu.firstPU->geoSplitDir, cu.firstPU->geoMergeIdx0, cu.firstPU->geoMergeIdx1, cu.firstPU->geoTmFlag0, cu.firstPU->geoMMVDFlag0, cu.firstPU->geoMMVDIdx0, cu.firstPU->geoTmFlag1, cu.firstPU->geoMMVDFlag1, cu.firstPU->geoMMVDIdx1);
+#else
+ m_pcInterPred->motionCompensationGeo(cu, m_geoMrgCtx);
+ PU::spanGeoMMVDMotionInfo(*cu.firstPU, m_geoMrgCtx, cu.firstPU->geoSplitDir, cu.firstPU->geoMergeIdx0, cu.firstPU->geoMergeIdx1, cu.firstPU->geoMMVDFlag0, cu.firstPU->geoMMVDIdx0, cu.firstPU->geoMMVDFlag1, cu.firstPU->geoMMVDIdx1);
+#endif
+#else
m_pcInterPred->motionCompensationGeo( cu, m_geoMrgCtx );
PU::spanGeoMotionInfo( *cu.firstPU, m_geoMrgCtx, cu.firstPU->geoSplitDir, cu.firstPU->geoMergeIdx0, cu.firstPU->geoMergeIdx1 );
+#endif
}
else
{
@@ -1285,6 +1295,50 @@ void DecCu::xDeriveCUMV( CodingUnit &cu )
if( pu.cu->geoFlag )
{
PU::getGeoMergeCandidates( pu, m_geoMrgCtx );
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+ if (pu.geoTmFlag0)
+ {
+ m_geoTmMrgCtx0.numValidMergeCand = m_geoMrgCtx.numValidMergeCand;
+ m_geoTmMrgCtx0.BcwIdx[pu.geoMergeIdx0] = BCW_DEFAULT;
+ m_geoTmMrgCtx0.useAltHpelIf[pu.geoMergeIdx0] = false;
+#if INTER_LIC
+ m_geoTmMrgCtx0.LICFlags[pu.geoMergeIdx0] = false;
+#endif
+ m_geoTmMrgCtx0.addHypNeighbours[pu.geoMergeIdx0] = m_geoMrgCtx.addHypNeighbours[pu.geoMergeIdx0];
+ m_geoTmMrgCtx0.interDirNeighbours[pu.geoMergeIdx0] = m_geoMrgCtx.interDirNeighbours[pu.geoMergeIdx0];
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1)].mv = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx0 << 1)].mv;
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1) + 1].mv = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx0 << 1) + 1].mv;
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1)].refIdx = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx0 << 1)].refIdx;
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1) + 1].refIdx = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx0 << 1) + 1].refIdx;
+
+ m_geoTmMrgCtx0.setMergeInfo(pu, pu.geoMergeIdx0);
+ pu.geoTmType = g_geoTmShape[0][g_GeoParams[pu.geoSplitDir][0]];
+ m_pcInterPred->deriveTMMv(pu);
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1)].mv.set(pu.mv[0].getHor(), pu.mv[0].getVer());
+ m_geoTmMrgCtx0.mvFieldNeighbours[(pu.geoMergeIdx0 << 1) + 1].mv.set(pu.mv[1].getHor(), pu.mv[1].getVer());
+ }
+ if (pu.geoTmFlag1)
+ {
+ m_geoTmMrgCtx1.numValidMergeCand = m_geoMrgCtx.numValidMergeCand;
+ m_geoTmMrgCtx1.BcwIdx[pu.geoMergeIdx1] = BCW_DEFAULT;
+ m_geoTmMrgCtx1.useAltHpelIf[pu.geoMergeIdx1] = false;
+#if INTER_LIC
+ m_geoTmMrgCtx1.LICFlags[pu.geoMergeIdx1] = false;
+#endif
+ m_geoTmMrgCtx1.addHypNeighbours[pu.geoMergeIdx1] = m_geoMrgCtx.addHypNeighbours[pu.geoMergeIdx1];
+ m_geoTmMrgCtx1.interDirNeighbours[pu.geoMergeIdx1] = m_geoMrgCtx.interDirNeighbours[pu.geoMergeIdx1];
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1)].mv = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx1 << 1)].mv;
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1) + 1].mv = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx1 << 1) + 1].mv;
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1)].refIdx = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx1 << 1)].refIdx;
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1) + 1].refIdx = m_geoMrgCtx.mvFieldNeighbours[(pu.geoMergeIdx1 << 1) + 1].refIdx;
+
+ m_geoTmMrgCtx1.setMergeInfo(pu, pu.geoMergeIdx1);
+ pu.geoTmType = g_geoTmShape[1][g_GeoParams[pu.geoSplitDir][0]];
+ m_pcInterPred->deriveTMMv(pu);
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1)].mv.set(pu.mv[0].getHor(), pu.mv[0].getVer());
+ m_geoTmMrgCtx1.mvFieldNeighbours[(pu.geoMergeIdx1 << 1) + 1].mv.set(pu.mv[1].getHor(), pu.mv[1].getVer());
+ }
+#endif
}
else
{
diff --git a/source/Lib/DecoderLib/DecCu.h b/source/Lib/DecoderLib/DecCu.h
index 91944c63a..33c761acf 100644
--- a/source/Lib/DecoderLib/DecCu.h
+++ b/source/Lib/DecoderLib/DecCu.h
@@ -105,6 +105,9 @@ private:
#endif
MergeCtx m_geoMrgCtx;
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+ MergeCtx m_geoTmMrgCtx0, m_geoTmMrgCtx1;
+#endif
};
//! \}
diff --git a/source/Lib/DecoderLib/VLCReader.cpp b/source/Lib/DecoderLib/VLCReader.cpp
index 96d62b5f8..c2836ec3b 100644
--- a/source/Lib/DecoderLib/VLCReader.cpp
+++ b/source/Lib/DecoderLib/VLCReader.cpp
@@ -3576,6 +3576,18 @@ void HLSyntaxReader::parsePictureHeader( PicHeader* picHeader, ParameterSetManag
picHeader->setDisFracMMVD(false);
}
+#if JVET_W0097_GPM_MMVD_TM
+ if (sps->getUseGeo() && (!pps->getRplInfoInPhFlag() || picHeader->getRPL(1)->getNumRefEntries() > 0))
+ {
+ READ_FLAG(uiCode, "ph_gpm_ext_mmvd_flag");
+ picHeader->setGPMMMVDTableFlag(uiCode);
+ }
+ else
+ {
+ picHeader->setGPMMMVDTableFlag(false);
+ }
+#endif
+
#if JVET_R0324_REORDER
// mvd L1 zero flag
if (!pps->getRplInfoInPhFlag() || picHeader->getRPL(1)->getNumRefEntries() > 0)
diff --git a/source/Lib/EncoderLib/CABACWriter.cpp b/source/Lib/EncoderLib/CABACWriter.cpp
index 4904308cf..ca65ed133 100644
--- a/source/Lib/EncoderLib/CABACWriter.cpp
+++ b/source/Lib/EncoderLib/CABACWriter.cpp
@@ -2655,6 +2655,57 @@ void CABACWriter::merge_idx( const PredictionUnit& pu )
{
if( pu.cu->geoFlag )
{
+#if JVET_W0097_GPM_MMVD_TM
+ m_BinEncoder.encodeBin(pu.geoMMVDFlag0, Ctx::GeoMmvdFlag());
+ if (pu.geoMMVDFlag0)
+ {
+ geo_mmvd_idx(pu, REF_PIC_LIST_0);
+ }
+ m_BinEncoder.encodeBin(pu.geoMMVDFlag1, Ctx::GeoMmvdFlag());
+ if (pu.geoMMVDFlag1)
+ {
+ geo_mmvd_idx(pu, REF_PIC_LIST_1);
+ }
+
+#if TM_MRG
+ if (!pu.geoMMVDFlag0 && !pu.geoMMVDFlag1)
+ {
+ tm_merge_flag(pu);
+ if (pu.tmMergeFlag)
+ {
+ CHECK(!pu.geoTmFlag0 || !pu.geoTmFlag1, "both must be true");
+ CHECK(pu.geoMergeIdx0 == pu.geoMergeIdx1, "Incorrect geoMergeIdx0 and geoMergeIdx1");
+ geo_merge_idx(pu);
+ }
+ else
+ {
+
+ CHECK(pu.geoTmFlag0 || pu.geoTmFlag1, "both must be false");
+ geo_merge_idx(pu);
+ }
+ }
+#else
+ if (!pu.geoMMVDFlag0 && !pu.geoMMVDFlag1)
+ {
+ geo_merge_idx(pu);
+ }
+#endif
+ else if (pu.geoMMVDFlag0 && pu.geoMMVDFlag1)
+ {
+ if (pu.geoMMVDIdx0 == pu.geoMMVDIdx1)
+ {
+ geo_merge_idx(pu);
+ }
+ else
+ {
+ geo_merge_idx1(pu);
+ }
+ }
+ else
+ {
+ geo_merge_idx1(pu);
+ }
+#else
uint8_t splitDir = pu.geoSplitDir;
uint8_t candIdx0 = pu.geoMergeIdx0;
uint8_t candIdx1 = pu.geoMergeIdx1;
@@ -2681,7 +2732,7 @@ void CABACWriter::merge_idx( const PredictionUnit& pu )
unary_max_eqprob(candIdx1 - 1, numCandminus2 - 1);
}
}
-
+#endif
return;
}
int numCandminus1;
@@ -2788,6 +2839,176 @@ void CABACWriter::mmvd_merge_idx(const PredictionUnit& pu)
DTRACE(g_trace_ctx, D_SYNTAX, "pos() pos=%d\n", var2);
DTRACE(g_trace_ctx, D_SYNTAX, "mmvd_merge_idx() mmvd_merge_idx=%d\n", pu.mmvdMergeIdx);
}
+#if JVET_W0097_GPM_MMVD_TM
+void CABACWriter::geo_mmvd_idx(const PredictionUnit& pu, RefPicList eRefPicList)
+{
+ int geoMMVDIdx = (eRefPicList == REF_PIC_LIST_0) ? pu.geoMMVDIdx0 : pu.geoMMVDIdx1;
+ bool extMMVD = pu.cs->picHeader->getGPMMMVDTableFlag();
+ CHECK(geoMMVDIdx >= (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM), "invalid GPM MMVD index exist");
+
+ int step = (extMMVD ? (geoMMVDIdx >> 3) : (geoMMVDIdx >> 2));
+ int direction = (extMMVD ? (geoMMVDIdx - (step << 3)) : (geoMMVDIdx - (step << 2)));
+
+ int mmvdStepToIdx[GPM_EXT_MMVD_REFINE_STEP] = { 5, 0, 1, 2, 3, 4, 6, 7, 8 };
+ step = mmvdStepToIdx[step];
+
+ int numCandminus1_step = (extMMVD ? GPM_EXT_MMVD_REFINE_STEP : GPM_MMVD_REFINE_STEP) - 1;
+ if (numCandminus1_step > 0)
+ {
+ if (step == 0)
+ {
+ m_BinEncoder.encodeBin(0, Ctx::GeoMmvdStepMvpIdx());
+ }
+ else
+ {
+ m_BinEncoder.encodeBin(1, Ctx::GeoMmvdStepMvpIdx());
+ for (unsigned idx = 1; idx < numCandminus1_step; idx++)
+ {
+ m_BinEncoder.encodeBinEP(step == idx ? 0 : 1);
+ if (step == idx)
+ {
+ break;
+ }
+ }
+ }
+ }
+
+ int maxMMVDDir = (extMMVD ? GPM_EXT_MMVD_REFINE_DIRECTION : GPM_MMVD_REFINE_DIRECTION);
+ m_BinEncoder.encodeBinsEP(direction, maxMMVDDir > 4 ? 3 : 2);
+}
+
+void CABACWriter::geo_merge_idx(const PredictionUnit& pu)
+{
+ uint8_t splitDir = pu.geoSplitDir;
+ uint8_t candIdx0 = pu.geoMergeIdx0;
+ uint8_t candIdx1 = pu.geoMergeIdx1;
+
+ xWriteTruncBinCode(splitDir, GEO_NUM_PARTITION_MODE);
+ candIdx1 -= candIdx1 < candIdx0 ? 0 : 1;
+ const int maxNumGeoCand = pu.cs->sps->getMaxNumGeoCand();
+
+ int numCandminus2 = maxNumGeoCand - 2;
+ m_BinEncoder.encodeBin(candIdx0 == 0 ? 0 : 1, Ctx::MergeIdx());
+ if (candIdx0 > 0)
+ {
+ unary_max_eqprob(candIdx0 - 1, numCandminus2);
+ }
+ if (numCandminus2 > 0)
+ {
+ m_BinEncoder.encodeBin(candIdx1 == 0 ? 0 : 1, Ctx::MergeIdx());
+ if (candIdx1 > 0)
+ {
+ unary_max_eqprob(candIdx1 - 1, numCandminus2 - 1);
+ }
+ }
+}
+
+void CABACWriter::geo_merge_idx1(const PredictionUnit& pu)
+{
+ uint8_t splitDir = pu.geoSplitDir;
+ uint8_t candIdx0 = pu.geoMergeIdx0;
+ uint8_t candIdx1 = pu.geoMergeIdx1;
+
+ xWriteTruncBinCode(splitDir, GEO_NUM_PARTITION_MODE);
+ const int maxNumGeoCand = pu.cs->sps->getMaxNumGeoCand();
+
+ int numCandminus2 = maxNumGeoCand - 2;
+ m_BinEncoder.encodeBin(candIdx0 == 0 ? 0 : 1, Ctx::MergeIdx());
+ if (candIdx0 > 0)
+ {
+ unary_max_eqprob(candIdx0 - 1, numCandminus2);
+ }
+ m_BinEncoder.encodeBin(candIdx1 == 0 ? 0 : 1, Ctx::MergeIdx());
+ if (candIdx1 > 0)
+ {
+ unary_max_eqprob(candIdx1 - 1, numCandminus2);
+ }
+}
+
+uint64_t CABACWriter::geo_mode_est(const TempCtx& ctxStart, const int geoMode)
+{
+ getCtx() = ctxStart;
+ resetBits();
+
+ xWriteTruncBinCode(geoMode, GEO_NUM_PARTITION_MODE);
+
+ return getEstFracBits();
+}
+
+uint64_t CABACWriter::geo_mergeIdx_est(const TempCtx& ctxStart, const int candIdx, const int maxNumGeoCand)
+{
+ getCtx() = ctxStart;
+ resetBits();
+
+ int numCandminus2 = maxNumGeoCand - 2;
+ m_BinEncoder.encodeBin(candIdx == 0 ? 0 : 1, Ctx::MergeIdx());
+ if (candIdx > 0)
+ {
+ unary_max_eqprob(candIdx - 1, numCandminus2);
+ }
+
+ return getEstFracBits();
+}
+
+uint64_t CABACWriter::geo_mmvdFlag_est(const TempCtx& ctxStart, const int flag)
+{
+ getCtx() = ctxStart;
+ resetBits();
+
+ m_BinEncoder.encodeBin(flag, Ctx::GeoMmvdFlag());
+
+ return getEstFracBits();
+}
+
+#if TM_MRG
+uint64_t CABACWriter::geo_tmFlag_est(const TempCtx& ctxStart, const int flag)
+{
+ getCtx() = ctxStart;
+ resetBits();
+
+ m_BinEncoder.encodeBin(flag, Ctx::TMMergeFlag());
+
+ return getEstFracBits();
+}
+#endif
+
+uint64_t CABACWriter::geo_mmvdIdx_est(const TempCtx& ctxStart, const int geoMMVDIdx, const bool extMMVD)
+{
+ getCtx() = ctxStart;
+ resetBits();
+
+ CHECK(geoMMVDIdx >= (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM), "invalid GPM MMVD index exist");
+ int step = (extMMVD ? (geoMMVDIdx >> 3) : (geoMMVDIdx >> 2));
+ int direction = (extMMVD ? (geoMMVDIdx - (step << 3)) : (geoMMVDIdx - (step << 2)));
+
+ int mmvdStepToIdx[GPM_EXT_MMVD_REFINE_STEP] = { 5, 0, 1, 2, 3, 4, 6, 7, 8 };
+ step = mmvdStepToIdx[step];
+
+ int numCandminus1_step = (extMMVD ? GPM_EXT_MMVD_REFINE_STEP : GPM_MMVD_REFINE_STEP) - 1;
+ if (numCandminus1_step > 0)
+ {
+ if (step == 0)
+ {
+ m_BinEncoder.encodeBin(0, Ctx::GeoMmvdStepMvpIdx());
+ }
+ else
+ {
+ m_BinEncoder.encodeBin(1, Ctx::GeoMmvdStepMvpIdx());
+ for (unsigned idx = 1; idx < numCandminus1_step; idx++)
+ {
+ m_BinEncoder.encodeBinEP(step == idx ? 0 : 1);
+ if (step == idx)
+ {
+ break;
+ }
+ }
+ }
+ }
+ int maxMMVDDir = (extMMVD ? GPM_EXT_MMVD_REFINE_DIRECTION : GPM_MMVD_REFINE_DIRECTION);
+ m_BinEncoder.encodeBinsEP(direction, maxMMVDDir > 4 ? 3 : 2);
+ return getEstFracBits();
+}
+#endif
void CABACWriter::inter_pred_idc( const PredictionUnit& pu )
{
if( !pu.cs->slice->isInterB() )
diff --git a/source/Lib/EncoderLib/CABACWriter.h b/source/Lib/EncoderLib/CABACWriter.h
index bae249e50..0b55a2d36 100644
--- a/source/Lib/EncoderLib/CABACWriter.h
+++ b/source/Lib/EncoderLib/CABACWriter.h
@@ -141,6 +141,19 @@ public:
#endif
#if TM_MRG
void tm_merge_flag ( const PredictionUnit& pu);
+#endif
+#if JVET_W0097_GPM_MMVD_TM
+ void geo_mmvd_idx(const PredictionUnit& pu, RefPicList eRefPicList);
+ void geo_merge_idx(const PredictionUnit& pu);
+ void geo_merge_idx1(const PredictionUnit& pu);
+
+ uint64_t geo_mode_est(const TempCtx& ctxStart, const int geoMode);
+ uint64_t geo_mergeIdx_est(const TempCtx& ctxStart, const int candIdx, const int maxNumGeoCand);
+ uint64_t geo_mmvdFlag_est(const TempCtx& ctxStart, const int flag);
+ uint64_t geo_mmvdIdx_est(const TempCtx& ctxStart, const int mmvdIdx, const bool extMMVD);
+#if TM_MRG
+ uint64_t geo_tmFlag_est(const TempCtx& ctxStart, const int flag);
+#endif
#endif
void imv_mode ( const CodingUnit& cu );
void affine_amvr_mode ( const CodingUnit& cu );
diff --git a/source/Lib/EncoderLib/EncCu.cpp b/source/Lib/EncoderLib/EncCu.cpp
index 214226047..c988f0054 100644
--- a/source/Lib/EncoderLib/EncCu.cpp
+++ b/source/Lib/EncoderLib/EncCu.cpp
@@ -85,6 +85,13 @@ EncCu::EncCu() : m_GeoModeTest
}
}
#endif
+#if JVET_W0097_GPM_MMVD_TM
+ fastGpmMmvdSearch = false;
+ fastGpmMmvdRelatedCU = false;
+ includeMoreMMVDCandFirstPass = false;
+ maxNumGPMDirFirstPass = 64;
+ numCandPerPar = 5;
+#endif
}
void EncCu::create( EncCfg* encCfg )
@@ -197,6 +204,35 @@ void EncCu::create( EncCfg* encCfg )
{
m_acGeoWeightedBuffer[ui].create( chromaFormat, Area( 0, 0, uiMaxWidth, uiMaxHeight ) );
}
+#if JVET_W0097_GPM_MMVD_TM
+ for (unsigned ui = 0; ui < MRG_MAX_NUM_CANDS; ui++)
+ {
+ for (unsigned vi = 0; vi < GPM_EXT_MMVD_MAX_REFINE_NUM; vi++)
+ {
+ m_acGeoMMVDBuffer[ui][vi].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
+ m_acGeoMMVDTmpBuffer[ui][vi].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
+ }
+ }
+ int sourceWidth = encCfg->getSourceWidth();
+ int sourceHeight = encCfg->getSourceHeight();
+ fastGpmMmvdSearch = (((encCfg->getIntraPeriod() > 0) && ((sourceWidth * sourceHeight) <= (1920 * 1080))) || ((encCfg->getIntraPeriod() < 0) && ((sourceWidth * sourceHeight) >= (1280 * 720)))) && !encCfg->getIBCMode();
+#if TM_MRG
+ fastGpmMmvdRelatedCU = ((encCfg->getIntraPeriod() < 0) && ((sourceWidth * sourceHeight) >= (1920 * 1080))) && !encCfg->getIBCMode();
+#else
+ fastGpmMmvdRelatedCU = ((encCfg->getIntraPeriod() < 0) && ((sourceWidth * sourceHeight) >= (1280 * 720))) && !encCfg->getIBCMode();
+#endif
+
+ includeMoreMMVDCandFirstPass = ((encCfg->getIntraPeriod() > 0) || ((encCfg->getIntraPeriod() < 0) && fastGpmMmvdSearch));
+ maxNumGPMDirFirstPass = ((encCfg->getIntraPeriod() < 0) ? 50 : (fastGpmMmvdSearch ? 36 : 64));
+ numCandPerPar = (fastGpmMmvdSearch ? 4 : 5);
+#if TM_MRG
+ for (uint16_t ui = 0; ui < GEO_TM_MAX_NUM_CANDS; ui++)
+ {
+ m_acGeoMergeTmpBuffer[ui].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
+ m_acGeoSADTmpBuffer[ui].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
+ }
+#endif
+#endif
m_ciipBuffer[0].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
m_ciipBuffer[1].create(chromaFormat, Area(0, 0, uiMaxWidth, uiMaxHeight));
@@ -307,6 +343,23 @@ void EncCu::destroy()
{
m_acGeoWeightedBuffer[ui].destroy();
}
+#if JVET_W0097_GPM_MMVD_TM
+ for (unsigned ui = 0; ui < MRG_MAX_NUM_CANDS; ui++)
+ {
+ for (unsigned vi = 0; vi < GPM_EXT_MMVD_MAX_REFINE_NUM; vi++)
+ {
+ m_acGeoMMVDBuffer[ui][vi].destroy();
+ m_acGeoMMVDTmpBuffer[ui][vi].destroy();
+ }
+ }
+#if TM_MRG
+ for (uint16_t ui = 0; ui < GEO_TM_MAX_NUM_CANDS; ui++)
+ {
+ m_acGeoMergeTmpBuffer[ui].destroy();
+ m_acGeoSADTmpBuffer[ui].destroy();
+ }
+#endif
+#endif
m_ciipBuffer[0].destroy();
m_ciipBuffer[1].destroy();
}
@@ -796,7 +849,9 @@ void EncCu::xCompressCU( CodingStructure*& tempCS, CodingStructure*& bestCS, Par
m_acGeoWeightedBuffer, GEO_MAX_TRY_WEIGHTED_SAD,
localUnitArea);
#endif
-
+#if JVET_W0097_GPM_MMVD_TM
+ m_mergeCandAvail = false;
+#endif
double bestIntPelCost = MAX_DOUBLE;
if (tempCS->slice->getSPS()->getUseColorTrans())
@@ -945,7 +1000,20 @@ void EncCu::xCompressCU( CodingStructure*& tempCS, CodingStructure*& bestCS, Par
}
else if( currTestMode.type == ETM_MERGE_GEO )
{
+#if JVET_W0097_GPM_MMVD_TM
+ CodedCUInfo &relatedCU = ((EncModeCtrlMTnoRQT *)m_modeCtrl)->getBlkInfo(partitioner.currArea());
+ if (!relatedCU.isGPMTested)
+ {
+ xCheckRDCostMergeGeoComb2Nx2N(tempCS, bestCS, partitioner, currTestMode);
+ relatedCU.isGPMTested = 1;
+ }
+ else
+ {
+ xCheckRDCostMergeGeoComb2Nx2N(tempCS, bestCS, partitioner, currTestMode, true);
+ }
+#else
xCheckRDCostMergeGeo2Nx2N( tempCS, bestCS, partitioner, currTestMode );
+#endif
}
#if MULTI_HYP_PRED
else if (currTestMode.type == ETM_INTER_MULTIHYP)
@@ -2746,6 +2814,10 @@ void EncCu::xCheckRDCostMerge2Nx2N( CodingStructure *&tempCS, CodingStructure *&
PU::getInterMergeCandidates(pu, mergeCtx
, 0
);
+#if JVET_W0097_GPM_MMVD_TM
+ m_mergeCand.copyMergeCtx(mergeCtx);
+ m_mergeCandAvail = true;
+#endif
PU::getInterMMVDMergeCandidates(pu, mergeCtx);
#if TM_MRG && MERGE_ENC_OPT
if (cu.cs->sps->getUseDMVDMode())
@@ -3893,22 +3965,73 @@ void EncCu::xCheckRDCostMerge2Nx2N( CodingStructure *&tempCS, CodingStructure *&
}
}
-void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode)
+#if JVET_W0097_GPM_MMVD_TM
+void EncCu::xCheckRDCostMergeGeoComb2Nx2N(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode, bool isSecondPass)
{
- const Slice &slice = *tempCS->slice;
- CHECK(slice.getSliceType() == I_SLICE, "Merge modes not available for I-slices");
+ int numSATDCands = ((fastGpmMmvdSearch && isSecondPass) ? 60 : 70);
tempCS->initStructData(encTestMode.qp);
-
+#if TM_MRG
+ MergeCtx mergeCtx[GEO_NUM_TM_MV_CAND];
+#else
MergeCtx mergeCtx;
+#endif
const SPS &sps = *tempCS->sps;
+ CodedCUInfo& relatedCU = ((EncModeCtrlMTnoRQT *)m_modeCtrl)->getBlkInfo(pm.currArea());
+ bool extMMVD = tempCS->picHeader->getGPMMMVDTableFlag();
if (sps.getSbTMVPEnabledFlag())
{
Size bufSize = g_miScaling.scale(tempCS->area.lumaSize());
+#if TM_MRG
+ for (int i = 0; i < GEO_NUM_TM_MV_CAND; i++)
+ {
+ mergeCtx[i].subPuMvpMiBuf = MotionBuf(m_SubPuMiBuf, bufSize);
+ }
+#else
mergeCtx.subPuMvpMiBuf = MotionBuf(m_SubPuMiBuf, bufSize);
+#endif
+ }
+
+ // 1. bit estimation
+ const double sqrtLambdaFracBits = m_pcRdCost->getMotionLambda() * FRAC_BITS_SCALE;
+ uint8_t maxNumMergeCandidates = tempCS->sps->getMaxNumGeoCand();
+ const TempCtx ctxStart(m_CtxCache, m_CABACEstimator->getCtx());
+
+ double geoModeCost[GEO_NUM_PARTITION_MODE], geoMergeIdxCost[MRG_MAX_NUM_CANDS], geoMMVDFlagCost[2], geoMMVDIdxCost[GPM_EXT_MMVD_MAX_REFINE_NUM];
+#if TM_MRG
+ double geoTMFlagCost[2];
+#endif
+ for (int idx = 0; idx < GEO_NUM_PARTITION_MODE; idx++)
+ {
+ uint64_t fracBits = m_CABACEstimator->geo_mode_est(ctxStart, idx);
+ geoModeCost[idx] = (double)fracBits * sqrtLambdaFracBits;
+ }
+ for (int idx = 0; idx < maxNumMergeCandidates; idx++)
+ {
+ uint64_t fracBits = m_CABACEstimator->geo_mergeIdx_est(ctxStart, idx, maxNumMergeCandidates);
+ geoMergeIdxCost[idx] = (double)fracBits * sqrtLambdaFracBits;
+ }
+ for (int idx = 0; idx < 2; idx++)
+ {
+ uint64_t fracBits = m_CABACEstimator->geo_mmvdFlag_est(ctxStart, idx);
+ geoMMVDFlagCost[idx] = (double)fracBits * sqrtLambdaFracBits;
+ }
+ for (int idx = 0; idx < (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM); idx++)
+ {
+ uint64_t fractBits = m_CABACEstimator->geo_mmvdIdx_est(ctxStart, idx, extMMVD);
+ geoMMVDIdxCost[idx] = (double)fractBits * sqrtLambdaFracBits;
+ }
+#if TM_MRG
+ for (int idx = 0; idx < 2 && sps.getUseDMVDMode(); idx++)
+ {
+ uint64_t fracBits = m_CABACEstimator->geo_tmFlag_est(ctxStart, idx);
+ geoTMFlagCost[idx] = (double)fracBits * sqrtLambdaFracBits;
}
+#endif
+ m_CABACEstimator->getCtx() = ctxStart;
+ // 2. get SAD for all candidates
CodingUnit &cu = tempCS->addCU(tempCS->area, pm.chType);
pm.setCUData(cu);
cu.predMode = MODE_INTER;
@@ -3937,53 +4060,88 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
#if TM_MRG
pu.tmMergeFlag = false;
#endif
- PU::getGeoMergeCandidates(pu, mergeCtx);
+ CHECK(!m_mergeCandAvail, "merge candidates are not available");
+#if TM_MRG
+ PU::getGeoMergeCandidates(pu, mergeCtx[GEO_TM_OFF], &m_mergeCand);
+ maxNumMergeCandidates = min((int)maxNumMergeCandidates, mergeCtx[GEO_TM_OFF].numValidMergeCand);
+#else
+ PU::getGeoMergeCandidates(pu, mergeCtx, &m_mergeCand);
+ maxNumMergeCandidates = min((int)maxNumMergeCandidates, mergeCtx.numValidMergeCand);
+#endif
- GeoComboCostList comboList;
- int bitsCandTB = floorLog2(GEO_NUM_PARTITION_MODE);
PelUnitBuf geoBuffer[GEO_MAX_NUM_UNI_CANDS];
PelUnitBuf geoTempBuf[GEO_MAX_NUM_UNI_CANDS];
+ PelUnitBuf geoMMVDBuf[GEO_MAX_NUM_UNI_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
+ PelUnitBuf geoMMVDTempBuf[GEO_MAX_NUM_UNI_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
PelUnitBuf geoCombinations[GEO_MAX_TRY_WEIGHTED_SAD];
- DistParam distParam;
+ DistParam distParam;
const UnitArea localUnitArea(tempCS->area.chromaFormat, Area(0, 0, tempCS->area.Y().width, tempCS->area.Y().height));
- const double sqrtLambdaForFirstPass = m_pcRdCost->getMotionLambda();
- uint8_t maxNumMergeCandidates = cu.cs->sps->getMaxNumGeoCand();
DistParam distParamWholeBlk;
m_pcRdCost->setDistParam(distParamWholeBlk, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y().buf, m_acMergeBuffer[0].Y().stride, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
- Distortion bestWholeBlkSad = MAX_UINT64;
- double bestWholeBlkCost = MAX_DOUBLE;
+ Distortion sadWholeBlk[GEO_MAX_NUM_UNI_CANDS], sadMMVDWholeBlk[GEO_MAX_NUM_UNI_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
- Distortion sadWholeBlk[GEO_MAX_NUM_UNI_CANDS];
- int pocMrg[GEO_MAX_NUM_UNI_CANDS];
- Mv MrgMv[GEO_MAX_NUM_UNI_CANDS];
- bool isSkipThisCand[GEO_MAX_NUM_UNI_CANDS] = { false };
+ int pocMrg[GEO_MAX_NUM_UNI_CANDS];
+ Mv mrgMv[GEO_MAX_NUM_UNI_CANDS];
+ bool mrgDuplicated[GEO_MAX_NUM_UNI_CANDS];
+
+ double bestMrgCost = MAX_DOUBLE;
+ double bestNormalMrgCost = MAX_DOUBLE;
+
+ bool isGeoChromaAvail[GEO_MAX_NUM_UNI_CANDS];
+ bool isGeoMMVDChromaAvail[GEO_MAX_NUM_UNI_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
+ memset(isGeoChromaAvail, false, sizeof(bool) * GEO_MAX_NUM_UNI_CANDS);
+ memset(isGeoMMVDChromaAvail, false, sizeof(bool) * GEO_MAX_NUM_UNI_CANDS * GPM_EXT_MMVD_MAX_REFINE_NUM);
+#if TM_MRG
+ bool isGeoTmChromaAvail[GEO_TM_MAX_NUM_CANDS];
+ memset(isGeoTmChromaAvail, false, sizeof(bool) * GEO_TM_MAX_NUM_CANDS);
+#endif
for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
{
- geoBuffer[mergeCand] = m_acMergeBuffer[mergeCand].getBuf(localUnitArea);
- mergeCtx.setMergeInfo(pu, mergeCand);
- int MrgList = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + 0].refIdx == -1 ? 1 : 0;
- RefPicList MrgeRefPicList = (MrgList ? REF_PIC_LIST_1 : REF_PIC_LIST_0);
- int MrgrefIdx = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + MrgList].refIdx;
- pocMrg[mergeCand] = tempCS->slice->getRefPic(MrgeRefPicList, MrgrefIdx)->getPOC();
- MrgMv[mergeCand] = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + MrgList].mv;
-
- for( int i = 0; i < mergeCand; i++ )
+#if TM_MRG
+ int mrgList = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(mergeCand << 1) + 0].refIdx == -1 ? 1 : 0;
+ int mrgRefIdx = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(mergeCand << 1) + mrgList].refIdx;
+#else
+ int mrgList = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + 0].refIdx == -1 ? 1 : 0;
+ int mrgRefIdx = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + mrgList].refIdx;
+#endif
+ pocMrg[mergeCand] = tempCS->slice->getRefPic((RefPicList)mrgList, mrgRefIdx)->getPOC();
+#if TM_MRG
+ mrgMv[mergeCand] = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(mergeCand << 1) + mrgList].mv;
+#else
+ mrgMv[mergeCand] = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + mrgList].mv;
+#endif
+ mrgDuplicated[mergeCand] = false;
+ if (mergeCand)
{
- if( pocMrg[mergeCand] == pocMrg[i] && MrgMv[mergeCand] == MrgMv[i] )
+ for (int i = 0; i < mergeCand; i++)
{
- isSkipThisCand[mergeCand] = true;
- break;
+ if (pocMrg[mergeCand] == pocMrg[i] && mrgMv[mergeCand] == mrgMv[i])
+ {
+ mrgDuplicated[mergeCand] = true;
+ break;
+ }
}
}
-
+#if !MULTI_HYP_PRED
+ if (mrgDuplicated[mergeCand])
+ {
+ continue;
+ }
+#endif
+ geoBuffer[mergeCand] = m_acMergeBuffer[mergeCand].getBuf(localUnitArea);
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setMergeInfo(pu, mergeCand);
+#else
+ mergeCtx.setMergeInfo(pu, mergeCand);
+#endif
if (m_pcEncCfg->getMCTSEncConstraint() && (!(MCTSHelper::checkMvBufferForMCTSConstraint(pu))))
{
tempCS->initStructData(encTestMode.qp);
return;
}
- m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand]);
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand], REF_PIC_LIST_X, true, false);
#if MULTI_HYP_PRED
geoTempBuf[mergeCand] = m_acRealMergeBuffer[MRG_MAX_NUM_CANDS + mergeCand].getBuf(localUnitArea);
#else
@@ -3994,35 +4152,45 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
distParamWholeBlk.cur.buf = geoTempBuf[mergeCand].Y().buf;
distParamWholeBlk.cur.stride = geoTempBuf[mergeCand].Y().stride;
sadWholeBlk[mergeCand] = distParamWholeBlk.distFunc(distParamWholeBlk);
- if (sadWholeBlk[mergeCand] < bestWholeBlkSad)
+ double curCost = sadWholeBlk[mergeCand] + geoMergeIdxCost[mergeCand];
+ if (curCost < bestNormalMrgCost)
{
- bestWholeBlkSad = sadWholeBlk[mergeCand];
- bestWholeBlkCost = ( double ) bestWholeBlkSad + ( mergeCand + 1 ) * sqrtLambdaForFirstPass;
+ bestNormalMrgCost = curCost;
+ }
+ curCost += geoMMVDFlagCost[0];
+ if (curCost < bestMrgCost)
+ {
+ bestMrgCost = curCost;
}
}
#if MULTI_HYP_PRED
+#if TM_MRG
+ m_pcInterSearch->setGeoTmpBuffer(mergeCtx[GEO_TM_OFF]);
+#else
m_pcInterSearch->setGeoTmpBuffer(mergeCtx);
#endif
- bool isGeo = true;
- for (uint8_t mergeCand = 1; mergeCand < maxNumMergeCandidates; mergeCand++)
- {
- isGeo &= isSkipThisCand[mergeCand];
- }
- if (isGeo)
- {
- return;
- }
+#endif
int wIdx = floorLog2(cu.lwidth()) - GEO_MIN_CU_LOG2;
int hIdx = floorLog2(cu.lheight()) - GEO_MIN_CU_LOG2;
+ Distortion sadSmall = 0, sadLarge = 0;
+ int maskStride = 0, maskStride2 = 0, stepX = 1;
+ Pel* SADmask;
+ static_vector<int, GEO_NUM_PARTITION_MODE> selGeoModeList;
+ static_vector<double, GEO_NUM_PARTITION_MODE> selGeoModeRDList;
+ static_vector<int, 5> mergeCandList0[GEO_NUM_PARTITION_MODE];
+ static_vector<int, 5> mergeCandList1[GEO_NUM_PARTITION_MODE];
+ static_vector<int, 5> mmvdCandList0[GEO_NUM_PARTITION_MODE];
+ static_vector<int, 5> mmvdCandList1[GEO_NUM_PARTITION_MODE];
+ static_vector<double, 5> sadCostList0[GEO_NUM_PARTITION_MODE];
+ static_vector<double, 5> sadCostList1[GEO_NUM_PARTITION_MODE];
+
for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
{
- int maskStride = 0, maskStride2 = 0;
- int stepX = 1;
- Pel* SADmask;
int16_t angle = g_GeoParams[splitDir][0];
if (g_angle2mirror[angle] == 2)
{
+ stepX = 1;
maskStride = -GEO_WEIGHT_MASK_SIZE;
maskStride2 = -(int)cu.lwidth();
SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][(GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][1]) * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
@@ -4036,97 +4204,155 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
}
else
{
+ stepX = 1;
maskStride = GEO_WEIGHT_MASK_SIZE;
maskStride2 = -(int)cu.lwidth();
SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
}
- Distortion sadSmall = 0, sadLarge = 0;
for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
{
- int bitsCand = mergeCand + 1;
-
+ if (mrgDuplicated[mergeCand])
+ {
+ continue;
+ }
m_pcRdCost->setDistParam(distParam, tempCS->getOrgBuf().Y(), geoTempBuf[mergeCand].Y().buf, geoTempBuf[mergeCand].Y().stride, SADmask, maskStride, stepX, maskStride2, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
sadLarge = distParam.distFunc(distParam);
- m_GeoCostList.insert(splitDir, 0, mergeCand, (double)sadLarge + (double)bitsCand * sqrtLambdaForFirstPass);
+ double tempCost = (double)sadLarge + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[0];
+ m_GeoMMVDCostList.insert(splitDir, 0, mergeCand, 0, tempCost);
+ sortCandList(tempCost, mergeCand, 0, sadCostList0[splitDir], mergeCandList0[splitDir], mmvdCandList0[splitDir], numCandPerPar);
sadSmall = sadWholeBlk[mergeCand] - sadLarge;
- m_GeoCostList.insert(splitDir, 1, mergeCand, (double)sadSmall + (double)bitsCand * sqrtLambdaForFirstPass);
+ tempCost = (double)sadSmall + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[0];
+ m_GeoMMVDCostList.insert(splitDir, 1, mergeCand, 0, tempCost);
+ sortCandList(tempCost, mergeCand, 0, sadCostList1[splitDir], mergeCandList1[splitDir], mmvdCandList1[splitDir], numCandPerPar);
}
+ updateCandList(splitDir, (sadCostList0[splitDir][0] + sadCostList1[splitDir][0]), selGeoModeList, selGeoModeRDList, GEO_NUM_PARTITION_MODE);
}
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoSplitDirList;
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoMergeCand0;
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoMergeCand1;
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoMmvdCand0;
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoMmvdCand1;
+ static_vector<double, GEO_MAX_TRY_WEIGHTED_SAD> geoSADCostList;
+
for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
{
- for (int GeoMotionIdx = 0; GeoMotionIdx < maxNumMergeCandidates * (maxNumMergeCandidates - 1); GeoMotionIdx++)
+ int numCandPart0 = min(numCandPerPar, (int)mergeCandList0[splitDir].size());
+ int numCandPart1 = min(numCandPerPar, (int)mergeCandList1[splitDir].size());
+ for (int candIdx0 = 0; candIdx0 < numCandPart0; candIdx0++)
{
- unsigned int mergeCand0 = m_GeoModeTest[GeoMotionIdx].m_candIdx0;
- unsigned int mergeCand1 = m_GeoModeTest[GeoMotionIdx].m_candIdx1;
- double tempCost = m_GeoCostList.singleDistList[0][splitDir][mergeCand0].cost + m_GeoCostList.singleDistList[1][splitDir][mergeCand1].cost;
- if( tempCost > bestWholeBlkCost )
+ for (int candIdx1 = 0; candIdx1 < numCandPart1; candIdx1++)
{
- continue;
+ int mergeCand0 = mergeCandList0[splitDir][candIdx0];
+ int mergeCand1 = mergeCandList1[splitDir][candIdx1];
+
+ if (mergeCand0 == mergeCand1)
+ {
+ continue;
+ }
+
+ double tempCost = m_GeoMMVDCostList.singleDistList[0][splitDir][mergeCand0][0].cost + m_GeoMMVDCostList.singleDistList[1][splitDir][mergeCand1][0].cost;
+ tempCost = tempCost + geoModeCost[splitDir];
+#if TM_MRG
+ if (sps.getUseDMVDMode())
+ {
+ tempCost += geoTMFlagCost[0];
+ }
+#endif
+ updateGeoMMVDCandList(tempCost, splitDir, mergeCand0, mergeCand1, 0, 0, geoSADCostList, geoSplitDirList, geoMergeCand0, geoMergeCand1, geoMmvdCand0, geoMmvdCand1, numSATDCands);
}
- tempCost = tempCost + (double)bitsCandTB * sqrtLambdaForFirstPass;
- comboList.list.push_back(GeoMergeCombo(splitDir, mergeCand0, mergeCand1, tempCost));
}
}
- if( comboList.list.empty() )
- {
- return;
- }
-
- comboList.sortByCost();
- bool geocandHasNoResidual[GEO_MAX_TRY_WEIGHTED_SAD] = { false };
- bool bestIsSkip = false;
- int geoNumCobo = (int)comboList.list.size();
- static_vector<uint8_t, GEO_MAX_TRY_WEIGHTED_SAD> geoRdModeList;
- static_vector<double, GEO_MAX_TRY_WEIGHTED_SAD> geocandCostList;
+ static_vector<uint8_t, GEO_MAX_TRY_WEIGHTED_SAD> geoRdModeList;
+ static_vector<bool, GEO_MAX_TRY_WEIGHTED_SAD> isNonMMVDListIdx;
+ static_vector<int, GEO_MAX_TRY_WEIGHTED_SAD> geoPartitionModeList;
+ static_vector<double, GEO_MAX_TRY_WEIGHTED_SAD> geocandCostList;
DistParam distParamSAD2;
const bool useHadamard = !tempCS->slice->getDisableSATDForRD();
m_pcRdCost->setDistParam(distParamSAD2, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, useHadamard);
- int geoNumMrgSATDCand = min(GEO_MAX_TRY_WEIGHTED_SATD, geoNumCobo);
- for (uint8_t candidateIdx = 0; candidateIdx < min(geoNumCobo, GEO_MAX_TRY_WEIGHTED_SAD); candidateIdx++)
+ int numberGeoCandChecked = (int)geoSADCostList.size();
+ int geoNumMrgSATDCand = min(GEO_MAX_TRY_WEIGHTED_SATD, numberGeoCandChecked);
+ int numStoredCands = geoNumMrgSATDCand;
+ for (uint8_t candidateIdx = 0; candidateIdx < numberGeoCandChecked; candidateIdx++)
{
- int splitDir = comboList.list[candidateIdx].splitDir;
- int mergeCand0 = comboList.list[candidateIdx].mergeIdx0;
- int mergeCand1 = comboList.list[candidateIdx].mergeIdx1;
+ int splitDir = geoSplitDirList[candidateIdx];
+ int mergeCand0 = geoMergeCand0[candidateIdx];
+ int mergeCand1 = geoMergeCand1[candidateIdx];
+ bool mmvdFlag0 = false;
+ bool mmvdFlag1 = false;
geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_LUMA, geoCombinations[candidateIdx], geoBuffer[mergeCand0], geoBuffer[mergeCand1]);
distParamSAD2.cur = geoCombinations[candidateIdx].Y();
Distortion sad = distParamSAD2.distFunc(distParamSAD2);
- int mvBits = 2;
- mergeCand1 -= mergeCand1 < mergeCand0 ? 0 : 1;
- mvBits += mergeCand0;
- mvBits += mergeCand1;
- double updateCost = (double)sad + (double)(bitsCandTB + mvBits) * sqrtLambdaForFirstPass;
- comboList.list[candidateIdx].cost = updateCost;
- updateCandList(candidateIdx, updateCost, geoRdModeList, geocandCostList, geoNumMrgSATDCand);
+
+ double updateCost = geoModeCost[splitDir] + geoMergeIdxCost[mergeCand0] + (fastGpmMmvdSearch ? geoMergeIdxCost[mergeCand1 > mergeCand0 ? (mergeCand1 - 1) : mergeCand1] : geoMergeIdxCost[mergeCand1]) + geoMMVDFlagCost[mmvdFlag0] + geoMMVDFlagCost[mmvdFlag1];
+#if TM_MRG
+ if (sps.getUseDMVDMode())
+ {
+ updateCost += geoTMFlagCost[0];
+ }
+#endif
+ updateCost += (double)sad;
+ orderCandList(candidateIdx, true, splitDir, updateCost, geoRdModeList, isNonMMVDListIdx, geoPartitionModeList, geocandCostList, numStoredCands);
}
- for (uint8_t i = 0; i < geoNumMrgSATDCand; i++)
+
+ for (uint8_t i = 1; i < geoNumMrgSATDCand; i++)
{
+#if MERGE_ENC_OPT
+ if (geocandCostList[i] > MRG_FAST_RATIO * geocandCostList[0] || geocandCostList[i] > getMergeBestSATDCost())
+#else
if (geocandCostList[i] > MRG_FAST_RATIO * geocandCostList[0] || geocandCostList[i] > getMergeBestSATDCost() || geocandCostList[i] > getAFFBestSATDCost())
+#endif
{
geoNumMrgSATDCand = i;
break;
}
}
+
for (uint8_t i = 0; i < geoNumMrgSATDCand && isChromaEnabled(pu.chromaFormat); i++)
{
uint8_t candidateIdx = geoRdModeList[i];
- int splitDir = comboList.list[candidateIdx].splitDir;
- int mergeCand0 = comboList.list[candidateIdx].mergeIdx0;
- int mergeCand1 = comboList.list[candidateIdx].mergeIdx1;
+ int splitDir = geoSplitDirList[candidateIdx];
+ int mergeCand0 = geoMergeCand0[candidateIdx];
+ int mergeCand1 = geoMergeCand1[candidateIdx];
+
+ if (!isGeoChromaAvail[mergeCand0])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setMergeInfo(pu, mergeCand0);
+#else
+ mergeCtx.setMergeInfo(pu, mergeCand0);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand0], REF_PIC_LIST_X, false, true);
+ isGeoChromaAvail[mergeCand0] = true;
+ }
+
+ if (!isGeoChromaAvail[mergeCand1])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setMergeInfo(pu, mergeCand1);
+#else
+ mergeCtx.setMergeInfo(pu, mergeCand1);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand1], REF_PIC_LIST_X, false, true);
+ isGeoChromaAvail[mergeCand1] = true;
+ }
+
geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_CHROMA, geoCombinations[candidateIdx], geoBuffer[mergeCand0], geoBuffer[mergeCand1]);
}
+ bool geocandHasNoResidual[GEO_MAX_TRY_WEIGHTED_SAD];
+ bool bestIsSkip = false;
+ std::memset(geocandHasNoResidual, false, GEO_MAX_TRY_WEIGHTED_SAD * sizeof(bool));
+
m_bestModeUpdated = tempCS->useDbCost = bestCS->useDbCost = false;
tempCS->initStructData(encTestMode.qp);
- uint8_t iteration;
- uint8_t iterationBegin = 0;
- iteration = 2;
+ uint8_t iteration = 2, iterationBegin = 0;
for (uint8_t noResidualPass = iterationBegin; noResidualPass < iteration; ++noResidualPass)
{
for (uint8_t mrgHADIdx = 0; mrgHADIdx < geoNumMrgSATDCand; mrgHADIdx++)
@@ -4154,20 +4380,30 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
cu.mmvdSkip = false;
cu.skip = false;
cu.mipFlag = false;
-#if JVET_V0130_INTRA_TMP
- cu.tmpFlag = false;
-#endif
cu.bdpcmMode = 0;
PredictionUnit &pu = tempCS->addPU(cu, pm.chType);
pu.mergeFlag = true;
pu.regularMergeFlag = false;
- pu.geoSplitDir = comboList.list[candidateIdx].splitDir;
- pu.geoMergeIdx0 = comboList.list[candidateIdx].mergeIdx0;
- pu.geoMergeIdx1 = comboList.list[candidateIdx].mergeIdx1;
+ pu.geoSplitDir = geoSplitDirList[candidateIdx];
+ pu.geoMergeIdx0 = geoMergeCand0[candidateIdx];
+ pu.geoMergeIdx1 = geoMergeCand1[candidateIdx];
+#if TM_MRG
+ pu.tmMergeFlag = false;
+ pu.geoTmFlag0 = false;
+ pu.geoTmFlag1 = false;
+#endif
+ pu.geoMMVDFlag0 = false;
+ pu.geoMMVDFlag1 = false;
+
pu.mmvdMergeFlag = false;
pu.mmvdMergeIdx = MAX_UCHAR;
-
- PU::spanGeoMotionInfo(pu, mergeCtx, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1);
+#if TM_MRG
+ MergeCtx *mergeTmCtx0 = nullptr;
+ MergeCtx *mergeTmCtx1 = nullptr;
+ PU::spanGeoMMVDMotionInfo(pu, mergeCtx[GEO_TM_OFF], *mergeTmCtx0, *mergeTmCtx1, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1, pu.geoTmFlag0, pu.geoMMVDFlag0, pu.geoMMVDIdx0, pu.geoTmFlag1, pu.geoMMVDFlag1, pu.geoMMVDIdx1);
+#else
+ PU::spanGeoMMVDMotionInfo(pu, mergeCtx, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1, pu.geoMMVDFlag0, pu.geoMMVDIdx0, pu.geoMMVDFlag1, pu.geoMMVDIdx1);
+#endif
tempCS->getPredBuf().copyFrom(geoCombinations[candidateIdx]);
#if ENABLE_OBMC
cu.isobmcMC = true;
@@ -4175,6 +4411,7 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
m_pcInterSearch->subBlockOBMC(pu);
cu.isobmcMC = false;
#endif
+
xEncodeInterResidual(tempCS, bestCS, pm, encTestMode, noResidualPass, (noResidualPass == 0 ? &geocandHasNoResidual[candidateIdx] : NULL));
if (m_pcEncCfg->getUseFastDecisionForMerge() && !bestIsSkip)
@@ -4184,28 +4421,1102 @@ void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure
tempCS->initStructData(encTestMode.qp);
}
}
- if (m_bestModeUpdated && bestCS->cost != MAX_DOUBLE)
+
+ CodingUnit *bestCU = bestCS->getCU(CHANNEL_TYPE_LUMA);
+ bool skipGPMMMVD = false;
+ if (geoNumMrgSATDCand > 0)
{
- xCalDebCost(*bestCS, pm);
+ if (bestCU->skip && !bestCU->geoFlag && !bestCU->affine && !bestCU->mmvdSkip && !bestCU->firstPU->mmvdMergeFlag)
+ {
+ skipGPMMMVD = true;
+ }
+ else if (bestCU->affine && bestCU->skip && (bestCU->lwidth() >= 16 || bestCU->lheight() >= 16))
+ {
+ skipGPMMMVD = true;
+ }
}
-}
-#if MERGE_ENC_OPT
-void EncCu::xCheckSATDCostRegularMerge(CodingStructure *&tempCS, CodingUnit &cu, PredictionUnit &pu, MergeCtx mergeCtx, PelUnitBuf *acMergeTempBuffer[MMVD_MRG_MAX_RD_NUM], PelUnitBuf *&singleMergeTempBuffer, PelUnitBuf acMergeTmpBuffer[MRG_MAX_NUM_CANDS]
-#if !MULTI_PASS_DMVR
- , Mv refinedMvdL0[MAX_NUM_PARTS_IN_CTU][MRG_MAX_NUM_CANDS]
-#endif
- , unsigned& uiNumMrgSATDCand, static_vector<ModeInfo, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &RdModeList, static_vector<double, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &candCostList, DistParam distParam, const TempCtx &ctxStart
-#if MULTI_PASS_DMVR
- , bool* applyBDMVR
-#endif
-)
-{
-#if INTER_LIC
- cu.LICFlag = false;
-#endif
- cu.mmvdSkip = false;
- cu.geoFlag = false;
+ bool isBaseMergeCandIncluded[GEO_MAX_NUM_UNI_CANDS];
+ std::memset(isBaseMergeCandIncluded, false, GEO_MAX_NUM_UNI_CANDS * sizeof(bool));
+ bool isGPMModeIncludedForMMVD[GEO_NUM_PARTITION_MODE];
+ std::memset(isGPMModeIncludedForMMVD, false, GEO_NUM_PARTITION_MODE * sizeof(bool));
+
+ if (!skipGPMMMVD)
+ {
+ skipGPMMMVD = (selGeoModeRDList[0] > (bestNormalMrgCost * 1.1));
+ }
+
+ if (!skipGPMMMVD)
+ {
+ if (isSecondPass)
+ {
+ for (int i = 0; i < relatedCU.numGeoDirCand; i++)
+ {
+ isGPMModeIncludedForMMVD[relatedCU.geoDirCandList[i]] = true;
+ if (fastGpmMmvdRelatedCU)
+ {
+ isBaseMergeCandIncluded[relatedCU.geoMrgIdx0List[i]] = true;
+ isBaseMergeCandIncluded[relatedCU.geoMrgIdx1List[i]] = true;
+ }
+ }
+ if (!fastGpmMmvdRelatedCU)
+ {
+ std::memset(isBaseMergeCandIncluded, true, GEO_MAX_NUM_UNI_CANDS * sizeof(bool));
+ }
+ }
+ else
+ {
+ double dirCostThresh = (selGeoModeRDList[0] * 1.2);
+ isGPMModeIncludedForMMVD[selGeoModeList[0]] = true;
+ isBaseMergeCandIncluded[mergeCandList0[selGeoModeList[0]][0]] = true;
+ isBaseMergeCandIncluded[mergeCandList1[selGeoModeList[0]][0]] = true;
+
+ for (int i = 1; i < maxNumGPMDirFirstPass; i++)
+ {
+ if (selGeoModeRDList[i] > dirCostThresh)
+ {
+ break;
+ }
+ else
+ {
+ isGPMModeIncludedForMMVD[selGeoModeList[i]] = true;
+ isBaseMergeCandIncluded[mergeCandList0[selGeoModeList[i]][0]] = true;
+ isBaseMergeCandIncluded[mergeCandList1[selGeoModeList[i]][0]] = true;
+ }
+ }
+ if (includeMoreMMVDCandFirstPass)
+ {
+ int num = 0;
+ // add more cands from best combo results obtained in weighted blended nonmmvd combo
+ num = min((int)geocandCostList.size(), GEO_MAX_TRY_WEIGHTED_SATD);
+ for (int i = 0; i < num; i++)
+ {
+ if (fastGpmMmvdSearch && (geocandCostList[i] > dirCostThresh))
+ {
+ break;
+ }
+ else
+ {
+ isGPMModeIncludedForMMVD[geoPartitionModeList[i]] = true;
+ isBaseMergeCandIncluded[geoMergeCand0[geoRdModeList[i]]] = true;
+ isBaseMergeCandIncluded[geoMergeCand1[geoRdModeList[i]]] = true;
+ }
+ }
+ }
+ }
+ }
+
+ if (!skipGPMMMVD)
+ {
+ CodingUnit &cu = tempCS->addCU(tempCS->area, pm.chType);
+ pm.setCUData(cu);
+ cu.predMode = MODE_INTER;
+ cu.slice = tempCS->slice;
+ cu.tileIdx = tempCS->pps->getTileIdx(tempCS->area.lumaPos());
+ cu.qp = encTestMode.qp;
+ cu.affine = false;
+ cu.mtsFlag = false;
+#if INTER_LIC
+ cu.LICFlag = false;
+#endif
+ cu.BcwIdx = BCW_DEFAULT;
+ cu.geoFlag = true;
+ cu.imv = 0;
+ cu.mmvdSkip = false;
+ cu.skip = false;
+ cu.mipFlag = false;
+ cu.bdpcmMode = 0;
+
+ PredictionUnit &pu = tempCS->addPU(cu, pm.chType);
+ pu.mergeFlag = true;
+ pu.regularMergeFlag = false;
+#if TM_MRG
+ pu.tmMergeFlag = false;
+#endif
+ bool simpleGPMMMVDStep = (m_pcEncCfg->getIntraPeriod() == -1);
+ double mmvdMrgCost[GEO_MAX_NUM_UNI_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
+ for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
+ {
+ if (mrgDuplicated[mergeCand])
+ {
+ continue;
+ }
+ if (!isBaseMergeCandIncluded[mergeCand])
+ {
+ continue;
+ }
+ for (uint8_t mmvdCand = 0; mmvdCand < (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM); mmvdCand++)
+ {
+ if (simpleGPMMMVDStep)
+ {
+ int mmvdStep = (extMMVD ? (mmvdCand >> 3) : (mmvdCand >> 2));
+ if (mmvdStep >= 5 && (!fastGpmMmvdSearch || (fastGpmMmvdSearch && !isSecondPass)))
+ {
+ continue;
+ }
+ }
+ geoMMVDBuf[mergeCand][mmvdCand] = m_acGeoMMVDBuffer[mergeCand][mmvdCand].getBuf(localUnitArea);
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setGeoMmvdMergeInfo(pu, mergeCand, mmvdCand);
+#else
+ mergeCtx.setGeoMmvdMergeInfo(pu, mergeCand, mmvdCand);
+#endif
+ if (m_pcEncCfg->getMCTSEncConstraint() && (!(MCTSHelper::checkMvBufferForMCTSConstraint(pu))))
+ {
+ tempCS->initStructData(encTestMode.qp);
+ return;
+ }
+ m_pcInterSearch->motionCompensation(pu, geoMMVDBuf[mergeCand][mmvdCand], REF_PIC_LIST_X, true, false);
+ geoMMVDTempBuf[mergeCand][mmvdCand] = m_acGeoMMVDTmpBuffer[mergeCand][mmvdCand].getBuf(localUnitArea);
+ geoMMVDTempBuf[mergeCand][mmvdCand].Y().copyFrom(geoMMVDBuf[mergeCand][mmvdCand].Y());
+ geoMMVDTempBuf[mergeCand][mmvdCand].Y().roundToOutputBitdepth(geoMMVDTempBuf[mergeCand][mmvdCand].Y(), cu.slice->clpRng(COMPONENT_Y));
+ distParamWholeBlk.cur.buf = geoMMVDTempBuf[mergeCand][mmvdCand].Y().buf;
+ distParamWholeBlk.cur.stride = geoMMVDTempBuf[mergeCand][mmvdCand].Y().stride;
+ sadMMVDWholeBlk[mergeCand][mmvdCand] = distParamWholeBlk.distFunc(distParamWholeBlk);
+ mmvdMrgCost[mergeCand][mmvdCand] = sadMMVDWholeBlk[mergeCand][mmvdCand] + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[1] + geoMMVDIdxCost[mmvdCand];
+ if (mmvdMrgCost[mergeCand][mmvdCand] < bestMrgCost)
+ {
+ bestMrgCost = mmvdMrgCost[mergeCand][mmvdCand];
+ }
+ }
+ }
+
+ double mrgCostThres = (bestMrgCost * 3.0);
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ if (!isGPMModeIncludedForMMVD[splitDir])
+ {
+ continue;
+ }
+ int16_t angle = g_GeoParams[splitDir][0];
+ if (g_angle2mirror[angle] == 2)
+ {
+ stepX = 1;
+ maskStride = -GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][(GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][1]) * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ else if (g_angle2mirror[angle] == 1)
+ {
+ stepX = -1;
+ maskStride2 = cu.lwidth();
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + (GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][0])];
+ }
+ else
+ {
+ stepX = 1;
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
+ {
+ if (mrgDuplicated[mergeCand])
+ {
+ continue;
+ }
+ if (!isBaseMergeCandIncluded[mergeCand])
+ {
+ continue;
+ }
+ for (uint8_t mmvdCand = 0; mmvdCand < (extMMVD ? GPM_EXT_MMVD_MAX_REFINE_NUM : GPM_MMVD_MAX_REFINE_NUM); mmvdCand++)
+ {
+ if (simpleGPMMMVDStep)
+ {
+ int mmvdStep = (extMMVD ? (mmvdCand >> 3) : (mmvdCand >> 2));
+ if (mmvdStep >= 5 && (!fastGpmMmvdSearch || (fastGpmMmvdSearch && !isSecondPass)))
+ {
+ continue;
+ }
+ }
+ if (mmvdMrgCost[mergeCand][mmvdCand] > mrgCostThres)
+ {
+ continue;
+ }
+ m_pcRdCost->setDistParam(distParam, tempCS->getOrgBuf().Y(), geoMMVDTempBuf[mergeCand][mmvdCand].Y().buf, geoMMVDTempBuf[mergeCand][mmvdCand].Y().stride, SADmask, maskStride, stepX, maskStride2, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
+ sadLarge = distParam.distFunc(distParam);
+ double tempCost = (double)sadLarge + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[1] + geoMMVDIdxCost[mmvdCand];
+ m_GeoMMVDCostList.insert(splitDir, 0, mergeCand, (mmvdCand + 1), tempCost);
+ sortCandList(tempCost, mergeCand, (mmvdCand + 1), sadCostList0[splitDir], mergeCandList0[splitDir], mmvdCandList0[splitDir], numCandPerPar);
+
+ sadSmall = sadMMVDWholeBlk[mergeCand][mmvdCand] - sadLarge;
+ tempCost = (double)sadSmall + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[1] + geoMMVDIdxCost[mmvdCand];
+ m_GeoMMVDCostList.insert(splitDir, 1, mergeCand, (mmvdCand + 1), tempCost);
+ sortCandList(tempCost, mergeCand, (mmvdCand + 1), sadCostList1[splitDir], mergeCandList1[splitDir], mmvdCandList1[splitDir], numCandPerPar);
+ }
+ }
+ }
+
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ int numCandPart0 = min(numCandPerPar, (int)mergeCandList0[splitDir].size());
+ int numCandPart1 = min(numCandPerPar, (int)mergeCandList1[splitDir].size());
+ for (int candIdx0 = 0; candIdx0 < numCandPart0; candIdx0++)
+ {
+ for (int candIdx1 = 0; candIdx1 < numCandPart1; candIdx1++)
+ {
+ int mergeCand0 = mergeCandList0[splitDir][candIdx0];
+ int mergeCand1 = mergeCandList1[splitDir][candIdx1];
+ int mmvdCand0 = mmvdCandList0[splitDir][candIdx0];
+ int mmvdCand1 = mmvdCandList1[splitDir][candIdx1];
+#if TM_MRG
+ bool geoTmFlag0 = (mmvdCand0 == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ bool geoTmFlag1 = (mmvdCand1 == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ CHECK(geoTmFlag0 || geoTmFlag1, "GPM TM has not been tested by far");
+#endif
+ if ((mmvdCand0 == 0) && (mmvdCand1 == 0))
+ {
+ continue;
+ }
+ if ((mmvdCand0 == mmvdCand1) && (mmvdCand0 > 0))
+ {
+ if (mergeCand0 == mergeCand1)
+ {
+ continue;
+ }
+ }
+
+ double tempCost = m_GeoMMVDCostList.singleDistList[0][splitDir][mergeCand0][mmvdCand0].cost + m_GeoMMVDCostList.singleDistList[1][splitDir][mergeCand1][mmvdCand1].cost;
+ tempCost = tempCost + geoModeCost[splitDir];
+#if TM_MRG
+ if (sps.getUseDMVDMode())
+ {
+ tempCost += geoTMFlagCost[0];
+ }
+#endif
+ updateGeoMMVDCandList(tempCost, splitDir, mergeCand0, mergeCand1, mmvdCand0, mmvdCand1,
+ geoSADCostList, geoSplitDirList, geoMergeCand0, geoMergeCand1, geoMmvdCand0, geoMmvdCand1, numSATDCands);
+ }
+ }
+ }
+
+#if TM_MRG
+ uint8_t maxNumTmMrgCand = maxNumMergeCandidates;
+ PelUnitBuf geoTmBuffer[GEO_TM_MAX_NUM_CANDS];
+ PelUnitBuf geoTmTempBuf[GEO_TM_MAX_NUM_CANDS];
+ if (sps.getUseDMVDMode())
+ {
+ for (int i = GEO_TM_SHAPE_AL; i < GEO_NUM_TM_MV_CAND; i++)
+ {
+ mergeCtx[i].numValidMergeCand = maxNumTmMrgCand;
+ for (int idx = 0; idx < maxNumTmMrgCand; idx++)
+ {
+ if (mrgDuplicated[idx])
+ {
+ continue;
+ }
+ mergeCtx[i].BcwIdx[idx] = BCW_DEFAULT;
+ mergeCtx[i].useAltHpelIf[idx] = false;
+#if INTER_LIC
+ mergeCtx[i].LICFlags[idx] = false;
+#endif
+ mergeCtx[i].interDirNeighbours[idx] = mergeCtx[GEO_TM_OFF].interDirNeighbours[idx];
+ mergeCtx[i].mvFieldNeighbours[(idx << 1)].mv = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(idx << 1)].mv;
+ mergeCtx[i].mvFieldNeighbours[(idx << 1) + 1].mv = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(idx << 1) + 1].mv;
+ mergeCtx[i].mvFieldNeighbours[(idx << 1)].refIdx = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(idx << 1)].refIdx;
+ mergeCtx[i].mvFieldNeighbours[(idx << 1) + 1].refIdx = mergeCtx[GEO_TM_OFF].mvFieldNeighbours[(idx << 1) + 1].refIdx;
+ }
+ }
+
+ pu.tmMergeFlag = true;
+ Distortion sadTmWholeBlk[GEO_TM_MAX_NUM_CANDS];
+ for (uint8_t tmType = GEO_TM_SHAPE_AL; tmType < GEO_NUM_TM_MV_CAND; tmType++)
+ {
+ pu.geoTmType = tmType;
+ for (uint8_t mrgIdx = 0; mrgIdx < maxNumTmMrgCand; mrgIdx++)
+ {
+ if (mrgDuplicated[mrgIdx])
+ {
+ continue;
+ }
+ uint8_t mergeCand = mrgIdx + (tmType - 1) * GEO_MAX_NUM_UNI_CANDS;
+ mergeCtx[tmType].setMergeInfo(pu, mrgIdx);
+ m_pcInterSearch->deriveTMMv(pu);
+ mergeCtx[tmType].mvFieldNeighbours[(mrgIdx << 1)].mv = pu.mv[0];
+ mergeCtx[tmType].mvFieldNeighbours[(mrgIdx << 1) + 1].mv = pu.mv[1];
+
+ geoTmBuffer[mergeCand] = m_acGeoMergeTmpBuffer[mergeCand].getBuf(localUnitArea);
+ m_pcInterSearch->motionCompensation(pu, geoTmBuffer[mergeCand]);
+
+ // calculate SAD for each candidate
+ geoTmTempBuf[mergeCand] = m_acGeoSADTmpBuffer[mergeCand].getBuf(localUnitArea);
+ geoTmTempBuf[mergeCand].Y().copyFrom(geoTmBuffer[mergeCand].Y());
+ geoTmTempBuf[mergeCand].Y().roundToOutputBitdepth(geoTmTempBuf[mergeCand].Y(), cu.slice->clpRng(COMPONENT_Y));
+ distParamWholeBlk.cur.buf = geoTmTempBuf[mergeCand].Y().buf;
+ distParamWholeBlk.cur.stride = geoTmTempBuf[mergeCand].Y().stride;
+ sadTmWholeBlk[mergeCand] = distParamWholeBlk.distFunc(distParamWholeBlk);
+ }
+ }
+ pu.tmMergeFlag = false;
+
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ int16_t angle = g_GeoParams[splitDir][0];
+ if (g_angle2mirror[angle] == 2)
+ {
+ stepX = 1;
+ maskStride = -GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][(GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][1]) * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ else if (g_angle2mirror[angle] == 1)
+ {
+ stepX = -1;
+ maskStride2 = cu.lwidth();
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + (GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][0])];
+ }
+ else
+ {
+ stepX = 1;
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ for (uint8_t mergeCand = 0; mergeCand < maxNumTmMrgCand; mergeCand++)
+ {
+ if (mrgDuplicated[mergeCand])
+ {
+ continue;
+ }
+
+ uint8_t mergeCand0 = mergeCand + (g_geoTmShape[0][g_GeoParams[splitDir][0]] - 1) * GEO_MAX_NUM_UNI_CANDS;
+ m_pcRdCost->setDistParam(distParam, tempCS->getOrgBuf().Y(), geoTmTempBuf[mergeCand0].Y().buf, geoTmTempBuf[mergeCand0].Y().stride, SADmask, maskStride, stepX, maskStride2, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
+ sadLarge = distParam.distFunc(distParam);
+ double tempCost = (double)sadLarge + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[0];
+ m_GeoMMVDCostList.insert(splitDir, 0, mergeCand, (GPM_EXT_MMVD_MAX_REFINE_NUM + 1), tempCost);
+
+ uint8_t mergeCand1 = mergeCand + (g_geoTmShape[1][g_GeoParams[splitDir][0]] - 1) * GEO_MAX_NUM_UNI_CANDS;
+ m_pcRdCost->setDistParam(distParam, tempCS->getOrgBuf().Y(), geoTmTempBuf[mergeCand1].Y().buf, geoTmTempBuf[mergeCand1].Y().stride, SADmask, maskStride, stepX, maskStride2, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
+ sadSmall = sadTmWholeBlk[mergeCand1] - distParam.distFunc(distParam);
+ tempCost = (double)sadSmall + geoMergeIdxCost[mergeCand] + geoMMVDFlagCost[0];
+ m_GeoMMVDCostList.insert(splitDir, 1, mergeCand, (GPM_EXT_MMVD_MAX_REFINE_NUM + 1), tempCost);
+ }
+ }
+
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ for (int mergeCand0 = 0; mergeCand0 < maxNumTmMrgCand; mergeCand0++)
+ {
+ if (mrgDuplicated[mergeCand0])
+ {
+ continue;
+ }
+ for (int mergeCand1 = 0; mergeCand1 < maxNumTmMrgCand; mergeCand1++)
+ {
+ if (mrgDuplicated[mergeCand1])
+ {
+ continue;
+ }
+ if (mergeCand0 == mergeCand1)
+ {
+ continue;
+ }
+ double tempCost = m_GeoMMVDCostList.singleDistList[0][splitDir][mergeCand0][GPM_EXT_MMVD_MAX_REFINE_NUM + 1].cost + m_GeoMMVDCostList.singleDistList[1][splitDir][mergeCand1][GPM_EXT_MMVD_MAX_REFINE_NUM + 1].cost;
+ tempCost = tempCost + geoModeCost[splitDir] + geoTMFlagCost[1];
+ updateGeoMMVDCandList(tempCost, splitDir, mergeCand0, mergeCand1, (GPM_EXT_MMVD_MAX_REFINE_NUM + 1), (GPM_EXT_MMVD_MAX_REFINE_NUM + 1),
+ geoSADCostList, geoSplitDirList, geoMergeCand0, geoMergeCand1, geoMmvdCand0, geoMmvdCand1, numSATDCands);
+ }
+ }
+ }
+ }
+#endif
+
+ int numberGeoCandChecked = (int)geoSADCostList.size();
+ if (numberGeoCandChecked == 0)
+ {
+ return;
+ }
+
+ geoNumMrgSATDCand = min(GEO_MAX_TRY_WEIGHTED_SATD, numberGeoCandChecked);
+ if (geoRdModeList.size() > geoNumMrgSATDCand)
+ {
+ geoRdModeList.resize(geoNumMrgSATDCand);
+ isNonMMVDListIdx.resize(geoNumMrgSATDCand);
+ geoPartitionModeList.resize(geoNumMrgSATDCand);
+ geocandCostList.resize(geoNumMrgSATDCand);
+ }
+ for (uint8_t candidateIdx = 0; candidateIdx < numberGeoCandChecked; candidateIdx++)
+ {
+ int splitDir = geoSplitDirList[candidateIdx];
+ int mergeCand0 = geoMergeCand0[candidateIdx];
+ int mergeCand1 = geoMergeCand1[candidateIdx];
+#if TM_MRG
+ bool tmFlag0 = (geoMmvdCand0[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ bool tmFlag1 = (geoMmvdCand1[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ bool mmvdFlag0 = (geoMmvdCand0[candidateIdx] >= 1 && geoMmvdCand0[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ bool mmvdFlag1 = (geoMmvdCand1[candidateIdx] >= 1 && geoMmvdCand1[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ int mmvdCand0 = (mmvdFlag0 ? (geoMmvdCand0[candidateIdx] - 1) : MAX_INT);
+ int mmvdCand1 = (mmvdFlag1 ? (geoMmvdCand1[candidateIdx] - 1) : MAX_INT);
+ CHECK(tmFlag0 != tmFlag1, "TM flag cannot be enabled/disabled for two partitions separately");
+
+ if (!tmFlag0 && !tmFlag1 && !mmvdFlag0 && !mmvdFlag1)
+ {
+ continue;
+ }
+ if (tmFlag0 && mergeCand0 == mergeCand1)
+ {
+ continue;
+ }
+#else
+ int mmvdCand0 = geoMmvdCand0[candidateIdx] - 1;
+ int mmvdCand1 = geoMmvdCand1[candidateIdx] - 1;
+ bool mmvdFlag0 = (mmvdCand0 >= 0);
+ bool mmvdFlag1 = (mmvdCand1 >= 0);
+
+ if (!mmvdFlag0 && !mmvdFlag1)
+ {
+ continue;
+ }
+#endif
+
+ geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
+#if TM_MRG
+ PelUnitBuf predSrc0, predSrc1;
+ if (tmFlag0)
+ {
+ int mrgTmCand0 = mergeCand0 + (g_geoTmShape[0][g_GeoParams[splitDir][0]] - 1) * GEO_MAX_NUM_UNI_CANDS;
+ predSrc0 = geoTmBuffer[mrgTmCand0];
+ }
+ else if (mmvdFlag0)
+ {
+ predSrc0 = geoMMVDBuf[mergeCand0][mmvdCand0];
+ }
+ else
+ {
+ predSrc0 = geoBuffer[mergeCand0];
+ }
+
+ if (tmFlag1)
+ {
+ int mrgTmCand1 = mergeCand1 + (g_geoTmShape[1][g_GeoParams[splitDir][0]] - 1) * GEO_MAX_NUM_UNI_CANDS;
+ predSrc1 = geoTmBuffer[mrgTmCand1];
+ }
+ else if (mmvdFlag1)
+ {
+ predSrc1 = geoMMVDBuf[mergeCand1][mmvdCand1];
+ }
+ else
+ {
+ predSrc1 = geoBuffer[mergeCand1];
+ }
+
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_LUMA, geoCombinations[candidateIdx], predSrc0, predSrc1);
+#else
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_LUMA, geoCombinations[candidateIdx], (mmvdFlag0 ? geoMMVDBuf[mergeCand0][mmvdCand0] : geoBuffer[mergeCand0]), (mmvdFlag1 ? geoMMVDBuf[mergeCand1][mmvdCand1] : geoBuffer[mergeCand1]));
+#endif
+ distParamSAD2.cur = geoCombinations[candidateIdx].Y();
+ Distortion sad = distParamSAD2.distFunc(distParamSAD2);
+
+ double updateCost = geoModeCost[splitDir] + geoMergeIdxCost[mergeCand0] + geoMergeIdxCost[mergeCand1] + geoMMVDFlagCost[mmvdFlag0] + geoMMVDFlagCost[mmvdFlag1];
+ if (mmvdFlag0)
+ {
+ updateCost += geoMMVDIdxCost[mmvdCand0];
+ }
+ if (mmvdFlag1)
+ {
+ updateCost += geoMMVDIdxCost[mmvdCand1];
+ }
+#if TM_MRG
+ if (sps.getUseDMVDMode())
+ {
+ if (!mmvdFlag0 && !mmvdFlag1)
+ {
+ updateCost += geoTMFlagCost[tmFlag0];
+ }
+ }
+#endif
+ updateCost += (double)sad;
+ orderCandList(candidateIdx, false, splitDir, updateCost, geoRdModeList, isNonMMVDListIdx, geoPartitionModeList, geocandCostList, geoNumMrgSATDCand);
+ }
+
+ if (fastGpmMmvdRelatedCU)
+ {
+ int cnt = 0;
+ for (uint8_t i = 0; i < geoNumMrgSATDCand; i++)
+ {
+ if (isNonMMVDListIdx[i] == false)
+ {
+ relatedCU.geoDirCandList[cnt] = geoPartitionModeList[i];
+ relatedCU.geoMrgIdx0List[cnt] = geoMergeCand0[geoRdModeList[i]];
+ relatedCU.geoMrgIdx1List[cnt] = geoMergeCand1[geoRdModeList[i]];
+ cnt++;
+ }
+ }
+ relatedCU.numGeoDirCand = cnt;
+ }
+ else
+ {
+ relatedCU.numGeoDirCand = geoNumMrgSATDCand;
+ for (uint8_t i = 0; i < geoNumMrgSATDCand; i++)
+ {
+ relatedCU.geoDirCandList[i] = geoPartitionModeList[i];
+ }
+ }
+
+ for (uint8_t i = 1; i < geoNumMrgSATDCand; i++)
+ {
+#if MERGE_ENC_OPT
+ if (geocandCostList[i] > MRG_FAST_RATIO * geocandCostList[0] || geocandCostList[i] > getMergeBestSATDCost())
+#else
+ if (geocandCostList[i] > MRG_FAST_RATIO * geocandCostList[0] || geocandCostList[i] > getMergeBestSATDCost() || geocandCostList[i] > getAFFBestSATDCost())
+#endif
+ {
+ geoNumMrgSATDCand = i;
+ break;
+ }
+ }
+ for (uint8_t i = 0; i < geoNumMrgSATDCand && isChromaEnabled(pu.chromaFormat); i++)
+ {
+ if (isNonMMVDListIdx[i])
+ {
+ continue;
+ }
+ uint8_t candidateIdx = geoRdModeList[i];
+ int splitDir = geoSplitDirList[candidateIdx];
+ int mergeCand0 = geoMergeCand0[candidateIdx];
+ int mergeCand1 = geoMergeCand1[candidateIdx];
+#if TM_MRG
+ bool tmFlag0 = (geoMmvdCand0[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ bool tmFlag1 = (geoMmvdCand1[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ bool mmvdFlag0 = (geoMmvdCand0[candidateIdx] >= 1 && geoMmvdCand0[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ bool mmvdFlag1 = (geoMmvdCand1[candidateIdx] >= 1 && geoMmvdCand1[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ int mmvdCand0 = (mmvdFlag0 ? (geoMmvdCand0[candidateIdx] - 1) : MAX_INT);
+ int mmvdCand1 = (mmvdFlag1 ? (geoMmvdCand1[candidateIdx] - 1) : MAX_INT);
+#else
+ int mmvdCand0 = geoMmvdCand0[candidateIdx] - 1;
+ int mmvdCand1 = geoMmvdCand1[candidateIdx] - 1;
+
+ bool mmvdFlag0 = (mmvdCand0 >= 0);
+ bool mmvdFlag1 = (mmvdCand1 >= 0);
+#endif
+
+#if TM_MRG
+ int mrgTmCand0 = MAX_INT, mrgTmCand1 = MAX_INT;
+ if (tmFlag0)
+ {
+ int geoTmType = g_geoTmShape[0][g_GeoParams[splitDir][0]];
+ mrgTmCand0 = mergeCand0 + (geoTmType - 1) * GEO_MAX_NUM_UNI_CANDS;
+ if (!isGeoTmChromaAvail[mrgTmCand0])
+ {
+ mergeCtx[geoTmType].setMergeInfo(pu, mergeCand0);
+ m_pcInterSearch->motionCompensation(pu, geoTmBuffer[mrgTmCand0], REF_PIC_LIST_X, false, true);
+ isGeoTmChromaAvail[mrgTmCand0] = true;
+ }
+ }
+ else
+#endif
+ if (mmvdFlag0)
+ {
+ if (!isGeoMMVDChromaAvail[mergeCand0][mmvdCand0])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setGeoMmvdMergeInfo(pu, mergeCand0, mmvdCand0);
+#else
+ mergeCtx.setGeoMmvdMergeInfo(pu, mergeCand0, mmvdCand0);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoMMVDBuf[mergeCand0][mmvdCand0], REF_PIC_LIST_X, false, true);
+ isGeoMMVDChromaAvail[mergeCand0][mmvdCand0] = true;
+ }
+ }
+ else
+ {
+ if (!isGeoChromaAvail[mergeCand0])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setMergeInfo(pu, mergeCand0);
+#else
+ mergeCtx.setMergeInfo(pu, mergeCand0);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand0], REF_PIC_LIST_X, false, true);
+ isGeoChromaAvail[mergeCand0] = true;
+ }
+ }
+
+#if TM_MRG
+ if (tmFlag1)
+ {
+ int geoTmType = g_geoTmShape[1][g_GeoParams[splitDir][0]];
+ mrgTmCand1 = mergeCand1 + (geoTmType - 1) * GEO_MAX_NUM_UNI_CANDS;
+ if (!isGeoTmChromaAvail[mrgTmCand1])
+ {
+ mergeCtx[geoTmType].setMergeInfo(pu, mergeCand1);
+ m_pcInterSearch->motionCompensation(pu, geoTmBuffer[mrgTmCand1], REF_PIC_LIST_X, false, true);
+ isGeoTmChromaAvail[mrgTmCand1] = true;
+ }
+ }
+ else
+#endif
+ if (mmvdFlag1)
+ {
+ if (!isGeoMMVDChromaAvail[mergeCand1][mmvdCand1])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setGeoMmvdMergeInfo(pu, mergeCand1, mmvdCand1);
+#else
+ mergeCtx.setGeoMmvdMergeInfo(pu, mergeCand1, mmvdCand1);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoMMVDBuf[mergeCand1][mmvdCand1], REF_PIC_LIST_X, false, true);
+ isGeoMMVDChromaAvail[mergeCand1][mmvdCand1] = true;
+ }
+ }
+ else
+ {
+ if (!isGeoChromaAvail[mergeCand1])
+ {
+#if TM_MRG
+ mergeCtx[GEO_TM_OFF].setMergeInfo(pu, mergeCand1);
+#else
+ mergeCtx.setMergeInfo(pu, mergeCand1);
+#endif
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand1], REF_PIC_LIST_X, false, true);
+ isGeoChromaAvail[mergeCand1] = true;
+ }
+ }
+
+ geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
+#if TM_MRG
+ PelUnitBuf predSrc0, predSrc1;
+ if (tmFlag0)
+ {
+ predSrc0 = geoTmBuffer[mrgTmCand0];
+ }
+ else if (mmvdFlag0)
+ {
+ predSrc0 = geoMMVDBuf[mergeCand0][mmvdCand0];
+ }
+ else
+ {
+ predSrc0 = geoBuffer[mergeCand0];
+ }
+
+ if (tmFlag1)
+ {
+ predSrc1 = geoTmBuffer[mrgTmCand1];
+ }
+ else if (mmvdFlag1)
+ {
+ predSrc1 = geoMMVDBuf[mergeCand1][mmvdCand1];
+ }
+ else
+ {
+ predSrc1 = geoBuffer[mergeCand1];
+ }
+
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_CHROMA, geoCombinations[candidateIdx], predSrc0, predSrc1);
+#else
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_CHROMA, geoCombinations[candidateIdx], (mmvdFlag0 ? geoMMVDBuf[mergeCand0][mmvdCand0] : geoBuffer[mergeCand0]), (mmvdFlag1 ? geoMMVDBuf[mergeCand1][mmvdCand1] : geoBuffer[mergeCand1]));
+#endif
+ }
+
+ std::memset(geocandHasNoResidual, false, GEO_MAX_TRY_WEIGHTED_SAD * sizeof(bool));
+ tempCS->initStructData(encTestMode.qp);
+ uint8_t iteration = 2, iterationBegin = 0;
+ for (uint8_t noResidualPass = iterationBegin; noResidualPass < iteration; ++noResidualPass)
+ {
+ for (uint8_t mrgHADIdx = 0; mrgHADIdx < geoNumMrgSATDCand; mrgHADIdx++)
+ {
+ if (isNonMMVDListIdx[mrgHADIdx])
+ {
+ continue;
+ }
+ uint8_t candidateIdx = geoRdModeList[mrgHADIdx];
+ if (((noResidualPass != 0) && geocandHasNoResidual[candidateIdx])
+ || ((noResidualPass == 0) && bestIsSkip))
+ {
+ continue;
+ }
+ CodingUnit &cu = tempCS->addCU(tempCS->area, pm.chType);
+ pm.setCUData(cu);
+ cu.predMode = MODE_INTER;
+ cu.slice = tempCS->slice;
+ cu.tileIdx = tempCS->pps->getTileIdx(tempCS->area.lumaPos());
+ cu.qp = encTestMode.qp;
+ cu.affine = false;
+ cu.mtsFlag = false;
+#if INTER_LIC
+ cu.LICFlag = false;
+#endif
+ cu.BcwIdx = BCW_DEFAULT;
+ cu.geoFlag = true;
+ cu.imv = 0;
+ cu.mmvdSkip = false;
+ cu.skip = false;
+ cu.mipFlag = false;
+ cu.bdpcmMode = 0;
+ PredictionUnit &pu = tempCS->addPU(cu, pm.chType);
+ pu.mergeFlag = true;
+ pu.regularMergeFlag = false;
+ pu.geoSplitDir = geoSplitDirList[candidateIdx];
+ pu.geoMergeIdx0 = geoMergeCand0[candidateIdx];
+ pu.geoMergeIdx1 = geoMergeCand1[candidateIdx];
+#if TM_MRG
+ pu.geoTmFlag0 = (geoMmvdCand0[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ pu.geoTmFlag1 = (geoMmvdCand1[candidateIdx] == (GPM_EXT_MMVD_MAX_REFINE_NUM + 1));
+ pu.geoMMVDFlag0 = (geoMmvdCand0[candidateIdx] >= 1 && geoMmvdCand0[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ pu.geoMMVDFlag1 = (geoMmvdCand1[candidateIdx] >= 1 && geoMmvdCand1[candidateIdx] <= GPM_EXT_MMVD_MAX_REFINE_NUM);
+ CHECK(pu.geoTmFlag0 != pu.geoTmFlag1, "TM flag cannot be enabled/disabled for two partitions separately");
+ pu.tmMergeFlag = pu.geoTmFlag0;
+#else
+ pu.geoMMVDFlag0 = (geoMmvdCand0[candidateIdx] > 0);
+ pu.geoMMVDFlag1 = (geoMmvdCand1[candidateIdx] > 0);
+#endif
+
+ if (pu.geoMMVDFlag0)
+ {
+ pu.geoMMVDIdx0 = (geoMmvdCand0[candidateIdx] - 1);
+ }
+ if (pu.geoMMVDFlag1)
+ {
+ pu.geoMMVDIdx1 = (geoMmvdCand1[candidateIdx] - 1);
+ }
+ pu.mmvdMergeFlag = false;
+ pu.mmvdMergeIdx = MAX_UCHAR;
+#if TM_MRG
+ MergeCtx* mrgTmCtx0 = (pu.geoTmFlag0 == 0 ? nullptr : &mergeCtx[g_geoTmShape[0][g_GeoParams[pu.geoSplitDir][0]]]);
+ MergeCtx* mrgTmCtx1 = (pu.geoTmFlag1 == 0 ? nullptr : &mergeCtx[g_geoTmShape[1][g_GeoParams[pu.geoSplitDir][0]]]);
+ PU::spanGeoMMVDMotionInfo(pu, mergeCtx[GEO_TM_OFF], *mrgTmCtx0, *mrgTmCtx1, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1, pu.geoTmFlag0, pu.geoMMVDFlag0, pu.geoMMVDIdx0, pu.geoTmFlag1, pu.geoMMVDFlag1, pu.geoMMVDIdx1);
+#else
+ PU::spanGeoMMVDMotionInfo(pu, mergeCtx, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1, pu.geoMMVDFlag0, pu.geoMMVDIdx0, pu.geoMMVDFlag1, pu.geoMMVDIdx1);
+#endif
+ tempCS->getPredBuf().copyFrom(geoCombinations[candidateIdx]);
+#if ENABLE_OBMC
+ cu.isobmcMC = true;
+ cu.obmcFlag = true;
+ m_pcInterSearch->subBlockOBMC(pu);
+ cu.isobmcMC = false;
+#endif
+
+ xEncodeInterResidual(tempCS, bestCS, pm, encTestMode, noResidualPass, (noResidualPass == 0 ? &geocandHasNoResidual[candidateIdx] : NULL));
+
+ if (m_pcEncCfg->getUseFastDecisionForMerge() && !bestIsSkip)
+ {
+ bestIsSkip = bestCS->getCU(pm.chType)->rootCbf == 0;
+ }
+ tempCS->initStructData(encTestMode.qp);
+ }
+ }
+ }
+ if (m_bestModeUpdated && bestCS->cost != MAX_DOUBLE)
+ {
+ xCalDebCost(*bestCS, pm);
+ }
+}
+#else
+void EncCu::xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode)
+{
+ const Slice &slice = *tempCS->slice;
+ CHECK(slice.getSliceType() == I_SLICE, "Merge modes not available for I-slices");
+
+ tempCS->initStructData(encTestMode.qp);
+
+ MergeCtx mergeCtx;
+ const SPS &sps = *tempCS->sps;
+
+ if (sps.getSbTMVPEnabledFlag())
+ {
+ Size bufSize = g_miScaling.scale(tempCS->area.lumaSize());
+ mergeCtx.subPuMvpMiBuf = MotionBuf(m_SubPuMiBuf, bufSize);
+ }
+
+ CodingUnit &cu = tempCS->addCU(tempCS->area, pm.chType);
+ pm.setCUData(cu);
+ cu.predMode = MODE_INTER;
+ cu.slice = tempCS->slice;
+ cu.tileIdx = tempCS->pps->getTileIdx(tempCS->area.lumaPos());
+ cu.qp = encTestMode.qp;
+ cu.affine = false;
+ cu.mtsFlag = false;
+#if INTER_LIC
+ cu.LICFlag = false;
+#endif
+ cu.BcwIdx = BCW_DEFAULT;
+ cu.geoFlag = true;
+ cu.imv = 0;
+ cu.mmvdSkip = false;
+ cu.skip = false;
+ cu.mipFlag = false;
+#if JVET_V0130_INTRA_TMP
+ cu.tmpFlag = false;
+#endif
+ cu.bdpcmMode = 0;
+
+ PredictionUnit &pu = tempCS->addPU(cu, pm.chType);
+ pu.mergeFlag = true;
+ pu.regularMergeFlag = false;
+#if TM_MRG
+ pu.tmMergeFlag = false;
+#endif
+ PU::getGeoMergeCandidates(pu, mergeCtx);
+
+ GeoComboCostList comboList;
+ int bitsCandTB = floorLog2(GEO_NUM_PARTITION_MODE);
+ PelUnitBuf geoBuffer[GEO_MAX_NUM_UNI_CANDS];
+ PelUnitBuf geoTempBuf[GEO_MAX_NUM_UNI_CANDS];
+ PelUnitBuf geoCombinations[GEO_MAX_TRY_WEIGHTED_SAD];
+ DistParam distParam;
+
+ const UnitArea localUnitArea(tempCS->area.chromaFormat, Area(0, 0, tempCS->area.Y().width, tempCS->area.Y().height));
+ const double sqrtLambdaForFirstPass = m_pcRdCost->getMotionLambda();
+ uint8_t maxNumMergeCandidates = cu.cs->sps->getMaxNumGeoCand();
+ DistParam distParamWholeBlk;
+ m_pcRdCost->setDistParam(distParamWholeBlk, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y().buf, m_acMergeBuffer[0].Y().stride, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
+ Distortion bestWholeBlkSad = MAX_UINT64;
+ double bestWholeBlkCost = MAX_DOUBLE;
+
+ Distortion sadWholeBlk[GEO_MAX_NUM_UNI_CANDS];
+ int pocMrg[GEO_MAX_NUM_UNI_CANDS];
+ Mv MrgMv[GEO_MAX_NUM_UNI_CANDS];
+ bool isSkipThisCand[GEO_MAX_NUM_UNI_CANDS] = { false };
+
+ for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
+ {
+ geoBuffer[mergeCand] = m_acMergeBuffer[mergeCand].getBuf(localUnitArea);
+ mergeCtx.setMergeInfo(pu, mergeCand);
+ int MrgList = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + 0].refIdx == -1 ? 1 : 0;
+ RefPicList MrgeRefPicList = (MrgList ? REF_PIC_LIST_1 : REF_PIC_LIST_0);
+ int MrgrefIdx = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + MrgList].refIdx;
+ pocMrg[mergeCand] = tempCS->slice->getRefPic(MrgeRefPicList, MrgrefIdx)->getPOC();
+ MrgMv[mergeCand] = mergeCtx.mvFieldNeighbours[(mergeCand << 1) + MrgList].mv;
+
+ for( int i = 0; i < mergeCand; i++ )
+ {
+ if( pocMrg[mergeCand] == pocMrg[i] && MrgMv[mergeCand] == MrgMv[i] )
+ {
+ isSkipThisCand[mergeCand] = true;
+ break;
+ }
+ }
+
+ if (m_pcEncCfg->getMCTSEncConstraint() && (!(MCTSHelper::checkMvBufferForMCTSConstraint(pu))))
+ {
+ tempCS->initStructData(encTestMode.qp);
+ return;
+ }
+ m_pcInterSearch->motionCompensation(pu, geoBuffer[mergeCand]);
+#if MULTI_HYP_PRED
+ geoTempBuf[mergeCand] = m_acRealMergeBuffer[MRG_MAX_NUM_CANDS + mergeCand].getBuf(localUnitArea);
+#else
+ geoTempBuf[mergeCand] = m_acMergeTmpBuffer[mergeCand].getBuf(localUnitArea);
+#endif
+ geoTempBuf[mergeCand].Y().copyFrom(geoBuffer[mergeCand].Y());
+ geoTempBuf[mergeCand].Y().roundToOutputBitdepth(geoTempBuf[mergeCand].Y(), cu.slice->clpRng(COMPONENT_Y));
+ distParamWholeBlk.cur.buf = geoTempBuf[mergeCand].Y().buf;
+ distParamWholeBlk.cur.stride = geoTempBuf[mergeCand].Y().stride;
+ sadWholeBlk[mergeCand] = distParamWholeBlk.distFunc(distParamWholeBlk);
+ if (sadWholeBlk[mergeCand] < bestWholeBlkSad)
+ {
+ bestWholeBlkSad = sadWholeBlk[mergeCand];
+ bestWholeBlkCost = ( double ) bestWholeBlkSad + ( mergeCand + 1 ) * sqrtLambdaForFirstPass;
+ }
+ }
+#if MULTI_HYP_PRED
+ m_pcInterSearch->setGeoTmpBuffer(mergeCtx);
+#endif
+ bool isGeo = true;
+ for (uint8_t mergeCand = 1; mergeCand < maxNumMergeCandidates; mergeCand++)
+ {
+ isGeo &= isSkipThisCand[mergeCand];
+ }
+ if (isGeo)
+ {
+ return;
+ }
+
+ int wIdx = floorLog2(cu.lwidth()) - GEO_MIN_CU_LOG2;
+ int hIdx = floorLog2(cu.lheight()) - GEO_MIN_CU_LOG2;
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ int maskStride = 0, maskStride2 = 0;
+ int stepX = 1;
+ Pel* SADmask;
+ int16_t angle = g_GeoParams[splitDir][0];
+ if (g_angle2mirror[angle] == 2)
+ {
+ maskStride = -GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][(GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][1]) * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ else if (g_angle2mirror[angle] == 1)
+ {
+ stepX = -1;
+ maskStride2 = cu.lwidth();
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + (GEO_WEIGHT_MASK_SIZE - 1 - g_weightOffset[splitDir][hIdx][wIdx][0])];
+ }
+ else
+ {
+ maskStride = GEO_WEIGHT_MASK_SIZE;
+ maskStride2 = -(int)cu.lwidth();
+ SADmask = &g_globalGeoEncSADmask[g_angle2mask[g_GeoParams[splitDir][0]]][g_weightOffset[splitDir][hIdx][wIdx][1] * GEO_WEIGHT_MASK_SIZE + g_weightOffset[splitDir][hIdx][wIdx][0]];
+ }
+ Distortion sadSmall = 0, sadLarge = 0;
+ for (uint8_t mergeCand = 0; mergeCand < maxNumMergeCandidates; mergeCand++)
+ {
+ int bitsCand = mergeCand + 1;
+
+ m_pcRdCost->setDistParam(distParam, tempCS->getOrgBuf().Y(), geoTempBuf[mergeCand].Y().buf, geoTempBuf[mergeCand].Y().stride, SADmask, maskStride, stepX, maskStride2, sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y);
+ sadLarge = distParam.distFunc(distParam);
+ m_GeoCostList.insert(splitDir, 0, mergeCand, (double)sadLarge + (double)bitsCand * sqrtLambdaForFirstPass);
+ sadSmall = sadWholeBlk[mergeCand] - sadLarge;
+ m_GeoCostList.insert(splitDir, 1, mergeCand, (double)sadSmall + (double)bitsCand * sqrtLambdaForFirstPass);
+ }
+ }
+
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ for (int GeoMotionIdx = 0; GeoMotionIdx < maxNumMergeCandidates * (maxNumMergeCandidates - 1); GeoMotionIdx++)
+ {
+ unsigned int mergeCand0 = m_GeoModeTest[GeoMotionIdx].m_candIdx0;
+ unsigned int mergeCand1 = m_GeoModeTest[GeoMotionIdx].m_candIdx1;
+ double tempCost = m_GeoCostList.singleDistList[0][splitDir][mergeCand0].cost + m_GeoCostList.singleDistList[1][splitDir][mergeCand1].cost;
+ if( tempCost > bestWholeBlkCost )
+ {
+ continue;
+ }
+ tempCost = tempCost + (double)bitsCandTB * sqrtLambdaForFirstPass;
+ comboList.list.push_back(GeoMergeCombo(splitDir, mergeCand0, mergeCand1, tempCost));
+ }
+ }
+ if( comboList.list.empty() )
+ {
+ return;
+ }
+
+ comboList.sortByCost();
+
+ bool geocandHasNoResidual[GEO_MAX_TRY_WEIGHTED_SAD] = { false };
+ bool bestIsSkip = false;
+ int geoNumCobo = (int)comboList.list.size();
+ static_vector<uint8_t, GEO_MAX_TRY_WEIGHTED_SAD> geoRdModeList;
+ static_vector<double, GEO_MAX_TRY_WEIGHTED_SAD> geocandCostList;
+
+ DistParam distParamSAD2;
+ const bool useHadamard = !tempCS->slice->getDisableSATDForRD();
+ m_pcRdCost->setDistParam(distParamSAD2, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, useHadamard);
+ int geoNumMrgSATDCand = min(GEO_MAX_TRY_WEIGHTED_SATD, geoNumCobo);
+
+ for (uint8_t candidateIdx = 0; candidateIdx < min(geoNumCobo, GEO_MAX_TRY_WEIGHTED_SAD); candidateIdx++)
+ {
+ int splitDir = comboList.list[candidateIdx].splitDir;
+ int mergeCand0 = comboList.list[candidateIdx].mergeIdx0;
+ int mergeCand1 = comboList.list[candidateIdx].mergeIdx1;
+
+ geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_LUMA, geoCombinations[candidateIdx], geoBuffer[mergeCand0], geoBuffer[mergeCand1]);
+ distParamSAD2.cur = geoCombinations[candidateIdx].Y();
+ Distortion sad = distParamSAD2.distFunc(distParamSAD2);
+ int mvBits = 2;
+ mergeCand1 -= mergeCand1 < mergeCand0 ? 0 : 1;
+ mvBits += mergeCand0;
+ mvBits += mergeCand1;
+ double updateCost = (double)sad + (double)(bitsCandTB + mvBits) * sqrtLambdaForFirstPass;
+ comboList.list[candidateIdx].cost = updateCost;
+ updateCandList(candidateIdx, updateCost, geoRdModeList, geocandCostList, geoNumMrgSATDCand);
+ }
+ for (uint8_t i = 0; i < geoNumMrgSATDCand; i++)
+ {
+ if (geocandCostList[i] > MRG_FAST_RATIO * geocandCostList[0] || geocandCostList[i] > getMergeBestSATDCost() || geocandCostList[i] > getAFFBestSATDCost())
+ {
+ geoNumMrgSATDCand = i;
+ break;
+ }
+ }
+ for (uint8_t i = 0; i < geoNumMrgSATDCand && isChromaEnabled(pu.chromaFormat); i++)
+ {
+ uint8_t candidateIdx = geoRdModeList[i];
+ int splitDir = comboList.list[candidateIdx].splitDir;
+ int mergeCand0 = comboList.list[candidateIdx].mergeIdx0;
+ int mergeCand1 = comboList.list[candidateIdx].mergeIdx1;
+ geoCombinations[candidateIdx] = m_acGeoWeightedBuffer[candidateIdx].getBuf(localUnitArea);
+ m_pcInterSearch->weightedGeoBlk(pu, splitDir, CHANNEL_TYPE_CHROMA, geoCombinations[candidateIdx], geoBuffer[mergeCand0], geoBuffer[mergeCand1]);
+ }
+
+ m_bestModeUpdated = tempCS->useDbCost = bestCS->useDbCost = false;
+ tempCS->initStructData(encTestMode.qp);
+ uint8_t iteration;
+ uint8_t iterationBegin = 0;
+ iteration = 2;
+ for (uint8_t noResidualPass = iterationBegin; noResidualPass < iteration; ++noResidualPass)
+ {
+ for (uint8_t mrgHADIdx = 0; mrgHADIdx < geoNumMrgSATDCand; mrgHADIdx++)
+ {
+ uint8_t candidateIdx = geoRdModeList[mrgHADIdx];
+ if (((noResidualPass != 0) && geocandHasNoResidual[candidateIdx])
+ || ((noResidualPass == 0) && bestIsSkip))
+ {
+ continue;
+ }
+ CodingUnit &cu = tempCS->addCU(tempCS->area, pm.chType);
+ pm.setCUData(cu);
+ cu.predMode = MODE_INTER;
+ cu.slice = tempCS->slice;
+ cu.tileIdx = tempCS->pps->getTileIdx(tempCS->area.lumaPos());
+ cu.qp = encTestMode.qp;
+ cu.affine = false;
+ cu.mtsFlag = false;
+#if INTER_LIC
+ cu.LICFlag = false;
+#endif
+ cu.BcwIdx = BCW_DEFAULT;
+ cu.geoFlag = true;
+ cu.imv = 0;
+ cu.mmvdSkip = false;
+ cu.skip = false;
+ cu.mipFlag = false;
+#if JVET_V0130_INTRA_TMP
+ cu.tmpFlag = false;
+#endif
+ cu.bdpcmMode = 0;
+ PredictionUnit &pu = tempCS->addPU(cu, pm.chType);
+ pu.mergeFlag = true;
+ pu.regularMergeFlag = false;
+ pu.geoSplitDir = comboList.list[candidateIdx].splitDir;
+ pu.geoMergeIdx0 = comboList.list[candidateIdx].mergeIdx0;
+ pu.geoMergeIdx1 = comboList.list[candidateIdx].mergeIdx1;
+ pu.mmvdMergeFlag = false;
+ pu.mmvdMergeIdx = MAX_UCHAR;
+
+ PU::spanGeoMotionInfo(pu, mergeCtx, pu.geoSplitDir, pu.geoMergeIdx0, pu.geoMergeIdx1);
+ tempCS->getPredBuf().copyFrom(geoCombinations[candidateIdx]);
+#if ENABLE_OBMC
+ cu.isobmcMC = true;
+ cu.obmcFlag = true;
+ m_pcInterSearch->subBlockOBMC(pu);
+ cu.isobmcMC = false;
+#endif
+ xEncodeInterResidual(tempCS, bestCS, pm, encTestMode, noResidualPass, (noResidualPass == 0 ? &geocandHasNoResidual[candidateIdx] : NULL));
+
+ if (m_pcEncCfg->getUseFastDecisionForMerge() && !bestIsSkip)
+ {
+ bestIsSkip = bestCS->getCU(pm.chType)->rootCbf == 0;
+ }
+ tempCS->initStructData(encTestMode.qp);
+ }
+ }
+ if (m_bestModeUpdated && bestCS->cost != MAX_DOUBLE)
+ {
+ xCalDebCost(*bestCS, pm);
+ }
+}
+#endif
+#if MERGE_ENC_OPT
+void EncCu::xCheckSATDCostRegularMerge(CodingStructure *&tempCS, CodingUnit &cu, PredictionUnit &pu, MergeCtx mergeCtx, PelUnitBuf *acMergeTempBuffer[MMVD_MRG_MAX_RD_NUM], PelUnitBuf *&singleMergeTempBuffer, PelUnitBuf acMergeTmpBuffer[MRG_MAX_NUM_CANDS]
+#if !MULTI_PASS_DMVR
+ , Mv refinedMvdL0[MAX_NUM_PARTS_IN_CTU][MRG_MAX_NUM_CANDS]
+#endif
+ , unsigned& uiNumMrgSATDCand, static_vector<ModeInfo, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &RdModeList, static_vector<double, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &candCostList, DistParam distParam, const TempCtx &ctxStart
+#if MULTI_PASS_DMVR
+ , bool* applyBDMVR
+#endif
+)
+{
+#if INTER_LIC
+ cu.LICFlag = false;
+#endif
+ cu.mmvdSkip = false;
+ cu.geoFlag = false;
cu.affine = false;
cu.imv = 0;
@@ -4820,7 +6131,7 @@ void EncCu::xCheckSATDCostAffineMmvdMerge( CodingStructure*& tempCS,
pu.afMmvdFlag = false;
}
#endif
-
+#if !JVET_W0097_GPM_MMVD_TM
void EncCu::xCheckSATDCostGeoMerge(CodingStructure *&tempCS, CodingUnit &cu, PredictionUnit &pu, MergeCtx geoMergeCtx, PelUnitBuf *acMergeTempBuffer[MMVD_MRG_MAX_RD_NUM], PelUnitBuf *&singleMergeTempBuffer
, unsigned& uiNumMrgSATDCand, static_vector<ModeInfo, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &RdModeList, static_vector<double, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &candCostList, DistParam distParam, const TempCtx &ctxStart)
{
@@ -5028,6 +6339,7 @@ void EncCu::xCheckSATDCostGeoMerge(CodingStructure *&tempCS, CodingUnit &cu, Pre
uiNumMrgSATDCand = (unsigned int)RdModeList.size();
}
}
+#endif
#else
void EncCu::xCheckRDCostAffineMerge2Nx2N( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner, const EncTestMode& encTestMode )
{
@@ -5122,7 +6434,12 @@ void EncCu::xCheckRDCostAffineMerge2Nx2N( CodingStructure *&tempCS, CodingStruct
{
RdModeList.clear();
mrgTempBufSet = true;
+#if JVET_W0097_GPM_MMVD_TM
+ const double sqrtLambdaForFirstPassIntra = m_pcRdCost->getMotionLambda() * FRAC_BITS_SCALE;
+ const TempCtx ctxStart(m_CtxCache, m_CABACEstimator->getCtx());
+#else
const double sqrtLambdaForFirstPass = m_pcRdCost->getMotionLambda( );
+#endif
CodingUnit &cu = tempCS->addCU( tempCS->area, partitioner.chType );
@@ -5189,12 +6506,18 @@ void EncCu::xCheckRDCostAffineMerge2Nx2N( CodingStructure *&tempCS, CodingStruct
m_pcInterSearch->motionCompensation( pu, acMergeBuffer[uiMergeCand], REF_PIC_LIST_X, true, false );
Distortion uiSad = distParam.distFunc( distParam );
+#if JVET_W0097_GPM_MMVD_TM
+ m_CABACEstimator->getCtx() = ctxStart;
+ uint64_t fracBits = m_pcInterSearch->xCalcPuMeBits(pu);
+ double cost = (double)uiSad + (double)fracBits * sqrtLambdaForFirstPassIntra;
+#else
uint32_t uiBitsCand = uiMergeCand + 1;
if ( uiMergeCand == tempCS->picHeader->getMaxNumAffineMergeCand() - 1 )
{
uiBitsCand--;
}
double cost = (double)uiSad + (double)uiBitsCand * sqrtLambdaForFirstPass;
+#endif
#if MULTI_HYP_PRED
if (testMHP && pu.addHypData.size() < tempCS->sps->getMaxNumAddHyps())
{
@@ -5225,7 +6548,9 @@ void EncCu::xCheckRDCostAffineMerge2Nx2N( CodingStructure *&tempCS, CodingStruct
tempCS->initStructData( encTestMode.qp );
setAFFBestSATDCost(candCostList[0]);
-
+#if JVET_W0097_GPM_MMVD_TM
+ m_CABACEstimator->getCtx() = ctxStart;
+#endif
}
else
{
diff --git a/source/Lib/EncoderLib/EncCu.h b/source/Lib/EncoderLib/EncCu.h
index 0f30ed338..46c3a0410 100644
--- a/source/Lib/EncoderLib/EncCu.h
+++ b/source/Lib/EncoderLib/EncCu.h
@@ -148,7 +148,61 @@ public:
}
int numGeoTemplatesInitialized;
};
+#if JVET_W0097_GPM_MMVD_TM
+struct SingleGeoMMVDMergeEntry
+{
+ int mergeIdx;
+ int mmvdIdx; // 0 - mmvd OFF; 1 - mmvdIdx = 0; 2 - mmvdIdx = 1; 3 - mmvdIdx = 2; ...
+ double cost;
+ SingleGeoMMVDMergeEntry() : mergeIdx(0), mmvdIdx(0), cost(MAX_DOUBLE) {};
+ SingleGeoMMVDMergeEntry(int _mergeIdx, int _mmvdIdx, double _cost) : mergeIdx(_mergeIdx), mmvdIdx(_mmvdIdx), cost(_cost) {};
+};
+class FastGeoMMVDCostList
+{
+public:
+ FastGeoMMVDCostList()
+ {
+ for (int partIdx = 0; partIdx < 2; partIdx++)
+ {
+ singleDistList[partIdx] = new SingleGeoMMVDMergeEntry**[GEO_NUM_PARTITION_MODE];
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ singleDistList[partIdx][splitDir] = new SingleGeoMMVDMergeEntry*[MRG_MAX_NUM_CANDS];
+ for (int candIdx = 0; candIdx < MRG_MAX_NUM_CANDS; candIdx++)
+ {
+#if JVET_W0097_GPM_MMVD_TM && TM_MRG
+ singleDistList[partIdx][splitDir][candIdx] = new SingleGeoMMVDMergeEntry[GPM_EXT_MMVD_MAX_REFINE_NUM + 2];
+#else
+ singleDistList[partIdx][splitDir][candIdx] = new SingleGeoMMVDMergeEntry[GPM_EXT_MMVD_MAX_REFINE_NUM + 1];
+#endif
+ }
+ }
+ }
+ }
+ ~FastGeoMMVDCostList()
+ {
+ for (int partIdx = 0; partIdx < 2; partIdx++)
+ {
+ for (int splitDir = 0; splitDir < GEO_NUM_PARTITION_MODE; splitDir++)
+ {
+ for (int candIdx = 0; candIdx < MRG_MAX_NUM_CANDS; candIdx++)
+ {
+ delete[] singleDistList[partIdx][splitDir][candIdx];
+ }
+ delete[] singleDistList[partIdx][splitDir];
+ }
+ delete[] singleDistList[partIdx];
+ singleDistList[partIdx] = nullptr;
+ }
+ }
+ SingleGeoMMVDMergeEntry*** singleDistList[2];
+ void insert(int geoIdx, int partIdx, int mergeIdx, int mmvdIdx, double cost)
+ {
+ singleDistList[partIdx][geoIdx][mergeIdx][mmvdIdx] = SingleGeoMMVDMergeEntry(mergeIdx, mmvdIdx, cost);
+ }
+};
+#endif
class EncCu
: DecCu
{
@@ -212,6 +266,20 @@ private:
PelStorage m_acGeoWeightedBuffer[GEO_MAX_TRY_WEIGHTED_SAD]; // to store weighted prediction pixels
FastGeoCostList m_GeoCostList;
+#if JVET_W0097_GPM_MMVD_TM
+ PelStorage m_acGeoMMVDBuffer[MRG_MAX_NUM_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
+ PelStorage m_acGeoMMVDTmpBuffer[MRG_MAX_NUM_CANDS][GPM_EXT_MMVD_MAX_REFINE_NUM];
+ FastGeoMMVDCostList m_GeoMMVDCostList;
+ bool fastGpmMmvdSearch;
+ bool fastGpmMmvdRelatedCU;
+ bool includeMoreMMVDCandFirstPass;
+ int maxNumGPMDirFirstPass;
+ int numCandPerPar;
+#if TM_MRG
+ PelStorage m_acGeoMergeTmpBuffer[GEO_TM_MAX_NUM_CANDS];
+ PelStorage m_acGeoSADTmpBuffer[GEO_TM_MAX_NUM_CANDS];
+#endif
+#endif
double m_AFFBestSATDCost;
double m_mergeBestSATDCost;
MotionInfo m_SubPuMiBuf [( MAX_CU_SIZE * MAX_CU_SIZE ) >> ( MIN_CU_LOG2 << 1 )];
@@ -240,6 +308,10 @@ private:
const bool updateRdCostLambda );
#endif
double m_sbtCostSave[2];
+#if JVET_W0097_GPM_MMVD_TM
+ MergeCtx m_mergeCand;
+ bool m_mergeCandAvail;
+#endif
public:
/// copy parameters from encoder class
void init ( EncLib* pcEncLib, const SPS& sps PARL_PARAM( const int jId = 0 ) );
@@ -333,9 +405,11 @@ protected:
#endif
);
#endif
+#if !JVET_W0097_GPM_MMVD_TM
void xCheckSATDCostGeoMerge
( CodingStructure *&tempCS, CodingUnit &cu, PredictionUnit &pu, MergeCtx geoMergeCtx, PelUnitBuf *acMergeTempBuffer[MMVD_MRG_MAX_RD_NUM], PelUnitBuf *&singleMergeTempBuffer
, unsigned& uiNumMrgSATDCand, static_vector<ModeInfo, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &RdModeList, static_vector<double, MRG_MAX_NUM_CANDS + MMVD_ADD_NUM> &candCostList, DistParam distParam, const TempCtx &ctxStart);
+#endif
#else
void xCheckRDCostAffineMerge2Nx2N
( CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &partitioner, const EncTestMode& encTestMode );
@@ -360,7 +434,11 @@ protected:
#if ENABLE_OBMC
void xCheckRDCostInterWoOBMC(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode);
#endif
+#if JVET_W0097_GPM_MMVD_TM
+ void xCheckRDCostMergeGeoComb2Nx2N(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode, bool isSecondPass = false);
+#else
void xCheckRDCostMergeGeo2Nx2N(CodingStructure *&tempCS, CodingStructure *&bestCS, Partitioner &pm, const EncTestMode& encTestMode);
+#endif
void xEncodeInterResidual( CodingStructure *&tempCS
, CodingStructure *&bestCS
, Partitioner &partitioner
diff --git a/source/Lib/EncoderLib/EncLib.cpp b/source/Lib/EncoderLib/EncLib.cpp
index 16965b1ac..55baaafb0 100644
--- a/source/Lib/EncoderLib/EncLib.cpp
+++ b/source/Lib/EncoderLib/EncLib.cpp
@@ -2133,6 +2133,30 @@ void EncLib::xInitPicHeader(PicHeader &picHeader, const SPS &sps, const PPS &pps
picHeader.setDisBdofFlag(false);
picHeader.setDisDmvrFlag(false);
picHeader.setDisProfFlag(false);
+#if JVET_W0097_GPM_MMVD_TM
+ if (sps.getUseGeo())
+ {
+#if TOOLS
+ if (getIntraPeriod() > 0)
+ {
+ if ((getSourceWidth() * getSourceHeight()) > (1920 * 1080))
+ {
+ picHeader.setGPMMMVDTableFlag(false);
+ }
+ else
+ {
+ picHeader.setGPMMMVDTableFlag(true);
+ }
+ }
+ else
+ {
+ picHeader.setGPMMMVDTableFlag(true);
+ }
+#else
+ picHeader.setGPMMMVDTableFlag(false);
+#endif
+ }
+#endif
}
void EncLib::xInitAPS(APS &aps)
diff --git a/source/Lib/EncoderLib/EncModeCtrl.cpp b/source/Lib/EncoderLib/EncModeCtrl.cpp
index 007b25ea1..dc656f338 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.cpp
+++ b/source/Lib/EncoderLib/EncModeCtrl.cpp
@@ -1251,6 +1251,9 @@ void EncModeCtrlMTnoRQT::initCULevel( Partitioner &partitioner, const CodingStru
cuECtx.set( DO_TRIV_SPLIT, 1 );
cuECtx.set( BEST_IMV_COST, MAX_DOUBLE * .5 );
cuECtx.set( BEST_NO_IMV_COST, MAX_DOUBLE * .5 );
+#if JVET_W0097_GPM_MMVD_TM
+ cuECtx.set( BEST_GPM_COST, MAX_DOUBLE * .5);
+#endif
cuECtx.set( QT_BEFORE_BT, qtBeforeBt );
cuECtx.set( DID_QUAD_SPLIT, false );
cuECtx.set( IS_BEST_NOSPLIT_SKIP, false );
@@ -1892,6 +1895,12 @@ bool EncModeCtrlMTnoRQT::tryMode( const EncTestMode& encTestmode, const CodingSt
{
return false;
}
+#if JVET_W0097_GPM_MMVD_TM
+ if (encTestmode.type == ETM_MERGE_GEO && relatedCU.skipGPM)
+ {
+ return false;
+ }
+#endif
return true;
}
else if( isModeSplit( encTestmode ) )
@@ -2131,6 +2140,27 @@ bool EncModeCtrlMTnoRQT::tryMode( const EncTestMode& encTestmode, const CodingSt
// assume the non-split modes are done and set the marks for the best found mode
if( bestCS && bestCU )
{
+#if JVET_W0097_GPM_MMVD_TM
+ if (!slice.isIntra())
+ {
+ double gpmCost = cuECtx.get<double>(BEST_GPM_COST);
+ if (gpmCost != (MAX_DOUBLE * .5))
+ {
+ if (gpmCost > (bestCS->cost * 1.2))
+ {
+ relatedCU.skipGPM = true;
+ }
+ else if (gpmCost > bestCS->cost && bestCU->skip)
+ {
+ relatedCU.skipGPM = true;
+ }
+ else if (gpmCost > bestCS->cost && (!cuECtx.bestTU->cbf[0] && !cuECtx.bestTU->cbf[1] && !cuECtx.bestTU->cbf[2]))
+ {
+ relatedCU.skipGPM = true;
+ }
+ }
+ }
+#endif
if( CU::isInter( *bestCU ) )
{
relatedCU.isInter = true;
@@ -2321,7 +2351,15 @@ bool EncModeCtrlMTnoRQT::useModeResult( const EncTestMode& encTestmode, CodingSt
}
}
}
-
+#if JVET_W0097_GPM_MMVD_TM
+ if (encTestmode.type == ETM_MERGE_GEO)
+ {
+ if (tempCS->cost < cuECtx.get<double>(BEST_GPM_COST))
+ {
+ cuECtx.set(BEST_GPM_COST, tempCS->cost);
+ }
+ }
+#endif
if( encTestmode.type == ETM_SPLIT_QT )
{
int maxQtD = 0;
diff --git a/source/Lib/EncoderLib/EncModeCtrl.h b/source/Lib/EncoderLib/EncModeCtrl.h
index 528ddb8da..ee90073a6 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.h
+++ b/source/Lib/EncoderLib/EncModeCtrl.h
@@ -462,6 +462,14 @@ struct CodedCUInfo
bool isSkip;
bool isMMVDSkip;
bool isIBC;
+#if JVET_W0097_GPM_MMVD_TM
+ bool skipGPM;
+ char isGPMTested;
+ int geoDirCandList[GEO_MAX_TRY_WEIGHTED_SATD];
+ int numGeoDirCand;
+ int geoMrgIdx0List[GEO_MAX_TRY_WEIGHTED_SATD];
+ int geoMrgIdx1List[GEO_MAX_TRY_WEIGHTED_SATD];
+#endif
bool validMv[NUM_REF_PIC_LIST_01][MAX_STORED_CU_INFO_REFS];
Mv saveMv [NUM_REF_PIC_LIST_01][MAX_STORED_CU_INFO_REFS];
@@ -512,7 +520,9 @@ public:
protected:
void touch ( const UnitArea& area );
#endif
-
+#if JVET_W0097_GPM_MMVD_TM
+public:
+#endif
CodedCUInfo& getBlkInfo( const UnitArea& area );
public:
@@ -609,13 +619,38 @@ public:
// EncModeCtrlMTnoRQT - allows and controls modes introduced by QTBT (inkl. multi-type-tree)
// - only 2Nx2N, no RQT, additional binary/triary CU splits
//////////////////////////////////////////////////////////////////////////
-
+#if JVET_W0097_GPM_MMVD_TM
+enum ExtraFeatures
+{
+ DID_HORZ_SPLIT = 0,
+ DID_VERT_SPLIT,
+ DID_QUAD_SPLIT,
+ BEST_HORZ_SPLIT_COST,
+ BEST_VERT_SPLIT_COST,
+ BEST_TRIH_SPLIT_COST,
+ BEST_TRIV_SPLIT_COST,
+ DO_TRIH_SPLIT,
+ DO_TRIV_SPLIT,
+ BEST_NON_SPLIT_COST,
+ BEST_NO_IMV_COST,
+ BEST_IMV_COST,
+ BEST_GPM_COST,
+ QT_BEFORE_BT,
+ IS_BEST_NOSPLIT_SKIP,
+ MAX_QT_SUB_DEPTH,
+#if REUSE_CU_RESULTS
+ IS_REUSING_CU,
+#endif
+ NUM_EXTRA_FEATURES
+};
+#endif
class EncModeCtrlMTnoRQT : public EncModeCtrl, public CacheBlkInfoCtrl
#if REUSE_CU_RESULTS
, public BestEncInfoCache
#endif
, public SaveLoadEncInfoSbt
{
+#if !JVET_W0097_GPM_MMVD_TM
enum ExtraFeatures
{
DID_HORZ_SPLIT = 0,
@@ -638,7 +673,7 @@ class EncModeCtrlMTnoRQT : public EncModeCtrl, public CacheBlkInfoCtrl
#endif
NUM_EXTRA_FEATURES
};
-
+#endif
unsigned m_skipThreshold;
public:
diff --git a/source/Lib/EncoderLib/EncReshape.cpp b/source/Lib/EncoderLib/EncReshape.cpp
index 1af97f01b..4bfa516f9 100644
--- a/source/Lib/EncoderLib/EncReshape.cpp
+++ b/source/Lib/EncoderLib/EncReshape.cpp
@@ -104,7 +104,6 @@ void EncReshape::createEnc(int picWidth, int picHeight, uint32_t maxCUWidth, ui
{
m_chromaAdjHelpLUT.resize(PIC_CODE_CW_BINS, 1<<CSCALE_FP_PREC);
}
-
m_sliceReshapeInfo.setUseSliceReshaper(true);
m_sliceReshapeInfo.setSliceReshapeChromaAdj(true);
m_sliceReshapeInfo.setSliceReshapeModelPresentFlag(true);
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index 2a7145c80..c4ce53564 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -3414,6 +3414,9 @@ void InterSearch::predInterSearchAdditionalHypothesis(PredictionUnit& pu, const
{
PU::getGeoMergeCandidates(fakePredData, m_geoMrgCtx);
}
+#if JVET_W0097_GPM_MMVD_TM
+ maxNumMergeCandidates = min((int)maxNumMergeCandidates, m_geoMrgCtx.numValidMergeCand);
+#endif
const auto savedAffine = pu.cu->affine;
const auto savedIMV = pu.cu->imv;
for (int i = 0; i < maxNumMergeCandidates; i++)
@@ -3443,6 +3446,12 @@ void InterSearch::predInterSearchAdditionalHypothesis(PredictionUnit& pu, const
}
setGeoTmpBuffer();
}
+#if JVET_W0097_GPM_MMVD_TM
+ else
+ {
+ maxNumMergeCandidates = min((int)maxNumMergeCandidates, m_geoMrgCtx.numValidMergeCand);
+ }
+#endif
for (int i = 0; i < maxNumMergeCandidates; i++)
{
int refList = m_geoMrgCtx.interDirNeighbours[i] - 1; CHECK(refList != 0 && refList != 1, "");
diff --git a/source/Lib/EncoderLib/VLCWriter.cpp b/source/Lib/EncoderLib/VLCWriter.cpp
index 19aeb75d3..cb4c13e9f 100644
--- a/source/Lib/EncoderLib/VLCWriter.cpp
+++ b/source/Lib/EncoderLib/VLCWriter.cpp
@@ -2265,6 +2265,12 @@ void HLSWriter::codePictureHeader( PicHeader* picHeader, bool writeRbspTrailingB
{
picHeader->setDisFracMMVD(false);
}
+#if JVET_W0097_GPM_MMVD_TM
+ if (sps->getUseGeo() && (!pps->getRplInfoInPhFlag() || picHeader->getRPL(1)->getNumRefEntries() > 0))
+ {
+ WRITE_FLAG(picHeader->getGPMMMVDTableFlag(), "ph_gpm_ext_mmvd_flag");
+ }
+#endif
#if JVET_R0324_REORDER
// mvd L1 zero flag
if (!pps->getRplInfoInPhFlag() || picHeader->getRPL(1)->getNumRefEntries() > 0)
--
GitLab