From 003824b2e90d9249756bdbc805fe4892da25da9a Mon Sep 17 00:00:00 2001
From: Xiaoyu Xiu <xiaoyuxiu@kwai.com>
Date: Wed, 3 Apr 2019 19:13:54 +0200
Subject: [PATCH] JVET-N0329: IBC Search Improvement
---
source/Lib/CommonLib/IbcHashMap.cpp | 99 +++++++++++
source/Lib/CommonLib/IbcHashMap.h | 4 +
source/Lib/CommonLib/MotionInfo.h | 7 +
source/Lib/CommonLib/RdCost.h | 4 +
source/Lib/CommonLib/Slice.h | 8 +-
source/Lib/CommonLib/TypeDef.h | 2 +
source/Lib/CommonLib/UnitTools.cpp | 113 ++++++++++++
source/Lib/CommonLib/UnitTools.h | 5 +
source/Lib/EncoderLib/EncCu.cpp | 16 ++
source/Lib/EncoderLib/EncModeCtrl.cpp | 4 +
source/Lib/EncoderLib/EncSlice.cpp | 12 ++
source/Lib/EncoderLib/InterSearch.cpp | 242 +++++++++++++++++++++++++-
source/Lib/EncoderLib/InterSearch.h | 20 +++
source/Lib/EncoderLib/IntraSearch.cpp | 8 +
14 files changed, 539 insertions(+), 5 deletions(-)
diff --git a/source/Lib/CommonLib/IbcHashMap.cpp b/source/Lib/CommonLib/IbcHashMap.cpp
index 9d876292de..0b2cce6eff 100644
--- a/source/Lib/CommonLib/IbcHashMap.cpp
+++ b/source/Lib/CommonLib/IbcHashMap.cpp
@@ -274,6 +274,9 @@ bool IbcHashMap::ibcHashMatch(const Area& lumaArea, std::vector<Position>& cand,
// find the block with least candidates
size_t minSize = MAX_UINT;
unsigned int targetHashOneBlock = 0;
+#if JVET_N0329_IBC_SEARCH_IMP
+ Position targetBlockOffsetInCu(0, 0);
+#endif
for (SizeType y = 0; y < lumaArea.height && minSize > 1; y += MIN_PU_SIZE)
{
for (SizeType x = 0; x < lumaArea.width && minSize > 1; x += MIN_PU_SIZE)
@@ -283,6 +286,9 @@ bool IbcHashMap::ibcHashMatch(const Area& lumaArea, std::vector<Position>& cand,
{
minSize = m_hash2Pos[hash].size();
targetHashOneBlock = hash;
+#if JVET_N0329_IBC_SEARCH_IMP
+ targetBlockOffsetInCu.repositionTo(Position(x, y));
+#endif
}
}
}
@@ -294,11 +300,20 @@ bool IbcHashMap::ibcHashMatch(const Area& lumaArea, std::vector<Position>& cand,
// check whether whole block match
for (std::vector<Position>::iterator refBlockPos = candOneBlock.begin(); refBlockPos != candOneBlock.end(); refBlockPos++)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ Position topLeft = refBlockPos->offset(-targetBlockOffsetInCu.x, -targetBlockOffsetInCu.y);
+ Position bottomRight = topLeft.offset(lumaArea.width - 1, lumaArea.height - 1);
+#else
Position bottomRight = refBlockPos->offset(lumaArea.width - 1, lumaArea.height - 1);
+#endif
bool wholeBlockMatch = true;
if (lumaArea.width > MIN_PU_SIZE || lumaArea.height > MIN_PU_SIZE)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (!cs.isDecomp(bottomRight, cs.chType) || bottomRight.x >= m_picWidth || bottomRight.y >= m_picHeight || topLeft.x < 0 || topLeft.y < 0)
+#else
if (!cs.isDecomp(bottomRight, cs.chType) || bottomRight.x >= m_picWidth || bottomRight.y >= m_picHeight)
+#endif
{
continue;
}
@@ -307,20 +322,33 @@ bool IbcHashMap::ibcHashMatch(const Area& lumaArea, std::vector<Position>& cand,
for (SizeType x = 0; x < lumaArea.width && wholeBlockMatch; x += MIN_PU_SIZE)
{
// whether the reference block and current block has the same hash
+#if JVET_N0329_IBC_SEARCH_IMP
+ wholeBlockMatch &= (m_pos2Hash[lumaArea.pos().y + y][lumaArea.pos().x + x] == m_pos2Hash[topLeft.y + y][topLeft.x + x]);
+#else
wholeBlockMatch &= (m_pos2Hash[lumaArea.pos().y + y][lumaArea.pos().x + x] == m_pos2Hash[refBlockPos->y + y][refBlockPos->x + x]);
+#endif
}
}
}
else
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ CHECK(topLeft != *refBlockPos, "4x4 target block should not have offset!");
+ if (abs(topLeft.x - lumaArea.x) > searchRange4SmallBlk || abs(topLeft.y - lumaArea.y) > searchRange4SmallBlk || !cs.isDecomp(bottomRight, cs.chType))
+#else
if (abs(refBlockPos->x - lumaArea.x) > searchRange4SmallBlk || abs(refBlockPos->y - lumaArea.y) > searchRange4SmallBlk || !cs.isDecomp(bottomRight, cs.chType))
+#endif
{
continue;
}
}
if (wholeBlockMatch)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ cand.push_back(topLeft);
+#else
cand.push_back(*refBlockPos);
+#endif
if (cand.size() > maxCand)
{
break;
@@ -349,5 +377,76 @@ int IbcHashMap::getHashHitRatio(const Area& lumaArea)
return 100 * hit / total;
}
+#if JVET_N0329_IBC_SEARCH_IMP
+int IbcHashMap::calHashBlkMatchPerc(const Area& lumaArea)
+{
+ int maxX = std::min((int)(lumaArea.x + lumaArea.width), m_picWidth);
+ int maxY = std::min((int)(lumaArea.y + lumaArea.height), m_picHeight);
+ int maxUsage[100];
+ unsigned int mostSelHash[100];
+
+ static int numExcludedHashValue = 36;
+
+ for (int i = 0; i < numExcludedHashValue; i++)
+ {
+ maxUsage[i] = 0;
+ mostSelHash[i] = 0;
+ }
+
+ for (std::unordered_map<unsigned int, std::vector<Position>>::iterator it = m_hash2Pos.begin(); it != m_hash2Pos.end(); ++it)
+ {
+ unsigned int hash = it->first;
+ int usage = (int)it->second.size();
+ assert(usage == m_hash2Pos[hash].size());
+
+ int insertPos = -1;
+ for (insertPos = 0; insertPos < numExcludedHashValue; insertPos++)
+ {
+ if (usage > maxUsage[insertPos])
+ {
+ break;
+ }
+ }
+ assert(insertPos <= numExcludedHashValue);
+
+ if (insertPos < numExcludedHashValue)
+ {
+ for (int i = (numExcludedHashValue - 1); i >= (insertPos + 1); i--)
+ {
+ maxUsage[i] = maxUsage[i - 1];
+ mostSelHash[i] = mostSelHash[i - 1];
+ }
+ maxUsage[insertPos] = usage;
+ mostSelHash[insertPos] = hash;
+ }
+ }
+
+ int hit = 0, total = 0;
+ for (int y = lumaArea.y; y < maxY; y += MIN_PU_SIZE)
+ {
+ for (int x = lumaArea.x; x < maxX; x += MIN_PU_SIZE)
+ {
+ unsigned int hash = m_pos2Hash[y][x];
+
+ bool excludedHash = false;
+ for (int i = 0; i < numExcludedHashValue && !excludedHash; i++)
+ {
+ if (hash == mostSelHash[i])
+ {
+ excludedHash = true;
+ }
+ }
+
+ if (excludedHash)
+ {
+ continue;
+ }
+ hit += (m_hash2Pos[hash].size() > 1);
+ total++;
+ }
+ }
+ return 100 * hit / total;
+}
+#endif
//! \}
diff --git a/source/Lib/CommonLib/IbcHashMap.h b/source/Lib/CommonLib/IbcHashMap.h
index e343aab6e0..46074a4b8c 100644
--- a/source/Lib/CommonLib/IbcHashMap.h
+++ b/source/Lib/CommonLib/IbcHashMap.h
@@ -82,6 +82,10 @@ public:
bool ibcHashMatch(const Area& lumaArea, std::vector<Position>& cand, const CodingStructure& cs, const int maxCand, const int searchRange4SmallBlk);
int getHashHitRatio(const Area& lumaArea);
+#if JVET_N0329_IBC_SEARCH_IMP
+ int calHashBlkMatchPerc(const Area& lumaArea);
+#endif
+
#ifdef TARGET_SIMD_X86
void initIbcHashMapX86();
template <X86_VEXT vext>
diff --git a/source/Lib/CommonLib/MotionInfo.h b/source/Lib/CommonLib/MotionInfo.h
index 020323575f..609fd14adf 100644
--- a/source/Lib/CommonLib/MotionInfo.h
+++ b/source/Lib/CommonLib/MotionInfo.h
@@ -224,4 +224,11 @@ struct LutMotionCand
static_vector<MotionInfo, MAX_NUM_HMVP_CANDS> lutShare;
static_vector<MotionInfo, MAX_NUM_HMVP_CANDS> lutShareIbc;
};
+#if JVET_N0329_IBC_SEARCH_IMP
+struct PatentBvCand
+{
+ Mv m_bvCands[IBC_NUM_CANDIDATES];
+ int currCnt;
+};
+#endif
#endif // __MOTIONINFO__
diff --git a/source/Lib/CommonLib/RdCost.h b/source/Lib/CommonLib/RdCost.h
index 64d2e64595..22e5b1b684 100644
--- a/source/Lib/CommonLib/RdCost.h
+++ b/source/Lib/CommonLib/RdCost.h
@@ -179,7 +179,11 @@ public:
inline Distortion getBvCostMultiplePreds(int x, int y, bool useIMV)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ return Distortion(m_dCost * getBitsMultiplePreds(x, y, useIMV));
+#else
return Distortion((m_dCost * getBitsMultiplePreds(x, y, useIMV)) / 65536.0);
+#endif
}
unsigned int getBitsMultiplePreds(int x, int y, bool useIMV)
diff --git a/source/Lib/CommonLib/Slice.h b/source/Lib/CommonLib/Slice.h
index de5979d8e6..f68f0f8aa9 100644
--- a/source/Lib/CommonLib/Slice.h
+++ b/source/Lib/CommonLib/Slice.h
@@ -1828,6 +1828,9 @@ private:
int m_apsId;
APS* m_aps;
bool m_tileGroupAlfEnabledFlag;
+#if JVET_N0329_IBC_SEARCH_IMP
+ bool m_disableSATDForRd;
+#endif
SliceReshapeInfo m_sliceReshapeInfo;
public:
Slice();
@@ -2114,7 +2117,10 @@ public:
bool getTileGroupAlfEnabledFlag() const { return m_tileGroupAlfEnabledFlag; }
void setTileGroupAlfEnabledFlag(bool b) { m_tileGroupAlfEnabledFlag = b; }
-
+#if JVET_N0329_IBC_SEARCH_IMP
+ void setDisableSATDForRD(bool b) { m_disableSATDForRd = b; }
+ bool getDisableSATDForRD() { return m_disableSATDForRd; }
+#endif
const SliceReshapeInfo& getReshapeInfo() const { return m_sliceReshapeInfo; }
SliceReshapeInfo& getReshapeInfo() { return m_sliceReshapeInfo; }
protected:
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index bd385f954b..a387b883c0 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -58,6 +58,8 @@
#define JVET_N0242_NON_LINEAR_ALF 1 // enable CE5-3.2, Non-linear ALF based on clipping function
+#define JVET_N0329_IBC_SEARCH_IMP 1 // IBC encoder-side improvement
+
#define JVET_N0325_BDOF 1 // unified right-shifts for BDOF derivation
#define JVET_N0449_MMVD_SIMP 1 // Configurable number of mmvd distance entries used
diff --git a/source/Lib/CommonLib/UnitTools.cpp b/source/Lib/CommonLib/UnitTools.cpp
index f1915892a3..4ecabb63e7 100644
--- a/source/Lib/CommonLib/UnitTools.cpp
+++ b/source/Lib/CommonLib/UnitTools.cpp
@@ -2186,9 +2186,121 @@ bool PU::isDiffMER(const PredictionUnit &pu1, const PredictionUnit &pu2)
return false;
}
+
+#if JVET_N0329_IBC_SEARCH_IMP
+bool PU::isAddNeighborMv(const Mv& currMv, Mv* neighborMvs, int numNeighborMv)
+{
+ bool existed = false;
+ for (uint32_t cand = 0; cand < numNeighborMv && !existed; cand++)
+ {
+ if (currMv == neighborMvs[cand])
+ {
+ existed = true;
+ }
+ }
+
+ if (!existed)
+ {
+ return true;
+ }
+ else
+ {
+ return false;
+ }
+}
+#endif
+
+#if JVET_N0329_IBC_SEARCH_IMP
+void PU::getIbcMVPsEncOnly(PredictionUnit &pu, Mv* mvPred, int& nbPred)
+#else
void PU::getIbcMVPsEncOnly(PredictionUnit &pu, Mv* MvPred, int& nbPred)
+#endif
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ const PreCalcValues &pcv = *pu.cs->pcv;
+ const int cuWidth = pu.blocks[COMPONENT_Y].width;
+ const int cuHeight = pu.blocks[COMPONENT_Y].height;
+ const int log2UnitWidth = g_aucLog2[pcv.minCUWidth];
+ const int log2UnitHeight = g_aucLog2[pcv.minCUHeight];
+ const int totalAboveUnits = (cuWidth >> log2UnitWidth) + 1;
+ const int totalLeftUnits = (cuHeight >> log2UnitHeight) + 1;
+
+ nbPred = 0;
+ Position posLT = pu.Y().topLeft();
+
+ // above-left
+ const PredictionUnit *aboveLeftPU = pu.cs->getPURestricted(posLT.offset(-1, -1), pu, pu.cs->chType);
+ if (aboveLeftPU && CU::isIBC(*aboveLeftPU->cu))
+ {
+ if (isAddNeighborMv(aboveLeftPU->bv, mvPred, nbPred))
+ {
+ mvPred[nbPred++] = aboveLeftPU->bv;
+ }
+ }
+
+ // above neighbors
+ for (uint32_t dx = 0; dx < totalAboveUnits && nbPred < IBC_NUM_CANDIDATES; dx++)
+ {
+ const PredictionUnit* tmpPU = pu.cs->getPURestricted(posLT.offset((dx << log2UnitWidth), -1), pu, pu.cs->chType);
+ if (tmpPU && CU::isIBC(*tmpPU->cu))
+ {
+ if (isAddNeighborMv(tmpPU->bv, mvPred, nbPred))
+ {
+ mvPred[nbPred++] = tmpPU->bv;
+ }
+ }
+ }
+
+ // left neighbors
+ for (uint32_t dy = 0; dy < totalLeftUnits && nbPred < IBC_NUM_CANDIDATES; dy++)
+ {
+ const PredictionUnit* tmpPU = pu.cs->getPURestricted(posLT.offset(-1, (dy << log2UnitHeight)), pu, pu.cs->chType);
+ if (tmpPU && CU::isIBC(*tmpPU->cu))
+ {
+ if (isAddNeighborMv(tmpPU->bv, mvPred, nbPred))
+ {
+ mvPred[nbPred++] = tmpPU->bv;
+ }
+ }
+ }
+
+ size_t numAvaiCandInLUT = pu.cs->motionLut.lutIbc.size();
+ for (uint32_t cand = 0; cand < numAvaiCandInLUT && nbPred < IBC_NUM_CANDIDATES; cand++)
+ {
+ MotionInfo neibMi = pu.cs->motionLut.lutIbc[cand];
+ if (isAddNeighborMv(neibMi.bv, mvPred, nbPred))
+ {
+ mvPred[nbPred++] = neibMi.bv;
+ }
+ }
+ bool isBvCandDerived[IBC_NUM_CANDIDATES];
+ ::memset(isBvCandDerived, false, IBC_NUM_CANDIDATES);
+
+ int curNbPred = nbPred;
+ if (curNbPred < IBC_NUM_CANDIDATES)
+ {
+ do
+ {
+ curNbPred = nbPred;
+ for (uint32_t idx = 0; idx < curNbPred && nbPred < IBC_NUM_CANDIDATES; idx++)
+ {
+ if (!isBvCandDerived[idx])
+ {
+ Mv derivedBv;
+ if (getDerivedBV(pu, mvPred[idx], derivedBv))
+ {
+ if (isAddNeighborMv(derivedBv, mvPred, nbPred))
+ {
+ mvPred[nbPred++] = derivedBv;
+ }
+ }
+ isBvCandDerived[idx] = true;
+ }
+ }
+ } while (nbPred > curNbPred && nbPred < IBC_NUM_CANDIDATES);
+ }
+#else
//-- Get Spatial MV
Position posLT = pu.Y().topLeft();
Position posRT = pu.Y().topRight();
@@ -2257,6 +2369,7 @@ void PU::getIbcMVPsEncOnly(PredictionUnit &pu, Mv* MvPred, int& nbPred)
if (getDerivedBV(pu, MvPred[nbPred - 1], MvPred[nbPred]))
nbPred++;
}
+#endif
}
bool PU::getDerivedBV(PredictionUnit &pu, const Mv& currentMv, Mv& derivedMv)
diff --git a/source/Lib/CommonLib/UnitTools.h b/source/Lib/CommonLib/UnitTools.h
index 0cb4a09d9b..45b749c7a2 100644
--- a/source/Lib/CommonLib/UnitTools.h
+++ b/source/Lib/CommonLib/UnitTools.h
@@ -184,7 +184,12 @@ namespace PU
bool isUniqueTriangleCandidates (const PredictionUnit &pu, MergeCtx &triangleMrgCtx);
void spanTriangleMotionInfo ( PredictionUnit &pu, MergeCtx &triangleMrgCtx, const bool splitDir, const uint8_t candIdx0, const uint8_t candIdx1);
int32_t mappingRefPic (const PredictionUnit &pu, int32_t refPicPoc, bool targetRefPicList);
+#if JVET_N0329_IBC_SEARCH_IMP
+ bool isAddNeighborMv (const Mv& currMv, Mv* neighborMvs, int numNeighborMv);
+ void getIbcMVPsEncOnly(PredictionUnit &pu, Mv* mvPred, int& nbPred);
+#else
void getIbcMVPsEncOnly(PredictionUnit &pu, Mv* MvPred, int& nbPred);
+#endif
bool getDerivedBV(PredictionUnit &pu, const Mv& currentMv, Mv& derivedMv);
bool isBlockVectorValid(PredictionUnit& pu, int xPos, int yPos, int width, int height, int picWidth, int picHeight, int xStartInCU, int yStartInCU, int xBv, int yBv, int ctuSize);
bool checkDMVRCondition(const PredictionUnit& pu);
diff --git a/source/Lib/EncoderLib/EncCu.cpp b/source/Lib/EncoderLib/EncCu.cpp
index dda140f085..faefbe832d 100644
--- a/source/Lib/EncoderLib/EncCu.cpp
+++ b/source/Lib/EncoderLib/EncCu.cpp
@@ -1803,7 +1803,11 @@ void EncCu::xCheckRDCostMerge2Nx2N( CodingStructure *&tempCS, CodingStructure *&
PredictionUnit &pu = tempCS->addPU( cu, partitioner.chType );
DistParam distParam;
+#if JVET_N0329_IBC_SEARCH_IMP
+ const bool bUseHadamard = !encTestMode.lossless && !tempCS->slice->getDisableSATDForRD();
+#else
const bool bUseHadamard= !encTestMode.lossless;
+#endif
m_pcRdCost->setDistParam (distParam, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth (CHANNEL_TYPE_LUMA), COMPONENT_Y, bUseHadamard);
const UnitArea localUnitArea( tempCS->area.chromaFormat, Area( 0, 0, tempCS->area.Y().width, tempCS->area.Y().height) );
@@ -2308,7 +2312,11 @@ void EncCu::xCheckRDCostMergeTriangle2Nx2N( CodingStructure *&tempCS, CodingStru
}
DistParam distParam;
+#if JVET_N0329_IBC_SEARCH_IMP
+ const bool useHadamard = !encTestMode.lossless && !tempCS->slice->getDisableSATDForRD();
+#else
const bool useHadamard = !encTestMode.lossless;
+#endif
m_pcRdCost->setDistParam( distParam, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth( CHANNEL_TYPE_LUMA ), COMPONENT_Y, useHadamard );
const UnitArea localUnitArea( tempCS->area.chromaFormat, Area( 0, 0, tempCS->area.Y().width, tempCS->area.Y().height) );
@@ -2627,7 +2635,11 @@ void EncCu::xCheckRDCostAffineMerge2Nx2N( CodingStructure *&tempCS, CodingStruct
PredictionUnit &pu = tempCS->addPU( cu, partitioner.chType );
DistParam distParam;
+#if JVET_N0329_IBC_SEARCH_IMP
+ const bool bUseHadamard = !encTestMode.lossless && !tempCS->slice->getDisableSATDForRD();
+#else
const bool bUseHadamard = !encTestMode.lossless;
+#endif
m_pcRdCost->setDistParam( distParam, tempCS->getOrgBuf().Y(), m_acMergeBuffer[0].Y(), sps.getBitDepth( CHANNEL_TYPE_LUMA ), COMPONENT_Y, bUseHadamard );
const UnitArea localUnitArea( tempCS->area.chromaFormat, Area( 0, 0, tempCS->area.Y().width, tempCS->area.Y().height ) );
@@ -2887,7 +2899,11 @@ void EncCu::xCheckRDCostIBCModeMerge2Nx2N(CodingStructure *&tempCS, CodingStruct
cu.mmvdSkip = false;
cu.triangle = false;
DistParam distParam;
+#if JVET_N0329_IBC_SEARCH_IMP
+ const bool bUseHadamard = !encTestMode.lossless && !cu.slice->getDisableSATDForRD();
+#else
const bool bUseHadamard = !encTestMode.lossless;
+#endif
PredictionUnit &pu = tempCS->addPU(cu, partitioner.chType); //tempCS->addPU(cu);
pu.mmvdMergeFlag = false;
Picture* refPic = pu.cu->slice->getPic();
diff --git a/source/Lib/EncoderLib/EncModeCtrl.cpp b/source/Lib/EncoderLib/EncModeCtrl.cpp
index 0c8a5f7c6d..32a940e073 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.cpp
+++ b/source/Lib/EncoderLib/EncModeCtrl.cpp
@@ -1533,7 +1533,11 @@ bool EncModeCtrlMTnoRQT::tryMode( const EncTestMode& encTestmode, const CodingSt
if( lastTestMode().type != ETM_INTRA && cuECtx.bestCS && cuECtx.bestCU && interHadActive( cuECtx ) )
{
// Get SATD threshold from best Inter-CU
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (!cs.slice->isIRAP() && m_pcEncCfg->getUsePbIntraFast() && !cs.slice->getDisableSATDForRD())
+#else
if( !cs.slice->isIRAP() && m_pcEncCfg->getUsePbIntraFast() )
+#endif
{
CodingUnit* bestCU = cuECtx.bestCU;
if (bestCU && !CU::isIntra(*bestCU))
diff --git a/source/Lib/EncoderLib/EncSlice.cpp b/source/Lib/EncoderLib/EncSlice.cpp
index 825d3dd624..0a36a5c041 100644
--- a/source/Lib/EncoderLib/EncSlice.cpp
+++ b/source/Lib/EncoderLib/EncSlice.cpp
@@ -738,6 +738,9 @@ void EncSlice::initEncSlice(Picture* pcPic, const int pocLast, const int pocCurr
rpcSlice->setMaxBTSizeIChroma( rpcSlice->getSPS()->getMaxBTSizeIChroma() );
rpcSlice->setMaxTTSizeIChroma( rpcSlice->getSPS()->getMaxTTSizeIChroma() );
}
+#if JVET_N0329_IBC_SEARCH_IMP
+ rpcSlice->setDisableSATDForRD(false);
+#endif
}
@@ -1584,11 +1587,20 @@ void EncSlice::encodeCtus( Picture* pcPic, const bool bCompressEntireSlice, cons
if ( pcSlice->getSPS()->getDisFracMmvdEnabledFlag() ||
(pcSlice->getSPS()->getIBCFlag() && m_pcCuEncoder->getEncCfg()->getIBCHashSearch()))
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_pcCuEncoder->getIbcHashMap().rebuildPicHashMap(cs.picture->getTrueOrigBuf());
+ if (m_pcCfg->getIntraPeriod() != -1)
+ {
+ int hashBlkHitPerc = m_pcCuEncoder->getIbcHashMap().calHashBlkMatchPerc(cs.area.Y());
+ cs.slice->setDisableSATDForRD(hashBlkHitPerc > 59);
+ }
+#else
if (pcSlice->getSPS()->getUseReshaper() && m_pcLib->getReshaper()->getCTUFlag() && pcSlice->getSPS()->getIBCFlag())
cs.picture->getOrigBuf(COMPONENT_Y).rspSignal(m_pcLib->getReshaper()->getFwdLUT());
m_pcCuEncoder->getIbcHashMap().rebuildPicHashMap( cs.picture->getOrigBuf() );
if (pcSlice->getSPS()->getUseReshaper() && m_pcLib->getReshaper()->getCTUFlag() && pcSlice->getSPS()->getIBCFlag())
cs.picture->getOrigBuf().copyFrom(cs.picture->getTrueOrigBuf());
+#endif
}
checkDisFracMmvd( pcPic, startCtuTsAddr, boundingCtuTsAddr );
// for every CTU in the slice segment (may terminate sooner if there is a byte limit on the slice-segment)
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index 182e93a7a6..25cb3e3495 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -198,7 +198,15 @@ void InterSearch::init( EncCfg* pcEncCfg,
{
CHECK(m_isInitialized, "Already initialized");
m_numBVs = 0;
+#if JVET_N0329_IBC_SEARCH_IMP
+ for (int i = 0; i < IBC_NUM_CANDIDATES; i++)
+ {
+ m_defaultCachedBvs.m_bvCands[i].setZero();
+ }
+ m_defaultCachedBvs.currCnt = 0;
+#else
m_numBV16s = 0;
+#endif
m_pcEncCfg = pcEncCfg;
m_pcTrQuant = pcTrQuant;
m_iSearchRange = iSearchRange;
@@ -747,7 +755,11 @@ Distortion InterSearch::xGetInterPredictionError( PredictionUnit& pu, PelUnitBuf
DistParam cDistParam;
cDistParam.applyWeight = false;
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_pcRdCost->setDistParam(cDistParam, origBuf.Y(), predBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, m_pcEncCfg->getUseHADME() && !pu.cu->transQuantBypass && !pu.cu->slice->getDisableSATDForRD());
+#else
m_pcRdCost->setDistParam( cDistParam, origBuf.Y(), predBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, m_pcEncCfg->getUseHADME() && !pu.cu->transQuantBypass );
+#endif
return (Distortion)cDistParam.distFunc( cDistParam );
}
@@ -805,6 +817,13 @@ int InterSearch::xIBCSearchMVChromaRefine(PredictionUnit& pu,
UnitArea allCompBlocks(pu.chromaFormat, (Area)pu.block(COMPONENT_Y));
for (int cand = 0; cand < CHROMA_REFINEMENT_CANDIDATES; cand++)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (sadBestCand[cand] == std::numeric_limits<Distortion>::max())
+ {
+ continue;
+ }
+#endif
+
if ((!cMVCand[cand].getHor()) && (!cMVCand[cand].getVer()))
continue;
@@ -869,10 +888,22 @@ int InterSearch::xIBCSearchMVChromaRefine(PredictionUnit& pu,
return bestCandIdx;
}
+#if JVET_N0329_IBC_SEARCH_IMP
+static unsigned int xMergeCandLists(Mv *dst, unsigned int dn, unsigned int dstTotalLength, Mv *src, unsigned int sn)
+{
+ for (unsigned int cand = 0; cand < sn && dn < dstTotalLength; cand++)
+#else
static unsigned int xMergeCandLists(Mv *dst, unsigned int dn, Mv *src, unsigned int sn)
{
for (unsigned int cand = 0; cand < sn && dn<IBC_NUM_CANDIDATES; cand++)
+#endif
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (src[cand] == Mv())
+ {
+ continue;
+ }
+#endif
bool found = false;
for (int j = 0; j<dn; j++)
{
@@ -942,7 +973,12 @@ void InterSearch::xIntraPatternSearch(PredictionUnit& pu, IntTZSearchStruct& cS
Distortion tempSadBest = 0;
int srLeft = srchRngHorLeft, srRight = srchRngHorRight, srTop = srchRngVerTop, srBottom = srchRngVerBottom;
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_numBVs = 0;
+ m_numBVs = xMergeCandLists(m_acBVs, m_numBVs, (2 * IBC_NUM_CANDIDATES), m_defaultCachedBvs.m_bvCands, m_defaultCachedBvs.currCnt);
+ Mv cMvPredEncOnly[IBC_NUM_CANDIDATES];
+#else
if (roiWidth>8 || roiHeight>8)
{
m_numBVs = 0;
@@ -953,9 +989,14 @@ void InterSearch::xIntraPatternSearch(PredictionUnit& pu, IntTZSearchStruct& cS
}
Mv cMvPredEncOnly[16];
+#endif
int nbPreds = 0;
PU::getIbcMVPsEncOnly(pu, cMvPredEncOnly, nbPreds);
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_numBVs = xMergeCandLists(m_acBVs, m_numBVs, (2 * IBC_NUM_CANDIDATES), cMvPredEncOnly, nbPreds);
+#else
m_numBVs = xMergeCandLists(m_acBVs, m_numBVs, cMvPredEncOnly, nbPreds);
+#endif
for (unsigned int cand = 0; cand < m_numBVs; cand++)
{
@@ -1196,6 +1237,23 @@ void InterSearch::xIntraPatternSearch(PredictionUnit& pu, IntTZSearchStruct& cS
ruiCost = sadBest;
end:
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_numBVs = 0;
+ m_numBVs = xMergeCandLists(m_acBVs, m_numBVs, (2 * IBC_NUM_CANDIDATES), m_defaultCachedBvs.m_bvCands, m_defaultCachedBvs.currCnt);
+
+ m_defaultCachedBvs.currCnt = 0;
+ m_defaultCachedBvs.currCnt = xMergeCandLists(m_defaultCachedBvs.m_bvCands, m_defaultCachedBvs.currCnt, IBC_NUM_CANDIDATES, cMVCand, CHROMA_REFINEMENT_CANDIDATES);
+ m_defaultCachedBvs.currCnt = xMergeCandLists(m_defaultCachedBvs.m_bvCands, m_defaultCachedBvs.currCnt, IBC_NUM_CANDIDATES, m_acBVs, m_numBVs);
+
+ for (unsigned int cand = 0; cand < CHROMA_REFINEMENT_CANDIDATES; cand++)
+ {
+ if (cMVCand[cand].getHor() == 0 && cMVCand[cand].getVer() == 0)
+ {
+ continue;
+ }
+ m_ctuRecord[pu.lumaPos()][pu.lumaSize()].bvRecord[cMVCand[cand]] = sadBestCand[cand];
+ }
+#else
if (roiWidth + roiHeight > 8)
{
m_numBVs = xMergeCandLists(m_acBVs, m_numBVs, cMVCand, CHROMA_REFINEMENT_CANDIDATES);
@@ -1205,6 +1263,7 @@ end:
m_numBV16s = m_numBVs;
}
}
+#endif
return;
}
@@ -1218,10 +1277,57 @@ void InterSearch::xIBCEstimation(PredictionUnit& pu, PelUnitBuf& origBuf,
Distortion &ruiCost, const int localSearchRangeX, const int localSearchRangeY
)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ const int iPicWidth = pu.cs->slice->getSPS()->getPicWidthInLumaSamples();
+ const int iPicHeight = pu.cs->slice->getSPS()->getPicHeightInLumaSamples();
+ const unsigned int lcuWidth = pu.cs->slice->getSPS()->getMaxCUWidth();
+ const int cuPelX = pu.Y().x;
+ const int cuPelY = pu.Y().y;
+ int iRoiWidth = pu.lwidth();
+ int iRoiHeight = pu.lheight();
+
+ PelUnitBuf* pBuf = &origBuf;
+
+ // Search key pattern initialization
+ CPelBuf tmpPattern = pBuf->Y();
+ CPelBuf* pcPatternKey = &tmpPattern;
+
+ if ((pu.cs->slice->getReshapeInfo().getUseSliceReshaper() && m_pcReshape->getCTUFlag()))
+ {
+ const CompArea &area = pu.blocks[COMPONENT_Y];
+ CompArea tmpArea(COMPONENT_Y, area.chromaFormat, Position(0, 0), area.size());
+ PelBuf tmpOrgLuma = m_tmpStorageLCU.getBuf(tmpArea);
+ tmpOrgLuma.copyFrom(tmpPattern);
+ tmpOrgLuma.rspSignal(m_pcReshape->getFwdLUT());
+ pcPatternKey = (CPelBuf*)&tmpOrgLuma;
+ }
+
+ m_lumaClpRng = pu.cs->slice->clpRng(COMPONENT_Y);
+ Picture* refPic = pu.cu->slice->getPic();
+ const CPelBuf refBuf = refPic->getRecoBuf(pu.blocks[COMPONENT_Y]);
+
+ IntTZSearchStruct cStruct;
+ cStruct.pcPatternKey = pcPatternKey;
+ cStruct.iRefStride = refBuf.stride;
+ cStruct.piRefY = refBuf.buf;
+ cStruct.imvShift = pu.cu->imv << 1;
+ cStruct.subShiftMode = 0; // used by intra pattern search function
+
+ // disable weighted prediction
+ setWpScalingDistParam(-1, REF_PIC_LIST_X, pu.cs->slice);
+
+ m_pcRdCost->getMotionCost(0, pu.cu->transQuantBypass);
+ m_pcRdCost->setPredictors(pcMvPred);
+ m_pcRdCost->setCostScale(0);
+
+ m_cDistParam.useMR = false;
+ m_pcRdCost->setDistParam(m_cDistParam, *cStruct.pcPatternKey, cStruct.piRefY, cStruct.iRefStride, m_lumaClpRng.bd, COMPONENT_Y, cStruct.subShiftMode);
+#endif
bool buffered = false;
if (m_pcEncCfg->getIBCFastMethod() & IBC_FAST_METHOD_BUFFERBV)
{
ruiCost = MAX_UINT;
+#if !JVET_N0329_IBC_SEARCH_IMP
const int iPicWidth = pu.cs->slice->getSPS()->getPicWidthInLumaSamples();
const int iPicHeight = pu.cs->slice->getSPS()->getPicHeightInLumaSamples();
const int cuPelX = pu.Y().x;
@@ -1229,8 +1335,11 @@ void InterSearch::xIBCEstimation(PredictionUnit& pu, PelUnitBuf& origBuf,
int iRoiWidth = pu.lwidth();
int iRoiHeight = pu.lheight();
+#endif
std::unordered_map<Mv, Distortion>& history = m_ctuRecord[pu.lumaPos()][pu.lumaSize()].bvRecord;
+#if !JVET_N0329_IBC_SEARCH_IMP
const unsigned int lcuWidth = pu.cs->slice->getSPS()->getMaxCUWidth();
+#endif
for (std::unordered_map<Mv, Distortion>::iterator p = history.begin(); p != history.end(); p++)
{
const Mv& bv = p->first;
@@ -1238,7 +1347,28 @@ void InterSearch::xIBCEstimation(PredictionUnit& pu, PelUnitBuf& origBuf,
int xBv = bv.hor;
int yBv = bv.ver;
if (PU::isBlockVectorValid(pu, cuPelX, cuPelY, iRoiWidth, iRoiHeight, iPicWidth, iPicHeight, 0, 0, xBv, yBv, lcuWidth))
- {
+ {
+#if JVET_N0329_IBC_SEARCH_IMP
+ buffered = true;
+ Distortion sad = m_pcRdCost->getBvCostMultiplePreds(xBv, yBv, pu.cs->sps->getAMVREnabledFlag());
+ m_cDistParam.cur.buf = cStruct.piRefY + cStruct.iRefStride * yBv + xBv;
+ sad += m_cDistParam.distFunc(m_cDistParam);
+ if (sad < ruiCost)
+ {
+ rcMv = bv;
+ ruiCost = sad;
+ }
+ else if (sad == ruiCost)
+ {
+ // stabilise the search through the unordered list
+ if (bv.hor < rcMv.getHor()
+ || (bv.hor == rcMv.getHor() && bv.ver < rcMv.getVer()))
+ {
+ // update the vector.
+ rcMv = bv;
+ }
+ }
+#else
if (p->second < ruiCost)
{
rcMv = bv;
@@ -1255,12 +1385,59 @@ void InterSearch::xIBCEstimation(PredictionUnit& pu, PelUnitBuf& origBuf,
rcMv = bv;
}
}
+#endif
+ }
+ }
+
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (buffered)
+ {
+ Mv cMvPredEncOnly[IBC_NUM_CANDIDATES];
+ int nbPreds = 0;
+ PU::getIbcMVPsEncOnly(pu, cMvPredEncOnly, nbPreds);
+
+ for (unsigned int cand = 0; cand < nbPreds; cand++)
+ {
+ int xPred = cMvPredEncOnly[cand].getHor();
+ int yPred = cMvPredEncOnly[cand].getVer();
+
+ if (PU::isBlockVectorValid(pu, cuPelX, cuPelY, iRoiWidth, iRoiHeight, iPicWidth, iPicHeight, 0, 0, xPred, yPred, lcuWidth))
+ {
+ Distortion sad = m_pcRdCost->getBvCostMultiplePreds(xPred, yPred, pu.cs->sps->getAMVREnabledFlag());
+ m_cDistParam.cur.buf = cStruct.piRefY + cStruct.iRefStride * yPred + xPred;
+ sad += m_cDistParam.distFunc(m_cDistParam);
+ if (sad < ruiCost)
+ {
+ rcMv.set(xPred, yPred);
+ ruiCost = sad;
+ }
+ else if (sad == ruiCost)
+ {
+ // stabilise the search through the unordered list
+ if (xPred < rcMv.getHor()
+ || (xPred == rcMv.getHor() && yPred < rcMv.getVer()))
+ {
+ // update the vector.
+ rcMv.set(xPred, yPred);
+ }
+ }
+
+ m_ctuRecord[pu.lumaPos()][pu.lumaSize()].bvRecord[Mv(xPred, yPred)] = sad;
+ }
}
}
+#endif
}
if (!buffered)
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ Mv cMvSrchRngLT;
+ Mv cMvSrchRngRB;
+
+ // assume that intra BV is integer-pel precision
+ xSetIntraSearchRange(pu, pu.lwidth(), pu.lheight(), localSearchRangeX, localSearchRangeY, cMvSrchRngLT, cMvSrchRngRB);
+#else
Mv cMvSrchRngLT;
Mv cMvSrchRngRB;
@@ -1304,9 +1481,9 @@ void InterSearch::xIBCEstimation(PredictionUnit& pu, PelUnitBuf& origBuf,
m_pcRdCost->getMotionCost(0, pu.cu->transQuantBypass);
m_pcRdCost->setPredictors(pcMvPred);
m_pcRdCost->setCostScale(0);
+#endif
// Do integer search
-
xIntraPatternSearch(pu, cStruct, rcMv, ruiCost, &cMvSrchRngLT, &cMvSrchRngRB, pcMvPred);
}
}
@@ -1469,7 +1646,9 @@ bool InterSearch::predIBCSearch(CodingUnit& cu, Partitioner& partitioner, const
pu.refIdx[REF_PIC_LIST_0] = MAX_NUM_REF;
pu.mv[REF_PIC_LIST_0].changePrecision(MV_PRECISION_QUARTER, MV_PRECISION_INTERNAL);
+#if !JVET_N0329_IBC_SEARCH_IMP
m_ctuRecord[cu.lumaPos()][cu.lumaSize()].bvRecord[pu.bv] = cost;
+#endif
}
return true;
@@ -2898,9 +3077,15 @@ Distortion InterSearch::xGetAffineTemplateCost( PredictionUnit& pu, PelUnitBuf&
}
// calc distortion
-
+#if JVET_N0329_IBC_SEARCH_IMP
+ enum DFunc distFunc = (pu.cu->transQuantBypass || pu.cs->slice->getDisableSATDForRD()) ? DF_SAD : DF_HAD;
+#endif
uiCost = m_pcRdCost->getDistPart( origBuf.Y(), predBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y
- , DF_HAD
+#if JVET_N0329_IBC_SEARCH_IMP
+ , distFunc
+#else
+ , DF_HAD
+#endif
);
uiCost += m_pcRdCost->getCost( m_auiMVPIdxCost[iMVPIdx][iMVPNum] );
DTRACE( g_trace_ctx, D_COMMON, " (%d) affineTemplateCost=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiCost );
@@ -3686,7 +3871,11 @@ void InterSearch::xPatternSearchIntRefine(PredictionUnit& pu, IntTZSearchStruct&
CHECK( amvpInfo.mvCand[riMVPIdx] != rcMvPred, "xPatternSearchIntRefine(): MvPred issue.");
const SPS &sps = *pu.cs->sps;
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_pcRdCost->setDistParam(m_cDistParam, *cStruct.pcPatternKey, cStruct.piRefY, cStruct.iRefStride, m_lumaClpRng.bd, COMPONENT_Y, 0, 1, m_pcEncCfg->getUseHADME() && !pu.cu->transQuantBypass && !pu.cs->slice->getDisableSATDForRD());
+#else
m_pcRdCost->setDistParam( m_cDistParam, *cStruct.pcPatternKey, cStruct.piRefY, cStruct.iRefStride, m_lumaClpRng.bd, COMPONENT_Y, 0, 1, m_pcEncCfg->getUseHADME() && !pu.cu->transQuantBypass );
+#endif
// input MV rcMV has integer resolution
// -> shift it to QPEL
@@ -3821,14 +4010,22 @@ void InterSearch::xPatternSearchFracDIF(
m_pcRdCost->setCostScale(0);
xExtDIFUpSamplingH(&cPatternRoi);
rcMvQter = rcMvInt; rcMvQter <<= 2; // for mv-cost
+#if JVET_N0329_IBC_SEARCH_IMP
+ ruiCost = xPatternRefinement(cStruct.pcPatternKey, baseRefMv, 1, rcMvQter, !bIsLosslessCoded && !pu.cs->slice->getDisableSATDForRD());
+#else
ruiCost = xPatternRefinement(cStruct.pcPatternKey, baseRefMv, 1, rcMvQter, !bIsLosslessCoded);
+#endif
return;
}
if (cStruct.imvShift || (m_useCompositeRef && cStruct.zeroMV))
{
+#if JVET_N0329_IBC_SEARCH_IMP
+ m_pcRdCost->setDistParam(m_cDistParam, *cStruct.pcPatternKey, cStruct.piRefY + iOffset, cStruct.iRefStride, m_lumaClpRng.bd, COMPONENT_Y, 0, 1, m_pcEncCfg->getUseHADME() && !bIsLosslessCoded && !pu.cs->slice->getDisableSATDForRD());
+#else
m_pcRdCost->setDistParam( m_cDistParam, *cStruct.pcPatternKey, cStruct.piRefY + iOffset, cStruct.iRefStride, m_lumaClpRng.bd, COMPONENT_Y, 0, 1, m_pcEncCfg->getUseHADME() && !bIsLosslessCoded );
+#endif
ruiCost = m_cDistParam.distFunc( m_cDistParam );
ruiCost += m_pcRdCost->getCostOfVectorWithPredictor( rcMvInt.getHor(), rcMvInt.getVer(), cStruct.imvShift );
return;
@@ -3840,7 +4037,11 @@ void InterSearch::xPatternSearchFracDIF(
rcMvHalf = rcMvInt; rcMvHalf <<= 1; // for mv-cost
Mv baseRefMv(0, 0);
+#if JVET_N0329_IBC_SEARCH_IMP
+ ruiCost = xPatternRefinement(cStruct.pcPatternKey, baseRefMv, 2, rcMvHalf, (!bIsLosslessCoded && !pu.cs->slice->getDisableSATDForRD()));
+#else
ruiCost = xPatternRefinement(cStruct.pcPatternKey, baseRefMv, 2, rcMvHalf, !bIsLosslessCoded);
+#endif
// quarter-pel refinement
m_pcRdCost->setCostScale( 0 );
@@ -3850,7 +4051,11 @@ void InterSearch::xPatternSearchFracDIF(
rcMvQter = rcMvInt; rcMvQter <<= 1; // for mv-cost
rcMvQter += rcMvHalf; rcMvQter <<= 1;
+#if JVET_N0329_IBC_SEARCH_IMP
+ ruiCost = xPatternRefinement(cStruct.pcPatternKey, baseRefMv, 1, rcMvQter, (!bIsLosslessCoded && !pu.cs->slice->getDisableSATDForRD()));
+#else
ruiCost = xPatternRefinement( cStruct.pcPatternKey, baseRefMv, 1, rcMvQter, !bIsLosslessCoded );
+#endif
}
Distortion InterSearch::xGetSymmetricCost( PredictionUnit& pu, PelUnitBuf& origBuf, RefPicList eCurRefPicList, const MvField& cCurMvField, MvField& cTarMvField, int gbiIdx )
@@ -3881,7 +4086,12 @@ Distortion InterSearch::xGetSymmetricCost( PredictionUnit& pu, PelUnitBuf& origB
bufTmp.Y().addAvg( predBufA.Y(), predBufB.Y(), pu.cu->slice->clpRng( COMPONENT_Y ) );
// calc distortion
+#if JVET_N0329_IBC_SEARCH_IMP
+ DFunc distFunc = (!pu.cu->transQuantBypass && !pu.cu->slice->getDisableSATDForRD()) ? DF_HAD : DF_SAD;
+ cost = m_pcRdCost->getDistPart(bufTmp.Y(), origBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
cost = m_pcRdCost->getDistPart(bufTmp.Y(), origBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, DF_HAD);
+#endif
return(cost);
}
@@ -4925,6 +5135,9 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
double fWeight = 1.0;
PelUnitBuf origBufTmp = m_tmpStorageLCU.getBuf( UnitAreaRelative( *pu.cu, pu ) );
+#if JVET_N0329_IBC_SEARCH_IMP
+ enum DFunc distFunc = (pu.cu->transQuantBypass || pu.cs->slice->getDisableSATDForRD()) ? DF_SAD : DF_HAD;
+#endif
// if Bi, set to ( 2 * Org - ListX )
if ( bBi )
@@ -5010,7 +5223,11 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
xPredAffineBlk( COMPONENT_Y, pu, refPic, acMvTemp, predBuf, false, pu.cs->slice->clpRng( COMPONENT_Y ) );
// get error
+#if JVET_N0329_IBC_SEARCH_IMP
+ uiCostBest = m_pcRdCost->getDistPart(predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
uiCostBest = m_pcRdCost->getDistPart( predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, DF_HAD );
+#endif
// get cost with mv
m_pcRdCost->setCostScale(0);
@@ -5202,7 +5419,11 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
xPredAffineBlk( COMPONENT_Y, pu, refPic, acMvTemp, predBuf, false, pu.cu->slice->clpRng( COMPONENT_Y ) );
// get error
+#if JVET_N0329_IBC_SEARCH_IMP
+ Distortion uiCostTemp = m_pcRdCost->getDistPart(predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
Distortion uiCostTemp = m_pcRdCost->getDistPart( predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, DF_HAD );
+#endif
DTRACE( g_trace_ctx, D_COMMON, " (%d) uiCostTemp=%d\n", DTRACE_GET_COUNTER(g_trace_ctx,D_COMMON), uiCostTemp );
// get cost with mv
@@ -5235,7 +5456,11 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
{
xPredAffineBlk(COMPONENT_Y, pu, refPic, ctrlPtMv, predBuf, false, pu.cu->slice->clpRng(COMPONENT_Y));
// get error
+#if JVET_N0329_IBC_SEARCH_IMP
+ Distortion costTemp = m_pcRdCost->getDistPart(predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
Distortion costTemp = m_pcRdCost->getDistPart(predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, DF_HAD);
+#endif
// get cost with mv
m_pcRdCost->setCostScale(0);
uint32_t bitsTemp = ruiBits;
@@ -5277,7 +5502,11 @@ void InterSearch::xAffineMotionEstimation( PredictionUnit& pu,
clipMv( acMvTemp[j], pu.cu->lumaPos(), pu.cu->lumaSize(), *pu.cs->sps );
xPredAffineBlk( COMPONENT_Y, pu, refPic, acMvTemp, predBuf, false, pu.cu->slice->clpRng( COMPONENT_Y ) );
+#if JVET_N0329_IBC_SEARCH_IMP
+ Distortion costTemp = m_pcRdCost->getDistPart(predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
Distortion costTemp = m_pcRdCost->getDistPart( predBuf.Y(), pBuf->Y(), pu.cs->sps->getBitDepth( CHANNEL_TYPE_LUMA ), COMPONENT_Y, DF_HAD );
+#endif
uint32_t bitsTemp = ruiBits;
bitsTemp += xCalcAffineMVBits( pu, acMvTemp, acMvPred, pu.cu->imv != 1 );
costTemp = ( Distortion ) ( floor( fWeight * ( double ) costTemp ) + ( double ) m_pcRdCost->getCost( bitsTemp ) );
@@ -7066,7 +7295,12 @@ void InterSearch::symmvdCheckBestMvp(
bufTmp.Y().addAvg(predBufA.Y(), predBufB.Y(), pu.cu->slice->clpRng(COMPONENT_Y));
// calc distortion
+#if JVET_N0329_IBC_SEARCH_IMP
+ DFunc distFunc = (!pu.cu->transQuantBypass && !pu.cu->slice->getDisableSATDForRD()) ? DF_HAD : DF_SAD;
+ Distortion cost = m_pcRdCost->getDistPart(bufTmp.Y(), origBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, distFunc);
+#else
Distortion cost = m_pcRdCost->getDistPart(bufTmp.Y(), origBuf.Y(), pu.cs->sps->getBitDepth(CHANNEL_TYPE_LUMA), COMPONENT_Y, DF_HAD);
+#endif
m_pcRdCost->setPredictor(amvpCur.mvCand[i]);
uint32_t bits = m_pcRdCost->getBitsOfVectorWithPredictor(curMv.hor, curMv.ver, (pu.cu->imv << 1));
diff --git a/source/Lib/EncoderLib/InterSearch.h b/source/Lib/EncoderLib/InterSearch.h
index d2a4bccb92..acc2861c1f 100644
--- a/source/Lib/EncoderLib/InterSearch.h
+++ b/source/Lib/EncoderLib/InterSearch.h
@@ -118,6 +118,9 @@ private:
int m_affMVListMaxSize;
Distortion m_hevcCost;
EncAffineMotion m_affineMotion;
+#if JVET_N0329_IBC_SEARCH_IMP
+ PatentBvCand m_defaultCachedBvs;
+#endif
protected:
// interface to option
EncCfg* m_pcEncCfg;
@@ -150,8 +153,14 @@ protected:
Mv m_integerMv2Nx2N [NUM_REF_PIC_LIST_01][MAX_NUM_REF];
bool m_isInitialized;
+
+#if JVET_N0329_IBC_SEARCH_IMP
+ Mv m_acBVs[2 * IBC_NUM_CANDIDATES];
+ unsigned int m_numBVs;
+#else
unsigned int m_numBVs, m_numBV16s;
Mv m_acBVs[IBC_NUM_CANDIDATES];
+#endif
bool m_useCompositeRef;
Distortion m_estMinDistSbt[NUMBER_SBT_MODE + 1]; // estimated minimum SSE value of the PU if using a SBT mode
uint8_t m_sbtRdoOrder[NUMBER_SBT_MODE]; // order of SBT mode in RDO
@@ -281,7 +290,18 @@ public:
bool predIBCSearch ( CodingUnit& cu, Partitioner& partitioner, const int localSearchRangeX, const int localSearchRangeY, IbcHashMap& ibcHashMap);
void xIntraPatternSearch ( PredictionUnit& pu, IntTZSearchStruct& cStruct, Mv& rcMv, Distortion& ruiCost, Mv* cMvSrchRngLT, Mv* cMvSrchRngRB, Mv* pcMvPred);
void xSetIntraSearchRange ( PredictionUnit& pu, int iRoiWidth, int iRoiHeight, const int localSearchRangeX, const int localSearchRangeY, Mv& rcMvSrchRngLT, Mv& rcMvSrchRngRB);
+#if JVET_N0329_IBC_SEARCH_IMP
+ void resetIbcSearch()
+ {
+ for (int i = 0; i < IBC_NUM_CANDIDATES; i++)
+ {
+ m_defaultCachedBvs.m_bvCands[i].setZero();
+ }
+ m_defaultCachedBvs.currCnt = 0;
+ }
+#else
void resetIbcSearch() { m_numBVs = m_numBV16s = 0; }
+#endif
void xIBCEstimation ( PredictionUnit& pu, PelUnitBuf& origBuf, Mv *pcMvPred, Mv &rcMv, Distortion &ruiCost, const int localSearchRangeX, const int localSearchRangeY);
void xIBCSearchMVCandUpdate ( Distortion uiSad, int x, int y, Distortion* uiSadBestCand, Mv* cMVCand);
int xIBCSearchMVChromaRefine( PredictionUnit& pu, int iRoiWidth, int iRoiHeight, int cuPelX, int cuPelY, Distortion* uiSadBestCand, Mv* cMVCand);
diff --git a/source/Lib/EncoderLib/IntraSearch.cpp b/source/Lib/EncoderLib/IntraSearch.cpp
index 1fcaedc19c..5d73a658c8 100644
--- a/source/Lib/EncoderLib/IntraSearch.cpp
+++ b/source/Lib/EncoderLib/IntraSearch.cpp
@@ -678,7 +678,11 @@ void IntraSearch::estIntraPredLumaQT( CodingUnit &cu, Partitioner &partitioner,
// after this point, don't use numModesForFullRD
// PBINTRA fast
+#if JVET_N0329_IBC_SEARCH_IMP
+ if (m_pcEncCfg->getUsePbIntraFast() && !cs.slice->isIntra() && uiRdModeList.size() < numModesAvailable && !cs.slice->getDisableSATDForRD())
+#else
if( m_pcEncCfg->getUsePbIntraFast() && !cs.slice->isIntra() && uiRdModeList.size() < numModesAvailable )
+#endif
{
if( CandHadList.size() < 3 || CandHadList[2] > cs.interHad * PBINTRA_RATIO )
{
@@ -976,7 +980,11 @@ void IntraSearch::estIntraPredChromaQT( CodingUnit &cu, Partitioner &partitioner
}
DistParam distParam;
+#if JVET_N0329_IBC_SEARCH_IMP
+ const bool useHadamard = !cu.transQuantBypass;
+#else
const bool useHadamard = true;
+#endif
pu.intraDir[1] = MDLM_L_IDX; // temporary assigned, just to indicate this is a MDLM mode. for luma down-sampling operation.
initIntraPatternChType(cu, pu.Cb());
--
GitLab