From 8e16ada0855a08e590ceb5770c2915a01bd7ae84 Mon Sep 17 00:00:00 2001
From: Frank Bossen <fbossen@gmail.com>
Date: Sun, 18 Aug 2019 11:47:06 -0400
Subject: [PATCH] Clean up: reduce memory reallocation in
CodingStructure::createCoeffs
Use vectors and resize them if needed instead of deleting and
reallocating memory for every picture. This appears to have a
significant impact on decoder run time.
Use PLTRunMode type where appropriate instead of bool since
using vector<bool> creates a number of issues
---
source/Lib/CommonLib/Buffer.h | 4 +--
source/Lib/CommonLib/CodingStructure.cpp | 40 +++++++-----------------
source/Lib/CommonLib/CodingStructure.h | 8 ++---
source/Lib/CommonLib/TypeDef.h | 6 +++-
source/Lib/CommonLib/Unit.cpp | 7 +++--
source/Lib/CommonLib/Unit.h | 6 ++--
source/Lib/DecoderLib/CABACReader.cpp | 13 +++++---
source/Lib/EncoderLib/CABACWriter.cpp | 7 +++--
source/Lib/EncoderLib/EncModeCtrl.cpp | 10 +++---
source/Lib/EncoderLib/EncModeCtrl.h | 2 +-
source/Lib/EncoderLib/IntraSearch.cpp | 6 ++--
11 files changed, 54 insertions(+), 55 deletions(-)
diff --git a/source/Lib/CommonLib/Buffer.h b/source/Lib/CommonLib/Buffer.h
index e76cdd7a5..895ad0023 100644
--- a/source/Lib/CommonLib/Buffer.h
+++ b/source/Lib/CommonLib/Buffer.h
@@ -177,8 +177,8 @@ typedef AreaBuf<const MotionInfo> CMotionBuf;
typedef AreaBuf< TCoeff> PLTescapeBuf;
typedef AreaBuf<const TCoeff> CPLTescapeBuf;
-typedef AreaBuf< bool> PLTtypeBuf;
-typedef AreaBuf<const bool> CPLTtypeBuf;
+typedef AreaBuf<PLTRunMode> PLTtypeBuf;
+typedef AreaBuf<const PLTRunMode> CPLTtypeBuf;
#endif
#define SIZE_AWARE_PER_EL_OP( OP, INC ) \
diff --git a/source/Lib/CommonLib/CodingStructure.cpp b/source/Lib/CommonLib/CodingStructure.cpp
index d8814defc..9f350676e 100644
--- a/source/Lib/CommonLib/CodingStructure.cpp
+++ b/source/Lib/CommonLib/CodingStructure.cpp
@@ -76,13 +76,6 @@ CodingStructure::CodingStructure(CUCache& cuCache, PUCache& puCache, TUCache& tu
{
for( uint32_t i = 0; i < MAX_NUM_COMPONENT; i++ )
{
- m_coeffs[ i ] = nullptr;
- m_pcmbuf[ i ] = nullptr;
-#if JVET_O0119_BASE_PALETTE_444
- m_runType[i] = nullptr;
- m_runLength[i] = nullptr;
-#endif
-
m_offsets[ i ] = 0;
}
@@ -640,7 +633,7 @@ TransformUnit& CodingStructure::addTU( const UnitArea &unit, const ChannelType c
TCoeff *coeffs[5] = { nullptr, nullptr, nullptr, nullptr, nullptr };
Pel *pcmbuf[5] = { nullptr, nullptr, nullptr, nullptr, nullptr };
#if JVET_O0119_BASE_PALETTE_444
- bool *runType[5] = { nullptr, nullptr, nullptr, nullptr, nullptr };
+ PLTRunMode *runType[5] = { nullptr, nullptr, nullptr, nullptr, nullptr };
Pel *runLength[5] = { nullptr, nullptr, nullptr, nullptr, nullptr };
#endif
@@ -677,11 +670,11 @@ TransformUnit& CodingStructure::addTU( const UnitArea &unit, const ChannelType c
}
}
- coeffs[i] = m_coeffs[i] + m_offsets[i];
- pcmbuf[i] = m_pcmbuf[i] + m_offsets[i];
+ coeffs[i] = m_coeffs[i].data() + m_offsets[i];
+ pcmbuf[i] = m_pcmbuf[i].data() + m_offsets[i];
#if JVET_O0119_BASE_PALETTE_444
- runType[i] = m_runType[i] + m_offsets[i];
- runLength[i] = m_runLength[i] + m_offsets[i];
+ runType[i] = m_runType[i].data() + m_offsets[i];
+ runLength[i] = m_runLength[i].data() + m_offsets[i];
#endif
unsigned areaSize = tu->blocks[i].area();
@@ -969,32 +962,23 @@ void CodingStructure::rebindPicBufs()
void CodingStructure::createCoeffs()
{
- const unsigned numCh = getNumberValidComponents( area.chromaFormat );
+ const size_t numCh = getNumberValidComponents(area.chromaFormat);
- for( unsigned i = 0; i < numCh; i++ )
+ for (size_t i = 0; i < numCh; i++)
{
- unsigned _area = area.blocks[i].area();
+ size_t _area = area.blocks[i].area();
- m_coeffs[i] = _area > 0 ? ( TCoeff* ) xMalloc( TCoeff, _area ) : nullptr;
- m_pcmbuf[i] = _area > 0 ? ( Pel* ) xMalloc( Pel, _area ) : nullptr;
+ m_coeffs[i].resize(_area);
+ m_pcmbuf[i].resize(_area);
#if JVET_O0119_BASE_PALETTE_444
- m_runType[i] = _area > 0 ? ( bool* ) xMalloc( bool, _area ) : nullptr;
- m_runLength[i] = _area > 0 ? ( Pel* ) xMalloc( Pel, _area ) : nullptr;
+ m_runType[i].resize(_area);
+ m_runLength[i].resize(_area);
#endif
}
}
void CodingStructure::destroyCoeffs()
{
- for( uint32_t i = 0; i < MAX_NUM_COMPONENT; i++ )
- {
- if( m_coeffs[i] ) { xFree( m_coeffs[i] ); m_coeffs[i] = nullptr; }
- if( m_pcmbuf[i] ) { xFree( m_pcmbuf[i] ); m_pcmbuf[i] = nullptr; }
-#if JVET_O0119_BASE_PALETTE_444
- if (m_runType[i]) { xFree(m_runType[i]); m_runType[i] = nullptr; }
- if (m_runLength[i]) { xFree(m_runLength[i]); m_runLength[i] = nullptr; }
-#endif
- }
}
void CodingStructure::initSubStructure( CodingStructure& subStruct, const ChannelType _chType, const UnitArea &subArea, const bool &isTuEnc )
diff --git a/source/Lib/CommonLib/CodingStructure.h b/source/Lib/CommonLib/CodingStructure.h
index 8a11a3bef..5ee873eba 100644
--- a/source/Lib/CommonLib/CodingStructure.h
+++ b/source/Lib/CommonLib/CodingStructure.h
@@ -247,11 +247,11 @@ private:
PelStorage m_reco;
PelStorage m_orgr;
- TCoeff *m_coeffs [ MAX_NUM_COMPONENT ];
- Pel *m_pcmbuf [ MAX_NUM_COMPONENT ];
+ std::vector<TCoeff> m_coeffs[MAX_NUM_COMPONENT];
+ std::vector<Pel> m_pcmbuf[MAX_NUM_COMPONENT];
#if JVET_O0119_BASE_PALETTE_444
- bool *m_runType [MAX_NUM_COMPONENT];
- Pel *m_runLength[MAX_NUM_COMPONENT];
+ std::vector<PLTRunMode> m_runType[MAX_NUM_COMPONENT];
+ std::vector<Pel> m_runLength[MAX_NUM_COMPONENT];
#endif
int m_offsets[ MAX_NUM_COMPONENT ];
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 2a72e25f2..1665bc551 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -1131,7 +1131,11 @@ enum PLTRunMode
PLT_RUN_INDEX = 0,
PLT_RUN_COPY = 1,
NUM_PLT_RUN = 2
-};
+}
+#if __GNUC__
+__attribute__((packed))
+#endif
+;
#endif
/// parameters for deblocking filter
struct LFCUParam
diff --git a/source/Lib/CommonLib/Unit.cpp b/source/Lib/CommonLib/Unit.cpp
index 05d854815..c2c18d717 100644
--- a/source/Lib/CommonLib/Unit.cpp
+++ b/source/Lib/CommonLib/Unit.cpp
@@ -755,7 +755,7 @@ void TransformUnit::initData()
}
#if JVET_O0119_BASE_PALETTE_444
-void TransformUnit::init(TCoeff **coeffs, Pel **pcmbuf, Pel **runLength, bool **runType)
+void TransformUnit::init(TCoeff **coeffs, Pel **pcmbuf, Pel **runLength, PLTRunMode **runType)
#else
void TransformUnit::init(TCoeff **coeffs, Pel **pcmbuf)
#endif
@@ -848,7 +848,10 @@ const CPLTescapeBuf TransformUnit::getescapeValue(const ComponentID id) const {
Pel* TransformUnit::getPLTIndex (const ComponentID id) { return m_pcmbuf[id]; }
Pel* TransformUnit::getRunLens (const ComponentID id) { return m_runLength[id]; }
- bool* TransformUnit::getRunTypes (const ComponentID id) { return m_runType[id]; }
+ PLTRunMode * TransformUnit::getRunTypes(const ComponentID id)
+ {
+ return m_runType[id];
+ }
#endif
void TransformUnit::checkTuNoResidual( unsigned idx )
diff --git a/source/Lib/CommonLib/Unit.h b/source/Lib/CommonLib/Unit.h
index f5c92b9d8..8a8deffeb 100644
--- a/source/Lib/CommonLib/Unit.h
+++ b/source/Lib/CommonLib/Unit.h
@@ -486,7 +486,7 @@ struct TransformUnit : public UnitArea
TransformUnit *prev;
#if JVET_O0119_BASE_PALETTE_444
- void init(TCoeff **coeffs, Pel **pcmbuf, Pel **runLength, bool **runType);
+ void init(TCoeff **coeffs, Pel **pcmbuf, Pel **runLength, PLTRunMode **runType);
#else
void init(TCoeff **coeffs, Pel **pcmbuf);
#endif
@@ -515,7 +515,7 @@ struct TransformUnit : public UnitArea
const CPLTescapeBuf getescapeValue(const ComponentID id) const;
Pel* getPLTIndex(const ComponentID id);
Pel* getRunLens(const ComponentID id);
- bool* getRunTypes(const ComponentID id);
+ PLTRunMode * getRunTypes(const ComponentID id);
#endif
#if ENABLE_SPLIT_PARALLELISM || ENABLE_WPP_PARALLELISM
@@ -527,7 +527,7 @@ private:
TCoeff *m_coeffs[ MAX_NUM_TBLOCKS ];
Pel *m_pcmbuf[ MAX_NUM_TBLOCKS ];
#if JVET_O0119_BASE_PALETTE_444
- bool *m_runType[MAX_NUM_TBLOCKS];
+ PLTRunMode *m_runType[MAX_NUM_TBLOCKS];
Pel *m_runLength[MAX_NUM_TBLOCKS];
#endif
};
diff --git a/source/Lib/DecoderLib/CABACReader.cpp b/source/Lib/DecoderLib/CABACReader.cpp
index 78e8ff012..41aedf216 100644
--- a/source/Lib/DecoderLib/CABACReader.cpp
+++ b/source/Lib/DecoderLib/CABACReader.cpp
@@ -1923,7 +1923,7 @@ void CABACReader::cu_palette_info(CodingUnit& cu, ComponentID compBegin, uint32_
uint32_t width = cu.block(compBegin).width;
int numCopyIndexRuns = -1;
- bool lastRunType = 0;
+ PLTRunMode lastRunType = PLT_RUN_INDEX;
uint32_t numIndices = 0;
uint32_t adjust = 0;
uint32_t symbol = 0;
@@ -1940,7 +1940,7 @@ void CABACReader::cu_palette_info(CodingUnit& cu, ComponentID compBegin, uint32_
adjust = 1;
parsedIdxList.push_back(symbol);
}
- lastRunType = m_BinDecoder.decodeBin(Ctx::RunTypeFlag());
+ lastRunType = m_BinDecoder.decodeBin(Ctx::RunTypeFlag()) ? PLT_RUN_COPY : PLT_RUN_INDEX;
parseScanRotationModeFlag(cu, compBegin);
adjust = 0;
}
@@ -2002,9 +2002,11 @@ void CABACReader::cu_palette_info(CodingUnit& cu, ComponentID compBegin, uint32_
}
else
{
- if (numCopyIndexRuns && strPos < endPos - 1) // JC: if numIndices (decoder will know this value) == 0 - > only CopyAbove, if strPos == endPos - 1, the last RunType was already coded
+ if (numCopyIndexRuns
+ && strPos < endPos - 1) // JC: if numIndices (decoder will know this value) == 0 - > only CopyAbove, if
+ // strPos == endPos - 1, the last PLTRunMode was already coded
{
- runType.at(posx, posy) = (m_BinDecoder.decodeBin(Ctx::RunTypeFlag()));
+ runType.at(posx, posy) = m_BinDecoder.decodeBin(Ctx::RunTypeFlag()) ? PLT_RUN_COPY : PLT_RUN_INDEX;
}
else
{
@@ -2046,7 +2048,8 @@ void CABACReader::cu_palette_info(CodingUnit& cu, ComponentID compBegin, uint32_
lastRun = numCopyIndexRuns == 0 && runType.at(posx, posy) == lastRunType;
if (!lastRun)
{
- runLength.at(posx, posy) = cu_run_val((PLTRunMode)runType.at(posx, posy), curLevel, endPos - strPos - numCopyIndexRuns - 1 - lastRunType) + 1;
+ runLength.at(posx, posy) =
+ cu_run_val(runType.at(posx, posy), curLevel, endPos - strPos - numCopyIndexRuns - 1 - lastRunType) + 1;
}
else
{
diff --git a/source/Lib/EncoderLib/CABACWriter.cpp b/source/Lib/EncoderLib/CABACWriter.cpp
index 8156fa3a1..82558c058 100644
--- a/source/Lib/EncoderLib/CABACWriter.cpp
+++ b/source/Lib/EncoderLib/CABACWriter.cpp
@@ -1821,7 +1821,9 @@ void CABACWriter::cu_palette_info(const CodingUnit& cu, ComponentID compBegin, u
}
else
{
- if (numIndices && strPos < endPos - 1) // if numIndices (decoder will know this value) == 0 - > only CopyAbove, if strPos == endPos - 1, the last RunType was already coded
+ if (numIndices
+ && strPos < endPos - 1) // if numIndices (decoder will know this value) == 0 - > only CopyAbove, if strPos
+ // == endPos - 1, the last PLTRunMode was already coded
{
m_BinEncoder.encodeBin((runType.at(posx, posy)), Ctx::RunTypeFlag());
}
@@ -1846,7 +1848,8 @@ void CABACWriter::cu_palette_info(const CodingUnit& cu, ComponentID compBegin, u
if (lastRunPos != strPos)
{
numIndices -= (runType.at(posx, posy) == PLT_RUN_INDEX);
- cu_run_val(runLength.at(posx, posy) - 1, (PLTRunMode)runType.at(posx, posy), curLevel, endPos - strPos - numIndices - 1 - lastRunType);
+ cu_run_val(runLength.at(posx, posy) - 1, runType.at(posx, posy), curLevel,
+ endPos - strPos - numIndices - 1 - lastRunType);
}
}
diff --git a/source/Lib/EncoderLib/EncModeCtrl.cpp b/source/Lib/EncoderLib/EncModeCtrl.cpp
index fc7a30a4d..2de1f8c89 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.cpp
+++ b/source/Lib/EncoderLib/EncModeCtrl.cpp
@@ -783,14 +783,14 @@ void BestEncInfoCache::init( const Slice &slice )
m_pCoeff = new TCoeff[numCoeff*MAX_NUM_TUS];
m_pPcmBuf = new Pel [numCoeff*MAX_NUM_TUS];
#if JVET_O0119_BASE_PALETTE_444
- m_runType = new bool[numCoeff*MAX_NUM_TUS];
+ m_runType = new PLTRunMode[numCoeff * MAX_NUM_TUS];
m_runLength = new Pel [numCoeff*MAX_NUM_TUS];
#endif
#else
m_pCoeff = new TCoeff[numCoeff];
m_pPcmBuf = new Pel [numCoeff];
#if JVET_O0119_BASE_PALETTE_444
- m_runType = new bool[numCoeff];
+ m_runType = new PLTRunMode[numCoeff];
m_runLength = new Pel [numCoeff];
#endif
#endif
@@ -798,7 +798,7 @@ void BestEncInfoCache::init( const Slice &slice )
TCoeff *coeffPtr = m_pCoeff;
Pel *pcmPtr = m_pPcmBuf;
#if JVET_O0119_BASE_PALETTE_444
- bool *runTypePtr = m_runType;
+ PLTRunMode *runTypePtr = m_runType;
Pel *runLengthPtr = m_runLength;
#endif
@@ -817,7 +817,9 @@ void BestEncInfoCache::init( const Slice &slice )
TCoeff *coeff[MAX_NUM_TBLOCKS] = { 0, };
Pel *pcmbf[MAX_NUM_TBLOCKS] = { 0, };
#if JVET_O0119_BASE_PALETTE_444
- bool *runType[MAX_NUM_TBLOCKS] = { 0, };
+ PLTRunMode *runType[MAX_NUM_TBLOCKS] = {
+ 0,
+ };
Pel *runLength[MAX_NUM_TBLOCKS] = { 0, };
#endif
diff --git a/source/Lib/EncoderLib/EncModeCtrl.h b/source/Lib/EncoderLib/EncModeCtrl.h
index 32dabb5d2..6bc5de1f9 100644
--- a/source/Lib/EncoderLib/EncModeCtrl.h
+++ b/source/Lib/EncoderLib/EncModeCtrl.h
@@ -480,7 +480,7 @@ private:
TCoeff *m_pCoeff;
Pel *m_pPcmBuf;
#if JVET_O0119_BASE_PALETTE_444
- bool *m_runType;
+ PLTRunMode * m_runType;
Pel *m_runLength;
#endif
CodingStructure m_dummyCS;
diff --git a/source/Lib/EncoderLib/IntraSearch.cpp b/source/Lib/EncoderLib/IntraSearch.cpp
index d9c2b821a..654fedfed 100644
--- a/source/Lib/EncoderLib/IntraSearch.cpp
+++ b/source/Lib/EncoderLib/IntraSearch.cpp
@@ -1780,7 +1780,7 @@ void IntraSearch::PLTSearch(CodingStructure &cs, Partitioner& partitioner, Compo
}
Pel *runLength = tu.getRunLens (compBegin);
- bool *runType = tu.getRunTypes(compBegin);
+ PLTRunMode *runType = tu.getRunTypes(compBegin);
cu.lastPLTSize[compBegin] = cs.prevPLT.curPLTSize[compBegin];
//derive palette
derivePLTLossy(cs, partitioner, compBegin, numComp);
@@ -1796,7 +1796,7 @@ void IntraSearch::PLTSearch(CodingStructure &cs, Partitioner& partitioner, Compo
deriveRunAndCalcBits(cs, partitioner, compBegin, numComp, PLT_SCAN_VERTRAV, bits);
}
cu.useRotation[compBegin] = m_bestScanRotationMode;
- memcpy(runType, m_runTypeRD, sizeof(bool)*width*height);
+ memcpy(runType, m_runTypeRD, sizeof(PLTRunMode) * width * height);
memcpy(runLength, m_runLengthRD, sizeof(Pel)*width*height);
//reconstruct pixel
PelBuf curPLTIdx = tu.getcurPLTIdx(compBegin);
@@ -1878,7 +1878,7 @@ void IntraSearch::deriveRunAndCalcBits(CodingStructure& cs, Partitioner& partiti
uint32_t height = cu.block(compBegin).height;
uint32_t width = cu.block(compBegin).width;
Pel *runLength = tu.getRunLens (compBegin);
- bool *runType = tu.getRunTypes(compBegin);
+ PLTRunMode * runType = tu.getRunTypes(compBegin);
cu.useRotation[compBegin] = (pltScanMode == PLT_SCAN_VERTRAV);
m_scanOrder = g_scanOrder[SCAN_UNGROUPED][(cu.useRotation[compBegin]) ? SCAN_TRAV_VER : SCAN_TRAV_HOR][gp_sizeIdxInfo->idxFrom(width)][gp_sizeIdxInfo->idxFrom(height)];
--
GitLab