From 2b9409c876ba4c4f5d888dbc57137c12befa1bd7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tom=C3=A1=C5=A1=20Oberhuber?= <oberhuber.tomas@gmail.com>
Date: Tue, 26 Nov 2019 19:14:53 +0100
Subject: [PATCH] Fixes after rebase.
---
src/Benchmarks/SpMV/spmv.h | 38 +++++++++++++-------------
src/TNL/Matrices/AdEllpack.h | 8 +++---
src/TNL/Matrices/AdEllpack_impl.h | 34 ++++-------------------
src/TNL/Matrices/BiEllpack_impl.h | 35 +++---------------------
src/TNL/Matrices/CSR_impl.h | 4 +--
src/TNL/Matrices/ChunkedEllpack_impl.h | 5 ----
src/TNL/Matrices/Dense_impl.h | 22 ---------------
src/TNL/Matrices/Ellpack_impl.h | 18 ++++++------
src/TNL/Matrices/SlicedEllpack_impl.h | 4 +--
9 files changed, 43 insertions(+), 125 deletions(-)
diff --git a/src/Benchmarks/SpMV/spmv.h b/src/Benchmarks/SpMV/spmv.h
index 6a9dab96a4..b7579386e5 100644
--- a/src/Benchmarks/SpMV/spmv.h
+++ b/src/Benchmarks/SpMV/spmv.h
@@ -74,8 +74,8 @@ template< typename Real,
template< typename, typename, typename > class Matrix,
template< typename, typename, typename > class Vector = Containers::Vector >
bool
-benchmarkSpMV( Benchmark & benchmark,
- const String & inputFileName,
+benchmarkSpMV( Benchmark& benchmark,
+ const String& inputFileName,
bool verboseMR )
{
// Setup CSR for cuSPARSE. It will compared to the format given as a template parameter to this function
@@ -100,11 +100,11 @@ benchmarkSpMV( Benchmark & benchmark,
return false;
}
+#ifdef HAVE_CUDA
// cuSPARSE handle setup
cusparseHandle_t cusparseHandle;
cusparseCreate( &cusparseHandle );
-#ifdef HAVE_CUDA
// cuSPARSE (in TNL's CSR) only works for device, copy the matrix from host to device
CSRdeviceMatrix = CSRhostMatrix;
@@ -185,9 +185,11 @@ benchmarkSpMV( Benchmark & benchmark,
auto spmvCuda = [&]() {
deviceMatrix.vectorProduct( deviceVector, deviceVector2 );
};
+#ifdef HAVE_CUDA
auto spmvCusparse = [&]() {
cusparseCSR.vectorProduct( deviceVector, deviceVector2 );
};
+#endif
benchmark.setOperation( datasetSize );
benchmark.time< Devices::Host >( reset, "CPU", spmvHost );
@@ -201,18 +203,6 @@ benchmarkSpMV( Benchmark & benchmark,
// Copy the values
resultHostVector2 = hostVector2;
-#ifdef HAVE_CUDA
- benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );
-
- // Initialize the device vector to be compared.
- // (The values in deviceVector2 will be reset when spmvCusparse starts)
- HostVector resultDeviceVector2;
- resultDeviceVector2.setSize( deviceVector2.getSize() );
- resultDeviceVector2.setValue( 0.0 );
-
- resultDeviceVector2 = deviceVector2;
-#endif
-
// Setup cuSPARSE MetaData, since it has the same header as CSR,
// and therefore will not get its own headers (rows, cols, speedup etc.) in log.
// * Not setting this up causes (among other undiscovered errors) the speedup from CPU to GPU on the input format to be overwritten.
@@ -223,8 +213,18 @@ benchmarkSpMV( Benchmark & benchmark,
{ "columns", convertToString( hostMatrix.getColumns() ) },
{ "matrix format", convertToString( "CSR-cuSPARSE" ) }
} ));
-
+
#ifdef HAVE_CUDA
+ benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );
+
+ // Initialize the device vector to be compared.
+ // (The values in deviceVector2 will be reset when spmvCusparse starts)
+ HostVector resultDeviceVector2;
+ resultDeviceVector2.setSize( deviceVector2.getSize() );
+ resultDeviceVector2.setValue( 0.0 );
+
+ resultDeviceVector2 = deviceVector2;
+
benchmark.time< Devices::Cuda >( reset, "GPU", spmvCusparse );
HostVector resultcuSPARSEDeviceVector2;
@@ -232,7 +232,6 @@ benchmarkSpMV( Benchmark & benchmark,
resultcuSPARSEDeviceVector2.setValue( 0.0 );
resultcuSPARSEDeviceVector2 = deviceVector2;
- #endif
// Difference between GPU (curent format) and GPU-cuSPARSE results
Real cuSparseDifferenceAbsMax = resultDeviceVector2.differenceAbsMax( resultcuSPARSEDeviceVector2 );
@@ -243,6 +242,7 @@ benchmarkSpMV( Benchmark & benchmark,
char *GPUcuSparse_absMax = &GPUxGPUcuSparse_resultDifferenceAbsMax[ 0u ];
char *GPUcuSparse_lpNorm = &GPUxGPUcuSparse_resultDifferenceLpNorm[ 0u ];
+ #endif
// Difference between CPU and GPU results for the current format
@@ -270,14 +270,14 @@ benchmarkSpMV( Benchmark & benchmark,
template< typename Real = double,
typename Index = int >
bool
-benchmarkSpmvSynthetic( Benchmark & benchmark,
+benchmarkSpmvSynthetic( Benchmark& benchmark,
const String& inputFileName,
bool verboseMR )
{
bool result = true;
result |= benchmarkSpMV< Real, Matrices::CSR >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, Matrices::Ellpack >( benchmark, inputFileName, verboseMR );
- result |= benchmarkSpMV< Real, SlicedEllpack >( benchmark, inputFileName, verboseMR );
+ result |= benchmarkSpMV< Real, Matrices::SlicedEllpack >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, Matrices::ChunkedEllpack >( benchmark, inputFileName, verboseMR );
// AdEllpack is broken
diff --git a/src/TNL/Matrices/AdEllpack.h b/src/TNL/Matrices/AdEllpack.h
index 34b0819146..f011e6c804 100644
--- a/src/TNL/Matrices/AdEllpack.h
+++ b/src/TNL/Matrices/AdEllpack.h
@@ -195,7 +195,7 @@ public:
typename = typename Enabler< Device2 >::type >
AdEllpack& operator=( const AdEllpack< Real2, Device2, Index2 >& matrix );
- bool save( File& file ) const;
+ void save( File& file ) const;
void load( File& file );
@@ -207,13 +207,13 @@ public:
bool balanceLoad( const RealType average,
ConstCompressedRowLengthsVectorView rowLengths,
- warpList< ThisType >* list );
+ warpList< AdEllpack >* list );
void computeWarps( const IndexType SMs,
const IndexType threadsPerSM,
- warpList< ThisType >* list );
+ warpList< AdEllpack >* list );
- bool createArrays( warpList< ThisType >* list );
+ bool createArrays( warpList< AdEllpack >* list );
void performRowTest();
diff --git a/src/TNL/Matrices/AdEllpack_impl.h b/src/TNL/Matrices/AdEllpack_impl.h
index b01e9041e6..510c1e19b1 100644
--- a/src/TNL/Matrices/AdEllpack_impl.h
+++ b/src/TNL/Matrices/AdEllpack_impl.h
@@ -162,28 +162,6 @@ AdEllpack< Real, Device, Index >::AdEllpack()
warpSize( 32 )
{}
-template< typename Real,
- typename Device,
- typename Index >
-String AdEllpack< Real, Device, Index >::getTypeVirtual() const
-{
- return this->getType();
-}
-
-template< typename Real,
- typename Device,
- typename Index >
-String AdEllpack< Real, Device, Index >::getType()
-{
- return String( "Matrices::AdEllpack< ") +
- String( TNL::getType< Real >() ) +
- String( ", " ) +
- String( Device::getDeviceType() ) +
- String( ", " ) +
- String( TNL::getType< Index >() ) +
- String( " >" );
-}
-
template< typename Real,
typename Device,
typename Index >
@@ -204,7 +182,7 @@ setCompressedRowLengths( ConstCompressedRowLengthsVectorView rowLengths )
average /= ( RealType ) this->getRows();
this->totalLoad = average;
- warpList< ThisType >* list = new warpList< ThisType >();
+ warpList< AdEllpack >* list = new warpList< AdEllpack >();
if( !this->balanceLoad( average, rowLengths, list ) )
throw 0; // TODO: Make better exception
@@ -766,7 +744,7 @@ template< typename Real,
typename Index >
bool AdEllpack< Real, Device, Index >::balanceLoad( const RealType average,
ConstCompressedRowLengthsVectorView rowLengths,
- warpList< ThisType >* list )
+ warpList< AdEllpack >* list )
{
IndexType offset, rowOffset, localLoad, reduceMap[ 32 ];
@@ -882,10 +860,10 @@ template< typename Real,
typename Index >
void AdEllpack< Real, Device, Index >::computeWarps( const IndexType SMs,
const IndexType threadsPerSM,
- warpList< ThisType >* list )
+ warpList< AdEllpack >* list )
{
IndexType averageLoad = 0;
- warpInfo< ThisType >* temp = list->getHead()->next;
+ warpInfo< AdEllpack >* temp = list->getHead()->next;
while( temp/*->next*/ != list->getTail() )
{
@@ -918,7 +896,7 @@ void AdEllpack< Real, Device, Index >::computeWarps( const IndexType SMs,
template< typename Real,
typename Device,
typename Index >
-bool AdEllpack< Real, Device, Index >::createArrays( warpList< ThisType >* list )
+bool AdEllpack< Real, Device, Index >::createArrays( warpList< AdEllpack >* list )
{
IndexType length = list->getNumberOfWarps();
@@ -928,7 +906,7 @@ bool AdEllpack< Real, Device, Index >::createArrays( warpList< ThisType >* list
this->reduceMap.setSize( length * this->warpSize );
IndexType iteration = 0;
- warpInfo< ThisType >* warp = list->getHead()->next;
+ warpInfo< AdEllpack >* warp = list->getHead()->next;
while( warp != list->getTail() )
{
this->offset.setElement( iteration, warp->offset );
diff --git a/src/TNL/Matrices/BiEllpack_impl.h b/src/TNL/Matrices/BiEllpack_impl.h
index e20b5cd230..5a8f67d7fc 100644
--- a/src/TNL/Matrices/BiEllpack_impl.h
+++ b/src/TNL/Matrices/BiEllpack_impl.h
@@ -45,28 +45,6 @@ BiEllpack< Real, Device, Index >::BiEllpack()
logWarpSize( 5 )
{}
-template< typename Real,
- typename Device,
- typename Index >
-String BiEllpack< Real, Device, Index >::getType()
-{
- return String( "Matrices::BiEllpack< ") +
- String( TNL::getType< Real >() ) +
- String( ", " ) +
- String( Device :: getDeviceType() ) +
- String( ", " ) +
- String( TNL::getType< Index >() ) +
- String( " >" );
-}
-
-template< typename Real,
- typename Device,
- typename Index >
-String BiEllpack< Real, Device, Index >::getTypeVirtual() const
-{
- return this->getType();
-}
-
template< typename Real,
typename Device,
typename Index >
@@ -715,18 +693,13 @@ BiEllpack< Real, Device, Index >::operator=( const BiEllpack< Real2, Device2, In
this->virtualRows = matrix.virtualRows;
this->rowPermArray = matrix.rowPermArray;
this->groupPointers = matrix.groupPointers;
-
- if( std::is_same< Device, Devices::MIC >::value ) {
- throw std::runtime_error("Not Implemented yet for MIC");
- }
-
return *this;
}
template< typename Real,
typename Device,
typename Index >
-bool BiEllpack< Real, Device, Index >::save( File& file ) const
+void BiEllpack< Real, Device, Index >::save( File& file ) const
{
Sparse< Real, Device, Index >::save( file );
file << this->groupPointers << this->rowPermArray;
@@ -735,7 +708,7 @@ bool BiEllpack< Real, Device, Index >::save( File& file ) const
template< typename Real,
typename Device,
typename Index >
-bool BiEllpack< Real, Device, Index >::load( File& file )
+void BiEllpack< Real, Device, Index >::load( File& file )
{
Sparse< Real, Device, Index >::load( file );
file >> this->groupPointers >> this->rowPermArray;
@@ -744,7 +717,7 @@ bool BiEllpack< Real, Device, Index >::load( File& file )
template< typename Real,
typename Device,
typename Index >
-bool BiEllpack< Real, Device, Index >::save( const String& fileName ) const
+void BiEllpack< Real, Device, Index >::save( const String& fileName ) const
{
Object::save( fileName );
}
@@ -752,7 +725,7 @@ bool BiEllpack< Real, Device, Index >::save( const String& fileName ) const
template< typename Real,
typename Device,
typename Index >
-bool BiEllpack< Real, Device, Index >::load( const String& fileName )
+void BiEllpack< Real, Device, Index >::load( const String& fileName )
{
Object::load( fileName );
}
diff --git a/src/TNL/Matrices/CSR_impl.h b/src/TNL/Matrices/CSR_impl.h
index 3164a7fff3..db31d6dcde 100644
--- a/src/TNL/Matrices/CSR_impl.h
+++ b/src/TNL/Matrices/CSR_impl.h
@@ -45,9 +45,7 @@ String CSR< Real, Device, Index >::getSerializationType()
{
return String( "Matrices::CSR< ") +
TNL::getType< Real>() +
- String( ", " ) +
- String( Device :: getDeviceType() ) +
- String( ", " ) +
+ ", [any_device], " +
String( TNL::getType< Index >() ) +
String( " >" );
}
diff --git a/src/TNL/Matrices/ChunkedEllpack_impl.h b/src/TNL/Matrices/ChunkedEllpack_impl.h
index 3826a8574f..9752ee4316 100644
--- a/src/TNL/Matrices/ChunkedEllpack_impl.h
+++ b/src/TNL/Matrices/ChunkedEllpack_impl.h
@@ -1278,11 +1278,6 @@ ChunkedEllpack< Real, Device, Index >::operator=( const ChunkedEllpack< Real2, D
}
}
}
-
- if( std::is_same< Device, Devices::MIC >::value ) {
- throw std::runtime_error("Not Implemented yet for MIC");
- }
-
return *this;
}
diff --git a/src/TNL/Matrices/Dense_impl.h b/src/TNL/Matrices/Dense_impl.h
index f690946e88..246bd09edb 100644
--- a/src/TNL/Matrices/Dense_impl.h
+++ b/src/TNL/Matrices/Dense_impl.h
@@ -24,28 +24,6 @@ Dense< Real, Device, Index >::Dense()
{
}
-template< typename Real,
- typename Device,
- typename Index >
-String Dense< Real, Device, Index >::getType()
-{
- return String( "Matrices::Dense< " ) +
- String( TNL::getType< Real >() ) +
- String( ", " ) +
- String( Device :: getDeviceType() ) +
- String( ", " ) +
- String( TNL::getType< Index >() ) +
- String( " >" );
-}
-
-template< typename Real,
- typename Device,
- typename Index >
-String Dense< Real, Device, Index >::getTypeVirtual() const
-{
- return this->getType();
-}
-
template< typename Real,
typename Device,
typename Index >
diff --git a/src/TNL/Matrices/Ellpack_impl.h b/src/TNL/Matrices/Ellpack_impl.h
index b99dbc88b5..5ae12f4087 100644
--- a/src/TNL/Matrices/Ellpack_impl.h
+++ b/src/TNL/Matrices/Ellpack_impl.h
@@ -16,7 +16,7 @@
#include <TNL/Exceptions/NotImplementedError.h>
namespace TNL {
-namespace Matrices {
+namespace Matrices {
template< typename Real,
typename Device,
@@ -33,9 +33,7 @@ String Ellpack< Real, Device, Index >::getSerializationType()
{
return String( "Matrices::Ellpack< " ) +
String( TNL::getType< Real >() ) +
- String( ", " ) +
- String( Device :: getDeviceType() ) +
- String( ", " ) +
+ ", [any device], " +
getType< Index >() +
String( " >" );
}
@@ -59,21 +57,21 @@ void Ellpack< Real, Device, Index >::setDimensions( const IndexType rows,
<< " columns = " << columns << std::endl );
this->rows = rows;
this->columns = columns;
-
+
if( std::is_same< Device, Devices::Cuda >::value )
{
- this->alignedRows = roundToMultiple( columns, Devices::Cuda::getWarpSize() );
+ this->alignedRows = roundToMultiple( columns, Cuda::getWarpSize() );
if( this->rows - this->alignedRows > 0 )
{
IndexType missingRows = this->rows - this->alignedRows;
-
- missingRows = roundToMultiple( missingRows, Devices::Cuda::getWarpSize() );
+
+ missingRows = roundToMultiple( missingRows, Cuda::getWarpSize() );
this->alignedRows += missingRows;
}
}
else this->alignedRows = rows;
-
+
if( this->rowLengths != 0 )
allocateElements();
}
@@ -87,7 +85,7 @@ void Ellpack< Real, Device, Index >::setCompressedRowLengths( ConstCompressedRow
TNL_ASSERT_GT( this->getColumns(), 0, "cannot set row lengths of an empty matrix" );
TNL_ASSERT_EQ( this->getRows(), rowLengths.getSize(), "wrong size of the rowLengths vector" );
- this->rowLengths = this->maxRowLength = rowLengths.max();
+ this->rowLengths = this->maxRowLength = max( rowLengths );
allocateElements();
}
diff --git a/src/TNL/Matrices/SlicedEllpack_impl.h b/src/TNL/Matrices/SlicedEllpack_impl.h
index 772360c8c0..8c629b563c 100644
--- a/src/TNL/Matrices/SlicedEllpack_impl.h
+++ b/src/TNL/Matrices/SlicedEllpack_impl.h
@@ -34,9 +34,7 @@ String SlicedEllpack< Real, Device, Index, SliceSize >::getSerializationType()
{
return String( "Matrices::SlicedEllpack< ") +
TNL::getType< Real >() +
- String( ", " ) +
- String( Device :: getDeviceType() ) +
- String( ", " ) +
+ ", [any_device], " +
String( TNL::getType< Index >() ) +
String( " >" );
}
--
GitLab