diff --git a/src/Benchmarks/SpMV/spmv.h b/src/Benchmarks/SpMV/spmv.h
index d8b20898310a1353cad1bced491410be620abc08..37b9bf7bbff15a28937af2c3b609b8ecb5a40bfe 100644
--- a/src/Benchmarks/SpMV/spmv.h
+++ b/src/Benchmarks/SpMV/spmv.h
@@ -40,11 +40,11 @@ using SlicedEllpackAlias = Matrices::SlicedEllpack< Real, Device, Index >;
std::string getMatrixFileName( const String& InputFileName )
{
std::string fileName = InputFileName;
-
+
const size_t last_slash_idx = fileName.find_last_of( "/\\" );
if( std::string::npos != last_slash_idx )
fileName.erase( 0, last_slash_idx + 1 );
-
+
return fileName;
}
@@ -55,7 +55,7 @@ std::string getMatrixFormat( const Matrix& matrix )
std::string mtrxFullType = getType( matrix );
std::string mtrxType = mtrxFullType.substr( 0, mtrxFullType.find( "<" ) );
std::string format = mtrxType.substr( mtrxType.find( ':' ) + 2 );
-
+
return format;
}
@@ -63,7 +63,7 @@ std::string getMatrixFormat( const Matrix& matrix )
template< typename Matrix >
void printMatrixInfo( const Matrix& matrix,
std::ostream& str )
-{
+{
str << "\n Format: " << getMatrixFormat( matrix ) << std::endl;
str << " Rows: " << matrix.getRows() << std::endl;
str << " Cols: " << matrix.getColumns() << std::endl;
@@ -81,67 +81,67 @@ benchmarkSpMV( Benchmark& benchmark,
// Setup CSR for cuSPARSE. It will compared to the format given as a template parameter to this function
typedef Matrices::CSR< Real, Devices::Host, int > CSR_HostMatrix;
typedef Matrices::CSR< Real, Devices::Cuda, int > CSR_DeviceMatrix;
-
+
CSR_HostMatrix CSRhostMatrix;
CSR_DeviceMatrix CSRdeviceMatrix;
-
+
// Read the matrix for CSR, to set up cuSPARSE
try
- {
+ {
if( ! MatrixReader< CSR_HostMatrix >::readMtxFile( inputFileName, CSRhostMatrix, verboseMR ) )
- {
+ {
throw std::bad_alloc();
return false;
}
}
- catch( std::bad_alloc e )
+ catch( std::bad_alloc& e )
{
e.what();
return false;
}
-
+
#ifdef HAVE_CUDA
// cuSPARSE handle setup
cusparseHandle_t cusparseHandle;
cusparseCreate( &cusparseHandle );
-
+
// cuSPARSE (in TNL's CSR) only works for device, copy the matrix from host to device
CSRdeviceMatrix = CSRhostMatrix;
-
+
// Delete the CSRhostMatrix, so it doesn't take up unnecessary space
CSRhostMatrix.reset();
-
+
// Initialize the cusparseCSR matrix.
TNL::CusparseCSR< Real > cusparseCSR;
cusparseCSR.init( CSRdeviceMatrix, &cusparseHandle );
#endif
-
+
// Setup the format which is given as a template parameter to this function
typedef Matrix< Real, Devices::Host, int > HostMatrix;
typedef Matrix< Real, Devices::Cuda, int > DeviceMatrix;
typedef Containers::Vector< Real, Devices::Host, int > HostVector;
typedef Containers::Vector< Real, Devices::Cuda, int > CudaVector;
-
+
HostMatrix hostMatrix;
DeviceMatrix deviceMatrix;
HostVector hostVector, hostVector2;
CudaVector deviceVector, deviceVector2;
-
+
// Load the format
try
- {
+ {
if( ! MatrixReader< HostMatrix >::readMtxFile( inputFileName, hostMatrix, verboseMR ) )
{
throw std::bad_alloc();
return false;
}
}
- catch( std::bad_alloc e )
+ catch( std::bad_alloc& e )
{
e.what();
return false;
}
-
+
// Setup MetaData here (not in tnl-benchmark-spmv.h, as done in Benchmarks/BLAS),
// because we need the matrix loaded first to get the rows and columns
@@ -191,17 +191,17 @@ benchmarkSpMV( Benchmark& benchmark,
benchmark.setOperation( datasetSize );
benchmark.time< Devices::Host >( reset, "CPU", spmvHost );
-
+
// Initialize the host vector to be compared.
// (The values in hostVector2 will be reset when spmvCuda starts)
HostVector resultHostVector2;
resultHostVector2.setSize( hostVector2.getSize() );
resultHostVector2.setValue( 0.0 );
-
+
// Copy the values
resultHostVector2 = hostVector2;
-
- // Setup cuSPARSE MetaData, since it has the same header as CSR,
+
+ // 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.
benchmark.setMetadataColumns( Benchmark::MetadataColumns({
@@ -211,7 +211,7 @@ benchmarkSpMV( Benchmark& benchmark,
{ "columns", convertToString( hostMatrix.getColumns() ) },
{ "matrix format", convertToString( "CSR-cuSPARSE" ) }
} ));
-
+
#ifdef HAVE_CUDA
benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );
@@ -220,51 +220,51 @@ benchmarkSpMV( Benchmark& benchmark,
HostVector resultDeviceVector2;
resultDeviceVector2.setSize( deviceVector2.getSize() );
resultDeviceVector2.setValue( 0.0 );
-
+
resultDeviceVector2 = deviceVector2;
-
+
benchmark.time< Devices::Cuda >( reset, "GPU", spmvCusparse );
-
+
HostVector resultcuSPARSEDeviceVector2;
resultcuSPARSEDeviceVector2.setSize( deviceVector2.getSize() );
resultcuSPARSEDeviceVector2.setValue( 0.0 );
-
+
resultcuSPARSEDeviceVector2 = deviceVector2;
-
+
// Difference between GPU (curent format) and GPU-cuSPARSE results
//Real cuSparseDifferenceAbsMax = resultDeviceVector2.differenceAbsMax( resultcuSPARSEDeviceVector2 );
Real cuSparseDifferenceAbsMax = max( abs( resultDeviceVector2 - resultcuSPARSEDeviceVector2 ) );
//Real cuSparseDifferenceLpNorm = resultDeviceVector2.differenceLpNorm( resultcuSPARSEDeviceVector2, 1 );
Real cuSparseDifferenceLpNorm = lpNorm( resultDeviceVector2 - resultcuSPARSEDeviceVector2, 1 );
-
+
std::string GPUxGPUcuSparse_resultDifferenceAbsMax = "GPUxGPUcuSPARSE differenceAbsMax = " + std::to_string( cuSparseDifferenceAbsMax );
std::string GPUxGPUcuSparse_resultDifferenceLpNorm = "GPUxGPUcuSPARSE differenceLpNorm = " + std::to_string( cuSparseDifferenceLpNorm );
-
+
char *GPUcuSparse_absMax = &GPUxGPUcuSparse_resultDifferenceAbsMax[ 0u ];
char *GPUcuSparse_lpNorm = &GPUxGPUcuSparse_resultDifferenceLpNorm[ 0u ];
-
-
+
+
// Difference between CPU and GPU results for the current format
//Real differenceAbsMax = resultHostVector2.differenceAbsMax( resultDeviceVector2 );
Real differenceAbsMax = max( abs( resultHostVector2 - resultDeviceVector2 ) );
//Real differenceLpNorm = resultHostVector2.differenceLpNorm( resultDeviceVector2, 1 );
Real differenceLpNorm = lpNorm( resultHostVector2 - resultDeviceVector2, 1 );
-
+
std::string CPUxGPU_resultDifferenceAbsMax = "CPUxGPU differenceAbsMax = " + std::to_string( differenceAbsMax );
std::string CPUxGPU_resultDifferenceLpNorm = "CPUxGPU differenceLpNorm = " + std::to_string( differenceLpNorm );
-
+
char *CPUxGPU_absMax = &CPUxGPU_resultDifferenceAbsMax[ 0u ];
char *CPUxGPU_lpNorm = &CPUxGPU_resultDifferenceLpNorm[ 0u ];
-
+
// Print result differences of CPU and GPU of current format
std::cout << CPUxGPU_absMax << std::endl;
std::cout << CPUxGPU_lpNorm << std::endl;
-
+
// Print result differences of GPU of current format and GPU with cuSPARSE.
std::cout << GPUcuSparse_absMax << std::endl;
std::cout << GPUcuSparse_lpNorm << std::endl;
#endif
-
+
std::cout << std::endl;
return true;
}
@@ -277,11 +277,11 @@ benchmarkSpmvSynthetic( Benchmark& benchmark,
bool verboseMR )
{
bool result = true;
- result |= benchmarkSpMV< Real, Matrices::CSR >( benchmark, inputFileName, verboseMR );
+ result |= benchmarkSpMV< Real, Matrices::CSR >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, Matrices::Ellpack >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, SlicedEllpackAlias >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, Matrices::ChunkedEllpack >( benchmark, inputFileName, verboseMR );
-
+
// AdEllpack is broken
// result |= benchmarkSpMV< Real, Matrices::AdEllpack >( benchmark, inputFileName, verboseMR );
result |= benchmarkSpMV< Real, Matrices::BiEllpack >( benchmark, inputFileName, verboseMR );