diff --git a/src/Benchmarks/SpMV/spmv.h b/src/Benchmarks/SpMV/spmv.h
index 7f47cf2518e37e8d5ecd4f70bb1e04483e05ddbb..883ea084f41defc939ff71cf0fc02580f0d7aeb0 100644
--- a/src/Benchmarks/SpMV/spmv.h
+++ b/src/Benchmarks/SpMV/spmv.h
@@ -20,7 +20,8 @@
#include <TNL/Matrices/SlicedEllpack.h>
#include <TNL/Matrices/ChunkedEllpack.h>
-//#include <TNL/Matrices/MatrixReader.h>
+#include <TNL/Matrices/MatrixReader.h>
+using namespace TNL::Matrices;
namespace TNL {
namespace Benchmarks {
@@ -29,16 +30,6 @@ namespace Benchmarks {
template< typename Real, typename Device, typename Index >
using SlicedEllpack = Matrices::SlicedEllpack< Real, Device, Index >;
-//template< typename Matrix >
-//void printMatrixInfo( const String& inputFileName,
-// const Matrix& matrix,
-// std::ostream& str )
-//{
-// str << " Rows: " << std::setw( 8 ) << matrix.getRows();
-// str << " Columns: " << std::setw( 8 ) << matrix.getColumns();
-// str << " Nonzero Elements: " << std::setw( 10 ) << matrix.getNumberOfNonzeroMatrixElements();
-//}
-
template< typename Matrix >
int setHostTestMatrix( Matrix& matrix,
const int elementsPerRow )
@@ -106,106 +97,75 @@ template< typename Real,
template< typename, typename, typename > class Vector = Containers::Vector >
bool
benchmarkSpMV( Benchmark & benchmark,
- const int & size,
- const int elementsPerRow = 5 )
+ const String & inputFileName )
{
- 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;
- Containers::Vector< int, Devices::Host, int > hostRowLengths;
- Containers::Vector< int, Devices::Cuda, int > deviceRowLengths;
- HostVector hostVector, hostVector2;
- CudaVector deviceVector, deviceVector2;
-
- // create benchmark group
- const std::vector< String > parsedType = parseObjectType( HostMatrix::getType() );
-#ifdef HAVE_CUDA
- benchmark.createHorizontalGroup( parsedType[ 0 ], 2 );
-#else
- benchmark.createHorizontalGroup( parsedType[ 0 ], 1 );
-#endif
-
- hostRowLengths.setSize( size );
- hostMatrix.setDimensions( size, size );
- hostVector.setSize( size );
- hostVector2.setSize( size );
-#ifdef HAVE_CUDA
- deviceRowLengths.setSize( size );
- deviceMatrix.setDimensions( size, size );
- deviceVector.setSize( size );
- deviceVector2.setSize( size );
-#endif
-
- hostRowLengths.setValue( elementsPerRow );
-#ifdef HAVE_CUDA
- deviceRowLengths.setValue( elementsPerRow );
-#endif
-
- hostMatrix.setCompressedRowLengths( hostRowLengths );
-#ifdef HAVE_CUDA
- deviceMatrix.setCompressedRowLengths( deviceRowLengths );
-#endif
-
- const int elements = setHostTestMatrix< HostMatrix >( hostMatrix, elementsPerRow );
- setCudaTestMatrix< DeviceMatrix >( deviceMatrix, elementsPerRow );
- const double datasetSize = (double) elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;
-
- // reset function
- auto reset = [&]() {
- hostVector.setValue( 1.0 );
- hostVector2.setValue( 0.0 );
-#ifdef HAVE_CUDA
- deviceVector.setValue( 1.0 );
- deviceVector2.setValue( 0.0 );
-#endif
- };
-
- // compute functions
- auto spmvHost = [&]() {
- hostMatrix.vectorProduct( hostVector, hostVector2 );
- };
- auto spmvCuda = [&]() {
- deviceMatrix.vectorProduct( deviceVector, deviceVector2 );
- };
-
- benchmark.setOperation( datasetSize );
- benchmark.time< Devices::Host >( reset, "CPU", spmvHost );
-#ifdef HAVE_CUDA
- benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );
-#endif
-
- return true;
+ 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;
+
+ if( ! MatrixReader< HostMatrix >::readMtxFile(inputFileName, hostMatrix ) )
+ std::cerr << "I am not able to read the matrix file " << inputFileName << "." << std::endl;
+ else
+ {
+ #ifdef HAVE_CUDA
+ if( ! MatrixReader< DeviceMatrix >::readMtxFile(inputFileName, deviceMatrix ) )
+ std::cerr << "I am not able to read the matrix file " << inputFileName << "." << std::endl;
+ #endif
+
+ hostVector.setSize( hostMatrix.getColumns() );
+ hostVector2.setSize( hostMatrix.getRows() );
+
+ #ifdef HAVE_CUDA
+ deviceVector.setSize( deviceMatrix.getColumns() );
+ deviceVector2.setSize( deviceMatrix.getRows() );
+ #endif
+
+ // reset function
+ auto reset = [&]() {
+ hostVector.setValue( 1.0 );
+ hostVector2.setValue( 0.0 );
+ #ifdef HAVE_CUDA
+ deviceVector.setValue( 1.0 );
+ deviceVector2.setValue( 0.0 );
+ #endif
+ };
+
+ const int elements = hostMatrix.getNumberOfNonzeroMatrixElements();
+
+ const double datasetSize = (double) elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;
+
+ // compute functions
+ auto spmvHost = [&]() {
+ hostMatrix.vectorProduct( hostVector, hostVector2 );
+ };
+ auto spmvCuda = [&]() {
+ deviceMatrix.vectorProduct( deviceVector, deviceVector2 );
+ };
+
+ benchmark.setOperation( datasetSize );
+ benchmark.time< Devices::Host >( reset, "CPU", spmvHost );
+ #ifdef HAVE_CUDA
+ benchmark.time< Devices::Cuda >( reset, "GPU", spmvCuda );
+ #endif
+ return true;
+ }
}
template< typename Real = double,
typename Index = int >
bool
benchmarkSpmvSynthetic( Benchmark & benchmark,
- const int & size,
- const int & elementsPerRow )
+ const String& inputFileName )
{
-// typedef Matrices::CSR< Real, Devices::Host, int > CSRType;
-// CSRType csrMatrix;
-// try
-// {
-// if( ! MatrixReader< CSRType >::readMtxFile( inputFileName, csrMatrix ) )
-// {
-// std::cerr << "I am not able to read the matrix file " << inputFileName << "." << std::endl;
-// return false;
-// }
-// }
-// catch( std::bad_alloc )
-// {
-// std::cerr << "Not enough memory to read the matrix." << std::endl;
-// return false;
-// }
bool result = true;
// TODO: benchmark all formats from tnl-benchmark-spmv (different parameters of the base formats)
- result |= benchmarkSpMV< Real, Matrices::CSR >( benchmark, size, elementsPerRow );
+ result |= benchmarkSpMV< Real, Matrices::CSR >( benchmark, inputFileName );
// result |= benchmarkSpMV< Real, Matrices::Ellpack >( benchmark, size, elementsPerRow );
// result |= benchmarkSpMV< Real, SlicedEllpack >( benchmark, size, elementsPerRow );
// result |= benchmarkSpMV< Real, Matrices::ChunkedEllpack >( benchmark, size, elementsPerRow );
diff --git a/src/Benchmarks/SpMV/tnl-benchmark-spmv.h b/src/Benchmarks/SpMV/tnl-benchmark-spmv.h
index b3ad3102e1da327585fcfb2d50562b68df834427..4493dd4cab921dc8bc426b7fb3ee21f359079608 100644
--- a/src/Benchmarks/SpMV/tnl-benchmark-spmv.h
+++ b/src/Benchmarks/SpMV/tnl-benchmark-spmv.h
@@ -21,43 +21,51 @@
#include <Benchmarks/BLAS/vector-operations.h>
#include "spmv.h"
+#include <TNL/Matrices/MatrixReader.h>
+using namespace TNL::Matrices;
+
using namespace TNL;
using namespace TNL::Benchmarks;
+//template< typename Matrix >
+//void printMatrixInfo( const String& inputFileName,
+// const Matrix& matrix,
+// std::ostream& str )
+//{
+// str << " Rows: " << std::setw( 8 ) << matrix.getRows();
+// str << " Columns: " << std::setw( 8 ) << matrix.getColumns();
+// str << " Nonzero Elements: " << std::setw( 10 ) << matrix.getNumberOfNonzeroMatrixElements();
+//}
template< typename Real >
void
runSpMVBenchmarks( Benchmark & benchmark,
Benchmark::MetadataMap metadata,
- const std::size_t & size,
- const int & elementsPerRow )
+ const String & inputFileName )
{
- const String precision = getType< Real >();
- metadata["precision"] = precision;
-
- // Array operations
- benchmark.newBenchmark( String("Array operations (") + precision + ")",
- metadata );
- benchmark.setMetadataColumns( Benchmark::MetadataColumns({
- { "size", convertToString( size ) }, } ));
- benchmarkArrayOperations< Real >( benchmark, size );
-
- // Vector operations
- benchmark.newBenchmark( String("Vector operations (") + precision + ")",
- metadata );
- benchmark.setMetadataColumns( Benchmark::MetadataColumns({
- { "size", convertToString( size ) }, } ));
- benchmarkVectorOperations< Real >( benchmark, size );
-
- // Sparse matrix-vector multiplication
- benchmark.newBenchmark( String("Sparse matrix-vector multiplication (") + precision + ")",
- metadata );
- benchmark.setMetadataColumns( Benchmark::MetadataColumns({
- { "rows", convertToString( size ) },
- { "columns", convertToString( size ) },
- { "elements per row", convertToString( elementsPerRow ) },
- } ));
- benchmarkSpmvSynthetic< Real >( benchmark, size, elementsPerRow );
+ // DO: get rows and cols from inputFileName (/TNL/Matrices/MatrixReader_impl.h)
+
+ typedef Matrices::CSR< Real, Devices::Host, int > CSRType;
+ CSRType csrMatrix;
+
+ if( ! MatrixReader< CSRType >::readMtxFile( inputFileName, csrMatrix ) )
+ std::cerr << "I am not able to read the matrix file " << inputFileName << "." << std::endl;
+ else
+ {
+ const std::size_t rows = csrMatrix.getRows();
+ const std::size_t cols = csrMatrix.getColumns();
+ const String precision = getType< Real >();
+ metadata["precision"] = precision;
+
+ // Sparse matrix-vector multiplication
+ benchmark.newBenchmark( String("Sparse matrix-vector multiplication (") + precision + ")",
+ metadata );
+ benchmark.setMetadataColumns( Benchmark::MetadataColumns({
+ { "rows", convertToString( rows ) },
+ { "columns", convertToString( cols ) }
+ } ));
+ benchmarkSpmvSynthetic< Real >( benchmark, inputFileName );
+ }
}
void
@@ -73,9 +81,7 @@ setupConfig( Config::ConfigDescription & config )
config.addEntryEnum( "float" );
config.addEntryEnum( "double" );
config.addEntryEnum( "all" );
- config.addEntry< int >( "size", "Size of arrays/vectors used in the benchmark.", 100000 );
config.addEntry< int >( "loops", "Number of iterations for every computation.", 10 );
- config.addEntry< int >( "elements-per-row", "Number of elements per row of the sparse matrix used in the matrix-vector multiplication benchmark.", 5 );
config.addEntry< int >( "verbose", "Verbose mode.", 1 );
config.addDelimiter( "Device settings:" );
@@ -100,6 +106,7 @@ main( int argc, char* argv[] )
! Devices::Cuda::setup( parameters ) )
return EXIT_FAILURE;
+ const String & inputFileName = parameters.getParameter< String >( "input-file" );
const String & logFileName = parameters.getParameter< String >( "log-file" );
const String & outputMode = parameters.getParameter< String >( "output-mode" );
const String & precision = parameters.getParameter< String >( "precision" );
@@ -108,9 +115,7 @@ main( int argc, char* argv[] )
// to pass 64-bit integer values
// const std::size_t minSize = parameters.getParameter< std::size_t >( "min-size" );
// const std::size_t maxSize = parameters.getParameter< std::size_t >( "max-size" );
- const std::size_t size = parameters.getParameter< int >( "size" );
const int loops = parameters.getParameter< int >( "loops" );
- const int elementsPerRow = parameters.getParameter< int >( "elements-per-row" );
const int verbose = parameters.getParameter< int >( "verbose" );
// open log file
@@ -124,16 +129,19 @@ main( int argc, char* argv[] )
// prepare global metadata
Benchmark::MetadataMap metadata = getHardwareMetadata();
-
+
+
+ // DO: Pass the inputFileName parameter and get rows and cols from it to create the cout GUI.
if( precision == "all" || precision == "float" )
- runSpMVBenchmarks< float >( benchmark, metadata, size, elementsPerRow );
+ runSpMVBenchmarks< float >( benchmark, metadata, inputFileName );
if( precision == "all" || precision == "double" )
- runSpMVBenchmarks< double >( benchmark, metadata, size, elementsPerRow );
+ runSpMVBenchmarks< double >( benchmark, metadata, inputFileName );
if( ! benchmark.save( logFile ) ) {
std::cerr << "Failed to write the benchmark results to file '" << parameters.getParameter< String >( "log-file" ) << "'." << std::endl;
return EXIT_FAILURE;
}
+ std::cout << "== BENCHMARK FINISHED ==" << std::endl;
return EXIT_SUCCESS;
}