From a147f47a056ff726a12b2383607f10eb6d55841a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tom=C3=A1=C5=A1=20Oberhuber?= <oberhuber.tomas@gmail.com>
Date: Thu, 5 Mar 2020 22:11:57 +0100
Subject: [PATCH] Improving SpMV benchmark with benchmark results.
---
src/Benchmarks/SpMV/SpmvBenchmarkResult.h | 52 +++++
src/Benchmarks/SpMV/spmv.h | 234 ++++++++++------------
2 files changed, 154 insertions(+), 132 deletions(-)
create mode 100644 src/Benchmarks/SpMV/SpmvBenchmarkResult.h
diff --git a/src/Benchmarks/SpMV/SpmvBenchmarkResult.h b/src/Benchmarks/SpMV/SpmvBenchmarkResult.h
new file mode 100644
index 0000000000..699be6efd1
--- /dev/null
+++ b/src/Benchmarks/SpMV/SpmvBenchmarkResult.h
@@ -0,0 +1,52 @@
+/***************************************************************************
+ SpmvBenchmarkResult.h - description
+ -------------------
+ begin : Mar 5, 2020
+ copyright : (C) 2020 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include "../Benchmarks.h"
+
+namespace TNL {
+namespace Benchmarks {
+
+template< typename Real = double,
+ typename Index = int >
+struct SpmvBenchmarkResult
+: public BenchmarkResult
+{
+ using RealType = Real;
+ using IndexType = Index;
+ using HostVector = Containers::Vector< Real, Devices::Host, Index >;
+ using CudaVector = Containers::Vector< Real, Devices::Cuda, Index >;
+
+ SpmvBenchmarkResult( CudaVector& cudaResult, HostVector& hostResult, CudaVector& cusparseResult )
+ : hostResult( hostResult ), cudaResult( cudaResult), cusparseResult( cusparseResult ){};
+
+ virtual HeaderElements getTableHeader() const override
+ {
+ return HeaderElements({"time", "stddev", "stddev/time", "speedup", "Host.Diff.Max", "Host.Diff.L2", "Cusparse.Diff.Max", "Cusparse.Diff.L2"});
+ }
+
+ virtual RowElements getRowElements() const override
+ {
+ HostVector cudaCopy, cusparseCopy, a, b;
+ cudaCopy = cudaResult;
+ cusparseCopy = cusparseResult;
+ a = cudaCopy - hostResult;
+ b = cudaCopy - cusparseCopy;
+ return RowElements({ time, stddev, stddev/time, speedup, max( abs( a ) ), lpNorm( a, 2.0 ), max( abs( b ) ), lpNorm( b, 2.0 ) });
+ }
+
+ HostVector &hostResult;
+
+ CudaVector &cudaResult, &cusparseResult;
+};
+
+} //namespace Benchmarks
+} //namespace TNL
diff --git a/src/Benchmarks/SpMV/spmv.h b/src/Benchmarks/SpMV/spmv.h
index 3cd5c19d2a..8851e41143 100644
--- a/src/Benchmarks/SpMV/spmv.h
+++ b/src/Benchmarks/SpMV/spmv.h
@@ -15,6 +15,7 @@
#pragma once
#include "../Benchmarks.h"
+#include "SpmvBenchmarkResult.h"
#include <TNL/Pointers/DevicePointer.h>
#include <TNL/Matrices/Legacy/CSR.h>
@@ -113,168 +114,137 @@ 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
- typedef Matrices::Legacy::CSR< Real, Devices::Host, int > CSR_HostMatrix;
- typedef Matrices::Legacy::CSR< Real, Devices::Cuda, int > CSR_DeviceMatrix;
+ // Setup CSR for cuSPARSE. It will compared to the format given as a template parameter to this function
+ using CSR_HostMatrix = Matrices::Legacy::CSR< Real, Devices::Host, int >;
+ using CSR_DeviceMatrix = Matrices::Legacy::CSR< Real, Devices::Cuda, int >;
- CSR_HostMatrix CSRhostMatrix;
- CSR_DeviceMatrix CSRdeviceMatrix;
+ CSR_HostMatrix CSRhostMatrix;
+ CSR_DeviceMatrix CSRdeviceMatrix;
- // Read the matrix for CSR, to set up cuSPARSE
- MatrixReader< CSR_HostMatrix >::readMtxFile( inputFileName, CSRhostMatrix, verboseMR );
+ // Read the matrix for CSR, to set up cuSPARSE
+ MatrixReader< CSR_HostMatrix >::readMtxFile( inputFileName, CSRhostMatrix, verboseMR );
#ifdef HAVE_CUDA
- // cuSPARSE handle setup
- cusparseHandle_t cusparseHandle;
- cusparseCreate( &cusparseHandle );
+ // 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;
+ // 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();
+ // 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 );
+ // 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;
+ // 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;
+ HostMatrix hostMatrix;
+ DeviceMatrix deviceMatrix;
+ HostVector hostVector, hostVector2;
+ CudaVector deviceVector, deviceVector2, cusparseVector;
- // Load the format
- MatrixReader< HostMatrix >::readMtxFile( inputFileName, hostMatrix, verboseMR );
+ // Load the format
+ MatrixReader< HostMatrix >::readMtxFile( inputFileName, hostMatrix, verboseMR );
- // 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
- benchmark.setMetadataColumns( Benchmark::MetadataColumns({
- { "matrix name", convertToString( getMatrixFileName( inputFileName ) ) },
- { "non-zeros", convertToString( hostMatrix.getNumberOfNonzeroMatrixElements() ) },
- { "rows", convertToString( hostMatrix.getRows() ) },
- { "columns", convertToString( hostMatrix.getColumns() ) },
- { "matrix format", MatrixInfo< HostMatrix >::getFormat() } //convertToString( getType( hostMatrix ) ) }
- } ));
+ // 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
+ benchmark.setMetadataColumns( Benchmark::MetadataColumns({
+ { "matrix name", convertToString( getMatrixFileName( inputFileName ) ) },
+ { "non-zeros", convertToString( hostMatrix.getNumberOfNonzeroMatrixElements() ) },
+ { "rows", convertToString( hostMatrix.getRows() ) },
+ { "columns", convertToString( hostMatrix.getColumns() ) },
+ { "matrix format", MatrixInfo< HostMatrix >::getFormat() } //convertToString( getType( hostMatrix ) ) }
+ } ));
- hostVector.setSize( hostMatrix.getColumns() );
- hostVector2.setSize( hostMatrix.getRows() );
+ hostVector.setSize( hostMatrix.getColumns() );
+ hostVector2.setSize( hostMatrix.getRows() );
#ifdef HAVE_CUDA
- deviceMatrix = hostMatrix;
- deviceVector.setSize( hostMatrix.getColumns() );
- deviceVector2.setSize( hostMatrix.getRows() );
+ deviceMatrix = hostMatrix;
+ deviceVector.setSize( hostMatrix.getColumns() );
+ deviceVector2.setSize( hostMatrix.getRows() );
+ cusparseVector.setSize( hostMatrix.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 );
+ // reset function
+ auto resetHostVectors = [&]() {
+ hostVector = 1.0;
+ hostVector2 = 0.0;
+ };
+#ifdef HAVE_CUDA
+ auto resetCudaVectors = [&]() {
+ deviceVector = 1.0;
+ deviceVector2 = 0.0;
+ };
+ auto resetCusparseVectors = [&]() {
+ deviceVector = 1.0;
+ cusparseVector == 0.0;
+ };
#endif
- };
-
- const int elements = hostMatrix.getNumberOfNonzeroMatrixElements();
- const double datasetSize = (double) elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;
+ 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 spmvHost = [&]() {
+ hostMatrix.vectorProduct( hostVector, hostVector2 );
+ };
#ifdef HAVE_CUDA
- auto spmvCuda = [&]() {
- deviceMatrix.vectorProduct( deviceVector, deviceVector2 );
- };
+ auto spmvCuda = [&]() {
+ deviceMatrix.vectorProduct( deviceVector, deviceVector2 );
+ };
- auto spmvCusparse = [&]() {
- cusparseCSR.vectorProduct( deviceVector, deviceVector2 );
- };
+ auto spmvCusparse = [&]() {
+ cusparseCSR.vectorProduct( deviceVector, cusparseVector );
+ };
#endif
- benchmark.setOperation( datasetSize );
- benchmark.time< Devices::Host >( reset, "CPU", spmvHost );
+ benchmark.setOperation( datasetSize );
+ benchmark.time< Devices::Host >( resetHostVectors, "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 );
+ // 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;
+ // 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;
-
- // 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({
- { "matrix name", convertToString( getMatrixFileName( inputFileName ) ) },
- { "non-zeros", convertToString( hostMatrix.getNumberOfNonzeroMatrixElements() ) },
- { "rows", convertToString( hostMatrix.getRows() ) },
- { "columns", convertToString( hostMatrix.getColumns() ) },
- { "matrix format", convertToString( "CSR-cuSPARSE-" + getFormatShort( hostMatrix ) ) }
- } ));
-
- benchmark.time< Devices::Cuda >( reset, "GPU", spmvCusparse );
-
- HostVector resultcuSPARSEDeviceVector2;
- resultcuSPARSEDeviceVector2.setSize( deviceVector2.getSize() );
- resultcuSPARSEDeviceVector2.setValue( 0.0 );
-
- resultcuSPARSEDeviceVector2 = deviceVector2;
-
- // Difference between GPU (current 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;
+ benchmark.time< Devices::Cuda >( resetCudaVectors, "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;
+
+ // 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({
+ { "matrix name", convertToString( getMatrixFileName( inputFileName ) ) },
+ { "non-zeros", convertToString( hostMatrix.getNumberOfNonzeroMatrixElements() ) },
+ { "rows", convertToString( hostMatrix.getRows() ) },
+ { "columns", convertToString( hostMatrix.getColumns() ) },
+ { "matrix format", convertToString( "CSR-cuSPARSE-" + getFormatShort( hostMatrix ) ) }
+ } ));
+
+ SpmvBenchmarkResult< Real, int > benchmarkResult( deviceVector2, hostVector2, cusparseVector );
+ benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", spmvCusparse, benchmarkResult );
+
#endif
std::cout << std::endl;
--
GitLab