Commit 92bb20c0 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber Committed by Jakub Klinkovský
Browse files

Added LightSpMV Warp kernel benchmark.

parent 751c3d1c

src/Benchmarks/SpMV/ReferenceFormats/LightSpMVBenchmark.h

+5 −0
Original line number Diff line number Diff line
@@ -126,6 +126,11 @@ struct LightSpMVBenchmark 
#endif

   }



   void setKernelType( LightSpMVBenchmarkKernelType type )

   {

      this->kernelType = type;

   }



   void resetVectors()

   {

      this->inVectorView = 1.0;

src/Benchmarks/SpMV/spmv.h

+21 −12
Original line number Diff line number Diff line
@@ -39,9 +39,9 @@ 
#include <TNL/Algorithms/Segments/BiEllpack.h>



// Comment the following to turn off some groups of SpMV benchmarks and speed-up the compilation

//#define WITH_TNL_BENCHMARK_SPMV_GENERAL_MATRICES

//#define WITH_TNL_BENCHMARK_SPMV_SYMMETRIC_MATRICES

//#define WITH_TNL_BENCHMARK_SPMV_LEGACY_FORMATS

#define WITH_TNL_BENCHMARK_SPMV_GENERAL_MATRICES

#define WITH_TNL_BENCHMARK_SPMV_SYMMETRIC_MATRICES

#define WITH_TNL_BENCHMARK_SPMV_LEGACY_FORMATS



// Uncomment the following line to enable benchmarking the sandbox sparse matrix.

//#define WITH_TNL_BENCHMARK_SPMV_SANDBOX_MATRIX
@@ -374,7 +374,7 @@ benchmarkSpMV( Benchmark& benchmark, 
template< typename Real = double,

          typename Index = int >

void

benchmarkSpmvSynthetic( Benchmark& benchmark,

benchmarkSpmv( Benchmark& benchmark,

               const String& inputFileName,

               const Config::ParameterContainer& parameters,

               bool verboseMR )
@@ -469,8 +469,8 @@ benchmarkSpmvSynthetic( Benchmark& benchmark, 
       cusparseMatrix.vectorProduct( cudaInVector, cudaOutVector );

   };



   SpmvBenchmarkResult< Real, Devices::Host, int > cusparseBenchmarkResults( hostOutVector, hostOutVector, csrHostMatrix.getNonzeroElementsCount() );

   benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", spmvCusparse, cusparseBenchmarkResults );

   SpmvBenchmarkResult< Real, Devices::Host, int > cudaBenchmarkResults( hostOutVector, hostOutVector, csrHostMatrix.getNonzeroElementsCount() );

   benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", spmvCusparse, cudaBenchmarkResults );



#ifdef HAVE_CSR5

   ////
@@ -489,7 +489,7 @@ benchmarkSpmvSynthetic( Benchmark& benchmark, 
   auto csr5SpMV = [&]() {

       csr5Benchmark.vectorProduct();

   };

   benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", csr5SpMV, cusparseBenchmarkResults );

   benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", csr5SpMV, cudaBenchmarkResults );

   std::cerr << "CSR5 error = " << max( abs( cudaOutVector - cudaOutVector2 ) ) << std::endl;

   csrCudaMatrix.reset();

#endif
@@ -501,7 +501,7 @@ benchmarkSpmvSynthetic( Benchmark& benchmark, 
      { "matrix name", convertToString( inputFileName ) },

      { "rows", convertToString( csrHostMatrix.getRows() ) },

      { "columns", convertToString( csrHostMatrix.getColumns() ) },

      { "matrix format", String( "LightSpMV" ) }

      { "matrix format", String( "LightSpMV Vector" ) }

   } ));



   LightSpMVCSRHostMatrix lightSpMVCSRHostMatrix;
@@ -514,7 +514,16 @@ benchmarkSpmvSynthetic( Benchmark& benchmark, 
   auto spmvLightSpMV = [&]() {

       lightSpMVBenchmark.vectorProduct();

   };

   benchmark.time< Devices::Cuda >( resetLightSpMVVectors, "GPU", spmvLightSpMV, cusparseBenchmarkResults );

   benchmark.time< Devices::Cuda >( resetLightSpMVVectors, "GPU", spmvLightSpMV, cudaBenchmarkResults );



   benchmark.setMetadataColumns( Benchmark::MetadataColumns({

      { "matrix name", convertToString( inputFileName ) },

      { "rows", convertToString( csrHostMatrix.getRows() ) },

      { "columns", convertToString( csrHostMatrix.getColumns() ) },

      { "matrix format", String( "LightSpMV Warp" ) }

   } ));

   lightSpMVBenchmark.setKernelType( LightSpMVBenchmarkKernelWarp );

   benchmark.time< Devices::Cuda >( resetLightSpMVVectors, "GPU", spmvLightSpMV, cudaBenchmarkResults );

#endif

   csrHostMatrix.reset();

src/Benchmarks/SpMV/tnl-benchmark-spmv.h

+1 −1
Original line number Diff line number Diff line
@@ -45,7 +45,7 @@ runSpMVBenchmarks( Benchmark & benchmark, 
                           metadata );

   // Start the actual benchmark in spmv.h

   try {

      SpMVLegacy::benchmarkSpmvSynthetic< Real >( benchmark, inputFileName, parameters, verboseMR );

      SpMVLegacy::benchmarkSpmv< Real >( benchmark, inputFileName, parameters, verboseMR );

   }

   catch( const std::exception& ex ) {

      std::cerr << ex.what() << std::endl;