Commit 0c6f26e5 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Programming SPMV benchmark.

parent 8b80f8f3

tests/benchmarks/tnl-benchmark-spmv.h

+105 −50
Original line number Diff line number Diff line
@@ -130,7 +130,7 @@ double computeThroughput( const long int nonzeroElements, 


template< typename Matrix,

          typename Vector >

void benchmarkHostMatrix( const Matrix& matrix,

void benchmarkMatrix( const Matrix& matrix,

                      const Vector& x,

                      Vector& b,

                      const long int nonzeroElements,
@@ -160,6 +160,12 @@ void benchmarkHostMatrix( const Matrix& matrix, 
   logFile << "  " << throughput << endl;

}



void writeTestFailed( fstream& logFile )

{

   logFile << "N/A" << endl;

   logFile << "N/A" << endl;

}



template< typename Real >

bool setupBenchmark( const tnlParameterContainer& parameters )

{
@@ -213,7 +219,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
      rowLengthsCuda.setSize( csrMatrix.getRows() );

      rowLengthsCuda = rowLengthsHost;

#endif

      benchmarkHostMatrix( csrMatrix,

      benchmarkMatrix( csrMatrix,

                       hostX,

                       hostB,

                       nonzeroElements,
@@ -226,7 +232,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
      typedef tnlEllpackMatrix< Real, tnlHost, int > EllpackMatrixType;

      EllpackMatrixType ellpackMatrix;

      ellpackMatrix.copyFrom( csrMatrix, rowLengthsHost );

      benchmarkHostMatrix( ellpackMatrix,

      benchmarkMatrix( ellpackMatrix,

                       hostX,

                       hostB,

                       nonzeroElements,
@@ -241,10 +247,12 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
      if( ! cudaEllpackMatrix.copyFrom( ellpackMatrix, rowLengthsCuda ) )

      {

         cerr << "I am not able to transfer the matrix on GPU." << endl;

         return false;

         writeTestFailed( logFile );

      }

      else

      {

         cout << " done." << endl;

      benchmarkHostMatrix( cudaEllpackMatrix,

         benchmarkMatrix( cudaEllpackMatrix,

                          cudaX,

                          cudaB,

                          nonzeroElements,
@@ -252,6 +260,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
                          stopTime,

                          verbose,

                          logFile );

      }

      cudaEllpackMatrix.reset();

#endif

      ellpackMatrix.reset();
@@ -259,7 +268,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
      typedef tnlSlicedEllpackMatrix< Real, tnlHost, int > SlicedEllpackMatrixType;

      SlicedEllpackMatrixType slicedEllpackMatrix;

      slicedEllpackMatrix.copyFrom( csrMatrix, rowLengthsHost );

      benchmarkHostMatrix( slicedEllpackMatrix,

      benchmarkMatrix( slicedEllpackMatrix,

                       hostX,

                       hostB,

                       nonzeroElements,
@@ -267,12 +276,35 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
                       stopTime,

                       verbose,

                       logFile );

#ifdef HAVE_CUDA

      typedef tnlSlicedEllpackMatrix< Real, tnlCuda, int > SlicedEllpackMatrixCudaType;

      SlicedEllpackMatrixCudaType cudaSlicedEllpackMatrix;

      cout << "Copying matrix to GPU... ";

      if( ! cudaSlicedEllpackMatrix.copyFrom( slicedEllpackMatrix, rowLengthsCuda ) )

      {

         cerr << "I am not able to transfer the matrix on GPU." << endl;

         writeTestFailed( logFile );

      }

      else

      {

         cout << " done." << endl;

         benchmarkMatrix( cudaSlicedEllpackMatrix,

                          cudaX,

                          cudaB,

                          nonzeroElements,

                          "SlicedEllpack Cuda",

                          stopTime,

                          verbose,

                          logFile );

      }

      cudaSlicedEllpackMatrix.reset();

#endif

      slicedEllpackMatrix.reset();



      typedef tnlChunkedEllpackMatrix< Real, tnlHost, int > ChunkedEllpackMatrixType;

      ChunkedEllpackMatrixType chunkedEllpackMatrix;

      chunkedEllpackMatrix.copyFrom( csrMatrix, rowLengthsHost );

      benchmarkHostMatrix( chunkedEllpackMatrix,

      benchmarkMatrix( chunkedEllpackMatrix,

                       hostX,

                       hostB,

                       nonzeroElements,
@@ -280,6 +312,29 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
                       stopTime,

                       verbose,

                       logFile );

#ifdef HAVE_CUDA

      typedef tnlChunkedEllpackMatrix< Real, tnlCuda, int > ChunkedEllpackMatrixCudaType;

      ChunkedEllpackMatrixCudaType cudaChunkedEllpackMatrix;

      cout << "Copying matrix to GPU... ";

      if( ! cudaChunkedEllpackMatrix.copyFrom( chunkedEllpackMatrix, rowLengthsCuda ) )

      {

         cerr << "I am not able to transfer the matrix on GPU." << endl;

         writeTestFailed( logFile );

      }

      else

      {

         cout << " done." << endl;

         benchmarkMatrix( cudaChunkedEllpackMatrix,

                          cudaX,

                          cudaB,

                          nonzeroElements,

                          "ChunkedEllpack Cuda",

                          stopTime,

                          verbose,

                          logFile );

      }

      cudaChunkedEllpackMatrix.reset();

#endif

      chunkedEllpackMatrix.reset();

   }

}