Commit 7b7a4f94 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Added cuBLAS benchmarks for scalar product

parent 88ac33c7

tests/benchmarks/CMakeLists.txt

+3 −0
Original line number Diff line number Diff line
@@ -2,6 +2,9 @@ ADD_SUBDIRECTORY( share ) 


IF( BUILD_CUDA )

    CUDA_ADD_EXECUTABLE( tnl-cuda-benchmarks${debugExt} tnl-cuda-benchmarks.cu )

    if( WITH_CUBLAS STREQUAL "yes" )

        CUDA_ADD_CUBLAS_TO_TARGET( tnl-cuda-benchmarks${debugExt} )

    endif()

    TARGET_LINK_LIBRARIES( tnl-cuda-benchmarks${debugExt} tnl${debugExt}-${tnlVersion} ${CUSPARSE_LIBRARY} )                        

    

    CUDA_ADD_EXECUTABLE( tnl-benchmark-spmv${debugExt} tnl-benchmark-spmv.cu )

tests/benchmarks/cublasWrappers.h

0 → 100644
+27 −0
Original line number Diff line number Diff line 
#pragma once



#ifdef HAVE_CUDA

#ifdef HAVE_CUBLAS



#include <cublas_v2.h>



inline cublasStatus_t

cublasGdot( cublasHandle_t handle, int n,

            const float        *x, int incx,

            const float        *y, int incy,

            float         *result )

{

    return cublasSdot( handle, n, x, incx, y, incy, result );

}



inline cublasStatus_t

cublasGdot( cublasHandle_t handle, int n,

            const double       *x, int incx,

            const double       *y, int incy,

            double        *result )

{

    return cublasDdot( handle, n, x, incx, y, incy, result );

}



#endif

#endif

tests/benchmarks/vector-operations.h

+21 −20
Original line number Diff line number Diff line
@@ -5,7 +5,7 @@ 
#include <core/vectors/tnlVector.h>



#ifdef HAVE_CUBLAS

//#include <cublas.h>

#include "cublasWrappers.h"

#endif



namespace tnl
@@ -36,6 +36,11 @@ benchmarkVectorOperations( const int & loops, 


    Real resultHost, resultDevice;



#ifdef HAVE_CUBLAS

    cublasHandle_t cublasHandle;

    cublasCreate( &cublasHandle );

#endif





    // reset functions

    // (Make sure to always use some in benchmarks, even if it's not necessary
@@ -200,29 +205,21 @@ benchmarkVectorOperations( const int & loops, 
    auto scalarProductCuda = [&]() {

        resultDevice = deviceVector.scalarProduct( deviceVector2 );

    };

    benchmarkOperation( "scalar product", 2 * datasetSize, loops, reset1,

                        "CPU", scalarProductHost,

                        "GPU", scalarProductCuda );



/* TODO

#ifdef HAVE_CUBLAS

   cout << "Benchmarking scalar product on GPU with Cublas: " << endl;

   cublasHandle_t handle;

   cublasCreate( &handle );

   timer.reset();

   timer.start();

   for( int i = 0; i < loops; i++ )

      cublasDdot( handle,

                  size,

                  deviceVector.getData(), 1,

    auto scalarProductCublas = [&]() {

        cublasGdot( cublasHandle, size,

                    deviceVector.getData(), 1,

                    deviceVector2.getData(), 1,

                    &resultDevice );

   cudaThreadSynchronize();

   timer.stop();

   bandwidth = 2 * datasetSize / timer.getTime();

   cout << "bandwidth: " << bandwidth << " GB/sec, time: " << timer.getTime() << " sec." << endl;

    };

#endif

*/

    benchmarkOperation( "scalar product", 2 * datasetSize, loops, reset1,

                        "CPU", scalarProductHost,

                        "GPU", scalarProductCuda

#ifdef HAVE_CUBLAS

                      , "cuBLAS", scalarProductCublas

#endif

                      );



    /*

    cout << "Benchmarking prefix-sum:" << endl;
@@ -253,6 +250,10 @@ benchmarkVectorOperations( const int & loops, 
       }

    */



#ifdef HAVE_CUBLAS

    cublasDestroy( cublasHandle );

#endif



    return true;

}