Commit 75b18f59 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Refactoring BLAS benchmarks

parent 9e6fe9a8

CMakeLists.txt

+2 −2
Original line number Diff line number Diff line
@@ -44,8 +44,8 @@ endif() 
# set Debug/Release options

set( CMAKE_CXX_FLAGS "-std=c++11 -Wall -Wno-unused-local-typedefs -Wno-unused-variable" )

set( CMAKE_CXX_FLAGS_DEBUG "-g" )

set( CMAKE_CXX_FLAGS_RELEASE "-O3 -march=native -DNDEBUG" )

#set( CMAKE_CXX_FLAGS_RELEASE "-O3 -march=native -DNDEBUG -ftree-vectorizer-verbose=1 -ftree-vectorize -fopt-info-vec-missed -funroll-loops" )

set( CMAKE_CXX_FLAGS_RELEASE "-O3 -march=native -mtune=native -DNDEBUG" )

#set( CMAKE_CXX_FLAGS_RELEASE "-O3 -march=native -mtune=native -DNDEBUG -ftree-vectorizer-verbose=1 -ftree-vectorize -fopt-info-vec-missed -funroll-loops" )

# pass -rdynamic only in Debug mode

set( CMAKE_SHARED_LIBRARY_LINK_C_FLAGS "" )

set( CMAKE_SHARED_LIBRARY_LINK_C_FLAGS_DEBUG "-rdynamic" )

src/TNL/String.cpp

+6 −0
Original line number Diff line number Diff line
@@ -55,6 +55,12 @@ String :: String( int number ) 
   this->setString( convertToString( number ).getString() );

}



String :: String( unsigned long int number )

: string( 0 ), length( 0 )

{

   this->setString( convertToString( number ).getString() );

}



String :: String( long int number )

: string( 0 ), length( 0 )

{

src/TNL/String.h

+2 −0
Original line number Diff line number Diff line
@@ -55,6 +55,8 @@ class String 


   String( int number );

 

   String( unsigned long int number );



   String( long int number );



   String( float number );

tests/benchmarks/CMakeLists.txt

+14 −19
Original line number Diff line number Diff line
@@ -2,11 +2,11 @@ ADD_SUBDIRECTORY( share ) 
ADD_SUBDIRECTORY( heat-equation-benchmark )



IF( BUILD_CUDA )

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

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

    if( HAVE_CUBLAS STREQUAL "yes" )

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

        CUDA_ADD_CUBLAS_TO_TARGET( tnl-benchmark-blas${debugExt} )

    endif()

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

    TARGET_LINK_LIBRARIES( tnl-benchmark-blas${debugExt} tnl${debugExt}-${tnlVersion} ${CUSPARSE_LIBRARY} )



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

    TARGET_LINK_LIBRARIES( tnl-benchmark-spmv${debugExt} tnl${debugExt}-${tnlVersion} ${CUSPARSE_LIBRARY} )
@@ -14,6 +14,9 @@ IF( BUILD_CUDA ) 
    CUDA_ADD_EXECUTABLE( tnl-benchmark-linear-solvers${debugExt} tnl-benchmark-linear-solvers.cu )

    TARGET_LINK_LIBRARIES( tnl-benchmark-linear-solvers${debugExt} tnl${debugExt}-${tnlVersion} ${CUSPARSE_LIBRARY} )

ELSE()

    ADD_EXECUTABLE( tnl-benchmark-blas${debugExt} tnl-benchmark-blas.cpp )

    TARGET_LINK_LIBRARIES( tnl-benchmark-blas${debugExt} tnl${debugExt}-${tnlVersion} )



    ADD_EXECUTABLE( tnl-benchmark-spmv${debugExt} tnl-benchmark-spmv.cpp )

    TARGET_LINK_LIBRARIES( tnl-benchmark-spmv${debugExt} tnl${debugExt}-${tnlVersion} )
@@ -21,16 +24,8 @@ ELSE() 
    TARGET_LINK_LIBRARIES( tnl-benchmark-linear-solvers${debugExt} tnl${debugExt}-${tnlVersion} )

ENDIF()



if( BUILD_CUDA )                                                              

   INSTALL( TARGETS

                tnl-cuda-benchmarks${debugExt}

            RUNTIME DESTINATION bin )

endif()



INSTALL( TARGETS

            tnl-benchmark-blas${debugExt}

            tnl-benchmark-spmv${debugExt}

            tnl-benchmark-linear-solvers${debugExt}

         RUNTIME DESTINATION bin )

tests/benchmarks/array-operations.h

+38 −18
Original line number Diff line number Diff line
@@ -14,7 +14,7 @@ template< typename Real = double, 
bool

benchmarkArrayOperations( Benchmark & benchmark,

                          const int & loops,

                          const int & size )

                          const long & size )

{

    typedef Containers::Array< Real, Devices::Host, Index > HostArray;

    typedef Containers::Array< Real, Devices::Cuda, Index > CudaArray;
@@ -25,9 +25,14 @@ benchmarkArrayOperations( Benchmark & benchmark, 
    HostArray hostArray, hostArray2;

    CudaArray deviceArray, deviceArray2;

    if( ! hostArray.setSize( size ) ||

        ! hostArray2.setSize( size ) ||

        ! hostArray2.setSize( size )

#ifdef HAVE_CUDA

        ||

        ! deviceArray.setSize( size ) ||

        ! deviceArray2.setSize( size ) )

        ! deviceArray2.setSize( size )

#endif

    )



    {

        const char* msg = "error: allocation of arrays failed";

        std::cerr << msg << std::endl;
@@ -41,11 +46,15 @@ benchmarkArrayOperations( Benchmark & benchmark, 
    // reset functions

    auto reset1 = [&]() {

        hostArray.setValue( 1.0 );

#ifdef HAVE_CUDA

        deviceArray.setValue( 1.0 );

#endif

    };

    auto reset2 = [&]() {

        hostArray2.setValue( 1.0 );

#ifdef HAVE_CUDA

        deviceArray2.setValue( 1.0 );

#endif

    };

    auto reset12 = [&]() {

        reset1();
@@ -63,9 +72,10 @@ benchmarkArrayOperations( Benchmark & benchmark, 
        resultDevice = (int) deviceArray == deviceArray2;

    };

    benchmark.setOperation( "comparison (operator==)", 2 * datasetSize );

    benchmark.time( reset1,

                    "CPU", compareHost,

                    "GPU", compareCuda );

    benchmark.time( reset1, "CPU", compareHost );

#ifdef HAVE_CUDA

    benchmark.time( reset1, "GPU", compareCuda );

#endif





    auto copyAssignHostHost = [&]() {
@@ -75,9 +85,10 @@ benchmarkArrayOperations( Benchmark & benchmark, 
        deviceArray = deviceArray2;

    };

    benchmark.setOperation( "copy (operator=)", 2 * datasetSize );

    double basetime = benchmark.time( reset1,

                    "CPU", copyAssignHostHost,

                    "GPU", copyAssignCudaCuda );

    benchmark.time( reset1, "CPU", copyAssignHostHost );

#ifdef HAVE_CUDA

    benchmark.time( reset1, "GPU", copyAssignCudaCuda );

#endif





    auto copyAssignHostCuda = [&]() {
@@ -86,10 +97,12 @@ benchmarkArrayOperations( Benchmark & benchmark, 
    auto copyAssignCudaHost = [&]() {

        hostArray = deviceArray;

    };

#ifdef HAVE_CUDA

    benchmark.setOperation( "copy (operator=)", datasetSize, basetime );

    benchmark.time( reset1,

                    "CPU->GPU", copyAssignHostCuda,

                    "GPU->CPU", copyAssignCudaHost );

#endif





    auto setValueHost = [&]() {
@@ -99,9 +112,10 @@ benchmarkArrayOperations( Benchmark & benchmark, 
        deviceArray.setValue( 3.0 );

    };

    benchmark.setOperation( "setValue", datasetSize );

    benchmark.time( reset1,

                    "CPU", setValueHost,

                    "GPU", setValueCuda );

    benchmark.time( reset1, "CPU", setValueHost );

#ifdef HAVE_CUDA

    benchmark.time( reset1, "GPU", setValueCuda );

#endif





    auto setSizeHost = [&]() {
@@ -112,12 +126,15 @@ benchmarkArrayOperations( Benchmark & benchmark, 
    };

    auto resetSize1 = [&]() {

        hostArray.reset();

#ifdef HAVE_CUDA

        deviceArray.reset();

#endif

    };

    benchmark.setOperation( "allocation (setSize)", datasetSize );

    benchmark.time( resetSize1,

                    "CPU", setSizeHost,

                    "GPU", setSizeCuda );

    benchmark.time( resetSize1, "CPU", setSizeHost );

#ifdef HAVE_CUDA

    benchmark.time( resetSize1, "GPU", setSizeCuda );

#endif





    auto resetSizeHost = [&]() {
@@ -128,12 +145,15 @@ benchmarkArrayOperations( Benchmark & benchmark, 
    };

    auto setSize1 = [&]() {

        hostArray.setSize( size );

#ifdef HAVE_CUDA

        deviceArray.setSize( size );

#endif

    };

    benchmark.setOperation( "deallocation (reset)", datasetSize );

    benchmark.time( setSize1,

                    "CPU", resetSizeHost,

                    "GPU", resetSizeCuda );

    benchmark.time( setSize1, "CPU", resetSizeHost );

#ifdef HAVE_CUDA

    benchmark.time( setSize1, "GPU", resetSizeCuda );

#endif



    return true;

}