Commit c4d018b8 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Fixed BLAS benchmark 

Calculation of bandwidth broke in commit fb97bcd0

Check of MpiCommunicator::IsInitialized is necessary because
non-distributed benchmarks like tnl-benchmark-blas do not initialize
the communicator.
parent 679d1979

src/Benchmarks/BLAS/array-operations.h

+1 −1
Original line number Diff line number Diff line
@@ -30,7 +30,7 @@ benchmarkArrayOperations( Benchmark & benchmark, 
   typedef Containers::Array< Real, Devices::Cuda, Index > CudaArray;

   using namespace std;



   double datasetSize = ( double ) ( loops * size ) * sizeof( Real ) / oneGB;

   double datasetSize = (double) size * sizeof( Real ) / oneGB;



   HostArray hostArray, hostArray2;

   CudaArray deviceArray, deviceArray2;

src/Benchmarks/BLAS/spmv.h

+1 −1
Original line number Diff line number Diff line
@@ -139,7 +139,7 @@ benchmarkSpMV( Benchmark & benchmark, 


   const int elements = setHostTestMatrix< HostMatrix >( hostMatrix, elementsPerRow );

   setCudaTestMatrix< DeviceMatrix >( deviceMatrix, elementsPerRow );

   const double datasetSize = ( double ) loops * elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;

   const double datasetSize = (double) elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;



   // reset function

   auto reset = [&]() {

src/Benchmarks/BLAS/vector-operations.h

+3 −3
Original line number Diff line number Diff line
@@ -36,7 +36,7 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   typedef Containers::Vector< Real, Devices::Cuda, Index > CudaVector;

   using namespace std;



   double datasetSize = ( double ) ( loops * size ) * sizeof( Real ) / oneGB;

   double datasetSize = (double) size * sizeof( Real ) / oneGB;



   HostVector hostVector, hostVector2;

   CudaVector deviceVector, deviceVector2;
@@ -252,7 +252,7 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   hostVector.computePrefixSum();

   timer.stop();

   timeHost = timer.getTime();

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

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

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



   timer.reset();
@@ -260,7 +260,7 @@ benchmarkVectorOperations( Benchmark & benchmark, 
   deviceVector.computePrefixSum();

   timer.stop();

   timeDevice = timer.getTime();

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

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

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

   std::cout << "  CPU/GPU speedup: " << timeHost / timeDevice << std::endl;

src/Benchmarks/Benchmarks.h

+3 −1
Original line number Diff line number Diff line
@@ -504,7 +504,9 @@ Benchmark::MetadataMap getHardwareMetadata() 
       { "system release", Devices::SystemInfo::getSystemRelease() },

       { "start time", Devices::SystemInfo::getCurrentTime() },

#ifdef HAVE_MPI

       { "number of MPI processes", Communicators::MpiCommunicator::GetSize( Communicators::MpiCommunicator::AllGroup ) },

       { "number of MPI processes", (Communicators::MpiCommunicator::IsInitialized())

                                       ? Communicators::MpiCommunicator::GetSize( Communicators::MpiCommunicator::AllGroup )

                                       : 1 },

#endif

       { "OpenMP enabled", Devices::Host::isOMPEnabled() },

       { "OpenMP threads", Devices::Host::getMaxThreadsCount() },