Commit 5b36e3c1 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Benchmarks: add a case with explicit cudaThreadsSynchronize() call

parent 742581b2

tests/benchmarks/benchmarks.h

+16 −1
Original line number Diff line number Diff line
@@ -62,11 +62,13 @@ benchmarkCuda( const int & loops, 
               CheckFunction check = trueFunc,

               ResetFunction reset = voidFunc )

{

    tnlTimerRT timerHost, timerCuda;

    tnlTimerRT timerHost, timerCuda, timerCudaSync;

    timerHost.reset();

    timerHost.stop();

    timerCuda.reset();

    timerCuda.stop();

    timerCudaSync.reset();

    timerCudaSync.stop();



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

        timerHost.start();
@@ -81,15 +83,28 @@ benchmarkCuda( const int & loops, 
            throw BenchmarkError();



        reset();



        // Compute again on CUDA, with explicit synchronization of threads

        timerCudaSync.start();

        computeCuda();

        cudaThreadSynchronize();

        timerCudaSync.stop();



        reset();

    }



    const double timeHost = timerHost.getTime();

    const double timeCuda = timerCuda.getTime();

    const double timeCudaSync = timerCudaSync.getTime();

    const double bandwidthHost = datasetSize / timeHost;

    const double bandwidthCuda = datasetSize / timeCuda;

    const double bandwidthCudaSync = datasetSize / timeCudaSync;

    std::cout << "  CPU: bandwidth: " << bandwidthHost << " GB/sec, time: " << timeHost << " sec." << std::endl;

    std::cout << "  GPU: bandwidth: " << bandwidthCuda << " GB/sec, time: " << timeCuda << " sec." << std::endl;

    std::cout << "  GPU (sync): bandwidth: " << bandwidthCudaSync << " GB/sec, time: " << timeCudaSync << " sec." << std::endl;

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

    std::cout << "  CPU/GPU (sync) speedup: " << timeHost / timeCudaSync << std::endl;

    std::cout << std::endl;

}



} // namespace benchmarks

tests/benchmarks/tnl-cuda-benchmarks.h

+3 −3
Original line number Diff line number Diff line
@@ -43,7 +43,9 @@ using SlicedEllpackMatrix = tnlSlicedEllpackMatrix< Real, Device, Index >; 
//   - reset() clears the timer and starts it again

//   - getTime() stops the timer and starts it again !!!

//   - data members are not zero-initialized - reset has to be called manually, but it immediately starts the timer

// FIXME: scalarProduct is not const method

// FIXME:

// - scalarProduct is not const method

// - cudaThreadSynchronize() should be called from all CUDA methods





template< typename Matrix >
@@ -183,8 +185,6 @@ benchmarkSpMV( const int & loops, 
   };

   auto spmvCuda = [&]() {

      deviceMatrix.vectorProduct( deviceVector, deviceVector2 );

      // TODO: tnlCSRMatrix does not synchronize

      cudaThreadSynchronize();

   };



   benchmarkCuda( loops, datasetSize, spmvHost, spmvCuda, check, reset );

tests/benchmarks/vector-operations.h

+0 −2
Original line number Diff line number Diff line
@@ -78,8 +78,6 @@ benchmarkVectorOperations( const int & loops, 
    };

    auto addVectorCuda = [&]() {

        deviceVector.addVector( deviceVector2 );

        // TODO: synchronization should be part of addVector

        cudaThreadSynchronize();

    };

    benchmarkCuda( loops, 3 * datasetSize, addVectorHost, addVectorCuda, compare1, reset1 );