Commit a2fb0bdf authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Optimized parallel reduction on CUDA

parent 592f6355

src/TNL/Containers/Algorithms/CudaReductionKernel.h

+1 −1
Original line number Diff line number Diff line
@@ -31,7 +31,7 @@ namespace Algorithms { 
 */

static constexpr int Reduction_maxThreadsPerBlock = 256;  // must be a power of 2

#if (__CUDA_ARCH__ >= 300 )

   static constexpr int Reduction_minBlocksPerMultiprocessor = 6;

   static constexpr int Reduction_minBlocksPerMultiprocessor = 8;

#else

   static constexpr int Reduction_minBlocksPerMultiprocessor = 4;

#endif

src/TNL/Devices/CudaDeviceInfo.cu

+11 −3
Original line number Diff line number Diff line
@@ -10,6 +10,8 @@ 


#ifdef HAVE_CUDA



#include <unordered_map>



#include <TNL/Devices/CudaDeviceInfo.h>

#include <TNL/Devices/Cuda.h>
@@ -101,10 +103,16 @@ int 
CudaDeviceInfo::

getCudaMultiprocessors( int deviceNum )

{

    // results are cached because they are used for configuration of some kernels

    static std::unordered_map< int, int > results;

    if( results.count( deviceNum ) == 0 ) {

        cudaDeviceProp properties;

        cudaGetDeviceProperties( &properties, deviceNum );

        results.emplace( deviceNum, properties.multiProcessorCount );

        return properties.multiProcessorCount;

    }

    return results[ deviceNum ];

}



int

CudaDeviceInfo::