Commit 28938623 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Non-fundamental types must be reduced completely on the CUDA device

parent 3c03735e

src/TNL/Containers/Algorithms/Reduction_impl.h

+66 −29
Original line number Diff line number Diff line
@@ -51,11 +51,18 @@ reductionOnCudaDevice( Operation& operation, 
   typedef typename Operation::ResultType ResultType;

   typedef typename Operation::LaterReductionOperation LaterReductionOperation;

 

   /***

    * Only fundamental and pointer types can be safely reduced on host. Complex

    * objects stored on the device might contain pointers into the device memory,

    * in which case reduction on host might fail.

    */

   constexpr bool can_copy_to_host = std::is_fundamental< RealType >::value || std::is_pointer< RealType >::value;



   /***

    * First check if the input array(s) is/are large enough for the reduction on GPU.

    * Otherwise copy it/them to host and reduce on CPU.

    */

   if( size <= minGPUReductionDataSize )

   if( can_copy_to_host && size <= minGPUReductionDataSize )

   {

      RealType hostArray1[ minGPUReductionDataSize ];

      RealType hostArray2[ minGPUReductionDataSize ];
@@ -92,6 +99,7 @@ reductionOnCudaDevice( Operation& operation, 
      timer.start();

   #endif



   if( can_copy_to_host ) {

      /***

       * Transfer the reduced data from device to host.

       */
@@ -102,9 +110,6 @@ reductionOnCudaDevice( Operation& operation, 
      #ifdef CUDA_REDUCTION_PROFILING

         timer.stop();

         std::cout << "   Transferring data to CPU took " << timer.getRealTime() << " sec. " << std::endl;

   #endif



   #ifdef CUDA_REDUCTION_PROFILING

         timer.reset();

         timer.start();

      #endif
@@ -121,6 +126,38 @@ reductionOnCudaDevice( Operation& operation, 
         timer.stop();

         std::cout << "   Reduction of small data set on CPU took " << timer.getRealTime() << " sec. " << std::endl;

      #endif

   }

   else {

      /***

       * Data can't be safely reduced on host, so continue with the reduction on the CUDA device.

       */

      LaterReductionOperation laterReductionOperation;

      while( reducedSize > 1 ) {

         // TODO: copy the intermediate result somewhere else, in-place reduction probably does not work

         reducedSize = CudaReductionKernelLauncher( laterReductionOperation,

                                                    reducedSize,

                                                    deviceAux1,

                                                    (ResultType*) 0,

                                                    deviceAux1 );

      }



      #ifdef CUDA_REDUCTION_PROFILING

         timer.stop();

         std::cout << "   Reduction of small data set on GPU took " << timer.getRealTime() << " sec. " << std::endl;

         timer.reset();

         timer.start();

      #endif



      ResultType resultArray[ 1 ];

      if( ! ArrayOperations< Devices::Host, Devices::Cuda >::copyMemory< ResultType, ResultType, IndexType >( resultArray, deviceAux1, reducedSize ) )

         return false;

      result = resultArray[ 0 ];



      #ifdef CUDA_REDUCTION_PROFILING

         timer.stop();

         std::cout << "   Transferring the result to CPU took " << timer.getRealTime() << " sec. " << std::endl;

      #endif

   }

 

   return checkCudaDevice;

#else