Commit 57d66051 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

CudaScanKernelLauncher: configuration of blockSize depending on the ValueType

parent 8accbc52

src/TNL/Algorithms/detail/CudaScanKernel.h

+7 −2
Original line number Diff line number Diff line
@@ -436,7 +436,10 @@ CudaScanKernelUniformShift( OutputView output, 
 */

template< ScanType scanType,

          ScanPhaseType phaseType,

          int blockSize = 256,

          typename ValueType,

          // use blockSize=256 for 32-bit value types, scale with sizeof(ValueType)

          // to keep shared memory requirements constant

          int blockSize = 256 * 4 / sizeof(ValueType),

          // valuesPerThread should be odd to avoid shared memory bank conflicts

          int valuesPerThread = 7 >

struct CudaScanKernelLauncher
@@ -520,6 +523,7 @@ struct CudaScanKernelLauncher 
                      Reduction&& reduction,

                      typename OutputArray::ValueType zero )

   {

      static_assert( std::is_same< ValueType, typename OutputArray::ValueType >::value, "invalid configuration of ValueType" );

      using Index = typename InputArray::IndexType;



      if( end - begin <= blockSize * valuesPerThread ) {
@@ -639,7 +643,7 @@ struct CudaScanKernelLauncher 


         // blockResults now contains scan results for each block. The first phase

         // ends by computing an exclusive scan of this array.

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase >::perform(

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::perform(

            blockResults,

            blockResults,

            0,
@@ -689,6 +693,7 @@ struct CudaScanKernelLauncher 
                       typename OutputArray::ValueType zero,

                       typename OutputArray::ValueType shift )

   {

      static_assert( std::is_same< ValueType, typename OutputArray::ValueType >::value, "invalid configuration of ValueType" );

      using Index = typename InputArray::IndexType;



      // if the input was already scanned with just one block in the first phase,

src/TNL/Algorithms/detail/Scan.hpp

+3 −3
Original line number Diff line number Diff line
@@ -278,7 +278,7 @@ perform( const InputArray& input, 
   if( end <= begin )

      return;



   detail::CudaScanKernelLauncher< Type, PhaseType >::perform(

   detail::CudaScanKernelLauncher< Type, PhaseType, typename OutputArray::ValueType >::perform(

      input,

      output,

      begin,
@@ -312,7 +312,7 @@ performFirstPhase( const InputArray& input, 
      return block_results;

   }



   return detail::CudaScanKernelLauncher< Type, PhaseType >::performFirstPhase(

   return detail::CudaScanKernelLauncher< Type, PhaseType, typename OutputArray::ValueType >::performFirstPhase(

      input,

      output,

      begin,
@@ -346,7 +346,7 @@ performSecondPhase( const InputArray& input, 
   if( end <= begin )

      return;



   detail::CudaScanKernelLauncher< Type, PhaseType >::performSecondPhase(

   detail::CudaScanKernelLauncher< Type, PhaseType, typename OutputArray::ValueType >::performSecondPhase(

      input,

      output,

      blockShifts,

src/UnitTests/Algorithms/distributedScanTest.h

+10 −10
Original line number Diff line number Diff line
@@ -87,14 +87,14 @@ protected: 
#ifdef HAVE_CUDA

      if( std::is_same< DeviceType, Devices::Cuda >::value )

      {

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::maxGridSize() = 3;

      }

#endif

   }
@@ -106,8 +106,8 @@ protected: 
      // skip the check for too small arrays

      if( check_cuda_grids && array.getLocalRange().getSize() > 256 ) {

         // we don't care which kernel launcher was actually used

         const auto gridsCount = TNL::max( CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInFirstPhase >::gridsCount(),

                                           CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInSecondPhase >::gridsCount() );

         const auto gridsCount = TNL::max( CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInFirstPhase, ValueType >::gridsCount(),

                                           CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInSecondPhase, ValueType >::gridsCount() );

         EXPECT_GT( gridsCount, 1 );

      }

#endif

src/UnitTests/Algorithms/scanTest.h

+10 −10
Original line number Diff line number Diff line
@@ -59,14 +59,14 @@ protected: 
#ifdef HAVE_CUDA

      if( std::is_same< DeviceType, Devices::Cuda >::value )

      {

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInFirstPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Inclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::maxGridSize() = 3;

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::resetMaxGridSize();

         CudaScanKernelLauncher< ScanType::Exclusive, ScanPhaseType::WriteInSecondPhase, ValueType >::maxGridSize() = 3;

      }

#endif

   }
@@ -78,8 +78,8 @@ protected: 
      // skip the check for too small arrays

      if( array.getSize() > 256 ) {

         // we don't care which kernel launcher was actually used

         const auto gridsCount = TNL::max( CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInFirstPhase >::gridsCount(),

                                           CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInSecondPhase >::gridsCount() );

         const auto gridsCount = TNL::max( CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInFirstPhase, ValueType >::gridsCount(),

                                           CudaScanKernelLauncher< ScanType, ScanPhaseType::WriteInSecondPhase, ValueType >::gridsCount() );

         EXPECT_GT( gridsCount, 1 );

      }

#endif