Commit 1743358a authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Moved synchronization of smart pointers from Devices::Cuda into TNL::Pointers... 

Moved synchronization of smart pointers from Devices::Cuda into TNL::Pointers namespace as free functions

synchronizeDevice() was renamed to synchronizeSmartPointersOnDevice()
for clarity - there are many similarly named functions in CUDA (e.g.
cudaDeviceSynchronize()).
parent 2d5176fb

src/Benchmarks/HeatEquation/HeatEquationBenchmarkProblem_impl.h

+2 −2
Original line number Diff line number Diff line
@@ -490,7 +490,7 @@ getExplicitUpdate( const RealType& time, 
         

         //std::cerr << "Setting boundary conditions..." << std::endl;



         Devices::Cuda::synchronizeDevice();

         Pointers::synchronizeSmartPointersOnDevice< Devices::Cuda >();

         for( IndexType gridYIdx = 0; gridYIdx < cudaYGrids; gridYIdx ++ )

            for( IndexType gridXIdx = 0; gridXIdx < cudaXGrids; gridXIdx ++ )

               boundaryConditionsTemplatedCompact< MeshType, CellType, BoundaryCondition, MeshFunctionType >
@@ -594,7 +594,7 @@ getExplicitUpdate( const RealType& time, 
                               gridYSize / 16 + ( gridYSize % 16 != 0 ) );

            */



            TNL::Devices::Cuda::synchronizeDevice();

            Pointers::synchronizeSmartPointersOnDevice< Devices::Cuda >();

            int cudaErr;

            Meshes::Traverser< MeshType, Cell > meshTraverser;

            meshTraverser.template processInteriorEntities< UserData,

src/Benchmarks/HeatEquation/Tuning/GridTraverser_impl.h

+2 −2
Original line number Diff line number Diff line
@@ -246,7 +246,7 @@ processEntities( 
      IndexType cudaThreadsCount = 2 * ( end.x() - begin.x() + end.y() - begin.y() + 1 );

      Cuda::setupThreads( cudaBlockSize, cudaBlocksCount, cudaGridsCount, cudaThreadsCount );

      dim3 gridIdx, cudaGridSize;

      Devices::Cuda::synchronizeDevice();

      Pointers::synchronizeSmartPointersOnDevice< Devices::Cuda >();

      for( gridIdx.x = 0; gridIdx.x < cudaGridsCount.x; gridIdx.x++ )

      {

         Cuda::setupGrid( cudaBlocksCount, cudaGridsCount, gridIdx, cudaGridSize );
@@ -273,7 +273,7 @@ processEntities( 
      auto& pool = Cuda::StreamPool::getInstance();

      const cudaStream_t& s = pool.getStream( stream );



      Devices::Cuda::synchronizeDevice();

      Pointers::synchronizeSmartPointersOnDevice< Devices::Cuda >();

      dim3 gridIdx, cudaGridSize;

      for( gridIdx.y = 0; gridIdx.y < cudaGridsCount.y; gridIdx.y ++ )

         for( gridIdx.x = 0; gridIdx.x < cudaGridsCount.x; gridIdx.x ++ )

src/Benchmarks/LinearSolvers/tnl-benchmark-linear-solvers.h

+1 −1
Original line number Diff line number Diff line
@@ -130,7 +130,7 @@ benchmarkIterativeSolvers( Benchmark& benchmark, 
   *cudaMatrixPointer = *matrixPointer;



   // synchronize shared pointers

   Devices::Cuda::synchronizeDevice();

   Pointers::synchronizeSmartPointersOnDevice< Devices::Cuda >();

#endif



   using namespace Solvers::Linear;

src/TNL/Devices/Cuda.h

+0 −21
Original line number Diff line number Diff line
@@ -10,13 +10,7 @@ 


#pragma once



#include <iostream>



#include <TNL/String.h>

#include <TNL/Assert.h>

#include <TNL/Pointers/SmartPointersRegister.h>

#include <TNL/Timer.h>

#include <TNL/Cuda/CudaCallable.h>

#include <TNL/Config/ConfigDescription.h>

#include <TNL/Config/ParameterContainer.h>
@@ -33,16 +27,6 @@ public: 


   static inline constexpr int getGPUTransferBufferSize();



   static inline void insertSmartPointer( Pointers::SmartPointer* pointer );



   static inline void removeSmartPointer( Pointers::SmartPointer* pointer );



   // Negative deviceId means that CudaDeviceInfo::getActiveDevice will be

   // called to get the device ID.

   static inline bool synchronizeDevice( int deviceId = -1 );



   static inline Timer& getSmartPointersSynchronizationTimer();



   ////

   // When we transfer data between the GPU and the CPU we use 5 MB buffer. This

   // size should ensure good performance -- see.
@@ -50,11 +34,6 @@ public: 
   // We use the same buffer size even for retyping data during IO operations.

   //

   static constexpr std::size_t TransferBufferSize = 5 * 2<<20;





   protected:



   static inline Pointers::SmartPointersRegister& getSmartPointersRegister();

};



#ifdef HAVE_CUDA

src/TNL/Devices/Cuda_impl.h

+5 −38
Original line number Diff line number Diff line
@@ -10,12 +10,15 @@ 


#pragma once



#include <iostream>



#include <TNL/Math.h>

#include <TNL/Devices/Cuda.h>

#include <TNL/Cuda/DeviceInfo.h>

#include <TNL/Exceptions/CudaBadAlloc.h>

#include <TNL/Exceptions/CudaSupportMissing.h>

#include <TNL/Exceptions/CudaRuntimeError.h>

#include <TNL/Pointers/SmartPointersRegister.h>



namespace TNL {

namespace Devices {
@@ -42,8 +45,8 @@ Cuda::setup( const Config::ParameterContainer& parameters, 
      std::cerr << "I cannot activate CUDA device number " << cudaDevice << "." << std::endl;

      return false;

   }

   getSmartPointersSynchronizationTimer().reset();

   getSmartPointersSynchronizationTimer().stop();

   Pointers::getSmartPointersSynchronizationTimer< Devices::Cuda >().reset();

   Pointers::getSmartPointersSynchronizationTimer< Devices::Cuda >().stop();

#endif

   return true;

}
@@ -53,42 +56,6 @@ inline constexpr int Cuda::getGPUTransferBufferSize() 
   return 1 << 20;

}



inline void Cuda::insertSmartPointer( Pointers::SmartPointer* pointer )

{

   getSmartPointersRegister().insert( pointer, TNL::Cuda::DeviceInfo::getActiveDevice() );

}



inline void Cuda::removeSmartPointer( Pointers::SmartPointer* pointer )

{

   getSmartPointersRegister().remove( pointer, TNL::Cuda::DeviceInfo::getActiveDevice() );

}



inline bool Cuda::synchronizeDevice( int deviceId )

{

#ifdef HAVE_CUDA

   if( deviceId < 0 )

      deviceId = TNL::Cuda::DeviceInfo::getActiveDevice();

   getSmartPointersSynchronizationTimer().start();

   bool b = getSmartPointersRegister().synchronizeDevice( deviceId );

   getSmartPointersSynchronizationTimer().stop();

   return b;

#else

   return true;

#endif

}



inline Timer& Cuda::getSmartPointersSynchronizationTimer()

{

   static Timer timer;

   return timer;

}



inline Pointers::SmartPointersRegister& Cuda::getSmartPointersRegister()

{

   static Pointers::SmartPointersRegister reg;

   return reg;

}



// double-precision atomicAdd function for Maxwell and older GPUs

// copied from: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomic-functions

#ifdef HAVE_CUDA