Commit 69488aac authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Moved Cuda::configSetup and Cuda::setup from Cuda.cpp to Cuda_impl.h 

This will fix the bugs described in the FIXME notes.

Compilation fails due to cyclic dependency between the headers, it will
be fixed later.
parent 674312d4

src/TNL/Devices/Cuda.cpp

+0 −30
Original line number Diff line number Diff line
@@ -10,8 +10,6 @@ 


#include <TNL/Devices/Cuda.h>

#include <TNL/Math.h>

#include <TNL/Config/ConfigDescription.h>

#include <TNL/Config/ParameterContainer.h>

#include <TNL/Devices/CudaDeviceInfo.h>



namespace TNL {
@@ -37,34 +35,6 @@ int Cuda::getNumberOfGrids( const int blocks, 
   return roundUpDivision( blocks, gridSize );

}



void Cuda::configSetup( Config::ConfigDescription& config,

                        const String& prefix )

{

// FIXME: HAVE_CUDA is never defined in .cpp files

#ifdef HAVE_CUDA

   config.addEntry< int >( prefix + "cuda-device", "Choose CUDA device to run the computation.", 0 );

#else

   config.addEntry< int >( prefix + "cuda-device", "Choose CUDA device to run the computation (not supported on this system).", 0 );

#endif

}



bool Cuda::setup( const Config::ParameterContainer& parameters,

                  const String& prefix )

{

// FIXME: HAVE_CUDA is never defined in .cpp files

#ifdef HAVE_CUDA

   int cudaDevice = parameters.getParameter< int >( prefix + "cuda-device" );

   if( cudaSetDevice( cudaDevice ) != cudaSuccess )

   {

      std::cerr << "I cannot activate CUDA device number " << cudaDevice << "." << std::endl;

      return false;

   }

   smartPointersSynchronizationTimer.reset();

   smartPointersSynchronizationTimer.stop();

#endif

   return true;

}



void Cuda::insertSmartPointer( SmartPointer* pointer )

{

   smartPointersRegister.insert( pointer, Devices::CudaDeviceInfo::getActiveDevice() );

src/TNL/Devices/Cuda_impl.h

+30 −0
Original line number Diff line number Diff line
@@ -14,6 +14,8 @@ 
#include <TNL/Exceptions/CudaBadAlloc.h>

#include <TNL/Exceptions/CudaSupportMissing.h>

#include <TNL/CudaSharedMemory.h>

#include <TNL/Config/ConfigDescription.h>

#include <TNL/Config/ParameterContainer.h>



namespace TNL {

namespace Devices {
@@ -160,6 +162,34 @@ __device__ Element* Cuda::getSharedMemory() 
   return CudaSharedMemory< Element >();

}



inline void

Cuda::configSetup( Config::ConfigDescription& config,

                   const String& prefix )

{

#ifdef HAVE_CUDA

   config.addEntry< int >( prefix + "cuda-device", "Choose CUDA device to run the computation.", 0 );

#else

   config.addEntry< int >( prefix + "cuda-device", "Choose CUDA device to run the computation (not supported on this system).", 0 );

#endif

}



inline bool

Cuda::setup( const Config::ParameterContainer& parameters,

             const String& prefix )

{

#ifdef HAVE_CUDA

   int cudaDevice = parameters.getParameter< int >( prefix + "cuda-device" );

   if( cudaSetDevice( cudaDevice ) != cudaSuccess )

   {

      std::cerr << "I cannot activate CUDA device number " << cudaDevice << "." << std::endl;

      return false;

   }

   smartPointersSynchronizationTimer.reset();

   smartPointersSynchronizationTimer.stop();

#endif

   return true;

}



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

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

#if __CUDA_ARCH__ < 600