Queue< PlatformKind::CUDA > Class Template Reference

Collaboration diagram for Queue< PlatformKind::CUDA >:

Public Member Functions
	Queue ()
	~Queue ()
template<class DSC >
void	enqueueH2D (DSC &dataset, typename DSC::size_type size=0, typename DSC::size_type offset=0)
template<class DSC >
void	enqueueD2H (DSC &dataset, typename DSC::size_type size=0, typename DSC::size_type offset=0)
void	sync ()
cudaStream_t	getNative ()

Private Attributes
cudaStream_t	hstream_

Detailed Description

template<>
class qmcplusplus::compute::Queue< PlatformKind::CUDA >

Definition at line 25 of file QueueCUDA.hpp.

Constructor & Destructor Documentation

Queue()

Queue ( )

inline

Definition at line 28 of file QueueCUDA.hpp.

References qmcplusplus::cudaErrorCheck(), and cudaStreamCreate.

{ cudaErrorCheck(cudaStreamCreate(&hstream_), "cudaStreamCreate failed!"); }

~Queue()

~Queue ( )

inline

Definition at line 30 of file QueueCUDA.hpp.

References qmcplusplus::cudaErrorCheck(), and cudaStreamDestroy.

{ cudaErrorCheck(cudaStreamDestroy(hstream_), "cudaStreamDestroy failed!"); }

Member Function Documentation

enqueueD2H()

void enqueueD2H ( DSC &  dataset, typename DSC::size_type  size = 0, typename DSC::size_type  offset = 0  )

inline

Definition at line 53 of file QueueCUDA.hpp.

References qmcplusplus::cudaErrorCheck(), cudaMemcpyAsync, and cudaMemcpyDeviceToHost.

  {
    if (dataset.data() == dataset.device_data()) return;

    if (size == 0)
    {
      cudaErrorCheck(cudaMemcpyAsync(dataset.data() + offset, dataset.device_data() + offset,
                                     dataset.size() * sizeof(typename DSC::value_type), cudaMemcpyDeviceToHost, hstream_),
                     "Queue<PlatformKind::CUDA>::enqueueD2H cudaMemcpyAsync failed!");
    }
    else
    {
      cudaErrorCheck(cudaMemcpyAsync(dataset.data() + offset, dataset.device_data() + offset,
                                     size * sizeof(typename DSC::value_type), cudaMemcpyDeviceToHost, hstream_),
                     "Queue<PlatformKind::CUDA>::enqueueD2H cudaMemcpyAsync failed!");
    }
  }

enqueueH2D()

void enqueueH2D ( DSC &  dataset, typename DSC::size_type  size = 0, typename DSC::size_type  offset = 0  )

inline

Definition at line 34 of file QueueCUDA.hpp.

References qmcplusplus::cudaErrorCheck(), cudaMemcpyAsync, and cudaMemcpyHostToDevice.

  {
    if (dataset.data() == dataset.device_data()) return;

    if (size == 0)
    {
      cudaErrorCheck(cudaMemcpyAsync(dataset.device_data() + offset, dataset.data() + offset,
                                     dataset.size() * sizeof(typename DSC::value_type), cudaMemcpyHostToDevice, hstream_),
                     "Queue<PlatformKind::CUDA>::enqueueH2D cudaMemcpyAsync failed!");
    }
    else
    {
      cudaErrorCheck(cudaMemcpyAsync(dataset.device_data() + offset, dataset.data() + offset,
                                     size * sizeof(typename DSC::value_type), cudaMemcpyHostToDevice, hstream_),
                     "Queue<PlatformKind::CUDA>::enqueueH2D cudaMemcpyAsync failed!");
    }
  }

getNative()

cudaStream_t getNative ( )

inline

Definition at line 73 of file QueueCUDA.hpp.

Referenced by DelayedUpdateCUDA< T, T_FP >::getInvRow(), DelayedUpdateCUDA< T, T_FP >::initializeInv(), and DelayedUpdateCUDA< T, T_FP >::updateInvMat().

{ return hstream_; }

sync()

void sync ( )

inline

Definition at line 71 of file QueueCUDA.hpp.

References qmcplusplus::cudaErrorCheck(), and cudaStreamSynchronize.

Referenced by DelayedUpdateCUDA< T, T_FP >::getInvRow(), DelayedUpdateCUDA< T, T_FP >::initializeInv(), and DelayedUpdateCUDA< T, T_FP >::updateInvMat().

{ cudaErrorCheck(cudaStreamSynchronize(hstream_), "cudaStreamSynchronize failed!"); }

Member Data Documentation

hstream_

cudaStream_t hstream_

private

Definition at line 76 of file QueueCUDA.hpp.

---

The documentation for this class was generated from the following file:

• /home/pk7/projects/qmc/for_cron_doxygen/qmcpack/src/Platforms/CUDA/QueueCUDA.hpp