Kokkos node interface to the Thrust library for NVIDIA CUDA-capable GPUs. More...

#include <Kokkos_ThrustGPUNode.hpp>

Inheritance diagram for Kokkos::ThrustGPUNode:

Public Member Functions
	ThrustGPUNode (Teuchos::ParameterList &pl)
	Constructor acceptings a list of parameters.
	~ThrustGPUNode ()
	Destructor has no effect.

template<class T >
ArrayRCP< T >	allocBuffer (size_t size)
	Allocate a parallel buffer, returning it as a pointer ecnapsulated in an ArrayRCP.
template<class T >
void	copyFromBuffer (size_t size, const ArrayRCP< const T > &buffSrc, const ArrayView< T > &hostDest)
	Copy data to host memory from a parallel buffer.
template<class T >
void	copyToBuffer (size_t size, const ArrayView< const T > &hostSrc, const ArrayRCP< T > &buffDest)
	Copy data to host memory from a parallel buffer.
template<class T >
void	copyBuffers (size_t size, const ArrayRCP< const T > &buffSrc, const ArrayRCP< T > &buffDest)
	Copy data between buffers.
template<class T >
ArrayRCP< const T >	viewBuffer (size_t size, ArrayRCP< const T > buff)
	Return a const view of a buffer for use on the host. This creates a `const` view of length `size`, constituting the first `size` entries of `buff`, as they exist at the time of view creation. host memory allocated for the creation of this view is automatically deleted when no refences to the view remain.
template<class T >
ArrayRCP< T >	viewBufferNonConst (ReadWriteOption rw, size_t size, const ArrayRCP< T > &buff)
	Return a non-const view of a buffer for use on the host.

void	printStatistics (const RCP< Teuchos::FancyOStream > &os) const
	Print some statistics regarding node allocation and memory transfer.
void	clearStatistics ()
	Clear all statistics on memory transfer.
Static Public Attributes
static const bool	isHostNode = false
	Indicates that parallel buffers allocated by this node are not available for use on the host thread.
void	sync () const
	Block until all node work is complete. Aids in accurate timing of multiple kernels.
template<class WDP >
static void	parallel_for (int begin, int end, WDP wdp)
	parallel for skeleton, a wrapper around thrust::for_each. See Kokkos Node API
template<class WDP >
static WDP::ReductionType	parallel_reduce (int begin, int end, WDP wd)
	parallel reduction skeleton, a wrapper around thrust::transform_reduce. See Kokkos Node API

Detailed Description

Kokkos node interface to the Thrust library for NVIDIA CUDA-capable GPUs.

Definition at line 54 of file Kokkos_ThrustGPUNode.hpp.

Constructor & Destructor Documentation

Kokkos::ThrustGPUNode::ThrustGPUNode ( Teuchos::ParameterList & pl )

Constructor acceptings a list of parameters.

This constructor accepts the parameters:

Parameters:

Device Number	[int] Specifies the CUDA device to which the node will attach.
Verbose	[int] Non-zero parameter specifies that the constructor is verbose, printing information about the the attached device. Default: 0.

The constructor throw std::runtime_error if "Device Number" is outside the range $[0,numDevices)$ , where numDevices is the number of CUDA devices reported by cudaGetDeviceCount().

Definition at line 49 of file Kokkos_ThrustGPUNode.cpp.

Kokkos::ThrustGPUNode::~ThrustGPUNode ( )

Destructor has no effect.

Definition at line 90 of file Kokkos_ThrustGPUNode.cpp.

Member Function Documentation

template<class WDP >

void Kokkos::ThrustGPUNode::parallel_for	(	int	begin,
		int	end,
		WDP	wdp
	)		`[static]`

parallel for skeleton, a wrapper around thrust::for_each. See Kokkos Node API

Definition at line 96 of file Kokkos_ThrustGPUNode.cuh.

template<class WDP >

WDP::ReductionType Kokkos::ThrustGPUNode::parallel_reduce	(	int	begin,
		int	end,
		WDP	wd
	)		`[static]`

parallel reduction skeleton, a wrapper around thrust::transform_reduce. See Kokkos Node API

Definition at line 120 of file Kokkos_ThrustGPUNode.cuh.

void Kokkos::ThrustGPUNode::sync ( ) const

Block until all node work is complete. Aids in accurate timing of multiple kernels.

Definition at line 92 of file Kokkos_ThrustGPUNode.cpp.

template<class T >

ArrayRCP< T > Kokkos::CUDANodeMemoryModel::allocBuffer ( size_t size ) [inline, inherited]

Allocate a parallel buffer, returning it as a pointer ecnapsulated in an ArrayRCP.

Dereferencing the returned ArrayRCP or its underlying pointer in general results in undefined behavior outside of parallel computations.

The buffer will be automatically freed by the Node when no more references remain.

Template Parameters:

T	The data type of the allocate buffer. This is used to perform alignment and determine the number of bytes to allocate.

Parameters:

[in] size The size requested for the parallel buffer, greater than zero.

Postcondition:: The method will return an ArrayRCP encapsulating a pointer. The underlying pointer may be used in parallel computation routines, and is guaranteed to have size large enough to reference size number of entries of type T.

Definition at line 66 of file Kokkos_CUDANodeMemoryModelImpl.hpp.

template<class T >

void Kokkos::CUDANodeMemoryModel::copyFromBuffer	(	size_t	size,
		const ArrayRCP< const T > &	buffSrc,
		const ArrayView< T > &	hostDest
	)		`[inline, inherited]`

Copy data to host memory from a parallel buffer.

Parameters:

[in]	size	The number of entries to copy from `buffSrc` to `hostDest`.
[in]	buffSrc	The parallel buffer from which to copy.
[out]	hostDest	The location in host memory where the data from `buffSrc` is copied to.

Precondition:: size is non-negative.; buffSrc has length at least size.; hostDest has length equal to size.

Postcondition:: On return, entries in the range [0 , size) of buffSrc have been copied to hostDest entries in the range [0 , size).

Definition at line 89 of file Kokkos_CUDANodeMemoryModelImpl.hpp.

template<class T >

void Kokkos::CUDANodeMemoryModel::copyToBuffer	(	size_t	size,
		const ArrayView< const T > &	hostSrc,
		const ArrayRCP< T > &	buffDest
	)		`[inline, inherited]`

Copy data to host memory from a parallel buffer.

Parameters:

[in]	size	The number of entries to copy from `hostSrc` to `buffDest`.
[in]	hostSrc	The location in host memory from where the data is copied.
[out]	buffDest	The parallel buffer to which the data is copied.

Precondition:: size is non-negative.; hostSrc has length equal to size.; buffSrc has length at least size.

Postcondition:: On return, entries in the range [0 , size) of hostSrc are allowed to be written to. The data is guaranteed to be present in buffDest before it is used in a parallel computation.

Definition at line 117 of file Kokkos_CUDANodeMemoryModelImpl.hpp.

template<class T >

void Kokkos::CUDANodeMemoryModel::copyBuffers	(	size_t	size,
		const ArrayRCP< const T > &	buffSrc,
		const ArrayRCP< T > &	buffDest
	)		`[inline, inherited]`

Copy data between buffers.

Parameters:

[in]	size	The size of the copy, greater than zero.
[in]	buffSrc	The source buffer, with length at least as large as `size`.
[in,out]	buffDest	The destination buffer, with length at least as large as `size`.

Postcondition:: The data is guaranteed to have been copied before any other usage of buffSrc or buffDest occurs.

Definition at line 145 of file Kokkos_CUDANodeMemoryModelImpl.hpp.

template<class T >

ArrayRCP< const T > Kokkos::CUDANodeMemoryModel::viewBuffer	(	size_t	size,
		ArrayRCP< const T >	buff
	)		`[inline, inherited]`

Return a const view of a buffer for use on the host. This creates a const view of length size, constituting the first size entries of buff, as they exist at the time of view creation. host memory allocated for the creation of this view is automatically deleted when no refences to the view remain.

Precondition:: buff.size() >= size

Definition at line 177 of file Kokkos_CUDANodeMemoryModelImpl.hpp.

template<class T >

ArrayRCP< T > Kokkos::CUDANodeMemoryModel::viewBufferNonConst	(	ReadWriteOption	rw,
		size_t	size,
		const ArrayRCP< T > &	buff
	)		`[inline, inherited]`

Return a non-const view of a buffer for use on the host.

Parameters:

[in] rw Specifies Kokkos::ReadWrite or Kokkos::WriteOnly. If Kokkos::WriteOnly, the contents of the view are undefined when it is created and must be initialized on the host. However, this prevents the potential need for a copy from device to host memory needed to set the view values as when Kokkos::ReadWrite is specified.

This creates a view of length size, constituting the first size entries of buff, as they exist at the time of view creation.

A non-const view permits changes, which must be copied back to the buffer. This does not occur until all references to the view are deleted. If the buffer is deallocated before the view is deleted, then the copy-back does not occur.