A templatized class to encapsulate a device global memory pointer in a std::vector like interface complete with most of the usual methods. This class is intended to be used only in host code and does not work device side; Passing the pointer to a kernel can be done with the data() method. This class works for any device side pointer, scalar or array valued. More...

#include <DeviceVector.h>

Public Types
typedef T	value_type

Public Member Functions
	DeviceVector () noexcept

	DeviceVector (size_t const size, bool const initialize=false)
	Construct a new Device Vector object by calling the `_allocate` private method.

	DeviceVector (DeviceVector< T > &&other) noexcept

DeviceVector &	operator= (DeviceVector< T > &&other) noexcept

	~DeviceVector ()
	Destroy the Device Vector object by calling the `_deAllocate` private method.

	DeviceVector (const DeviceVector< T > &)=delete

DeviceVector< T > &	operator= (const DeviceVector< T > &other)=delete

T *	data ()
	Get the raw device pointer.

size_t	size ()
	Get the number of elements in the array.

T	operator[] (size_t const &index)
	Overload the [] operator to return a value from device memory. This method performs a cudaMemcpy to copy the desired element to the host then returns it. Unlike the `at()` method this method does not perform bounds checking.

T	at (size_t const index)
	Return a value from device memory. This method performs a cudaMemcpy to copy the desired element to the host then returns it. Unlike the `[]` overload this method perform bounds checking.

void	assign (T const &hostValue, size_t const &index=0)
	Assign a single value in the array. Should generally only be used when the pointer points to a scalar value. By default this writes `hostValue` to the 0th element of the array.

void	resize (size_t const newSize)
	Resize the device container to contain `newSize` elements. If `newSize` is greater than the current size then all the values are kept and the rest of the array is default initialized. If `newSize` is smaller than the current size then the array is truncated and values at locations greater than `newSize` are lost. Keeping the values in the array requires that the new array be allocated, the values be copied, then the old array be freed; as such this method is quite slow and can use a large amount of memory. If you don't care about the values in the array then use the `reset` method.

void	reset (size_t const newSize)
	Reset the size of the array. This frees the old array and allocates a new one; all values in the array may be lost. The values in memory are not initialized and therefore the behaviour of the default values is undefined.

void	cpyHostToDevice (const T *arrIn, size_t const &arrSize)
	Copy the first `arrSize` elements of `arrIn` to the device.

void	cpyHostToDevice (std::vector< T > const &vecIn)
	Copy the contents of a std::vector to the device.

void	cpyDeviceToHost (T *arrOut, size_t const &arrSize)
	Copy the array from the device to a host array. Checks if the host array is large enough based on the `arrSize` parameter.

void	cpyDeviceToHost (std::vector< T > &vecOut)
	Copy the array from the device to a host std::vector. Checks if the host array is large enough.

Detailed Description

template<typename T>
class cuda_utilities::DeviceVector< T >

A templatized class to encapsulate a device global memory pointer in a std::vector like interface complete with most of the usual methods. This class is intended to be used only in host code and does not work device side; Passing the pointer to a kernel can be done with the data() method. This class works for any device side pointer, scalar or array valued.

Template Parameters

T	Any trivially copyable type where `sizeof(T)` returns correct results should work, but non-primitive types have not been tested.

Constructor & Destructor Documentation

◆ DeviceVector() [1/3]

template<typename T >

cuda_utilities::DeviceVector< T >::DeviceVector ( )

inlinenoexcept

Construct an empty DeviceVector

◆ DeviceVector() [2/3]

template<typename T >

cuda_utilities::DeviceVector< T >::DeviceVector	(	size_t const	size,
		bool const	initialize = `false`
	)

Construct a new Device Vector object by calling the _allocate private method.

Parameters

[in]	size	The number of elements desired in the array. Can be any positive integer.
[in]	initialize	(optional) If true then initialize the GPU memory to int(0)

◆ DeviceVector() [3/3]

template<typename T >

cuda_utilities::DeviceVector< T >::DeviceVector ( DeviceVector< T > && other )

inlinenoexcept

Define a move constructor

Member Function Documentation

◆ assign()

template<typename T >

void cuda_utilities::DeviceVector< T >::assign	(	T const &	hostValue,
		size_t const &	index = `0`
	)

Assign a single value in the array. Should generally only be used when the pointer points to a scalar value. By default this writes hostValue to the 0th element of the array.

Parameters

[in]	hostValue	The value to write to the device array
[in]	index	The location to write the value to, defaults to zero.

◆ at()

template<typename T >

T cuda_utilities::DeviceVector< T >::at ( size_t const index )

Return a value from device memory. This method performs a cudaMemcpy to copy the desired element to the host then returns it. Unlike the [] overload this method perform bounds checking.

Parameters

[in] index The index of the desired value

Returns: T The value at dev_ptr[index]

◆ cpyDeviceToHost() [1/2]

template<typename T >

void cuda_utilities::DeviceVector< T >::cpyDeviceToHost ( std::vector< T > & vecOut )

inline

Copy the array from the device to a host std::vector. Checks if the host array is large enough.

Parameters

[out] vecOut The std::vector to copy the device array into

◆ cpyDeviceToHost() [2/2]

template<typename T >

void cuda_utilities::DeviceVector< T >::cpyDeviceToHost	(	T *	arrOut,
		size_t const &	arrSize
	)

Copy the array from the device to a host array. Checks if the host array is large enough based on the arrSize parameter.

Parameters

[out]	arrOut	The pointer to the host array
[in]	arrSize	The number of elements allocated in the host array

◆ cpyHostToDevice() [1/2]

template<typename T >

void cuda_utilities::DeviceVector< T >::cpyHostToDevice	(	const T *	arrIn,
		size_t const &	arrSize
	)

Copy the first arrSize elements of arrIn to the device.

Parameters

[in]	arrIn	The pointer to the array to be copied to the device
[in]	arrSize	The number of elements/size of the array to copy to the device

◆ cpyHostToDevice() [2/2]

template<typename T >

void cuda_utilities::DeviceVector< T >::cpyHostToDevice ( std::vector< T > const & vecIn )

inline

Copy the contents of a std::vector to the device.

Parameters

[in] vecIn The array whose contents are to be copied

◆ data()

template<typename T >

T * cuda_utilities::DeviceVector< T >::data ( )

inline

Get the raw device pointer.

Returns: T* The pointer for the array in global memory

◆ operator=()

template<typename T >

DeviceVector & cuda_utilities::DeviceVector< T >::operator= ( DeviceVector< T > && other )

inlinenoexcept

Define move assignment

◆ operator[]()

template<typename T >

T cuda_utilities::DeviceVector< T >::operator[] ( size_t const & index )

Overload the [] operator to return a value from device memory. This method performs a cudaMemcpy to copy the desired element to the host then returns it. Unlike the at() method this method does not perform bounds checking.

Parameters

[in] index The index of the desired value

Returns: T The value at dev_ptr[index]

◆ reset()

template<typename T >

void cuda_utilities::DeviceVector< T >::reset ( size_t const newSize )

Reset the size of the array. This frees the old array and allocates a new one; all values in the array may be lost. The values in memory are not initialized and therefore the behaviour of the default values is undefined.

Parameters

newSize

◆ resize()

template<typename T >

void cuda_utilities::DeviceVector< T >::resize ( size_t const newSize )

Resize the device container to contain newSize elements. If newSize is greater than the current size then all the values are kept and the rest of the array is default initialized. If newSize is smaller than the current size then the array is truncated and values at locations greater than newSize are lost. Keeping the values in the array requires that the new array be allocated, the values be copied, then the old array be freed; as such this method is quite slow and can use a large amount of memory. If you don't care about the values in the array then use the reset method.

Parameters

[in] newSize The desired size of the array

◆ size()

template<typename T >

size_t cuda_utilities::DeviceVector< T >::size ( )

inline

Get the number of elements in the array.

Returns: size_t The number of elements in the array

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

DeviceVector.h

Public Types

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ DeviceVector() [1/3]

◆ DeviceVector() [2/3]

◆ DeviceVector() [3/3]

Member Function Documentation

◆ assign()

◆ at()

◆ cpyDeviceToHost() [1/2]

◆ cpyDeviceToHost() [2/2]

◆ cpyHostToDevice() [1/2]

◆ cpyHostToDevice() [2/2]

◆ data()

◆ operator=()

◆ operator[]()

◆ reset()

◆ resize()

◆ size()