GPUArray< T > Class Template Reference

A class for handling data simultaneously on the CPU and GPU. More...

#include <gpuarray.h>

Collaboration diagram for GPUArray< T >:

Public Member Functions
	GPUArray (bool neverGPU=false, bool _register=false)
	GPUArray (unsigned int num_elements, bool neverGPU=false, bool _register=false)
	The most common constructor takes in the desired size of the array. More...
virtual	~GPUArray ()
	GPUArray (const GPUArray &from)
GPUArray &	operator= (const GPUArray &rhs)
void	swap (GPUArray &from)
	Swap two GPUarrays efficiently. More...
unsigned int	getNumElements () const
	Get the size of the array. More...
void	setRegistered (bool _reg)
	Switch from simple memcpys to HostRegister pinned memory copies. Not currently fully functional. More...
virtual void	resize (unsigned int num_elements)
	Resize the array...performs operations on both the CPU and GPU. More...

Public Attributes
data_location::Enum	Data_location
	Tracks the current location of the data. More...
bool	noGPU

Protected Member Functions
void	memclear (unsigned int first=0)
T *	acquire (const access_location::Enum location, const access_mode::Enum mode) const
void	release () const

Protected Attributes
T *	d_data
	pointer to memory on device More...
T *	h_data
	pointer to memory on host More...

Private Member Functions
void	allocate ()
void	deallocate ()
void	memcpyDeviceToHost () const
void	memcpyHostToDevice () const
T *	resizeHostArray (unsigned int num_elements)
T *	resizeDeviceArray (unsigned int num_elements)

Private Attributes
unsigned int	Num_elements
	Number of elements. More...
bool	Acquired
	Tracks whether the data has been acquired. More...
bool	RegisterArray
	Tracks whether the data has been acquired. More...

Friends
class	ArrayHandle< T >

Detailed Description

template<class T>
class GPUArray< T >

A class for handling data simultaneously on the CPU and GPU.

This class and accessor are based on GPUArray.h, from the HOOMD-Blue simulation package. It is, however, simplified. It takes care of cuda memory copying for templated arrays. A flag (default to false) when declaring a GPUArray controls whether the memory is HostRegistered but only handles synchronous copy operatations (no Asynch, no HostRegister, etc.) It is also only for 1D arrays of data. Importantly, the user accesses and handles data through the ArrayHandle class.

Constructor & Destructor Documentation

GPUArray() [1/3]

template<class T >

GPUArray< T >::GPUArray	(	bool	neverGPU = false,
		bool	_register = false
	)

GPUArray() [2/3]

template<class T >

GPUArray< T >::GPUArray	(	unsigned int	num_elements,
		bool	neverGPU = false,
		bool	_register = false
	)

The most common constructor takes in the desired size of the array.

~GPUArray()

template<class T >

GPUArray< T >::~GPUArray ( )

virtual

GPUArray() [3/3]

template<class T >

GPUArray< T >::GPUArray

(

const GPUArray< T > &

from

)

Member Function Documentation

operator=()

template<class T >

GPUArray< T > & GPUArray< T >::operator=

(

const GPUArray< T > &

rhs

)

swap()

template<class T >

void GPUArray< T >::swap ( GPUArray< T > &  from )

inline

Swap two GPUarrays efficiently.

a.swap(b) is: GPUArray c(a); a = b; b = c; It just swaps internal pointers

getNumElements()

template<class T>

unsigned int GPUArray< T >::getNumElements ( ) const

inline

Get the size of the array.

Referenced by hyperrectangularCellList::computeAdjacentCells(), neighborList::computeCPU(), hyperrectangularCellList::computeCPU(), landauDeGennesLC::computeFirstDerivatives(), neighborList::computeGPU(), hyperrectangularCellList::computeGPU(), landauDeGennesLC::computeObjectForces(), landauDeGennesLC::computeStressTensors(), cubicLattice::createBoundaryObject(), equationOfMotion::getClassSize(), multirankQTensorLatticeModel::getClassSize(), energyMinimizerGradientDescent::getClassSize(), energyMinimizerLoLBFGS::getClassSize(), energyMinimizerFIRE::getClassSize(), force::getClassSize(), cubicLattice::getClassSize(), landauDeGennesLC::getClassSize(), simpleModel::getClassSize(), gpu_copy_gpuarray(), gpu_gpuarray_dVec_dot_products(), noiseSource::initializeGPURNGs(), hyperrectangularCellList::resetCellSizes(), hyperrectangularCellList::resetCellSizesCPU(), neighborList::resetNeighborsCPU(), neighborList::resetNeighborsGPU(), and simpleModel::setParticlePositions().

setRegistered()

template<class T>

void GPUArray< T >::setRegistered ( bool  _reg )

inline

Switch from simple memcpys to HostRegister pinned memory copies. Not currently fully functional.

resize()

template<class T >

void GPUArray< T >::resize ( unsigned int  num_elements )

virtual

Resize the array...performs operations on both the CPU and GPU.

Referenced by hyperrectangularCellList::computeAdjacentCells(), landauDeGennesLC::computeFirstDerivatives(), landauDeGennesLC::computeObjectForces(), landauDeGennesLC::computeStressTensors(), cubicLattice::createBoundaryObject(), multirankQTensorLatticeModel::determineBufferLayout(), cubicLattice::fillNeighborLists(), gpu_copy_gpuarray(), gpu_gpuarray_dVec_dot_products(), noiseSource::initialize(), energyMinimizerAdam::initializeFromModel(), energyMinimizerNesterovAG::initializeFromModel(), energyMinimizerLoLBFGS::initializeFromModel(), equationOfMotion::initializeFromModel(), energyMinimizerGradientDescent::initializeFromModel(), energyMinimizerFIRE::initializeFromModel(), noiseSource::initializeGPURNGs(), energyMinimizerNesterovAG::initializeParameters(), energyMinimizerLoLBFGS::initializeParameters(), simpleModel::initializeSimpleModel(), multirankQTensorLatticeModel::multirankQTensorLatticeModel(), qTensorLatticeModel::qTensorLatticeModel(), hyperrectangularCellList::resetCellSizes(), hyperrectangularCellList::resetCellSizesCPU(), neighborList::resetNeighborsCPU(), neighborList::resetNeighborsGPU(), hyperrectangularCellList::setGridSize(), energyMinimizerLoLBFGS::setLoLBFGSParameters(), and landauDeGennesLC::setModel().

memclear()

template<class T >

void GPUArray< T >::memclear ( unsigned int  first = 0 )

inlineprotected

Referenced by GPUArray< curandState >::GPUArray().

acquire()

template<class T >

T * GPUArray< T >::acquire ( const access_location::Enum  location, const access_mode::Enum  mode  ) const

inlineprotected

Acquire does all the work, keeping track of when data needs to be copied, etc. It is called by the ArrayHandle class

release()

template<class T>

void GPUArray< T >::release ( ) const

inlineprotected

allocate()

template<class T >

void GPUArray< T >::allocate ( )

inlineprivate

Referenced by GPUArray< curandState >::GPUArray().

deallocate()

template<class T >

void GPUArray< T >::deallocate ( )

inlineprivate

memcpyDeviceToHost()

template<class T >

void GPUArray< T >::memcpyDeviceToHost ( ) const

inlineprivate

memcpyHostToDevice()

template<class T >

void GPUArray< T >::memcpyHostToDevice ( ) const

inlineprivate

resizeHostArray()

template<class T >

T * GPUArray< T >::resizeHostArray ( unsigned int  num_elements )

inlineprivate

resizeDeviceArray()

template<class T >

T * GPUArray< T >::resizeDeviceArray ( unsigned int  num_elements )

inlineprivate

Friends And Related Function Documentation

ArrayHandle< T >

template<class T>

friend class ArrayHandle< T >

friend

Member Data Documentation

Data_location

template<class T>

data_location::Enum GPUArray< T >::Data_location

mutable

Tracks the current location of the data.

Referenced by GPUArray< curandState >::swap().

noGPU

template<class T>

bool GPUArray< T >::noGPU

Referenced by cubicLattice::createBoundaryObject(), equationOfMotion::initializeFromModel(), energyMinimizerGradientDescent::initializeFromModel(), energyMinimizerFIRE::initializeFromModel(), cubicLattice::initializeNSites(), simpleModel::initializeSimpleModel(), landauDeGennesLC::landauDeGennesLC(), multirankQTensorLatticeModel::multirankQTensorLatticeModel(), and qTensorLatticeModel::qTensorLatticeModel().

Num_elements

template<class T>

unsigned int GPUArray< T >::Num_elements

mutableprivate

Number of elements.

Referenced by GPUArray< curandState >::getNumElements(), GPUArray< curandState >::GPUArray(), GPUArray< curandState >::operator=(), GPUArray< curandState >::setRegistered(), and GPUArray< curandState >::swap().

Acquired

template<class T>

bool GPUArray< T >::Acquired

mutableprivate

Tracks whether the data has been acquired.

Referenced by GPUArray< curandState >::release(), and GPUArray< curandState >::swap().

RegisterArray

template<class T>

bool GPUArray< T >::RegisterArray

private

Tracks whether the data has been acquired.

Referenced by GPUArray< curandState >::operator=(), GPUArray< curandState >::setRegistered(), and GPUArray< curandState >::swap().

d_data

template<class T>

T* GPUArray< T >::d_data

mutableprotected

pointer to memory on device

Referenced by GPUArray< curandState >::swap().

h_data

template<class T>

T* GPUArray< T >::h_data

mutableprotected

pointer to memory on host

Referenced by GPUArray< curandState >::GPUArray(), GPUArray< curandState >::setRegistered(), and GPUArray< curandState >::swap().

---

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

• gpuarray.h