This class provides an operation for evaluating multiple grid points in the domain and doing least squares data mining. More...

#include <OperationMultiEvalStreamingOCLMultiPlatform.hpp>

Inheritance diagram for sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >:

Public Member Functions
double	getDuration () override

void	mult (base::DataVector &alpha, base::DataVector &result) override
	Performs the MultiEval operation \(v:= B^T \alpha\). More...

void	mult (base::DataVector &alpha, base::DataVector &result, size_t startIndexData, size_t endIndexData) override
	Multiplication of \(B^T\) with vector \(\alpha\). More...

void	multTranspose (base::DataVector &source, base::DataVector &result) override
	Performs the transposed MultiEval operation \(v':= B v\). More...

void	multTranspose (base::DataVector &source, base::DataVector &result, size_t startIndexGrid, size_t endIndexGrid) override
	Multiplication of \(B\) with vector \(\alpha\). More...

	OperationMultiEvalStreamingOCLMultiPlatform (base::Grid &grid, base::DataMatrix &dataset, std::shared_ptr< base::OCLManagerMultiPlatform > manager, std::shared_ptr< base::OCLOperationConfiguration > parameters)
	Creates a new instance of the OperationMultiEvalStreamingOCLMultiPlatform class. More...

void	prepare () override
	Creates the internal data structures used by the algorithm. More...

	~OperationMultiEvalStreamingOCLMultiPlatform ()
	Destructor. More...

Public Member Functions inherited from sgpp::base::OperationMultipleEval
void	eval (DataVector &alpha, DataVector &result)
	Evaluate multiple datapoints with the specified grid. More...

virtual std::string	getImplementationName ()
	Name of this implementation of the operation. More...

	OperationMultipleEval (sgpp::base::Grid &grid, DataMatrix &dataset)
	Constructor. More...

virtual	~OperationMultipleEval ()
	Destructor. More...

Protected Attributes
size_t	commonDatasetPadding

size_t	commonGridPadding

size_t	datasetSizePadded

size_t	datasetSizeUnpadded

std::vector< std::shared_ptr< base::OCLDevice > >	devices

size_t	dims

double	duration

size_t	gridSizePadded

size_t	gridSizeUnpadded

std::vector< T >	index

std::vector< T >	kernelDataset

std::vector< T >	level

std::shared_ptr< base::OCLManagerMultiPlatform >	manager

std::vector< StreamingOCLMultiPlatform::KernelMult< T > >	multKernels

std::vector< StreamingOCLMultiPlatform::KernelMultTranspose< T > >	multTransposeKernels

base::SGppStopwatch	myTimer

std::shared_ptr< base::OCLOperationConfiguration >	parameters

base::DataMatrix	preparedDataset

std::shared_ptr< base::QueueLoadBalancerOpenMP >	queueLoadBalancerMult

std::shared_ptr< base::QueueLoadBalancerOpenMP >	queueLoadBalancerMultTranspose

bool	verbose

Protected Attributes inherited from sgpp::base::OperationMultipleEval
DataMatrix &	dataset

Grid &	grid

bool	isPrepared

Detailed Description

template<typename T>
class sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >

This class provides an operation for evaluating multiple grid points in the domain and doing least squares data mining.

This algorithmic variant uses the streaming algorithm for evaluation. It uses high performance OpenCL kernels and is well-suited for large irregular datasets and grids. This class manages one OpenCL kernel for each devices configured using the OCLOperationConfiguration. When a operation is called it triggers the device work by using OpenMP and delegating the work to instances of the kernels. Furthermore, this class converts the received grid and dataset into a representation that is suited for the streaming algorithm.

See also: base::OperationMultipleEval; StreamingOCLMultiPlatform::KernelMult; StreamingOCLMultiPlatform::KernelMultTranspose

Constructor & Destructor Documentation

◆ OperationMultiEvalStreamingOCLMultiPlatform()

template<typename T >

sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::OperationMultiEvalStreamingOCLMultiPlatform	(	base::Grid &	grid,
		base::DataMatrix &	dataset,
		std::shared_ptr< base::OCLManagerMultiPlatform >	manager,
		std::shared_ptr< base::OCLOperationConfiguration >	parameters
	)

inline

Creates a new instance of the OperationMultiEvalStreamingOCLMultiPlatform class.

This class should not be created directly, instead the datadriven operator factory should be used or at least the factory method.

See also: createStreamingOCLMultiPlatformConfigured

Parameters

grid	The grid to evaluate
dataset	The datapoints to evaluate
manager	The OpenCL manager that manages OpenCL internels for this kernel
parameters	The configuration of the kernel leading to different compute kernels

References sgpp::datadriven::StreamingOCLMultiPlatform::Configuration::getKernelName(), sgpp::base::DataMatrix::getNcols(), sgpp::base::DataMatrix::getNrows(), sgpp::base::DataMatrix::getSize(), python.statsfileInfo::i, sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::prepare(), and sgpp::base::DataMatrix::transpose().

◆ ~OperationMultiEvalStreamingOCLMultiPlatform()

template<typename T >

sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::~OperationMultiEvalStreamingOCLMultiPlatform ( )

inline

Destructor.

Member Function Documentation

◆ getDuration()

template<typename T >

double sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::getDuration ( )

inlineoverridevirtual

Returns: The duration of the last call to mult or multTranspose

Implements sgpp::base::OperationMultipleEval.

References sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::duration.

◆ mult() [1/2]

template<typename T >

void sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::mult	(	base::DataVector &	alpha,
		base::DataVector &	result
	)

inlineoverridevirtual

Performs the MultiEval operation \(v:= B^T \alpha\).

Parameters

alpha	The surpluses of the grid
result	A vector that contains the result in the order of the dataset

Implements sgpp::base::OperationMultipleEval.

◆ mult() [2/2]

template<typename T >

void sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::mult	(	base::DataVector &	alpha,
		base::DataVector &	result,
		size_t	startIndexData,
		size_t	endIndexData
	)

inlineoverridevirtual

Multiplication of \(B^T\) with vector \(\alpha\).

This implementation variant produces a partial result passed on the start and end index provided. Not every MultiEval operation implements this method, as it only useful for distributed algorithms.

Parameters

alpha	vector, to which \(B\) is applied. Typically the coefficient vector
result	the result vector of the matrix vector multiplication
startIndexData	begin of the fragment of the dataset to be evaluated
endIndexData	end of the fragment of the dataset to be evaluated

Returns: The results of the evaluation for the specified range of the dataset

Reimplemented from sgpp::base::OperationMultipleEval.

References sgpp::base::DataVector::getSize(), sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::gridSizePadded, python.statsfileInfo::i, sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::prepare(), sgpp::base::SGppStopwatch::start(), and sgpp::base::SGppStopwatch::stop().

◆ multTranspose() [1/2]

template<typename T >

void sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::multTranspose	(	base::DataVector &	source,
		base::DataVector &	result
	)

inlineoverridevirtual

Performs the transposed MultiEval operation \(v':= B v\).

Parameters

source	The vector \(v\)
result	The result of the matrix vector multiplication in the order of grid (of the alpha vector)

Implements sgpp::base::OperationMultipleEval.

◆ multTranspose() [2/2]

template<typename T >

void sgpp::datadriven::StreamingOCLMultiPlatform::OperationMultiEvalStreamingOCLMultiPlatform< T >::multTranspose	(	base::DataVector &	source,
		base::DataVector &	result,
		size_t	startIndexGrid,
		size_t	endIndexGrid
	)

inlineoverridevirtual

Multiplication of \(B\) with vector \(\alpha\).

This implementation variant produces a partial result passed on the start and end index provided. Not every MultiEval operation implements this method, as it only useful for distributed algorithms.

Parameters

source	vector, to which \(B^T\) is applied. Typically the coefficient vector
result	the result vector of the matrix vector multiplication
startIndexGrid	begin of the fragment of the grid to apply the operator to
endIndexGrid	end of the fragment of the grid to be apply the operator to