Base and interface class for communicators used in LAMA. More...

#include <Communicator.hpp>

Inheritance diagram for lama::Communicator:

Public Types
enum	ThreadSafetyLevel { Funneled = 1, Serialized = 2, Multiple = 3 }
	Enumeration type for supported thread safety levels. More...
Public Member Functions
virtual	~Communicator ()
void	factorize2 (const double sizeX, const double sizeY, PartitionId procgrid[2]) const
void	factorize3 (const double sizeX, const double sizeY, const double sizeZ, PartitionId procgrid[3]) const
void	getGrid2Rank (PartitionId pos[2], const PartitionId procgrid[2]) const
void	getGrid3Rank (PartitionId pos[3], const PartitionId procgrid[3]) const
bool	operator== (const Communicator &other) const
	Equality operator for two communicators.
bool	operator!= (const Communicator &other) const
virtual bool	isEqual (const Communicator &other) const =0
	Virtual method used for the equality operator.
virtual ThreadSafetyLevel	getThreadSafetyLevel () const =0
virtual PartitionId	getSize () const =0
	Number of partitions.
virtual PartitionId	getRank () const =0
	Id of this partitinon.
PartitionId	getNeighbor (int pos) const
	Help routine to get the rank of a neighbored position.
virtual void	all2all (int recvValues, const int sendValues) const =0
	All-to-all exchange of an integer value between all processors.
virtual void	exchangeByPlan (int const recvData, const CommunicationPlan &recvPlan, const int const sendData, const CommunicationPlan &sendPlan) const =0
	Exchange data between all processors by communication plans.
virtual void	exchangeByPlan (float const recvData, const CommunicationPlan &recvPlan, const float const sendData, const CommunicationPlan &sendPlan) const =0
	Exchange of float values.
virtual void	exchangeByPlan (double const recvData, const CommunicationPlan &recvPlan, const double const sendData, const CommunicationPlan &sendPlan) const =0
	Exchange of double values.
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (int const recvData, const CommunicationPlan &recvPlan, const int const sendData, const CommunicationPlan &sendPlan) const =0
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (float const recvData, const CommunicationPlan &recvPlan, const float const sendData, const CommunicationPlan &sendPlan) const =0
	Exchange of float values.
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (double const recvData, const CommunicationPlan &recvPlan, const double const sendData, const CommunicationPlan &sendPlan) const =0
	Exchange of double values.
template<typename T >
void	exchangeByPlan (LAMAArray< T > &recvArray, const CommunicationPlan &recvPlan, const LAMAArray< T > &sendArray, const CommunicationPlan &sendPlan) const
template<typename T >
std::auto_ptr< SyncToken >	exchangeByPlanAsync (LAMAArray< T > &recvArray, const CommunicationPlan &recvPlan, const LAMAArray< T > &sendArray, const CommunicationPlan &sendPlan) const
template<typename T >
void	updateHalo (LAMAArray< T > &haloValues, const LAMAArray< T > &localValues, const Halo &halo) const
	Update of halo array via Halo object.
template<typename T >
std::auto_ptr< SyncToken >	updateHaloAsync (LAMAArray< T > &haloValues, const LAMAArray< T > &localValues, const Halo &halo) const
	Asynchronous update of halo array via Halo object.
template<typename T >
void	shift (LAMAArray< T > &recv, const LAMAArray< T > &send, const int direction) const
	Shift on LAMA arrays.
template<typename T >
std::auto_ptr< SyncToken >	shiftAsync (LAMAArray< T > &recvArray, const LAMAArray< T > &sendArray, const int direction) const
	Asychronous shift on LAMA arrays.
void	computeOwners (const std::vector< IndexType > &requiredIndexes, const Distribution &distribution, std::vector< PartitionId > &owners) const
	Ownership computation for indexes where only each partition individually can determine whether an index is local or not.
virtual void	bcast (double val[], const IndexType n, const PartitionId root) const =0
	Broadcast a typed array from root to all other processors.
virtual void	bcast (float val[], const IndexType n, const PartitionId root) const =0
virtual void	bcast (int val[], const IndexType n, const PartitionId root) const =0
virtual void	scatter (double myvals[], const IndexType n, const PartitionId root, const double allvals[]) const =0
	Scatter of an array of values from root to all other processors.
virtual void	scatter (float myvals[], const IndexType n, const PartitionId root, const float allvals[]) const =0
virtual void	scatter (int myvals[], const IndexType n, const PartitionId root, const int allvals[]) const =0
virtual void	scatter (double myvals[], const IndexType n, const PartitionId root, const double allvals[], const IndexType sizes[]) const =0
	Scatter of an array of values from root to all other processors.
virtual void	scatter (float myvals[], const IndexType n, const PartitionId root, const float allvals[], const IndexType sizes[]) const =0
virtual void	scatter (int myvals[], const IndexType n, const PartitionId root, const int allvals[], const IndexType sizes[]) const =0
virtual void	gather (double allvals[], const IndexType n, const PartitionId root, const double myvals[]) const =0
	Gather of an array of values from all processors to root.
virtual void	gather (float allvals[], const IndexType n, const PartitionId root, const float myvals[]) const =0
virtual void	gather (int allvals[], const IndexType n, const PartitionId root, const int myvals[]) const =0
virtual void	gather (double allvals[], const IndexType n, const PartitionId root, const double myvals[], const IndexType sizes[]) const =0
	Gather of an array of values from all processors to root.
virtual void	gather (float allvals[], const IndexType n, const PartitionId root, const float myvals[], const IndexType sizes[]) const =0
virtual void	gather (int allvals[], const IndexType n, const PartitionId root, const int myvals[], const IndexType sizes[]) const =0
virtual IndexType	shift (double newVals[], const IndexType newSize, const double oldVals[], const IndexType oldSize, const int direction) const =0
	This routine shifts data between neighbored processors.
virtual IndexType	shift (float newVals[], const IndexType newSize, const float oldVals[], const IndexType oldSize, const int direction) const =0
virtual IndexType	shift (int newVals[], const IndexType newSize, const int oldVals[], const IndexType oldSize, const int direction) const =0
virtual std::auto_ptr< SyncToken >	shiftAsync (double newVals[], const double oldVals[], const IndexType size, const int direction) const
	Asynchronous version of shift.
virtual std::auto_ptr< SyncToken >	shiftAsync (float newVals[], const float oldVals[], const IndexType size, const int direction) const
virtual std::auto_ptr< SyncToken >	shiftAsync (int newVals[], const int oldVals[], const IndexType size, const int direction) const
virtual float	sum (const float value) const =0
	Sum operations sum up one single value from each partition to a global value.
virtual double	sum (const double value) const =0
virtual int	sum (const int value) const =0
virtual size_t	sum (const size_t value) const =0
virtual float	min (const float value) const =0
virtual float	max (const float value) const =0
virtual double	min (const double value) const =0
virtual double	max (const double value) const =0
virtual int	min (const int value) const =0
virtual int	max (const int value) const =0
virtual void	maxloc (double &val, int &location, const PartitionId root) const =0
	Maximal value combined with a location value where maximum was found.
virtual void	maxloc (float &val, int &location, const PartitionId root) const =0
virtual void	maxloc (int &val, int &location, const PartitionId root) const =0
virtual void	swap (double val[], const IndexType n, const PartitionId partner) const =0
	Swap of an array with another processor.
virtual void	swap (float val[], const IndexType n, const PartitionId partner) const =0
virtual void	swap (int val[], const IndexType n, const PartitionId partner) const =0
virtual void	gather (std::vector< float > &values, float value) const =0
	Gather single float value from each processor into a vector.
virtual void	synchronize () const =0
	Barrier synchronization between all processors.
virtual void	writeAt (std::ostream &stream) const
	Override virtual method of base class Printable.
const std::string &	getType () const
	getter for the type of a communicator.
Protected Member Functions
	Communicator (const std::string &type)
	LAMA_LOG_DECL_STATIC_LOGGER (logger)
template<typename T >
IndexType	shift0 (T newVals[], const IndexType newSize, const T oldVals[], const IndexType oldSize) const
	Shift implementation for direction == 0, just copies values.
template<typename T >
std::auto_ptr< SyncToken >	defaultShiftAsync (T newVals[], const T oldVals[], const IndexType size, const int direction) const
	Default asynchronous shift uses synchronous shift.
virtual ContextPtr	getCommunicationContext () const =0
	getter for Context needed for Communication
Static Protected Member Functions
static void	getUserProcArray (PartitionId userProcArray[3])
	Read in the environment variable NP4LAMA for user processor array.
Protected Attributes
std::string	mCommunicatorType

Detailed Description

Base and interface class for communicators used in LAMA.

This base class defines pure methods for communication and data exchange between the different partitions of a distributed application. These pure methods must be provided by base classes that implement a full communicator.

Furthermore, this class provides also some higher functionality communication methods especially for LAMA arrays that will use the pure virtual methods.

Communicator objects will be managed by the Communicator factory. All operations on a communicator are const methods. Communicators are referenced via shared pointers so that resources are freed only if no more reference to the communicator exists.

Default copy constructor and assignment operator are disabled.

Member Enumeration Documentation

enum lama::Communicator::ThreadSafetyLevel

Enumeration type for supported thread safety levels.

Enumerator:

Funneled	The Program can be multithreaded but only the master thread should communicate.
Serialized	The Program can be multithreaded but the communicator usage is serialized.
Multiple	The Program can be multithreaded and the communicator can be used by all threads concurrently.

Constructor & Destructor Documentation

lama::Communicator::~Communicator ( ) [virtual]

lama::Communicator::Communicator ( const std::string & type ) [protected]

Member Function Documentation

virtual void lama::Communicator::all2all	(	int *	recvValues,
		const int *	sendValues
	)		const `[pure virtual]`

All-to-all exchange of an integer value between all processors.

Parameters:

[out]	recvValues	will contain one value from each processor
[in]	sendValues	must contain one value for each processor

recvValues and sendValues must both have a size of communicator size. recvValues[i] on processor j contains sendValues[j] of processor i.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by lama::CommunicationPlan::allocateTranspose().

virtual void lama::Communicator::bcast	(	double	val[],
		const IndexType	n,
		const PartitionId	root
	)		const `[pure virtual]`

Broadcast a typed array from root to all other processors.

Parameters:

[in,out]	val	in on root, out on all other processors
[in]	n	number of elements in vector val
[in]	root	processor with correct values of val

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by lama::SparseMatrix< T >::getRow(), and lama::DenseMatrix< T >::getRow().

virtual void lama::Communicator::bcast	(	float	val[],
		const IndexType	n,
		const PartitionId	root
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::bcast	(	int	val[],
		const IndexType	n,
		const PartitionId	root
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

void lama::Communicator::computeOwners	(	const std::vector< IndexType > &	requiredIndexes,
		const Distribution &	distribution,
		std::vector< PartitionId > &	owners
	)		const

Ownership computation for indexes where only each partition individually can determine whether an index is local or not.

This method should not be called for distributions where the owner can be computed by a closed formula.

References lama::WriteAccess< T >::get(), lama::Distribution::getCommunicator(), getRank(), getSize(), lama::Distribution::isLocal(), LAMA_ASSERT_ERROR, LAMA_THROWEXCEPTION, max(), lama::WriteAccess< T >::release(), shift(), and lama::LAMAArray< T >::swap().

Referenced by lama::HaloBuilder::build().

template<typename T >

template LAMA_DLL_IMPORTEXPORT auto_ptr< SyncToken > lama::Communicator::defaultShiftAsync	(	T	newVals[],
		const T	oldVals[],
		const IndexType	size,
		const int	direction
	)		const `[protected]`

Default asynchronous shift uses synchronous shift.

References LAMA_ASSERT_ERROR, and shift().

Referenced by lama::PGASCommunicator::shiftAsync(), lama::MPICommunicator::shiftAsync(), and shiftAsync().

virtual void lama::Communicator::exchangeByPlan	(	int *const	recvData,
		const CommunicationPlan &	recvPlan,
		const int *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Exchange data between all processors by communication plans.

Parameters:

[out]	recvData	buffer for data received from other processors
[in]	recvPlan	number of elements and offsets for receiving
[in]	sendData	buffer for data to send to other processors
[in]	sendPlan	contains number of elements and offsets for sending data

All send and receive data between each pair of processors must be a contiguous part of the sendData or recvData.

The size of recvData must be recvPlan.totalQuantity(). The size of sendData must be sendPlan.totalQuantity().

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by exchangeByPlan(), lama::StorageMethods< T >::exchangeHaloCSR(), lama::Redistributor::exchangeHaloN(), lama::SpecializedJacobi::iterateTyped(), and updateHalo().

virtual void lama::Communicator::exchangeByPlan	(	float *const	recvData,
		const CommunicationPlan &	recvPlan,
		const float *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Exchange of float values.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::exchangeByPlan	(	double *const	recvData,
		const CommunicationPlan &	recvPlan,
		const double *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Exchange of double values.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

template<typename T >

void lama::Communicator::exchangeByPlan	(	LAMAArray< T > &	recvArray,
		const CommunicationPlan &	recvPlan,
		const LAMAArray< T > &	sendArray,
		const CommunicationPlan &	sendPlan
	)		const

References lama::WriteAccess< T >::clear(), exchangeByPlan(), lama::ReadAccess< T >::get(), lama::WriteAccess< T >::get(), getCommunicationContext(), LAMA_ASSERT_ERROR, lama::WriteAccess< T >::resize(), lama::_LAMAArray::size(), and lama::CommunicationPlan::totalQuantity().

virtual std::auto_ptr<SyncToken> lama::Communicator::exchangeByPlanAsync	(	int *const	recvData,
		const CommunicationPlan &	recvPlan,
		const int *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by lama::HaloBuilder::build(), exchangeByPlanAsync(), lama::Redistributor::exchangeHalo(), lama::Redistributor::exchangeVHalo(), and updateHaloAsync().

virtual std::auto_ptr<SyncToken> lama::Communicator::exchangeByPlanAsync	(	float *const	recvData,
		const CommunicationPlan &	recvPlan,
		const float *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Exchange of float values.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual std::auto_ptr<SyncToken> lama::Communicator::exchangeByPlanAsync	(	double *const	recvData,
		const CommunicationPlan &	recvPlan,
		const double *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[pure virtual]`

Exchange of double values.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

template<typename T >

std::auto_ptr< SyncToken > lama::Communicator::exchangeByPlanAsync	(	LAMAArray< T > &	recvArray,
		const CommunicationPlan &	recvPlan,
		const LAMAArray< T > &	sendArray,
		const CommunicationPlan &	sendPlan
	)		const

References exchangeByPlanAsync(), LAMA_ASSERT_ERROR, lama::_LAMAArray::size(), and lama::CommunicationPlan::totalQuantity().

void lama::Communicator::factorize2	(	const double	sizeX,
		const double	sizeY,
		PartitionId	procgrid[2]
	)		const

References getSize(), getUserProcArray(), and LAMA_THROWEXCEPTION.

void lama::Communicator::factorize3	(	const double	sizeX,
		const double	sizeY,
		const double	sizeZ,
		PartitionId	procgrid[3]
	)		const

References getSize(), getUserProcArray(), and LAMA_THROWEXCEPTION.

virtual void lama::Communicator::gather	(	double	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const double	myvals[]
	)		const `[pure virtual]`

Gather of an array of values from all processors to root.

Parameters:

[out]	allvals	values that I receive (size must be n * size() )
[in]	n	number of elements in vector val
[in]	root	processor with values for all processors
[in]	myvals	values for all processors

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by updateHalo().

virtual void lama::Communicator::gather	(	float	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const float	myvals[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::gather	(	int	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const int	myvals[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::gather	(	double	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const double	myvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Gather of an array of values from all processors to root.

Parameters:

[out]	allvals	values that I receive (size must be sum(sizes) )
[in]	n	number of elements in vector val
[in]	root	processor with values for all processors
[in]	myvals	values for all processors
[in]	sizes	number of total values for all processors

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::gather	(	float	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const float	myvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::gather	(	int	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const int	myvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::gather	(	std::vector< float > &	values,
		float	value
	)		const `[pure virtual]`

Gather single float value from each processor into a vector.

Implemented in lama::NoCommunicator, lama::MPICommunicator, and lama::PGASCommunicator.

virtual ContextPtr lama::Communicator::getCommunicationContext ( ) const [protected, pure virtual]

getter for Context needed for Communication

Returns:: the ContextPtr

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

Referenced by exchangeByPlan().

void lama::Communicator::getGrid2Rank	(	PartitionId	pos[2],
		const PartitionId	procgrid[2]
	)		const

References getRank(), getSize(), and LAMA_ASSERT_ERROR.

void lama::Communicator::getGrid3Rank	(	PartitionId	pos[3],
		const PartitionId	procgrid[3]
	)		const

References getRank(), getSize(), and LAMA_ASSERT_ERROR.

PartitionId lama::Communicator::getNeighbor ( int pos ) const [inline]

Help routine to get the rank of a neighbored position.

Parameters:

[in] pos is the distance to the neighbor (also negative)

Returns:: rank of the neighbored processor

This method assumes a circular ring formed by all processors.

References lama::abs(), getRank(), getSize(), and LAMA_ASSERT.

Referenced by lama::DenseMatrix< T >::matrixTimesVectorImpl(), lama::PGASCommunicator::shift(), lama::MPICommunicator::shift(), lama::PGASCommunicator::shiftAsync(), and lama::MPICommunicator::shiftAsync().

virtual PartitionId lama::Communicator::getRank ( ) const [pure virtual]

Id of this partitinon.

Returns:: rank of this partition, 0 <= rank < getSize()

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

Referenced by lama::HaloBuilder::build(), computeOwners(), lama::MatrixCreator< T >::fillRandom(), getGrid2Rank(), getGrid3Rank(), getNeighbor(), lama::SparseMatrix< T >::getRow(), lama::DenseMatrix< T >::getRow(), lama::DenseMatrix< T >::getValue(), lama::OpenMPSCALAPACK::inverse(), lama::DenseMatrix< T >::matrixTimesVectorImpl(), lama::DenseMatrix< T >::setIdentity(), lama::DenseMatrix< T >::writeToFile(), and lama::SparseMatrix< T >::writeToFile().

virtual PartitionId lama::Communicator::getSize ( ) const [pure virtual]

Number of partitions.

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

Referenced by lama::CommunicationPlan::allocateTranspose(), computeOwners(), factorize2(), factorize3(), getGrid2Rank(), getGrid3Rank(), getNeighbor(), lama::OpenMPSCALAPACK::inverse(), lama::Distribution::replicate(), lama::Distribution::replicateN(), lama::DenseMatrix< T >::setIdentity(), shift(), updateHalo(), updateHaloAsync(), and lama::SparseMatrix< T >::writeToFile().

virtual ThreadSafetyLevel lama::Communicator::getThreadSafetyLevel ( ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

const std::string& lama::Communicator::getType ( ) const [inline]

getter for the type of a communicator.

void lama::Communicator::getUserProcArray ( PartitionId userProcArray[3] ) [static, protected]

Read in the environment variable NP4LAMA for user processor array.

Parameters:

[out] userProcArray specifies the user processor array.

Referenced by factorize2(), and factorize3().

virtual bool lama::Communicator::isEqual ( const Communicator & other ) const [pure virtual]

Virtual method used for the equality operator.

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

Referenced by operator!=(), and operator==().

lama::Communicator::LAMA_LOG_DECL_STATIC_LOGGER ( logger ) [protected]

Reimplemented in lama::MPICommunicator, and lama::PGASCommunicator.

virtual float lama::Communicator::max ( const float value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

Referenced by computeOwners(), lama::DenseMatrix< T >::matrixTimesVectorImpl(), lama::DenseMatrix< T >::maxDiffNormImpl(), lama::SparseMatrix< T >::maxDiffNormImpl(), lama::DenseVector< T >::maxNorm(), lama::Distribution::replicate(), and lama::Distribution::replicateN().

virtual double lama::Communicator::max ( const double value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual int lama::Communicator::max ( const int value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual void lama::Communicator::maxloc	(	double &	val,
		int &	location,
		const PartitionId	root
	)		const `[pure virtual]`

Maximal value combined with a location value where maximum was found.

Parameters:

[in,out]	val	is a value on each processor, only out for root with maximal value
[in,out]	location	is an additional int value, only out for root

Only root processor will contain the maximal value and the location loc.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::maxloc	(	float &	val,
		int &	location,
		const PartitionId	root
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::maxloc	(	int &	val,
		int &	location,
		const PartitionId	root
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual float lama::Communicator::min ( const float value ) const [pure virtual]

Implemented in lama::NoCommunicator, lama::MPICommunicator, and lama::PGASCommunicator.

virtual double lama::Communicator::min ( const double value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual int lama::Communicator::min ( const int value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

bool lama::Communicator::operator!= ( const Communicator & other ) const

References isEqual().

bool lama::Communicator::operator== ( const Communicator & other ) const

Equality operator for two communicators.

Parameters:

[in] other communicator for comparison

Returns:: true if this communicator is equal to other communicator.

Note: Logic of this operator is implemented via the virtual functions isEqual.

References isEqual().

virtual void lama::Communicator::scatter	(	double	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const double	allvals[]
	)		const `[pure virtual]`

Scatter of an array of values from root to all other processors.

Parameters:

[out]	myvals	values that I receive
[in]	n	number of elements in vector val
[in]	root	processor with values for all processors
[in]	allvals	values for all processors (size must be n * size() )

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::scatter	(	float	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const float	allvals[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::scatter	(	int	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const int	allvals[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::scatter	(	double	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const double	allvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Scatter of an array of values from root to all other processors.

Parameters:

[out]	myvals	values that I receive
[in]	n	number of elements in vector val
[in]	root	processor with values for all processors
[in]	allvals	values for all processors (size must be sum(sizes) )
[in]	sizes	number of total values for all processors

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::scatter	(	float	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const float	allvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::scatter	(	int	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const int	allvals[],
		const IndexType	sizes[]
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

template<typename T >

void lama::Communicator::shift	(	LAMAArray< T > &	recv,
		const LAMAArray< T > &	send,
		const int	direction
	)		const

Shift on LAMA arrays.

Parameters:

[in]	recv	array to receive for this partition
[out]	send	array to send from this partition
[in]	direction	number of positions to shift, e.g. 1 or -1

Note: The recv array must have a capacity that is sufficent to receive all the data.

References lama::WriteAccess< T >::capacity(), lama::ReadAccess< T >::get(), lama::WriteAccess< T >::get(), getSize(), LAMA_ASSERT_ERROR, lama::WriteAccess< T >::resize(), and lama::ReadAccess< T >::size().

Referenced by computeOwners(), defaultShiftAsync(), lama::DenseMatrix< T >::matrixTimesVectorImpl(), lama::Distribution::replicate(), and lama::Distribution::replicateN().

virtual IndexType lama::Communicator::shift	(	double	newVals[],
		const IndexType	newSize,
		const double	oldVals[],
		const IndexType	oldSize,
		const int	direction
	)		const `[pure virtual]`

This routine shifts data between neighbored processors.

Parameters:

[out]	newVals	array with data this partition get from neighbored partition
[in]	newSize	allocated size of array newVals
[in]	oldVals	array with data this partition sends to neighbored partition
[in]	oldSize	number of elements of array oldVals that will be sent
[in]	direction	specifies the neighbored partitions to send and receive

Returns:: number of values received in newVals (newSize must have been larger before)

Each partition sends to rank() + direction and receives from rank() - direction.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual IndexType lama::Communicator::shift	(	float	newVals[],
		const IndexType	newSize,
		const float	oldVals[],
		const IndexType	oldSize,
		const int	direction
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual IndexType lama::Communicator::shift	(	int	newVals[],
		const IndexType	newSize,
		const int	oldVals[],
		const IndexType	oldSize,
		const int	direction
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

template<typename T >

template LAMA_DLL_IMPORTEXPORT IndexType lama::Communicator::shift0	(	T	newVals[],
		const IndexType	newSize,
		const T	oldVals[],
		const IndexType	oldSize
	)		const `[protected]`

Shift implementation for direction == 0, just copies values.

References LAMA_ASSERT_ERROR.

Referenced by lama::NoCommunicator::shift(), lama::PGASCommunicator::shift(), and lama::MPICommunicator::shift().

template<typename T >

auto_ptr< SyncToken > lama::Communicator::shiftAsync	(	LAMAArray< T > &	recvArray,
		const LAMAArray< T > &	sendArray,
		const int	direction
	)		const

Asychronous shift on LAMA arrays.

Parameters:

[out]	recvArray	array to receive for this partition
[in]	sendArray	array to send from this partition
[in]	direction	number of positions to shift, e.g. 1 or -1

Note: All partitions must have the same size for send/recv array

References lama::LAMAArray< T >::clear(), LAMA_ASSERT_DEBUG, and LAMA_ASSERT_ERROR.

Referenced by lama::DenseMatrix< T >::matrixTimesVectorImpl().

std::auto_ptr< SyncToken > lama::Communicator::shiftAsync	(	double	newVals[],
		const double	oldVals[],
		const IndexType	size,
		const int	direction
	)		const `[virtual]`

Asynchronous version of shift.

As there is no information about the received size this routine can only be called for arrays that have the same size on all partitions.

A default implementation is provided that returns a NoSyncToken. Derived classes should override this method if there is a benefit of using asynchronous transfers.

Reimplemented in lama::MPICommunicator, and lama::PGASCommunicator.

References defaultShiftAsync().

std::auto_ptr< SyncToken > lama::Communicator::shiftAsync	(	float	newVals[],
		const float	oldVals[],
		const IndexType	size,
		const int	direction
	)		const `[virtual]`

Reimplemented in lama::MPICommunicator, and lama::PGASCommunicator.

References defaultShiftAsync().

std::auto_ptr< SyncToken > lama::Communicator::shiftAsync	(	int	newVals[],
		const int	oldVals[],
		const IndexType	size,
		const int	direction
	)		const `[virtual]`

Reimplemented in lama::MPICommunicator, and lama::PGASCommunicator.

References defaultShiftAsync().

virtual float lama::Communicator::sum ( const float value ) const [pure virtual]

Sum operations sum up one single value from each partition to a global value.

Parameters:

[in] value value on the calling partition

Returns:: global value, available for all partitions.

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

Referenced by lama::SparseMatrix< T >::getRow(), lama::DenseMatrix< T >::getRow(), lama::DenseMatrix< T >::getValue(), and lama::SparseMatrix< T >::getValue().

virtual double lama::Communicator::sum ( const double value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual int lama::Communicator::sum ( const int value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual size_t lama::Communicator::sum ( const size_t value ) const [pure virtual]

Implemented in lama::MPICommunicator, lama::NoCommunicator, and lama::PGASCommunicator.

virtual void lama::Communicator::swap	(	double	val[],
		const IndexType	n,
		const PartitionId	partner
	)		const `[pure virtual]`

Swap of an array with another processor.

Parameters:

[in,out]	val	is the data array to be swapped
[in]	n	is the number of entries in array val
[in]	partner	is the rank of partition with which this partition swaps

This method can also be used if partner is same as this processor.

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::swap	(	float	val[],
		const IndexType	n,
		const PartitionId	partner
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::swap	(	int	val[],
		const IndexType	n,
		const PartitionId	partner
	)		const `[pure virtual]`

Implemented in lama::MPICommunicator, lama::PGASCommunicator, and lama::NoCommunicator.

virtual void lama::Communicator::synchronize ( ) const [pure virtual]

Barrier synchronization between all processors.

Implemented in lama::NoCommunicator, lama::MPICommunicator, and lama::PGASCommunicator.

Referenced by lama::CommunicationPlan::allocateTranspose(), lama::DenseMatrix< T >::writeToFile(), and lama::SparseMatrix< T >::writeToFile().

template<typename T >

template LAMA_DLL_IMPORTEXPORT void lama::Communicator::updateHalo	(	LAMAArray< T > &	haloValues,
		const LAMAArray< T > &	localValues,
		const Halo &	halo
	)		const

Update of halo array via Halo object.

Parameters:

[out]	haloValues	will contain the non-local values from other processors
[in]	localValues	is the local part of the array on each processor
[in]	halo	is the Halo object containing all information about exchange

This method is not virtual but will use the pure virtual methods of base classes.

< temporary array for send communication

References lama::CommunicationPlan::allocated(), exchangeByPlan(), gather(), lama::Halo::getProvidesIndexes(), lama::Halo::getProvidesPlan(), lama::Halo::getRequiredPlan(), getSize(), LAMA_ASSERT_ERROR, LAMA_REGION, and lama::CommunicationPlan::size().

template<typename T >

template LAMA_DLL_IMPORTEXPORT auto_ptr< SyncToken > lama::Communicator::updateHaloAsync	(	LAMAArray< T > &	haloValues,
		const LAMAArray< T > &	localValues,
		const Halo &	halo
	)		const