Communicator class that implements communication and data exchange via MPI. More...

#include <MPICommunicator.hpp>

Inheritance diagram for lama::MPICommunicator:

Public Types
enum	ThreadSafetyLevel { Funneled = 1, Serialized = 2, Multiple = 3 }
	Enumeration type for supported thread safety levels. More...
Public Member Functions
virtual	~MPICommunicator ()
virtual bool	isEqual (const Communicator &other) const
	Virtual method used for the equality operator.
virtual ThreadSafetyLevel	getThreadSafetyLevel () const
virtual PartitionId	getSize () const
	Number of partitions.
virtual PartitionId	getRank () const
	Id of this partitinon.
MPI_Comm	getMPIComm () const
virtual void	all2all (int recvValues, const int sendValues) const
	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
	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
	Exchange of float values.
virtual void	exchangeByPlan (double const recvData, const CommunicationPlan &recvPlan, const double const sendData, const CommunicationPlan &sendPlan) const
	Exchange of double values.
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (int const recvData, const CommunicationPlan &recvPlan, const int const sendData, const CommunicationPlan &sendPlan) const
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (float const recvData, const CommunicationPlan &recvPlan, const float const sendData, const CommunicationPlan &sendPlan) const
	Exchange of float values.
virtual std::auto_ptr< SyncToken >	exchangeByPlanAsync (double const recvData, const CommunicationPlan &recvPlan, const double const sendData, const CommunicationPlan &sendPlan) const
	Exchange of double values.
template<typename T >
MPI_Request	startrecv (T *buffer, int count, int source) const
template<typename T >
MPI_Request	startsend (const T *buffer, int count, int target) const
virtual float	sum (const float value) const
	Sum operations sum up one single value from each partition to a global value.
virtual double	sum (const double value) const
virtual int	sum (const int value) const
virtual float	min (const float value) const
virtual size_t	sum (const size_t value) const
virtual float	max (const float value) const
virtual double	min (const double value) const
virtual double	max (const double value) const
virtual int	min (const int value) const
virtual int	max (const int value) const
virtual void	gather (std::vector< float > &values, float value) const
	Gather single float value from each processor into a vector.
virtual void	synchronize () const
	Barrier synchronization between all processors.
virtual IndexType	shift (double newVals[], const IndexType newSize, const double oldVals[], const IndexType oldSize, const int direction) const
	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
virtual IndexType	shift (int newVals[], const IndexType newSize, const int oldVals[], const IndexType oldSize, const int direction) const
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 void	bcast (double val[], const IndexType n, const PartitionId root) const
	Broadcast an array of doubles from root to all processors in simulation.
virtual void	bcast (int val[], const IndexType n, const PartitionId root) const
	Broadcast an array of int from root to all processors in simulation.
virtual void	bcast (float val[], const IndexType n, const PartitionId root) const
	Broadcast an array of float from root to all processors in simulation.
virtual void	scatter (double myvals[], const IndexType n, const PartitionId root, const double allvals[]) const
	scatter
virtual void	scatter (float myvals[], const IndexType n, const PartitionId root, const float allvals[]) const
virtual void	scatter (int myvals[], const IndexType n, const PartitionId root, const int allvals[]) const
virtual void	scatter (double myvals[], const IndexType n, const PartitionId root, const double allvals[], const IndexType sizes[]) const
	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
virtual void	scatter (int myvals[], const IndexType n, const PartitionId root, const int allvals[], const IndexType sizes[]) const
virtual void	gather (double allvals[], const IndexType n, const PartitionId root, const double myvals[]) const
	gather
virtual void	gather (float allvals[], const IndexType n, const PartitionId root, const float myvals[]) const
virtual void	gather (int allvals[], const IndexType n, const PartitionId root, const int myvals[]) const
virtual void	gather (double allvals[], const IndexType n, const PartitionId root, const double myvals[], const IndexType sizes[]) const
	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
virtual void	gather (int allvals[], const IndexType n, const PartitionId root, const int myvals[], const IndexType sizes[]) const
virtual void	maxloc (double &val, int &location, const PartitionId root) const
	Maximal value combined with a location value where maximum was found.
virtual void	maxloc (float &val, int &location, const PartitionId root) const
virtual void	maxloc (int &val, int &location, const PartitionId root) const
virtual void	swap (double val[], const IndexType n, const PartitionId partner) const
	This routine swaps an array with a partner processor.
virtual void	swap (float val[], const IndexType n, const PartitionId partner) const
virtual void	swap (int val[], const IndexType n, const PartitionId partner) const
virtual void	writeAt (std::ostream &stream) const
	Override virtual method of base class Printable.
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
PartitionId	getNeighbor (int pos) const
	Help routine to get the rank of a neighbored position.
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.
const std::string &	getType () const
	getter for the type of a communicator.
Protected Member Functions
virtual ContextPtr	getCommunicationContext () const
	getter for Context needed for Communication
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.
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
Private Member Functions
	MPICommunicator (int &argc, char **&argv)
template<typename T >
T	sumImpl (T myVal) const
template<typename T >
T	maxval (T myVal) const
template<typename T >
T	minval (T myVal) const
template<typename T >
void	send (const T *buffer, int count, int target) const
template<typename T >
int	getCount (MPI_Status &status) const
template<typename T >
void	scatterImpl (T myvals[], const IndexType n, const PartitionId root, const T allvals[]) const
template<typename T >
void	scatterImpl (T myvals[], const IndexType n, const PartitionId root, const T allvals[], const IndexType sizes[]) const
template<typename T >
void	gatherImpl (T allvals[], const IndexType n, const PartitionId root, const T myvals[]) const
template<typename T >
void	gatherImpl (T allvals[], const IndexType n, const PartitionId root, const T myvals[], const IndexType sizes[]) const
template<typename T >
IndexType	shiftImpl (T newvals[], const IndexType newSize, const PartitionId source, const T oldVals[], const IndexType oldSize, const PartitionId dest) const
template<typename T >
std::auto_ptr< SyncToken >	shiftAsyncImpl (T newvals[], const PartitionId source, const T oldVals[], const PartitionId dest, const IndexType size) const
template<typename T >
void	swapImpl (T val[], const IndexType n, PartitionId partner) const
template<typename T >
void	maxlocImpl (T &val, int &location, PartitionId root) const
template<typename T >
void	exchangeByPlanImpl (T const recvData, const CommunicationPlan &recvPlan, const T const sendData, const CommunicationPlan &sendPlan) const
template<typename T >
std::auto_ptr< SyncToken >	exchangeByPlanAsyncImpl (T const recvData, const CommunicationPlan &recvPlan, const T const sendData, const CommunicationPlan &sendPlan) const
MPI_Comm	selectMPIComm () const
	LAMA_LOG_DECL_STATIC_LOGGER (logger)
Static Private Member Functions
template<typename T >
static MPI_Datatype	getMPIType ()
template<typename T1 , typename T2 >
static MPI_Datatype	getMPI2Type ()
Private Attributes
MPI_Comm	mCommWorld
MPI_Comm	mComm
MPI_Comm	mCommTask
const boost::thread	mMainThread
Communicator::ThreadSafetyLevel	mThreadSafetyLevel
int	mRank
int	mSize
bool	mExternInitialization
Static Private Attributes
static const int	defaultTag = 1
Friends
class	MPICommunicatorManager

Detailed Description

Communicator class that implements communication and data exchange via MPI.

MPI_Init is called in the constructor, MPI_Finalize is called in the destructor.

Member Enumeration Documentation

enum lama::Communicator::ThreadSafetyLevel [inherited]

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::MPICommunicator::~MPICommunicator ( ) [virtual]

References LAMA_MPICALL, and mExternInitialization.

lama::MPICommunicator::MPICommunicator	(	int &	argc,
		char **&	argv
	)		`[private]`

References lama::Communicator::Funneled, LAMA_MPICALL, LAMA_THROWEXCEPTION, mComm, mCommTask, mCommWorld, mExternInitialization, mRank, mSize, mThreadSafetyLevel, lama::Communicator::Multiple, and lama::Communicator::Serialized.

Member Function Documentation

void lama::MPICommunicator::all2all	(	int *	recvValues,
		const int *	sendValues
	)		const `[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.

Implements lama::Communicator.

References LAMA_ASSERT_ERROR, LAMA_MPICALL, and selectMPIComm().

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

Broadcast an array of doubles from root to all processors in simulation.

Implements lama::Communicator.

References LAMA_MPICALL, and selectMPIComm().

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

Broadcast an array of int from root to all processors in simulation.

Implements lama::Communicator.

References LAMA_MPICALL, and selectMPIComm().

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

Broadcast an array of float from root to all processors in simulation.

Implements lama::Communicator.

References LAMA_MPICALL, and selectMPIComm().

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

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(), lama::Communicator::getRank(), lama::Communicator::getSize(), lama::Distribution::isLocal(), LAMA_ASSERT_ERROR, LAMA_THROWEXCEPTION, lama::Communicator::max(), lama::WriteAccess< T >::release(), lama::Communicator::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, inherited]`

Default asynchronous shift uses synchronous shift.

References LAMA_ASSERT_ERROR, and lama::Communicator::shift().

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

void lama::MPICommunicator::exchangeByPlan	(	int *const	recvData,
		const CommunicationPlan &	recvPlan,
		const int *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[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

Implements lama::Communicator.

References exchangeByPlanImpl().

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

Exchange of float values.

Implements lama::Communicator.

References exchangeByPlanImpl().

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

Exchange of double values.

Implements lama::Communicator.

References exchangeByPlanImpl().

template<typename T >

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

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

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

Implements lama::Communicator.

References exchangeByPlanAsyncImpl().

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

Exchange of float values.

Implements lama::Communicator.

References exchangeByPlanAsyncImpl().

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

Exchange of double values.

Implements lama::Communicator.

References exchangeByPlanAsyncImpl().

template<typename T >

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

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

template<typename T >

auto_ptr< SyncToken > lama::MPICommunicator::exchangeByPlanAsyncImpl	(	T *const	recvData,
		const CommunicationPlan &	recvPlan,
		const T *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[private]`

References lama::CommunicationPlan::allocated(), LAMA_ASSERT_DEBUG, LAMA_ASSERT_ERROR, mRank, lama::MPISyncToken::pushRequest(), lama::CommunicationPlan::size(), startrecv(), and startsend().

Referenced by exchangeByPlanAsync().

template<typename T >

void lama::MPICommunicator::exchangeByPlanImpl	(	T *const	recvData,
		const CommunicationPlan &	recvPlan,
		const T *const	sendData,
		const CommunicationPlan &	sendPlan
	)		const `[private]`

References lama::CommunicationPlan::allocated(), LAMA_ASSERT_DEBUG, LAMA_ASSERT_ERROR, LAMA_MPICALL, LAMA_REGION, mRank, send(), lama::CommunicationPlan::size(), and startrecv().

Referenced by exchangeByPlan().

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

References lama::Communicator::getSize(), lama::Communicator::getUserProcArray(), and LAMA_THROWEXCEPTION.

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

References lama::Communicator::getSize(), lama::Communicator::getUserProcArray(), and LAMA_THROWEXCEPTION.

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

Gather single float value from each processor into a vector.

Implements lama::Communicator.

References LAMA_MPICALL, mSize, and selectMPIComm().

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

gather

Implements lama::Communicator.

References gatherImpl().

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

Implements lama::Communicator.

References gatherImpl().

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

Implements lama::Communicator.

References gatherImpl().

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

Implements lama::Communicator.

References gatherImpl().

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

Implements lama::Communicator.

References gatherImpl().

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

Implements lama::Communicator.

References gatherImpl().

template<typename T >

void lama::MPICommunicator::gatherImpl	(	T	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const T	myvals[]
	)		const `[private]`

References getSize(), LAMA_ASSERT_DEBUG, LAMA_MPICALL, and selectMPIComm().

Referenced by gather().

template<typename T >

void lama::MPICommunicator::gatherImpl	(	T	allvals[],
		const IndexType	n,
		const PartitionId	root,
		const T	myvals[],
		const IndexType	sizes[]
	)		const `[private]`

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

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

getter for Context needed for Communication

Returns:: the ContextPtr

Implements lama::Communicator.

References lama::ContextFactory::getContext(), and lama::Context::Host.

template<typename T >

int lama::MPICommunicator::getCount ( MPI_Status & status ) const [inline, private]

References LAMA_MPICALL.

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

References lama::Communicator::getRank(), lama::Communicator::getSize(), and LAMA_ASSERT_ERROR.

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

References lama::Communicator::getRank(), lama::Communicator::getSize(), and LAMA_ASSERT_ERROR.

MPI_Datatype lama::MPICommunicator::getMPI2Type< int, int > ( ) [inline, static, private]

References LAMA_THROWEXCEPTION.

MPI_Comm lama::MPICommunicator::getMPIComm ( ) const

References mComm.

MPI_Datatype lama::MPICommunicator::getMPIType< unsigned long > ( ) [inline, static, private]

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

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(), lama::Communicator::getRank(), lama::Communicator::getSize(), and LAMA_ASSERT.

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

PartitionId lama::MPICommunicator::getRank ( ) const [virtual]

Id of this partitinon.

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

Implements lama::Communicator.

References mRank.

Referenced by gatherImpl(), scatterImpl(), shiftAsyncImpl(), and shiftImpl().

PartitionId lama::MPICommunicator::getSize ( ) const [virtual]

Number of partitions.

Implements lama::Communicator.

References mSize.

Referenced by gatherImpl(), scatterImpl(), shift(), and shiftAsync().

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

Implements lama::Communicator.

References mThreadSafetyLevel.

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

getter for the type of a communicator.

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

Read in the environment variable NP4LAMA for user processor array.

Parameters:

[out] userProcArray specifies the user processor array.

Referenced by lama::Communicator::factorize2(), and lama::Communicator::factorize3().

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

Virtual method used for the equality operator.

Implements lama::Communicator.

References mComm.

lama::MPICommunicator::LAMA_LOG_DECL_STATIC_LOGGER ( logger ) [private]

Reimplemented from lama::Communicator.

float lama::MPICommunicator::max ( const float value ) const [virtual]

Implements lama::Communicator.

References maxval().

double lama::MPICommunicator::max ( const double value ) const [virtual]

Implements lama::Communicator.

References maxval().

int lama::MPICommunicator::max ( const int value ) const [virtual]

Implements lama::Communicator.

References maxval().

void lama::MPICommunicator::maxloc	(	double &	val,
		int &	location,
		const PartitionId	root
	)		const `[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.

Implements lama::Communicator.

References maxlocImpl().

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

Implements lama::Communicator.

References maxlocImpl().

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

Implements lama::Communicator.

References maxlocImpl().

template<typename T >

void lama::MPICommunicator::maxlocImpl	(	T &	val,
		int &	location,
		PartitionId	root
	)		const `[private]`

References mRank, and selectMPIComm().

Referenced by maxloc().

template<typename T >

T lama::MPICommunicator::maxval ( T myVal ) const [private]

References LAMA_MPICALL, and selectMPIComm().

Referenced by max().

float lama::MPICommunicator::min ( const float value ) const [virtual]

Implements lama::Communicator.

References minval().

double lama::MPICommunicator::min ( const double value ) const [virtual]

Implements lama::Communicator.

References minval().

int lama::MPICommunicator::min ( const int value ) const [virtual]

Implements lama::Communicator.

References minval().

template<typename T >

T lama::MPICommunicator::minval ( T myVal ) const [private]

References LAMA_MPICALL, and selectMPIComm().

Referenced by min().

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

References lama::Communicator::isEqual().

bool lama::Communicator::operator== ( const Communicator & other ) const [inherited]

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 lama::Communicator::isEqual().

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

scatter

Implements lama::Communicator.

References scatterImpl().

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

Implements lama::Communicator.

References scatterImpl().

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

Implements lama::Communicator.

References scatterImpl().

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

Implements lama::Communicator.

References scatterImpl().

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

Implements lama::Communicator.

References scatterImpl().

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

Implements lama::Communicator.

References scatterImpl().

template<typename T >

void lama::MPICommunicator::scatterImpl	(	T	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const T	allvals[]
	)		const `[private]`

References getSize(), LAMA_ASSERT_DEBUG, LAMA_MPICALL, and selectMPIComm().

Referenced by scatter().

template<typename T >

void lama::MPICommunicator::scatterImpl	(	T	myvals[],
		const IndexType	n,
		const PartitionId	root,
		const T	allvals[],
		const IndexType	sizes[]
	)		const `[private]`

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

MPI_Comm lama::MPICommunicator::selectMPIComm ( ) const [inline, private]

References mComm, mCommTask, mMainThread, mThreadSafetyLevel, and lama::Communicator::Multiple.

Referenced by all2all(), bcast(), gather(), gatherImpl(), maxlocImpl(), maxval(), minval(), scatterImpl(), send(), shiftImpl(), startrecv(), startsend(), sumImpl(), swapImpl(), and synchronize().

template<typename T >

void lama::MPICommunicator::send	(	const T *	buffer,
		int	count,
		int	target
	)		const `[inline, private]`

References defaultTag, LAMA_MPICALL, and selectMPIComm().

Referenced by exchangeByPlanImpl().

IndexType lama::MPICommunicator::shift	(	double	newVals[],
		const IndexType	newSize,
		const double	oldVals[],
		const IndexType	oldSize,
		const int	direction
	)		const `[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.

Implements lama::Communicator.

References lama::Communicator::getNeighbor(), getSize(), lama::Communicator::shift0(), and shiftImpl().

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

Implements lama::Communicator.

References lama::Communicator::getNeighbor(), getSize(), lama::Communicator::shift0(), and shiftImpl().

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

Implements lama::Communicator.

References lama::Communicator::getNeighbor(), getSize(), lama::Communicator::shift0(), and shiftImpl().

template<typename T >

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

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

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

template<typename T >

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

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

References LAMA_ASSERT_ERROR.

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

auto_ptr< SyncToken > lama::MPICommunicator::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 from lama::Communicator.

References lama::Communicator::defaultShiftAsync(), lama::Communicator::getNeighbor(), getSize(), and shiftAsyncImpl().

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

Reimplemented from lama::Communicator.

References lama::Communicator::defaultShiftAsync(), lama::Communicator::getNeighbor(), getSize(), and shiftAsyncImpl().

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

Reimplemented from lama::Communicator.

References lama::Communicator::defaultShiftAsync(), lama::Communicator::getNeighbor(), getSize(), and shiftAsyncImpl().

template<typename T >

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

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().

template<typename T >

auto_ptr< SyncToken > lama::MPICommunicator::shiftAsyncImpl	(	T	newvals[],
		const PartitionId	source,
		const T	oldVals[],
		const PartitionId	dest,
		const IndexType	size
	)		const `[private]`

References getRank(), LAMA_ASSERT_ERROR, startrecv(), and startsend().

Referenced by shiftAsync().

template<typename T >

IndexType lama::MPICommunicator::shiftImpl	(	T	newvals[],
		const IndexType	newSize,
		const PartitionId	source,
		const T	oldVals[],
		const IndexType	oldSize,
		const PartitionId	dest
	)		const `[private]`

References getRank(), LAMA_ASSERT_ERROR, LAMA_MPICALL, and selectMPIComm().

Referenced by shift().

template<typename T >

MPI_Request lama::MPICommunicator::startrecv	(	T *	buffer,
		int	count,
		int	source
	)		const

References defaultTag, LAMA_MPICALL, and selectMPIComm().

Referenced by exchangeByPlanAsyncImpl(), exchangeByPlanImpl(), and shiftAsyncImpl().

template<typename T >

MPI_Request lama::MPICommunicator::startsend	(	const T *	buffer,
		int	count,
		int	target
	)		const

References defaultTag, LAMA_MPICALL, and selectMPIComm().

Referenced by exchangeByPlanAsyncImpl(), and shiftAsyncImpl().

float lama::MPICommunicator::sum ( const float value ) const [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.

Implements lama::Communicator.

References sumImpl().

Referenced by sumImpl().

double lama::MPICommunicator::sum ( const double value ) const [virtual]

Implements lama::Communicator.

References sumImpl().

int lama::MPICommunicator::sum ( const int value ) const [virtual]

Implements lama::Communicator.

References sumImpl().

size_t lama::MPICommunicator::sum ( const size_t value ) const [virtual]

Implements lama::Communicator.

template<typename T >

T lama::MPICommunicator::sumImpl ( T myVal ) const [private]

References LAMA_MPICALL, selectMPIComm(), and sum().

Referenced by sum().

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

This routine swaps an array with a partner processor.

Implements lama::Communicator.

References swapImpl().

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

Implements lama::Communicator.

References swapImpl().

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

Implements lama::Communicator.

References swapImpl().

template<typename T >

void lama::MPICommunicator::swapImpl	(	T	val[],
		const IndexType	n,
		PartitionId	partner
	)		const `[private]`

References defaultTag, LAMA_ASSERT_ERROR, LAMA_MPICALL, mRank, and selectMPIComm().

Referenced by swap().

void lama::MPICommunicator::synchronize ( ) const [virtual]

Barrier synchronization between all processors.

Implements lama::Communicator.

References LAMA_MPICALL, and selectMPIComm().

template<typename T >

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

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(), lama::Communicator::exchangeByPlan(), lama::Communicator::gather(), lama::Halo::getProvidesIndexes(), lama::Halo::getProvidesPlan(), lama::Halo::getRequiredPlan(), lama::Communicator::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 `[inherited]`