Wrapper class around an MPI_Comm object. More...

#include <MPIWrapper.hpp>

Public Member Functions
	MPIComm ()

	MPIComm (MPI_Comm c)

	MPIComm (const MPIComm &c)

	MPIComm (MPIComm &&c) noexcept

virtual	~MPIComm ()

MPIComm &	operator= (const MPIComm &c)

MPIComm &	operator= (MPIComm &&c) noexcept

MPI_Comm	comm () const

bool	is_null () const

int	rank () const

int	size () const

bool	is_root () const

void	barrier () const

template<typename T >
void	broadcast (std::vector< T > &sbuf) const

template<typename T >
void	broadcast_from (std::vector< T > &sbuf, int src) const

template<typename T , std::size_t N>
void	broadcast (std::array< T, N > &sbuf) const

template<typename T >
void	broadcast (T &data) const

template<typename T >
void	broadcast_from (T &data, int src) const

template<typename T >
void	broadcast (T *sbuf, std::size_t ssize) const

template<typename T >
void	broadcast_from (T *sbuf, std::size_t ssize, int src) const

template<typename T >
void	all_gather (T *buf, std::size_t rsize) const

void	all_gather (std::pair< long int, long int > *buf, std::size_t rsize) const

void	all_gather (std::pair< long long int, long long int > *buf, std::size_t rsize) const

template<typename T >
void	all_gather_v (T buf, const int rcnts, const int *displs) const

template<typename T >
void	gather (T sbuf, int ssize, int rbuf, int rsize, int root) const

template<typename T >
void	gather_v (T sbuf, int scnts, T rbuf, const int rcnts, const int displs, int root) const

template<typename T >
MPIRequest	isend (const std::vector< T > &sbuf, int dest, int tag) const

template<typename T >
void	isend (const std::vector< T > &sbuf, int dest, int tag, MPI_Request *req) const

template<typename T >
void	isend (const T sbuf, std::size_t ssize, int dest, int tag, MPI_Request req) const

template<typename T >
void	send (const T *sbuf, std::size_t ssize, int dest, int tag) const

template<typename T >
void	isend (const T &buf, int dest, int tag, MPI_Request *req) const

template<typename T >
void	send (const std::vector< T > &sbuf, int dest, int tag) const

template<typename T >
std::vector< T >	recv (int src, int tag) const

template<typename T >
std::pair< int, std::vector< T > >	recv_any_src (int tag) const

template<typename T >
T	recv_one (int src, int tag) const

template<typename T >
void	irecv (const T rbuf, std::size_t rsize, int src, int tag, MPI_Request req) const

template<typename T >
void	recv (const T *rbuf, std::size_t rsize, int src, int tag) const

template<typename T >
T	all_reduce (T t, MPI_Op op) const

template<typename T >
T	reduce (T t, MPI_Op op) const

template<typename T >
void	all_reduce (T *t, int ssize, MPI_Op op) const

template<typename T >
void	all_reduce (std::vector< T > &t, MPI_Op op) const

template<typename T >
void	reduce (T *t, int ssize, MPI_Op op, int dest=0) const

template<typename T >
void	all_to_all (const T sbuf, int scnt, T rbuf) const

template<typename T , typename A = std::allocator<T>>
std::vector< T, A >	all_to_allv (const T sbuf, int scnts, int sdispls, int rcnts, int *rdispls) const

template<typename T >
void	all_to_allv (const T sbuf, int scnts, int sdispls, T rbuf, int rcnts, int rdispls) const

template<typename T , typename A = std::allocator<T>>
void	all_to_all_v (std::vector< std::vector< T > > &sbuf, std::vector< T, A > &rbuf, std::vector< T * > &pbuf) const

template<typename T , typename A = std::allocator<T>>
std::vector< T, A >	all_to_all_v (std::vector< std::vector< T > > &sbuf) const

template<typename T , typename A = std::allocator<T>>
void	all_to_all_v (std::vector< std::vector< T > > &sbuf, std::vector< T, A > &rbuf, std::vector< T * > &pbuf, const MPI_Datatype Ttype) const

MPIComm	sub (int P0, int P, int stride=1) const

MPIComm	sub_self (int p) const

Static Public Member Functions
static void	control_start (const std::string &name)

static void	control_stop (const std::string &name)

static bool	initialized ()

Detailed Description

Wrapper class around an MPI_Comm object.

This is a simple wrapper around a MPI_Comm object. The main reason for this class is to simplify resource management of the MPI_Comm object. An object of class MPIComm owns the MPI_Comm that it stores, and is responsible for freeing it (in the destructor).

A number of simple wrappers around basic MPI calls are provided.

Constructor & Destructor Documentation

◆ MPIComm() [1/4]

strumpack::MPIComm::MPIComm ( )

inline

Default constructor. This will initialize the encapsulated MPI_Comm to MPI_COMM_WORLD.

◆ MPIComm() [2/4]

strumpack::MPIComm::MPIComm ( MPI_Comm c )

inline

Constructor using a MPI_Comm. This will DUPLICATE the input communicator c!

Parameters

c	the input MPI communicator, this will be duplicated internally and can thus be freed immediately

◆ MPIComm() [3/4]

strumpack::MPIComm::MPIComm ( const MPIComm & c )

inline

Copy constructor. Will DUPLICATE the underlying MPI_Comm object!

Parameters

c	will be copied, not changed

◆ MPIComm() [4/4]

strumpack::MPIComm::MPIComm ( MPIComm && c )

inlinenoexcept

Move constructor.

Parameters

c	will be moved from, will be reset to MPI_COMM_NULL

See also: operator=(MPIComm&& c)

◆ ~MPIComm()

virtual strumpack::MPIComm::~MPIComm ( )

inlinevirtual

Virtual destructor. Free the MPI_Comm object, unless it is MPI_COMM_NULL or MPI_COMM_WORLD.

Member Function Documentation

◆ all_reduce() [1/2]

template<typename T >

void strumpack::MPIComm::all_reduce	(	T *	t,
		int	ssize,
		MPI_Op	op
	)		const

inline

Compute the reduction of op(t[]_i) over all processes i, t[] is an array, and where op can be any MPI_Op, on all processes. This routine is collective on all ranks in this communicator. See documentation for MPI_Allreduce. Every element of the array will be reduced over the different processes. The operation is performed in-place.

Template Parameters

T	type of variables to reduce, should have a corresponding mpi_type<T>() implementation

Parameters

t	pointer to array of variables to reduce
ssize	size of array to reduce
op	reduction operator

◆ all_reduce() [2/2]

template<typename T >

T strumpack::MPIComm::all_reduce	(	T	t,
		MPI_Op	op
	)		const

inline

Compute the reduction of op(t_i) over all processes i, where op can be any MPI_Op, on all processes. This routine is collective on all ranks in this communicator. See documentation for MPI_Allreduce.

Template Parameters

T	type of variable to reduce, should have a corresponding mpi_type<T>() implementation

Parameters

t	variable to reduce, passed by value
op	reduction operator

Returns: variable of type T, result of the reduction, available on each rank

◆ all_to_all_v() [1/3]

template<typename T , typename A = std::allocator<T>>

std::vector< T, A > strumpack::MPIComm::all_to_all_v ( std::vector< std::vector< T > > & sbuf ) const

inline

Perform an MPI_Alltoallv. Each rank sends sbuf[i] to process i. The results are received in a single contiguous vector which is returned. This is collective on this MPI communicator.

Template Parameters

T	type of data to send, this should have a corresponding mpi_type<T>() implementation or should define T::mpi_type()
A	allocator to be used for the receive buffer

Parameters

sbuf	send buffers (should be size this->size())

Returns: receive buffer

See also: all_to_all_v

◆ all_to_all_v() [2/3]

template<typename T , typename A = std::allocator<T>>

void strumpack::MPIComm::all_to_all_v	(	std::vector< std::vector< T > > &	sbuf,
		std::vector< T, A > &	rbuf,
		std::vector< T * > &	pbuf
	)		const

inline

Perform an MPI_Alltoallv. Each rank sends sbuf[i] to process i. The results are received in a single contiguous vector rbuf. pbuf has pointers into rbuf, with pbuf[i] pointing to the data received from rank i. This is collective on this MPI communicator.

Template Parameters

T	type of data to send, this should have a corresponding mpi_type<T>() implementation or should define T::mpi_type()
A	allocator to be used for the rbuf

Parameters

sbuf	send buffers (should be size this->size())
rbuf	receive buffer, can be empty, will be allocated
pbuf	pointers (to positions in rbuf) to where data received from different ranks start

See also: all_to_all_v

◆ all_to_all_v() [3/3]

template<typename T , typename A = std::allocator<T>>

void strumpack::MPIComm::all_to_all_v	(	std::vector< std::vector< T > > &	sbuf,
		std::vector< T, A > &	rbuf,
		std::vector< T * > &	pbuf,
		const MPI_Datatype	Ttype
	)		const

inline

Perform an MPI_Alltoallv. Each rank sends sbuf[i] to process i. The results are received in a single contiguous vector rbuf. pbuf has pointers into rbuf, with pbuf[i] pointing to the data received from rank i. This is collective on this MPI communicator.

Template Parameters

T	type of data to send
A	allocator to be used for the receive buffer

Parameters

sbuf	send buffers (should be size this->size())
rbuf	receive buffer, can be empty, will be allocated
pbuf	pointers (to positions in rbuf) to where data received from different ranks start
Ttype	MPI_Datatype corresponding to the template parameter T

See also: all_to_all_v

◆ barrier()

void strumpack::MPIComm::barrier ( ) const

inline

Perform a barrier operation. This operation is collective on all the ranks in this communicator.

◆ comm()

MPI_Comm strumpack::MPIComm::comm ( ) const

inline

Returns the underlying MPI_Comm object.

◆ control_start()

static void strumpack::MPIComm::control_start ( const std::string & name )

inlinestatic

Call MPI_Pcontrol with level 1, and string name

◆ control_stop()

static void strumpack::MPIComm::control_stop ( const std::string & name )

inlinestatic

Call MPI_Pcontrol with level -1, and string name

◆ is_null()

bool strumpack::MPIComm::is_null ( ) const

inline

Checks whether the underlying MPI_Comm object is MPI_COMM_NULL.

◆ is_root()

bool strumpack::MPIComm::is_root ( ) const

inline

Check whether the current process is the root of this MPI communicator.

◆ isend() [1/2]

template<typename T >

MPIRequest strumpack::MPIComm::isend	(	const std::vector< T > &	sbuf,
		int	dest,
		int	tag
	)		const

inline

Non-blocking send of a vector to a destination process, with a certain tag.

Template Parameters

T	template type of the send buffer, should have a corresponding mpi_type<T>() implementation

Parameters

sbuf	buffer of type T to be send
dest	rank of destination process in this MPI communicator
tag	tag to use in MPI message

Returns: request object, use this to wait for completion of the non-blocking send

◆ isend() [2/2]

template<typename T >

void strumpack::MPIComm::isend	(	const std::vector< T > &	sbuf,
		int	dest,
		int	tag,
		MPI_Request *	req
	)		const

inline

Non-blocking send of a vector to a destination process, with a certain tag.

Template Parameters

T	template type of the send buffer, should have a corresponding mpi_type<T>() implementation

Parameters

sbuf	buffer of type T to be send
dest	rank of destination process in this MPI communicator
tag	tag to use in MPI message
req	MPI request object

◆ operator=() [1/2]

MPIComm & strumpack::MPIComm::operator= ( const MPIComm & c )

inline

Assignement operator. This will DUPLICATE the MPI_Comm object.

Parameters

c	object to copy from, will not be modified.

◆ operator=() [2/2]

MPIComm & strumpack::MPIComm::operator= ( MPIComm && c )

inlinenoexcept

Move operator.

Parameters

c	the object ro be moved from, will be reset to MPI_COMM_NULL

◆ rank()

int strumpack::MPIComm::rank ( ) const

inline

Return the current rank in this communicator.

◆ recv()

template<typename T >

std::vector< T > strumpack::MPIComm::recv	(	int	src,
		int	tag
	)		const

inline

Receive a vector of T's from process src, with tag. The message size does not need to be known in advance.

Template Parameters

T	template parameter of vector to receive, should have a corresponding mpi_type<T>() implementation

Parameters

src	process to receive from
tag	tag to match the message

Returns: std::vector<T> with the data to be received.

See also: isend, send

◆ reduce() [1/2]

template<typename T >

void strumpack::MPIComm::reduce	(	T *	t,
		int	ssize,
		MPI_Op	op,
		int	dest = `0`
	)		const

inline

Compute the reduction of op(t[]_i) over all processes i, t[] is an array, and where op can be any MPI_Op, on the root process. This routine is collective on all ranks in this communicator. See documentation for MPI_Reduce. Every element of the array will be reduced over the different processes. The operation is performed in-place.

Template Parameters

T	type of variables to reduce, should have a corresponding mpi_type<T>() implementation

Parameters

t	pointer to array of variables to reduce
ssize	size of array to reduce
op	reduction operator
dest	where to reduce to

◆ reduce() [2/2]

template<typename T >

T strumpack::MPIComm::reduce	(	T	t,
		MPI_Op	op
	)		const

inline

Compute the reduction of op(t_i) over all processes i, where op can be any MPI_Op, on the root. This routine is collective on all ranks in this communicator. See documentation for MPI_Reduce (with dest == 0).

Template Parameters

T	type of variable to reduce, should have a corresponding mpi_type<T>() implementation

Parameters

t	variable to reduce, passed by value
op	reduction operator

Returns: variable of type T, result of the reduction, available only on the root process.

◆ send()

template<typename T >

void strumpack::MPIComm::send	(	const std::vector< T > &	sbuf,
		int	dest,
		int	tag
	)		const

inline

Blocking send of a vector to a destination process, with a certain tag.

Template Parameters

T	template type of the send buffer, should have a corresponding mpi_type<T>() implementation

Parameters

sbuf	buffer of type T to be send
dest	rank of destination process in this MPI communicator
tag	tag to use in MPI message

See also: recv, isend

◆ size()

int strumpack::MPIComm::size ( ) const

inline

Return the size of, ie, number of processes in, this communicator. This communicator should not be MPI_COMM_NULL.

◆ sub()

MPIComm strumpack::MPIComm::sub	(	int	P0,
		int	P,
		int	stride = `1`
	)		const

inline

Return a subcommunicator with P ranks, starting from rank P0, using stride stride. Ie., ranks (relative to this communicator) [P0:stride:P0+stride*P) will be included in the new communicator. This operation is collective on all the processes in this communicator.

Parameters

P0	first rank in the new communicator
P	number of ranks in the new communicator
stride	stride between ranks in this communicator determining which ranks will go into new communicator

Returns: new communicator containing P ranks from this communicator [P0:stride:P0+stride*P)

See also: sub_self

◆ sub_self()

MPIComm strumpack::MPIComm::sub_self ( int p ) const

inline

Returns a communicator with only rank p, or an MPIComm wrapping MPI_COMM_NULL if my rank in the current MPIComm != p. This is collective on the current communicator.

Parameters

p	rank to be included in new MPIComm

Returns: a new communicator containing only rank p (from the original MPIComm)

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

/home/pieterg/LBL/STRUMPACK/src/misc/MPIWrapper.hpp

Public Member Functions

Static Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ MPIComm() [1/4]

◆ MPIComm() [2/4]

◆ MPIComm() [3/4]

◆ MPIComm() [4/4]

◆ ~MPIComm()

Member Function Documentation

◆ all_reduce() [1/2]

◆ all_reduce() [2/2]

◆ all_to_all_v() [1/3]

◆ all_to_all_v() [2/3]

◆ all_to_all_v() [3/3]

◆ barrier()

◆ comm()

◆ control_start()

◆ control_stop()

◆ is_null()

◆ is_root()

◆ isend() [1/2]

◆ isend() [2/2]

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ rank()

◆ recv()

◆ reduce() [1/2]

◆ reduce() [2/2]

◆ send()

◆ size()

◆ sub()

◆ sub_self()