#include <deal.II/base/mpi_noncontiguous_partitioner.h>

Inheritance diagram for Utilities::MPI::NoncontiguousPartitioner:

Public Member Functions
	NoncontiguousPartitioner ()=default

	NoncontiguousPartitioner (const IndexSet &indexset_locally_owned, const IndexSet &indexset_ghost, const MPI_Comm &communicator)

	NoncontiguousPartitioner (const std::vector< types::global_dof_index > &indices_locally_owned, const std::vector< types::global_dof_index > &indices_ghost, const MPI_Comm &communicator)

template<typename Number >
void	export_to_ghosted_array (const ArrayView< const Number > &locally_owned_array, const ArrayView< Number > &ghost_array) const

template<typename Number >
void	export_to_ghosted_array (const unsigned int communication_channel, const ArrayView< const Number > &locally_owned_array, const ArrayView< Number > &temporary_storage, const ArrayView< Number > &ghost_array, std::vector< MPI_Request > &requests) const

template<typename Number >
void	export_to_ghosted_array_start (const unsigned int communication_channel, const ArrayView< const Number > &locally_owned_array, const ArrayView< Number > &temporary_storage, std::vector< MPI_Request > &requests) const

template<typename Number >
void	export_to_ghosted_array_finish (const ArrayView< const Number > &temporary_storage, const ArrayView< Number > &ghost_array, std::vector< MPI_Request > &requests) const

std::pair< unsigned int, unsigned int >	n_targets ()

types::global_dof_index	memory_consumption ()

const MPI_Comm &	get_mpi_communicator () const override

void	reinit (const IndexSet &indexset_locally_owned, const IndexSet &indexset_ghost, const MPI_Comm &communicator) override

void	reinit (const std::vector< types::global_dof_index > &indices_locally_owned, const std::vector< types::global_dof_index > &indices_ghost, const MPI_Comm &communicator)

Public Member Functions inherited from LinearAlgebra::CommunicationPatternBase
virtual	~CommunicationPatternBase ()=default

Private Attributes
MPI_Comm	communicator

std::vector< unsigned int >	send_ranks

std::vector< types::global_dof_index >	send_ptr

std::vector< types::global_dof_index >	send_indices

std::vector< unsigned int >	recv_ranks

std::vector< types::global_dof_index >	recv_ptr

std::vector< types::global_dof_index >	recv_indices

std::vector< uint8_t >	buffers

std::vector< MPI_Request >	requests

Detailed Description

A flexible Partitioner class, which does not impose restrictions regarding the order of the underlying index sets.

Author: Peter Munch, 2020

Definition at line 41 of file mpi_noncontiguous_partitioner.h.

Constructor & Destructor Documentation

◆ NoncontiguousPartitioner() [1/3]

Utilities::MPI::NoncontiguousPartitioner::NoncontiguousPartitioner ( )

default

Default constructor. Requires calling one of the reinit() functions to create a valid object.

◆ NoncontiguousPartitioner() [2/3]

Utilities::MPI::NoncontiguousPartitioner::NoncontiguousPartitioner	(	const IndexSet &	indexset_locally_owned,
		const IndexSet &	indexset_ghost,
		const MPI_Comm &	communicator
	)

Constructor. Set up point-to-point communication pattern based on the IndexSets arguments indexset_locally_owned and indexset_ghost for the MPI communicator communicator.

◆ NoncontiguousPartitioner() [3/3]

Utilities::MPI::NoncontiguousPartitioner::NoncontiguousPartitioner	(	const std::vector< types::global_dof_index > &	indices_locally_owned,
		const std::vector< types::global_dof_index > &	indices_ghost,
		const MPI_Comm &	communicator
	)

Constructor. Same as above but for vectors of indices indices_locally_owned and indices_ghost. This allows the indices to not be sorted and the values are read and written automatically at the right position of the vector during update_values(), update_values_start(), and update_values_finish(). It is allowed to include entries with the value numbers::invalid_dof_index which do not take part of the index exchange but are present in the data vectors as padding.

Member Function Documentation

◆ export_to_ghosted_array() [1/2]

template<typename Number >

void Utilities::MPI::NoncontiguousPartitioner::export_to_ghosted_array	(	const ArrayView< const Number > &	locally_owned_array,
		const ArrayView< Number > &	ghost_array
	)		const

Fill the vector ghost_array according to the precomputed communication pattern with values from locally_owned_array.

Precondition: The vectors only have to provide a method begin(), which allows to access their raw data.; The size of both vectors must be at least as large as the number of entries in the index sets passed to the constructors or the reinit() functions.

Note: This function calls the methods update_values_start() and update_values_finish() in sequence. Users can call these two functions separately and hereby overlap communication and computation.

◆ export_to_ghosted_array() [2/2]

template<typename Number >

void Utilities::MPI::NoncontiguousPartitioner::export_to_ghosted_array	(	const unsigned int	communication_channel,
		const ArrayView< const Number > &	locally_owned_array,
		const ArrayView< Number > &	temporary_storage,
		const ArrayView< Number > &	ghost_array,
		std::vector< MPI_Request > &	requests
	)		const

Same as above but with an interface similar to Utilities::MPI::Partitioner::export_to_ghosted_array_start and Utilities::MPI::Partitioner::export_to_ghosted_array_finish. In this function, the user can provide the temporary data structures to be used.

Precondition: The size of the temporary_storage vector has to be at least as large as the sum of the number of entries in the index sets passed to the constructor and the reinit() functions. The reason for this is that this vector is used as buffer for both sending and receiving data.

◆ export_to_ghosted_array_start()

template<typename Number >

void Utilities::MPI::NoncontiguousPartitioner::export_to_ghosted_array_start	(	const unsigned int	communication_channel,
		const ArrayView< const Number > &	locally_owned_array,
		const ArrayView< Number > &	temporary_storage,
		std::vector< MPI_Request > &	requests
	)		const

Start update: Data is packed, non-blocking send and receives are started.

Note: In contrast to the function Utilities::MPI::Partitioner::export_to_ghosted_array_start, the user does not pass a reference to the destination vector, since the data is received into a designated part of the buffer temporary_storage. This allows for padding and other post-processing of the received data.

Precondition: The required size of the vectors are the same as in the functions above.

◆ export_to_ghosted_array_finish()

template<typename Number >

void Utilities::MPI::NoncontiguousPartitioner::export_to_ghosted_array_finish	(	const ArrayView< const Number > &	temporary_storage,
		const ArrayView< Number > &	ghost_array,
		std::vector< MPI_Request > &	requests
	)		const

Finish update. The method waits until all data has been sent and received. Once data from any process is received it is processed and placed at the right position of the vector dst.

Note: In contrast to the function Utilities::MPI::Partitioner::export_to_ghosted_array_finish, the user also has to pass a reference to the buffer temporary_storage, since the data has been received into the buffer and not into the destination vector.

Precondition: The required size of the vectors are the same as in the functions above.

◆ n_targets()

std::pair<unsigned int, unsigned int> Utilities::MPI::NoncontiguousPartitioner::n_targets ( )

Returns the number of processes this process sends data to and the number of processes this process receives data from.

◆ memory_consumption()

types::global_dof_index Utilities::MPI::NoncontiguousPartitioner::memory_consumption ( )

Return memory consumption in Byte.

◆ get_mpi_communicator()

const MPI_Comm& Utilities::MPI::NoncontiguousPartitioner::get_mpi_communicator ( ) const

overridevirtual

Return the underlying communicator.

Implements LinearAlgebra::CommunicationPatternBase.

◆ reinit() [1/2]

void Utilities::MPI::NoncontiguousPartitioner::reinit	(	const IndexSet &	indexset_locally_owned,
		const IndexSet &	indexset_ghost,
		const MPI_Comm &	communicator
	)

overridevirtual

Initialize the inner data structures.

Implements LinearAlgebra::CommunicationPatternBase.

◆ reinit() [2/2]

void Utilities::MPI::NoncontiguousPartitioner::reinit	(	const std::vector< types::global_dof_index > &	indices_locally_owned,
		const std::vector< types::global_dof_index > &	indices_ghost,
		const MPI_Comm &	communicator
	)

Initialize the inner data structures.

Member Data Documentation

◆ communicator

MPI_Comm Utilities::MPI::NoncontiguousPartitioner::communicator

private

MPI communicator.

Definition at line 199 of file mpi_noncontiguous_partitioner.h.

◆ send_ranks

std::vector<unsigned int> Utilities::MPI::NoncontiguousPartitioner::send_ranks

private

The ranks this process sends data to.

Definition at line 204 of file mpi_noncontiguous_partitioner.h.

◆ send_ptr

std::vector<types::global_dof_index> Utilities::MPI::NoncontiguousPartitioner::send_ptr

private

Offset of each process within send_buffer.

Note: Together with send_indices this forms a CRS data structure.

Definition at line 211 of file mpi_noncontiguous_partitioner.h.

◆ send_indices

std::vector<types::global_dof_index> Utilities::MPI::NoncontiguousPartitioner::send_indices

private

Local index of each entry in send_buffer within the destination vector.

Note: Together with send_ptr this forms a CRS data structure.

Definition at line 219 of file mpi_noncontiguous_partitioner.h.

◆ recv_ranks

std::vector<unsigned int> Utilities::MPI::NoncontiguousPartitioner::recv_ranks

private

The ranks this process receives data from.

Definition at line 224 of file mpi_noncontiguous_partitioner.h.

◆ recv_ptr

std::vector<types::global_dof_index> Utilities::MPI::NoncontiguousPartitioner::recv_ptr

private

Offset of each process within recv_buffer.

Note: Together with recv_indices this forms a CRS data structure.

Definition at line 231 of file mpi_noncontiguous_partitioner.h.

◆ recv_indices

std::vector<types::global_dof_index> Utilities::MPI::NoncontiguousPartitioner::recv_indices

private

Local index of each entry in recv_buffer within the destination vector.

Note: Together with recv_ptr this forms a CRS data structure.

Definition at line 239 of file mpi_noncontiguous_partitioner.h.

◆ buffers

std::vector<uint8_t> Utilities::MPI::NoncontiguousPartitioner::buffers

mutableprivate

Buffer containing the values sorted by rank for sending and receiving.

Note: Only allocated if not provided externally by user.; At this place we do not know the type of the data to be sent. So we use an arbitrary type of size 1 byte. The type is cast to the requested type in the relevant functions.

Definition at line 250 of file mpi_noncontiguous_partitioner.h.

◆ requests

std::vector<MPI_Request> Utilities::MPI::NoncontiguousPartitioner::requests

mutableprivate

MPI requests for sending and receiving.

Note: Only allocated if not provided externally by user.

Definition at line 257 of file mpi_noncontiguous_partitioner.h.

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

include/deal.II/base/mpi_noncontiguous_partitioner.h

Public Member Functions

Private Attributes

Detailed Description

Constructor & Destructor Documentation

◆ NoncontiguousPartitioner() [1/3]

◆ NoncontiguousPartitioner() [2/3]

◆ NoncontiguousPartitioner() [3/3]

Member Function Documentation

◆ export_to_ghosted_array() [1/2]

◆ export_to_ghosted_array() [2/2]

◆ export_to_ghosted_array_start()

◆ export_to_ghosted_array_finish()

◆ n_targets()

◆ memory_consumption()

◆ get_mpi_communicator()

◆ reinit() [1/2]

◆ reinit() [2/2]

Member Data Documentation

◆ communicator

◆ send_ranks

◆ send_ptr

◆ send_indices

◆ recv_ranks

◆ recv_ptr

◆ recv_indices

◆ buffers

◆ requests