gko::experimental::mpi Namespace Reference

The mpi namespace, contains wrapper for many MPI functions. More...

Classes
class	CollectiveCommunicator
class	communicator
class	contiguous_type
class	DenseCommunicator
class	environment
class	NeighborhoodCommunicator
class	request
struct	status
struct	type_impl
struct	type_impl< char >
struct	type_impl< double >
struct	type_impl< float >
struct	type_impl< half >
struct	type_impl< int >
struct	type_impl< long >
struct	type_impl< long double >
struct	type_impl< long long >
struct	type_impl< std::complex< double > >
struct	type_impl< std::complex< float > >
struct	type_impl< std::complex< half > >
struct	type_impl< unsigned >
struct	type_impl< unsigned char >
struct	type_impl< unsigned long >
struct	type_impl< unsigned long long >
struct	type_impl< unsigned short >
class	window

Enumerations
enum class	thread_type { serialized = MPI_THREAD_SERIALIZED , funneled = MPI_THREAD_FUNNELED , single = MPI_THREAD_SINGLE , multiple = MPI_THREAD_MULTIPLE }

Functions
constexpr bool	is_gpu_aware ()
int	map_rank_to_device_id (MPI_Comm comm, int num_devices)
std::vector< status >	wait_all (std::vector< request > &req)
bool	requires_host_buffer (const std::shared_ptr< const Executor > &exec, const communicator &comm)
double	get_walltime ()

Detailed Description

The mpi namespace, contains wrapper for many MPI functions.

Enumeration Type Documentation

thread_type

enum class gko::experimental::mpi::thread_type

strong

This enum specifies the threading type to be used when creating an MPI environment.

Function Documentation

get_walltime()

double gko::experimental::mpi::get_walltime ( )

inline

Get the rank in the communicator of the calling process.

Parameters

comm	the communicator

is_gpu_aware()

constexpr bool gko::experimental::mpi::is_gpu_aware ( )

inlineconstexpr

Return if GPU aware functionality is available

map_rank_to_device_id()

int gko::experimental::mpi::map_rank_to_device_id	(	MPI_Comm	comm,
		int	num_devices
	)

Maps each MPI rank to a single device id in a round robin manner.

Parameters

comm	used to determine the node-local rank, if no suitable environment variable is available.
num_devices	the number of devices per node.

Returns

device id that this rank should use.

requires_host_buffer()

bool gko::experimental::mpi::requires_host_buffer	(	const std::shared_ptr< const Executor > &	exec,
		const communicator &	comm
	)

Checks if the combination of Executor and communicator requires passing MPI buffers from the host memory.

wait_all()

std::vector< status > gko::experimental::mpi::wait_all ( std::vector< request > &  req )

inline

Allows a rank to wait on multiple request handles.

Parameters

req	The vector of request handles to be waited on.

Returns

status The vector of status objects that can be queried.