comm Module Reference

Data Types
interface	neko_comm_wrapper_init

Functions/Subroutines
pure logical function, public	is_mpmd ()
	Whether this is one of several simulations coupled in an MPMD run, that is, whether Neko's communicator is a subset of the global one.
subroutine, public	comm_init
subroutine	shmem_check_thread_level (provided)
	Decide whether the thread level an OpenSHMEM library handed back is usable, given how Neko drives it.
subroutine, public	comm_free

Variables
type(mpi_comm), public	neko_comm
	MPI communicator.
type(mpi_comm), public	neko_global_comm
type(mpi_datatype), public	mpi_real_precision
	MPI type for working precision of REAL types.
type(mpi_datatype), public	mpi_extra_precision
integer, public	pe_rank
	MPI rank.
integer, public	pe_size = 0
	MPI size of communicator.
logical, public	nio
	I/O node.
integer, public	global_pe_rank
	Global MPI rank.
integer, public	global_pe_size = 0
	Global MPI size of communicator.
integer, public	neko_mpi_thread_provided
	Thread support provided by the MPI library.

Function/Subroutine Documentation

comm_free()

subroutine, public comm::comm_free

Definition at line 276 of file comm.F90.

Here is the caller graph for this function:

comm_init()

subroutine, public comm::comm_init

Definition at line 89 of file comm.F90.

Here is the call graph for this function:

Here is the caller graph for this function:

is_mpmd()

pure logical function, public comm::is_mpmd

Definition at line 82 of file comm.F90.

Here is the caller graph for this function:

shmem_check_thread_level()

subroutine comm::shmem_check_thread_level ( integer, intent(in)  provided )

private

Only the device backends need SHMEM_THREAD_MULTIPLE, where several host threads drive their own streams. On the host, gs_shmem issues every SHMEM call from the OpenMP master thread – which is the initial thread, the one that called shmem_init_thread – and only lets the team drive the puts when the library reports MULTIPLE, so no level is actually unusable. That matters in practice: Cray LIBSMA reports SHMEM_THREAD_SINGLE, and refusing anything below FUNNELED would rule out hybrid MPI+OpenMP runs there altogether.

Parameters

provided

thread level reported by shmem_init_thread

Definition at line 265 of file comm.F90.

Here is the caller graph for this function:

Variable Documentation

global_pe_rank

integer, public comm::global_pe_rank

Definition at line 68 of file comm.F90.

global_pe_size

integer, public comm::global_pe_size = 0

Definition at line 71 of file comm.F90.

mpi_extra_precision

type(mpi_datatype), public comm::mpi_extra_precision

Definition at line 55 of file comm.F90.

mpi_real_precision

type(mpi_datatype), public comm::mpi_real_precision

Definition at line 54 of file comm.F90.

neko_comm

type(mpi_comm), public comm::neko_comm

Definition at line 46 of file comm.F90.

neko_global_comm

type(mpi_comm), public comm::neko_global_comm

Definition at line 47 of file comm.F90.

neko_mpi_thread_provided

integer, public comm::neko_mpi_thread_provided

Definition at line 74 of file comm.F90.

nio

logical, public comm::nio

Definition at line 65 of file comm.F90.

pe_rank

integer, public comm::pe_rank

Definition at line 59 of file comm.F90.

pe_size

integer, public comm::pe_size = 0

Definition at line 62 of file comm.F90.