gs_mpi_rma.f90 File Reference

Go to the source code of this file.

Data Types
type	gs_mpi_rma::gs_mpi_rma_t
	Gather-scatter communication using MPI one-sided puts into a passive target window, with per-rank signalling for completion. More...

Modules
module	gs_mpi_rma
	Defines MPI one-sided (RMA) gather-scatter communication.

Functions/Subroutines
subroutine	gs_mpi_rma::gs_mpi_rma_init (this, send_pe, recv_pe)
	Initialise MPI RMA based communication method.
subroutine	gs_mpi_rma::gs_mpi_rma_free (this)
	Deallocate MPI RMA based communication method.
integer(kind=i8) function	gs_mpi_rma::gs_mpi_rma_load (s, slot)
	Reload a signal counter through a VOLATILE dummy, so the spin in gs_mpi_rma_wait_ge reads memory on every iteration rather than a value the compiler hoisted out of the loop.
subroutine	gs_mpi_rma::gs_mpi_rma_wait_ge (this, slot, val)
	Spin until the local signal counter in slot slot has reached val. Called by the master thread only.
subroutine	gs_mpi_rma::gs_mpi_rma_nbsend (this, u, n, tag, deps, strm)
	Pack the gathered shared dofs and put them into each neighbour's receive window, then announce the puts. See the type comment for why the puts, the flush and the signals are separate phases.
subroutine	gs_mpi_rma::gs_mpi_rma_nbrecv (this, tag)
	No-op: receives are completed by the remote put and its signal.
subroutine	gs_mpi_rma::gs_mpi_rma_nbwait (this, u, n, op, strm)
	Wait per neighbour for the signal that its data has landed, reduce the slab into u, and ack the sender so it may overwrite the slab next round.
subroutine	gs_mpi_rma::gs_mpi_rma_nbsend_vec (this, u, n, nc, tag, deps, strm)
	Fused nc-component send: pack nc contiguous component blocks per peer slab and put nc*ndofs reals. Buffer indexing and put sizes scale by nc; the signalling is unchanged.
subroutine	gs_mpi_rma::gs_mpi_rma_nbrecv_vec (this, tag, nc)
	No-op: receives are completed by the remote put and its signal.
subroutine	gs_mpi_rma::gs_mpi_rma_nbwait_vec (this, u, n, nc, op, strm)
	Fused nc-component wait/reduce: per peer, wait on the data signal and reduce nc component blocks into u, then ack the sender.

Variables
logical, parameter, public	gs_mpi_rma::gs_mpi_rma_avail = .true.
	MPI RMA needs nothing beyond MPI-3, so the backend is always built. Kept as a parameter for symmetry with the other one-sided backends, which the gs comm. autotuner queries before benchmarking them.
integer, parameter	gs_mpi_rma::gs_mpi_rma_poke = 64
	Number of spin iterations between the progress pokes in gs_mpi_rma_wait_ge.
logical, save	gs_mpi_rma::gs_mpi_rma_flush_all = .true.
	Whether to complete the payload puts with a single MPI_Win_flush_all rather than one MPI_Win_flush per peer. Batched is the default, set NEKO_GS_RMA_FLUSH_ALL=0 for the per-peer form.
logical, save	gs_mpi_rma::gs_mpi_rma_flush_bound = .false.
	Whether the flush strategy has been read from the environment. It is a program-wide binding, read once on the first init.