MPI one-sided communication and shared memory

Detailed Description

Abstraction over the MPI_Win object that facilitates remote memory access and node-local shared memory.

Two move-only class templates are provided:

• mpi::window wraps MPI_Win for generic one-sided communication. It either takes ownership of a user-provided buffer or allocates one internally via MPI_Win_allocate. Data movement is done with mpi::window::get and mpi::window::put; synchronization is done with mpi::window::fence, mpi::window::flush, mpi::window::sync, mpi::window::lock / mpi::window::unlock (passive target) or mpi::window::post / mpi::window::start / mpi::window::complete / mpi::window::wait (general active target).
• mpi::shared_window is a MPI_Win_allocate_shared-backed specialization built on top of mpi::shared_communicator. It is intended for deduplicating large read-mostly data across MPI ranks that share an SMP node. The per-rank shared base pointer, byte size and displacement unit can be retrieved with mpi::shared_window::query (or the convenience accessors mpi::shared_window::base, mpi::shared_window::size and mpi::shared_window::disp_unit).

Classes
class	mpi::shared_communicator
	C++ wrapper around MPI_Comm that is a result of the mpi::communicator::split_shared operation. More...
class	mpi::shared_window< BaseType >
	A C++ wrapper around MPI_Win representing a shared memory window. More...
class	mpi::window< BaseType >
	A C++ wrapper around MPI_Win providing convenient memory window management. More...