gko::experimental::mpi::request#

Move-only handle representing an outstanding nonblocking MPI operation. Returned by the asynchronous send/receive helpers on communicator and by the nonblocking variants on the collective communicators.

auto exec = gko::ReferenceExecutor::create();
std::vector<double> send_buf(100, 1.0), recv_buf(100);
int next = (comm.rank() + 1) % comm.size();
int prev = (comm.rank() + comm.size() - 1) % comm.size();
auto send_req = comm.i_send(exec, send_buf.data(), 100, next, 0);
auto recv_req = comm.i_recv(exec, recv_buf.data(), 100, prev, 0);
// ... do other work ...
auto stat = recv_req.wait();
send_req.wait();

wait() blocks until the operation completes and returns the matching status. The request is consumed by wait().

class request#

The request class is a light, move-only wrapper around the MPI_Request handle.

Public Functions

inline request()#

The default constructor. It creates a null MPI_Request of MPI_REQUEST_NULL type.

inline MPI_Request *get()#

Get a pointer to the underlying MPI_Request handle.

Returns:

a pointer to MPI_Request handle

inline status wait()#

Allows a rank to wait on a particular request handle.

Parameters:

  • req – The request to wait on.

  • status – The status variable that can be queried.