(j3.2006) (SC22WG5.3873) [ukfortran] MPI non-blocking transfers
Wed Jan 21 16:52:09 EST 2009
On Wed, 2009-01-21 at 13:23 -0800, N.M. Maclaren wrote:
> >We needed one for I/O, but we ought not to need one for interaction
> >procedures having defective interfaces.
> I am not sure what you mean by this. There is nothing conceptually
> defective about MPI's non-blocking interfaces, and it has the decency
> state most clearly that they require semantics that are outside the
> standard. The issue is about how to marry them.
What I meant is that MPI procedures either drag around the address of
the buffer in an opaque data structure, or keep hold of it, so you don't
have to mention it in all the calls where it might be useful to the
compiler to know you're affecting it, such as the wait call. This
"feature" might be helpful when used from a language where the pointer
semantics paralyze the optimizer, but it is unhelpful in Fortran.
More information about the J3