[Mpi-forum] Fortran change for discussion
nmm1 at cam.ac.uk
Tue Sep 11 07:30:29 CDT 2012
On Sep 11 2012, Jeremiah Willcock wrote:
>On Tue, 11 Sep 2012, Jeff Squyres (jsquyres) wrote:
>> FWIW, we chose the copying language not because of the c language, but
>> rather because we have to explain that the MPI data type is applied, and
>> that seemed far easier to explain against a (virtual) contiguous buffer.
>Maybe the issue is that I'm not a Fortran programmer, but the copying
>language has always been confusing to me (its effect on what user-defined
>datatypes need to look like is another thing I asked about before). Am I
>to understand the interface as something like this?
> For purposes of interpreting derived datatypes and other information
> in communication operations, array slices are treated as if they are
> stored contiguously in memory. I.e., it is "as if" some kind of
> magic memory remapping occurs by the MPI implementation that maps the
> user's non-contiguous array into a contiguous one without copying the
> elements (in either direction).
No, that's not the reason, because your wording is considerably closer
to the Fortran model than the current wording. What you have said is
a fairly precise paraphrase of what sequence association is.
Realistically, I agree with your concerns - this is confusing. Your
addendum would help, as would adding the phrase "at an unspecified time
between the initial call and the wait". Personally, I would do both.
More information about the mpi-forum