[MPI3 Fortran] MPI 3 text is ready
rabenseifner at hlrs.de
Wed Sep 22 16:58:15 CDT 2010
Nick and Craig,
I attached the result-slides of our discussion at the MPI Forum meeting in
The tickets https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/229
through 253 reflect these decisions only partially.
I have to correct these tickets. This will need more time.
The major ticket
is already mainly correct.
As far as we know, Fortran will directly pass the reference of a
standardised "dope-vector" that describes the buffer.
The currently in this ticket mentioned CFI_update_cdesc() is not needed.
The "dope-vector" has enough information that the MPI wrappers can make
a contiguous copy of the possibly non-contiguous buffer (if the actual argument
is a multi-dimensional triplet-subsript).
This buffer is produced and under control of MPI. Call by reference is performed
by the compiler.
In the case of a vector-subscript (in some dimension, i.e. Fortran rank),
the Fortran compiler will generate code that copies the
non-contiguous buffer into a contiguous scratch
buffer and the dope-vector describes this contiguous scratch buffer.
This buffer is under control of the Fortran runtime.
Call by in-copy and out-copy is performed.
Nick and Craig, are these asumptions fully correct?
----- Original Message -----
> From: "N.M. Maclaren" <nmm1 at cam.ac.uk>
> To: "MPI-3 Fortran working group" <mpi3-fortran at lists.mpi-forum.org>
> Sent: Wednesday, September 22, 2010 9:10:05 PM
> Subject: Re: [MPI3 Fortran] MPI 3 text is ready
> On Sep 13 2010, Jeff Squyres (jsquyres) wrote:
> > Rolf and Craig and I want to present the entirety of the MPI 3
> > fortran
> > stuff to the WG this week, in preparation for presenting it to the
> > forum
> > in the oct meeting. If all goes well, we could have a formal first
> > reading in dec.
> > Will you all be able to attend the fortran WG meeting this week and
> > give
> > us first-level feedback?
> No. Where is it? I shall probably not get time to analyse it in
> but I am horribly afraid that there are going to be serious problems
> the new Fortran bindings. There are a lot of incompatible assumptions
> doing the rounds in this area. Both MPI and the Fortran Interop need
> much analysis as possible.
> mpi3-fortran mailing list
> mpi3-fortran at lists.mpi-forum.org
Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530
University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner
Nobelstr. 19, D-70550 Stuttgart, Germany . (Office: Allmandring 30)
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 672256 bytes
Desc: not available
More information about the mpiwg-fortran