RE: [mpi-21] F90 bindings

["Cownie, James H" <james.h.cownie@xxxxxxxxx>]

Aside from the number of function prototypes (a syntactic effect), there
is also the issue that the non-blocking MPI calls hang onto pointers to
their arguments and use them after the call has returned (a semantic
problem). That can certainly confuse Fortran compilers (which may create
temporaries to hold subroutine arguments).

To handle this you need something closer to the asynchronous I/O
additions to Fortran.

-- Jim

James Cownie <james.h.cownie@xxxxxxxxx>
SSG/DPD/PAT
Tel: +44 117 9071438




> -----Original Message-----
> From: owner-mpi-21@xxxxxxxxxxxxx [mailto:owner-mpi-21@xxxxxxxxxxxxx]
On
> Behalf Of Jeff Squyres
> Sent: 07 December 2007 19:57
> To: mpi-21@xxxxxxxxxxxxx
> Subject: Re: [mpi-21] F90 bindings
> 
> On Dec 7, 2007, at 2:32 PM, Steven Ericsson-Zenith wrote:
> 
> > I fear the by virtue of the beast Fortran will have to live with the
> > large number of function profiles. 7 million, however, is
> > surprisingly large. I assume that this is for all the permutations
> > of type.
> 
> Yes -- mainly because of the collectives that have 2 choice buffers.
> The math goes something like this:
> 
> - 15 fortran intrinsic types
> - max of 7 array dimensions
> - 50 MPI functions with 1 choice buffer
> - 25 MPI functions with 2 choice buffers
> 
> Total = (one choice buffer combos) + (two choice buffer combos)
>        = (15 * 7 * 50) + (15 * 7 * 25)^2
>        = 5,250 + 6,890,625
>        = 6,895,875
> 
> Yoinks.  :-(
> 
> > If that is the case, then it simply is what it is.
> 
> Unfortunately, it's not tenable.  There's no F90 compiler that can
> handle ~7M functions in a single module (not even close).  Even if
> they did, I can't imagine trying to compile a large F90 app that has
> "use mpi" all over the place -- the compiler would continually have to
> load ~7M signatures from the mpi module file, search to find the
> desired signature, etc.
> 
> In short: 7M signatures is just [way] too many.
> 
> Sun's F90 compiler has a "ignore the type" attribute similar to what
> was proposed for NEC.  Others may, too (I'm not a Fortran expert).
> But it doesn't seem good form that the MPI Standard *requires* a
> vendor-specific addition to the F90 compiler in order to work...
> 
> --
> Jeff Squyres
> Cisco Systems
---------------------------------------------------------------------
Intel Corporation (UK) Limited
Registered No. 1134945 (England)
Registered Office: Pipers Way, Swindon SN3 1RJ
VAT No: 860 2173 47

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.