[MPI3 Fortran] ASYNC attribute

[Bill Long]

Jeff Squyres wrote:
> On May 6, 2009, at 9:21 AM, Bill Long wrote:
>
>> > 2. Yes, the MPI_Request argument to MPI_Wait is inout.
>>
>> Question: Is MPI_Request actually a Fortran derived TYPE, or is it just
>> a KIND value for an INTEGER?  I had thought that all of these flag and
>> handle-like arguments were treated as integers in the Fortran interface.
>>   Making it an actual defined type would have some advantages, but would
>> that cause a problem with existing codes?
>>
>
>
> In MPI-1 and MPI-2, they're INTEGERs.  But we're toying with the idea 
> of using ISO_C_BIND for MPI-3 and therefore tying the MPI_Request in 
> Fortran to be whatever it is in C (likely either the equivalent of a C 
> int or a pointer).
>

OK. Ignoring the compatibility problem mentioned above, the advantage of 
a structure is the ability to add extra components.  If your interface 
uses dimension(..) for the buf argument, then it is passed as a (pointer 
to) a descriptor.  Copy-in/copy-out would never occur.  The Isend/Irecv 
routines could be written as wrappers, and use two extra pointer 
elements in the request structure to hold the addresses of the original 
descriptor and the local temp.  This would allow a safe copy-in into the 
low level C transfer routine that requires contiguous arguments.  The 
copy-out step would occur inside the MPI_Wait wrapper using the 
information delivered in the request struct.  I can provide a detailed 
scheme if you like. Such wrappers would solve the copy-in/copy-out 
problem by allowing it to be done safely, rather than prohibiting it.

Cheers,
Bill




-- 
Bill Long                                   longb at cray.com
Fortran Technical Support    &              voice: 651-605-9024
Bioinformatics Software Development         fax:   651-605-9142
Cray Inc., 1340 Mendota Heights Rd., Mendota Heights, MN, 55120