(j3.2006) (SC22WG5.3764) Response on the TR29113 draft N1761

Craig Rasmussen crasmussen
Mon Dec 8 14:03:38 EST 2008

On Dec 8, 2008, at 11:40 AM, Reinhold Bader wrote:

> Bill Long schrieb:
>> Reinhold Bader wrote:
> Note that the
>> main purpose of adding this feature was for MPI-like interfaces,  
>> where
>> the only objective was to pass the object to a C function and give  
>> the
>> function tools to correctly interpret the object.
> Yes. Note that I'm not yet sure the MPI Forum is going to swallow  
> this since
> the MPI vendors apparently don't want to rewrite their whole  
> infrastructure to use
> descriptors ... even though it provides the technically better  
> solution.

Since the MPI vendors will have to write wrappers for the MPI  
functions, there is no real technical problem for them and I don't  
think they will complain too loudly (I will ask at the MPI Forum next  
week).  The main requirement is that existing _users code_ won't need  
to change.  The way the TR is currently written, I think the MPI 3.0  
Fortran interfaces won't use assumed shape dummy variables and so  
vendors won't need to access descriptors.


More information about the J3 mailing list