[MPI3 Fortran] MPI buffer arguments - void and call-by-reference

[Craig Rasmussen]

On Feb 19, 2009, at 9:00 AM, Bill Long wrote:

> Aleksandar Donev wrote:
>> Hi Craig,
>>
>> Since I wasn't in Vegas I will read the papers and try to  
>> reconstruct what you guys talked about, and in particular, to  
>> understand why "TYPE(*), DIMENSION(*)" was not a possibility  
>> (excluding scalar actuals if necessary---handle that by a separate  
>> procedure if need be).
>>
>>> interface
>>>    subroutine MPI_Send(buf, ...) BIND(C)
>>>        TYPE(*), DIMENSION(..), INTENT(in) :: buf
>>>        ...
>>>    end subroutine
>>> end interface
>>> I think using TYPE(*) and DIMENSION(..) is an excellent solution  
>>> for MPI 3.0.  The advantages are:
>> I am in agreement that this is a good solution.
>> BUT:
>>> 1. Current user code won't have to change beyond using the MPI3  
>>> module.
>> That makes no sense. If the array is passed by descriptor, there is  
>> absolutely no need to also pass its size as an argument. Even the  
>> type, at least for the set of predefined intrinsic interoperable  
>> types, does not need to be passed.
>>
>> Either we keep the interfaces as close to what they are (pass-by- 
>> address), or, we refine them. Otherwise you get something that  
>> neither satisfies existing codes/implementations nor does it  
>> satisfy demanding users (read: me :-)
>
> I would note (for the MPI folks) that discussions at J3 meetings  
> suggest that Aleks is not the only one who would like a "modern" MPI  
> interface for Fortran.  J3 has no official position on this one way  
> or the other.  We are working to make such an interface possible (as  
> part of a larger scope C interoperability TR), but it is up to the  
> MPI group to decide what to support.  The hard reality is that there  
> is a enormous amount of legacy MPI code out there and the "don't  
> make me change my code" crowd is a significant fraction (most likely  
> the majority) of MPI/Fortran users.  So, I think that continuing to  
> support the current style of interface is not really optional.  The  
> question is whether to support an additional interface that is more  
> in tune with modern Fortran, and offers definite advantages to  
> authors of new codes. I don't think it is either/or, but rather one  
> or two.

This topic has come up frequently at MPI Forum meetings and in  
discussions between Jeff Squyres and myself.  The answer to the  
question is that there will be a low level interface that follows the  
current C standard as closely as possible, but provides type checking  
where feasible and solves/provide guidance on the asynchronous problem.

There is also interest in providing a high level (real Fortran)  
interface that takes advantage of the full features in Fortran.  This  
would be a separate effort and not part of the MPI standardization  
effort.  I'll still awaiting news whether there will be funding for  
this or not.

Regards,
Craig