[MPI3 Fortran] Summary of items on the table

[Bill Long]

Craig Rasmussen wrote:
> There has been quite a lot of traffic in the past couple of months 
> regarding possible directions for a new MPI binding for Fortran.  I 
> thought now would be a good time to summarize and try to get a 
> consensus on what direction we should pursue.  Some choices will 
> require changes to the Fortran standard and this input we need to get 
> to the Fortran standards committee within the next two weeks.
>
> The major choice we need to make is how to declare an MPI "choice 
> buffer" (void *) in a Fortran interface.  I'll list the options below 
> with advantages and disadvantages as I see them.  Please rank these 
> options (they all will use F2003 BIND(C)).
>
> 1. Map void* to C_PTR.  Upside: this will work now in most compilers 
> and likely all soon.  Downside: users must change the call site to use 
> C_LOC to obtain the C address; C_LOC requires the POINTER or TARGET 
> attribute and may reduce performance and require additional code changes.
>

Changing the actual argument to a C_LOC reference also has the side 
effect of making the actual argument an expression whose value is not a 
Fortran pointer.  Thus, the corresponding dummy argument cannot be 
defined.  A problem for about half of the interesting MPI cases.


> 2. Map void* to a new Fortran type C_VOID.  Upside: no changes to 
> current code (I think). Downside: would require a change in the 
> Fortran standard and couldn't be done until Fortran 20xx (a long time 
> before we see it in compilers).

I think the downside here is significant.  Adding a new intrinsic type 
has significant ramifications for compiler internals.   I think chances 
for this to be added even in 20xx are slim.


>
> 3. Map void* to type(C_VOID) defined in the ISO_C_BINDING intrinsic 
> module and add LANG=C to the Fortran standard.  Upside: no changes to 
> current code other than to USE the MPI3 module.  Downside: it will 
> require changes to the Fortran standard but can get to users 
> relatively soon (before F2008 appears in compilers).
>
> 4. Add an IGNORE attribute with parameters T (type) K (kind) R (rank) 
> for BIND(C) interfaces.  This pretty much (as I understand it) does 
> the same thing as #3, just with different syntax and with slightly 
> more flexibility (partial type checking).  Upside: see #3.  Downside: 
> see #3.

Option 3 is basically a subset of 4 that allows only IGNORE(TKR) and 
restricts the usage to BIND(C) procedures that are written in C.  Option 
4 is more general, and does not require breaking the F/C symmetry in the 
current rules.  Perhaps a more compelling case for 4 is that many 
compilers already support this functionality through compiler 
directives.  (How do you think we implement our MPI modules today?)    
Use of 4 would not have to be limited to BIND(C) interfaces.  Because of 
existing directives, I think 4 will be easy for most vendors to accept.  
One of the J3 alternates from Intel has already stated this is his 
preferred choice.


>
> Aleks has submitted a proposal to the Fortran J3 committee to improve 
> on the ASYNCHRONOUS attribute.  Hopefully this can make it into the 
> F2008 standard and fix problems associated with copyin/copyout 
> semantics and with code motion associated with asynchronous MPI 
> calls.  A question I have is there a problem with copyin/copyout for 
> synchronous MPI?

At least there is a performance degradation.   For codes that use a lot 
of memory already, you increase the risk of going over some limit.

Cheers,
Bill

>
> I'm currently at the April MPI Forum where I'll collect votes.  Others 
> please add your feedback (rank the options) in email (actually it 
> would be helpful for the record to get email from everyone within 
> shouting distance).
>
> A question just came up about passing CHARACTER parameters.  Comments?
>
> Regards,
> Craig
>

-- 
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