[MPI3 Fortran] MPI-2.1: Fortran 90 bindings

[Jeff Squyres]

On Mar 8, 2008, at 2:03 AM, Harald Klimach wrote:

>> It is required that the send and receive buffers have
>> compatible type
>> maps.  See MPI-1:4.1:

Note that this is the only requirement -- compatible type *maps*.  Not  
compatible *types*.

> So this is about derived types? I don't think it is
> possible to catch derived types within the scope of F90,
> you have to fall back to the implicit F77 interface for
> this anyway.
> Even so, as you said compatible type maps are required,
> which means, as far as I understand it, that there are the
> same base types on each side.
> Looking at MPI-1: 3.3: "Data type matching and data
> conversion", I even think it is required that the two
> arguments are of the same type.

Not really.  You could send N COMPLEX items and receive N*2 REAL  
items.  You can get more creative with C structs and Fortran derived  
types, with the already-mentioned restriction that you might (* see  
below) have to fall back to the F77 bindings for Fortran derived types.

> Observations so far:
> It would be very nice, if Fortran would allow the assumed
> size notation in a generic interface to consume all ranks.

It was pointed out on the call on Friday that most F90 compilers have  
some kind of directive that effectively allows a Fortran subroutine/ 
function parameter to be the equivalent of C's (void*), thereby making  
much of this conversation moot.  We were going to investigate this a  
bit more deeply; this option was not [widely] available when MPI-2 was  
written (nor when the Rasmussen/Squyres paper was written a few years  
ago).

> Which would mean that you'd have to use assumed shape
> arrays if you want to explicitly use different routines
> for different ranks and both options would be mutually
> exclusive.
>
> I think the MPI standard really prohibits different types
> in the case where there are 2 choice arguments.

Per above, I disagree.  :-)

> I can't find anywhere in the standard where it is required
> that arrays of all ranks are allowed to be passed as
> actual arguments. But I can't find where it states that
> scalars are allowed either and in MPI-2 it is explicitly
> stated that this is an inconsistency of MPI with F90.
> Both things (scalars and arrays of different ranks) are
> rather a observed behaviour of the F77 binding, which
> comes from the implicit interfaces.
> Maybe the F90 binding shouldn't exactly rebuild the
> behaviour of the F77 binding?

This is a point that was explicitly brought up when designing the C++  
bindings (that perhaps the C++ bindings should not be effectively a 1- 
to-1 mapping of the C bindings).  The Forum overwhelmingly decided to  
keep them simple and pretty much the same as the C bindings (they do  
take advantage of a small number of C++ language features, to be more  
natural to the C++ developer community).

The rationale is that if you design an alternate set of bindings, it  
is effectively a new and different standard that requires its own,  
separate maintenance, and require a potentially substantially  
different implementation (than the C bindings).

> Having to use only arrays of rank 1 certainly is quite
> inconvinient for the user, but I don't see where it is
> required by the standard to support all ranks in the F90
> binding.

Open MPI defaults to supporting scalar plus ranks 1-4; you can change  
this with a configure argument when building Open MPI.  We did this  
because we noticed that in some compilers, supporting 5-7 ranks made  
compilation of F90 MPI apps "slow" (and users complained).  We found  
"4" to be a "sweet spot".  But we had to offer the *possibility* of  
all 7 for anyone who wanted it.

> To allow scalars and all ranks for all types I think in
> total there are 8 * 15 * 75 = 9000 subroutines needed.
>
> If you restrict yourself to the types, that are at least
> required by MPI (MPI-1: 3.2.2) and don't support any other
> ranks besides rank 1, you're down at: 6*75 = 825
> subroutines.

Based on my disagreements above, I think the above formulae are  
incorrect.

> In MPI-2 Page 13 Line 33 it says:
> All MPI names have an MPI_ prefix, and all characters are
> capitals.
> However this statement "and all characters are capitals"
> hasn't any meaning in the Fortran context, right?


My understanding (and I'll repeat again: I'm nowhere near a fortran  
expert) is that fortran compilers will change subroutine/function/ 
variable symbols to either all upper case or all lower case (possibly  
with some additional underscores).  I don't know if this is a Fortran  
standard thing (e.g., if the Fortran standard says that symbols are  
case-insensitive), or whether it's an implementation thing.

-- 
Jeff Squyres
Cisco Systems