Parallel Fortran and PFA

[Bron Campbell Nelson]

In article <7542 at castle.ed.ac.uk>, cmb at castle.ed.ac.uk (Colin Brough) writes:
> In the IRIS-4D Series FORTRAN 77 Programmer's Guide, section 5. FORTRAN
> Enhancements for Multiprocessors, it states:
> 
> 	"To provide compatibility for existing parallel programs,
> 	Silicon Graphics has chosen to adopt the syntax for parallelism
> 	used by Sequent Computer Corporation.  This syntax...."
> 
> Unfortunately the compiler directives subsequently described do not
> match those described in the Sequent ATS Fortran User's Manual (ATS
> Fortran version 1.0).
[excellent comparison deleted]
> 
> Have Sequent recently changed their syntax (ATS is their latest Fortran
> compiler), or have SGI not actually implemented total compatability?

SGI has not implemented total compatibility.  In our first release, we
only implemented a subset of the capabilities found in the Sequent
compiler, intending to extend this in later releases.  However, we found
that our customers were pulling us in slightly different directions, and
so we worked on other extensions, rather than fully implementing Sequent
compatibilty.  As a result, there is a "core" subset that should be
common on both machines, but neither company fully implements all of the
other's commands.

I would at this point like to publicly acknowledge and thank Gary Fielland
of Sequent Computer Systems, Inc. for giving SGI permission to use the
Sequent parallel command syntax in our compiler.  Thanks, Gary.

> 
> On a related topic, does the PFA parallel processing preprocessor only
> produce code with a subset of these parallel processing directives in
> it, or does it make use of them all?  In particular, I understand that
> Kuck and Associates produce both PFA for SGI and the KAP preprocesor for
> Sequent and I was wondering whether both these preprocessors used the
> common subset of facilities provided by both SGI and Sequnet Parallel
> Fortrans's.

Kuck&Associates does produce KAP/PFA.  I believe it is the case that
the only commands that PFA will generate that are *not* in the Sequent
compatible subset are (1) the IF clause, and (2) allowing array variables
to be LASTLOCAL.  You can prevent (1) by setting the command line switch
"-minconc=0", and can prevent (2) by leaving the optimization level at 4
(i.e. "-o=4" is the default for PFA; LASTLOCAL arrays will only happen at
"-o=5").  If you do this, then the results of a PFA run should be
compatible/compilable with Sequent Fortran.  I do not have a Sequent
machine to test this out on however, so I cannot guarentee success, but it
should work (if it doesn't, let me know).

As for the KAP/Sequent product: I suspect that if you set "-cthreshold=-100"
and "-unroll=1" you will get KAP/Sequent to produce the SGI compatible subset,
but again I don't have a Sequent machine, so I can't guarentee this.

--
Bron Campbell Nelson
bron at sgi.com  or possibly  ..!ames!sgi!bron
These statements are my own, not those of Silicon Graphics.