Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!samsung!uunet!munnari.oz.au!brolga!bunyip.cc.uq.oz.au!marlin.jcu.edu.au!csrdh
From: csrdh@marlin.jcu.edu.au (Rowan Hughes)
Newsgroups: comp.benchmarks
Subject: Re: A benchmark
Keywords: snake performance
Message-ID: <1991May3.124937.17144@marlin.jcu.edu.au>
Date: 3 May 91 12:49:37 GMT
References: <1991May3.023705.5616@marlin.jcu.edu.au> <1991May3.053053.29174@rice.edu>
Organization: James Cook University
Lines: 52

In <1991May3.053053.29174@rice.edu> preston@ariel.rice.edu (Preston Briggs) writes:

>It's important to remember that code that is ideal for a vector machine
>is not necessarily ideal for a scalar (or super-scalar) machine.
>Yes, the IBM and HP machines can cook on vector code,
>but often it can be rearranged for even better performance.
...
>This is ok for vector machines


>	DO j = 1, n
>	    DO i = 1, n
>		A(i) = A(i) + B(j)
>	    ENDDO
>	ENDDO

>but this is better for scalar machines
>(and terrible for vector machines, because of the recurrence on A(i))

>	DO i = 1, n
>	    DO j = 1, n
>		A(i) = A(i) + B(j)
>	    ENDDO
>	ENDDO

I do see your point about scalar machines. A quick scan through the
hot spots in the code showed that potential chaining couldn't be had.

Any sensible vector compiler would do a loop inversion for the second
example, or possibly a chain with A(I) treated as a scalar. 
The GFDL code is essentially dot and vector products, and triads,
which should be OK for both types of machines.

Scalar chaining in the cache should be regarded as a bonus, not a necessity,
and benchmarks which exclusively use it (eg linpack300x300 on HP9000-750)
should be treated with contempt.

Chaining is also quite useful for vector machines too, and necessary
in a case like SUM=SUM+A(I). You should be aware that the backwards
recurrence problem has been solved for vector compilers, and Fujitsu's
latest vector compiler permitts 1st order forward recurrence.

Philosophical note:
I really hate the idea of butchering code to suite some particular
machine; computer scientists were invented to write compilers that
took care of that! The UNIX curly bracket bozos can't even write
a bug-free f77 compiler !!

-- 
Rowan Hughes                                James Cook University
Marine Modelling Unit                       Townsville, Australia.
Dept. Civil and Systems Engineering         csrdh@marlin.jcu.edu.au