Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!ub!uhura.cc.rochester.edu!rochester!pt.cs.cmu.edu!gandalf.cs.cmu.edu!lindsay
From: lindsay@gandalf.cs.cmu.edu (Donald Lindsay)
Newsgroups: comp.arch
Subject: Re: Vector vs Cache/Superscalar
Message-ID: <13012@pt.cs.cmu.edu>
Date: 12 May 91 05:57:59 GMT
References: <2646@fornax.UUCP>
Organization: Carnegie Mellon
Lines: 20

In article <2646@fornax.UUCP> bremner@cs.sfu.ca (David Bremner) writes:
>>I think this is partly because the vector model of parallelism is
>>so rigid; optimization for the superscalars involves a bigger bag
>>of tricks.
>I think the vector model is actually quite a bit nicer.

There's yet another model - the data parallel one. By this, I mean
that one says e.g. "a=5", and 64,000 a's are set. This is obviously
the natural programming style of SIMDs like the Connection Machine.

- or so I thought, until I heard about the CM Fortran 1.0 compiler
from Thinking Machines. It uses the Fortran-90 style of vector
operations, e.g.
	x(1:100:2) = y(1:50) + z(1:50)

and (for scientific problems) it seems to be generating faster code
than ever before. TMI is claiming that they commonly see 5 GFLOPS,
wheras data-parallel Paris was getting more like 1.5 GFLOPS.
-- 
Don		D.C.Lindsay 	Carnegie Mellon Robotics Institute