Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!uunet!validgh!dgh
From: dgh@validgh.com (David G. Hough on validgh)
Newsgroups: comp.arch
Subject: More on Linpack pivoting: isamax and instruction set design
Message-ID: <396@validgh.com>
Date: 13 Jun 91 21:08:18 GMT
Organization: validgh, PO Box 20370, San Jose, CA 95160
Lines: 31

Refering to isamax in my previous Linpack posting brought to mind
the quandary that it represents for computer instruction set architects.

I believe that Cray first noticed that when LU-factoring matrices of dimension
n, although saxpy operations
( X += c * Y for scalar c and vector X and Y) are n times more frequent
than isamax operations (find i such that Xi has largest magnitude in X),
you can gain a lot more from agressive vector hardware design and compiler
technology on saxpy than on isamax, so that eventually isamax becomes
the bottleneck.  The heart of isamax is

      do 30 i = 2,n
         if(abs(dx(i)).le.dmax) go to 30
         isamax = i
         dmax = abs(dx(i))
   30 continue

What kinds of additional
instructions belong in an instruction set architecture that is
intended to be scalable, from one-chip systems with inexpensive memory,
to very high performance systems implemented with multiple paths to memory
and perhaps multiple functional units (integer, float, branch)
that may, however, be relatively distant, so that conditional
branches become quite expensive?
A simple max/min won't help.   How can multiple comparisons be overlapped?
Can the "abs" be concealed?
-- 

David Hough

dgh@validgh.com		uunet!validgh!dgh	na.hough@na-net.ornl.gov