Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.csd.uwm.edu!cs.utexas.edu!uunet!crdgw1!CRD.GE.COM
From: oconnordm@CRD.GE.COM (Dennis M. O'Connor)
Newsgroups: comp.arch
Subject: Re: John von Neumann, sqrt instr
Message-ID: <1732@crdgw1.crd.ge.com>
Date: 18 Aug 89 02:20:03 GMT
References: <21353@cup.portal.com> <25643@obiwan.mips.COM>
Sender: news@crdgw1.crd.ge.com
Reply-To: oconnordm@CRD.GE.COM (Dennis M. O'Connor)
Distribution: usa
Organization: GE Corporate R&D Center
Lines: 29
In-reply-to: mark@mips.COM (Mark G. Johnson)

mark@mips (Mark G. Johnson) writes:
]In article <21353@cup.portal.com> mmm@cup.portal.com (Mark Robert Thorson) writes:
]    >I remember reading in an old AFIPS paper that von Neumann believed
]    >computers of the future would all have the SQRT instruction, because of
]    >the importance of square root in coordinate geometry.
] 
]Several of the RISC camps, armed with gigabytes of traces and simulation
]results, have decided to include FP square root as well.  I suspect it is
]because of the importance of square root in computing the width of the
]depletion layer at a semiconductor junction.

Three years ago, a team working on a RISC at GE went and asked the
real-time-control people at GE what they need in the way of complicated
floating-point operations. Other than +-*/, the answer wasn't
square root. It was ( drum roll ... )

	Inverse Square Root ( i.e. X to the negative one-half power ).

Seems control theory is riddled with inverse square roots.
And it's a lot faster to do Inv.SQRT than SQRT followed by divide,
as you would expect. However, by putting a 64-entry "first-guess"
table in memory, and doing Newton-Raphson, you can do this
with multiplies almost as quickly :-) as doing it in microcode.

There ARE other considerations, of course. Right and wrong
aren't applicable judgements to make in this domain.
--
 Dennis O'Connor      OCONNORDM@CRD.GE.COM       UUNET!CRD.GE.COM!OCONNORDM
 Rimmer : "This isn't part of my fantasy!" Cat : "No, it's part of mine!"