Path: utzoo!utgpu!watmath!clyde!att!alberta!ubc-cs!uw-beaver!cornell!mailrus!purdue!i.cc.purdue.edu!k.cc.purdue.edu!l.cc.purdue.edu!cik
From: cik@l.cc.purdue.edu (Herman Rubin)
Newsgroups: comp.arch
Subject: Re: Quadruple-Precision Floating Point ?
Summary: That something is not provided is no reason to believe it should not be
Keywords: REAL*16 hardware
Message-ID: <1080@l.cc.purdue.edu>
Date: 26 Dec 88 11:08:49 GMT
References: <8561@alice.UUCP> <3688@s.cc.purdue.edu> <285@loligo.fsu.edu> <6132@ecsvax.uncecs.edu>
Organization: Purdue University Statistics Department
Lines: 40

In article <6132@ecsvax.uncecs.edu>, urjlew@ecsvax.uncecs.edu (Rostyk Lewyckyj) writes:
> Since none of the supercomputers i.e. Cray, ETA, CONVEX,
> (By definition of most who claim to be supercomputer experts,
> IBM does not qualify ) provide quadruple precision hardware
> support. (Cray double precision is 96 bits. IBM Q precision
> with some hardware support is 112 bits.), there is by common
> concensus no need for reasonable speed Q precision calculations.
> So you should either find another way of doing it, after all
> extended accuracy is less important than speed, as evidenced
> by the great emphasis placed on Cray speed and none placed on
> the other companies arithmetic libraries, OR you may need
> to do your work on that off brand IBM equipment.

This insistence that the hardware and software provided is the best possible,
and that if it is not there one should not want it, is totalitarian and stupid.

What is needed is input from those who can think and can find uses for
operations which can be efficiently done in hardware, but not efficiently
in software, to make these needs known to the manufacturers.  There is need
for the software operations which those who understand the use of bit patterns
find useful to be included in the languages.

That these operations are not much used is no evidence.  If I know that a
square root is 10 times as expensive as a division, and 60 times as expensive
as a multiplication, you can be sure that I will choose my algorithms with
this in mind.  If I know that Cray double precision is many times as expensive
as single precision, I may come up with tricky algorithms to get around it.
If the Cray had an additional operation to get the least significant part of
a product, it would probably speed up multiple precision by at least a factor
of 5.  And sometimes this is needed for other purposes.

I know of no computer reasonably designed for multiple precision operations.
What is needed is fast unsigned integer arithmetic.  The hardware for 
integer multiplication is certainly no more difficult than for floating
point, and for unsigned actually easier than signed..  With the CPU such
a small price of the computer, get those instructions in.
-- 
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907
Phone: (317)494-6054
hrubin@l.cc.purdue.edu (Internet, bitnet, UUCP)