Newsgroups: comp.lang.lisp
Path: cantaloupe.srv.cs.cmu.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!yeshua.marcam.com!usc!elroy.jpl.nasa.gov!decwrl!netcomsv!netcom.com!hbaker
From: hbaker@netcom.com (Henry G. Baker)
Subject: Re: C is faster than lisp (lisp vs c++ / Rick Graham...)
Message-ID: <hbakerCvnzvE.LJw@netcom.com>
Organization: nil
References: <345du6$g4h@cantaloupe.srv.cs.cmu.edu> <TMB.94Sep2023108@arolla.idiap.ch> <LGM.94Sep4131854@polaris.ih.att.com>
Date: Mon, 5 Sep 1994 16:16:26 GMT
Lines: 40

In article <LGM.94Sep4131854@polaris.ih.att.com> lgm@polaris.ih.att.com (Lawrence G. Mayka) writes:
>In article <TMB.94Sep2023108@arolla.idiap.ch> tmb@arolla.idiap.ch (Thomas M. Breuel) writes:
>
>   This "frivolous consing" is absolutely unacceptable in numerical
>   code.  Even if collecting the floating point garbage were completely
>   free, just the cost of allocating the values and dereferencing them is
>   often much too expensive.  And, in fact, even systems with excellent
>   generational GC become very slow for numerical code whenever there is
>   any consing going on in the numerical code.
>
>I've sometimes wondered: Does floating-point numerical code, of the
>kind you describe, typically require IEEE precision (double or
>single)?  Or can such code often make do with, say, the same 7 bits of
>exponent as IEEE single-precision but only 19 bits of significand
>(FLOAT-DIGITS) instead of 24?  My real question is, How often can a
>cons-less SHORT-FLOAT meet this need for fast floating-point numerics?

IEEE 'single precision' (32 bits) is considered essentially worthless by
most numerical analysts.  Even John Neumann's original article on his
'Von Neumann architecture' argued for a minimum of 38-40 bits.

I've been told that IEEE single precision is only good for glorified
signal processing applications, and even then, the lack of a decent
double precision accumulator limits its usefulness.

Even IEEE double (64 bits) is minimal, with most microprocessors opting
for 80 bits for their 'double' size.

If storage for these boxed floats is managed using some of the techniques
of 'linear logic', the cost of storage management should still be dominated
by (or overlapped with) the computational cost, on all but the fastest
implementations.

Refs on linear logic storage management.

'Sparse Polynomials and Linear Logic'.  ACM Sigsam  Bull.  27,4 (Dec. 1993),
10-14.

'A Linear Logic Quicksort'.  ACM Sigplan Notices 29,2 (Feb. 1994), 13-18.

