Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!maverick.ksu.ksu.edu!zaphod.mps.ohio-state.edu!usc!randvax!edhall
From: edhall@rand.org (Ed Hall)
Newsgroups: comp.sys.ibm.pc.hardware
Subject: Re: 249 hours to generate MANDELBROT image on 286
Message-ID: <1991May4.003029.16362@rand.org>
Date: 4 May 91 00:30:29 GMT
References: <14697@encore.Encore.COM> <1991May2.124542.21865@engin.umich.edu> <1991May2.142752.23003@msuinfo.cl.msu.edu>
Sender: usenet@rand.org
Reply-To: edhall@rand.org
Distribution: usa
Organization: The RAND Corporation
Lines: 32
Originator: edhall@ives
Nntp-Posting-Host: ives

In article <1991May2.142752.23003@msuinfo.cl.msu.edu> draper@buster.cps.msu.edu (Patrick J Draper) writes:
>Isn't there an option to use either integer or floating point math? If
>so, a coprocessor wouldn't save you anything in the integer mode.
>       .       .       .       .
>I'll bet that your friend's 386 is coonfigured for integer math, and the
>286 machine is configured for floating point math.

I believe that the later versions of fractint will use 32-bit integer
instructions on a '386.  Even without a coprocessor, all other things
being equal this can easily be a factor of 2 (adds) to 4 (multiplies)
faster.  Also, the '286 is hard pressed to hold in registers
everything needed in a mandelbrot inner loop (at two 16-bit registers
per value); the 32-bit registers on the '386 can do this easily.
Register-to-register operations are about twice as fast as
register-to-memory, on average.  Finally, the shifts necessary for
such fixed-point arithmetic are a lot faster on a '386, which has a
barrel-shifter, than the '286, which has to consume extra cycles for
each bit shifted.  Oh, and one last thing: the '386 uses an
"early-out" multiplication algorithm, where multiply cycles are only
performed until the multiplier runs out of significant bits.

So I'd give the '386 a factor of eight or so based upon architecture
alone.  That times 4 for the clock still falls far short of explaining
a 100:1 difference.  However, if the calculations were floating-point,
100:1 sounds small for a non-coprocessor 8MHz '286 versus a 33MHz '386
with coprocessor.  On some floating-point code I once ran for waveform
synthesis, a non-CP 10MHz '286 was about 200 times slower than a
25MHz '386 with CP.  (I'm assuming that the mandelbrot code is as
FP-intensive, which it should be.)

		-Ed Hall
		edhall@rand.org