Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!wuarchive!udel!nelson
From: nelson@udel.edu (Mark Nelson)
Newsgroups: comp.arch
Subject: Re: Cray divide speed
Message-ID: <2923@nigel.udel.EDU>
Date: 30 Oct 89 19:29:25 GMT
References: <MCCALPIN.89Oct29174054@masig3.masig3.ocean.fsu.edu>
Sender: usenet@udel.EDU
Reply-To: nelson@udel.edu (Mark Nelson)
Distribution: usa
Organization: University of Delaware
Lines: 36


In article <MCCALPIN.89Oct29174054@masig3.masig3.ocean.fsu.edu> you write:
>Does anyone have the figures handy for the peak speed of the Cray
>X/MP, the Cray Y/MP, and the Cray 2 for vector divide operations?
>
>I know that Cray uses an approximate reciprocal scheme, but I need to
>know how many cycles per result are required for a plain vector
>divide, i.e.
>
>	DO 100 I=1,N
>	    A(I) = B(I)/C(I)
>  100	CONTINUE
>
For all Cray machines, a full precision divide requires one
pass through the reciprocal approximation functional unit and
three passes through the fp multiply functional unit.  For the
X/MP and Y/MP, these are distinct functional units and the
reciprocal can be chained with one of the multiplies, so the
peak speed is one result/three clock periods.

For the Cray 2, both operations are handled by the same functional
unit, and there isn't any chaining, so the peak speed is
one result/four clock periods.

If you only need half precision (26 bits of accurate mantissa)
the divide can be done in one multiply, so peak speed becomes:
X/MP Y/MP: one result/clock period
Cray 2:    one result/two clock periods.

However, I don't think you can persuade the FORTRAN compilers
to generate half precision divides.
--

Mark Nelson                 ...!rutgers!udel!nelson or nelson@udel.edu
This function is occasionally useful as an argument to other functions
that require functions as arguments. -- Guy Steele