Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!ccu.umanitoba.ca!herald.usask.ca!alberta!brazeau.ucs.ualberta.ca!unixg.ubc.ca!ubc-cs!van-bc!zaphod.mps.ohio-state.edu!samsung!noose.ecn.purdue.edu!mentor.cc.purdue.edu!mace.cc.purdue.edu!bfp
From: bfp@mace.cc.purdue.edu (Bryan Putnam)
Newsgroups: comp.lang.fortran
Subject: Re: Need complex matrix solver for CM
Message-ID: <7902@mace.cc.purdue.edu>
Date: 28 Jun 91 20:06:52 GMT
References: <1991Jun27.190724.26161@ariel.unm.edu> <7887@mace.cc.purdue.edu> <1991Jun28.061454.29461@fs7.ece.cmu.edu>
Organization: Purdue University Computing Center
Lines: 24

In article <1991Jun28.061454.29461@fs7.ece.cmu.edu> winstead@faraday.ece.cmu.edu (Charles Holden Winstead) writes:
>
>Yo dude, grow up!  I agree that it's a simple task of changing a complex 
>linear system into a real one, but with a language that deals with complex
>numbers as nicely as fortran does, let's agree that there are some benefits
>to not switching from complex to real and back again repeatedly.  Why not
>

There can also be some benefits to separating the data into real and
imaginary parts, because on some machines the vector hardware
instructions do not operate on complex data. If you the programmer do
not separate the data, then the compiler has too.  For example, the
highly optimized FFT routines written for the CYBER 205, and which
were used by the National Weather Service, start off by SEPARATING
the real and imaginary parts simply because the vector hardware is
more efficient at performing real arithmetic. Too many programmers
get into trouble by ignoring the architecture of specific machines.

>Why not
>just look through the code that works for real matrices and change the 
>corresponding variable declaration from real to complex.

I suppose the obvious answer would be that the source code is seldom
available, unless of course this is a trick question.