Re: [AMBER-Developers] PMEMD now built by default

From: Robert Duke <>
Date: Wed, 3 Mar 2010 08:56:56 -0500

Hi Dave,
Yes, FFTW has never been the default, nor would I recommend it. In my
opinion, you have to be desperate for a little bit of speedup on a small
number of processors (say 1 to at most 4) for it to be useful, and there are
hassles with having to get it yourself and build it yourself. It is a nice
package, but after I reworked our "public" fft code (I think it came from
netlib by way of Darden), the advantage is really small. Also, to get the
advantage, I call it with its "find the best algorithm" setup code turned
on. This results in it sometimes producing slightly different results on
the same machine (apparently two algorithms are sufficiently close in
performance that OS scheduling can influence which one is cheapest during a
setup). This can be annoying in test.
Regards - Bob
----- Original Message -----
From: "case" <>
To: "AMBER Developers Mailing List" <>
Sent: Wednesday, March 03, 2010 8:40 AM
Subject: Re: [AMBER-Developers] PMEMD now built by default

> On Tue, Mar 02, 2010, Ross Walker wrote:
>> I know there are things like FFTW but
>> I never saw much speedup with that so figured it wasn't worth the hassle.
> I recommend against using FFTW by default. There are weird license issues
> involved: FFTW is GPL, but pmemd is not. This doesn't prevent an
> individual
> user from making the link themselves (since they can't redistribute the
> code
> anyway), but I'd like to avoid having it be the default, where possible.
> To everyone: be aware of license issues if you use any outside code.
> (Tyler: don't let me forget that we need a final word from NINT before
> releasing RISM as GPL...)
> ...thx...dac
> _______________________________________________
> AMBER-Developers mailing list

AMBER-Developers mailing list
Received on Wed Mar 03 2010 - 06:00:30 PST
Custom Search