RE: [AMBER-Developers] Help with NVIDIA Compilation using ifort

From: Ross Walker <ross.rosswalker.co.uk>
Date: Tue, 30 Mar 2010 20:35:30 -0700

Hi Bob et al,

I have tracked this down to an issue with the -fast option, specifically the
'-ipo' option that this implies. If one removes this then things link
properly. From what I can gather -ipo is corrupting the symbols in
cuda/gpu.o which leads to the link issues, even though there is no compiler
error.

I will update the configure file to fix this and send it to Dave.

That said I am still seeing problems when the icc command line has
optimization other than -O0 on it. :-(

All the best
Ross

> -----Original Message-----
> From: amber-developers-bounces.ambermd.org [mailto:amber-developers-
> bounces.ambermd.org] On Behalf Of Robert Duke
> Sent: Tuesday, March 30, 2010 5:44 AM
> To: AMBER Developers Mailing List
> Subject: Re: [AMBER-Developers] Help with NVIDIA Compilation using
> ifort
>
> Okay, I still am not set up to build cuda, but I plowed through the
> logic on
> all this, and it is sort of strange. What we are dealing with here is
> NOT
> under the control of CLINK* (so scratch the whole discussion of how to
> get a
> c name decorated in our source - that is not the problem). This is
> logically equivalent to having to build mpi libs (also an imported .a)
> such
> that the name decoration matches what the fortran compiler is
> expecting. So
> the question becomes, "what do you get when you do a 'nm cuda.a' and
> grep
> for the undecorated missing symbols (or just gpu)"? I am puzzled by
> all
> this because it looks like gnu-based compiles are set up in your
> generic
> configure to produce a single underscore, so if cuda.a works with gnu
> fortran, I would expect it to work with ifort (so this is maybe
> slightly
> convoluted logic, but I am saying that when we build c for either ifort
> or
> gfortran ourselves, we use single underscore decoration, the default if
> there is not a CLINK* definition, so what we give to gfortran
> presumably is
> the same thing we give to ifort (in terms of name decoration of our own
> c
> source).
> Regards - Bob
> ----- Original Message -----
> From: "Ross Walker" <ross.rosswalker.co.uk>
> To: "'AMBER Developers Mailing List'" <amber-developers.ambermd.org>
> Sent: Tuesday, March 30, 2010 12:54 AM
> Subject: RE: [AMBER-Developers] Help with NVIDIA Compilation using
> ifort
>
>
> >> What platform are you building on ?
> >
> > RedHat EL 4U8, x86_64.
> >
> > Gnu works fine. Intel is currently broken at link stage.
> >
> >> And is all this cuda stuff in the latest release candidate ?
> >
> > Yes. Try:
> >
> > ./configure -cuda intel
> >
> >> Yes, Bob is also mostly using CLINK... in pmemd.
> >> I am not aware of any platform with such a terribly broken C
> >> preprocessor;
> >> this includes over the years ibms, hps, suns, sgis, crays, etc.
> >
> > I'll take a look at Bob's CLINK and see if I can figure something
> simple
> > out.
> >
> >> Here's my comment that we are too close to release for all this
> >> activity...
> >
> > Fine. Release it broken, see if I care. I'll be sure to forward all
> the
> > user
> > comments to your cell phone's email to sms gateway. ;-)
> >
> > All the best
> > Ross
> >
> > /\
> > \/
> > |\oss Walker
> >
> > | Assistant Research Professor |
> > | San Diego Supercomputer Center |
> > | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
> > | http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
> >
> > Note: Electronic Mail is not secure, has no guarantee of delivery,
> may not
> > be read every day, and should not be used for urgent or sensitive
> issues.
> >
> >
> >
> >
> >
> >
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
> >
> >
>
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers


_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Mar 30 2010 - 21:00:04 PDT
Custom Search