Re: [AMBER-Developers] Re: problem installing pmemd.cuda

From: Jason Swails <>
Date: Fri, 2 Apr 2010 21:51:26 -0400

On Fri, Apr 2, 2010 at 9:11 PM, case <> wrote:
> On Fri, Apr 02, 2010, Jason Swails wrote:
>> >
>> > Lesson learned ---> google first.  It's a problem with GNU 4.4
>> > compilers i'm led to believe (CUDA issue, not Amber).  I'll try again
>> > after I install older compilers.
>> Rolling back to 4.2.4 worked great.  Is this documented anywhere?
> I've only found it at the NVIDIA forums.  Emails indicate that Scott LeGrand
> is not aware of this [I found the same error as you, and emailed Ross and
> Scott about it.]  According to the infallible Web, anything before 4.4
> is likely to work.  [My WAG: g++ keeps getting more and more strict about
> the language, and NVIDIA headers that used to be OK are now flagged as
> errors.]

That's what was happening to me as well. I also tried installing gcc
4.3.4 from aptitude on my Ubuntu box and compile with gfortran4.4 with
gcc4.3/g++4.3, but I got the same errors. Then I installed gcc 4.2.4
from aptitude and retried (this time using gfortran4.2 with gcc/g++
4.2), and I finally succeeded. I don't know if this means that gcc4.3
is no good, because I still tried to use gfortran4.4 that time and I
don't know if that was the source of the errors. Anyway, I would
prefer to avoid rampaging around in my /usr/bin directory since it's a
great way of shooting myself in the foot, so I didn't go back and try
4.3.4 the same way I got 4.2 to work.

On another note, I installed pmemd.cuda on the local HPC with intel
compilers (I've installed it with gnu compilers there before), and it
worked fine and passed all tests (data on wiki).

On another note, I've been trying very hard to work cuda magic on Mac
OS X 10.6, but to no avail. I've gotten it as far as the link stage
when it dies with a handful of errors like this:
      ___gxx_personality_v0$non_lazy_ptr in cuda.a(cudpp_scan.o)
I do know, however, that currently the -m32 flag has to be passed to
both gfortran and gcc/g++ since nvcc will only build 32-bit binaries
and you'll get architecture complaints without it

> It's not yet documented in the Amber manuals, but then the entire chapter
> on CUDA does not yet exist....

I'd be willing to help if there's any way that I can, even if it's
just reading it over... (I'm certainly not an expert here, but I've
used it :) ).


Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Graduate Student
AMBER-Developers mailing list
Received on Fri Apr 02 2010 - 19:00:03 PDT
Custom Search