[petsc-dev] PETSc-Cuda with gfortran 4.7

[Nystrom, William D]

Satish,

Cuda 5 has the same check and error out.  Only Cuda 5 supports up to gcc 4.6.x.

As far as I know, the same issue will be relevant for future cuda releases.  I talked to
an Nvidia person at SC12 who indicated there were issues that were not that easy to
work around.  Based on that, I would guess that the same restriction on gcc version
will be in Cuda 5.1 - except that maybe it will support gcc 4.7.x.

I was trying to understand why Asmund has problems building petsc with 4.7.x gnu
compilers when I am able to build fine with cuda support using 4.7.2 as long as I
force nvcc to use 4.6.3 or earlier as mentioned below.

Dave

________________________________________
From: petsc-dev-bounces at mcs.anl.gov [petsc-dev-bounces at mcs.anl.gov] on behalf of Satish Balay [balay at mcs.anl.gov]
Sent: Wednesday, March 13, 2013 8:43 AM
To: For users of the development version of PETSc
Subject: Re: [petsc-dev] PETSc-Cuda with gfortran 4.7

Since CUDA is very specific about the version of gcc - should we
bother supporting such combinations [that cuda doesn't support?]

For eg: for a very long time - you needed gcc-4.4 for cuda 4.x series.
[it would do a version check and give error].

Does cuda-5 do the same check and error out? Do you do someting to
supress this error?

Will this problem stay with say cuda-5.1 or future releases?  So I'm
not sure if this should go into include/petsc-private/petscimpl.h.

Perhaps a configure test for the appropriate case [of cuda version +
gcc version that requires it - in petscconf.h or CXXFLAGS?]

Satish

On Wed, 13 Mar 2013, Åsmund Ervik wrote:

> Hi Karli,
>
> I don't think this interferes with anything else, since all it does is
> undef some C++ stuff which conflicts with CUDA, and was introduced in
> gcc 4.7. That is, the things it undefines are also undefined when one
> uses gcc 4.6 or lower. It is general for gcc >= 4.7.
>
> Best regards,
> \AA smund
>
> On 12. mars 2013 15:13, Karl Rupp wrote:
> > Hi Ervik,
> >
> > thanks for the hint. Do you happen to have any information on whether
> > the suggested fix interferes with other versions of GCC or CUDA? Is this
> > specific to version 4.7.2 or a 'general fix' for GCC?
> >
> > Best regards,
> > Karli
> >
> >
> > On 03/12/2013 02:35 AM, Åsmund Ervik wrote:
> >> Hi all,
> >>
> >> (This suggestion is two lines of code, so I hope it will survive the git/hg apocalypse/transition today. If not, I'll resend it tomorrow.)
> >>
> >> Trying to compile PETSc with Cuda enabled, following the example in config/examples/arch-cuda-double.py, I get an error with gfortran 4.7.2 when making (configure works fine).
> >> This error is apparently well-known with Cuda and gcc 4.7.2. To fix it, I added two lines to include/petsc-private/petscimpl.h:
> >> #undef _GLIBCXX_ATOMIC_BUILTINS
> >> #undef _GLIBCXX_USE_INT128
> >> and everything works fine. I guess you should add this to petsc-dev, possibly with some if-statement around these lines.
> >>
> >> Best regards,
> >> Åsmund Ervik
> >>
> >
> >
>
> --
> Sendt fra Thunderbird
>