[Q-e-developers] Compile Quantum ESPRESSO with GPU support, the "libcuda" linking issue

[Filippo Spiga]

Dear all,
    a lot of people usually compile Quantum ESPRESSO on HPC clusters
running configure and make on the login nodes (this is the way that should
be!). Unfortunately a lot of people experience problems compiling the QE
with GPU support because the linking phase fails due to "missing libcuda"
(or similar messages).

The problem is well-know and very easy to avoid. The "missing file" is part
of the CUDA driver. That missing library represents the layer
interface between the CUDA runtime environment and the GPU hardware. If you
do not have it on the login node is perfectly fine (and it makes perfectly
sense because you do not need it if you do not have a GPU device attached
to the login node). QE+GPU is designed to do not require at compile time
that library (usually it is installed under /urs/lib64 since it is part of
the operative system). Unfortunately the MAGMA library (the equivalent for
GPU of the LAPACK library) needs it.

So, until future changes in the MAGMA library (changes not so complex,
IMHO), if the linking phase fails the only way to finalize the compilation
is perform the compilation by submitting a job to a compute node. And
everything will go smooth.

Hope this trick helps.

Regards,
F.

--
Filippo SPIGA

«Nobody will drive us out of Cantor's paradise.» ~ David Hilbert
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/developers/attachments/20120328/78a06bca/attachment.html>