On Dienstag 23 Dezember 2008, Jean-Christophe Penalva wrote:
at the end of the install :
$ cd PyCuda-VERSION/test
$ python test_driver.py
need a gcc on the node with the GPU ?
Here, we have a cluster with login node (on wich i've compil+link) and
executions nodes with GPU (but whithout compiler).
For now, you definitely need nvcc on that machine if you're going to use
SourceModule, and I think nvcc calls out to gcc unconditionally--even though
I'm not sure why. If you'd like to avoid that, you have a number of
- don't use SourceModule (and gpuarray, which uses SourceModule)
- make sure that PyCuda finds compiled versions of every module it needs in
the cache directory
- hack SourceModule to specifically look for precompiled kernels elsewhere.