[PyCUDA] using libraries with pycuda

[Andreas Klöckner]

On Dienstag 29 September 2009, Dan Goodman wrote:
> Hi all,
> 
> Has anyone tried using pycuda with a library of GPU code such as CUDPP
> [1] or chag:pp [2]? How do you use these in pycuda? Is it just a matter
> of passing some header and lib file directories to nvcc via the
> options=... keyword for SourceModule?

SourceModule won't work because it's meant only for device-side code, and most 
of these libraries expose host-side interfaces.

If the library is written with the driver API in mind, it's as simple as 
wrapping it into Python using, e.g. Boost.Python, cython, etc. If it is using 
the CUDA "runtime" API, it's not quite as simple, because then Nvidia's 
idiotic driver/runtime dichotomy kicks in and prevents direct communication. 
Luckily, however, I learned at the Nvidia conference last week that a) they're 
aware that this is causing problems, and b) this issue will go away in CUDA 
3.0, which will be released along with their new hardware.

HTH,
Andreas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.tiker.net/pipermail/pycuda/attachments/20091004/588b14ce/attachment-0001.pgp>