I am doing a fair bit of Python programming using the Enthought python
distribution and the Komodo IDE and wanted to give either PyCUDA or
PyOpenCL a try, but have had NIGHTMARES installing either. I thought
PyOpenCL would be the easier of the two. The installation
instructions seem straightforward. I installed boost (which is pretty
rough imo), but when I try to install PyOpenCL one of the many errors
I get is:
Compiling with an SDK that doesn't seem to exist: /Developer/SDKs/
MacOSX10.4u.sdk'
Why would you need the old 10.4u SDK? I can even trace through the
labyrinthine to find where it actually invokes that once I invoke that
black hole command that make usually is.
Any help or ideas?
Dear Romain,
On Mittwoch 02 September 2009, you wrote:
> We use your package PyCuda, which is very nice, for our neural simulator
> Brian (http://www.briansimulator.org). I noticed you recently released
> PyOpenCL, and I was wondering whether you would have any advice on which
> one we should choose for our project. In fact I just set up a google
> group to discuss GPU issues for our simulations and I would very happy
> if you could give us some insight about it:
> http://groups.google.fr/group/brian-on-gpu. Thanks a lot and
> congratulations for your great work!
>
> Best wishes,
> Romain Brette
I put together a wiki page with the best answer I can provide, here:
http://wiki.tiker.net/CudaVsOpenCL
As I've mentioned on the page, I deliberately put this on a Wiki--if you feel
like you can contribute, please don't hesitate to do so.
The fact that I've released PyOpenCL doesn't mean that I'm abandoning PyCUDA,
quite the contrary actually. Writing PyOpenCL was a way for me to take a close
look at the CL spec, and I liked what I saw. CUDA on the other hand is nice,
too. Both CL and CUDA have their advantages. Neither PyCUDA nor PyOpenCL is
going away--I expect there'll be a niche for both of them.
HTH,
Andreas
