On Wed, 1 Jun 2011 09:58:43 +0200, Riaan van den Dool <[email protected]> 
wrote:
> I assume some of the new CUDA 4.0 features such as cross-device memory
> referencing (I think they call it DirectAccess) will only be available in
> PyOpenCL once the OpenCL specification includes them?

It turns out that CL is actually already prepared to do this, at least
in a sense. Memory allocations in CL are not tied to devices (but rather
contexts, which can contain many devices). There's thus a way to call a
kernel on device A with memory that resides on device B. Unfortunately,
you don't get a lot of control over what memory resides where--I imagine
that might be in the next standard. I.e. it's up to the implementation
to decide whether to access the memory in place or copy it over to
device A.

Andreas

Attachment: pgp0R12Asd4Bl.pgp
Description: PGP signature

_______________________________________________
PyOpenCL mailing list
[email protected]
http://lists.tiker.net/listinfo/pyopencl

Reply via email to