This is a neat feature, but do you really need to call from a kernel
directly? If all you want to do is call CUBLAS functions on pycuda vector
types, then you can use scikits.cuda. The driver and runtime api
are inter-operable on a program level, just not between kernels, yet.

http://www.scribd.com/doc/88255296/228/G-4-Interoperability-between-Runtime-and-Driver-APIs

http://lebedov.github.com/scikits.cuda/

On Fri, Aug 17, 2012 at 5:12 PM, Stanley Seibert <[email protected]>wrote:

> I too am looking forward to trying this out, but it is important to note
> that this feature is limited to compute capability 3.5 devices, none of
> which have been released yet...
>
> On Aug 17, 2012, at 4:36 PM, andrew cooke <[email protected]> wrote:
>
> >
> > Hi,
> >
> > At http://developer.nvidia.com/cuda/cublas it says:
> >
> >  *New in CUDA 5.0* Device API that can be called from CUDA kernels
> >
> > Which I take to mean that it is now possibel to use CUBLAS while using
> the
> > Driver API.
> >
> > Meanwhile :o) at
> >
> http://wiki.tiker.net/PyCuda/FrequentlyAskedQuestions#Are_the_CUBLAS_APIs_available_via_PyCUDA.3F
> > it says, while explaining why PyCUDA doesn't dupport CUBLAS:
> >
> >  PyCUDA is based on the driver API. CUBLAS uses the high-level API.
> >
> > So... have you noticed this?  Have you found any more information (I
> can't
> > find anything on CUBLAS 5 although I've got CUDA 5 installed and
> running)?  Do
> > you have any plans?
> >
> > I am not sure PyCUDA will need to do anything - it looks like I can
> simply
> > call CUBLAS from a kernel - although I guess you could auto-generate
> simple
> > kernels that wrapper each call.
> >
> > Most of all I'd like to find more docs.  Anyone?
> >
> > Thanks,
> > Andrew
> >
> > _______________________________________________
> > PyCUDA mailing list
> > [email protected]
> > http://lists.tiker.net/listinfo/pycuda
>
>
> _______________________________________________
> PyCUDA mailing list
> [email protected]
> http://lists.tiker.net/listinfo/pycuda
>
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda

Reply via email to