On Thu, 18 Jun 2015 11:44:03 +0000
<[email protected]> wrote:

> Note: This is an existence proof, not an approach.  I noticed some people 
> getting halfway there. This gets all the way, but somebody with expert 
> knowledge of pyOpenCL design would be best to integrate clMath into pyOpenCL.

Andreas is probably the best person to answer but he did not include 
"scikit-cuda" which provides bindings for cuFFT. I guess his point of view is 
the same for a similar issue.

> Also AMD told me clAmdBlas is deprecated in favor of open source clMath on 
> github, so clMath is the thing to incorporate. 

> ClMath has a pyopencl wrapper in the distribution, but there is not a soup to 
> nuts example of how to use it in a purely pyOpenCL and pure Python way. Also 
> other features aren't wrapped.

It is a good move from AMD: for the FFT part, it is an external project which 
is providing the bindings.

> Also there is githuib clRnd as well as clFFT mentioned.

I did not know this one. We did some work with the debian-science community 
(thanks Ghislain) to package the blas and fft part. My idea was to make them 
available  from python.

> Not clear to me where wrappers should go so that user in Python (not writing 
> app code in Cython),can see all of these libraries in pyOpenCL. Should 
> wrappers go in each package or in pyOpenCL package?
As the library is C++, the python wrapper needs to be either in clMath either 
in a separate project. In pyOpenCL it is likely to de-stabilize pyopencl 
(seg-fault, library not found, ...)

Cheers,

Jerome

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

Reply via email to