[Numpy-discussion] magma?

Lev Givon lev@columbia....
Fri Nov 18 13:21:30 CST 2011


Received from Robert Kern on Fri, Nov 18, 2011 at 08:38:48AM EST:
> On Fri, Nov 18, 2011 at 13:12, Neal Becker <ndbecker2@gmail.com> wrote:
> > Is there any possibility of incorporating this work into numpy?
> >
> > http://icl.cs.utk.edu/magma/software/index.html
> 
> Into numpy itself? Very low probability, I think. The dependency on
> nVidia hardware and proprietary compilers makes supporting it as a
> build option in numpy itself too much effort for too little gain, not
> to mention the heuristics necessary to decide when to switch from the
> CPU implementations to the GPU implementations. The best value/cost
> ratio can probably be achieved by a separate package that wraps these
> routines with functions signature-compatible with those in
> numpy.linalg.

Apropos of the above, I've written a BSD-licensed scikit [1] that
contains a variety of numpy/scipy-like functions that can take
advantage of NVIDIA GPUs via PyCUDA and ctypes interfaces to several
libraries provided by NVIDIA and others. I would like to add
interfaces to the MAGMA library at some point, but I'm afraid I don't
have time to do so in the immediate future (I'm currently adding
support for some library interface changes introduced in CUDA 4.0). If
anyone is interested in contributing MAGMA wrappers to the scikit, I
would be happy to consider incorporating them.

						L.G.

[1] http://github.com/lebedov/scikits.cuda 


More information about the NumPy-Discussion mailing list