[SciPy-Dev] playing with intel ipp (signal processing) library

[Robert Kern]

On Thu, May 2, 2013 at 3:32 PM, Neal Becker <ndbecker2 at gmail.com> wrote:
> I spent a little time wrapping some of intel ipp signal processing library and
> benchmarking against my hand-written c++ code.  So far I've tried some of the
> FIR filtering (single and multi-rate), and correlation.
>
> I was surprised to find that the ipp components were 3-10 times the speed of my
> c++ code.
>
> Perhaps others might want to look at using these components.  Of course, they
> are not free - but this is in the same class as building numpy/scipy using MKL.

It's not quite the same class as building numpy/scipy with MKL. The
parts of the MKL that numpy/scipy use are the well-standardized,
vendor-neutral BLAS/LAPACK interfaces. We can make use of the MKL
mostly by linking unmodified numpy/scipy sources with the MKL. But if
you are just talking about whether individuals might make the decision
to use it in their own application code, yeah, the decision is about
the same.

FWIW, AMD has an Apache-licensed competitor to the IPP that seems to
closely follow the IPP API with mostly just name changes:

  http://framewave.sourceforge.net/

I had a devil of a time getting it built, though (i.e. I didn't
succeed), so I cannot comment much beyond that.

--
Robert Kern