Refactoring pairwise into pairwise and gemm

[jaberg]

The pairwise benchmark was contaminated by gemm-based solutions. I made a gemm benchmark dir, which tests the ability of various compilers to compete with BLAS libraries, and removed the gemm-based solutions to the pairwise benchmark. Currently the BLAS libs destroy the compiled functions, but I am curious if a readable Python base implementation that actually exposes a fast blocked algorithm could be compiled to run at a more competitive speed.

For the pyopencl stuff, I moved it to a new repo (https://github.com/jaberg/python-benchmarks-pyopencl) because I am interested in developing pyopencl auto-tunable code generators, and development along that line should not clutter up this repo.

[ogrisel]

Looks good but github says there are conflicts. Could you please solve them and force push a branch rebased on the current master?

[ogrisel]

Please add a description of the benchmark and some motivation in a the module level docstring.

[jaberg]

@ogrisel Thanks for the feedback, I think I fixed it all up.

[ogrisel]

Actually I had thought that python-benchmarks-pyopencl would be an independent project and python-benchmarks would no longer have any direct pyopencl benchmarks included. Do you really want make python-benchmarks depend on another benchmark project? I think it adding such a dependency on another benchmark repo makes it unnecessarily complex for people to run the benchmark on their own hardware.

[jaberg]

How is that different from just failing and ignoring it if
the python-benchmarks-pyopencl repo isn't installed (current behaviour?)

I'm not really sure what to do with that PyOpenCL code to be honest. One
idea is to try and merge the code generators in to Reikna or some other
project, another idea is to leave them where they are, give them more
generic interfaces, and rename the repo. I think there's room for a suite
of OpenCL code generators to support pipelines of the sort used by Theano,
scikit-learn, scikit-image.

Anyway, in terms of helping people run the benchmarks on their own
hardware, how about just printing out the necessary pip command on
ImportError?

On Mon, Aug 19, 2013 at 12:25 PM, Olivier Grisel
[email protected]:

Actually I had thought that python-benchmarks-pyopencl would be an
independent project and python-benchmarks would no longer have any direct
pyopencl benchmarks included. Do you really want make python-benchmarks
depend on another benchmark project? I think it adding such a dependency on
another benchmark repo makes it unnecessarily complex for people to run the
benchmark on their own hardware.

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/24#issuecomment-22884488
.