[AstroPy] Fast(er) Deeming Periodograms from python via OpenCL

Ewald Zietsman ewald.zietsman at gmail.com
Thu Sep 25 09:37:06 EDT 2014


Hi All,

I'm working with some Kepler data from which I need to extract amplitudes
and frequencies. The dataset is really long and hence the required DFT
takes an awfully long time to calculate (16 hours per DFT if I run 4
threads on my Macbook Pro). I'm getting around this problem by running it
in parallel on my Graphics Processing Unit (on my Linux Box) via PyopenCL.
I did a blog post about the kind of performance increase I get on my setup
doing this. You can read it here:

http://ezietsman.github.io/python/2014/09/06/parallel-python-on-a-gpu-with-opencl/

TL;DR; My Macbook will take around 16 hours to do one DFT which my will
take my GPU about 50 minutes.

I'm also working on a smallish library that will allow me to find and
prewhiten the lightcurve using robust sinusoid fitting after finding
amplitudes/periods using DFTs. That repo is still an infant and will change
quickly but it is here:

https://github.com/ezietsman/seismo

It at least shows (in the tests/ folder) how to run my DFT code. There are
3 DFT implementations at the moment, all three using exactly the same slow
algorithm, numpy serial, numpy/fortran/openmp parallel and opencl parallel.

I could work towards getting the numpy serial version into astropy somehow,
it there is interest?

Cheers!
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/astropy/attachments/20140925/1abdfda3/attachment.html>


More information about the AstroPy mailing list