Sebastian Berg wrote:

On Mo, 2014-06-02 at 07:27 -0400, Neal Becker wrote:

...

I just d/l numpy-1.8.1 and try to build. I uncomment:

[fftw] libraries = fftw3

This is fedora 20. fftw3 (and devel) is installed as fftw.

I see nothing written to stderr during the build that has any reference to fftw.

I don't know the details, but this is not supported currently. It did work for some old versions of numpy I think.

- Sebastian

If fftw is not supported anymore, then the comments should be removed from site.cfg.example

On Mon, Jun 2, 2014 at 2:26 PM, David Cournapeau wrote:

FFTW is not used anymore in neither numpy or scipy (has not been for several years). If you want to use fftw with numpy, there are 3rd party extensions to do it, like pyfftw

If it was once supported in numpy, does someone remember the reason why it was removed? Was it the license or a technical reason? In the long term I would like to add fftw support to numpy, at least for the source distribution where the licensing is no concern. Some testing I did a while ago showed fftpack and fftw have different numerical behaviour and depending on the size fftpack can be faster due to fftws overheads when not using pregenerated plans, so to avoid breaking user code the implementation should probably be selectable at runtime by the user instead of hardwiring the implementation at compile time. But I don't know if I'll ever find the time to do so :(

On 02.06.2014 14:26, David Cournapeau wrote:

FFTW is not used anymore in neither numpy or scipy (has not been for several years). If you want to use fftw with numpy, there are 3rd party extensions to do it, like pyfftw

If you feel pyfftw bothers you with too many FFTW details, you may try something like https://github.com/aeberspaecher/transparent_pyfftw (be careful, it's a hack that has seen only little testing). Alex

On 05.06.2014 11:13, Daπid wrote:

pyFFTW provides a drop-in replacement for Numpy and Scipy's fftw:

https://hgomersall.github.io/pyFFTW/pyfftw/interfaces/interfaces.html

Sure. But if you want use multi-threading and the wisdom mechanisms, you have to take care of it yourself. You didn't have to with anfft.

You can still set the number of threads and other advanced parameters, but you can just ignore them and view it as a very simple library.

Sure.

Does your wrapper set these to reasonable values? If not, I am missing completely the point.

You have to decide on the number of threads when you configure tfftw. Anyway it is well possible that tfftw is completely useless for you.

I am starting to use pyFFTW, and maybe I can help you test tfftw.

Contributions or bug reports are welcome! Alex