nomkl & pyfftw on Windows

[i3v]

conda install nomkl and conda install -c conda-forge pyfftw currently fail for me on Windows with PackageNotFoundError. But at least some stuff from the "part1" seem to work. Are they optional actually?

[adler-j]

• nomkl is optional on windows.
• pyfftw can be replaced by the numpy fft for all exercises, however if you want to do three dimensional stuff it might become somewhat slow. You should be able to install it using pip install pyfftw (see here)

[kohr-h]

Actually, nomkl fails on Windows, and it seems that the conda package has an issue with a false nomkl dependency in it. We need to fix that.

[i3v]

Indeed, pyfftw seem to be a good choice. It's about 5x faster than numpy (default version from conda) on my machine (for single-axis fft on 512x512x512 array). But, just-in-case: plain pip install pyfftw still may not work on Windows. Manually downloading correct version do works for me.

[kohr-h]

You can also use conda install -c conda-forge fftw pyfftw. I don't know if they have the packages in the windows channel, but it's easier to do it manually.

[i3v]

Sadly, conda install -c conda-forge fftw pyfftw doesn't work. It looks like only osx and linux-64 versions are currently available in "conda-forge" channel.

There is one channel that got recent pyfftw version (v0.10.4, they say) for windows - "ccpi". I've decided to test it, and... ran into "PyFFTW < 0.10.4" warning. Somehow, pyfftw.__version__ gives 0.10.3.dev0+e827cb5. Thus, installing PyFFTW from that channel doesn't look like a good idea either.

Thus, it looks like a quick-and-dirty workaround with manually downloading whl from bintray is the simplest approach for now.

[kohr-h]

There is one channel that got recent pyfftw version (v0.10.4, they say) for windows - "ccpi". I've decided to test it, and... ran into "PyFFTW < 0.10.4" warning. Somehow, pyfftw.version gives 0.10.3.dev0+e827cb5. Thus, installing PyFFTW from that channel doesn't look like a good idea either.

Don't worry about that. The pyfftw devs messed up their versioning, and the 0.10.3 release has the exact same version string as the 0.10.4 release, which makes them indistinguishable at runtime. See also this issue.

But if you found a variant that works for you, it's all good.