Hi, I did a few simple timing tests (see comment in PR), which suggests it is hardly worth having the cache. Indeed, if one really worries about speed, one should probably use pyFFTW (scipy.fft is a bit faster too, but at least for me the way real FFT values are stored is just too inconvenient). So, my suggestion would be to do away with the cache altogether. If we do keep it, I think the approach in the PR is nice, but I would advocate setting both a size and number limit (e.g., by default no more than 8 entries or so, which should cover most repetitive use cases). All the best, Marten p.s. I do like having a quick fft routine in numpy. My main gripe is that it always casts to float64/complex128 rather than sticking with the input dtype. Hope to get around to making a PR for that...

On 30/05/16 10:07, Joseph Martinot-Lagarde wrote:

Marten van Kerkwijk writes:

...

I did a few simple timing tests (see comment in PR), which suggests it is hardly worth having the cache. Indeed, if one really worries about speed, one should probably use pyFFTW (scipy.fft is a bit faster too, but at least for me the way real FFT values are stored is just too inconvenient). So, my suggestion would be to do away with the cache altogether.

I added a slightly more comprehensive benchmark to the PR. Please have a look. It tests the total time for 100 FFTs with and without cache. It is over 30 percent faster with cache which it totally worth it in my opinion as repeated FFTs of the same size are a very common use case. Also many people will not have enough knowledge to use FFTW or some other FFT implementation.

Joseph Martinot-Lagarde wrote:

The problem with FFTW is that its license is more restrictive (GPL), and because of this may not be suitable everywhere numpy.fft is.

A lot of us use NumPy linked with MKL or Accelerate, both of which have some really nifty FFTs. And the license issue is hardly any worse than linking with them for BLAS and LAPACK, which we do anyway. We could extend numpy.fft to use MKL or Accelerate when they are available. Sturla

Am 31.05.2016 um 23:36 schrieb Sturla Molden :

Joseph Martinot-Lagarde wrote:

...

The problem with FFTW is that its license is more restrictive (GPL), and because of this may not be suitable everywhere numpy.fft is.

A lot of us use NumPy linked with MKL or Accelerate, both of which have some really nifty FFTs. And the license issue is hardly any worse than linking with them for BLAS and LAPACK, which we do anyway. We could extend numpy.fft to use MKL or Accelerate when they are available.

It seems the anaconda numpy binaries do already use MKL for fft: In [2]: np.fft.using_mklfft Out[2]: True Is this based on a proprietary patch of numpy? Gregor

Sturla

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org https://mail.scipy.org/mailman/listinfo/numpy-discussion

On Tue, May 31, 2016 at 10:36 PM, Sturla Molden wrote:

Joseph Martinot-Lagarde wrote:

...

The problem with FFTW is that its license is more restrictive (GPL), and because of this may not be suitable everywhere numpy.fft is.

A lot of us use NumPy linked with MKL or Accelerate, both of which have some really nifty FFTs. And the license issue is hardly any worse than linking with them for BLAS and LAPACK, which we do anyway. We could extend numpy.fft to use MKL or Accelerate when they are available.

That's what we used to do in scipy, but it was a PITA to maintain. Contrary to blas/lapack, fft does not have a standard API, hence exposing a consistent API in python, including data layout involved quite a bit of work. It is better to expose those through 3rd party APIs. David

Sturla

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org https://mail.scipy.org/mailman/listinfo/numpy-discussion

You can backport the pure Python version of lru_cache for Python 2 (or vendor the backport done here: https://pypi.python.org/pypi/backports.functools_lru_cache/). The advantage is that lru_cache is C-accelerated in Python 3.5 and upwards...

That's a pretty big back-port. The speed also does not matter for this particular use-case: Time for the actual FFT will dominate by far. The lru_cache decorator can furthermore only limit the cache size by item count and not size in memory as the proposed solution does. I think the downsides outweigh the advantages of being able to use functionality from the stdlib.