Fix FFT norm (pyfftw & numpy)

[leonroussel]

Fixing problem on normalization of FFT

Usual normalization is on backward FFT ( ./1 on forward & ./N on backward) but we can expect to have a non-normalize FFT (with parameter normalize="none"), this option is not implemented correctly

Solution & modifications :

• numpy : norm="backward" on FFT & norm="forward" on IFFT
• pyfftw : normalise_idft=True on FFT & normalise_idft=False on IFFT

(There is also a problem with pyfftw implementation, it does not have the expected behavior with the last version (0.13), https://github.com/pyFFTW/pyFFTW/issues/341 but the solution normally bypass this problem)

TODO

• [x] ~~requirements for numpy version ? >= 1.12 ? or different in setup.py ? (if >= 1.10, a case could be avoided in npfft.py)~~
• [ ] normalization "by hand" if numpy version do not allow normalization="none" ? or keep a warning ?
• [ ] need to be tested with both version of pyfftw 0.12 & 0.13
• [x] ~~more efficient solution than np_version in ["1.8", ...]~~

[leonroussel]

Failure on old numpy implem ? --> "backward" & "forward" keywords only for version after 1.20

================================== FAILURES =================================== ... venv_test\lib\site-packages\numpy\fft_pocketfft.py:367: in rfft if norm is not None and _unitary(norm): ... E ValueError: Invalid norm value backward, should be None or "ortho". venv_test\lib\site-packages\numpy\fft_pocketfft.py:88: ValueError

next commit fixes it

[leonroussel]

Still CI failures : ================================== FAILURES =================================== _____________________ TestMaskToolsWidget.testWithAnImage _____________________ ...

[pierrepaleo]

Thanks!

From what I see in requirements.txt and pyproject.toml, the minimum required numpy version is 1.12, not 1.8.

To simplify version comparison a little bit, you can use from pkg_resources import parse_version, eg. parse_version("1.10.2") > parse_version("1.8.0").

The latest CI failures seem related to GUI (happens from time to time)

[leonroussel]

Ok it seems the same for me, so I removed code that take into account version 1.8 & 1.9.

Ho ! from pkg_resources import parse_version was exactly what I was looking for. Thanks

[pierrepaleo]

Test fails for numpy == 1.19.5. Seems to be because not all normalization were available for this version.

Just tested on my machine:

• test_norms fails with numpy==1.19.5
• test_norms passes with numpy==1.20.0

[pierrepaleo]

Tested with pyfftw 0.12 and 0.13.