loading FFTW wisdom from disk is unreliable?

[mperrin]

Noticed by @douglase . When you import poppy it doesn't seem to be loading the saved wisdom as it should.

I tried changing __init__.py to make sure that fftw_load_wisdom is being called. It is. But the calculation still seems to take longer the first time after reloading.

And, even more confusingly - then it gets slower on subsequent sessions?!? Something is buggy and weird here.

(But we're moving away from FFTW anyway so this may be less important to debug.)

In [1]: import poppy

In [2]: poppy.accel_math._FFTW_INIT
Out[2]: {}

In [3]: poppy.accel_math.benchmark_fft(npix=4096)
Timing performance of FFT for 4096 x 4096, complex128, with 20 iterations
Timing performance in plain numpy:
  0.170 s
Timing performance with FFTW:
  0.619 s
Timing performance with Numexpr + FFTW:
  0.169 s
Timing performance with OpenCL:
  0.299 s
Out[3]:
{'numpy': 0.17025854700041237,
 'fftw': 0.6192405729001621,
 'numexpr': 0.1686754029011354,
 'cuda': nan,
 'opencl': 0.2994044750012108}

In [4]: poppy.utils.fftw_save_wisdom()

In [5]:
Do you really want to exit ([y]/n)? y


(astroconda) mperrin@phoenix ~  > ipython
Python 3.6.5 |Anaconda, Inc.| (default, Apr 26 2018, 08:42:37)
Type 'copyright', 'credits' or 'license' for more information
IPython 6.3.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import poppy

In [2]: poppy.accel_math.benchmark_fft(npix=4096)
Timing performance of FFT for 4096 x 4096, complex128, with 20 iterations
Timing performance in plain numpy:
  0.168 s
Timing performance with FFTW:
  0.453 s
Timing performance with Numexpr + FFTW:
  0.466 s
Timing performance with OpenCL:
  0.291 s
Out[2]:
{'numpy': 0.16818084940023253,
 'fftw': 0.4526968545498676,
 'numexpr': 0.46563783830060856,
 'cuda': nan,
 'opencl': 0.2912771725503262}