Convolution routine not thread safe

[maahn]

For the radar simulator and the radar moments estimator, we cannot release the GIL as for gaseous absorption (35f59b0), because FFTW is not thread safe. See http://www.fftw.org/fftw3_doc/Thread-safety.html

Workaround: Use subprocesses instead of threads http://dask.pydata.org/en/latest/setup/single-machine.html

[DaveOri]

Only the "planning" part of fftw is not thread safe
For the future it might be beneficial to explore the possibility to plan fft outside the routine for the entire dataset.
If the data is already available before running convolution it is possible to run fftw_plan in advance and execute threaded iterations of the same plan on different data
http://www.fftw.org/fftw3_doc/New_002darray-Execute-Functions.html#New_002darray-Execute-Functions
http://www.fftw.org/fftw3_doc/Advanced-Complex-DFTs.html#Advanced-Complex-DFTs

[maahn]

As of know, the plan is different every time you call fftw

       1      140323642926016
       2      140323642941984
       3      140323642890512
       1      140323640285792
       2      140323640352688
       3      140323600266464
       1      140323642814960
       2      140323642962464
       3      140323640104624
       1      140323640460960
       2      140323640470720
       3      140323640347520
       1      140323642965184
       2      140323642969216
       3      140323642972000
       1      140323642934560
       2      140323642973904
       3      140323642980288
       1      140323642973840
       2      140323642925984
       3      140323642971488
       1      140323644961808
       2      140323638110368
       3      140323645020800
       1      140323645013840
       2      140323644944544
       3      140323645013952
       1      140323642934336
       2      140323642989344
       3      140323640787584
       1      140323645010944
       2      140323644980176
       3      140323644982176
       1      140323642986592
       2      140323642996768
       3      140323642994816

This is the result of the three planing steps for convoluting 12 spectra. Since it is different for every call we would need to store all of them in advance... If a general plan would work we could indeed estimate it quiet easily in advance. E.g. by using a Fortran module with a separate plan subroutine and storing the plan value as global variable.