Fixing API change in pixell 0.19.0

[xgarrido]

pixell is broken with the latest version of pillow (see #65 (comment))

[xzackli]

@xgarrido should we bound pixell versions to < 0.18?

[xgarrido]

Actually this branch is for trying to use pixell latest version 0.19.0 and thus fix the unit tests. There is a small change in the pixell API but the tests remains broken.

[xgarrido]

@xzackli in the current master branch we now downgrade pixell to fix the unit test. We will work on new pixell version later

[codecov-commenter]

Codecov Report

Merging #69 (398bdf4) into master (1d5d873) will not change coverage.
The diff coverage is 60.00%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #69   +/-   ##
=======================================
  Coverage   33.62%   33.62%           
=======================================
  Files          19       19           
  Lines        2091     2091           
=======================================
  Hits          703      703           
  Misses       1388     1388           

Files	Coverage Δ
pspy/sph_tools.py	28.91% <60.00%> (ø)

[xgarrido]

Here is a summary just to not to forget what have been done:

• the current pixell version 0.19.2 has fixed different bugs that makes the unit test on ubuntu passing without problem (and with no changes or no new generation of reference data)
• on mac os, the issue is not the tests by themself but the way ducc is installed. The error is how the threads are used and I haven't found a way (compile ducc from sources for instance) or an option (see https://gitlab.mpcdf.mpg.de/mtr/ducc/-/blob/ducc0/setup.py#L22) to make it working.

[xgarrido]

Unit tests are all passing on linux as well as on mac. When using different C++ compilers namely gcc vs. clang it raises an error due to the conversion of OMP_NUM_THREADS env. variable into long int by ducc. Basically this line in pixell set the DUCC0_NUM_THREADS variable to OMP_NUM_THREADS and, in case the OMP_NUM_THREADS is not set the utils.getenv of pixell sets the value to None. When ducc converts the string None into long int via this function https://gitlab.mpcdf.mpg.de/mtr/ducc/-/blob/ducc0/src/ducc0/infra/threading.cc#L108 it's actually impossible : with GNU compiler on linux, no error is raised and actually the value returned by mystrtol is zero but with clang, it raises an error.

Within our unit test, we set the OMP_NUM_THREADS to zero but I think pixell should set the default returned value from utils.getenv to zero something like

utils.setenv("DUCC0_NUM_THREADS", utils.getenv("OMP_NUM_THREADS", 0), keep=True)

[amaurea]

This should be fixed in pixell 0.20.3. It's fixed a bit differently, by making setenv not try to initialize DUCC0_NUM_THREADS at all if OMP_NUM_THREADS is undefined. That's how it was supposed to work in the first place, but there was a logic error in utils.setenv.

[xgarrido]

@thibautlouis I think we are good to go and to move pspy to pixell 0.20.3 and ducc. We will update numpy when the f2py bug will be fixed (next release 1.26.1 I guess, see #78 (comment))

[xgarrido]

As expected numpy 1.26.1 fixes the issues with fortran compilation. There is no more stopper.