Numba-based StdDevUDF

[uellue]

This version is slightly about 2x faster compared to the version in #625 for the benign case of signal buffers that fit comfortably into the CPU caches.

For the malign case of signal buffers that don't fit the caches well,
it is about 10x faster since it eliminates large intermediate buffers and
calculates each pixel in one pass the result in cache-friendly blocks and CPU -friendly access patterns.

• Added test for numerical stability.
• Added citation in bibtex

Contributor Checklist:

• I have added myself to the creators.json file
• I have added/updated documentation for all user-facing changes
• I have added/updated test cases

[codecov]

Codecov Report

Merging #640 into master will increase coverage by 0.35%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #640      +/-   ##
==========================================
+ Coverage   58.13%   58.49%   +0.35%     
==========================================
  Files         210      210              
  Lines        8012     8068      +56     
  Branches     1049     1065      +16     
==========================================
+ Hits         4658     4719      +61     
+ Misses       3163     3160       -3     
+ Partials      191      189       -2     

Impacted Files	Coverage Δ
src/libertem/web/messages.py	84.48% <0.00%> (-1.73%)	⬇️
src/libertem/web/jobs.py	65.83% <0.00%> (-0.84%)	⬇️
client/src/dataset/helpers.ts	0.00% <0.00%> (ø)	⬆️
client/src/dataset/components/SERParamsForm.tsx	0.00% <0.00%> (ø)	⬆️
client/src/dataset/components/FRMS6ParamsForm.tsx	0.00% <0.00%> (ø)	⬆️
...lient/src/dataset/components/RawFileParamsForm.tsx	0.00% <0.00%> (ø)	⬆️
src/libertem/api.py	87.78% <0.00%> (+0.09%)	⬆️
src/libertem/io/dataset/base.py	89.26% <0.00%> (+0.93%)	⬆️
src/libertem/io/dataset/raw.py	92.68% <0.00%> (+8.75%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 05edd76...8dfc2ec. Read the comment docs.

[sk1p]

I added some comments - most importantly I think the docstrings, variable naming and thus readability can be improved - the general structure and implementation looks good!

[uellue]

Thx @sk1p for the feedback! This should be ready for a second round of review.

[uellue]

The Numba coverage for process_tile was not counted for some reason. It did work before, with the exception of not passing through the blocked part of the loop since the input data was too small.

[sk1p]

Well done, and impressive speed-up! With the improved doc strings and variable names, the code is much better readable. Two more suggestions added, other than those this is ready to be merged.

[sk1p]

The Numba coverage for process_tile was not counted for some reason. It did work before, with the exception of not passing through the blocked part of the loop since the input data was too small.

coverage issues are best debugged locally (I don't trust codecov... the travis job says coverage was uploaded, but apparently it was not correctly merged into the full coverage report, sigh.) - did you try tox -e numba_coverage and checking the result in the htmlcov directory?

[uellue]

did you try tox -e numba_coverage and checking the result in the htmlcov directory?

Good point! The function was actually never called since the partition did fit in a single tile. I've adapted the parameters accordingly. At least locally, the coverage is 100 % now.

[sk1p]

Perfect, thanks, merged!