Create numba ufunc for sum of samples within charge extraction window

[watsonjj]

This PR replaces the extract_charge_from_peakpos_array function with sum_samples_around_peakpos. This is a numba function that more efficiently extracts the charge from the waveform within the window defined by the peak position, window width, and window size.

It is defined using the @guvectorize decorator, which creates a numpy universal function, enabling the passing of both scalars and arrays for the peak position, window width, and window shift arguments.

The way the @guvectorize works (in this case) is that you define the operation that is applied for each channel and pixel. The operation is then optimised using numba's just-in-time compilation.

I have profiled sum_samples_around_peakpos against the previous extract_charge_from_peakpos_array function. The new function provides a factor of 60 reduction in execution time (from a couple of ms, to tens of µs).

It also provides a factor of 2 reduction in execution time upon the waveforms[:, :, start:end].sum(2) operation performed in GlobalPeakWindowSum.

The execution time for the extractors before this change were:

GlobalPeakWindowSum
6.15 ms ± 57.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
LocalPeakWindowSum
12 ms ± 209 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
NeighborPeakWindowSum
14.7 ms ± 293 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

The execution time for the extractors following this change are:

GlobalPeakWindowSum
3.26 ms ± 30.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
LocalPeakWindowSum
3.28 ms ± 28.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
NeighborPeakWindowSum
5.67 ms ± 33 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

The bottleneck is now the pulse_time calculation, which I could address in a different PR.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (master@24707ab). Click here to learn what that means.
The diff coverage is 88.23%.

@@           Coverage Diff            @@
##             master   #1038   +/-   ##
========================================
  Coverage          ?   83.2%           
========================================
  Files             ?     186           
  Lines             ?   10583           
  Branches          ?       0           
========================================
  Hits              ?    8806           
  Misses            ?    1777           
  Partials          ?       0

Impacted Files	Coverage Δ
ctapipe/image/tests/test_extractor.py	100% <100%> (ø)
ctapipe/image/extractor.py	67.67% <53.84%> (ø)

---

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 24707ab...cc3b326. Read the comment docs.

[thomasarmstrong]

I have had a look over the code and have run some tests to compare these changes to the extraction methods in the current master branch. Everything seems good to me, in fact this pull request also fixes a current bug in the GlobalPeakWindowSum which causes it to fail if there is only one channel. I would suggest that the PR is approved.

(One note, there is a typo in the sum_samples_around_peak description "The ret argument is required by numpy to creae the numpy array which is")

[maxnoe]

Other than a typo looks good

[watsonjj]

@kosack @maxnoe Could you restore your review here