Use sparse matrix approach for neighbor window sum

[maxnoe]

Use same neighbor approach like number_of_islands

This results in a speedup from:

166.10 ± 66.14 µs

to

69.12 ± 4.01 µs

using:

import numpy as np
from ctapipe.instrument import TelescopeDescription, SubarrayDescription
import astropy.units as u
from ctapipe.image.extractor import NeighborPeakWindowSum
from timeit import repeat


tel = TelescopeDescription.from_name('LST', 'LSTCam')
geom = tel.camera.geometry
subarray = SubarrayDescription(
    'perf',
    tel_positions={1: [0, 0, 0] * u.m},
    tel_descriptions={1: tel},
)


extractor = NeighborPeakWindowSum(subarray)

waveforms = np.random.normal(size=geom.n_pixels).astype('float32')
selected_gain_channels = np.zeros(geom.n_pixels, dtype=int)


# number_of_islands(geom, mask)
code = '''
extractor(waveforms, 1, selected_gain_channels)
'''

exec(code)

N = 200
times = np.array(repeat(code, number=N, repeat=10, globals=globals()))
print(f'{np.mean(times / N) * 1e6:.2f} ± {np.std(times / N) * 1e6:.2f} µs')

[codecov]

Codecov Report

Merging #1398 into master will decrease coverage by 0.04%.
The diff coverage is 60.00%.

@@            Coverage Diff             @@
##           master    #1398      +/-   ##
==========================================
- Coverage   90.90%   90.86%   -0.05%     
==========================================
  Files         184      184              
  Lines       12715    12735      +20     
==========================================
+ Hits        11559    11572      +13     
- Misses       1156     1163       +7     

Impacted Files	Coverage Δ
ctapipe/instrument/camera/geometry.py	90.24% <ø> (-0.16%)	⬇️
ctapipe/tools/tests/test_tools.py	97.40% <ø> (-0.04%)	⬇️
ctapipe/image/extractor.py	85.37% <14.28%> (-1.23%)	⬇️
ctapipe/image/muon/intensity_fitter.py	90.54% <20.00%> (-2.47%)	⬇️
ctapipe/containers.py	100.00% <100.00%> (ø)
ctapipe/image/tests/test_extractor.py	100.00% <100.00%> (ø)
ctapipe/io/simteleventsource.py	94.64% <100.00%> (ø)
ctapipe/tools/stage1.py	85.77% <100.00%> (-0.06%)	⬇️
ctapipe/core/tests/test_component.py	97.71% <0.00%> (+0.16%)	⬆️
ctapipe/core/component.py	96.72% <0.00%> (+0.23%)	⬆️

---

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 597c1dc...834097c. Read the comment docs.

[watsonjj]

Nice. From trying this PR it all seems to be working correctly and 2x faster than before

[maxnoe]

Side point: numba jitted functions do not contribute to coverage, we might want to create one instance of the tests with NUMBA_DISABLE_JIT=1 for the coverage report.