Optimize com_psf

[cgarling]

If you copy-paste the revised function definitions from this PR into your REPL you can run the following comparison which will show the new implementation is ~3x faster and allocation free EDIT: If you have to collect the input (as we do because of how Photometry.jl passes us the cutout) then it will be 2 allocations and adds ~100 ns to the base case speed. I chose to do com_psf(T::Type{<:AbstractFloat}, img_ap, rel_thresh) because Float64 is not really any slower on my testing so this way it makes it easier for us to switch to Float64 in the future if we every wanted to.

using PSFModels: gaussian
using BenchmarkTools
import Astroalign

const x = 1:20
const y = 1:20
const T = Float32
model(x, y, amp) = gaussian(T, 4, 4; x, y, fwhm=3) * amp
data = model.(x, y', 10) .+ T(0.1) * randn(T, length(x), length(y))
result_orig = Astroalign.com_psf(data; rel_thresh=0.1f0)
result_revised = com_psf(data; rel_thresh=0.1f0)
println("Original COM: ", (result_orig.psf_params.x, result_orig.psf_params.y))
println("Revised COM: ", (result_revised.psf_params.x, result_revised.psf_params.y))
println("Original FWHM: ", result_orig.psf_params.fwhm)
println("Revised FWHM: ", result_revised.psf_params.fwhm)
println("Original benchmark: ")
display(@benchmark Astroalign.com_psf($data; rel_thresh=$0.1f0))
println("Revised benchmark: ")
display(@benchmark com_psf($data; rel_thresh=$0.1f0))

My result:

EDIT: Because Photometry.jl will pass us an object from Transducers.jl we have to call collect, giving us 2 allocations and runtime +100 ns over not collecting (if we had a pure matrix input).

julia> println("Original COM: ", (result_orig.psf_params.x, result_orig.psf_params.y))
Original COM: (3.9974918f0, 3.9858212f0)

julia> println("Revised COM: ", (result_revised.psf_params.x, result_revised.psf_params.y))
Revised COM: (3.9974945f0, 3.9858336f0)

julia> println("Original FWHM: ", result_orig.psf_params.fwhm)
Original FWHM: (2.31581660320762, 2.3766457470426725)

julia> println("Revised FWHM: ", result_revised.psf_params.fwhm)
Revised FWHM: (2.3164616f0, 2.3772223f0)

julia> display(@benchmark Astroalign.com_psf($data; rel_thresh=$0.1f0))
BenchmarkTools.Trial: 10000 samples with 10 evaluations per sample.
 Range (min … max):  1.408 μs …  1.050 ms  ┊ GC (min … max):  0.00% … 99.19%
 Time  (median):     1.544 μs              ┊ GC (median):     0.00%
 Time  (mean ± σ):   2.372 μs ± 15.628 μs  ┊ GC (mean ± σ):  21.39% ±  4.62%

  ▆█▇▅▃▁▁▁                ▃▄▃▂                               ▂
  █████████▇▇▆▇▇▇▇▆▇▆▆▆▇▆██████▆▆▆▅▄▅▄▅▅▃▅▆███▇▅▄▄▄▁▁▄▁▆▅▇██ █
  1.41 μs      Histogram: log(frequency) by time     5.56 μs <

 Memory estimate: 10.41 KiB, allocs estimate: 20.

julia> println("Revised benchmark: ")
Revised benchmark:

julia> @benchmark com_psf($data; rel_thresh=$0.1f0)
BenchmarkTools.Trial: 10000 samples with 183 evaluations per sample.
 Range (min … max):  571.716 ns …   9.490 μs  ┊ GC (min … max): 0.00% … 88.05%
 Time  (median):     600.563 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   678.928 ns ± 532.693 ns  ┊ GC (mean ± σ):  8.48% ±  9.61%

  █▅▂▁▁                                                         ▁
  ██████▇▇▆▆▄▅▁▄▃▁▃▁▁▁▁▁▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▁▃▄▃▄▄▄▄▄▄▄▃▄▅ █
  572 ns        Histogram: log(frequency) by time        4.5 μs <

 Memory estimate: 1.64 KiB, allocs estimate: 2.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.49%. Comparing base (02a35cd) to head (6fc8e45).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #59      +/-   ##
==========================================
+ Coverage   99.45%   99.49%   +0.03%     
==========================================
  Files           5        5              
  Lines         185      197      +12     
==========================================
+ Hits          184      196      +12     
  Misses          1        1              

Flag	Coverage Δ
unittests	99.49% <100.00%> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.