General numeric type support

[graeme-a-stewart]

Support is added for general numerical types in this PR. EEjet and PseudoJet become parametrised on T <: Real. ClusterSequence stays parametrised on J <: FourMomentum, but is it really important to pass the parametrised jet collection into the constructor otherwise performance greatly suffers! (i.e., ClusterSequence{EEjet{Float64}} not ClusterSequence{EEjet}. To support this the eltype method is defined for both jet types to return the internal type used.

The instrumented-jetreco.jl script gets a new --type option that allows a numeric type to be passed for the reconstruction. Added a quick benchmark.sh script to run all three main use cases.

Current performance tests shows that we are losing a bit in the $pp$ N2Tiled reconstruction, so this needs to be fixed before merging. Performance for Durham $e^+e^-$ and $pp$ N2Plain are the same as the current v0.4.3 release.

Testing on M2pro gives

This branch:

 ~/.julia/dev/JetReconstruction/examples/ [general-numeric-types*] ./benchmark.sh 
pp 14TeV Tiled
Lowest time per event 189.185 μs
pp 14 TeV Plain
Lowest time per event 678.68667 μs
ee H Durham
Lowest time per event 28.38375 μs

v0.4.3:

 ~/.julia/dev/JetReconstruction/examples/ [general-numeric-types] ~/tmp/benchmark.sh
pp 14TeV Tiled
Lowest time per event 176.06666 μs
pp 14 TeV Plain
Lowest time per event 677.14125 μs
ee H Durham
Lowest time per event 27.92792 μs

Closes #91

---

To do:

• Convert EEjet to T <: Real
• Convert PseudoJet to T <: Real
• Convert TiledJet to T <: Real
• Add option to examples to switch types for the internal representation
  • instrumented_jetreco.jl
  • jetreco.jl
• Ensure performance is maintained in Float64 case
  • pp Tiled
  • pp Plain
  • ee
• Add unit tests for Float32 reconstruction

[codecov]

Codecov Report

Attention: Patch coverage is 95.12195% with 2 lines in your changes missing coverage. Please review.

Project coverage is 71.30%. Comparing base (70cf5f1) to head (79036ee).

Files with missing lines	Patch %	Lines
src/EEjet.jl	83.33%	1 Missing ⚠️
src/Pseudojet.jl	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #97      +/-   ##
==========================================
- Coverage   71.52%   71.30%   -0.22%     
==========================================
  Files          17       17              
  Lines        1166     1178      +12     
==========================================
+ Hits          834      840       +6     
- Misses        332      338       +6     

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

[graeme-a-stewart]

For completeness, here are the benchmarks on my x86 Ryzen 7 5700G:

pc-sft-fa-0a :: dev/JetReconstruction/examples ‹general-numeric-types› » ./benchmark.sh
pp 14TeV Tiled
Lowest time per event 199.00421999999998 μs
pp 14 TeV Plain
Lowest time per event 506.19272000000007 μs
ee H Durham
Lowest time per event 31.21525 μs

pc-sft-fa-0a :: dev/JetReconstruction/examples ‹main› » ~/tmp/benchmark.sh
pp 14TeV Tiled
Lowest time per event 177.75058 μs
pp 14 TeV Plain
Lowest time per event 502.63006 μs
ee H Durham
Lowest time per event 31.40393 μs

So, again, the significant performance loss is pp tiled (and a small loss for pp plain).