use SIMD.jl directly instead of LV.jl for fast_findmin()

[Moelf]

fix #83

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 70.59%. Comparing base (eab62d1) to head (fa51402).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #84      +/-   ##
==========================================
+ Coverage   70.18%   70.59%   +0.40%     
==========================================
  Files          17       17              
  Lines        1164     1180      +16     
==========================================
+ Hits          817      833      +16     
  Misses        347      347              

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

---

🚨 Try these New Features:

• Flaky Tests Detection - Detect and resolve failed and flaky tests

[graeme-a-stewart]

I will test this, but initial indications are that the performance on Apple silicon is rather bad

[Moelf]

hmm, I don't have access to Apple Silicon, but luckily we don't run JetReco on aarch64 in HEP yet

[graeme-a-stewart]

Yeah, but it's where I develop! And people testing FCC workflows, for example, do use their Mac laptops...

[graeme-a-stewart]

Yeah, it's quite a significant regression on Apple silicon. On my M2...

main

 ~/.julia/dev/JetReconstruction/examples/ [main*] julia --project instrumented-jetreco.jl --algorithm=AntiKt -R 0.4 ../test/data/events.pp13TeV.hepmc3.gz -m 32
Processed 100 events 32 times
Average time per event 170.6484634375 ± 4.057617987022607 μs
Lowest time per event 166.17292 μs

exorcise_lv

 ~/.julia/dev/JetReconstruction/examples/ [exorcise_lv] julia --project instrumented-jetreco.jl --algorithm=AntiKt -R 0.4 ../test/data/events.pp13TeV.hepmc3.gz -m 32
Processed 100 events 32 times
Average time per event 201.44234250000002 ± 3.8225514883321954 μs
Lowest time per event 197.07167 μs

I need to test also on x86 and follow-up with further benchmarks of suggestions which came up in the Discourse thread.

[graeme-a-stewart]

Benchmarks for x86_64, AMD Ryzen 7 5700G.

main

pc-sft-fa-0a :: dev/JetReconstruction/examples ‹main› » julia --project instrumented-jetreco.jl --algorithm=AntiKt -R 0.4 ../test/data/events.pp13TeV.hepmc3.gz -m 32
Processed 100 events 32 times
Average time per event 184.5104475 ± 6.358035705629633 μs
Lowest time per event 176.97703 μs

exorcise_lv

pc-sft-fa-0a :: dev/JetReconstruction/examples ‹exorcise_lv› » julia --project instrumented-jetreco.jl --algorithm=AntiKt -R 0.4 ../test/data/events.pp13TeV.hepmc3.gz -m 32
Processed 100 events 32 times
Average time per event 184.13992875 ± 6.5172848018939655 μs
Lowest time per event 177.66260999999997 μs

So it's really doing a good job on x86. We just have to find a way to make it also work well on Apple silicon.

[Moelf]

do you ever have NaN in this function? because:

julia> @fastmath foldl(min, [1.0, NaN, 0.5])
0.5

julia> foldl(min, [1.0, NaN, 0.5])
NaN

[graeme-a-stewart]

No, there can't be a NaN there. I think if you can have NaN then fast math's assumptions are violated and bad things happen.

[Moelf]

ok yeah if you don't have NaN and don't care about -0.0 vs. 0.0 I think you can use ~fastmath

[graeme-a-stewart]

ok yeah if you don't have NaN and don't care about -0.0 vs. 0.0 I think you can use ~fastmath

There shouldn't be any zeros at all, so -0.0 vs. +0.0 should be moot!

[Moelf]

honestly, I might need to buy a Mac Mini just so I can test aarch64 performance....

[graeme-a-stewart]

honestly, I might need to buy a Mac Mini just so I can test aarch64 performance....

They are quite sweet little machines (especially now they have M4s) - I am thinking about getting one myself!

[graeme-a-stewart]

So I was thinking that if we do not find a generic solution, we can tolerate two ways to implement fast_findmin, switching on Sys.KERNEL (specifically isapple() and friends).

Though it would be good to understand if it's OS we should switch on or CPU arch - @aoanla would be able to help us run a few tests.

[carstenbauer]

honestly, I might need to buy a Mac Mini just so I can test aarch64 performance....

Poor man's workaround: ssh to a aarch64 MacOS GitHub Runner. ;)

[Moelf]

btw, here's what LV.jl is doing:
https://gist.github.com/Moelf/7432073603f10d7718ef82552d5362dd