Refactor Reactant microphysics tests + test P3 compilation

[dkytezab]

No description provided.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[giordano]

I reverted the changes to the source code, but this is failing to raise a function (was failing before also before my change), probably needs more work on the Reactant side?

[Pangoraw]

I think the raising issue is with this function that iterates over a tuple of offset arrays. i rewrote it in a similar manner to zero_orphaned_numbers! so it becomes unrolled.

[giordano]

I guess this change is the cause of the huge performance drop in Reactant benchmarks. Edit: maybe not, after all.

[numterra-bot]

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'Breeze.jl Benchmarks'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.10.

Benchmark suite	Current: 028de58	Previous: 82d9622	Ratio
ModelTendency; Grid: 256x256x128/Advection: WENO7/NVIDIA L4/F32 reactant raise=true	131183253.64778502 points/s	693851521.5735954 points/s	5.29
ModelTendency; Grid: 256x256x128/Advection: WENO9/NVIDIA L4/F32 reactant raise=true	18996418.14392214 points/s	52738861.65621344 points/s	2.78
ScalarTendency; Grid: 256x256x128/Advection: WENO5/NVIDIA L4/BF16 reactant raise=true	10836174802.681719 points/s	11987613125.97085 points/s	1.11
ScalarTendency; Grid: 256x256x128/Advection: WENO7/NVIDIA L4/BF16 reactant raise=true	5972524675.461524 points/s	6953248841.623632 points/s	1.16
ScalarTendency; Grid: 256x256x128/Advection: WENO9/NVIDIA L4/F32 reactant raise=true	94645195.31787553 points/s	273872334.52941924 points/s	2.89

This comment was automatically generated by workflow using github-action-benchmark.

[giordano]

Oh wow, that's a massive slowdown

[giordano]

This is good to go from my point of view, I'll leave it to @Pangoraw and @dkytezab to decide what to do with the performance regression (merge now the tests and deal with the regression later, or wait to resolve the regression before pushing this forward).

[giordano]

I should note that at the same time ScalarTendency; Grid: 256x256x128/Advection: WENO9/NVIDIA L4/BF16 reactant raise=true got a 90% speedup, recovering from the drop in #923 (but WENO9 with bfloat16 is probably an unusual combination in practice). No other benchmarks got significant improvements.

[giordano]

Uh, based on #1036 (review) the regressions may come from Reactant v0.2.289 and not changes in this PR, so this should be good to go.

[Pangoraw]

I'll investigate the regression, I don't think this was caused by the code change in this PR.

[giordano]

No, it's happening elsewhere as well, so it's probably new Reactant version. This is ready from my point of view