Amanda Lund

I had run TestEm3 with Celeritas with refined energy grids (1e7 primaries, 1mm production cuts):  Default is 7 bins/decade, and the error is relative to the 112 bins/decade result.

I think they are very similar, but cubic splines have the constraints that the first and second derivatives be continuous everywhere? A downside is this would require storing the second...

Great summary @whokion! I think we can check off the tracking cut which we should already support in Celeritas (#648 let us kill tracks with energies below `LowestElectronEnergy` in along-step...

I'd forgotten we had this one... sounds good to me.

@sethrj @esseivaju I think this is ready for a look and some testing/profiling now. I haven't done much myself yet, but in some quick initial comparisons saw ~25% improvement on...

Thanks @sethrj! I'd modified this to try to reuse the parent's geo state when possible for improved performance, but got sidetracked when I added an assertion that the parent's position...

Alright, I've updated this so we're now partitioning an array of indices (rather than track initializers) which we can also use to access the parent track slot IDs to copy...

I also ran it actually (on an A100 + AMD EPYC 7532): 

It's possible the worse performance for cms+msc could still be the penalty of having to reinitialize the geometry state more often than on develop...

Yeah it definitely seems worthwhile to explore ways of speeding up that initialization.