Seth R. Johnson

From the LZ GDML file we have: Surface Name | Surface Type | Finish -- | -- | -- diffuseReflectorOpSurface | dielectric_dielectric | groundfrontpainted gasXeSteelSurface | dielectric_metal | polished aluminumQuartzSurface...

@esseivaju 's runs on perlmutter confirm this is an ORANGE issue: 

OK, the difference was from changing from the pregenerated `.org.json` to the built-in GDML generator... not clear why

@esseivaju has been doing some profiling of the along-step kernels. Long story short: the T- cases ("EM flat") are doing the same number of operations but the memory access are...

BTW I found this using `git bisect`: very glad (@pcanal ) that we use a squash merge strategy 😉 ```console $ git bisect log # bad: [21940769b1bfba623289ad1ebb3fe30b81d1b5e4] Update frontier toolchain...

Yeah the ugly length-111 cell is the "background" volume (world).

It seems that the only real slowdown is because of the less compact data layout for the temporary states: I don't think the background cell is even being reached in...

The plan discussed with @elliottbiondo: 1. Refactor the background intersect from "all surfaces then volumes" to "all volumes then surfaces", using the BIH bounding boxes as a first pass for...

@elliottbiondo Here's a great example for optimization: `LArMgr__LAr__EMB__STAC` (placed as `LArMgr__LAr__EMB__STAC0x3a8a8870`) in `atlas-full-lar.gdml` has 584734 elements in the CSG tree, which (with inlining disabled) turns into: ```python {'surface_counts': {'pz': 7,...

Worked around with using #1318