Roundtable
Alex
Luthien
Lightning framework after c-tight.
Benjamin
traccc-GNN-connection: fucntion to separate input from framework. Plan to check with people on jagged vectors for returns.
Attila: Found tracks already work with jagged vectors but are planned to be reworked.
-> Maybe not touch this for now and work on the other things first.
Maybe start with a label list. This won't work when more tracks to each track-candidate will be assigned.
Plan to prepare input for summer student.
Public API-definition in ACTS are maybe not well defined.
Paul: Everything that can be used is public API. The read-the-docs is currently wrong on this.
Andreas Stefl: Currently there are also things in detail that are used publicly.
Andreas Salzburger: We need to do a clean-up in the future.
Doga
Unblinding of the data set should come next week.
Work on poster.
Dimitra
Worked on issues with geometry identifiers. Changed the order in the visiting and it didn't break anything but the assignment from the parents was not optimal. It would be better if we could apply it from the inner volumes.
Paul: Portal surfaces are shared between volumes, so there is no right way to label them correctly. Now we get numbering from parent volumes. This breaks when we have more than 256 parents. Now they get it from layer ID in has no limit anymore. Portals now point away from volumes.
Attila: Why do we need to know? -> Everything needs an ID.
Unfortunately some unit-tests break, mostly when we do hard-value checks on the ID. This can maybe changed easily.
Paul: The change seems fine, there is not a lot of existing API and no reason to do it otherwise.
Overlap removal tool. Try another cut. Yields seem a bit better than in run2. Events need to be rejected because the phase space does not match.
Stephen
traccc performance:
- change of branching factor added a huge boost.
- CUDA had a performance cut, when we changed the compiled architecture (there is real and virtual architecture. many can be chosen for compilation an it is not straight forward which is fastest). This happens only in the fitter, probably because it is the most complicated component.
New ambiguity resolution algorithm: It looks promising but physics still needs to be validated.
Compilation target issue: In KF there is a kernel fusion which seems to be a bad idea. Not sure if it is physics related or occupancy bounded.
Somehow also the forward step takes longer which is unclear why.
PF
chi2cuts for measurement association.
7-hits-cut: should work fine for muons but maybe for pions wrong.
Anton
Started implementing transformer model. Looked into flow-matching.
Paul
detray-conversion: material conversion seem s to work now. not fully extensible to new material out, still needs some work. Try out with ITk geometry soon.
How difficult to translate the ODD? -> We outsourced to 2 other people. It seems to work fine so far. TrackML detector is already converted. Ran the conversion with an even more simplified detector.
Examples framework? -> Probably we should not put any time in building this. There is already a script, but it won't be the same as Athena because it is different.
Preparing slide for the workshop. Currently not able to to show a lot of plots because most of them are labeled as "internal". Just put it there if it is really internal. Just don't put ATLAS on the plot.
Attila
EDM development.
Preparing for ISC next week. There will be discussions with intel. Maybe nvidia too. sycl tutorials coming.
Neza
Put all the files for digitization and similar into one place. Now ready for H100. Found a bug in the move track function. The fix should be now merged with athena latest.
Plan: Do first measurements tomorrow morning.
Andreas Salzburger
Started series of alignment PRs. Misalignment showcase can be then moved to examples.
Big chain of PRs necessary to avoid overlaps in names.
Plan to finish step 5 and 6 this week.
In general following how Paul did ITk misalignment in Athena. Just alignment context os replaced with an equivalent. Noe there is a way for experiments to misaligned without touching the core code.
Goal: have misalignment with just a single mechanism that replaces aligned detector elements with misaligned ones.
Andreas Stefl
Worked on seeding in ACTS. Plan to make it a bit faster. The ITk reconstruction change. Found out so far, that there are many indirections in spacepoint EDM. There is a proxy that points into a container. Then we have a pointer to the proxy. so the proxy needs to stay in memory. Not a huge performance increase by removing the indirection. But it made it easier to optimise smaller components. Now planned to rearrange some ifs. Looks like 15% faster in ACTS but Athena did not improve yet.
I looks like the speed depends mostly o the configuration andnless on the code itself. Potentially the easiest way out would be to optimise these cuts.
PF: Tried vectorised code instead of ifs. Brought in a mask path. Maybe could be interesting to have a look into that. But if-breaks have been still faster than bringing all the code along.
CPU will figure out which branches are hot and avoids them. We can help by presorting, but in the end, when we hit them we will hit them. We also did most of these improvements already.
Tomorrow, we will discuss on how to go on. Then move back to Athena.
