Roundtable
Benjamin
CMake integration of TensorRT. New flags that setup of covfie works automatically. The goal was to make the change as minimal as possible. Just a small macro.
(Not everyone is sure, if this is the best way to do it. -> Will have a second look into exact implementation.)
Can stripspace points in traccc mergd already, at elast a preliminary version?
Stephen: ODD creates accidential space points because the description seems to be wrong.
Attila: Track candidates are not yet array of indeces but copies of measurements.
Noemi
ACTS-ITk side. Finally merged MR to get it more consistant.
Found bug in Athena calibrators. Now results are consistent with expectation.
Anton
Plans some discussion with supervisor.
Doga
Prepared files for summer student. Goal is to prepare input-nTuples in the mask former training. nTuples vontain truth tracks, truth jets, and their connection. Will write new algorithms. Start with separate algorithms and have one that connects them.
Truth jets in ACTS can be described with fastJet to do the jet-clustering. No existing jet class, planning to write one.
Matching them with deltaR (the euclidean distance).
FastJet-core has a restricitve license, not sp easy to connect it directly.
Stephen
traccc-seeding. Currently the biggest performance problem. Discussed on spacepoint confirmation refinement. Triplets lose a tiny bit of performance. Trying to write analysis tool to figure out why they are missing.
Dimitra
Gen3 work, implemented navigation policy. Started from beginning, because gen1 was not a great starting point. Rewrote policy with index surface for gen2. Not all gen2 methods are useful for gen3, but at least not harmful. Needs some discussion with andi to not break anything in gen3.
Already getting expected candidates from the grid.
Upcoming muon workshop tomorrow.
In gen2 we assigned more fields (spread up like a tree). Each identifier should be unique. Worked in gen2 because we modified manually, gen3 should handle this over the blueprint. Needs some discussion with Paul on how to create blueprint nodes.
Suggestion: try first with manually assigned gemetry modifier.
Plan to revisit the seeding algorithm.
Alex
Discussions with Johannes on GX2F.
Cleaning through old PRs.
Plans: University work, discussions with Benjamin, writing.
Attila
Admin stuff: Benjamin's machine will stay here, even when communication was confusing. WIll also be great for summer student work. Should we put them on A100? -> Maybe good for development work but large scale training should be outsourced to not block development for a few days.
traccc development. First implementation for output as structured arrays. Right now each track has own copy of measurement. This would allow for inplace calibration. But indeces to measurements would be more memory efficient.
Andreas Stefl
Started to look into material map performance in ACTS. Losing single muons because of not directly accounting for material interaction. In general looks good but fine grained, Geant4 is much better.
Seeding: Compute performance lags. External space points EDM. In seeding we use pointers to externals. There is at least one indirection that could be removed. Initial tests show an improvement of 5%. No virtual calls at least. Big aim would be to move the seeding code to core.
