PF
2.5% branch mis prediction. (0.1% is already huge)
Attila: Please connect with Joana, since she is a specialist on vectorisation. Loading the registers is a large overhead, look into it to have any impact.
Paul: autovectorisation often does not work
Deadline for midcycle review is on Tuesday.
Luthien
Did software tutorial. Most things were already known, still some new things. Especially systematic handling was new and useful.
Looked into efficiency after material layers. Low eta has better efficiency. (10 GeV muons)
Noemi: need to do 2 GeV
Neza
Prepared slides for Tuesday and incorporated some comments already.
ttbar and single muon plots are remade
Attila
Busy with offline IO.
The NGT cluster is nearing its start of life. Announcement coming in the next days for general use. Also CMS and ATLAS presentations are planned. Useful for everyone, who wants to use a few H100s.
github-ci-manager: current setup: application for github. Implementation is not clear, for more details refere to Riccardo.
Paul is interested in using the implementation, if it fits.
Paul
Geometry construction for CMS revealed some painpoints. Were fixable.
Internal event generation. hepMC read-in is a bit faster than simulation.
Tried to checkup with spack's infrastructure changes. Not finished yet.
Compilers are not a build dependecy anymore. Now a bit better.
Spack has many mac-issues, but Paul fixes many of these.
Attila: What happens when root updates their configuration? -> Spack has a way to update it.
We need to think of an offline only mode, because some users have a offline-only requirement.
2 incomming issues with IC community. Needs a proper look. Getting feed back earlier, would help. Now they hand back half a year later.
Stephen
KF: Branching factor already reduced, now looking at the physics performance for different settings.
Attila: Do we need an extra code pass for branching factor (BF) = 1? -> Probbaly not yet.
Code to output the clustering data.
Dimitra
Geometry validation and checking MDTs. Bending in the same detector volume. Maybe during the propagation some surface for intersection are missed. (about 0.1% of the events)
Andreas Stefl: Put more in the navigation state and do a relaxed bounds check? Similar to the inner detector. Only do a boundary check, if one is targeted.
Andreas Salzburger: Adapted boundary check, because the distance is already known before the check.
Joana: detray has already a similar solution, have a look into that one.
Start to write navigation for Gen3
Analysis work.
Alex
GX2F:
refitting takes shape. A few differences to KF.
Fixed the efficincy problem https://github.com/acts-project/acts/pull/4167
Benjamin
Midcycle review in 2 weeks, working towards this. Fixing edge cases in GNN workflow for single muon rconstruction.
ONNX runtimes inferences. Load from device is supported. Maybe we can compile the whole GNN workflow with only ONNX and CUDA as dependencies. Only a bit of libtorch still in. Maybe add it to the ATHENA build soon?
Andreas Salzburger
Clustering: 2D/3D-problem in the ODD. For some cutoffs the pathlength needs to be known for some 3D cases. Therefore, the angle of the track is needed. Projection worked, but for weighting the real path is needed. Now only doing 3D.
Joana
Validating navigation in traccc. Energyloss and scattering. 10GeV: surfaces can be recovered, but many surfaces are checked for hits without having anything on them.
Material mapping on the toy detector works better. Maybe because of a lot of scattering?
Cross-product sped up.
Improved vectorisation by reordering multiplications from 6x8 to 8x6.
Stephen: flag vectorise-verbose tells why things were not vectorised.
autovectorisation is good at choosing the most efficient register.
Andreas Stefl
ODD perfromance
tarckfitting and finding. Pulls are up to 20% in range. Inefficencies for finding single muons.
Tried to map material on the back with Noemi, but this didn't improve the situation. Shifted too much in one direction.
Andreas Salzburger: CMS used geani (geant material). We could invest the time to map that in the navigator interface.
vertexing is missing.
