Roundtable
Alex
Investigated pytest for lcg CI. Solved most problems (onnx), but still two missing with clang19. segfaults with 107 and empty track with 108.
Trying to figure out the whole chain why it fails and maybe catch that problem earlier in the seeding.
Doga
poster.
agdt for acts-athena build. it worked fine. different flags for addt. planning to add timing information.
Noemi: if we change the decoration, we also need to understand, are clients still looking into previously generated examples or is it just us? Also the issue that it might stay forever, if we do not change soon.
Anton
Bug hunting after training NN. there were some NaNs. Maybe found the issue but still testing.
Joana
Speeding up intersection code. No need to go into local coordinates. Intersector code is not memory bound anymore. Next step: no branching for different shapes and only coordinate systems. (rectangular and trapezoid should not branch). Annulus bounds are still factor 10 slower than the others. Hopefully can also be extended to the Jacobians.
Stephen also merged sympy stuff which made a huge speedup on the GPU.
Andreas Stefl: Maybe add a dummy type for r-phi-z? -> could add a lot of complexity but might be the solution too.
Noemi
ACTS tracking chain in athena. one thing mentioned was change in grid structure for ITk. Bumpy structure in input parameters responsible for missing hits in inner section.
First checked and counts were wrong but locally it was seeming fine. Paul will do rebase today.
Next step: check with PF on compatibility of cuts.
Paul
Frustum navigation can be used to see which portals we are going too. If we leave it due to curvature, there is no way to recover. There is a way to store arbitrary date on the track and can then be used later.
Gen1 to Gen3 migration: Different copy methods for track containers (shallow, full, ...). Needs to taken care of in API because semi-public.
Andreas Salzburger
Maybe hackathon in december to finalise geometry.
First tests to run full python examples. Still will take a bit until it is fully done. Maybe show tomorrow. Gen2 will also not anymore be supported in python.
Carlo
Doublets seeding. ttbar pu200 generates millions of doublets but only 15k triples. started to kill many doublets before.
Suggestion to get rid of space points. We were too optimistic when discussing if it should go into athena. Estimated gain would be 1% (10% on data preparation).
Andreas Stefl: Probably easier to get rid of xAOD space points
Noemi: if we spend too much time on the formation because of xAOD, get rid of that and convert later.
Neza
summer school last week
tried some truth particles for acts standalone.
xAOD cluster container: the reported number in C100, does it include the writing? -> likely included.
Andreas Stefl
Truth tracking with the KF: where is the lower performance bound from track finding? On threadripper we are at 100 us/track. gain matrix smoother or nbf smoother gives 20% more. We are quite generous on finding surfaces. Trying to make the grid-cell smaller than the surface itself. Does not work with current setup. Trying surface to grid and grid to surface. with the neighborhood lookup it should just populate.
Bug in reverse KF, where smoother is not used. Trackcontainer was always copied, could be solved by a single character in the code and sped up 100x.
