Roundtable
Stephen
Truthseeding in tracc. close to getting it worked. Aiming to some strategy discussions soon.
Eigenissue: There seems no single great solution with fool proof and simple.
Problem is in memory alignment. If eg something would be instantiated in the athena tracking library and then handed over to acts, there could be a mismatch.
Current solution is not comprehensive enough, that we want to go into that direction.
Andreas Stefl: Maybe drop text file in the install directory to document the current setup?
It seems to be prefered to do it in a non-header-way.
The alignment is necessary to not overallocate memory.
If we add the flag, we also should enforce it in the configuration.
Alex
GX2F:
After revisiting the mathematics, it seems, that we cannot avoid hard propagations. Since we linearise our system, we should converge in the newton-subsystem in a single iteration. Now checking the athena-implementation, how that problem is solved there. (And cleaning up there on the way)
Joana
debugging in detray. investigated segfault in toydetetector benchmarks. Smoother gives out inf-states. This already happens in the KF-forward pass, but only in single precission. Not clear yet why but probably in covariance. It could be that the single precission is less stable, but probably could also happen in double precission.
segfault: cov-elemts are inf -> navigator tries to go to inf -> segfault
How should we handle matrices with det(M) = 0.
Fixed the hole counting in the KF and saw that efficiency is bad. Started to fit only first branch. Need to check, if this also fixes the segfault.
Attila
Beomki suggested to improve the API. Attila tested something out and it could work now conveniently. Now the space point switch-over is now in a semifunctional state. Code has now a similar speed as the old one.
traccc spacepoint claim to have variances on r and z, but we hardcoded 0 to these. There is also a getter for this. Now, the variance has also an influence. But in EDM inactive.
In Athena, the values are hard-coded too.
Paul: Do we assume a different variance for all space-points? -> Attila: Should not be a big issue at runtime.
Paul
codegeneration: UV for python looks nice, easily configurable and is really fast. This could remove the particle data files when we build on the fly. Later, we might also be able to add better math documentation in the code.
Doga
Analysis.
Working on Giovannis GNN. In 2 weeks, it should be finished.
Noemi
PR on updating seeding.
Trying to insert passive layers into geometry. covariance does not have time to expand. start to change with barrel.
PF: maybe add local chi2 to the layers?
Paul: we also need intersections for all sort of intersections -> Benjamin already started with that with a triangular mesh.
PF
ACTS-presentation in Vienna.
Clusterisation: Code optimisation also improves legacy, so ratio ACTS/legacy didn't improve.
Dimitra
micromega seeding. ended up in xy-layer but grew larger for next layers.
break loops on layers before sorting in y.
test gen3 navigation on plane surfaces.
analysis work to do.
Luthien
Looking into track-less events.
Andreas Salzburger: Maybe put material logger in to see, what has been seen?
Some hits don't get reconstructed correctly.
Andreas Salzburger: Reason to have the mapping so close is that the local navigation has it easier.
Andreas Stefl: The track looks also strange. beginning is straight end then bends. Maybe there is a lot of energy loss? Should be weven recover that?
Noemi: Maybe put material on backside? -> In gen1 we cannot propagate to the backside
Andreas Salzburger
Get proto-axis in to get layerstructures into gen3.
Andreas Stefl
Parameter changes (inward first vs outward first) changes also the estimation. Bottom spacepoint does not need to be on the surface. A few cases, where we won't end up on the surface. This happens only on top and bottom of local position. We might need extended bounds to tackle this.
Andreas Salzburger: Maybe a magnetic field issue?
