Roundtable
Alex
nothing interesting last week on ACTS
Doga
truth particle matching. small hiccup with theta and phi resolution. started with debugging and it seem to be under control.
Paul: in truth-tracking the hit ordering is not necessarily correct, since the seeding doesn't give a good ordering. So measurements are jumping around. Also truth-based parameter estimation does not check how to get them in a sensible order. Tried solution in the current branch uses sim-hit time to order hits by distance to origin. Hits are also sometimes ordered by hypotenuses which gives wrong ordering. We don't tak into account yet when a spacepoint comes from multiple clusters.
Andreas Stefl: Have a postcondition of the simulation that the simhits are always sorted by time?
Benjamin: But this needs to be implemented by all readers to make imports and similar stable, right?
Andreas Stefl: Maybe it should be readers job to sort them on the fly.
After the temporary fix the jet-issue had been solved.
Still looking at secondary vertices and why there are so many. Plan to work on it this week.
Carlo
ACAT-plots last week. plan to discuss them today.
Doublets-creation: 7kk might be created but only 10k really needed. not sure how to tackle yet. makes us slower than Athena.
Benjamin
poster for fastML conference is almost done. time to tackle actual problems: 1-2 days for sympy multistepper in GSF. catch up also with tomohiro.
Andreas Salzburger
found out, that one of the spack libraries points to the wrong xcode picked up from the cache.
Joana
Worked on the hole-issue and tried adaptive tolerance. Essentially using the covariance. But it turned out in the exactly wrong direction.
added one hip-build to the CI. previously most of the benchmarks haven't been build in the CI.
Paul
Spent some time to fix the truth-tracking stuff mentioned before.
detray: conversion: only add the end of the configuration, a wrong input would be detected.
CI is still a bit complicated with gitlab. try to structure stuff more efficiently now.
Andreas Stefl
Plots for ACAT. Present them today and circulate on Thursday.
Had issues with resolution. Figured out that we had hits in the center. This comes from the projection onto representative surfaces. shallow angles create offsets. Instead interface surface could be intersected in direction of hit and approach layer. otherwise we would need to increase binsize in z again.
measured speed up for seeding. reorganisation of the points give a 12% speed up and gets us closer to Athena.
A few ideas on how to improve the seeding to look at less combinatorics maybe with module map or compatible doublets. Mainly looking at module centers. Could be extended to triples, but centers might not be enough for this.
Stephen
traccc: performance got better but still a long way to go.
