Roundtable
Paul
Tuning of the spack build. CI building spack dependencies seems to work ok. Unpacking takes a bit longer than with tar-ball (1min -> 2min). Should be fine, since the rest of the build takes 20mins.
Get ITk geometry from readout. Blueprint got merged late, but now starting to get this out.
Andreas Salzburger: Roman will come for 1 week. Use that week to work on this projected.
Good telescope construction from code so far. DD4hep might have progressed too. DD4hep-plugin has not progressed yet.
Uniformity of telescope detectors: They are quite different, but we can still share a lot of infrastructure. Especially for the tracking. Problem is our cylindrical seeding method.
X
Finished some Analysis work in the last weeks.
Plan to discuss QP in the next days.
Andreas Salzburger
Dates for US-ACTS-topical workshop are getting clearer. Plan is for 3+1 days. US-ATLAS might be involved but primarly Berkley based. Probably 14.-16.5.
Connecting the dots conference also might happen this year, but no fixed plans yet.
Dimitra
Worked on analysis in the last weeks with MC samples.
Start soon with the migration to hook up the volumes. Theoretically, it could be done outside of Athena. But we need to see how it works. Idea is to have main tracking geometry being capable of launching the blueprint node construction and attach additional nodes.
Paul: GeoModel-tuning is already merged, but plugin-execution is still missing ("publishing"). The readout-geo-information is already in place and converted xml->readout. (not in memory yet, just in files) ITk split rolls cannot produced until we have this in. Paul maybe manages to solve this until end of January. Also hooking up the geometry should only take a few days. Hopefully, we just need a single flag for the navigator. Caching might not be necessary but important for the future.
Attila: How to figure out which Geant-version is used? What to mention in a paper? -> No-one in the room is sure.
Patternrecognision for micromegas need a bit more debugging.
Attila
Tried to make sycl work on Intel. Still full chain is not working. The code seems to be fine on nvidia and amd hardware.
AMD behaves well with sycl. But if AMD is the goal, we should do directly HIP. Maybe a technical student would be enough. Job would be well defined but a bit of work.
traccc: Thrust always comes with CUDA, but we also use it for other stuff. nvcc complains on the version, when we pre-download. Maybe remove it from our public headers? Mainly used for pair since std::pair is not enough for us.
Problem: nvidia is repacking/renaming things. thrust is now called CCCL. But intel does the same.
Joana
thrust: tuple works on device.
Looked at detray propagation benchmark plotting scripts. Needs some hardware to finally test. -> Hardware will be organised. Maybe get an AMD-gpu for ep-rnd?
Started to restructer the navigation code.
Alex
Plans for the next year:
- Finish thesis
- Some more exams at University
- University job
Some questions that came up during discussions at university.
"Why not Julia? GPU works out of the box"
NGT looks at Julia, CMS also looked at it. It is not a silver bullet. It is worth looking at.
cpp-interopt is not so great and we want to deploy in more experiments. Also not so widely supported as other languages, so maybe not sure if it is stable for the future.
No ahead of time compiler and cpp needs t
Works on gpu is similar to python numpy. Multiple kernels are launched and do a parallel for and return.
"Why not solve as a giant system and skip track finding/selection/clustering?"
We have a similar approach with the full graph.
We have around 500.000 measurements which make the system quite large. see also https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PAPERS/IDTR-2023-01/
The integrated magnetic field depends on the the particle hypothesis (type and momentum)
It could work but maybe not useful approach.
