Software R&D Working Meeting Minutes

2024-12-11

Room: Graeme, Joana, Aurora, Andi, Witek, Piyush, Pere, Severin, Florine, Vincenzo, Alberto

Remote: Joshua, Peter, Felice, Paul

Apologies: Andre, Mateusz, Anna, Juan, Swathi (Key4hep in person meeting)

News

EP R&D day for 2025 has been announced for 7 May - save the date!

Track Reconstruction Plans - Joana

Q. Pere: Does the goemetry support boolean solids?
- No, this was avoided. So far the need hasn't arisen, even ITk was implementable without them. It would really not work well on a GPU.

Q. Severin: Are you memory transfer limited?
- Not really. Sometimes updates are needed (IoVs, alignments) during a job. The issues are mainly memory access bound (Severin, this was the same for AdePT).
- Important to have these alignments as it is important not to compromise on

Q. Felice: is transform refering to alignement per event? Do you correct event by event, or do you recalculate and reupload?
- It's a recalculation and reupload - this is granular at the IoV level. (Would sometimes be ignored in HLT, but for offline need the full accuracy.)
- Q. Would a unified memory approach help?
- In the worst case these changes would happen on 60s lumiblocks, so there should be enough events to justify reuploading.

Q. Felice: Could you use 2D texture memory for some data?
- For the B field this acceleration is there. For other use cases, not yet.

Q. Severin: How is the performance with SYCL/Intel GPUs?
- The full chain isn't supported at the moment, but early results were quite promising?

Q. Severin: How are the results in single precision?
- Losing a few % in single precision in efficiency. Not clear where the loss is coming from yet.
- Optimising this will be use case driven. e.g., in HLT maybe ok to find tracks, but not go for maximum resolution that might be done offline? Or track finding on GPU, then refitting on CPU in double precision?

AOB

Have a great holiday everyone! See you next year.