NextGenTrigger Task 1.7

• Implementing our SoA code in MadGraph, together with
  • Stefan Roiser
  • Daniele Massaro
• No news

Implement NGT SoA Code in O2 standalone benchmark

• Working on this fork of the AliceO2 repo, with a CI pipeline:
  • Running on NGT hardware with 4 different GPUs (Nvidia and AMD)
  • Extended CI-pipline to fail if GPU.out changes
• Implemented SoA in:
  
  • SectorTracker
    • GPUTPCBaseTrackParam
    • GPUTPCTrackParam
    • GPUTPCTracklet
    • GPUTPCTrack
• Went back from CRTP to classical inheritance
• Implemented new way of providing iterators in O2: https://godbolt.org/z/haas6YP6c
• Issues since the latest updates of O2:
  • CI-pipelines on the NGT cluster fail (problem of missing libraries)
  • GPUs that are not inherently supported by the new O2 fail (thanks Gabriele for the workaround)
  • Missing build dependencies on CVMFS (Giulio is on it)
• Next Steps:
  • Adapt unit test to the changes
  • Adapt benchmarks to the changes
  • Make better use of SoA to improve performance
  • Try David's suggestion

Issues with AMD W7900 after updating to latest O2

1. Providing a .par file with certain Parameters results in non-deterministic behavior, despite deterministic mode.
2. Can't create my own .par files with dumpGPUDefParam.C of newest O2. Yields error incompatible launch bounds.
3. Changing FindO2GPU.cmake so that the VEGA or TAHITI setting are applied, results in compilation error: static_assert(NTHREADS == GPUCA_WARP_SIZE)
4. The only parameters that worked on W7900 are the ones for MI100 with WARP_SIZE changed to 32.
5. Other parameters (e.g. RDNA) yield deterministic behavior, but with a (slightly) different GPU.out than obtained wit other GPUs.

Moreover:

1. In my setting, I couldn't just add new GPUs to GPUParameters.csv: The headers of the newly added GPUs were created, but not the corresponding .par files.
2. In GPUParameters.csv, the AMD GPUs RDNA and MI210 are missing, but they show up in FindO2GPU.cmake. Is this intended?