ITS GPU tracking

• General priorities:
  • F. Schlepper found some rare (more often in PbPb data) deviations in the deterministic mode, investigations restricted to a few blocks of code (findCellsNeighbours, post-CHEP). ==> Fixed: https://github.com/AliceO2Group/AliceO2/pull/14200
  • [WIP] Thrust allocator with external memory management: for the moment it compiles but it does not work, needs dedicated discussion.
  • Focusing on porting all of what is possible on the device, extending the state of the art, and minimising computing on the host.
    • Moving vertexing routines to the externally managed memory system. -> WIP
      • Currently optimising the o2::its::timeframeGPU intialisation GPU transfers, trying to maximise data reuse & anticipating loads when it is possible.
• Optimizations:
  • Asynchronous parallelisation in the tracklet finding, i.e. Multi-streaming for obvious parallelisations.
  • intelligent scheduling and multi-streaming can happen right after.
  • Kernel-level optimisations to be investigated.

TODO:

• • Reproducer for HIP bug on multi-threaded track fitting: no follow-up yet.
  • Fix possible execution issues and known discrepancies when using gpu-reco-workflow: no progress.

DCAFitterGPU

• Deterministic approach via using SMatrixGPU on the host, under particular configuration: no progress.