The motivation for the faster reconstructions with accelerated k-tracker approach (for online monitoring and detector studies): 

  1. Improving the speed of the basic tracking algorithm while also making the I/O at every stage easy to access and test. 

  2. Writing modularized Python scripts to enable the use of CUDA/NUMBA for enhanced performance.

  3. The primary goal is not to use AI, but try to get close to the original k-tracker with only minor AI if needed.
  4. RUS file will used for I/O.
  5. Focusing on speed and stand-alone reliance, so no database reading/writing or anything leading to further latency, avoiding Fun4All and other dependencies.

We will achieve this using vectorized python, compiled python, JIT-compiled, broadcasting, dataflow optimization,...

Intended libraries: numpy, pandas, numba, tensorflow, pytorch 

Documentation: 

The workflow of the kTracker is documented in kTracker_Collaboration_20140212.pdf

List of Modules: 

  1. Event Reductions. 
    1. acceptEvent(SRawEvent* rawEvent)
  2. Hit Reduction and Declusturing:  EventReducer
  3. Quality Cuts.
    1. acceptTracklet(Tracklet& tracklet) 
    2. hodoMask(Tracklet& tracklet)
    3. muonID_comp(Tracklet& tracklet)
    4. muonID_search(Tracklet& tracklet)
    5. muonID_hodoAid(Tracklet& tracklet)
  4. Bulding Tracklets in St2 & 3 and project to St1, and building tracklet at St1.
    1. buildTrackletsInStation(int stationID, int listID, double* pos_exp = nullptr, double* window = nullptr)
    2. getSagittaWindowsInSt1(Tracklet& tracklet, double* pos_exp, double* window, int st1ID)
    3. getExtrapoWindowsInSt1(Tracklet& tracklet, double* pos_exp, double* window, int st1ID)
    4. buildBackPartialTracks() 
  5. Constructing global track.
    1. getExtrapoWindowsInSt1
    2. buildGlobalTracks()
    3. processOneTracklet(Tracklet& tracklet)
    4. fitTrack(KalmanTrack& kmtrk)



Critical info we want for commissioning, detector studies, and fast tests: track hit positions in each station, tracklets hit lists, back partial tracks hit list, tracks hit list, reconstructed momentum and vertex, test for alignment and efficiency very quickly.

  • No labels