Improve Scaling of Stereo Candidate Search in PrForwardTracking
Throughput tests running over lead lead data show that the stereo hit finding scales horribly with the number of hits present in an event. This is solved by using the HoughSearch
object also used by the Hybrid Seeding which now checks the deviation of stereo hits from the already found x candidate in bins and thus makes it possible to spot the right candidate without testing too many combinations. The HoughSearch
class is updated to fit the new needs: the search can optionally check the total number of layers present in a cluster and the invalid cluster entries are made generic according to the underlying element type of the Hough search. Along the way the data structures are updated and explicit types are preferred over templates. The new stereo hit finding improves throughput of the Forward Tracking by ~13% for proton-proton and by ~30% for lead-lead. The efficiencies stay roughly constant while the ghost rate is slightly smaller.
tl;dr
- use Hough search for stereo hit finding
- Forward Tracking throughput up 11% @ constant efficiencies