Reference Kalman filter: tests and development

[asalzburger]

Follo-up the discussion of the acts-parallelisation meeting from 23/08/204, see https://indico.cern.ch/event/1435327

The proposal is to change the progressive/combinatorial filter (PKF/CKF) for track finding with an (alternative) reference trajectory based track finding (RKF) approach. This should be seen as an alternative implementation of the the current CKF baseline, the latter should still go through the foreseen validation and optimisation.

• Algorithmic flow of PKF/CKF with mixed navigation + Kalman formalism (not entirely correct, but schematically right):

• Algorithmic flow of RKF track finding:

The intention is to break the navigation/propagation loop from the Kalman math and hence have the chance to increase the thread efficiency and reduce desynchronisation. The RKF can thus be seen as two independent tasks of preparing the reference trajectories (a) and then run the Kalman updates (b) - on differences to the reference trajectory.

Strategy

Pre-tests that should enhance our current understanding of thread desynchronisation, this part is largely covered by the already foreseen/planneed KF/CKF optimisation campaign.

• check if in the current implementation thread efficiency can be improved in case n identical seeds are followed
• check if in the current KF implementation (refitting mode) thread efficiency can be improved in case n identical tracks are refitted
• evaluate what degradation is introduced by randomising the tracks successively

(a) Reference trajectory building:

• Run propagation through the detector and start performance analysis (@beomki-yeo has already started this)
• Check if the thread efficiency can be improved if identical/similar trajectories are followed in parallel
• Evaluate and optimise the detray propagation performance
• If necessary, implement an acceleration structure/material update method with less branching
• develop a truth based reference trajectory building module to help develop and optimise (b)

(b) Referene Kalman filtering

• Implement a first prototype for RKF
• Run on optimal reference trajectories, repeating the tests: identical - similar - random tracks in parallel