PrLongLivedTracking speedup with SIMD (!4427) · Merge requests · LHCb / Rec

This MR modifies part of the PrLongLivedTracking algorithm to achieve an execution speedup using SIMD instructions.

Optimisation of `addUhits`

The biggest change lies in the addUhits function. The current logic:

Preselect U hits using the distance to the track
Copy a track candidate and add a U hit (loop over all the U-hits)
Run the simplyFit function, extract chi2.
If chi2 is too big or the track is not y-compatible, delete the track.

The proposed idea is to parallelise over U hits using SIMD. Therefore:

Preselect U hits using the distance to the track. Copy the preselected hits to the SOA container.
Run simplyFitSIMD. This function will fit the track candidates with different U hits using SIMD. In addition, it will check the chi2 and y-compatibility requirements. Only the accepted tracks are stored in the output vector

Affected parts of the code:

PrDownTrack.h: Downstream::SOA::Hits
PrLongLivedTracking.cpp: simplyFitSIMD(), addUHits()

Optimisation of X-track fitting

Current behaviour (scalar):

Find all the hits that satisfy the tolerance window in findMatchingHits
For every X hit, create a track candidate, add the X hit (fitXProjection), and fit (xFit).
Check the track quality and delete the track if it's bad.

The proposed idea is to parallelise over X hits using SIMD. Therefore:

Collapse the three previous functions into one createXTracks.
Fitting using SIMD instructions and parallelizing over the second X hits.

Affected parts of the code:

PrDownTrack.h: Downstream::SOA::Hits
PrLongLivedTracking.cpp: findMatchingHits(), fitXProjection(), xFit()

Buffering of additional variables in `Downstream::Hit`

This MR adds two more variables to the Downstream::Hit object: xAtYEq0 and dxDy. Previously, these quantities were extracted from the hitHandler object every time they were needed for the computations.

Affected parts of the code: