Draft: Adapt matching with UT for 2024-data taking

Need to be tested with lhcb-datapkg/ParamFiles!105 (merged)

Introduction

Following changes are included:

• Tune the UT Hit search window size based on 2024 real data
• Re-train the V2 NN-based ghost killer to adapt upstream changes for saturation fix
• Loose the clone killing conditions for matching with UT algorithm, in such way different VELO segment can associate to same UT segement without being killed.
• Extend the max size of container from 1000 to 2000 in order to find UT segments
• Replace the risk atomicAdd by safer one, once overflow happens will reset the container size to zero to skip the event
• Add correct UT hit plane code to the matching consolidation

Search window definition

Plots are created running MatchingWithUT with very wide search window.

First hit

Rest hits

Performance

Adding UT Hits in real data we lost only about 1.07% KsToPiPi signal yield.

Following trigger rate changes is observed using default tuning:

Without UT

Hlt1TrackMVA:                                     545/ 90000, (  181.67 +/-     7.76) kHz
Hlt1TwoTrackMVA:                                 1639/ 90000, (  546.33 +/-    13.37) kHz
Hlt1D2KK:                                         176/ 90000, (   58.67 +/-     4.42) kHz
Hlt1D2KPi:                                        229/ 90000, (   76.33 +/-     5.04) kHz
Hlt1D2PiPi:                                        87/ 90000, (   29.00 +/-     3.11) kHz
Hlt1Dst2D0Pi:                                      33/ 90000, (   11.00 +/-     1.91) kHz
Hlt1KsToPiPi:                                       3/ 90000, (    1.00 +/-     0.58) kHz
Hlt1KsToPiPiDoubleMuonMisID:                        4/ 90000, (    1.33 +/-     0.67) kHz
Hlt1TwoTrackKs:                                   163/ 90000, (   54.33 +/-     4.25) kHz
Hlt1TwoKs:                                         23/ 90000, (    7.67 +/-     1.60) kHz

Inclusive:                                       4665/ 90000, ( 1555.00 +/-    22.17) kHz

With UT

Hlt1TrackMVA:                                     496/ 90000, (  165.33 +/-     7.40) kHz
Hlt1TwoTrackMVA:                                 1465/ 90000, (  488.33 +/-    12.65) kHz
Hlt1D2KK:                                         157/ 90000, (   52.33 +/-     4.17) kHz
Hlt1D2KPi:                                        189/ 90000, (   63.00 +/-     4.58) kHz
Hlt1D2PiPi:                                        70/ 90000, (   23.33 +/-     2.79) kHz
Hlt1Dst2D0Pi:                                      34/ 90000, (   11.33 +/-     1.94) kHz
Hlt1KsToPiPi:                                       4/ 90000, (    1.33 +/-     0.67) kHz
Hlt1KsToPiPiDoubleMuonMisID:                        4/ 90000, (    1.33 +/-     0.67) kHz
Hlt1TwoTrackKs:                                   145/ 90000, (   48.33 +/-     4.01) kHz
Hlt1TwoKs:                                         17/ 90000, (    5.67 +/-     1.37) kHz

Inclusive:                                       4335/ 90000, ( 1445.00 +/-    21.41) kHz

TODO

• Plots for UT hit search window definition based on real data
• Plots for new NN performance based on both data and MC
• Throughput test result with testbench

FIY: @dtou @dovombru @adeoyang @cagapopo @gligorov