Skip to content
Snippets Groups Projects

Fix hardcoded geometry in forward/seeding algorithms

Merged Alessandro Scarabotto requested to merge ascarabo_forward_fix_scifi_geometry into master

Fixing the geometry hardcoding in forward/seeding algorithms: #463 (closed)

Changing DumpFTGeometry to dump in the .bin geometry files also the info of the average z and average dxdy per layer.

Through Moore, dumped the geometry in scifi_geometry.bin.

./Moore/run gaudirun.py Moore/Hlt/Moore/tests/options/default_input_and_conds_hlt1_retinacluster.py Allen/Dumpers/BinaryDumpers/options/dump_geometry.py

The Consumers at the beginning of the Allen sequence take these constants from the geometry and store it in the GPU. The constants are accessible in each Allen kernel using constants.dev_scifi_geometry in the operator.

Implemented new versioning of scifi geometry for retro-compatibility:

  • v0 hardcoded - v2 read-in geometry: for scifi decoding v4,5,6
  • v1 hardcoded - v3 read in geometry: for scifi decoding v7,8

All new dumped geometries will be either v2 or v3 depending on the decoding. If Allen reads a geometry v0 or v1 then hardcoded scifi geometry values are used (retro-compatibility)

Testing the effect of going for layer-level to scifi quarter-level geometry constants on a second MR to factorise the work: #497

@dovombru @gligorov @lcalefic

Edited by Alessandro Scarabotto

Merge request reports

Loading
Loading

Activity

Filter activity
  • Approvals
  • Assignees & reviewers
  • Comments (from bots)
  • Comments (from users)
  • Commits & branches
  • Edits
  • Labels
  • Lock status
  • Mentions
  • Merge request status
  • Tracking
  • added 1 commit

    • fc927cfe - remove hardcoded geometry in seeding

    Compare with previous version

  • Added the same fixes on the scifi seeding algorithm (average z per layer and average dxdy per layer from geometry). One issue I found is the hardcoded dxdy value the seeding is using is 0.086 while from the geometry we get something more correct at 0.087489. This causes an increase on the ghost rate which by a quick check is due to search windows which are tuned to be too large in the uv layers. One quick fix is to reduce the tolerance of the uv search windows from TUNING_TOL = 2.f to TUNING_TOL = 1.f. Of course this would need an optimisation. I tag here @lohenry @ahennequ @cagapopo @lcalefic for opinions. Checking the effect on data as I did above for the forward (I think I have no time today I will do it in January when I come back)

    Allen Checker:

    standard minbias sample: MiniBrunel_2018_MinBias_FTv4_DIGI_retinacluster_v1.mdf (she should use a more recent sample for the tuning, here is just to show you the effect)

    MASTER (hardcoded geometry)

    seeding_matching_MASTER

    GEOMETRY INPUT (average z and dxdy per layer)

    Screenshot_2023-12-19_at_17.26.49

    QUICK TUNING (reduce uv window tolerance to TUNING_TOL = 1.f) Screenshot_2023-12-19_at_17.28.21

  • mentioned in issue #463 (closed)

  • added 1 commit

    • 1cd69ccd - add consumer for geometry versioning

    Compare with previous version

  • added 152 commits

    Compare with previous version

  • added 1 commit

    Compare with previous version

  • added 1 commit

    Compare with previous version

  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Loading
  • Please register or sign in to reply
    Loading