Geant4 External Propagation

Remit

This page gives JUST an overview/summary. For details, code listings etc.. see:

• DetSim GiGa Geant4 Handoff
• DetSim DsPmtSensDet

External Propagation Requirements

1. collect optical photon data for all generated optical photons
2. suspend other processing of optical photons beyond collection
3. once all (or some configured number) of optical photons are collected dispatch the external propagation and get the results back
4. merge the propagated photon (or maybe hit data) back into geant4
5. proceed with geant4 processing as if the propagation was standard geant4

Alternate Early Photon Collection

Collection of photon data within the processes before becomes a G4Track and being stacked would avoid track stacking processing and associated memory expense/concerns.

• to support this split off photon collection into G4DAEChromaManager

GPU Hit formation Approach

• StackAction::ClassifyNewTrack(G4Track*)

  • collect photon data (TrackID?) and kill G4Track

• StackAction::NewStage

  • batched external GPU propagation
  • from ZMQ response form the PMT hits and fill hit collection (requires detector and PMT id info to be gleaned GPU side)
  • try to keep detector specifics CPU side, keep GPU code as generic as possible

Pros

Avoids complications:

• G4Track memory constraints (diddling with placeholder duplicate objects to avoid)
• how to merge back

Cons

• detector specific code duplication needed from DsPmtSensDet::ProcessHit

How to implement ?

Main issue is forming PMTID

• look into general SDManager traverse to harvest PMTIDs and incorporate into the exported COLLADA geometry
  • remember emergent nature of the geometry traverse, cannot apply identifier labels to all PMTs within the recursive structure but easy to do in a simple separate flat array
• there is a Chroma slot for a solid identifier ?

Propagated Photon Mergeback approach

The Geant4 “allowed” way to move photons gets complicated with creation of a pseudo process and creating particle changes. That seems expensive and pointless.

Initially preferred this as could allow to avoid a reimplementation if ProcessHits.

pseudo process complications

Processes operate at track level, need to propagate at event (or track collection level)

• obvious optimization is to throw photons that dont get to PMTs, how to handle