feat(src): add G4 validation executables for aligned-mode comparison#271
feat(src): add G4 validation executables for aligned-mode comparison#271ggalgoczi wants to merge 15 commits into
Conversation
There was a problem hiding this comment.
How is this new file used? Is it supposed to replace the same functionality in QCurandStateMonolithicTest https://github.com/BNLNPPS/eic-opticks/blob/main/tests/test_opticks.sh#L5-L11?
|
Please update or rebase the branch, and change the PR title to better reflect the intended change using conventional commit format. |
I'll rebase 271-273 once 269 and 270 are merged — otherwise they'll just need rebasing again. |
|
…ise hit paths Three corrections to test_wavelength_shifting.sh: 1. The GPU hit-file path was a literal "/tmp/MISSING_USER/opticks/GEOM/GEOM/...". "MISSING_USER" is opticks' fallback when $USER is unset (sysrap/spath.h:294), and the trailing "GEOM" is the unset $GEOM env var. The script never exported either, so the GPUPhotonSourceMinimal run wrote to a different path than the one the script then loaded. Set USER=fakeuser and GEOM=fakegeom (matching tests/test_GPURaytrace.sh and tests/test_GPUPhotonSource_8x8SiPM.sh) and reference the path through the env vars. 2. Drop the hard-coded "/opt/eic-opticks/bin/" prefix on both binaries. The install dir is on $PATH in the CI image and in the local eic-opticks-env.sh, matching the convention of every other test. 3. StandAloneGeant4Validation is provided by PR #271, not by anything currently on main. SKIP cleanly rather than failing when the binary isn't installed, so this test is benign on branches where #271 has not yet landed. Also pre-clean stale GPU/G4 hit files at the top of the run so a silently-failing binary cannot make the comparison succeed against leftover data.
…hmark Pure Geant4 optical photon simulation with no opticks/GPU dependencies. Loads GDML, fires torch photons, collects hits via G4VSensitiveDetector, saves g4_hits.npy for comparison with GPU output. Supports multi-threaded mode via G4MTRunManager (-t N flag) by splitting photons across multiple events. Sequential mode (-t 0) for reproducibility.
Adds photon-by-photon aligned comparison with GPU via U4Random precooked curand sequences. Includes PhotonFateAccumulator for indexed g4_photon.npy output, AlignedOpticalPhysics with InstrumentedG4OpBoundaryProcess, and U4Recorder integration for SEvt lifecycle and random alignment.
…ess dependency Use standard G4OpticalPhysics and direct U4Random::SetSequenceIndex. Only requires linking U4 (for U4Random) — no u4/ source modifications needed.
Uses ShimG4OpAbsorption/ShimG4OpRayleigh instead of standard G4OpticalPhysics when --aligned. Improves match from 95.7% to 97.7% by matching the GPU's RNILL random consumption pattern.
Generates Philox random streams matching GPU seeding for U4Random aligned mode. 113 lines, compiles with nvcc, no opticks dependencies beyond NP.hh.
Compares GPU and G4 photon.npy arrays element-by-element: flag match, position match, distributions, divergent photon listing.
Fires configurable electron in LAr, G4 produces scintillation/Cerenkov photons and tracks them to SiPM detection. For comparison with GPURaytrace which uses G4CXOpticks to hand gensteps to GPU for optical propagation. Tested: 1 MeV e- in det.gdml produces ~170 SiPM hits/event.
Relax MaxRecord assert to permit values above sseq::SLOTS (32). In DebugLite mode, respect OPTICKS_MAX_RECORD env var if set, falling back to the default record_limit (32) otherwise. This enables full step history recording (e.g. 1000 steps) for photon path analysis without affecting production performance.
G4OpticalPhysics defaults to delta time profile for G4OpWLS, which applies a fixed delay equal to WLSTIMECONSTANT. The physically correct model is exponential decay sampling: dt = -WLSTIMECONSTANT * log(u). The GPU implementation already uses exponential. This fix aligns G4 with GPU and with the correct stochastic decay physics.
SetWLSTimeProfile in GPURaytrace changed G4OpticalParameters during physics list initialization, inadvertently altering the electron tracking (different genstep count). The WLS time profile setting only matters for StandAloneGeant4Validation where G4 tracks optical photons directly. In GPURaytrace, the GPU handles optical physics so the G4 time profile is irrelevant.
Apply repository .clang-format (Microsoft base) to the four files introduced in this PR. Resolves the cpp-linter complaint from the PR-open check; sysrap/SEventConfig.cc keeps its existing aligned-assert style for consistency with surrounding code (the new lines already conform to the project style).
…Test Expand the file's leading docblock to make the distinction explicit: QCurandStateMonolithicTest produces curand internal STATE structs used to initialise the GPU kernel; generate_precooked_rng produces pre- evaluated random SAMPLES that U4Random reads in --aligned mode so each G4 draw matches the corresponding GPU draw byte-for-byte. Different artifacts, different consumers; both needed for the validation workflow.
ggalgoczi
force-pushed
the
feat/aligned-validation
branch
from
71da757 to
5888c19
Compare
|
Pushed a force-with-lease rebase + two follow-up commits. The rebase onto current main auto-cleaned several problems in |
ggalgoczi
changed the title
cpp-linter runs with lines-changed-only=true, which checks only PR- modified lines against the project .clang-format. The previous push left line 777 in the surrounding spaced-paren `assert(...)` style for local consistency with lines 778-779, but the bot rightly flags this line because *it* was the one the PR changed. Apply the project style to that single line. Surrounding pre-existing code keeps its style; cpp-linter does not check unchanged lines.
plexoos
left a comment
plexoos
left a comment
There was a problem hiding this comment.
I still do not quite understand how this study is intended to be performed, or what it is meant to demonstrate.
A few points are unclear to me:
- Where is fresnel_single.gdml? Why is it not included?
- Why is there not a 100% match when using pre-cooked random numbers?
- Why is the CUDA file not compiled with CMake?
- In what order should the executables be run in order to verify the claims?
- Should the new executables become part of the test suite?
2 participants
Add infrastructure for photon-by-photon aligned comparison between GPU and G4. StandAloneGeant4Validation provides a CPU-only G4 simulation with --aligned mode using precooked curand sequences via U4Random. Includes AlignedOpticalPhysics with ShimG4Op* wrappers for random number consumption alignment. G4ValidationGenstep enables electron-primary genstep validation. Achieves 97.7% flag match on fresnel_single.gdml.