Convert Geant4 geometry to VecGeom in memory

[mrguilima]

This is a first "working" version of the Geant4-to-VecGeom in-memory geometry conversion, for a first round of comments.

Known issues:

• Geant4 geometry could not be deleted. Attempt to call G4PhysicalVolumeStore::GetInstance()->Clean() results into a crash at the second test Setup()
• Similarly, there are duplicated materials from the second Setup() onwards
• Geant4 units are properly read in mm units, but VecGeom model is built without a mm->cm conversion in volume units. To be checked and discussed.

[sethrj]

Nice @mrguilima ! I've left a first round of comments.

[sethrj]

FYI the CI is failing, possibly because of a version difference between the VecGeom versions?

FAILED: src/celeritas/CMakeFiles/celeritas_objects.dir/ext/detail/G4VecgeomConverter.cc.o 
/usr/bin/c++ -DG4LIB_BUILD_DLL -DJSON_DIAGNOSTICS=0 -DJSON_USE_IMPLICIT_CONVERSIONS=1 -DVECCORE_ENABLE_CUDA -I/var/jenkins/workspace/Celeritas_PR-557/src -I/var/jenkins/workspace/Celeritas_PR-557/build/include -isystem /opt/view/include/Geant4 -isystem /usr/local/cuda/include -isystem /opt/view/include -isystem /opt/vecgeom/include -Wall -Wextra -pedantic -Werror -Wno-error=deprecated-declarations -O3 -DNDEBUG -fPIC -fopenmp -faligned-new -std=c++14 -MD -MT src/celeritas/CMakeFiles/celeritas_objects.dir/ext/detail/G4VecgeomConverter.cc.o -MF src/celeritas/CMakeFiles/celeritas_objects.dir/ext/detail/G4VecgeomConverter.cc.o.d -o src/celeritas/CMakeFiles/celeritas_objects.dir/ext/detail/G4VecgeomConverter.cc.o -c /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/G4VecgeomConverter.cc
In file included from /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/G4VecgeomConverter.cc:193:
/var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/GenericSolid.hh: In member function 'vecgeom::cxx::VPlacedVolume* GenericSolid<S>::SpecializedVolume(const vecgeom::cxx::LogicalVolume*, const vecgeom::cxx::Transformation3D*, TranslationCode, RotationCode, vecgeom::cxx::VPlacedVolume*) const':
/var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/GenericSolid.hh:186:62: error: invalid new-expression of abstract class type 'GenericPlacedVolume'
  186 |         return new GenericPlacedVolume(volume, transformation);
      |                                                              ^
In file included from /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/GenericSolid.hh:17,
                 from /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/G4VecgeomConverter.cc:193:
/var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/GenericPlacedVolume.hh:30:7: note:   because the following virtual functions are pure within 'GenericPlacedVolume':
   30 | class GenericPlacedVolume : public vecgeom::VPlacedVolume
      |       ^~~~~~~~~~~~~~~~~~~
In file included from /opt/vecgeom/include/VecGeom/management/GeoManager.h:14,
                 from /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/G4VecgeomConverter.hh:20,
                 from /var/jenkins/workspace/Celeritas_PR-557/src/celeritas/ext/detail/G4VecgeomConverter.cc:14:
/opt/vecgeom/include/VecGeom/volumes/PlacedVolume.h:407:16: note: 	'virtual void vecgeom::cxx::VPlacedVolume::SafetyToInMinimize(const vecgeom::cxx::SOA3D<double>&, vecgeom::cxx::Precision*) const'
  407 |   virtual void SafetyToInMinimize(SOA3D<Precision> const &points, Precision *const safeties) const = 0;
      |                ^~~~~~~~~~~~~~~~~~

[sethrj]

OK, there is still plenty that can be improved, but since the plan is to move this to a separate library (and because it's derivative code) let's just call it good. I'm going to add to the code list to LICENSE before merging.

[sethrj]

@mrguilima I built your branch and spent some detailed time looking it over last night and fixing issues.

• It didn't build on my mac at first (because vecgeom doesn't define DeviceVec if CUDA isn't enabled)
• I didn't catch that these detail classes weren't in the detail namespace
• By removing geant4 includes from the header I'm able to directly include the file from VecgeomParams
• Several functions were unused and several unused macro-blocked pieces of code had errors in them; I deleted both
• Temporarily changed the test results to pass on my mac: particles are getting stuck in the torus

The last one is particularly worrisome because it means we're getting different results from different builds of VecGeom. Since the torus tracking fails both in direct vecgeom and on the G4 constructed one, perhaps it's better to punt on that, remove the torus from our unit test, and report a bug on vecgeom jira?

Three other big things:

1. I think the converter is leaking a ton of memory. It's unclear from the code which of the bajillion pointers take ownership.
2. Constructing boolean volumes (addition, subtraction) results in duplicate name warnings like warning: Geometry contains duplicate volume names: "placedB_bool_right" (10), "placedB_bool_right" (12). These duplicate names are all over the place in the CMS run 3 geometry (even with VGDML) and seem to break our ability to transport on it. Can you figure out why these are being constructed and how? Is one a temporary?
3. It looks like the world pointer that we pass to geo_manager.SetWorld get copied instead of owned, so we won't be able to play any tricks with the pointer maps we build so the vecgeom pointers are useless. We'll need to build a separate map of g4_logical_volume.GetInstanceID() to vecgeom_logical_volume.id().

[mrguilima]

Commit e220a9d breaks my VecGeom tests in Linux (as that commit message implies).

[sethrj]

Following up from @mrguilima 's tests: building vecgeom (can you confirm it's 1.2.2?) with:

/opt/homebrew/bin/cmake \
    -G Ninja \
    -DBUILD_SHARED_LIBS="OFF" \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_INSTALL_PREFIX="/Users/lima/work/cele/install" \
    -DCMAKE_PREFIX_PATH="/Users/lima/work/cele/install" \
    -DVECGEOM_BACKEND="scalar" \
    -DVECGEOM_BUILTIN_VECCORE="OFF" \
    -DVECGEOM_GDML="ON" \
    -DVECGEOM_GEANT4="OFF" \
    -DVECGEOM_NO_SPECIALIZATION="ON" \
    -DVECGEOM_ROOT="OFF" \
    -DVECGEOM_TEST_BENCHMARK="OFF" \
    -DVECGEOM_TEST_VALIDATION="OFF" \
    -DVECGEOM_USE_NAVINDEX="ON" \
    -DXercesC_INCLUDE_DIR="/Users/lima/work/cele/install/include" \
    -DXercesC_LIBRARY="/Users/lima/work/cele/install/lib/libxerces-c.dylib" \
    /Users/lima/work/cele/deps/vecgeom

the tests pass without the REVERTME changes I added... i.e. they don't show the epsilon bumps through the torus. This is with an M1 laptop running MacOS Monterey 12.6.4 and clang 14.0.0.

[sethrj]

Once the CI passes I'm going to merge! 🎉