POW VecGeom 2026: restructure VecGeom to deliver higher GPU navigation performance, consistent behavior across backends, and a maintainable, portable code base (CUDA/HIP). The detailed plan is drafted here
-
Simplify and standardize the navigation interface (clear inputs/outputs; consistent boundary behavior; consistency of VNavigator/GPU interfaces).
-
Improve GPU navigation performance with a more locality-aware accelerator (beyond today’s BVH)
-
Replace recursion-heavy algorithms (esp. Boolean navigation) with iterative, stack-friendly implementations
-
Remove legacy vector/SIMD remnants in kernels to simplify code and improve scalar performance
-
Add a portability layer for CUDA/HIP and refactor device code toward POD + explicit dispatch (no device virtuals)
Mid-term items:
- Mixed-precision implementation (global API in double; per-solid precision policies)
- Documentation: user guides + algorithm notes to match the restructured core
Long-term goal: Unified solids common library providing solid-level and global navigation algorithms for Geant4, ROOT, reconstruction programs.