xdg-event-particle-sim

A basic GPU event-based pseudo particle transport simulation application for mocking neutral-particle transport against CAD models with the XDG library. This application acts as a mock interface for XDG's GPU API for future particle transport codes wishing to use XDG for CAD transport. GPU sorting of volumes is provided via CUDA thrust or HIP thrust sorting.

Building

This application is built against an installed XDG package. Build and install XDG with cuBQL support first, then point this project at that installation. The XDG build uses XDG's cubql_llvm_ada preset, while this standalone application uses its own llvm_ada preset.

Example XDG install:

cd ~/xdg
cmake --preset cubql_llvm_ada \
  -DCMAKE_INSTALL_PREFIX=$HOME/xdg-install/cubql_llvm_ada
cmake --build --preset cubql_llvm_ada --target xdg -j
cmake --install build/cubql_llvm_ada \
  --prefix $HOME/xdg-install/cubql_llvm_ada

Then build this application with the standalone LLVM/Ada preset:

cd ~/xdg-event-particle-sim
git submodule update --init --recursive

export XDG_INSTALL_PREFIX=$HOME/xdg-install/cubql_llvm_ada

cmake --preset llvm_ada
cmake --build --preset llvm_ada

The standalone llvm_ada preset enables CUDA Thrust sorting and uses LLVM OpenMP offload flags for an NVIDIA RTX 2000 Ada GPU.

Run example:

./build/llvm_ada/xdg-particle-sim-event <mesh.h5m> \
  --mesh-library MOAB \
  --rt-library CUBQL \
  --sort-by-volume

The --sort-by-volume option enables the Thrust queue sort at runtime.

Plotting Ray Launch Data

The plot_ray_launches.py script reads the per-launch CSV written by --enable-profiling-ray-launch and produces basic ray-tracing performance plots.

Generate ray launch data:

./build/llvm_ada/xdg-particle-sim-event <mesh.h5m> \
  --mesh-library MOAB \
  --rt-library CUBQL \
  --enable-profiling-ray-launch \
  --ray-launch-profile-output ray-launches.csv

Then run:

python3 plot_ray_launches.py ray-launches.csv

Mean Free Path Sweeps

The sweep_mean_free_path.py script runs the particle simulation over a list of mean free paths, writes one consolidated CSV row per run, and automatically plots particle outcomes, ray throughput, runtime, and transport workload.

For example:

python3 sweep_mean_free_path.py run atr.h5m \
  --mean-free-paths 0.1 0.25 0.5 1 2 5 10 \
  --output-directory mfp_sweep \
  --sort-by-volume \
  --profile-ray-launches

The main outputs are:

mfp_sweep/mean_free_path_sweep.csv
mfp_sweep/mean_free_path_sweep_summary.png
mfp_sweep/mean_free_path_sweep_particle_states.png
mfp_sweep/mean_free_path_sweep_ray_performance.png
mfp_sweep/mean_free_path_sweep_workload.png

With --profile-ray-launches, the script also retains a ray-launch CSV and generates the existing ray-launch summary plot for every mean free path.

Plots can be regenerated without rerunning transport:

python3 sweep_mean_free_path.py plot mfp_sweep/mean_free_path_sweep.csv