cuda-grmonty

CUDA implementation of relativistic Monte Carlo code. Based on grmonty and grmonty: A Monte Carlo Code for Relativistic Radiative Transport .

Building

Build dependencies:

• C++20 compiler (gcc/clang)
• CUDA
• Bazel
• Python 3 (optional, for spectrum plotting)

CPU build

To build CPU-only version, run:

bazel build -c opt //cuda_grmonty:main

then run it with:

./bazel-bin/cuda_grmonty/main --harm_dump_path ./dump019 --spectrum_path ./spectrum -photon_n 5000000

The HARM dump file can be downloaded from grmonty repository.

GPU build

To build the program with CUDA acceleration, run:

bazel build -c opt //cuda_grmonty:main --config=cuda

and run it the same as above.

Spectrum plot

To plot the spectrum, run plot_spectrum.py, preferably using uv as follows:

uv run ./plot_spectrum.py --spectrum_path ./spectrum --plot_path ./spectrum.png

If you want to use Python directly instead, install dependencies specified in plot_spectrum.py#4. And run:

python3 ./plot_spectrum.py --spectrum_path ./spectrum --plot_path ./spectrum.png

Visualizations

To enable recording photons positions for visualizations install OpenCV and add --trajectories_recording_path parameter.

./bazel-bin/cuda_grmonty/main --harm_dump_path ./dump019 --spectrum_path ./spectrum -photon_n 5000000 --trajectories_recording_path ./trajectories.mp4

Example result

Benchmarks

Input parameters:

• HARM dump: dump019
• Estimate of photon number: 1000000
• Mass unit: 4.e19

System:

• Arch Linux (6.18.6-arch1-1)
• CUDA 13.1
• clang 21.1.6

Hardware:

• CPU: i9-14900k
• RAM: 2x32 GB (4800 Mhz)
• GPU: RTX 3060 12 GB

Version	Other parameters / notes	Photon rate [1/s]	Total duration [s]
grmonty	OMP_NUM_THREADS=1	59743	272.51
grmonty	OMP_NUM_THREADS=32	289866	55.36
v0.0.1	CPU only	51429	317.02
v0.1.0		65119	241.84
v0.1.1		137939	114.95
v0.1.2		246538	64.91
v0.1.3		307186	52.41
v0.1.4		352152	45.88

Development

Testing

The tests can be ran with:

bazel test //cuda_grmonty/tests/...