villar-pso

Joint two-band Villar light-curve fitter using Particle Swarm Optimisation (PSO).

Using as a dependency

Add to your project's Cargo.toml:

Without GPU (CPU only)

[dependencies]
villar-pso = { git = "https://github.com/frenbox/villar-pso.git" }

No feature flags needed. This gives you the core library (fit_lightcurve, preprocess, etc.) and Rayon-based CPU parallelism.

With GPU (CUDA)

villar-pso = { git = "https://github.com/frenbox/villar-pso.git", features = ["cuda"] }

Requires nvcc and CUDA toolkit on the build machine. This enables the gpu module (GpuContext, GpuBatchData, batch_pso_multi_seed, etc.).

Python bindings (PyO3)

villar-pso = { git = "https://github.com/frenbox/villar-pso.git", features = ["python"] }

Build with maturin:

maturin develop --release --features python         # CPU only
maturin develop --release --features "python,cuda"   # CPU + GPU

Benchmarks

Two benchmark binaries are included:

Binary	What it tests
cpu-dispatch-bench	Compares sequential dispatch (par_iter, 1 file at a time) vs batch dispatch (par_chunks(500)) across 2, 4, 6, 8, 10 Rayon threads
gpu-scaling-bench	Scales from 1 to N GPUs, one Rayon thread per GPU, processing all sources in chunks of 500

Running benchmarks locally

CPU benchmark (no GPU needed):

cargo build --release --bin cpu-dispatch-bench
./target/release/cpu-dispatch-bench /path/to/photometry/

GPU benchmark (requires CUDA):

CUDA_HOME=/usr/local/cuda cargo build --release --features cuda --bin gpu-scaling-bench
./target/release/gpu-scaling-bench /path/to/photometry/

Running benchmarks via Apptainer

For machines where CUDA is available only through containers.

1. Build the container (from the villar-pso/ directory):

apptainer build --fakeroot rustgp.sif rustgp.def

2. Run CPU benchmark:

apptainer exec \
    --bind /path/to/project:/path/to/project \
    rustgp.sif \
    /app/cpu-dispatch-bench /path/to/photometry/

3. Run GPU benchmark:

apptainer exec --nv \
    --bind /path/to/project:/path/to/project \
    rustgp.sif \
    /app/gpu-scaling-bench /path/to/photometry/

The --nv flag exposes host NVIDIA drivers to the container.

4. Limit GPU count (optional):

apptainer exec --nv \
    --bind /path/to/project:/path/to/project \
    rustgp.sif \
    /app/gpu-scaling-bench /path/to/photometry/ --gpus 4

Project structure

villar-pso/
├── Cargo.toml
├── build.rs              # CUDA kernel compilation (when cuda feature enabled)
├── rustgp.def            # Apptainer container definition
├── cuda/
│   └── villar_joint.cu   # GPU cost function kernel
└── src/
    ├── lib.rs            # Core library: preprocessing, PSO, fitting, PyO3 bindings
    ├── gpu.rs            # GPU batch PSO, multi-GPU support
    └── bin/
        ├── cpu_dispatch_bench.rs
        └── gpu_scaling_bench.rs