ising-cuda

Learning project using the Ising model as a testing ground to cover the following topics:

• Writing efficient GPU kernel code to handle local interactions in $n$ dimensions
• Wrapping C++/CUDA in Python extension modules using nanobind
• Writing GPU kernels in JAX/Pallas

Project map

• ./ contains simple standalone implementations of Monte Carlo simulations of the Ising model. There are C++ reference and CUDA versions for each of the 2-d and general n-d cases.
  • python/ contains a Python project (ising_mcmc) that implements versions of the above wrapped in an extension module using nanobind; additionally, this includes a JAX and experimental Pallas backend.
    • python/src/ising_mcmc/ contains implementations of the sweeps function using various backends (currently CPU, CUDA, JAX, and Pallas)
    • python/notebooks/ contains some analysis and benchmarking, e.g. comparing the different backends

Python quickstart

The Python package is pip-installable from python/:

cd python
pip install .

Developing

The Python module can be pip-installed in editable mode

cd python
pip install -e .

allowing changes to Python files to be reflected in the installed module without running pip again. (Note that changes to the C++ code will require running the above command again; this should only recompile the changes, since scikit-build-core with CMake supports incremental builds.)

Nix Flakes defining development environments are provided. E.g. in python/notebooks, you can run

nix develop

to enter a development shell containing Python with ising_mcmc and plotting packages installed.