The project contains four small applications to simulate abelian sandpiles. Each one has its own directory.

In Atom/. This is a version in 6502 assembly language for the Acorn Atom. Simply load it up into an Acorn Atom (or emulator) and run.

In Metal/

Rewritten for Swift/Metal. Extremely unpolished. Read the source code in main.swift. Change parameters by modifying the source :) Build with make. Run with ./pile <n> and it'll do 2^n grains dropped in the centre. It just renders the bottom right corner. Use stitch.sh <in> <out> to construct the full image by symmetry. (I'm too lazy to study ImageMagick. So it doubles the middle row and column.)

In Jax/. Ported to jax. Much faster than TF version. I think it's still slower than raw CUDA.

Install jax with pip install jax jaxlib. Install PIL with pip install pillow.

Run with python jaxpile.py.

In TensorFlow/. Options and instructions in the comments.

You'll need Tensorflow 2. I installed it with the command

pip install tensorflow==2.0.0-beta1 or
pip install tensorflow-gpu==2.0.0-beta1

On OSX it'll run slow as TF doesn't support GPUs on OSX. It does use multiple cores though, if you have them. It should output messages like

    Wrote to 'out.0000.0000.jpg'
    Wrote to 'out.0000.0001.jpg'

These are checkpoints in the process. The last image is the final reduced sandpile.

On Linux with an nvidia GPU it should be faster. Still way slower than hand constructed CUDA code.

I think you can run TF on Windows.

Note: if the doublings parameter is 0, then the algorithm is a straight run and the intermediate files produced can be interpreted as the story from beginning to end. If doublings is > 0 then it's a slightly faster algorithm but the intermediates aren't so nicely ordered. So set doublings to zero (and read the accompanying comments) if you want to make a video of the evolution of a sandpile.