Skip to content

Latest commit

 

History

History
36 lines (29 loc) · 1.93 KB

README.md

File metadata and controls

36 lines (29 loc) · 1.93 KB

JAX OpenGL Height-Field Water Simulator

Simulation YouTube Video

This is a simple height-field water simulator written in Python using JAX and OpenGL. It is based on the awesome work of Matthias which was implemented using three.js.

The main differences:

  • Python
  • Uses JAX so simulator code can run on CPU, GPU, METAL and TPU (TPU untested)
  • Uses pure OpenGL with no external libraries

This application has been tested with Python 3.10 on CPU and METAL GPU on MacOS. It should work on any platform that supports JAX and OpenGL. On my MacBook Pro (M1) it runs at 350fps with 101x101 grid when using JAX CPU. On a Linux x86_64 machine with a NVIDIA 1080, I was also able to achieve ~350fps with 101x101 grid when using JAX CUDA. Tested with an NVIDIA 4090 and Ryzen 7950X3D, I achieved ~770 FPS with GPU and ~363 FPS with CPU-only. Since a32fe50, I was able to profile and optimise the code a bit and can now achieve ~2000 FPS with the NVIDIA 4090.

Run Instructions

First make sure you have Python. You may want to change how you install JAX so that you can run on CPU, GPU or METAL GPU. By default I have assumed you want to run the simulator code CPU which is easiest and fastest.

pip install git+https://github.com/cemlyn007/water-simulator "jax[cpu]"
water-simulator --n 101

Ubuntu

To install this to run, I found that I needed the following packages from apt:

apt install python3-dev gcc g++

Profiling

Using JAX native profiling

water-simulator --n 501 --enable_jax_profiling

Using nsys

nsys profile --capture-range=cudaProfilerApi --cuda-graph-trace=node --capture-range-end=stop -e JAX_TRACEBACK_IN_LOCATIONS_LIMIT=-1,DISPLAY=:1 water-simulator --n 501 --enable_nvidia_profiling