Self-nominating collision2.

It's an experiment in 2D GPU-based particle system simulation powered by wgpu and the awesome wgsl-bindgen.

Everything is done on the GPU: no CPU<->GPU data transfer after the data is uploaded at startup.

• SDF-based rendering: 6 million particles at 60 FPS on Radeon RX 560. 6M is only rendering though, physics performance is nowhere near that.
• Grid-based collision broad phase, currently implemented using WGSL atomics (working on a more efficient prefix sum implementation utilising subgroup operations).
• Efficient collision narrow phase employing bitcasts and CAS loop to store f32 forces.
• Symplectic Euler integrator.
• "Black holes"! Only gravitational attraction and frame dragging (Lense-Thirring) are implemented, no Minkowsky space-time or anything like that. Check out integrator.rs and the corresponding integrate.wgsl for details.
• Zoom with mouse wheel.
• Headless mode for benchmarking and recording purposes
• Can record the simulation to a series of PNGs
• Simulation step and other parameters are configurable via environment variables corresponsing to the fields of the Config struct in src/config.rs.

Notes

• Grid-based broad phase assumes particles of similar sizes. Performance degrades quickly if they differ significantly. There was a BVH-based broad phase, but I removed it for now. Look up the repo history if you want to check it out.
• The code barely has any comments, because it requires some proficiency in GPU programming and the lion's share of it is self-explanatory if you have it. And I'm lazy.
• Indirect dispatch is utilised where appropriate to avoid CPU<->GPU data transfer.
• Not published on crates.io since it's not a tool.
• Only tested on Manjaro Linux with RADV Vulkan driver on RX 560.
• Definitely has bugs. If I had a dollar for each one I've found over the course of development, I could easily fund my beer-brewing hobby 😆.

Any feedback is welcome, especially from GPU folks.