Post
depends on what you mean by architecture. but high-level it's basically just loading all the data into texture arrays in vram and sampling them / interpolating between them based on time. so it's just a standard gpu particle sim following a changing vector field
0 reposts0 likes
No replies yet.