metal-LIC-demo-small.mp4
GPU implementation of the line integral convolution (LIC) algorithm on Apple Silicon, written in Swift + Metal. Capable of real-time performance at 4k resolution (~40 fps in simple tests on my Apple M1 Pro machine).
It takes an input texture and a vector field, traces streamlines via RK2 integration with a Hann-windowed kernel, and outputs a grayscale LIC image. The entire pipeline runs as a Metal compute shader with no CPU readback.
Thanks to the performance of this implementation, it is possible to assign time-dynamics to the input texture that is getting convolved, giving a host of possible dynamic visualizations of the (static) vector field.
Measured GPU time on M1 Pro, single pass, kernel length 30 pixels:
| Resolution | Uniform field | Vortex field |
|---|---|---|
| 1080p | 238 fps | 179 fps |
| 2048x2048 | 118 fps | 88 fps |
| 3840x2160 | 60 fps | 46 fps |
Throughput scales linearly with kernel length.
Swift Package Manager library. Add as a dependency:
.package(url: "https://github.com/<you>/metal-LIC", branch: "main")The package provides LICEncoder for single-shot encoding and LICDispatcher for pipelined multi-buffered dispatch.
LICDemo is a SwiftUI app that pairs the LIC engine with a forest fire cellular automaton to give time-dependent noise.
This lets you run real-time simulations. Run it from Xcode or swift run LICDemo.