See original MPR paper by Matt Keeter: Massively Parallel Rendering of Complex Closed-Form Implicit Surfaces. Also see reference implementation in CUDA .
This is a Metal port of MPR, with following goals:
- Explore MPR implementations outside of CUDA (it's easier to port Metal -> WebGPU than from CUDA)
- Benchmark framerate and energy efficiency of MPR in macOS and iOS environments
- Experiment integrating implicit shapes into traditional rendering pipelines such as SceneKit or RealityKit.
It's recommended to read Keeter's MPR paper and review the CUDA reference implementation first (see above). This sections documents deviations from the reference implementation.
- No IEEE 754 floating-point arithmetic support in interval math. This is unsupported in Metal 2.
- No tile packing. I'm not sure if gains from less threads is worth the atomic counters and round trips between CPU <-> GPU. More benchmarking is needed here. Removing tile packing simplified a number of things:
- Tile struct is now just a tape pointer. No need to track position and next.
- No need to unpack xy from position. Just use 2/3D thread index.
- Tile buffer size and thread grid size can be pre-calculated. All stages of tile evaluation can be packed into a single command buffer.
- interval math
- 2D tiles evaluation
- 2D pixel evaluation
- 3D tiles evaluation
- 3D voxel evaluation
- RealityKit integration