compute_rasterizer VS webgpu-compute-rasterizer

Compute rasterization has the potential to be faster if you do it correctly but unless you're drawing a lot of points (hundreds of millions or even billions) just use GL_POINTS.

Write the application first and find out; you're not going to be bottlenecked by the GPU for a very long time, as even several hundred bodies will become a major CPU bottleneck until you implement some acceleration structures. If you do get to the point of being able to simulate more bodies than you can render (for what it's worth, I've never gotten close even with a 2-body approximation), you can explicitly double-buffer your application or look into persistent mapping. Beyond that, you can actually achieve faster point rendering by writing your own compute-based point cloud rasterizer, but obviously this is not something to take on lightly.

If you find traditional OpenGL techniques too slow, you can gain a 10-100 fold speedup by using compute shaders to manually rasterize your point cloud: https://github.com/m-schuetz/compute_rasterizer

Another example: https://github.com/m-schuetz/compute_rasterizer

well said! I think UE5 does use compute shaders for very small triangles. The other use case I know of is rendering dense point clouds (with more than 100 million points in the scene, see: https://github.com/m-schuetz/compute_rasterizer)

This is beyond your needs right now (you should take care of the basic first), but you might be interested in this. Turns out, a compute shader can rasterize big point clouds way more efficiently than using GL_POINTS, but don't look into that until you've identified a performance issue first.

OmarShehata How to Build a Compute Rasterizer with WebGPU https://github.com/OmarShehata/webgpu-compute-rasterizer/blob/main/how-to-build-a-compute-rasterizer.md Clearly I am not good enough with WebGpu to do it. Anyone interested?

Thank you, possibly, but ideally I would not want to write the dispatch logic myself because I would not know how to dispatch "triangles". I know that it can be done (see https://github.com/OmarShehata/webgpu-compute-rasterizer/blob/main/how-to-build-a-compute-rasterizer.md) but the thing is, I am fine with pretty much everything that a render pipeline would do for me; attribute setup, primitive assembly, interpolation, scanline conversion... I just don't want to write to a target.

I would be curious to benchmark an equivalent traditional renderer and be able to toggle back and forth on the web page just out of curiosity, I have an issue here on GitHub if anyone is interested in exploring that: https://github.com/OmarShehata/webgpu-compute-rasterizer/issues/1