Is there a diagram for the backends, and what are they compiled to?

[iampi31415]

Just took a look at this post and was wondering whether burn gpu mostly shaders within the WGPU API that compile to the actual platform backend (SPIR-V, Metal, etc.) ?

I get that you also support other backends but I wonder whether there is some flow diagram or chart that gives an overview?

For example, I wondered if something similar is true for the burn project as for the link above:

The compute logic is fully generic and shared across all targets, running the same code on both CPU and GPU.

[nathanielsimard]

There is a table in CubeCL that might help you: https://github.com/tracel-ai/cubecl

Platform	Runtime	Compiler	Hardware
WebGPU	wgpu	WGSL	Most GPUs
CUDA	CUDA	C++ (CUDA)	NVIDIA GPUs
ROCm	HIP	C++ (HIP)	AMD GPUs
Metal	wgpu	C++ (Metal)	Apple GPUs
Vulkan	wgpu	SPIR-V	Most GPUs on Linux & Windows

[iampi31415]

Thanks yes I meant that sort of table! Though WGSL is not a compiler, do you mean Naga?

I think one needs two translations one is the API commands say from WebGPU to Metal (by wgpu) and the other translation are the shaders say from WGSL to MSL (by Naga). Or I am misunderstanding it.

[nathanielsimard]

It's our compiler (transpiler) that outputs WGSL.

[iampi31415]

Interesting, so you have a compiler from Rust using the Burn API to WGSL, and I guess the same for each shader code. And then each part is run with the runtime.

So for example, the neural nets are converted to WGSL, I interpret.

[laggui]

So for example, the neural nets are converted to WGSL, I interpret.

That might be a bit of an oversimplification, the whole neural net is not converted to WGSL. Here's a more accurate interpretation:

The model can be seen as a graph of operations (e.g., matmul, conv, etc.), each mapped to some GPU kernels compiled by CubeCL to WGSL/CUDA/etc. The backend runtime executes them to run the model. (Side note: with fusion, multiple ops can be merged into a single kernel)