Plan for rust-gpu and rust-cuda

[JackWolfard]

In the transition announcement, it was announced

If instead you wish to stick to the NVIDIA ecosystem, stay tuned as the Rust CUDA project is in the process of being rebooted and possibly integrated with Rust GPU.
...
rust-cuda: Compiles unmodified Rust code to NVVM (CUDA) so that it can run on the GPU. Currently unmaintained but stay tuned!

I am interested in hearing more about this reboot effort and who might be the new maintainers. Is there development already happening in the open somewhere?

[JackWolfard]

@LegNeato @eddyb can you comment?

[LegNeato]

Sure! I have recently become a maintainer of rust-cuda and behind the scenes I have been talking to the previous maintainer as well as the person working on the rustc ptx backend about the best way forward for the project technically.

rust-cuda in the past has looked at the compiler infra bits rust-gpu has and mirrored some of them. It might make sense to merge codebases where we can and have cuda vs vulkan be more like arm vs x86 in rustc (that is, separate arch/platform specific implementations sharing overall project and infrastructure).

We've been focused on this project's transition before working on rebooting rust-cuda, but in a couple weeks once we get a release out and hit our stride I'll start to focus on rust-cuda.

[rydb]

Question. Why CUDA?

[JackWolfard]

Why CUDA?

My original question was b/c CUDA often has the most advanced GPU compute features of any language. If I wanted to make use of something like TensorCores, then they're directly programmable with CUDA, but in SPIR-V or OpenCL or another language, they're a little harder to access or are more limited in feature set.