vuh VS cuda-api-wrappers

The only one that is relatively lightweight (and doesn't have dependencies) is vuh (https://github.com/Glavnokoman/vuh) that looks unsupported.

Don't try to use it directly, you need a higher level library.

https://github.com/KomputeProject/kompute seems like what's taking off. There are also smaller ones like https://github.com/Glavnokoman/vuh etc.

1. This implements the clunky C-ish API; there's also the Modern-C++ API wrappers, with automatic error checking, RAII resource control etc.; see: https://github.com/eyalroz/cuda-api-wrappers (due disclosure: I'm the author)

2. Implementing the _runtime_ API is not the right choice; it's important to implement the _driver_ API, otherwise you can't isolate contexts, dynamically add newly-compiled JIT kernels via modules etc.

3. This is less than 3000 lines of code. Wrapping all of the core CUDA APIs (driver, runtime, NVTX, JIT compilation of CUDA-C++ and of PTX) took me > 14,000 LoC.

> since the underlying API's are still C/C++,

If the use of GPUs is via CUDA, there are my https://github.com/eyalroz/cuda-api-wrappers/ which are RAII/CADRe, and therefore less unsafe. And on the Rust side - don't you need a bunch of unsafe code in the library enabling GPU support?

Given your criteria, you might want to consider (modern) C++.

* Fast - in many cases faster than Rust, although the difference is inconsequential relative to Python-to-Rust improvement I guess.

* _Really_ utilize CUDA, OpenCL, Vulcan etc. Specifically, Rust GPU is limited in its supported features, see: https://github.com/Rust-GPU/Rust-CUDA/blob/master/guide/src/... ...

* Host-side use of CUDA is at least as nice, and probably nicer, than what you'll get with Rust. That is, provided you use my own Modern C++ wrappers for the CUDA APIs: https://github.com/eyalroz/cuda-api-wrappers/ :-) ... sorry for the shameless self-plug.

* ... which brings me to another point: Richer offering of libraries for various needs than Rust, for you to possibly utilize.

* Easier to share than Rust. A target system is less likely to have an appropriate version of Rust and the surrounding ecosystem.

There are downsides, of course, but I was just applying your criteria.

https://github.com/eyalroz/cuda-api-wrappers

I try to address these and some other issues.

We should also remember that NVIDIA artificially prevents its profiling tools from supporting OpenCL kernels - with no good reason.

On the host side however - the API you use to orchestrate execution of kernels on GPUs, data transfers etc. - the official API is very C'ish, annoying and confusing. I have written C++'ish wrappers for it which many enjoy but are of course not officially supported or endorsed: https://github.com/eyalroz/cuda-api-wrappers

Oh, I really wish those hackers would release the sources rather than pursue their dumbass crypto-mining demands... "We decided to help mining and gaming community" - hurting the gaming community, helping the get-rich-quick "community".

My own C++ wrappers for the CUDA APIs (shameless self-plug: https://github.com/eyalroz/cuda-api-wrappers/) would really benefit a lot from behind-the-curtains access to the driver; and even if I just know how the internal logic of the driver and the runtime works, without actually being able to hook into that logic - I would already be able to leverage this somewhat in my design considerations.

As a person making a living from being the "GPU guy" - I definitely agree.

The ecosystem around AMD GPUs is quite small - and now that they seem to have abandoned OpenCL (possibly not their own fault though) - even that is put into question.

But things are bad even on the NVIDIA side. Example of how bad: I had to write my own C++ bindings for the CUDA runtime API (https://github.com/eyalroz/cuda-api-wrappers/). You'd think they would have that after 13 years of CUDA being available, right? Wrong. I repeatedly tried to pitch this to them, but they seem to suffer from the "Not Invented Here" syndrome (https://learnosity.com/not-invented-here-syndrome-explained/). This despite me having a lot of respect for people like Mark Harris, Bryce Lelbach, Duane Merrill et alia, and their work.

You're also rights about the "two kinds of brains" - or rather, it's not clear to me that the brains creating the silicon and the brains creating the software are in close enough cooperation.

By the way - it is possible to extract a pretty distribution of CUDA to justify run 20 lines of GPGPU code, from their installer. But they won't be bothered to package this nicely for you.

Try Vulcan, or OpenCL. There are tons of wrappers for CUDA to make coding simpler ie https://github.com/eyalroz/cuda-api-wrappers