vuda VS autograph

There is actually a project on Github which tries to mimic the CPU-side code from CUDA to Vulkan projects, called VUDA.

There is something, though. I remember looking at VUDA which, while it's not a drop-in replacement, is a pretty good proof of concept

I'm working on a machine learning library https://github.com/charles-r-earp/autograph implemented in Rust that uses rust-gpu to compile Rust compute shaders to spirv, and then gfx_hal to target metal and dx12. Training performance is currently about 2x slower than pytorch (cuda) on my laptop but I've made significant progress recently and I am targeting 1.5x. While rust-gpu itself has it's own restrictions, it does support inline spirv assembly, which provides direct access to operations not provided in its std lib, thus it's lower level than GLSL. For example, it should be possible to target cuda tensor cores via cooperative matrix operations (I believe Metal supports these as well but this may not be implemented in spirv-cross and certainly isn't in naga). Once I have things a bit more stabilized I'd like to provide more examples, like porting from cuda / opencl, but I'm still figuring out patterns like how to work with 16 and 8 bit types in a nice and portable way.

autograph v0.1.0

Working on autograph, for machine learning and neural networks. Unlike CUDA / HIP it's threadsafe, but doesn't expose low level things like multiple streams. Most of the shaders are glsl but I'm now using rust_gpu for pure rust gpu code.

Maybe not for learning but as inspiration I have to plug this amazing effort for ML with (vulkan) shaders: https://github.com/charles-r-earp/autograph

Maybe you'd be interested in checking out my library, https://github.com/charles-r-earp/autograph?