What's the plan?

[Jakdaw]

Hi,

This project looks very exciting - do you mind me asking what your intentions are for it?

Is it just a proof of concept hobby project that it would be fun to extend to further architectures? Handy for those who want to learn more about compilers and GPUs?

or a means to help people who want to learn CUDA but don't have any Nvidia hardware? (aside: do you want to encourage those without Nvidia hardware to learn CUDA rather than something more open?)

or are there some existing CUDA code bases that you'd like to see being BarraCUDA being able to compile & run? Maybe something like the popsift project as a means towards having the alicevision stack running on more hardware, since there's little alternative to NV at present -- but this might require more in the direction of textures/surfaces since popsift makes use of these?

Just interested to know!

[Zaneham]

Hello @Jakdaw,

Thank you for your questions and apologies for the late reply.

Software is a hobby of mine however I do find quite a bit of enjoyment making things like this. My current plan is trying to get GPT2 kernels from Tinygrad to compile. Currently having some great success and looking at merging my branch soon.

In terms of CUDA itself this compiler is language agnostic. Each backend doesn't know the language the frontend is parsing and lexing. I'm looking at integrating more CUDA/c++ into this compiler and then I'll work on adding an OpenCL frontend. You're right on textures/surfaces currently I'm working on a way to get those working but I work on it when I have time.

On other architectures I'm absolutely open to anyone submitting a pr or discussing with me what GPU's they want supported. The more people that can use this and hopefully run programs on older hardware has the dual benefit of being a fun thing to make/read about and improving access to people with older/different hardware.

Thanks again for your question!

:-)