Hey, I am actually working on making this compatible on earlier AMD's as well because I have an old gaming laptop with an RX5700m which is GFX10. I'm reading up on the ISA documentation to see where the differences are, and I'll have to adjust some binary encoding to get it to work.
I mean this with respect to the other person though please don't vibe code this if you want to contribute or keep the compiler for yourself. This isn't because I'm against using AI assistance when it makes sense it's because LLMs will really fail in this space. Theres's things in the specs you won't find until you try it and LLMs find it really hard to get things right when literal bits matter.
I really like the minimal approach you've taken here - it's refreshing to see this built completely from the ground up and it's clearly readable and for me, very educational.
But help me understand something. BarraCuda does its own codegen and therefore has to implement its own optimisation layer? It's increbibly impressive to get "working" binaries, but will it ever become a "viable" alternative to nvidia's CUDA if it has to re-invent decades of optimisation techniques? Is there a performance comparison between the binaries produced by this compiler and the nvidia one? Is this something you working on as an interesting technical project to learn from and prove that this "can be done"? Or are you trying to create something that can make CUDA a realistic option on AMD GPUs?
If it helps: I'm actively considering how to make LLMs more helpful in GPU programming. Recently I converted the RDNA 3.5 ISA manual from PDF to Markdown for LLM retrieval (because I have a Strix Halo). Notably, I converted every bitfield diagram to plain text table and manually checked its correctness. Everyone can comment there if they find a typo. I guess someone can also do it for RDNA 2 and other manuals.
I mean this with respect to the other person though please don't vibe code this if you want to contribute or keep the compiler for yourself. This isn't because I'm against using AI assistance when it makes sense it's because LLMs will really fail in this space. Theres's things in the specs you won't find until you try it and LLMs find it really hard to get things right when literal bits matter.