Great read about the tradeoffs of Mesa building its own GPU-focused compiler vs. using LLVM https://www.jlekstrand.net/jason/blog/2022/01/in-defense-of-nir/
Great read about the tradeoffs of Mesa building its own GPU-focused compiler vs. using LLVM https://www.jlekstrand.net/jason/blog/2022/01/in-defense-of-nir/
@glaebhoerl I know this is one of the points he makes in passing in the article, but LLVM doesn't even model SIMT convergence correctly. So besides not having much shared GPU-specific open source code to build on, you're also starting from a semantic disadvantage rather than neutral ground.