Hi everyone,
I would like to submit a Roundtable proposal for EuroLLVM '23. This is just small heads-up to see whether there would be any interest and whether there’s something specific that you’d like to see discussed.
Title: Targeting CPUs from ML frameworks/compilers
Some terminology to avoid confusion (these terms can mean different things depending on context):
- frontend - consumes high level representation and generates LLVM IR,
- backend - consumes LLVM IR and generates machine code for the specified CPU.
There’s quite a few of us using MLIR and MLIR-based compilers (e.g. IREE/OpenXLA) that target various CPU backends in LLVM. There are other frontends, outside of LLVM, as well. In fact, Elen Kalda will be giving a presentation based on her experience with TVM. Is there anything that we can do to integrate both frontend and backend technologies better?
I am particularly interested in vectorisation and various CPU extensions, e.g. Arm’s Scalable Matrix Extension [1] as well as scalable vectors in general. I work for Arm, hence referring to SME, but this roundtable is meant to cover all CPUs that folks care about. Can you think of any other topics that you’d like to discuss?
Ideally, we would gather together backend and frontend developers and discuss what the ideal interface should be or what might be the barriers to enabling new CPU extensions. How does it sound?
Hope to see you in Glasgow!
-Andrzej