This impromptu talk introduces CubeCL, a Rust framework for writing portable parallel compute kernels targeting CPUs and GPUs. It explores how a shared programming model maps to different hardware and the challenges of supporting both execution environments. Drawing on the development of a JIT CPU backend, the session highlights compiler design choices and lessons from experimenting with LLVM, MLIR, and Pliron. Attendees will gain an overview of CubeCL's approach to portable computation and the trade-offs involved in bringing GPU-style parallelism to CPUs.
