Reproduce “Silent Degradation during Auto-Vectorization of bf16/fp16 on RISC-V Under LLVM” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Vaishnavi Kouru, Robert Chun

The authors compare LLVM-compiled multiply-accumulate loops using fp32, int8, bf16, and fp16 inputs, testing floating-point types with and without RISC-V vector extensions. Without extensions, bf16 and fp16 loops are reported as vectorized but perform extra work. Under QEMU emulation, baseline builds execute 17.0× and 75.9× as many dynamic guest instructions as enabled builds; these results cover tested kernels, not hardware speed.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.