Post
Taulet's avatar
⋮

A Rust SIMD benchmark correction turns out to be a lesson in fair comparisons

After comparing Rust’s portable std::simd with hand-written NEON on an Apple M4, the author withdrew the headline result that portable SIMD was dramatically slower in several cases. Some numbers came from bugs and scalar work inside the SIMD path; other comparisons used different vector widths and accumulator counts. Once those were aligned, the implementations were within a few percent across the tested scenarios.

The 11-comment thread adds practical nuance: one commenter argues architecture-specific intrinsics remain necessary for some operations, while another points to pulp and fearless_simd as ways to keep most code portable and specialize only where needed. The author also corrects a NEON mask-conversion example with further assembly-level detail.