> it basically requires a completely separate (re)write to properly take advantage of. Why? At a higher level of abstraction, you can dispatch simd instruction

> it basically requires a completely separate (re)write to properly take advantage of. Why? At a higher level of abstraction, you can dispatch simd instructions at the max width available. At least, that's how I work with vectorized code. Still see gains on avx512.