conference logo

Playlist "BOB Konferenz 2021"

SIMD in Higher Level Programming Languages

Matthias Wahl

Modern processors support Single Instruction Multiple Data techniques since 1997 when intel introduced MMX. Every modern CPU has physical registers for performing vectorized SIMD operations on up to 512 bits with a single instruction. There is a part of the CPU which is most of the time only used by compiler optimizations invisible and unbeknownst to users or by low-level systems programmers comfortable with assembly and processor architecture. Those magicians have gone through unspeakable pain and know ancient dark magic from unreadable runes chiseled into unreachable cave walls.

This is a sad state. Given our very CPUs could do so much more and be up to 8x as efficient. We actively waste resources.

Having established that we need to do something about it, Matthias Wahl explores what we can do to expose SIMD in higher level languages, what steps are necessary to do something about it, and shows us what is being done about it in existing and emerging languages.