1868103 - Use avx512vnni instructions to speed up ML model inference

Status: RESOLVED FIXED

(Firefox :: Translations, enhancement)

Description

[Marco Castelluccio [:marco]]

If bug 1868102 shows avx512vnni is supported on many of our users' machines, it makes sense to add support for it.

Comment 1

[Gian-Carlo Pascutto [:gcp]]

I suspect we have more machines with AVX512VNNI than we have with AVXVNNI (cfr bug 1868949), because AVXVNNI is much newer (even though it's the same operation, sigh). I wonder if there was a misunderstanding here that lead to bug 1871151.

The real question is, is it best to run the AVX512VNNI code in 512-bit mode or 256-bit mode (the latter of which is still different from AVXVNNI!) :-)

Comment 2

[[:sergesanspaille]]

gemmology support on going: https://github.com/mozilla/gemmology/pull/13

Comment 3

[Marco Castelluccio [:marco]]

All dependencies fixed.