@waz @soaska There are other reasons to be cautious about using AVX - older CPUs that support it are forcibly downclocked when the instructions are used so you'd better make sure that your vectorised loops are delivering enough benefits for the 10-20% penalty for all the rest of the code

https://stackoverflow.com/questions/35663635/why-do-processors-with-only-avx-out-perform-avx2-processors-for-many-simd-algori/44353690#44353690

https://stackoverflow.com/questions/56852812/simd-instructions-lowering-cpu-frequency