On: tolower() with AVX-512

[TinkersW]

I don't get saying mask operations are more important than width?

Mask operations can be trivially emulated with vblend, it is one extra instruction..

Width can't be emulated, you just are stuck running half speed.

This take keeps getting repeated, but doesn't appear to be backed up by reality.

Intel hasn't even put AVX10 on their upcoming chips(skymont), so it appears to be going nowhere.

[account42 → TinkersW]

> Mask operations can be trivially emulated with vblend, it is one extra instruction..

For unaligned loads where you can't guarantee that the entire vector is on a mapped page?

[fanf2 → TinkersW]

The important feature of AVX-512 demonstrated in my blog post is masked loads and stores, which can't be emulated with vblend.