Maynard Handley (name99.delete@this.name99.org) on July 12, 2020 10:30 am wrote:
> Jukka Larja (roskakori2006.delete@this.gmail.com) on July 12, 2020 8:51 am wrote:
> >
> > What sort of things are you seeing in regular code that you'd
> > get some normal programmers casually converting to SIMD?
> >
> > -JLarja
>
> I would have thought parallel small-table lookup is a not uncommon sort of task. Then you
> get various things that branch off from that basic idea, like various text processing.
> But, as was mentioned, you do have to somewhat reconceptualize what you are
> doing; it's frequently not a trivial rewrite of an existing kernel.

I have only a very, very vague idea of what that means and can't think of any way to trivially simdify typical lookups (which are usually either find from ordered/hashmap or scan a vector for value). If it's too hard, it's not practical. By default, programmers haven't ever done any SIMD programming. I have, but almost none for work (and what I did was mostly debugging why SSE2 and Neon gave different results).

(As a side note, I don't think we do any performance sensitive text processing even in any of the tools we've written. It's sadly very typical that in this kind of conversations the examples people come up with aren't relevant (at least not obviously), even if the speedup would be an order of magnitude.)

-JLarja