AVX512 plans? about xsimd HOT 3 CLOSED

Hey @bluescarni thanks for posting. We do plan on supporting AVX512. There is ongoing work on this being done this week! You can check out #101 for more details.

If you have experience with AVX512 and this type of work, don't hesitate to comment on the implementation.

from xsimd.

Would you like to review my PR? #101

It's almost done, I'm working on fixing all the tests today (basic and basic math already passing) and then it should be merged probably tomorrow!

I'll also take a look at yours to find if there are some ideas! One difference is the way we implemented the batch book (i am keeping the "mask")

Cheers and looking forward!

from xsimd.

@SylvainCorlay @wolfv That's great news! Looking forward to the merge of #101.

As I mentioned, I am just a newbie SIMD-wise, so it's unlikely you will find something interesting in my implementation :) I did notice the changes from AVX to AVX512 regarding the comparison intrinsics, and I kinda tried to shoehorn the AVX semantics in the AVX512 wrappers (which resulted in some overhead). I think working directly with masks (as you are doing in #101) is much better, so +1 from me for what it's worth.

What I can do is to run my code on the Skylake-512 server with both my partial implementation and the implementation from #101, so maybe I can provide some performance cross-checking on real hardware. I'll ping back when I am done, closing this in the meantime.

from xsimd.