[11u] RFR 8245512: CRC32 optimization using AVX512 instructions

[Andrew Haley]

On 1/7/21 10:26 AM, Doerr, Martin wrote:

> I think 2x performance improvement for CRC32 computations is a
> good argument.

It's not 2x for everybody, just the relatively small number of
AVX-512 users, and as we've seen AVX 512 throttling can slow down
other operations. That 2x isn't free: it is accompanied by
increased power consumption.

We don't know the future of AVX-512. Rumours say it's not going
to be included in Alder Lake-S, Intel’s first 10nm mainstream
desktop processor. I don't think the 10900K Comet Lake supports
it either. But I admit it does look like Xeons are going to
support AVX-512 for some time, so perhaps.

The question is this: is it appropriate to backport a
special-case accelerator for CRC32C that works on a smallish
segment of deployed systems to an LTS release of OpenJDK?
I could go either way.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671