Re: CRC32C Parallel Computation Optimization on ARM

On Thu, Mar 13, 2025 at 12:50 AM Devulapalli, Raghuveer
<raghuveer(dot)devulapalli(at)intel(dot)com> wrote:
>
> > > Intel has contributed SSE4.2 CRC32C [1] and AVX-512 CRC32C [2] based on
> > similar techniques to postgres.
> >
> > ...this is a restatement of facts we already know. I'm guessing the intended
> > takeaway is "since Intel submitted an implementation to us based on paper A,
> > then we are free to separately also use a technique from paper B (which cites
> > patents)".
>
> Yes.
>
> > The original proposal that started this thread is below, and I'd like to give that
> > author credit for initiating that work
>
> Yup, that should be fine.

Thank you for confirming. I've attached v10, which has mostly
polishing and comment writing, and a draft commit message. The lookup
table and software carryless multiplication routine are still in
pg_crc32c_sb.c , which is now built unconditionally. That's good
foreshadowing of future pclmul/pmull support, as I've found building
that file everywhere makes some things simpler anyway. That file has
become a bit of a misnomer, and I've thought of renaming it to
*_common.c or perhaps *_fallback.c , since the addition from this
patch is still kind of a fallback where we won't have the hardware
needed for faster algorithms, as discussed elsewhere.

0002-3 puts the relevant parts into a header so that the hardware
details can be abstracted away. These would be squashed, but I've kept
them separate here for comparison.

--
John Naylor
Amazon Web Services