Horton

Re: Improve CRC32C performance on SSE4.2

John Naylor <johncnaylorls@gmail.com>

View thread
From: John Naylor <johncnaylorls@gmail.com>
To: "Devulapalli, Raghuveer" <raghuveer.devulapalli@intel.com>
Cc: Nathan Bossart <nathandbossart@gmail.com>, "pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>, "Shankaran, Akash" <akash.shankaran@intel.com>
Date: 2025-02-21T01:28:08Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Include _mm512_zextsi128_si512() in AVX-512 configure probes.

  2. Properly fix AVX-512 CRC calculation bug

  3. Workaround code generation bug in clang

  4. Compute CRC32C using AVX-512 instructions where available

  5. Inline CRC computation for small fixed-length input on x86

  6. Be more paranoid in configure's checks for CRC and POPCNT intrinsics.

On Fri, Feb 21, 2025 at 1:24 AM Devulapalli, Raghuveer
<raghuveer.devulapalli@intel.com> wrote:
>
>
> > Now, there is no equivalent on MSVC and workarounds are fragile [1].
> > Maybe we could only assert initialization happened for the backend and for
> > frontend either
> > - add a couple manual initializations to to the frontend programs where we don't
> > want to lose performance for non-gcc/clang.
> > - require CRC on x86-64 MSVC since Windows 10 is EOL soon, going by Thomas
> > M.'s earlier findings on popcount (also SSE4.2) [2]
> >
> > The first is less risky but less tidy.
>
> Agree, let me think about this but not sure if I have any useful suggestions here. MSVC is unfortunately not my strong suit :/

Here's another idea to make it more automatic: Give up on initializing
every capability at once. The first time we call CRC, it will be
uninitialized, so this part:

if (pg_cpucap & PGCPUCAP_CRC32C)
  return COMP_CRC32C_HW(crc, data, len);
else
  return pg_comp_crc32c_sb8(crc, data, len);

...will call the SB8 path. Inside there, do the check:

#if defined(HAVE_CRC_RUNTIME)
// separate init bit for each capability
if (unlikely(pg_cpucap & PGCPUCAP_CRC32C_INIT == 0))
{
  pg_cpucap_crc32c(); // also sets PGCPUCAP_CRC32C_INIT
  if (pg_cpucap & PGCPUCAP_CRC32C)
    return COMP_CRC32C_HW(crc, data, len);
}
#endif
// ...fallthrough to SB8

--
John Naylor
Amazon Web Services