Horton

Re: Popcount optimization using AVX512

Nathan Bossart <nathandbossart@gmail.com>

View thread
From: Nathan Bossart <nathandbossart@gmail.com>
To: Ants Aasma <ants.aasma@cybertec.at>
Cc: Alvaro Herrera <alvherre@alvh.no-ip.org>, "Amonson, Paul D" <paul.d.amonson@intel.com>, Tom Lane <tgl@sss.pgh.pa.us>, David Rowley <dgrowleyml@gmail.com>, Andres Freund <andres@anarazel.de>, "Shankaran, Akash" <akash.shankaran@intel.com>, Noah Misch <noah@leadboat.com>, Matthias van de Meent <boekewurm+postgres@gmail.com>, "pgsql-hackers@lists.postgresql.org" <pgsql-hackers@lists.postgresql.org>
Date: 2024-04-02T15:53:01Z
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. Fix __attribute__((target(...))) usage.

  2. Use __attribute__((target(...))) for AVX-512 support.

  3. Fix code for probing availability of AVX-512.

  4. Optimize visibilitymap_count() with AVX-512 instructions.

  5. Optimize pg_popcount() with AVX-512 instructions.

  6. Inline pg_popcount() for small buffers.

  7. Avoid function call overhead of pg_popcount() in syslogger.c.

  8. Refactor code for setting pg_popcount* function pointers.

  9. Inline pg_popcount{32,64} into pg_popcount().

  10. Remove MSVC scripts

  11. Use ARMv8 CRC instructions where available.

  12. Use Intel SSE 4.2 CRC instructions where available.

Attachments

On Mon, Apr 01, 2024 at 05:11:17PM -0500, Nathan Bossart wrote:
> Here is a v19 of the patch set.  I moved out the refactoring of the
> function pointer selection code to 0001.  I think this is a good change
> independent of $SUBJECT, and I plan to commit this soon.  In 0002, I
> changed the syslogger.c usage of pg_popcount() to use pg_number_of_ones
> instead.  This is standard practice elsewhere where the popcount functions
> are unlikely to win.  I'll probably commit this one soon, too, as it's even
> more trivial than 0001.
>
> 0003 is the AVX512 POPCNT patch.  Besides refactoring out 0001, there are
> no changes from v18.  0004 is an early proof-of-concept for using AVX512
> for the visibility map code.  The code is missing comments, and I haven't
> performed any benchmarking yet, but I figured I'd post it because it
> demonstrates how it's possible to build upon 0003 in other areas.

I've committed the first two patches, and I've attached a rebased version
of the latter two.

> AFAICT the main open question is the function call overhead in 0003 that
> Alvaro brought up earlier.  After 0002 is committed, I believe the only
> in-tree caller of pg_popcount() with very few bytes is bit_count(), and I'm
> not sure it's worth expending too much energy to make sure there are
> absolutely no regressions there.  However, I'm happy to do so if folks feel
> that it is necessary, and I'd be grateful for thoughts on how to proceed on
> this one.

Another idea I had is to turn pg_popcount() into a macro that just uses the
pg_number_of_ones array when called for few bytes:

	static inline uint64
	pg_popcount_inline(const char *buf, int bytes)
	{
		uint64		popcnt = 0;

		while (bytes--)
			popcnt += pg_number_of_ones[(unsigned char) *buf++];

		return popcnt;
	}

	#define pg_popcount(buf, bytes) \
		((bytes < 64) ? \
		 pg_popcount_inline(buf, bytes) : \
		 pg_popcount_optimized(buf, bytes))

But again, I'm not sure this is really worth it for the current use-cases.

-- 
Nathan Bossart
Amazon Web Services: https://aws.amazon.com