From: Nathan Bossart <nathandbossart@gmail.com>
To: John Naylor <johncnaylorls@gmail.com>
Cc: Ants Aasma <ants@cybertec.at>
Cc: pgsql-hackers@postgresql.org
Subject: Re: add AVX2 support to simd.h
Date: Tue, 9 Jan 2024 10:20:09 -0600
Message-ID: <20240109162009.GA3033933@nathanxps13> (raw)
In-Reply-To: <CANWCAZYsnwxT2YjUQBWcc2QqOoE6xyEtTkqD6kw2pKCX1xevrg@mail.gmail.com>
References: <20231129171526.GA857928@nathanxps13>
<CANWCAZZgOS+f2W8+bFwHjYYT-6Fx4yP6dFpUy4oLfbXSgfTKLw@mail.gmail.com>
<20240108173715.GC2611898@nathanxps13>
<CANWCAZYsnwxT2YjUQBWcc2QqOoE6xyEtTkqD6kw2pKCX1xevrg@mail.gmail.com>
On Tue, Jan 09, 2024 at 09:20:09AM +0700, John Naylor wrote:
> On Tue, Jan 9, 2024 at 12:37 AM Nathan Bossart <nathandbossart@gmail.com> wrote:
>>
>> > I suspect that there could be a regression lurking for some inputs
>> > that the benchmark doesn't look at: pg_lfind32() currently needs to be
>> > able to read 4 vector registers worth of elements before taking the
>> > fast path. There is then a tail of up to 15 elements that are now
>> > checked one-by-one, but AVX2 would increase that to 31. That's getting
>> > big enough to be noticeable, I suspect. It would be good to understand
>> > that case (n*32 + 31), because it may also be relevant now. It's also
>> > easy to improve for SSE2/NEON for v17.
>>
>> Good idea. If it is indeed noticeable, we might be able to "fix" it by
>> processing some of the tail with shorter vectors. But that probably means
>> finding a way to support multiple vector sizes on the same build, which
>> would require some work.
>
> What I had in mind was an overlapping pattern I've seen in various
> places: do one iteration at the beginning, then subtract the
> aligned-down length from the end and do all those iterations. And
> one-by-one is only used if the total length is small.
Sorry, I'm not sure I understood this. Do you mean processing the first
several elements individually or with SSE2 until the number of remaining
elements can be processed with just the AVX2 instructions (a bit like how
pg_comp_crc32c_armv8() is structured for memory alignment)?
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-hackers@postgresql.org
Cc: nathandbossart@gmail.com, johncnaylorls@gmail.com, ants@cybertec.at
Subject: Re: add AVX2 support to simd.h
In-Reply-To: <20240109162009.GA3033933@nathanxps13>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox