public inbox for glibc-bugs@sourceware.org
help / color / mirror / Atom feed
From: "hjl.tools at gmail dot com" <sourceware-bugzilla@sourceware.org>
To: glibc-bugs@sourceware.org
Subject: [Bug string/27457] vzeroupper use in AVX2 multiarch string functions cause HTM aborts
Date: Mon, 01 Mar 2021 14:37:33 +0000 [thread overview]
Message-ID: <bug-27457-131-GaRqP7vqgT@http.sourceware.org/bugzilla/> (raw)
In-Reply-To: <bug-27457-131@http.sourceware.org/bugzilla/>
https://sourceware.org/bugzilla/show_bug.cgi?id=27457
--- Comment #18 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to rguenther from comment #16)
> On Mon, 1 Mar 2021, hjl.tools at gmail dot com wrote:
>
> > https://sourceware.org/bugzilla/show_bug.cgi?id=27457
> >
> > --- Comment #15 from H.J. Lu <hjl.tools at gmail dot com> ---
> > (In reply to Richard Biener from comment #14)
> > >
> > > Note according to Agner vzeroall, for example on Haswell, decodes to
> > > 20 uops while vzeroupper only requires 4. On Skylake it's even worse
> > > (34 uops). For short sizes (as in our benchmark which had 16-31 byte
> > > strcmp) this might be a bigger difference than using the SSE2 variant
> > > off an early xtest result. That said, why not, for HTM + AVX2 CPUs,
> > > have an intermediate dispatcher between the AVX2 and the SSE variant
> > > using xtest? That leaves the actual implementations unchanged and thus
> > > with known performance characteristic?
> >
> > It is implemented on users/hjl/pr27457/wrapper branch:
> >
> > https://gitlab.com/x86-glibc/glibc/-/tree/users/hjl/pr27457/wrapper
> >
> > There are 2 problems:
> >
> > 1. Many RTM tests failed for other reasons.
> > 2. Even with vzeroall overhead, AVX version may still be faster than
> > SSE version.
>
> And the SSE version may still be faster than the AVX version with
> vzeroall.
Here is some data:
Function: strcmp
Variant: default
__strcmp_avx2 __strcmp_sse2_unaligned
length=14, align1=14, align2=14: 11.36 17.50
length=14, align1=14, align2=14: 11.36 15.59
length=14, align1=14, align2=14: 11.43 15.55
length=15, align1=15, align2=15: 11.36 17.42
length=15, align1=15, align2=15: 11.96 17.41
length=15, align1=15, align2=15: 11.36 16.97
length=16, align1=16, align2=16: 11.36 18.58
length=16, align1=16, align2=16: 11.36 17.41
length=16, align1=16, align2=16: 11.43 17.34
length=17, align1=17, align2=17: 11.36 21.37
length=17, align1=17, align2=17: 11.36 18.52
length=17, align1=17, align2=17: 11.36 17.94
length=18, align1=18, align2=18: 11.36 19.73
length=18, align1=18, align2=18: 11.36 19.20
length=18, align1=18, align2=18: 11.36 19.13
length=19, align1=19, align2=19: 11.36 20.38
length=19, align1=19, align2=19: 11.36 19.39
length=19, align1=19, align2=19: 11.36 20.39
length=20, align1=20, align2=20: 11.36 21.53
length=20, align1=20, align2=20: 11.36 20.98
length=20, align1=20, align2=20: 11.36 20.93
length=21, align1=21, align2=21: 11.36 22.83
length=21, align1=21, align2=21: 11.36 22.26
length=21, align1=21, align2=21: 11.36 22.25
length=22, align1=22, align2=22: 11.43 23.37
length=22, align1=22, align2=22: 11.36 22.78
length=22, align1=22, align2=22: 12.29 22.12
length=23, align1=23, align2=23: 11.36 24.63
length=23, align1=23, align2=23: 12.53 23.97
length=23, align1=23, align2=23: 11.36 23.97
length=24, align1=24, align2=24: 11.36 24.52
length=24, align1=24, align2=24: 11.36 43.47
length=24, align1=24, align2=24: 11.36 44.47
length=25, align1=25, align2=25: 11.36 39.50
length=25, align1=25, align2=25: 11.36 48.97
length=25, align1=25, align2=25: 11.36 48.53
length=26, align1=26, align2=26: 11.36 47.87
length=26, align1=26, align2=26: 11.36 47.20
length=26, align1=26, align2=26: 11.36 47.15
length=27, align1=27, align2=27: 11.36 50.90
length=27, align1=27, align2=27: 11.44 49.98
length=27, align1=27, align2=27: 11.36 49.77
length=28, align1=28, align2=28: 11.36 49.74
length=28, align1=28, align2=28: 11.36 48.86
length=28, align1=28, align2=28: 11.36 49.08
length=29, align1=29, align2=29: 11.36 52.74
length=29, align1=29, align2=29: 11.36 54.04
length=29, align1=29, align2=29: 11.36 29.49
length=30, align1=30, align2=30: 11.36 50.91
length=30, align1=30, align2=30: 11.36 51.09
length=30, align1=30, align2=30: 11.36 51.13
length=31, align1=31, align2=31: 12.36 54.33
length=31, align1=31, align2=31: 11.36 53.49
length=31, align1=31, align2=31: 11.36 53.29
length=16, align1=0, align2=0: 11.36 18.02
length=16, align1=0, align2=0: 11.36 18.58
length=16, align1=0, align2=0: 11.36 17.34
length=16, align1=0, align2=0: 11.44 19.88
length=16, align1=0, align2=0: 11.36 16.74
length=16, align1=0, align2=0: 11.36 17.42
length=16, align1=0, align2=3: 11.36 17.34
length=16, align1=3, align2=4: 11.36 17.34
length=32, align1=0, align2=0: 12.29 61.07
length=32, align1=0, align2=0: 12.63 61.08
length=32, align1=0, align2=0: 11.36 60.48
length=32, align1=0, align2=0: 11.36 60.48
length=32, align1=0, align2=0: 11.36 60.40
length=32, align1=0, align2=0: 11.36 60.40
length=32, align1=0, align2=4: 11.36 60.40
length=32, align1=4, align2=5: 12.10 59.72
> I guess we should mostly care about optimizing for "modern" CPUs
> which likely means HTM + AVX512 which should be already optimal
> on your branches by using %ymm16+. So we're talking about
> the "legacy" AVX2 + HTM path.
>
> And there I think we should optimize the path that is _not_ in
> a transaction since that will be 99% of the cases. Which to
> me means using the proven tuned (on their respective ISA subsets)
> SSE2 and AVX2 variants and simply switch between them based on
> xtest. Yeah, so strcmp of a large string inside an transaction
I tried it and I got RTM abort for other reasons.
> might not run at optimal AVX2 speed. But it will be faster
> than before the xtest dispatch since before that it would have
> aborted the transaction.
Please give my current approach is a try.
--
You are receiving this mail because:
You are on the CC list for the bug.
next prev parent reply other threads:[~2021-03-01 14:37 UTC|newest]
Thread overview: 42+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-02-22 12:40 [Bug string/27457] New: " rguenth at gcc dot gnu.org
2021-02-22 12:40 ` [Bug string/27457] " rguenth at gcc dot gnu.org
2021-02-22 14:50 ` matz at suse dot de
2021-02-22 15:00 ` rguenth at gcc dot gnu.org
2021-02-22 15:26 ` hjl.tools at gmail dot com
2021-02-22 15:26 ` hjl.tools at gmail dot com
2021-02-22 18:45 ` fweimer at redhat dot com
2021-02-23 9:44 ` roman.dementiev at intel dot com
2021-02-27 2:39 ` hjl.tools at gmail dot com
2021-02-27 7:34 ` rguenther at suse dot de
2021-02-28 14:53 ` hjl.tools at gmail dot com
2021-03-01 11:32 ` fweimer at redhat dot com
2021-03-01 12:24 ` mliska at suse dot cz
2021-03-01 12:47 ` rguenther at suse dot de
2021-03-01 13:13 ` roman.dementiev at intel dot com
2021-03-01 13:19 ` fweimer at redhat dot com
2021-03-01 13:21 ` hjl.tools at gmail dot com
2021-03-01 13:24 ` hjl.tools at gmail dot com
2021-03-01 13:27 ` hjl.tools at gmail dot com
2021-03-01 13:29 ` hjl.tools at gmail dot com
2021-03-01 13:44 ` rguenth at gcc dot gnu.org
2021-03-01 14:05 ` hjl.tools at gmail dot com
2021-03-01 14:14 ` rguenther at suse dot de
2021-03-01 14:25 ` rguenth at gcc dot gnu.org
2021-03-01 14:37 ` hjl.tools at gmail dot com [this message]
2021-03-01 14:47 ` hjl.tools at gmail dot com
2021-03-01 14:49 ` rguenth at gcc dot gnu.org
2021-03-01 14:53 ` rguenth at gcc dot gnu.org
2021-03-01 15:19 ` hjl.tools at gmail dot com
2021-03-01 23:39 ` hjl.tools at gmail dot com
2021-03-05 16:54 ` hjl.tools at gmail dot com
2021-03-11 10:42 ` rguenth at gcc dot gnu.org
2021-03-16 13:53 ` rguenth at gcc dot gnu.org
2021-03-16 14:12 ` hjl.tools at gmail dot com
2021-03-29 23:00 ` hjl.tools at gmail dot com
2022-01-27 20:21 ` cvs-commit at gcc dot gnu.org
2022-01-27 20:23 ` cvs-commit at gcc dot gnu.org
2022-01-27 20:47 ` cvs-commit at gcc dot gnu.org
2022-01-27 20:47 ` cvs-commit at gcc dot gnu.org
2022-01-27 20:48 ` cvs-commit at gcc dot gnu.org
2022-01-27 22:41 ` cvs-commit at gcc dot gnu.org
2022-01-28 2:24 ` hjl.tools at gmail dot com
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=bug-27457-131-GaRqP7vqgT@http.sourceware.org/bugzilla/ \
--to=sourceware-bugzilla@sourceware.org \
--cc=glibc-bugs@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).