From: "H.J. Lu" <hjl.tools@gmail.com>
To: Noah Goldstein <goldstein.w.n@gmail.com>
Cc: GNU C Library <libc-alpha@sourceware.org>,
"Carlos O'Donell" <carlos@systemhalted.org>
Subject: Re: [PATCH v2 1/2] x86: Optimize less_vec evex and avx512 memset-vec-unaligned-erms.S
Date: Mon, 19 Apr 2021 13:39:03 -0700 [thread overview]
Message-ID: <CAMe9rOpxiWWmPVkc-EGAtgKWp20Po3aDc5o8ZEANLwEE1uz01w@mail.gmail.com> (raw)
In-Reply-To: <CAFUsyfJXLM-mHd=hBHV6PWoSVgUbZdmfBZwBtcqdoEm4Jjtk2Q@mail.gmail.com>
On Mon, Apr 19, 2021 at 12:35 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Mon, Apr 19, 2021 at 2:45 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Mon, Apr 19, 2021 at 9:30 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > >
> > > No bug. This commit adds optimized cased for less_vec memset case that
> > > uses the avx512vl/avx512bw mask store avoiding the excessive
> > > branches. test-memset and test-wmemset are passing.
> > >
> > > Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> > > ---
> > > sysdeps/x86_64/multiarch/ifunc-memset.h | 6 ++-
> > > .../multiarch/memset-avx512-unaligned-erms.S | 2 +-
> > > .../multiarch/memset-evex-unaligned-erms.S | 2 +-
> > > .../multiarch/memset-vec-unaligned-erms.S | 52 +++++++++++++++----
> > > 4 files changed, 47 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/sysdeps/x86_64/multiarch/ifunc-memset.h b/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > index 502f946a84..eda5640541 100644
> > > --- a/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > +++ b/sysdeps/x86_64/multiarch/ifunc-memset.h
> > > @@ -54,7 +54,8 @@ IFUNC_SELECTOR (void)
> > > && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))
> > > {
> > > if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
> > > - && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
> > > + && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
> > > + && CPU_FEATURE_USABLE_P (cpu_features, BMI2))
> > > {
> > > if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > return OPTIMIZE (avx512_unaligned_erms);
> > > @@ -68,7 +69,8 @@ IFUNC_SELECTOR (void)
> > > if (CPU_FEATURE_USABLE_P (cpu_features, AVX2))
> > > {
> > > if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
> > > - && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
> > > + && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
> > > + && CPU_FEATURE_USABLE_P (cpu_features, BMI2))
> >
> > Please also update ifunc-impl-list.c.
>
> Done.
>
> >
> > > {
> > > if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
> > > return OPTIMIZE (evex_unaligned_erms);
> > > diff --git a/sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S
> > > index 22e7b187c8..d03460be93 100644
> > > --- a/sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S
> > > +++ b/sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S
> > > @@ -19,6 +19,6 @@
> > > # define SECTION(p) p##.evex512
> > > # define MEMSET_SYMBOL(p,s) p##_avx512_##s
> > > # define WMEMSET_SYMBOL(p,s) p##_avx512_##s
> > > -
> > > +# define USE_LESS_VEC_MASKMOV 1
> >
> > USE_LESS_VEC_MASKED_STORE
>
> Done.
>
> >
> > > # include "memset-vec-unaligned-erms.S"
> > > #endif
> > > diff --git a/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S
> > > index ae0a4d6e46..eb3541ef60 100644
> > > --- a/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S
> > > +++ b/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S
> > > @@ -19,6 +19,6 @@
> > > # define SECTION(p) p##.evex
> > > # define MEMSET_SYMBOL(p,s) p##_evex_##s
> > > # define WMEMSET_SYMBOL(p,s) p##_evex_##s
> > > -
> > > +# define USE_LESS_VEC_MASKMOV 1
> > > # include "memset-vec-unaligned-erms.S"
> > > #endif
> > > diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > > index 584747f1a1..6b02e87f48 100644
> > > --- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > > +++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
> > > @@ -63,6 +63,9 @@
> > > # endif
> > > #endif
> > >
> > > +#define PAGE_SIZE 4096
> > > +#define LOG_PAGE_SIZE 12
> > > +
> > > #ifndef SECTION
> > > # error SECTION is not defined!
> > > #endif
> > > @@ -213,11 +216,38 @@ L(loop):
> > > cmpq %rcx, %rdx
> > > jne L(loop)
> > > VZEROUPPER_SHORT_RETURN
> > > +
> > > + .p2align 4
> > > L(less_vec):
> > > /* Less than 1 VEC. */
> > > # if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64
> > > # error Unsupported VEC_SIZE!
> > > # endif
> > > +# ifdef USE_LESS_VEC_MASKMOV
> > > + /* Clear high bits from edi. Only keeping bits relevant to page
> > > + cross check. Using sall instead of andl saves 3 bytes. Note
> > > + that we are using rax which is set in
> > > + MEMSET_VDUP_TO_VEC0_AND_SET_RETURN as ptr from here on out. */
> > > + sall $(32 - LOG_PAGE_SIZE), %edi
> > > + /* Check if VEC_SIZE load cross page. Mask loads suffer serious
> > > + performance degradation when it has to fault supress. */
> > > + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %edi
> >
> > Please use AND and CMP since AND has higher throughput.
>
> AND uses more code size for VEC_SIZE=16/32 and just barely pushes the
> L(cross_page) to the next 16 byte chunk so the extra 3 bytes from AND
> end up costing 16 bytes. Not aligning L(cross_page) to 16 also
> introduces higher variance to benchmarks so I think it has to be all 16 bytes.
>
> As is I don't think throughput of AND / SAL is on the critical
> path so code size should win out. (We can also decode MOV -1, ecx
> first cycle with SAL as opposed to AND).
>
> What do you think?
I prefer AND over SAL. Something like
diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
index 3a59d39267..763fb907b9 100644
--- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
@@ -217,21 +217,17 @@ L(loop):
jne L(loop)
VZEROUPPER_SHORT_RETURN
- .p2align 4
+ /* NB: Don't align this branch target to reduce code size. */
L(less_vec):
/* Less than 1 VEC. */
# if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64
# error Unsupported VEC_SIZE!
# endif
# ifdef USE_LESS_VEC_MASK_STORE
- /* Clear high bits from edi. Only keeping bits relevant to page
- cross check. Using sall instead of andl saves 3 bytes. Note
- that we are using rax which is set in
- MEMSET_VDUP_TO_VEC0_AND_SET_RETURN as ptr from here on out. */
- sall $(32 - LOG_PAGE_SIZE), %edi
- /* Check if VEC_SIZE load cross page. Mask loads suffer serious
+ /* Check if VEC_SIZE store cross page. Mask stores suffer serious
performance degradation when it has to fault supress. */
- cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %edi
+ andl $(PAGE_SIZE - 1), %edi
+ cmpl $(PAGE_SIZE - VEC_SIZE), %edi
ja L(cross_page)
# if VEC_SIZE > 32
movq $-1, %rcx
Thanks.
--
H.J.
next prev parent reply other threads:[~2021-04-19 20:39 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-04-19 16:30 Noah Goldstein
2021-04-19 16:30 ` [PATCH v2 2/2] x86: Expand test-memset.c and bench-memset.c Noah Goldstein
2021-04-19 18:44 ` [PATCH v2 1/2] x86: Optimize less_vec evex and avx512 memset-vec-unaligned-erms.S H.J. Lu
2021-04-19 19:35 ` Noah Goldstein
2021-04-19 20:39 ` H.J. Lu [this message]
2021-04-19 21:07 ` Noah Goldstein
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=CAMe9rOpxiWWmPVkc-EGAtgKWp20Po3aDc5o8ZEANLwEE1uz01w@mail.gmail.com \
--to=hjl.tools@gmail.com \
--cc=carlos@systemhalted.org \
--cc=goldstein.w.n@gmail.com \
--cc=libc-alpha@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).