From: Szabolcs Nagy via Libc-alpha <libc-alpha@sourceware.org>
To: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Cc: 'GNU C Library' <libc-alpha@sourceware.org>
Subject: Re: [PATCH v4 5/5] AArch64: Improve A64FX memset medium loops
Date: Tue, 10 Aug 2021 10:44:29 +0100 [thread overview]
Message-ID: <20210810094428.GG20410@arm.com> (raw)
In-Reply-To: <VE1PR08MB5599EFE4EC93A5A2D268893383F69@VE1PR08MB5599.eurprd08.prod.outlook.com>
The 08/09/2021 13:15, Wilco Dijkstra via Libc-alpha wrote:
> v4: minor loop change
>
> Simplify the code for memsets smaller than L1. Improve the unroll8 and L1_prefetch loops.
OK to commit, but keep
Reviewed-by: Naohiro Tamura <naohirot@fujitsu.com>
(further tweaks can go into follwup commits.)
>
> ---
>
> diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
> index 89dba912588c243e67a9527a56b4d3a44659d542..318c6350a31e0fad788b5f2139de645ddc51493f 100644
> --- a/sysdeps/aarch64/multiarch/memset_a64fx.S
> +++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
> @@ -30,7 +30,6 @@
> #define L2_SIZE (8*1024*1024) // L2 8MB
> #define CACHE_LINE_SIZE 256
> #define PF_DIST_L1 (CACHE_LINE_SIZE * 16) // Prefetch distance L1
> -#define rest x2
> #define vector_length x9
>
> #if HAVE_AARCH64_SVE_ASM
> @@ -89,29 +88,19 @@ ENTRY (MEMSET)
>
> .p2align 4
> L(vl_agnostic): // VL Agnostic
> - mov rest, count
> mov dst, dstin
> - add dstend, dstin, count
> - // if rest >= L2_SIZE && vector_length == 64 then L(L2)
> - mov tmp1, 64
> - cmp rest, L2_SIZE
> - ccmp vector_length, tmp1, 0, cs
> - b.eq L(L2)
> - // if rest >= L1_SIZE && vector_length == 64 then L(L1_prefetch)
> - cmp rest, L1_SIZE
> - ccmp vector_length, tmp1, 0, cs
> - b.eq L(L1_prefetch)
> -
> + cmp count, L1_SIZE
> + b.hi L(L1_prefetch)
>
> + // count >= 8 * vector_length
> L(unroll8):
> - lsl tmp1, vector_length, 3
> - .p2align 3
> -1: cmp rest, tmp1
> - b.cc L(last)
> - st1b_unroll
> + sub count, count, tmp1
> + .p2align 4
> +1: st1b_unroll 0, 7
> add dst, dst, tmp1
> - sub rest, rest, tmp1
> - b 1b
> + subs count, count, tmp1
> + b.hi 1b
> + add count, count, tmp1
>
> L(last):
> cmp count, vector_length, lsl 1
> @@ -129,18 +118,22 @@ L(last):
> st1b z0.b, p0, [dstend, -1, mul vl]
> ret
>
> -L(L1_prefetch): // if rest >= L1_SIZE
> + // count >= L1_SIZE
> .p2align 3
> +L(L1_prefetch):
> + cmp count, L2_SIZE
> + b.hs L(L2)
> + cmp vector_length, 64
> + b.ne L(unroll8)
> 1: st1b_unroll 0, 3
> prfm pstl1keep, [dst, PF_DIST_L1]
> st1b_unroll 4, 7
> prfm pstl1keep, [dst, PF_DIST_L1 + CACHE_LINE_SIZE]
> add dst, dst, CACHE_LINE_SIZE * 2
> - sub rest, rest, CACHE_LINE_SIZE * 2
> - cmp rest, L1_SIZE
> - b.ge 1b
> - cbnz rest, L(unroll8)
> - ret
> + sub count, count, CACHE_LINE_SIZE * 2
> + cmp count, PF_DIST_L1
> + b.hs 1b
> + b L(unroll8)
>
> // count >= L2_SIZE
> .p2align 3
>
--
prev parent reply other threads:[~2021-08-10 9:45 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-08-09 13:15 [PATCH v4 5/5] AArch64: Improve A64FX memset medium loops Wilco Dijkstra via Libc-alpha
2021-08-10 9:44 ` Szabolcs Nagy via Libc-alpha [this message]
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
List information: https://www.gnu.org/software/libc/involved.html
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20210810094428.GG20410@arm.com \
--to=libc-alpha@sourceware.org \
--cc=Wilco.Dijkstra@arm.com \
--cc=szabolcs.nagy@arm.com \
/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).