From: Wilco Dijkstra via Libc-alpha <libc-alpha@sourceware.org>
To: "naohirot@fujitsu.com" <naohirot@fujitsu.com>
Cc: 'GNU C Library' <libc-alpha@sourceware.org>
Subject: [PATCH v3 1/5] AArch64: Improve A64FX memset
Date: Thu, 22 Jul 2021 15:59:19 +0000 [thread overview]
Message-ID: <VE1PR08MB5599A80A0480330F1CB5CC0083E49@VE1PR08MB5599.eurprd08.prod.outlook.com> (raw)
Improve performance of small copies by reducing instruction counts and improving
alignment. Bench-memset shows 35-45% performance gain for small sizes.
---
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
index ce54e5418b08c8bc0ecc7affff68a59272ba6397..f7fcc7b323e1553f50a2e005b8ccef344a08127d 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 - 1MB
#define CACHE_LINE_SIZE 256
#define PF_DIST_L1 (CACHE_LINE_SIZE * 16) // Prefetch distance L1
-#define ZF_DIST (CACHE_LINE_SIZE * 21) // Zerofill distance
#define rest x8
#define vector_length x9
#define vl_remainder x10 // vector_length remainder
@@ -51,78 +50,54 @@
.endm
.macro st1b_unroll first=0, last=7
- st1b z0.b, p0, [dst, #\first, mul vl]
+ st1b z0.b, p0, [dst, \first, mul vl]
.if \last-\first
st1b_unroll "(\first+1)", \last
.endif
.endm
- .macro shortcut_for_small_size exit
- // if rest <= vector_length * 2
- whilelo p0.b, xzr, count
- whilelo p1.b, vector_length, count
- b.last 1f
- st1b z0.b, p0, [dstin, #0, mul vl]
- st1b z0.b, p1, [dstin, #1, mul vl]
- ret
-1: // if rest > vector_length * 8
- cmp count, vector_length, lsl 3 // vector_length * 8
- b.hi \exit
- // if rest <= vector_length * 4
- lsl tmp1, vector_length, 1 // vector_length * 2
- whilelo p2.b, tmp1, count
- incb tmp1
- whilelo p3.b, tmp1, count
- b.last 1f
- st1b z0.b, p0, [dstin, #0, mul vl]
- st1b z0.b, p1, [dstin, #1, mul vl]
- st1b z0.b, p2, [dstin, #2, mul vl]
- st1b z0.b, p3, [dstin, #3, mul vl]
- ret
-1: // if rest <= vector_length * 8
- lsl tmp1, vector_length, 2 // vector_length * 4
- whilelo p4.b, tmp1, count
- incb tmp1
- whilelo p5.b, tmp1, count
- b.last 1f
- st1b z0.b, p0, [dstin, #0, mul vl]
- st1b z0.b, p1, [dstin, #1, mul vl]
- st1b z0.b, p2, [dstin, #2, mul vl]
- st1b z0.b, p3, [dstin, #3, mul vl]
- st1b z0.b, p4, [dstin, #4, mul vl]
- st1b z0.b, p5, [dstin, #5, mul vl]
- ret
-1: lsl tmp1, vector_length, 2 // vector_length * 4
- incb tmp1 // vector_length * 5
- incb tmp1 // vector_length * 6
- whilelo p6.b, tmp1, count
- incb tmp1
- whilelo p7.b, tmp1, count
- st1b z0.b, p0, [dstin, #0, mul vl]
- st1b z0.b, p1, [dstin, #1, mul vl]
- st1b z0.b, p2, [dstin, #2, mul vl]
- st1b z0.b, p3, [dstin, #3, mul vl]
- st1b z0.b, p4, [dstin, #4, mul vl]
- st1b z0.b, p5, [dstin, #5, mul vl]
- st1b z0.b, p6, [dstin, #6, mul vl]
- st1b z0.b, p7, [dstin, #7, mul vl]
- ret
- .endm
-ENTRY (MEMSET)
+#undef BTI_C
+#define BTI_C
+ENTRY (MEMSET)
PTR_ARG (0)
SIZE_ARG (2)
- cbnz count, 1f
- ret
-1: dup z0.b, valw
cntb vector_length
- // shortcut for less than vector_length * 8
- // gives a free ptrue to p0.b for n >= vector_length
- shortcut_for_small_size L(vl_agnostic)
- // end of shortcut
+ dup z0.b, valw
+ whilelo p0.b, vector_length, count
+ b.last 1f
+ whilelo p1.b, xzr, count
+ st1b z0.b, p1, [dstin, 0, mul vl]
+ st1b z0.b, p0, [dstin, 1, mul vl]
+ ret
+
+ // count >= vector_length * 2
+1: cmp count, vector_length, lsl 2
+ add dstend, dstin, count
+ b.hi 1f
+ st1b z0.b, p0, [dstin, 0, mul vl]
+ st1b z0.b, p0, [dstin, 1, mul vl]
+ st1b z0.b, p0, [dstend, -2, mul vl]
+ st1b z0.b, p0, [dstend, -1, mul vl]
+ ret
+
+ // count > vector_length * 4
+1: lsl tmp1, vector_length, 3
+ cmp count, tmp1
+ b.hi L(vl_agnostic)
+ st1b z0.b, p0, [dstin, 0, mul vl]
+ st1b z0.b, p0, [dstin, 1, mul vl]
+ st1b z0.b, p0, [dstin, 2, mul vl]
+ st1b z0.b, p0, [dstin, 3, mul vl]
+ st1b z0.b, p0, [dstend, -4, mul vl]
+ st1b z0.b, p0, [dstend, -3, mul vl]
+ st1b z0.b, p0, [dstend, -2, mul vl]
+ st1b z0.b, p0, [dstend, -1, mul vl]
+ ret
+ .p2align 4
L(vl_agnostic): // VL Agnostic
mov rest, count
mov dst, dstin
next reply other threads:[~2021-07-22 16:00 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-07-22 15:59 Wilco Dijkstra via Libc-alpha [this message]
2021-07-28 8:10 ` [PATCH v3 1/5] AArch64: Improve A64FX memset naohirot--- via Libc-alpha
2021-08-02 13:53 ` naohirot--- via Libc-alpha
2021-08-02 14:38 ` Wilco Dijkstra via Libc-alpha
2021-08-02 14:50 ` Szabolcs Nagy via Libc-alpha
2021-08-03 2:57 ` naohirot--- via Libc-alpha
2021-08-03 8:01 ` Szabolcs Nagy via Libc-alpha
2021-09-24 7:56 ` naohirot--- via Libc-alpha
2021-08-03 2:56 ` naohirot--- via Libc-alpha
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=VE1PR08MB5599A80A0480330F1CB5CC0083E49@VE1PR08MB5599.eurprd08.prod.outlook.com \
--to=libc-alpha@sourceware.org \
--cc=Wilco.Dijkstra@arm.com \
--cc=naohirot@fujitsu.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).