c5d972
From 9bc2ed8f46d80859a5596789cc9e8cc2de84b0e7 Mon Sep 17 00:00:00 2001
c5d972
From: Wilco Dijkstra <wdijkstr@arm.com>
c5d972
Date: Tue, 10 Aug 2021 13:39:37 +0100
c5d972
Subject: [PATCH] [2/5] AArch64: Improve A64FX memset for large sizes
c5d972
c5d972
Improve performance of large memsets. Simplify alignment code. For zero memset
c5d972
use DC ZVA, which almost doubles performance. For non-zero memsets use the
c5d972
unroll8 loop which is about 10% faster.
c5d972
c5d972
Reviewed-by: Naohiro Tamura <naohirot@fujitsu.com>
c5d972
---
c5d972
 sysdeps/aarch64/multiarch/memset_a64fx.S | 85 +++++++-----------------
c5d972
 1 file changed, 25 insertions(+), 60 deletions(-)
c5d972
c5d972
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
c5d972
index cf3d402ef6..75cf43ae79 100644
c5d972
--- a/sysdeps/aarch64/multiarch/memset_a64fx.S
c5d972
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
c5d972
@@ -27,14 +27,11 @@
c5d972
  */
c5d972
 
c5d972
 #define L1_SIZE		(64*1024)	// L1 64KB
c5d972
-#define L2_SIZE         (8*1024*1024)	// L2 8MB - 1MB
c5d972
+#define L2_SIZE         (8*1024*1024)	// L2 8MB
c5d972
 #define CACHE_LINE_SIZE	256
c5d972
 #define PF_DIST_L1	(CACHE_LINE_SIZE * 16)	// Prefetch distance L1
c5d972
-#define ZF_DIST		(CACHE_LINE_SIZE * 21)	// Zerofill distance
c5d972
-#define rest		x8
c5d972
+#define rest		x2
c5d972
 #define vector_length	x9
c5d972
-#define vl_remainder	x10	// vector_length remainder
c5d972
-#define cl_remainder	x11	// CACHE_LINE_SIZE remainder
c5d972
 
c5d972
 #if HAVE_AARCH64_SVE_ASM
c5d972
 # if IS_IN (libc)
c5d972
@@ -42,14 +39,6 @@
c5d972
 
c5d972
 	.arch armv8.2-a+sve
c5d972
 
c5d972
-	.macro dc_zva times
c5d972
-	dc	zva, tmp1
c5d972
-	add	tmp1, tmp1, CACHE_LINE_SIZE
c5d972
-	.if \times-1
c5d972
-	dc_zva "(\times-1)"
c5d972
-	.endif
c5d972
-	.endm
c5d972
-
c5d972
 	.macro st1b_unroll first=0, last=7
c5d972
 	st1b	z0.b, p0, [dst, \first, mul vl]
c5d972
 	.if \last-\first
c5d972
@@ -188,54 +177,30 @@ L(L1_prefetch): // if rest >= L1_SIZE
c5d972
 	cbnz	rest, L(unroll32)
c5d972
 	ret
c5d972
 
c5d972
-L(L2):
c5d972
-	// align dst address at vector_length byte boundary
c5d972
-	sub	tmp1, vector_length, 1
c5d972
-	ands	tmp2, dst, tmp1
c5d972
-	// if vl_remainder == 0
c5d972
-	b.eq	1f
c5d972
-	sub	vl_remainder, vector_length, tmp2
c5d972
-	// process remainder until the first vector_length boundary
c5d972
-	whilelt	p2.b, xzr, vl_remainder
c5d972
-	st1b	z0.b, p2, [dst]
c5d972
-	add	dst, dst, vl_remainder
c5d972
-	sub	rest, rest, vl_remainder
c5d972
-	// align dstin address at CACHE_LINE_SIZE byte boundary
c5d972
-1:	mov	tmp1, CACHE_LINE_SIZE
c5d972
-	ands	tmp2, dst, CACHE_LINE_SIZE - 1
c5d972
-	// if cl_remainder == 0
c5d972
-	b.eq	L(L2_dc_zva)
c5d972
-	sub	cl_remainder, tmp1, tmp2
c5d972
-	// process remainder until the first CACHE_LINE_SIZE boundary
c5d972
-	mov	tmp1, xzr       // index
c5d972
-2:	whilelt	p2.b, tmp1, cl_remainder
c5d972
-	st1b	z0.b, p2, [dst, tmp1]
c5d972
-	incb	tmp1
c5d972
-	cmp	tmp1, cl_remainder
c5d972
-	b.lo	2b
c5d972
-	add	dst, dst, cl_remainder
c5d972
-	sub	rest, rest, cl_remainder
c5d972
-
c5d972
-L(L2_dc_zva):
c5d972
-	// zero fill
c5d972
-	mov	tmp1, dst
c5d972
-	dc_zva	(ZF_DIST / CACHE_LINE_SIZE) - 1
c5d972
-	mov	zva_len, ZF_DIST
c5d972
-	add	tmp1, zva_len, CACHE_LINE_SIZE * 2
c5d972
-	// unroll
c5d972
+	// count >= L2_SIZE
c5d972
 	.p2align 3
c5d972
-1:	st1b_unroll 0, 3
c5d972
-	add	tmp2, dst, zva_len
c5d972
-	dc	 zva, tmp2
c5d972
-	st1b_unroll 4, 7
c5d972
-	add	tmp2, tmp2, CACHE_LINE_SIZE
c5d972
-	dc	zva, tmp2
c5d972
-	add	dst, dst, CACHE_LINE_SIZE * 2
c5d972
-	sub	rest, rest, CACHE_LINE_SIZE * 2
c5d972
-	cmp	rest, tmp1	// ZF_DIST + CACHE_LINE_SIZE * 2
c5d972
-	b.ge	1b
c5d972
-	cbnz	rest, L(unroll8)
c5d972
-	ret
c5d972
+L(L2):
c5d972
+	tst	valw, 255
c5d972
+	b.ne	L(unroll8)
c5d972
+        // align dst to CACHE_LINE_SIZE byte boundary
c5d972
+	and	tmp2, dst, CACHE_LINE_SIZE - 1
c5d972
+	st1b	z0.b, p0, [dst, 0, mul vl]
c5d972
+	st1b	z0.b, p0, [dst, 1, mul vl]
c5d972
+	st1b	z0.b, p0, [dst, 2, mul vl]
c5d972
+	st1b	z0.b, p0, [dst, 3, mul vl]
c5d972
+	sub	dst, dst, tmp2
c5d972
+	add	count, count, tmp2
c5d972
+
c5d972
+	// clear cachelines using DC ZVA
c5d972
+	sub	count, count, CACHE_LINE_SIZE * 2
c5d972
+	.p2align 4
c5d972
+1:	add	dst, dst, CACHE_LINE_SIZE
c5d972
+	dc	zva, dst
c5d972
+	subs	count, count, CACHE_LINE_SIZE
c5d972
+	b.hi	1b
c5d972
+	add	count, count, CACHE_LINE_SIZE
c5d972
+	add	dst, dst, CACHE_LINE_SIZE
c5d972
+	b	L(last)
c5d972
 
c5d972
 END (MEMSET)
c5d972
 libc_hidden_builtin_def (MEMSET)
c5d972
-- 
c5d972
2.31.1
c5d972