|
 |
a41fe4 |
From e69d9981f858a38e19304e6ff5ebdf89f2cb0ba0 Mon Sep 17 00:00:00 2001
|
|
|
a41fe4 |
From: Wilco Dijkstra <wdijkstr@arm.com>
|
|
|
a41fe4 |
Date: Tue, 10 Aug 2021 13:44:27 +0100
|
|
|
a41fe4 |
Subject: [PATCH] [4/5] AArch64: Improve A64FX memset by removing unroll32
|
|
|
a41fe4 |
|
|
|
a41fe4 |
Remove unroll32 code since it doesn't improve performance.
|
|
|
a41fe4 |
|
|
|
a41fe4 |
Reviewed-by: Naohiro Tamura <naohirot@fujitsu.com>
|
|
|
a41fe4 |
---
|
|
|
a41fe4 |
sysdeps/aarch64/multiarch/memset_a64fx.S | 18 +-----------------
|
|
|
a41fe4 |
1 file changed, 1 insertion(+), 17 deletions(-)
|
|
|
a41fe4 |
|
|
|
a41fe4 |
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
index 337c86be6f..ef0315658a 100644
|
|
|
a41fe4 |
--- a/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
|
|
|
a41fe4 |
@@ -102,22 +102,6 @@ L(vl_agnostic): // VL Agnostic
|
|
|
a41fe4 |
ccmp vector_length, tmp1, 0, cs
|
|
|
a41fe4 |
b.eq L(L1_prefetch)
|
|
|
a41fe4 |
|
|
|
a41fe4 |
-L(unroll32):
|
|
|
a41fe4 |
- lsl tmp1, vector_length, 3 // vector_length * 8
|
|
|
a41fe4 |
- lsl tmp2, vector_length, 5 // vector_length * 32
|
|
|
a41fe4 |
- .p2align 3
|
|
|
a41fe4 |
-1: cmp rest, tmp2
|
|
|
a41fe4 |
- b.cc L(unroll8)
|
|
|
a41fe4 |
- st1b_unroll
|
|
|
a41fe4 |
- add dst, dst, tmp1
|
|
|
a41fe4 |
- st1b_unroll
|
|
|
a41fe4 |
- add dst, dst, tmp1
|
|
|
a41fe4 |
- st1b_unroll
|
|
|
a41fe4 |
- add dst, dst, tmp1
|
|
|
a41fe4 |
- st1b_unroll
|
|
|
a41fe4 |
- add dst, dst, tmp1
|
|
|
a41fe4 |
- sub rest, rest, tmp2
|
|
|
a41fe4 |
- b 1b
|
|
|
a41fe4 |
|
|
|
a41fe4 |
L(unroll8):
|
|
|
a41fe4 |
lsl tmp1, vector_length, 3
|
|
|
a41fe4 |
@@ -155,7 +139,7 @@ L(L1_prefetch): // if rest >= L1_SIZE
|
|
|
a41fe4 |
sub rest, rest, CACHE_LINE_SIZE * 2
|
|
|
a41fe4 |
cmp rest, L1_SIZE
|
|
|
a41fe4 |
b.ge 1b
|
|
|
a41fe4 |
- cbnz rest, L(unroll32)
|
|
|
a41fe4 |
+ cbnz rest, L(unroll8)
|
|
|
a41fe4 |
ret
|
|
|
a41fe4 |
|
|
|
a41fe4 |
// count >= L2_SIZE
|
|
|
a41fe4 |
--
|
|
|
a41fe4 |
2.31.1
|
|
|
a41fe4 |
|