* [PATCH v4 3/5] AArch64: Improve A64FX memset for remaining bytes
@ 2021-08-09 13:11 Wilco Dijkstra via Libc-alpha
2021-08-10 9:39 ` Szabolcs Nagy via Libc-alpha
0 siblings, 1 reply; 2+ messages in thread
From: Wilco Dijkstra via Libc-alpha @ 2021-08-09 13:11 UTC (permalink / raw)
To: naohirot@fujitsu.com; +Cc: 'GNU C Library'
v4: no changes
Simplify handling of remaining bytes. Avoid lots of taken branches and complex
whilelo computations, instead unconditionally write vectors from the end.
---
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
index 6bc8ef5e0c84dbb59a57d114ae6ec8e3fa3822ad..55f28b644defdffb140c88da0635ef099235546c 100644
--- a/sysdeps/aarch64/multiarch/memset_a64fx.S
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
@@ -130,38 +130,19 @@ L(unroll8):
b 1b
L(last):
- whilelo p0.b, xzr, rest
- whilelo p1.b, vector_length, rest
- b.last 1f
- st1b z0.b, p0, [dst, #0, mul vl]
- st1b z0.b, p1, [dst, #1, mul vl]
- ret
-1: lsl tmp1, vector_length, 1 // vector_length * 2
- whilelo p2.b, tmp1, rest
- incb tmp1
- whilelo p3.b, tmp1, rest
- b.last 1f
- st1b z0.b, p0, [dst, #0, mul vl]
- st1b z0.b, p1, [dst, #1, mul vl]
- st1b z0.b, p2, [dst, #2, mul vl]
- st1b z0.b, p3, [dst, #3, mul vl]
- ret
-1: lsl tmp1, vector_length, 2 // vector_length * 4
- whilelo p4.b, tmp1, rest
- incb tmp1
- whilelo p5.b, tmp1, rest
- incb tmp1
- whilelo p6.b, tmp1, rest
- incb tmp1
- whilelo p7.b, tmp1, rest
- st1b z0.b, p0, [dst, #0, mul vl]
- st1b z0.b, p1, [dst, #1, mul vl]
- st1b z0.b, p2, [dst, #2, mul vl]
- st1b z0.b, p3, [dst, #3, mul vl]
- st1b z0.b, p4, [dst, #4, mul vl]
- st1b z0.b, p5, [dst, #5, mul vl]
- st1b z0.b, p6, [dst, #6, mul vl]
- st1b z0.b, p7, [dst, #7, mul vl]
+ cmp count, vector_length, lsl 1
+ b.ls 2f
+ add tmp2, vector_length, vector_length, lsl 2
+ cmp count, tmp2
+ b.ls 5f
+ st1b z0.b, p0, [dstend, -8, mul vl]
+ st1b z0.b, p0, [dstend, -7, mul vl]
+ st1b z0.b, p0, [dstend, -6, mul vl]
+5: st1b z0.b, p0, [dstend, -5, mul vl]
+ st1b z0.b, p0, [dstend, -4, mul vl]
+ st1b z0.b, p0, [dstend, -3, mul vl]
+2: st1b z0.b, p0, [dstend, -2, mul vl]
+ st1b z0.b, p0, [dstend, -1, mul vl]
ret
L(L1_prefetch): // if rest >= L1_SIZE
@@ -199,7 +180,6 @@ L(L2):
subs count, count, CACHE_LINE_SIZE
b.hi 1b
add count, count, CACHE_LINE_SIZE
- add dst, dst, CACHE_LINE_SIZE
b L(last)
END (MEMSET)
^ permalink raw reply related [flat|nested] 2+ messages in thread
* Re: [PATCH v4 3/5] AArch64: Improve A64FX memset for remaining bytes
2021-08-09 13:11 [PATCH v4 3/5] AArch64: Improve A64FX memset for remaining bytes Wilco Dijkstra via Libc-alpha
@ 2021-08-10 9:39 ` Szabolcs Nagy via Libc-alpha
0 siblings, 0 replies; 2+ messages in thread
From: Szabolcs Nagy via Libc-alpha @ 2021-08-10 9:39 UTC (permalink / raw)
To: Wilco Dijkstra; +Cc: 'GNU C Library'
The 08/09/2021 13:11, Wilco Dijkstra via Libc-alpha wrote:
> v4: no changes
>
> Simplify handling of remaining bytes. Avoid lots of taken branches and complex
> whilelo computations, instead unconditionally write vectors from the end.
OK for commit, but keep
Reviewed-by: Naohiro Tamura <naohirot@fujitsu.com>
>
> ---
>
> diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
> index 6bc8ef5e0c84dbb59a57d114ae6ec8e3fa3822ad..55f28b644defdffb140c88da0635ef099235546c 100644
> --- a/sysdeps/aarch64/multiarch/memset_a64fx.S
> +++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
> @@ -130,38 +130,19 @@ L(unroll8):
> b 1b
>
> L(last):
> - whilelo p0.b, xzr, rest
> - whilelo p1.b, vector_length, rest
> - b.last 1f
> - st1b z0.b, p0, [dst, #0, mul vl]
> - st1b z0.b, p1, [dst, #1, mul vl]
> - ret
> -1: lsl tmp1, vector_length, 1 // vector_length * 2
> - whilelo p2.b, tmp1, rest
> - incb tmp1
> - whilelo p3.b, tmp1, rest
> - b.last 1f
> - st1b z0.b, p0, [dst, #0, mul vl]
> - st1b z0.b, p1, [dst, #1, mul vl]
> - st1b z0.b, p2, [dst, #2, mul vl]
> - st1b z0.b, p3, [dst, #3, mul vl]
> - ret
> -1: lsl tmp1, vector_length, 2 // vector_length * 4
> - whilelo p4.b, tmp1, rest
> - incb tmp1
> - whilelo p5.b, tmp1, rest
> - incb tmp1
> - whilelo p6.b, tmp1, rest
> - incb tmp1
> - whilelo p7.b, tmp1, rest
> - st1b z0.b, p0, [dst, #0, mul vl]
> - st1b z0.b, p1, [dst, #1, mul vl]
> - st1b z0.b, p2, [dst, #2, mul vl]
> - st1b z0.b, p3, [dst, #3, mul vl]
> - st1b z0.b, p4, [dst, #4, mul vl]
> - st1b z0.b, p5, [dst, #5, mul vl]
> - st1b z0.b, p6, [dst, #6, mul vl]
> - st1b z0.b, p7, [dst, #7, mul vl]
> + cmp count, vector_length, lsl 1
> + b.ls 2f
> + add tmp2, vector_length, vector_length, lsl 2
> + cmp count, tmp2
> + b.ls 5f
> + st1b z0.b, p0, [dstend, -8, mul vl]
> + st1b z0.b, p0, [dstend, -7, mul vl]
> + st1b z0.b, p0, [dstend, -6, mul vl]
> +5: st1b z0.b, p0, [dstend, -5, mul vl]
> + st1b z0.b, p0, [dstend, -4, mul vl]
> + st1b z0.b, p0, [dstend, -3, mul vl]
> +2: st1b z0.b, p0, [dstend, -2, mul vl]
> + st1b z0.b, p0, [dstend, -1, mul vl]
> ret
>
> L(L1_prefetch): // if rest >= L1_SIZE
> @@ -199,7 +180,6 @@ L(L2):
> subs count, count, CACHE_LINE_SIZE
> b.hi 1b
> add count, count, CACHE_LINE_SIZE
> - add dst, dst, CACHE_LINE_SIZE
> b L(last)
>
> END (MEMSET)
--
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2021-08-10 9:41 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-09 13:11 [PATCH v4 3/5] AArch64: Improve A64FX memset for remaining bytes Wilco Dijkstra via Libc-alpha
2021-08-10 9:39 ` Szabolcs Nagy via Libc-alpha
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).