This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH RFC] Imporve 64bit memset performance for Haswell CPU with AVX2 instruction
- From: OndÅej BÃlka <neleai at seznam dot cz>
- To: Ling Ma <ling dot ma dot program at gmail dot com>
- Cc: "H.J. Lu" <hjl dot tools at gmail dot com>, GNU C Library <libc-alpha at sourceware dot org>, Richard Henderson <rth at twiddle dot net>, Andreas Jaeger <aj at suse dot com>, Liubov Dmitrieva <liubov dot dmitrieva at gmail dot com>, Ling Ma <ling dot ml at alibaba-inc dot com>
- Date: Thu, 5 Jun 2014 18:32:24 +0200
- Subject: Re: [PATCH RFC] Imporve 64bit memset performance for Haswell CPU with AVX2 instruction
- Authentication-results: sourceware.org; auth=none
- References: <1396850238-29041-1-git-send-email-ling dot ma at alipay dot com> <20140513173616 dot GC5047 at domone dot podge> <20140515201458 dot GA24885 at domone dot podge> <CAOGi=dNmn2bPfB65VoXUGjQ7t6RLVJ2hj2QDarrUjZV75kTbDA at mail dot gmail dot com> <20140530113041 dot GB26528 at domone dot podge> <CAOGi=dPdWegEo1s8=wG4WzOANaQ3x=boLFitQ_wBp+Xf+hxexQ at mail dot gmail dot com> <CAMe9rOqv5RYK1MO2M098n3o50-KmmZJuvsvMmXqkBBt0g3OY_g at mail dot gmail dot com> <CAOGi=dMNyzckY8s3uF0qRpKuqUwHHhzQeyy-j29ydLNn_s9Bog at mail dot gmail dot com>
On Wed, Jun 04, 2014 at 03:00:05PM +0800, Ling Ma wrote:
> H.J
>
> The website changed IP, now the code is available again:
> http://www.yunos.org/tmp/memset-avx2.patch ,
> and also gziped as attachment in this mail.
>
> Thanks
> Ling
>
Now performance looks ok for me, but few formating problems.
With these fixed I would be satisfied H.J do you have comments?
There is possible followup to also optimize __bzero like we do in
general case.
Then second followup would be decrease function size by reshuffling
blocks, on several places there are 15/16 free bytes due alignment.
Formatting problems are here:
+ vpxor %xmm0, %xmm0, %xmm0
+ vmovd %esi, %xmm1
+ mov %rdi, %rsi
+ mov %rdi, %rax
here
+L(less_16bytes):
+ vmovd %xmm0, %rcx
+ cmp $8, %dl
+ jb L(less_8bytes)
+ mov %rcx, (%rdi)
+ mov %rcx, -0x08(%rsi)
+ ret
+
+ .p2align 4
+L(less_8bytes):
+ cmp $4, %dl
+ jb L(less_4bytes)
+ mov %ecx, (%rdi)
+ mov %ecx, -0x04(%rsi)
+ ret
and here
+ mov %rax, %rsi
+ vmovd %xmm0, %eax
+ mov %rdx, %rcx
As I mentioned code size one trick is that instructions
with -128 argument are shorter than with 128. You could save 16
bytes with following modification, however it must be tested if
it improves performance.
--- x 2014-06-05 18:20:35.313645591 +0200
+++ sysdeps/x86_64/multiarch/memset-avx2.S 2014-06-05
18:22:25.068642767 +0200
@@ -95,7 +95,6 @@
.p2align 4
L(256bytesormore):
vinserti128 $1, %xmm0, %ymm0, %ymm0
- mov $0x80, %rcx
add %rdx, %rsi
mov %rdi, %r9
vmovdqu %ymm0, (%rdi)
@@ -105,15 +104,15 @@
add %r9, %rdx
cmp $4096, %rdx
ja L(gobble_data)
- sub %ecx, %edx
+ add $-128, %edx
L(gobble_128_loop):
vmovdqa %ymm0, (%rdi)
vmovdqa %ymm0, 0x20(%rdi)
vmovdqa %ymm0, 0x40(%rdi)
vmovdqa %ymm0, 0x60(%rdi)
- add %rcx, %rdi
- sub %ecx, %edx
- jae L(gobble_128_loop)
+ sub $-128, %rdi
+ add $-128, %edx
+ jb L(gobble_128_loop)
vmovdqu %ymm0, -0x80(%rsi)
vmovdqu %ymm0, -0x60(%rsi)
vmovdqu %ymm0, -0x40(%rsi)