This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] ARM: Add optimized ARMv7 strcmp implementation
- From: Richard Earnshaw <rearnsha at arm dot com>
- To: Matt Turner <mattst88 at gmail dot com>
- Cc: "Joseph S. Myers" <joseph at codesourcery dot com>, Will Newton <will dot newton at linaro dot org>, "libc-alpha at sourceware dot org" <libc-alpha at sourceware dot org>
- Date: Mon, 28 Apr 2014 10:00:18 +0100
- Subject: Re: [PATCH] ARM: Add optimized ARMv7 strcmp implementation
- Authentication-results: sourceware.org; auth=none
- References: <1398256338-24784-1-git-send-email-will dot newton at linaro dot org> <Pine dot LNX dot 4 dot 64 dot 1404241437500 dot 6535 at digraph dot polyomino dot org dot uk> <535A42E8 dot 906 at arm dot com> <CAEdQ38FZz-NzNiiA5h4Rb-E2asuVCoN=H0b+O15-vjVXx6E+iA at mail dot gmail dot com>
On 25/04/14 20:30, Matt Turner wrote:
> On Fri, Apr 25, 2014 at 4:11 AM, Richard Earnshaw <rearnsha@arm.com> wrote:
>> These are the improvements for a production board with cortex-a15. For
>> very short strings there's a fair amount of noise in the measurements,
>> but the regressions are generally very small, while the improvements can
>> be significant. Yes, there are cases where the new code is more than
>> three times faster.
>>
> [snip]
>>
>> Length 1024 alignment 0/0: 238.38%
>> Length 1024 alignment 0/0: 239.21%
>> Length 1024 alignment 0/0: 247.45%
>
> While for these cases it's clear from the surrounding context what
> these numbers mean, I recommend not using percentages to describe
> performance improvements. I'd be nice for glibc to adopt this policy
> for sake of consistency and clarity.
>
> If I saw in isolation
>
>> Length 16 alignment 0/0: 105.97%
>
> I wouldn't know whether this was measured relative to the original
> performance (and was 5% faster) or as a speed up over the original
> performance (and was a bit more than twice as fast).
>
Yes, I pondered this before posting the numbers. I concluded that the
presence of a few small negative numbers (for the regressions) made the
whole table unambiguous.
> This could be unambiguously written as
>
>> Length 16 alignment 0/0: 2.0597x
>
If there's a need to re-submit, I'll do that.
R.