This is the mail archive of the
libc-help@sourceware.org
mailing list for the glibc project.
Re: Testsuite for glibc mathematics.
- From: OndÅej BÃlka <neleai at seznam dot cz>
- To: "Ryan S. Arnold" <ryan dot arnold at gmail dot com>
- Cc: Huan Luo <luo_huan_na at 126 dot com>, libc-help <libc-help at sourceware dot org>
- Date: Tue, 27 Nov 2012 00:23:07 +0100
- Subject: Re: Testsuite for glibc mathematics.
- References: <5cca9c63.dd7d.13b3d54bf37.Coremail.luo_huan_na@126.com><CAAKybw-zqFdFQMrt6bpOJjkbsC+Jr3HOUahc8EV9SyrZzoEhJw@mail.gmail.com>
Hi,
I am working on analyzing libc performance. I started at string
functions and broadened my framework also to math functions. Here are
data that I measured in my system. I unify similar functions so for
example sin,cos,sincos,tan are for now labeled as sin.
http://kam.mff.cuni.cz/~ondra/benchmark_string/profile/result.html
There math functions are not called frequently on my computer so about
90% of spent time is caused by factors like loading them code memory.
Any program that frequently uses math functions could solve as
benchmark.
My tool is available on
git clone https://github.com/neleai/benchmark_string
directory profile.
I now work on making it more user-friendly and allowing to compare
several implementations.
A simple example that I just pushed is to test how is sin faster than
sinl on same output, that currently are displayed as sin2/sin3
On Mon, Nov 26, 2012 at 09:48:45AM -0600, Ryan S. Arnold wrote:
> Hi, Huan Luo,
>
> libc-help@sourceware.org is probably the correct place for this question.
>
> None-the-less, run "make check" after doing your build and it will run
> the integrated math/ portion of the glibc test-suite. This is a
> "correctness" test-suite.
>
> As far as a performance test-suite, we don't yet have anything like
> that codified. In general, the contributors that are interested in
> the performance of portions of glibc tend to target them individually
> as they show up hot in workloads. There has been a lot of work lately
> in the math functions in order to make them full featured with regard
> to precision and rounding. This may have regressed performance.
>
> Here is an example of how Adhemerval evaluated (and improved) the
> performance of __ieee754_hypot for the Power Architecture:
>
> http://www.ibm.com/developerworks/opensource/library/l-evaluatelinuxonpower/index.html
>
> In general, performance improvement patches are accompanied by a
> baseline analysis as well as a discussion of the preconditions and
> assumptions that influenced the patches/changes.
>
> Regards,
> Ryan S. Arnold
>
> On Mon, Nov 26, 2012 at 9:27 AM, Huan Luo <luo_huan_na@126.com> wrote:
> > Hi,
> > Lately we've been studying the mathematics part in Glibc.
> > And we made a few changes to some of the functions.
> > The problem is we need a testsuite to tell us how these
> > functions' performance is affected.
> > Can anybody introduce us a testsuite or tell us how
> > people usually test their mathematical functions?
> > Thanks very much.
> >
> >
> >
> >
> > --
> > Best wishes.
> >
> > Huan Luo
--
SCSI Chain overterminated