From: Jim Gettys <jg@freedesktop.org>
To: Dave Taht <dave.taht@gmail.com>
Cc: codel@lists.bufferbloat.net
Subject: Re: [Codel] fp sqrt vis int sqrt?
Date: Fri, 04 May 2012 11:26:05 -0400 [thread overview]
Message-ID: <4FA3F50D.7080406@freedesktop.org> (raw)
In-Reply-To: <CAA93jw4wzVLrMOq+RzP2FTHCHaPrFAY06xJX=Q94nPO0z2-OcQ@mail.gmail.com>
On 05/04/2012 11:19 AM, Dave Taht wrote:
> On Fri, May 4, 2012 at 8:14 AM, Jim Gettys <jg@freedesktop.org> wrote:
>> On 05/04/2012 11:11 AM, Dave Taht wrote:
>>> The linux kernel has no floating point in it, so I'd substituted the
>>> internal int_sqrt as a substitute,
>>> just to get something to work.
>>>
>>> static inline ktime_t control_law(const struct codel_sched_data *q, ktime_t t)
>>> {
>>> return ktime_add_ns(t, q->interval / int_sqrt(q->count));
>>> }
>>>
>>> Often ns2 models use floating point. Having not seen the model I don't
>>> know that for sure.
>>>
>>> The series for an integer sqrt is far more 'chunky' than a fp one.
>>>
>>> int sqrt 1,2,3,4 = 1 1 1 2
>>> fp sqrt 1 2 3 4 = 1 1.4.1 1.73 2
>>>
>>> and gets even more chunky as you get larger values, eg, sqrt(36
>>> through 48) = 6, sqrt(49) = 7
>>>
>>> So we could precalculate the interval/sqrt(count) using floating point
>>> in the control law calculation,
>>> early, during qdisc setup, thus neatly avoiding both the divide and
>>> sqrt in the control law path.
>>>
>>> Or we could do fixed point.
>>>
>>> Kathie: In the sim, in various simulations, what is the dynamic range
>>> of count? And is it floating point sqrt?
>>>
>> The Wikipedia article on fast inverse sqrt is enlightening on computing
>> this really fast.
>>
>> See:
>>
>> http://en.wikipedia.org/wiki/Fast_inverse_square_root
>>
>> This has little to do with whether an integer or floating point value is
>> better algorithmically.
> Neat trick, that.
>
> I'm delighted the wikipedia article kept the fully commented original
> code around!
I don't think that is the "original", which predates quake by a lot
according to the article.
>
> That said, you can't use FP at all without special handling in the
> kernel. It's certainly a terrible idea on a hot path...
I suspect there are fully integer implementations kicking around...
It's just that x86 finally got fast enough FP that doing one multiply in
floating point beat the alternative.
- Jim
next prev parent reply other threads:[~2012-05-04 15:26 UTC|newest]
Thread overview: 18+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-05-04 15:11 Dave Taht
2012-05-04 15:14 ` Jim Gettys
2012-05-04 15:19 ` Dave Taht
2012-05-04 15:26 ` Jim Gettys [this message]
2012-05-04 15:56 ` Eric Dumazet
2012-05-04 17:23 ` Dave Taht
2012-05-04 17:43 ` Dave Taht
2012-05-04 17:44 ` Dave Taht
2012-05-04 17:46 ` Dave Taht
2012-05-04 23:42 ` Kathleen Nichols
2012-05-05 0:01 ` Dave Taht
2012-05-04 17:47 ` Eric Dumazet
2012-05-04 18:26 ` Eric Dumazet
2012-05-04 18:39 ` Dave Taht
2012-05-04 18:50 ` Eric Dumazet
2012-05-04 19:04 ` Dave Taht
2012-05-04 19:16 ` Eric Dumazet
2012-05-05 8:14 ` Eric Dumazet
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
List information: https://lists.bufferbloat.net/postorius/lists/codel.lists.bufferbloat.net/
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=4FA3F50D.7080406@freedesktop.org \
--to=jg@freedesktop.org \
--cc=codel@lists.bufferbloat.net \
--cc=dave.taht@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox