From: Joe Buck <Joe.Buck@synopsys.COM>
To: Paolo Carlini <pcarlini@suse.de>
Cc: Michael Veksler <VEKSLER@il.ibm.com>, gcc@gcc.gnu.org
Subject: Re: tr1::unordered_set<double> bizarre rounding behavior (x86)
Date: Tue, 05 Jul 2005 18:08:00 -0000 [thread overview]
Message-ID: <20050705180759.GA2315@synopsys.com> (raw)
In-Reply-To: <42CABDCC.3070508@suse.de>
On Tue, Jul 05, 2005 at 07:05:16PM +0200, Paolo Carlini wrote:
> Michael Veksler wrote:
>
> > std::tr1::hash<dobule> is implemented in a very bad way.
> > it casts double to size_t, which of course does a very poor job on big
> > values (is the result of 1.0e100 cast to size_t defined ?).
> >
> >
> A possible solution would be using frexp & co to extract the mantissa
> and then work on it, one way or the other. You can find it proposed
> around. Then, often the exponent is simply discarded.
>
> What do you think?
Close, but not quite.
Hash functions are, by nature, many-to-one. A good hash function has
few collisions for values that frequently appear; the program will preform
very poorly if many inputs hash to the same value. The existing function
will make all values over max(size_t) collide (assuming the cast clips).
Using only the mantissa is better, but if doubles are used to
represent fixed-point, then common values like 0.25, 0.5, 1, 2, 4 might
all be in the hash table, and they will collide.
You could do frexp, scale the mantissa to form an integer (e.g. multiply
by a large integer), then add it (modulo 2**n) to some prime number
multiplied by the exponent. This should give good spreading for both
large values and exactly representable values.
next prev parent reply other threads:[~2005-07-05 18:08 UTC|newest]
Thread overview: 32+ messages / expand[flat|nested] mbox.gz Atom feed top
2005-07-05 13:05 Michael Veksler
2005-07-05 13:32 ` Paolo Carlini
2005-07-05 15:43 ` Michael Veksler
2005-07-05 15:57 ` Gabriel Dos Reis
2005-07-05 16:04 ` Paolo Carlini
2005-07-05 20:18 ` Michael Veksler
2005-07-05 20:22 ` Paolo Carlini
2005-07-05 20:36 ` Michael Veksler
2005-07-05 20:41 ` Paolo Carlini
2005-07-05 20:47 ` Paolo Carlini
2005-07-05 21:12 ` Michael Veksler
2005-07-05 21:21 ` Gabriel Dos Reis
2005-07-05 14:28 ` Gabriel Dos Reis
2005-07-05 17:04 ` Paolo Carlini
2005-07-05 18:06 ` Gabriel Dos Reis
2005-07-05 18:10 ` Joe Buck
2005-07-05 18:32 ` Gabriel Dos Reis
2005-07-05 18:42 ` Paolo Carlini
2005-07-05 19:18 ` Gabriel Dos Reis
2005-07-05 19:52 ` Paolo Carlini
2005-07-05 20:24 ` Gabriel Dos Reis
2005-07-05 19:01 ` Michael Veksler
2005-07-05 19:24 ` Gabriel Dos Reis
2005-07-06 12:38 ` Avi Kivity
2005-07-06 12:54 ` Michael Veksler
2005-07-06 13:01 ` Avi Kivity
2005-07-06 13:50 ` Gabriel Dos Reis
2005-07-06 12:54 ` Gabriel Dos Reis
2005-07-05 18:08 ` Joe Buck [this message]
2005-07-05 18:12 ` Paolo Carlini
2005-07-05 19:58 ` Joe Buck
2005-07-05 19:59 ` Paolo Carlini
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
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20050705180759.GA2315@synopsys.com \
--to=joe.buck@synopsys.com \
--cc=VEKSLER@il.ibm.com \
--cc=gcc@gcc.gnu.org \
--cc=pcarlini@suse.de \
/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;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).