On 11/14/05, The Rasterman Carsten Haitzler <[EMAIL PROTECTED]> wrote:
> On Tue, 15 Nov 2005 05:58:52 +0000 Mike Frysinger <[EMAIL PROTECTED]> babbled:
>
> > On Tue, Nov 15, 2005 at 06:38:33AM +0100, Vincent Torri wrote:
> > >
> > > maybe one should test in configure.in if the processor is an mad64 or not,
> > > then define a specific flag according tot the arch and add it to the flags
> > > when one compiles.
> >
> > gcc bugs change greatly between releases ... trying to track which CFLAGS
> > are 'safe' and which ones are not is a huge pita
> >
> > a lot easier to say to the user:
> > dont use stupid optimizations
>
> -funroll-loops... ummm (i just HAD to say that!)
>
> um.. the main problem is gcc4's amd64 build usies sse math by DEFAULT - thats
> the problem.

not just gcc4, i have gcc 3.4.4 here, and it compiles sse by default.

> i frankly dont know what exactly the problewm is - someone being
> able to print a list of results, variables over a series of calculations and
> plot the difference (errors) in them all at all stages (all intermediate
> variable values too) would be useful.

this is exactly what i did. i'll attach my boneheaded program and
results, including binaries in case you want to disassemble or
something. Hopefully it will be useful to you.

Right off the bat, the first thing i notice is that all errors are
introduced in divisions, but not all divisions.

> i personalyl suspect its a bug in the sse
> math either in the processor or compiler simply losing excess precision
> somewhere - either just by the way the fpmath sse is used or by a real "silly"
> bug. as the fp math seems fine on ARM (emulated fp math), ppc fp math, x87
> fpmath etc. and i dont really use extremes of the fp range in calculation as
> best i know... but then again - a good analysis of the situation might be
> useful.
>
>
> --
> ------------- Codito, ergo sum - "I code, therefore I am" --------------
> The Rasterman (Carsten Haitzler)    [EMAIL PROTECTED]
> 裸好多
> Tokyo, Japan (東京 日本)
>
>
> -------------------------------------------------------
> This SF.Net email is sponsored by the JBoss Inc.  Get Certified Today
> Register for a JBoss Training Course.  Free Certification Exam
> for All Training Attendees Through End of 2005. For more info visit:
> http://ads.osdn.com/?ad_id=7628&alloc_id=16845&op=click
> _______________________________________________
> enlightenment-devel mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/enlightenment-devel
>

Attachment: floattest.tar.bz2
Description: BZip2 compressed data

Reply via email to