[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: pb with -ffast-math
From: |
Jaroslav Hajek |
Subject: |
Re: pb with -ffast-math |
Date: |
Wed, 15 Apr 2009 19:56:30 +0200 |
On Tue, Apr 14, 2009 at 9:54 AM, Alain Baeckeroot
<address@hidden> wrote:
>
> Hi
>
> i builded octave with additional optimisation flags
>> FFLAGS = -O -mieee-fp -march=native
>> XTRA_CFLAGS = -mieee-fp -march=native
>> XTRA_CXXFLAGS = -mieee-fp -march=native
>
> Adding just -march=native is fine, and it seems that the build
> is significantly faster (10-20%) than the one shipped with my distro
> (debian lenny 32bit, on core2duo T5550)
>
That's highly probable. The distro binaries are meant to be universal
and so are not typically compiled with march.
> All tests are OK (but 2 expected failures).
>
> when i add the -ffast-math option, it builds, but segfault at runtime:
Currently, the flags are used for nearly all the code, and ffast-math
breaks certain parts, especially those examining fp environment at
runtime. Maybe this will be resolved in the future.
> $ make
> ...
> Octave successfully built. Now choose from the following:
>
> ./run-octave - to run in place to test before installing
> make check - to run the tests
> make install - to install
>
> make[1]: quittant le répertoire « /opt/octave-3.0.5 »
>
> $ ./run-octave
> ** On entry to octave parameter number 13 had an illegal value
> Erreur de segmentation
>
>
> 1/ Is it normal ?
Yes, see above.
> 2/ Are there some optimisation flags i should set to
> improve speed. We have numerous independant short computations, and
> need no precision (2 significant digit is good enought)
> I thought to add -funsafe-math-optimizations -ffinite-math-only ,
> but only after trying simpler stuff.
>
I use -O3 -march=native -funroll-loops and get good results.
> And this is really not important, i doubt we can get another 20%
> speed improvement.
>
> Alain.
Depending on your code, you may get even much better improvements by
switching to the development sources (it seems you use 3.0.5). An
instance is the experience of F. Potorti:
http://www.nabble.com/thanks-for-the-indexing-improvements-td22494818.html#a22494818
regards
--
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz
Re: pb with -ffast-math, Michael Creel, 2009/04/19