fast-math question.

neuron · Advocate Joined: 28 May 2002 Posts: 2371

Was just reading through http://www.coyotegulch.com/reviews/gcc4/index.html
and noticed:

gentoo_lan · l33t Joined: 08 Sep 2004 Posts: 889 Location: Denver, CO

Well maybe GCC 4 has improved the use of ffast-math. I don't think that I would use it for GCC 3.3 or 3.4.
_________________
Registered Linux user# 375038.

lightvhawk0 · Guru Joined: 07 Nov 2003 Posts: 388

Lol the funny part is on the new opteron/gcc 4.0 test mfpmath=387 with -funsafe-math-optimizations held a %50 improvement on fpu intensive apps. So yeah they've improved on it.
_________________
If God has made us in his image, we have returned him the favor. - Voltaire

neuron · Advocate Joined: 28 May 2002 Posts: 2371

nbkolchin · Posted: Sat Aug 27, 2005 10:29 pm Post subject:

What problems have "fast-math" with 2.9, 3.x GCC versions? It always generate faster code.

rhill · Developer Joined: 22 Oct 2004 Posts: 1629 Location: sk.ca

er, yes, using -ffast-math generates faster running FP code. this has always been true. the problem is it breaks a few standards to do so, which some programs depend on, specifically assumptions about precision on certain architectures. the end result is anything that requires floating point operations to do what they're expected to do doesn't work properly (and that's quite a few packages).

btw, i'd take anything found on scott's site with a grain of salt.
_________________
by design, by neglect
for a fact or just for effect

perga · n00b Joined: 18 Sep 2004 Posts: 68 Location: SWEDEN

I recently put together a small mandelbrot set viewer, mainly for benchmarking purposes. Compiling it with -ffast-math produced some impressive speedups, generally the rendering times were reduced by 50 - 66 %. This was using the Gentoo vanilla gcc-3.5.5.
_________________
perga

nxsty · Veteran Joined: 23 Jun 2004 Posts: 1556 Location: .se

Most apps in portage compiles and runs with -ffast-math and those who breaks usually filters it since it's a commonly used flag. The only packages I know that doesn't filter it is perl and libperl which fails a few test when compiled with -ffast-math.

This is a quote from one of the gcc devs about it from http://gcc.gnu.org/ml/gcc/2005-05/msg01283.html:
"- simple builtin functions (sqrt, sin and cos, etc) can be used with hardware
implemented fsqrt, fsin and fcos insn instead of calling library functions.
These can handle arguments up to (something)*2^63, so it is quite enough for
normal apps. This way, a call overhead (saving all FP registers, overhead of
call insn itself) is eliminated, and as an added bonus, sin and cos of the same
argument are combined as fsincos insn."

So I guess that quote from coyotegulch.com might be at least partly true.