Floating point results change with different compilation options

Floating point results change with different compilation options

[Cristea Bogdan]

Hi
   I use gcc compiler in order to study the behavior of some iterative
maps (x_n=f(x_{n-1})) using single floating point precision (following
IEEE 754 standard). I have noticed that using different compiler
options, the output sequence could be very different. My programs are
written in C++ using 'float' type and I use 'gcc version 4.1.2
20061115 (prerelease) (Debian 4.1.1-21)'. Compilation options are:
'-Wall -O3 -march=pentium4' or '-Wall'. The same program compiled with
gcc 4.2.2 gives the same results regardless of the compilation options
(on an athlon64 processor).
   Is this a bug of the compiler or the floating point representation
differs from one compiler version to the other?

-- 
Bogdan Cristea

Re: Floating point results change with different compilation options

[John Love-Jensen]

Hi Cristea,

> Is this a bug of the compiler...?

No.  Certain optimizations can exacerbate code that poorly accommodates
float peculiarities.

> ...or the floating point representation differs from one compiler version to
the other?

No, not the representation.  Probably the calculation.

The problem is one-or-both of:
+ not all hardware is compliant with IEEE 754 calculations
+ not all code is written carefully to accommodate float peculiarities

Please read the article What Every Computer Scientist Should Know About
Floating-Point Arithmetic, by David Goldberg (1991).

http://docs.sun.com/source/806-3568/ncg_goldberg.html

HTH,
--Eljay

Re: Floating point results change with different compilation options

[burlen]

Hi,
I think the behavior you described can be expected in some cases. It is 
most likely not a difference in floating point representation(thanks to 
ieee754).  different instruction sequences are expected to produce 
different results thanks to rounding errors, especially in your case 
where the result is used over and over. the errors grow with each 
step!!! (if you haven't seen it before) You have to check out Goldberg 
1992 "What every scientist should know about floating point" he shows 
you how to estimate an upper bound on the rounding error. The bound 
depends on the number and type of and ordering of flops. That is why 
different sequences can produce different results.  Also for curiosity 
sake you might take a close look at the fpu control word, and verify 
that the same mode is being used on both builds. Setting the control 
word is something that the compiler is responsible for but not part of 
the ieee754 standard. It is conceivable(but unlikely) that the FPU is 
being configured differently on the two systems. (see Intel "Programming 
with the x87 fpu"), specifically the rounding control field.

HTH

-- 
Burlen Loring
Information Technologist III
Space Science Center
Institute for the Study of Earth, Oceans, and Space
University of New Hampshire
39 College Road, Durham, NH 03824
Phone: 603-862-1140



Cristea Bogdan wrote:
> Hi
>    I use gcc compiler in order to study the behavior of some iterative
> maps (x_n=f(x_{n-1})) using single floating point precision (following
> IEEE 754 standard). I have noticed that using different compiler
> options, the output sequence could be very different. My programs are
> written in C++ using 'float' type and I use 'gcc version 4.1.2
> 20061115 (prerelease) (Debian 4.1.1-21)'. Compilation options are:
> '-Wall -O3 -march=pentium4' or '-Wall'. The same program compiled with
> gcc 4.2.2 gives the same results regardless of the compilation options
> (on an athlon64 processor).
>    Is this a bug of the compiler or the floating point representation
> differs from one compiler version to the other?
>
>