Re: floating point exceptions: Linux kernel feature or bug?

["Wes Bauske" <wsb paralleldata com>]

Aaron M. Johnson wrote:

> Hi,
>
>         Your message about the floating point exception sounds related
> to a problem I have.  For me it appears to be an underflow.  Keep me
> posted if you find a solution.
>
> Thank you,
> aaron
>
> I am running Persistance of Vision ray tracer (povray) compiled on
>
> [amj@testnode branch]$ uname -a
> Linux testnode.math.ukans.edu 2.1.96 #2 SMP Thu Apr 16 19:05:54 CDT 1998 alpha unknown
> [amj@testnode branch]$ gcc --version
> 2.7.2.3
>
> This code generates a picture as output using Intel-Linux and IRIX 6.3
>
> [amj@testnode branch]$ gdb ~/bin/povray.debug
> (gdb) run +L/home/amj/thesis/povray3/include +Imov +KFF1 +FP
> Starting program: /home/amj/bin/povray.debug +L/home/amj/thesis/povray3/include +Imov +KFF1 +FP
>
>  <irrelevant povray output deleted>
>
> Parsing...........
>
> Creating bounding slabs.
> Scene contains 1 frame level objects; 0 infinite.
>
>  00:00:00 so far, Rendering frame 1, going to 1.
>
> Rendering...
> Program received signal SIGFPE, Arithmetic exception.
>
> 0x120045dcc in do_specular (Finish=0x120211ce0, Light_Source_Ray=0x11fffec38,
>     REye=0x11ffff5d8, Layer_Normal=0x11ffff6e0, Colour=0x11ffff758,
>     Light_Colour=0x11ffff5f0, Layer_Pigment_Colour=0x11ffff710)
>     at ../lighting.c:1688
> 1688          Intensity = Finish->Specular * pow(Cos_Angle_Of_Incidence, Finish->Roughness);
> (gdb) list 1680
> 1675
> 1676      VHalf(Halfway, REye, Light_Source_Ray->Direction);
> 1677
> 1678      VLength(Halfway_Length, Halfway);
> 1679
> 1680      if (Halfway_Length > 0.0)
> 1681      {
> 1682        VDot(Cos_Angle_Of_Incidence, Halfway, Layer_Normal);
> 1683
> 1684        Cos_Angle_Of_Incidence /= Halfway_Length;
> (gdb)
> 1685
> 1686        if (Cos_Angle_Of_Incidence > 0.0)
> 1687        {
> 1688          Intensity = Finish->Specular * pow(Cos_Angle_Of_Incidence, Finish->Roughness);
> 1689
> 1690          if (Finish->Metallic > 0.0)
> 1691          {
> 1692            /*
> 1693             * Calculate the reflected color by interpolating between
> 1694             * the light source color and the surface color according
> (gdb) p Halfway_Length
> $1 = 0.7136288436493563
> (gdb) p Cos_Angle_Of_Incidence
> $2 = 0.49138409073828376
> (gdb) p Intensity
> $3 = 2.3872427115046736e-314
> (gdb)
>
> Don't know the gdb syntax to print values within structure Finish->

Use "p *structptrname" or just 'p structname".

Be careful about assuming there is no bug in the code. What you show
here is similar to what I found in my own code when I first ported it
to Alpha. It looks like an underflow but the cause could be more than
one thing. For example, somehow pointers get screwed up and you
attempt to use a int when it was supposed to be a float. Course it
could also be a true underflow. Recursive agorithms can easily
underflow. Choices are to build using the ieee flags or figure out
how to silently handle the underflows. A while back, someone
talked about setting FP modes on the Alpha processor. I started
to look at it but found it was easier to debug my code, plus the
code is now correct vs previously using garbage as valid data.


Wes