[Bf-committers] Floating point compares

[Ton Roosendaal]

Hi,

Floats can guaranteed express values that are powers of 2. I mean a  
range of:
... 0.25 - 0.5 - 1.0 - 2.0 - 4.0 - ..... including zero.

No reason to obsfucate the code for these cases.
Different is when you want to check on values that come back from float  
math. A float can express a minus zero! I did a lot of fixing on this  
when porting to linux, freebsd, solaris, beos, 64 bits alpha, and  
windows. Just something to be aware of... it's the reason we have a  
'safe acos' for example.

I also noticed someone adding an 'f' to float constants. I don't know  
what this is good for... the compiler really can do that thinking for  
you, or not?
Just always use a point precision to express it's a float constant.

-Ton-

(btw; inline is C++ only)


On Wednesday, Oct 29, 2003, at 12:25 Europe/Amsterdam, Kester Maddock  
wrote:

> Hi guys,
>
> I've just noticed that there are a lot of floating point compares in  
> the code, ie:
>        if(tex->cropxmin==0.0) ret++;
>        if(tex->cropymin==0.0) ret++;
>        if(tex->cropxmax==1.0) ret++;
>        if(tex->cropymax==1.0) ret++;
>
> Isn't this bad? (because of the finite floating point precision.)
>
> #define FPRECISION 1e-10
> static inline int fzero(float x)
> {
>    return fabs(x) < FPRECISION;
> }
>
> static inline int fequal(float x, float y)
> {
>    return fzero(x - y);
> }
>
>        if(fzero(tex->cropxmin)) ret++;
>        if(fzero(tex->cropymin)) ret++;
>        if(fequal(tex->cropxmax, 1.0)) ret++;
>        if(fequal(tex->cropymax, 1.0)) ret++;
>
> GCC will catch this with -Wfloat-equal.
>
> Should I start fixing this or is there a valid reason to do it this  
> way?
>
> Kester
>
> _______________________________________________
> Bf-committers mailing list
> Bf-committers@blender.org
> http://www.blender.org/mailman/listinfo/bf-committers
>
>
------------------------------------------------------------------------ 
--
Ton Roosendaal  Blender Foundation ton@blender.org  
http://www.blender.org