Re: [avr-libc-dev] tanh in libm

[René Liebscher]

E. Weddington wrote:
> René Liebscher wrote:
>
> > Hi,
> >
> > I have question about the tanh function in libm.
> >
> > As I can see in
> > http://savannah.nongnu.org/cgi-bin/viewcvs/avr-libc/avr-libc/libm/fplib/tanh.S?rev=1.2&content-type=text/vnd.viewcvs-markup 
> >
> > there are two ways to calculate the result.
> >
> > The questions is why was sinh/cosh choosen. (The other way is 
> > commented out.)
> > Comparing which functions are called to calculate tanh I find 
> > sinh/cosh needs one call to exp, two calls to inverse and one call to 
> > ldexp more than the other way. So I think the current solution is 
> > slower than the other.
> > (And it takes more memory if you don't use sinh or cosh in your program.)
>
> AFAIK, the whole math library in avr-libc was a port of Michael Stumpf's 
> FPLib library to the AVR. See the readme.fplib file in CVS:
> <http://savannah.nongnu.org/cgi-bin/viewcvs/avr-libc/avr-libc/libm/fplib/readme.fplib?rev=1.1&content-type=text/vnd.viewcvs-markup> 
>
> So a lot of the decisions on how to do things were very likely his.

I have tried the part he has not used anymore. It results always in -1, 
but I cannot see what is wrong with it. Writing it in C++, it works.
---------
inline double tanh(double x){
    double e=exp(ldexp(x,1));
    return (e-1)/(e+1);
}
---------
and my program is now about 40% faster. (it uses the tanh function many 
times, online learning of a neural network)

> > If the other solution brings also correct results and is faster, is 
> > there a simple way to integrate it in my program without to recompile 
> > the math library.
>
> Probably only if you're making your own tanh function then you wouldn't 
> have to recompile the library. Obviously if you wanted it to be a part 
> of the library then it would have to be rebuilt.
Maybe someone should check if it wouldn't be better replace the tanh in 
the math library.

> HTH
> Eric

Kind regards
Rene liebscher