I have looked at this paper here which describes how to approximate the exponential function for double precision. After a little research on Wikipedia about single precision floating point representation I have worked out the equivalent algorithms. They only implemented the exp function, so I found an inverse function for the log and then simply did
POW(a, b) = EXP(LOG(a) * b).
compiling this gcc4.6.2 yields a pow function almost 4 times faster than the standard library’s implementation (compiling with O2).
Note: the code for EXP is copied almost verbatim from the paper I read and the LOG function is copied from here.
Here is the relevant code:
#define EXP_A 184
#define EXP_C 16249
float EXP(float y)
{
union
{
float d;
struct
{
#ifdef LITTLE_ENDIAN
short j, i;
#else
short i, j;
#endif
} n;
} eco;
eco.n.i = EXP_A*(y) + (EXP_C);
eco.n.j = 0;
return eco.d;
}
float LOG(float y)
{
int * nTemp = (int*)&y;
y = (*nTemp) >> 16;
return (y - EXP_C) / EXP_A;
}
float POW(float b, float p)
{
return EXP(LOG(b) * p);
}
There is still some optimization you can do here, or perhaps that is good enough.
This is a rough approximation but if you would have been satisfied with the errors introduced using the double representation, I imagine this will be satisfactory.