Very fast accurate approximation to Math.exp in java using floating point bit hacks.

fastexp.java

/* fast floating point exp function
 * must initialize table with buildexptable before using
 
Based on 
 A Fast, Compact Approximation of the Exponential Function
 Nicol N. Schraudolph 1999

Adapted to single precision to improve speed and added adjustment table to improve accuracy.
Alrecenk 2014

 * i = ay + b
 * a = 2^(mantissa bits) / ln(2)   ~ 12102203
 * b = (exponent bias) * 2^ ( mantissa bits) ~ 1065353216
 */
public static float fastexp(float x){
	final int temp = (int)(12102203 * x + 1065353216) ;
	return Float.intBitsToFloat(temp)*expadjust[(temp>>15)&0xff] ;
}

static float expadjust[] ;
//build correction table to improve result in region of interest
//if region of interest is large enough then improves result everywhere
public static void buildexptable(double min, double max, double step){
	expadjust = new float[256];
	int amount[] = new int[256] ;
	//calculate what adjustments should have been for values in region
	for(double x=min; x < max;x+=step){
		double exp = Math.exp(x);
		int temp = (int)(12102203 * x + 1065353216) ;
		int index = (temp>>15)&0xff ;
	    double fexp = Float.intBitsToFloat(temp);
		expadjust[index]+= exp/fexp ;
		amount[index]++;
	}
	//average them out to get adjustment table
	for(int k=0;k<amount.length;k++){
		expadjust[k]/=amount[k];
	}
}

Looks sorta crazy, are you using this in other projects?

Planning to use it for neural network stuff. Sigmoid and RBF activation functions both make heavy use of the exp function.