Performance Comparison in Computing Exponential Function

If your computation is dominated by exponential function evaluations, then it makes a significant difference whether you evaluate the exponential function exp() in single precision or in double precision. You can reduce your computing time by roughly 25% when moving from double precision (double) to single precision (float). Evaluation in quadruple precision is more than six times more expensive than evaluation in double precision.

Changing from double precision to single precision also halves the amount of storage needed. On x86_64 Linux float usually occupies 4 bytes, double occupies 8 bytes, and long double needs 16 bytes.

1. Result. Here are the runtime numbers of a test program.

1. Single precision (float): 2.44s
2. Double precision (double): 3.32s
3. Quadruple precision (long double): 22.88s

These numbers are dependant on CPU internal scheduling, see CPU Usage Time Is Dependant on Load.

2. Test program. The test program is essentially as below:

long i, rep=1024, n=65000;
int c, precision='d';
float sf = 0;
double sd = 0;
long double sq = 0;
...
switch(precision) {
case 'd':
        while (rep-- > 0)
                for (i=0; i<n; ++i)
                        sd += exp(i % 53) - exp((i+1) % 43) - exp((i+2) % 47) - exp((i+3) % 37);
        printf("sd = %f\n",sd);
        break;
case 'f':
        while (rep-- > 0)
                for (i=0; i<n; ++i)
                        sf += expf(i % 53) - expf((i+1) % 43) - expf((i+2) % 47) - expf((i+3) % 37);
        printf("sf = %f\n",sf);
        break;
case 'q':
        while (rep-- > 0)
                for (i=0; i<n; ++i)
                        sq += expl(i % 53) - expl((i+1) % 43) - expl((i+2) % 47) - expl((i+3) % 37);
        printf("sq = %Lf\n",sq);
        break;
}

Full source code is in GitHub, file in question is called exptst.c.

3. Environment. AMD Bulldozer FX-8120, 3.1 GHz, Arch Linux 5.6.8, gcc version 9.3.0. Compiled the code with -O3 -march=native