… = lim_{n->inf} (1+1/n)^n, which uses 1 double division, 1 double addition and 12 double multiplications, with Shraudolph's approximation from https://www.schraudolph.org/pubs/Schraudolph99.pdf, which uses 1 double multiplication, 1 integer addition, 1 integer shift.

This makes exp2 disappear from the profile entirely in our application, and reduced the time spent in sta::findRoot by 12%.

For our application, this reduces

Signed-off-by: Rasmus Munk Larsen <rmlarsen@google.com>

…lph's

paper for easier reference.

Signed-off-by: Rasmus Munk Larsen <rmlarsen@google.com>

…roximation is accurate. Fortunately, this is almost the entire range in which exp(x) is finite.

Signed-off-by: Rasmus Munk Larsen <rmlarsen@google.com>

Signed-off-by: Rasmus Munk Larsen <rmlarsen@google.com>

…nitude in the common case.

Signed-off-by: Rasmus Munk Larsen <rmlarsen@google.com>

Signed-off-by: Matt Liberty <mliberty@precisioninno.com>