Source code is from Boost:
http://www.boost.org/doc/libs/1_46_1/boost/math/special_functions/erf.hpp
with appropriate modifications for FFmpeg.

Tested on interval -6 to 6 (beyond which it saturates), +/-NAN, +/-INFINITY
under -fsanitize=undefined on clang to test for possible undefined behavior.

This function turns out to actually be essentially as accurate and faster than the
libm (GNU/BSD's/Mac OS X), and I can think of 3 reasons why upstream
does not use this:
1. They are not aware of it.
2. They are concerned about licensing - this applies especially to GNU
libm.
3. They do not know and/or appreciate the benefits of rational
approximations over polynomial approximations. Boost uses them to great
effect, see e.g swr/resample for bessel derived from them, which is also
similarly superior to libm variants.

First, performance.
sample benchmark (clang -O3, Haswell, GNU/Linux):

3e8 values evenly spaced from 0 to 6
time (libm):
./test  13.39s user 0.00s system 100% cpu 13.376 total
time (boost based):
./test  9.20s user 0.00s system 100% cpu 9.190 total

Second, accuracy.
1e8 eval pts from 0 to 6
maxdiff (absolute): 2.2204460492503131e-16
occuring at point where libm erf is correctly rounded, this is not.

Illustration of superior rounding of this function:
arg   : 0.83999999999999997
erf   : 0.76514271145499457
boost : 0.76514271145499446
real  : 0.76514271145499446

i.e libm is actually incorrectly rounded. Note that this is clear from:
https://github.com/JuliaLang/openlibm/blob/master/src/s_erf.c (the Sun
implementation used by both BSD and GNU libm's), where only 1 ulp is
guaranteed.

Reasons it is not easy/worthwhile to create a "correctly rounded"
variant of this function (i.e 0.5ulp):
1. Upstream libm's don't do it anyway, so we can't guarantee this unless
we force this implementation on all platforms. This is not easy, as the
linker would complain unless measures are taken.
2. Nothing in FFmpeg cares or can care about such things, due to the
above and FFmpeg's nature.
3. Creating a correctly rounded function will in practice need some use of long
double/fma. long double, although C89/C90, unfortunately has problems on
ppc. This needs fixing of toolchain flags/configure. In any case this
will be slower for miniscule gain.
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Ganesh Ajjanagadde <gajjanagadde@gmail.com>