Normal random number generators, by J. W. Rider

The NRAND.ARC file contains code for two different Turbo Pascal
units.  One (NRAND0.PAS) contains a Pascal implementation of
algorithms that can be found in the November 1988 issue of
Communications of the ACM.  The other (NRAND1.PAS) contains the
"conventional" methods of generating the same random variates.

In the process of testing NRAND0, I discovered that the
algorithms are extremely sensitive to the use of the 8087 numeric
coprocessor.  I was surprised to discover that the conventional
methods worked BETTER than the advertised "fast" algorithms from
CACM, but only when the 8087 was used.  When the floating point
processing was done in software, the CACM algorithms performed
significantly better.

If you use hardware floating point, use NRAND1 to generate your
normal random variates. (A 2:1 speed advantage over NRAND0.)

If you use software floating point, with or without the Borland
emulation package, use NRAND0. (A 2.3:1 speed advantage over
NRAND1.)

If you have a choice, use hardware floating point (NRAND1).
(A better than 5:1 speed advantage over NRAND0 with software
processing.)

Both units have identical interfaces.

In the interest of portability, both units use the SYSTEM random
function to generate uniformly distributed random variates.  This
could easily be changed to incorporate the RAND unit, which can
be found in RANDOM.ARC, BPROGA DL2.

Copyright 1988, by J. W. Rider.