You don't because I haven't <g>.  Actually, I've got one just about
ready.  It is Linear regression test that I stripped out of one of my
simulators.  I'll try and get it up by the end of the day.  Basically,
I'm thinking that each version should be a unit with a 'test' procedure
that can be called and timed.  That way one test rig can handle any of
the tests.  Also I need to throw together a quickie timing mechanism
that can is more accurate than GetTickCount and can be implemented in
all enviroments without pulling in all of the Windows API.

Sounds good, we can organize as we go.  

Based on my own prelim. test from about a year ago.  It is very easy to
skew the results merely by including or excluding specific functions.
This is the "trick" that unscrupulus VC'ers have used to make Borland
look really bad at FP.  They always bury an ATAN in it somewhere that VC
code does comparatively fast (but with some sloppiness) but is still so
slow that it swamps all the remaining FP computations.  This is the
reason that I wanted to base this on "real" world code.  What I'm trying
to say is that if you just cook up your benchmark, your are likely to
also be cooking your biases into the results.

--
Bob Lee
High Performance Delphi - http://www.econos.com/optimize/
Updated January 20