rumble -
No I haven't tried a GPU version, my general usage pattern for this library is real-time generation of geometry, and my GPU is already loaded down pretty heavily drawing said geometry! But I will try it at some point.
Andy -
From what I can make out, performing float operations during an SSE task causes the processor to switch from SSE mode to FP mode. Is this correct? So far I have found that anywhere I use a float variable it will cause the next SSE operation to produce indefinite or NaN results.
Yes Core 2 Duos far out perform Pentium D's so I'm sure its not your algorithm that is at fault!
As far as enabling SSE2 making my C++ version run slower I have no idea, but its definitly happening! I have had problems in the past with enabling SSE2 causing projects to crash when they otherwise would not with either SSE1 or no extensions. The final speed figures I quoted were the optimal ones I could get from MSVC2005 sp1 which was using:
Optimization : Maximize Speed
Inline : Any suitable
Favor : Fast Code
Enhanced instruction set: Not set/SSE1 (no difference)

/edit
Just thought I would leave a note here to say that the remainder part of the calculation was bugged, I have fixed this, along with all the type cast warnings and uploaded a new version under the same file name, so the above link will still work.

[Edited by - bluntman on June 17, 2008 8:13:06 PM]