I’m looking for a good online reference explaining the extended precision multiplication and addition algorithms used in the Mandelbrot demo in the SDK. All I’ve found so far are pointers to journals and some random discussions along the lines of “it works because Dekker and Knuth said so”.

Some pretty simple testing with some fractal programs I wrote for SSE2 show that the performance is “better” (seems faster, haven’t done the measurements yet) than a 3.4 Ghz Pentium D with same calcs run in tandem on both cores (i.e. 4 double precision “threads” in parallel), but 100x less precise. Assuming that 2xsingle precision gives about 2*7=14 decimal digits of precision vs about 16 for true double the100x factor is right on.

I’d still like to see a good reference for this. Really…