Threaded Mode | Linear Mode

Claudio L. · (This post was last modified: 07-07-2017 07:11 PM by Claudio L..)

(06-29-2017 10:30 PM)Claudio L. Wrote: Regarding speed vs size: I was aiming for both! On the PC the new algorithms are 3 times faster than the old ones at 32 digits, but when I finally tested on real hardware it was very disappointing to find gains that were much less. I think it's because the PC can multiply in one cycle, while the small ARM chips take a few cycles per multiplication, but they always add in 1 cycle (so it favors CORDIC). Can't always get what you want, but still the Savage benchmark went from 31 seconds to 20.5 seconds at the default 32 digits.

I tried a few improvements in the multiplication routine (the new algorithms are heavy on multiplications). Speed of multiplication doubled on the PC. On the other hand, on the calculator it didn't even move the needle, the Savage benchmark went from 20.3 sec to 20.0 seconds. I don't think the compiler will do better, even when taken by the hand. Time to try some hand-optimized assembler... just to prove a point.