newRPL: Now for 39GS and 40GS! [June 29-2017]

[Claudio L.]

(06-30-2017 12:01 AM)Helix Wrote:  There is the possibility to implement conventional routines for the default 32 digits precision.
NewRPL is noticeably faster than RPL, but I think it still doesn't do justice to the ARM processor for 99.9% of calculations, when more than 32 digits are not required.

True, next time I decide to spend 2 months rewriting those routines, I'll take that into account and code a special case for 32 digits. For now the objective was achieved: 20 sec vs. 31 sec is a decent improvement at 32 digits (1.5x), and those huge tables (900 kbytes) are gone for good.
The comparison at 12 digits precision isn't that meaningful due to overhead and lack of tuning (why would you use 16 guard digits at 12-digit precision?) It does the benchmark in 15 seconds, but at 32 digits you get 2.7x digits at only 33% more time.