Hello Regina;
I looked into the AOO code and the situation is actually a less critical than I thought: we don't use the system libraries for the hyperbolic functions but instead we have our own implementations in the SAL layer. ----- Messaggio originale ----- > Da: Regina Henschel > > Hi Pedro, hi Andrew, > > Pedro Giffuni schrieb: ... >> >> I did just a very small set of changes for asinh, acosh, atanh and a some >> internal power functions .. just for testing. >> >> Since there is interest in this I opened a Bugzilla issue with the patch: >> https://issues.apache.org/ooo/show_bug.cgi?id=121561 >> >> >> There's also a spreadsheet with some basic tests. I haven't been able to look at the boost implementation. They do have some tests but from the implementation description we could improve our testing.. I will probably look at that next year :). >>> Would want to compare expected group actual and old to new. >>> >>> Would want to devise test cases against both common and edge cases. >>> >>> Also curios about time impact, does it take more time or less. >>> > > Yes, the needed calculation time is critical. There are some places in > the code, where I had stopped a loop at a fixed count, so that the > calculation time is not too long. > To be honest, computers are so fast these days that I think most people may sacrifice some performance in exchange for accuracy. >> >> >> The differences are insignificant comparing FreeBSD amd64 (with boost) >> vs a VM running Windows XP with Symphony, but I am sure someone >> can come up with more creative tests :). > > For accuracy you need a comparison to a value which is calculated > without the restriction to data type double. For single values Wolfram > Alpha might work, for a larger set of test cases a computer algebra > system might be appropriate. > That's a good idea, thanks! While in Germany I actually met someone from REDUCE, I 'll have to get back to that. cheers, Pedro.