Quality comparison of floating-point maths libraries
r-bloggers.com
r-bloggers.com
"The results for R/Linux C were almost identical and a quick check of the R source tree showed that R calls the C library function to evaluate log"
thus making the whole article an exercise in redundancy.
Sure, the linked article scans all the single-precision floats in a certain range to do the test, but the function eventually called (using that list of floats) is log in libm.
In any event, no discrepancy is noted, so what was the point?
Additionally, TFA does not make clear that there should be no errors in the computed function affecting more than the least sig bit. It uses absolute errors (up to 50 e-17) rather than looking at the relative error (or ULP, in the jargon). It seems like a mess.
Here's a link with much more useful information:
http://www.gnu.org/s/libc/manual/html_node/Errors-in-Math-Fu...
Another common library is: http://www.ibm.com/software/awdtools/mass/ and http://www.ibm.com/systems/software/essl/ both from IBM.
Writing good math libraries is a huge undertaking, and there aren't that many versions out there.