On Thu, 21 Sep 2023 19:19:34 GMT, Joe Darcy <da...@openjdk.org> wrote:
> The Math.hypot method claims its error bound is one ulp. > > The paper > > "Accuracy of Mathematical Functions in Single, Double, Double > Extended, and Quadruple Precision" > Brian Gladman, Vincenzo Innocente and Paul Zimmermann > September 21, 2023 > https://members.loria.fr/PZimmermann/papers/accuracy.pdf > > lists a known worst-case error of 1.21 ulps for hypot for the "OpenLibm" math > library, which is a derivative of FDLIBM. > > The specification of Math.hypot should be updated to acknowledge the wider > error bound. I changed the allowable error bound to 1.5 ulps is give a bit of > cushion. This pull request has now been integrated. Changeset: b66ded9a Author: Joe Darcy <da...@openjdk.org> URL: https://git.openjdk.org/jdk/commit/b66ded9a5b699e4936db25b58944587432e64f46 Stats: 1 line in 1 file changed: 0 ins; 0 del; 1 mod 8316688: Widen allowable error bound of Math.hypot Reviewed-by: bpb, rgiulietti ------------- PR: https://git.openjdk.org/jdk/pull/15868
