31 Aug
2007
31 Aug
'07
4:45 p.m.
Stjepan Rajko wrote:
On 8/31/07, John Maddock <john@johnmaddock.co.uk> wrote:
Looks like I quashed most of the problems, but still missed a few, also looks like expl is inaccurate on that platform? Can you check what std::exp(-8072.0L) evaluates to: should be 2.371057450122979885196291188186517859e-3506 according to functions.wolfram.com.
It's:
2.371057450122979876482639519674937e-3506
OK that's about 33 eps, which seems to account for the bulk of the error in this case: so that would appear to be why erfc and bessel_k are less accurate on that platform. Shame :-( John. P.S. Are you able to test with Apple's 128-bit "double double" at all? Or is than only on Mach systems? Thanks.