Hi Julian, On 15.10.19 19:30, Julian Taylor wrote: > This would require disabling the testsuite on these architectures as > with no real long doubles they will likely fail (which I think is the > original reason the package is not built, but I'd have to look it up).
I think this would be a minor problem. > Imo it is also dangerous to provide a long double numerical library on > platforms that do not really support it. You likely will just get wrong > results. >From my C understanding, "long double" is always implementation dependent and is not guaranteed to be longer than "double". So, with a "long double" library one gets only the accuracy that is expected for the number of bits on that platform; this seems obvious to me. > I think it would be better to just not build cubature on these platforms > depending on how significant the long double support is. > Have you engaged with upstream to check if they want to make long double > support optional? The nominal precision (printed with 18 digits) will be a bit better. However, when limiting to 15-16 digits (as they are finally converted to double), almost all numbers are equal (a few deviate in the last digit by 1). This looks acceptable for me. Best Ole
