Paris (U.E.), le 23/03/2005 Bonjour In article <989aceac050322131460c7f704@mail.gmail.com>, Caleb Epstein <caleb.epstein@gmail.com> wrote:

On Wed, 16 Mar 2005 15:58:37 +0100, Hubert Holin <Hubert.Holin@meteo.fr> wrote:

...

I am currently investigating some regressions for my quaternion and octonion libraries, which appear in the report of the tests you run under SunOS5.

I have the feeling that for "long double" variables, numeric_limits<long double>::epsilon() returns the correct value, but standard transcendental functions (atan at least) behave as if their input were "double" and not "long double".

Do you know if this is the case?

Sorry for the late reply.

I believe you're correct. I don't see any prototypes for atan() that take a long-double or that are templatized, but indeed std::numeric_limits behaves correctly:

numeric_limits<double>: digits=53 digits10=15 max_exponent=1024 max_exponent10=308 min_exponent=-1021 min_exponent10=-307 radix=2 episilon=2.22045e-16 numeric_limits<long double>: digits=113 digits10=33 max_exponent=16384 max_exponent10=4932 min_exponent=-16381 min_exponent10=-4931 radix=2 episilon=1.92593e-34

Thanks for the information. This is the first platform where I see this actually happening. I will modify my tests so quaternions and octonions pass for that platform, even for "long double", though with a message saying that quality may be degraded for that type. Merci Hubert Holin