jzmaddock b4ec109b83 tanh_sinh: change calculation method used for scaling ranges: ...

"naive" method is prone to generating a point at the endpoint if the endpoint is large, subtract scaled complement from the endpoint instead.
Open questions:
1) Can this still fail if the endpoint is a very large value?  For example if we have endpoints of [+large, +larger] where the difference between them is small?
2) Do we still need the asserts, or should they be replaced by some other check?
3) If the endpoints are replaced by some other check we could either massage the position to be one representable value below the endpoint, or else simply return 0.  The former risks evaluating at the same point over and over as we go down levels, the latter would seem to be incorrect for functions which do not go to zero at the endpoint.

2018-04-22 19:46:34 +01:00

..

boost

tanh_sinh: change calculation method used for scaling ranges:

2018-04-22 19:46:34 +01:00