On Sat, Jan 26, 2008 at 01:16:39PM -0500, S. Newhouse wrote:
> On the need for bfloat, the issue is speed.
As a user (not that I consider this kind of interval arithmetic
useful, but that is another issue), I would expect to get exact
results for exact inputs, to get floats for float inputs, and to get
bfloats for bfloat inputs, with all intervening computations done with
the same fpprec.
So in my view it is a mistake to allow the user to enter exact
rationals when you actually want bfloats. If you really think that
automatic conversion might be nice for some particular application,
you can still define some frontend function that performs the
conversion:
frontend_function(a, ..) := the_real_function(bfloat(a), ...)
where the_real_function does all the work but knows that it can rely
on consistent numeric types.
Regards,
Albert.