> Barton> I don't think CL supports setting the floating point
> Barton> rounding mode (maybe CMUCL does as an extension); at
> Definitely true. Plus getting rounding right on x86 is pretty
> difficult with its 80-bit registers.
Yes, it would be nice to use the IEEE floating-point rounding modes, but
unless/until we get full IEEE floating-point support out of CL (it would
be nice to have extended precision and NaNs, too), there are
workarounds:
1) Do the round-up/down "by hand" by explicit calculation.
2) Only do the approximate thing for machine floats, but implement full
rounding modes for bfloats.
3) Just use bfloats that are long enough that you don't need to bother
with rounding modes. This is trivial to implement, though slow....
-s