RE: [boost] Re: Math Constants and intervals and rounding

From: Paul A. Bristow (pbristow_at_[hidden])
Date: 2001-03-27 11:13:48

I have explored the possibility of using the 'better' extended precision
offered by the IEEE standard and the x87 chips.

One can indeed control this using MS MACROs _controlfp()
(or on other system sby fiddling with the floating point control word
directly.
I have not done this but the NTL code shows how for gcc/Linux).
(The actual MS code is hidden, but not too hard to replicate if necessary).

This shows:

1 one can change the precision, but the extended values are hidden from view
and are rounded on output (and I can't see any use being made of them).

2  One cannot change the value stored, for example pi, by rounding,
so I feel one cannot get the pi plus least significant bit
and pi minus least significant bit for an interval.

So I conclude that switching to standard 53-bit mantissa precision (a la
NTL)
is the best for better portability.

And because this is simplest, I propose to use the compiler to fill all the
constants.

Paul