New fpprec default value (Was Re: Precision of float and bfloat)
Subject: New fpprec default value (Was Re: Precision of float and bfloat)
From: John Ogilvie
Date: Tue, 27 Oct 2009 18:12:50 -0700 (PDT)
On Mon, 26 Oct 2009, Stavros Macrakis wrote:
> Unlike machine float implementations which pack mantissa and exponent into
> an integral number of machine words, the current bfloat implementation uses
> Lisp bignum integers for the mantissa and stores the exponent separately, so
> by the "optimal use of the processor" argument the "natural" precisions are
> 18 decimal digits (64 bits), 27 decimal digits (96 bits), 37 decimal digits
> (128 bits), etc. (Though calculation actually uses numbers up to twice that
> big.) That said, bignum arithmetic is making far from optimal use of the
> processor!
>
> But I think we should start with mathematical requirements, not with machine
> requirements. Is there a principled reason, for example, to want twice as
> many digits as the base precision? Is there some important or common class
> of calculation where that is the natural next step?
>
> -s
Is there a principled reason? No, not one that comes to mind. Is
there some important or common class .. ? No, not particularly, although
in Maple I tend to adopt 32 decimal digits for calculations that might
benefit from more than 16 digits.
If 96 bits equivalent to 27 decimal digits is natural, why not adopt
27 as default instead of 24? I have no objections either way.