[Cfp-interest 2852] Re: CFP review of TS-4 and TS-5 revisions
Damian McGuckin
damianm at esi.com.au
Thu Aug 24 20:41:48 PDT 2023
David,
(... augmented operations)
On Thu, 24 Aug 2023, David Hough CFP wrote:
> Double-double represents numbers as a pair of doubles, the second much
> smaller in magnitude than the first. Ideally, in a pair (a.h,a.t), a.h
> would contain the correctly rounded result of computed(a.h+a.t), and a.t
> would contain the correctly rounded result of computed(a.h+a.t)-a.h.
> Performance considerations often compromise this ideal.
Should it be mandated that a.h be correctly rounded where I assume that
correctly rounded means round-to-nearest, ties to even.
Or, should the sign of a.h and a.t be mandated to match?
Obviously, given say two doubles, x and y, and (a.h,a.t) the result of
some augmented operation OP then obviously
(x OP y) correctly-rounded using the non-augmented OP
should match exactly
(a.h OP a.t) correctly-rounded using the non-augmented OP
I think to have the same sign, you need to round OP as round-towards-zero.
That probably makes it more complicated. There are advantages to having
them match but that means that a.h is then not the correctedly rounded
result (of say) the sum to the same precision as the original operands.
I tried to read Jason Riedy's and Jim Demmel's paper on this but I am sure
that it has not all permeated my brain yet.
Thanks - Damian
Pacific Engineering Systems International ..... 20D Grose St, Glebe NSW 2037
Ph:+61-2-8571-0847 .. Fx:+61-2-9692-9623 | unsolicited email not wanted here
Views & opinions here are mine and not those of any past or present employer
More information about the Cfp-interest
mailing list