On Wednesday, 17 February 2016 at 08:11:21 UTC, Nick B wrote:
Hi
John Gustafson was in town (Wellington, NZ) for the Multicore
World Conference 2016 ( http://www.multicoreworld.com/)
conference. I caught up with him, tonight, and spoke to him for
about two hours. Here is a quick summary of what we discussed.
John has just redesigned Unums, to address the design issues in
version 1.0. He presented his Powerpoint presentation to the
conference, with the details of Unums 2.0 (this is a tentative
name at present). Its a improved design, but I will only brief
detail it: "It will have more dynamic range with 16-bit
values than IEEE half-precision, but only by a small amount.
Still remarkable to be uniformly better in dynamic range and
precision, with support for inexact values and perfect
reciprocation. If a language supports just one unum data type,
John believes it should be the 16-bit one". John has agreed to
provide a link to the Powerpoint presentation, in a couple of
weeks, and then later, a link to his new published paper on the
subject, when it is ready. There will likely be a new book,
building on version 1.0, and, again, tentatively titled 'Unums
2.0'. I also discussed with him, about integrating it with D.
At the present, there is a 'C' codebase under construction, but
this could be rewritten in D in the future. D may require some
language changes, and a new phobos library, to support this
advanced functionality. Of course Walter will have decide if he
wants this advanced numbering system as a part of D.
I would be interested in the Powerpoint when it becomes available.
As for getting it to work in D, I'm not sure how much language
changes would be necessary. If they can get it to work in C, then
surely it would work in D. After all, the book has an
implementation in python (albeit this is not the latest and great
version apparently).
Wrt phobos, I would just recommend that whatever unum library
gets eventually written has a companion with the equivalent of
the functions from std.math.