Re: Motion P1788/M0050:EDP-Without-CA: voting period begins: YES and example
I vote YES for a number format FP(F) that can store result of EDP
of vectors with components of 754 format F.
Its value set consists of k*sqr(minPositiveValue_F).
That looks like fixed-point format, but any implementation is possible.
CA is a simple and nice implemetation, but I don't think that
standard should recommend any implementation.
It can describe CA in an annex without "recommended" status.
I want to describe an example that shows why such number format FP(F)
may be more convinient that correctly rounded dot product.
Suppose that I have two F-vectors a_i and b_i of the same length n.
I want to enclose their exact dot product by a form "m + [l,u]"
where m, l, u are of F format.
Yes, I can do it with correctly rounded dot-product.
1) m = rnd-near( Sum: 1 <= i <= n: a_i*b_i )
2) Build vectors a' and b' of length n+1 from vectors a and b
by appending values:
a_{n+1} = m b_{n+1} = -1
3) Compute l = rnd-to-neg-inf( Sum: 1 <= i <= n+1: a'_i*b'_i )
4) Compute u = rnd-to-pos-inf( Sum: 1 <= i <= n+1: a'_i*b'_i )
But see how it looks with a fixed-point number format FP(F).
var exactSum: FP(F) = ( Sum: 1 <= i <= n: a_i*b_i )
m = rnd-near( exactSum );
exactSum = exactSum - m;
l = rnd-to-neg-inf( exactSum );
u = rnd-to-pos-inf( exactSum );
We don't need to compute dot-product three times with fixed-point format FP(F) available.
-Dima
----- Исходное сообщение -----
От: rbk5287@xxxxxxxxxxxxx
Кому: stds-1788@xxxxxxxx
Отправленные: Суббота, 14 Сентябрь 2013 г 15:37:24 GMT +04:00 Абу-Даби, Маскат
Тема: Motion P1788/M0050:EDP-Without-CA: voting period begins
P-1788,
Unless there are objections, the voting period for this motion
herewith begins, and will continue until after Saturday, October 4.
This motion will proceed according to the rules for position papers.
The motion is as follows:
=======================================================================
IEEE P1788 requires an EDP. Providing it via CA is recommended.
=======================================================================
(Recall that "EDP" means "exact dot product" and "CA" means
"complete arithmetic".)
Juergen: Please update the page of motions with this information.
Best regards,
Baker (acting chair)
--
---------------------------------------------------------------
R. Baker Kearfott, rbk@xxxxxxxxxxxxx (337) 482-5346 (fax)
(337) 482-5270 (work) (337) 993-1827 (home)
URL: http://interval.louisiana.edu/kearfott.html
Department of Mathematics, University of Louisiana at Lafayette
(Room 217 Maxim D. Doucet Hall, 1403 Johnston Street)
Box 4-1010, Lafayette, LA 70504-1010, USA
---------------------------------------------------------------