Re: new motion
Ulrich,
The motion suggests that CA is recommended but the attached document
starts with the phrase
"The exact dot product is computed by use of complete arithmetic and
a complete format associated with a floating-point format.".
I think that this motion would have more chances if the document is reworked
so that it says about abstract number format without describing its CA implementation.
CA implementation could be in appendix as an example of implementation.
The name of abstract data type could be "fixed-point number format".
It is parameterized by two parameters
h - positive real;
K - positive integer.
"Fixed-point number format" FP(h,K) is comprised of such values:
1) k * h, where k - integer, -K <= k <= +K ;
2) negative and positive infinity;
3) NaN.
So it satisfies requirements of P1788 number format.
A list of operations on FP is small:
1) Constructor zero: -> FP .
It returns a value of FP with the value 0.
2) Addition/subtraction: FP x FP -> FP
3) Addition/subtraction with other number format: FP x F -> FP
4) Fuse Multiply-Add with other number format: FP x F x F -> FP
5) Rounding to other number format: FP -> F .
In P1788 the synposis of operations doesn't say either they are inplace or not.
For example, an implementation may supply any variant of P1788 fma operation: F x F x F -> F
either
r := x + y * z
or
x := x + y * z.
So synopsis of "fixed-poit number format" doesn't say about this too,
though usual implementation is inplace of course.
If the document is reworked in such way I will not mind against requirement that
754-conforming implementation shall supply the fixed point number format
FP(h,K) with h = sqr(minPositiveValue_b64) and with K such that
h * K / sqr(maxValue_b64) is sufficiently large (say 2^88).
Such "fixed point number format" can be easily implemented in software,
so I don't expect that it complicates an implementation of P1788 much.
I don't think that any implementation of "fixed-point number format" should be
recommended in P1788. However, CA can be described in appendix as a simple
software implementation of "fixed-point number format".
We leave to manufactures a decision either to implement CA in hardware or not.
Best Regards
-Dima
----- Исходное сообщение -----
От: ulrich.kulisch@xxxxxxx
Кому: stds-1788@xxxxxxxx
Отправленные: Четверг, 5 Сентябрь 2013 г 13:30:40 GMT +04:00 Абу-Даби, Маскат
Тема: new motion
On Sept. 4 we got the folloqwing mail:
| Voting on Motion 47 to amend Motion 45 ended with 3 Sept., 2013.
| Tally: Yes - 6; No - 30; Needed for quorum - 32.
| The motion FAILS.
I am a bit surprised about this result. On Sept. 2 we got the following mail:
| Voting on Motion 47 - Motion to amend Motion 45, closes TOMORROW, Tuesday, 3 September.
| Current tally: Yes - 7; No - 18; Needed for quorum - 32
After that we got 4 No votes and one withdrawal giving a total of 28 votes.
Anyhow without the YES votes the motion did not have reach the quorum.
The comments coming with the No votes on motion 47 (see, for instance, Svetoslav Markov's mail of Aug. 27 which is copied below) suggest that another formulation of the motion would lead to another result. So let me try another motion:
IEEE P1788 requires an EDP. Providing it via CA is recommended. (For CA see the attachment).
With best regards
Ulrich
ON Aug. 27 Swetoslav Markov wrote:
My vote on Motion 47 is NO.
I have changed my mind on Motion 45 on the base of the
last discussions. In particular,
I agree with Michel Hack :
"I would vote YES if P1788 included a complete standard for
Complete Arithmetic, ..."
and I agree with Ian McIntosh :
" I support having EDP but not mandatorily implemented via CA ...."
IMO the CA as formulated by Motion 45 is too much hardware
oriented. Exact Dot Product should be required in P1788 but
not mandatory based on CA. I think that dot product is as fundamental
operation as are addition and multiplication.
S Markov
--
Karlsruher Institut für Technologie (KIT)
Institut für Angewandte und Numerische Mathematik
D-76128 Karlsruhe, Germany
Prof. Ulrich Kulisch
Telefon: +49 721 608-42680
Fax: +49 721 608-46679
E-Mail:ulrich.kulisch@xxxxxxx
www.kit.edu
www.math.kit.edu/ianm2/~kulisch/
KIT - Universität des Landes Baden-Württemberg
und nationales Großforschungszentrum in der
Helmholtz-Gesellschaft