partial minutes of WG11 meeting - LIA-2

[David G. Hough at validgh]

I've extracted the LIA-related parts.   I was planning to attend but an
agenda change occurred to late for me to accommodate... which was just as well
since I was developing a flu.

To: uunet!dkuug.dk!sc22, uunet!dkuug.dk!sc22wg11
Subject: (SC22WG11.445) WG11 N407 - Minutes of WG11 meeting, January 1995
Date: Mon, 20 Feb 95 12:37:11 N
From: Willem Wakker <uunet!komp.ace.nl!willemw>

DRAFT MINUTES OF MEETING		      SC22/WG11/N407

Committee:   ISO/IEC JTC1/SC22/WG11

Place:	     Digital Equipment Corporation
	     Maynard, U.S.A.

Date:	     January 30 - February 3, 1995

Attendees:   Mr. Ed Barkmeyer	      NIST, USA.
	     Mr. Robert Eachus	      MITRE Corporation, USA
	     Mr. Kevin Harris	      DEC, USA
	     Mr. Randy Hudson	      Intermetrics Inc, USA
	     Prof. Katsuhiko Kakehi   Waseda University, Japan
	     Mr. Kent Karlsson	      Chalmers University of Technology, Sweden
	     Mr. Brian Meek	      King's College London, UK
	     Ms. Mary Payne	      DEC, USA
	     Mr. Paul Rabin	      OSF, USA
	     Mr. Craig Schaffert      DEC, USA
	     Mr. Willem Wakker	      ACE, Netherlands (Convener)


7.  Work Item 22.28 - Language-Independent Arithmetic, Part 1: Integer and
    Floating Point Arithmetic

The IS was published by ITTF in 1994.


8.  Work Item 22.33 - Language-Independent Arithmetic, Part 2: Language
    Independent Mathematical Procedure Standard

A first working draft of LIA-2 (WG11/N404) was discussed extensively.
Topics were:

   - change the title of part 2 to "Elementary Numeric Functions."  The
     title of part 3 would presumably be changed to match.  This will be
     further discussed via email.

   - standardize the evaluation of expressions.  This is a major work item,
     not currently in the LIA-2 scope, and should thus be developed as a
     separate work item.  Maybe LIA-4?

   - Conformity: allow implementations to conform to individual functions
     or groups of functions.  The conformity clause should be base the
     conformity clause on the LIA-1 text.

   - Scope / What functions to include:

	o Include the elementary numerical functions from existing or
	  proposed standard languages.
	o Do not include intricate functions such as matrix inversion --
	  even though Basic does.
	o Large separable groups of functions (such as matrix arithmetic or
	  statistical functions) deserve to be handled as separate parts of
	  LIA.
	o Include detailed radix and type conversions.
	o Include functions to help define conversions to and from textual
	  forms.  Definitely decimal, possibly other radices.  Note that
	  these functions are intended to handle the radix conversion and
	  rounding issues, not formatting.
	o Include reasons in the rationale.
	o Exclude functions that just "move values around" rather than
	  compute.  Example: min and max.
	o Generally avoid simple compositions of existing functions.
	  Particularly mixed mode functions.  We don't need to get into
	  defining all the possible combinations.
	o Compositions that add something (like guaranteed single rounding)
	  are reasonable candidates.
	o Try to use single definitions and variants when possible.
	  Example: the trig functions in degrees in the first draft.
	o Generalize hypot to euclidean-norm?  How do we handle varying
	  numbers of arguments?

   - Required accuracy: there was a general feeling that 1/2 ulp accuracy
     on all functions would be nice, but very expensive.  We would need a
     strong argument that acceptably efficient algorithms are available for
     each function so specified.

     The opposite extreme is to permit any accuracy to conform, and require
     runtime accessible parameters which characterize the maximum error.
     This was also rejected as unnecessarily permissive.

     LIA-2 will stipulate a maximum permitted error for each function.
     These will be stipulated independently, but may fall into natural
     groups.  Runtime information will be required.

     The Ada functions package will provide some information on achievable
     accuracy.

   - Runtime parameters: a simple "max error over the whole domain"
     parameter seems feasible for all the functions that we are currently
     considering.  A more complex approach such as an "error at a given
     point" function is much harder to provide.  Since simple solutions are
     preferred, we will require a single max-error parameter for now.

   - Form of specifications: don't define mathematical functions such as
     sine.  Assume them instead.

     Specify how functions behave when given NaNs and Infinities as
     arguments.  Specify when NaNs and Infinities shall be produced as
     results.  Such specifications shall hold whenever the implemented
     floating point types have such values.

     Be compatible with the NCEG work as much as possible.

     Axioms requiring certain exact values to be returned are OK.

     Monotonicity falls out naturally from many implementations of single
     argument functions and is useful.

   - Notification: adopt the LIA-1 model and requirements with additional
     exception names.

     An "argument too large" notification is tentatively accepted for the
     basic six trig functions.	More discussion is needed.

     Most overflow and underflow boundaries appear to be exact.  If so, no
     provision for "fuzzy" notification boundaries is needed.  But this
     needs further analysis.

It is planned to have an updated version before the next meeting, and a WD
for CD registration by September 1995

Milestones for the LIA Part 2 project:

2.1   91-09   WD study initiated
      94-11   First draft circulated
2.8   95-09   WD draft for CD registration


9.  Work Item 22.34 - Language-Independent Arithmetic, Part 3: Language
    Independent Complex Arithmetic and Procedure Standard

No progress made.  The planning of this work follows the planning for LIA-
2, with a delay of one year.

Milestones for the LIA Part 3 project:

2.1   91-09   WD study initiated
      95-11   First draft circulated
2.8   96-09   WD draft for CD registration


11.  Planning and Future Meetings

May 29 - June 2, 1995	Amsterdam   Issues: LIA-2, LISS

September 1995		USA	    Date and place to be fixed.
				    Preferably immediately before or
				    after the SC22 plenary in Annapolis
				    (September 18-22).
				    Issues: ballot comments on LIPC,
				    LIA-2, LISS