Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Level 1 and 2 text (revised) for discussion and vote

Dear P1788

At last I attach a revised version of
- Chapter 1,
- Levels 1 and 2 of the set-based flavor Chapter 2.
They are almost complete, however:
- Ch 1 needs the Definitions updated, and Level 2 definitions 
  that were temporarily suppressed reinstated. 
  (I seem to recall Christian Keil had done this, maybe I have 
  mis-handled the files on the SVN. Christian?)
- I think it is now time to work seriously on the Conformance 
  clause §3. Christian?
- I see some gaps in the Accuracy Requirements §11.10 and 
  some other places in Level 2. No doubt people will see 
  further gaps.
- I have not revisited I/O and interchange types. If anyone is
  able to help me there, it will be a great help.

The main changes are as follows.

1. The only change of substance in already accepted text, IMO, 
  is that, after much discussion among members of the language 
  subgroup (Michel Hack, Vincent Lefevre, Jürgen Wolff von 
  Gudenberg & others) there was a consensus that it is not 1788's
  role to define what an "expression" is, as Level 1 previously
  tried to do. Doing so formally, raises semantic issues that 
  should be left to language definitions.

  As a result, that part of Level 1 in Ch 2 was essentially removed.
  Instead, a clause "Expressions and the functions they define" was
  added to the flavor-independent part, Ch 1. It aims to describe
  *informally* what an expression is, with enough detail to give
  context for the FTIA, which is inherently about expressions.

2. Rearranging Ch 1, including putting the Introduction into the 
  front-matter to meet IEEE's style guidelines, meant that the FTIA 
  didn't get properly introduced. To cure this I moved "Expressions 
  and the functions they define" before "Flavors" in Ch 1. I now
  introduce the FTIA, and essentially the FTDIA, as part of describing
  expressions (Clause 6).

3. The decoration system is much revised, and shorter.
- I was convinced by Guillaume Melquiond's view that the "emp"
  decoration is redundant, and removed it.
- Making "com" optional for single-flavor implementations seemed
  more trouble than it's worth, so it is mandatory in all cases.
  The current rewrite makes §10.11 "Notes on the com decoration"
  look out of place however.
- "Compressed arithmetic" is rewritten; it seemed necessary to
  reintroduce "emp" for it, or rather, to say that the empty set
  'needs to be regarded as a new "emp" decoration' within the
  compressed framework.

4. In Level 2, "Accuracy requirements" is much rewritten. I felt
  the majority view is that 1788 should make no requirements on
  accuracy of operations at Level 2. As Bill Walster maintains, 
  this should be a Quality of Implementation (QoI) matter.

  However that puts the onus on implementations to be transparent
  on what accuracy they *do* achieve, otherwise one can't compare
  one implementation with another. It is thus made a requirement
  that an implementation document a warranted accuracy for each of
  its operations, in terms of the "tightest", "accurate" and "valid"
  modes that originate from the Vienna proposal. The table of 
  accuracies now has "recommended" status, and applies to operations
  on 754-conforming types.

  The "accurate" mode was rather loosely defined by Arnold Neumaier.
  It took me quite a bit of thought to make a definition that works
  (a) for an arbitrary inf-sup type; (b) for reverse arithmetic
  operations as well as forward ones; (c) for operations where the
  input interval types need not be the same as the output (destina-
  tion) type. I think I have cracked this, but the definition should
  be regarded as "in beta", and subject to revision by the elementary
  function experts.

There are numerous other changes, but IMO they are editorial rather than matters of content.

I believe the Chair's view is that this text should be divided up into chunks for voting on as accepted text, after a discussion period and resulting revision. If anyone feels strongly that I have the wrong end of the stick on some issue, they may wish to submit their own motion on it. However time is moving on fast, especially as my recent illness put me back about a month.

As for Level 3 of the standard, I hope Jürgen and his team will be able to take editorial charge of this. Jürgen, will you confirm? There will be no Level 4 material.

Regards

John Pryce

Attachment: 20130402Level1and2textV7.1Sent.pdf
Description: Adobe PDF document