message from IFIP Working Group WG 2.5

[David Hough 754R work <754r@xxxxxxxxxxx>]

Date: Fri, 18 Aug 2006 09:25:03 +0200
From: Bo Einarsson <boein@xxxxxxxxx>
Subject: [Stds-754] IEEE 754 Revision, comments from IFIP WG 2.5

To the IEEE 754 Revision Committee

The IFIP Working Group WG 2.5 at its meeting July 15, 2006, in Prescott,
Arizona, discussed certain features of the proposed revision of IEEE 754.

We appreciate what the present standard has achieved in assisting the
possibilities in writing truly portable numerical software and making
it easy to transfer binary data between computers from different
manufacturers.

We have two suggestions to further improve the standard.

The presentation below is based on Draft P754 1.1.4 of August 12, 2006.

1. NaNs (8.2.1 Binary encodings of NaN)

We have noted the problem at the transfer of binary data including
NaNs between computers from different manufacturers, since the meaning
of a specific NaN is not standardized.

We therefore suggest that the binary encoding of the NaNs should be as
follows in the trailing significand:

First bit: 1 for quiet NaN, 0 for signaling NaN
Following five bits: Contains coding of the main NaNs, compare the
list in 9.2 Invalid operation. Interpreting the bits in decimal form,
0     Disallowed (to surely distinguish a NaN from infinity)
1-30  Thirty different alternatives for NaN
31    Undefined NaN

2. Rounding alternatives (6. Modes and rounding)

Of interest to the numerical community is the use of interval
arithmetic.  In order to obtain as narrow intervals as possible, it is
necessary to have directed rounding in hardware.

The standard permits switching of rounding mode at the systems level,
either in constant-mode or in dynamic mode. Doing interval arithmetic
will then require switching the rounding mode almost after every
arithmetic operation, which will take a lot of time. For further
information see the letters from Ulrich Kulish of May and August,
2005, and June 2006.

We would therefore prefer to have operations like
AddBinaryRoundingDirectionTowardPositive(a,b) and
AddBinaryRoundingDirectionTowardNegative(a,b), as well as similar
for the other elementary operations, also available.

---------------------------

For IFIP Working Group 2.5 on Numerical Software

Bo Einarsson

-- 
Bo Einarsson
Ekholmsvägen 249
SE-589 29 LINKÖPING
SWEDEN
Tel. +46 13 151896