[Cfp-interest] WG14 IEEE 754-C binding meeting minutes 2017/03/21
Rajan Bhakta
rbhakta at us.ibm.com
Tue Mar 21 10:43:38 PDT 2017
Attendees: Rajan, Jim, Fred, Ian, David H., Mike, David C.
New agenda items:
None.
Last meeting action items:
Rajan: Submit DRS2:DDR9 example as a paper for a supplement to the STC
in the DR and change all the %A to %a (ensuring it conforms to the
specification for %a). - Done (N2126)
Jim: Create a DR against part 2 to specify the capitalization effects
of %a. - Done (N2125)
Jim: DRs2:DDR2: DECIMAL_DIG: Have DECIMAL_DIG to be the largest
supported floating type as a new replacement Suggested TC for DRs2:DDR2. -
Not done due to further discussion. Remove from list.
Jim: Add in the encoding *_DIG macros (as per
http://wiki.edg.com/pub/CFP/WebHome/characteristics_for_non-arith_formats.pdf
) and obsolescing DECIMAL_DIG as a proposal to WG14 as a new enhancement.
- Not done due to further discussion. Remove from list.
Jim: Include Type-generic macros for functions that round result to
narrower type as part of the third set of DR's. - Done (N2125).
Fred: Propose adding in #pragma STDC FENV_ROUND DEFAULT as a means to
set the static rounding mode to the default rounding mode. - Done (N2128)
via a different method.
New action items:
Rajan: Bring up "Understanding" points 4 and 5 when discussing DR501
(from Jim's email on 2017/03/17).
David H: Ask about the IEEE dependencies (rounding mode, infinities)
for the augmented functions and how they could be used for non-IEEE
formats.
Jim: Ask David Keaton what we should be doing to close off the group
or extend it (IEEE-754:2018 binding for example).
Next Meeting:
Tuesday April 25th, 2017, 12:00 EDT, 9:00 PST
Same teleconference number.
Discussion:
IEEE 754 revision:
Near the end.
Couple of major items without text: Max/min, augmented.
What to do the old min/max? Remove? Deprecate? Currently
deprecated is not a concept. Instead we say you can do x and y, but you
should do y.
Currently everything added is a recommendation. Next revision may
consider making changes that are required and may cause changes.
C++ liaison:
No update.
Proposals for the C standard (C2X):
Links in email sent by Jim today.
DRs:
N2127 (http://wiki.edg.com/pub/CFP/WebHome/n2127.pdf):
Proposed change to Part 3 for DR 501.
This paper changes part 2 in a simpler way which means part 3 does
not need to change anymore.
N2126 (
http://wiki.edg.com/pub/CFP/WebHome/N2126__Example_of_the_effect_of_the_change_for_CFP_DR11.htm
):
No capital A to avoid the %A specification issue (which is handled
via another paper).
Discuss DECIMAL_DIG issues (Jim's 3/17/17 email):
"Understanding":
Point 2:
Ian: Regarding extension floating-point types vs extended
integer types: We should see about changing the C standard to support
this.
Jim: This is essentially what we're doing in part 3.
Point 3:
Common extensions are not necessarily conforming.
Point 4:
Since it doesn't say long double, this follows.
"Alternatives":
Point 2: Seems to be more acceptable to be a C2X proposal since
WG14 did not seem to like making changes to C due to a TS.
Point 4: Will have to be a proposal for a new version of part 3
(or part of C if the Part 3 proposal for C2X gets in).
Alternative 2 seems to be the consensus.
Can bring up "Understanding" points 4 and 5 when discussing
DR501.
Binding for IEEE 754-2018 (Jim's email on 2017/03/20):
Augmented operations: Return two results, head and tail.
WG14 said a struct is better to return two results.
Non-IEEE format needs to be looked at. Rounding mode, infinities
for example do not always exist.
David H: Perhaps say the two parts need to add up to the correct
result to leave slack in terms of rounding.
Jim: There are algorithms for reproducible summations (the purpose
of these functions). Those algorithms likely need this rounding property
to work correctly. It does not need to be a rounding mode, just the result
has to follow that property.
David H: The property was an issue of efficiency. May not be
possible for non-IEEE due to implicit dependencies.
Ian: For AIX double-double (vs Linux double-double), having the
high part being close to infinity may have issues as rounding away from
zero may result in an infinity.
Ian: No rounding mode that works in general for IEEE.
David H: It does for Decimal.
It was not intended that they add another mode, just that the
semantics have this.
Ian: Most implementations do not have the rounding mode needed.
David H: Expected that we can implement this in hardware.
Ian: If we are expecting hardware to add this rounding for these
operations, would it work for a general rounding mode?
David H: No, since most hardware only has 2 bits for rounding
mode and adding this would result in another bit.
Non-IEEE implementations: double-double to get quad-quad,
short/half (hard to imagine), IBM Hex format (only has round to zero),
packed decimal?
Jim: Annex F allows double-double for 'long double'.
*David H: Ask about the IEEE dependencies (rounding mode,
infinities) for the augmented functions and how they could be used for
non-IEEE formats.
Min/Max:
NaN preference or number preference.
Can do fminISO, fminIEC, fminSTD or something like that.
Mike: fminimum can be used. No reason to abbreviate.
Fred: In Annex F we don't say anything about zeros or SNaNs.
Jim: In the TS we do say SNaNs always signal. The signed zero
will likely break a lot of implementations of the existing fmin (which is
more like minimumNumber).
Rajan: There are long names already for the atomic functions. Can
just do fminimum_magnitude_number.
We can discuss this next meeting after thinking about it.
Other:
Marius (pre-meeting question): When will things be finalized?
Fred: After C2X?
Jim: We need to stay with it until it is in the standard. This
will take a while.
If parts go into the standard but not all of them, we need to
revise the TS parts that didn't go in (Ex. rebase on C2X).
Rajan: Current WG14 mandate is for IEEE-754:2008 binding. The 2018
binding may be something we add on or a new group created.
*Jim: Ask David Keaton what we should be doing to close off the
group or extend it (IEEE-754:2018 binding for example).
Jim: WG14 will likely be happy to have us help integrate any parts
into C2X.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20170321/6dc521fe/attachment.html
More information about the Cfp-interest
mailing list