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

Re: Do I have a second? (Motion to not support mid-rad



> From: "Nate Hayes" <nh@xxxxxxxxxxxxxxxxx>
> To: <rbk@xxxxxxxxxxxx>,
> 	"Dan Zuras Intervals" <intervals08@xxxxxxxxxxxxxx>
> Cc: "Arnold Neumaier" <Arnold.Neumaier@xxxxxxxxxxxx>,
> 	<stds-1788@xxxxxxxxxxxxxxxxx>
> Subject: Re: Do I have a second? (Motion to not support mid-rad
> Date: Tue, 14 Sep 2010 09:54:37 -0500
> 
> I second.
> 
> Nate Hayes

	Thank you Nate.

	I guess this puts us in the discussion phase so let
	me clarify as I see it.

> 
> 
> P.S. I agree with Baker the motion is a bit ambiguous and
> needs some clarification during the discussion phase.
> I also believe if this motion passes that the title of
> this working group needs to change.

	One step at a time.
	One step at a time. :-)

	(BTW, changing the title of either the standard or
	the working group has PAR implications.  We can do
	it but we have to get the IEEE involved when we do.)

> 
> 
> 
> 
> ----- Original Message ----- 
> From: "Ralph Baker Kearfott" <rbk@xxxxxxxxxxxx>
> To: "Dan Zuras Intervals" <intervals08@xxxxxxxxxxxxxx>
> Cc: "Arnold Neumaier" <Arnold.Neumaier@xxxxxxxxxxxx>; 
> <stds-1788@xxxxxxxxxxxxxxxxx>
> Sent: Tuesday, September 14, 2010 9:04 AM
> Subject: Do I have a second? (Motion to not support mid-rad
> 
> 
> > P-1788:
> >
> > Arnold Neumaier and Dan Zuras have jointly put forward the
> > following motion:
> >
> > The standard shall not support a midrad interval format or
> > >     nonstandard intervals, beyond providing conversion support,
> > >     approximately to the extent specified in the Vienna Proposal.
> > >
> >
> > Do I have a second?
> >
> > Baker
> >
> > P.S. I'm wondering precisely what "support" means here.  Perhaps,
> >      if the motion is seconded, that could be clarified during the
> >      discussion.  (I'm assuming it means explicitly specifying the
> >      behavior of mid/rad operations.)
> >
> > P.P.S. The voting period for Motion 19 is continuing.
> >
> > -- 
> >
> > ---------------------------------------------------------------
> > R. Baker Kearfott,    rbk@xxxxxxxxxxxxx   (337) 482-5346 (fax)

	I believe you are both correct.  The notion of 'support'
	here is ambiguous.

	The motion itself is unusual but not without precedent.
	The motion states that a conforming implementation
	shall NOT do something in order to conform.  That is,
	an otherwise conforming implementation could lose
	that status if it chooses to do something extra.

	<begin old war story>

	An analogy from the floating-point world comes from
	Cray.

	Cray existed prior to 754.  Like everyone else at the
	time, they rolled their own when it came to floating-
	point.  Needless to say it had some interesting
	properties.  Let me concentrate on just a few.

	Their equivalent to our double was a 64-bit datatype
	they called single.  (They came from Minnesota but
	they thought like Texans. :-)  It had a 56-bit
	mantissa & a 7 bit exponent (with one bit for sign).

	But when they multiplied two such numbers together
	they only computed 48 bits of the result on the
	grounds that Seymour wanted to get rid of the low
	24x24 circuit & just use the remaining 3.

	As a result of this & booth encoding one side, their
	multiplies were not only inaccurate, they had the
	property that A*B != B*A.

	Then, adds were inexact when they might be exact
	due to a lack of a sticky bit.

	Finally, their divider (which used a Newton circuit
	to approximate x/y by x*(1/y)) suffered from the
	problems it inherited from the multiplier.  As a
	result you found that 18.0/3.0 != 6.0.  (When most
	of the Pentium designers were still in kindergarten.)

	This is all preamble because some time after 754 hit
	the streets Cray went to highly parallel commercial
	microprocessor based machines which used 754 based
	floating-point.  (Basically boatloads of overdriven
	Intel & AMD chips.)

	But, in order to support dusty decks they provided
	an 'old Cray' floating-point datatype that allowed
	people to do it the old crappy way if they had
	coded for it.

	Those Crays WITHOUT support for old floating-point
	could be considered conforming.  WITH that support,
	they became non-conforming for all the obvious
	reasons.

	<end old war story>

	So there is precedent for this.  It is even a good
	analogy.  For if Arnold is correct & doing things
	in the mid-rad fashion violates even containment
	then providing such a capability invalidates all
	else that the standard is there to provide.

	Therefore, I would say that 'support' in this context
	means providing the means to do ANYTHING with a
	mid-rad value other than to convert to & from it.

	I suppose this puts the library writer into a grey
	area.  In that if they want to provide a canned
	package that works internally in mid-rad but that
	they have gone to a lot of trouble to make sure is
	correct & containing they must do so in a manner
	that never allows anyone to look into the can.

	So arithmetic is out.  Conversions are in.  And
	a canned routine could be acceptable under a sort
	of don't ask don't tell policy.

	That's the way I figure it.

	Of course, Arnold can give his own opinion on the
	matter.

	Does that clarify things?

	This decision is a stark one.

	I think that is as it should be.


				Dan