Re: Motion 4
Folks,
I second this motion again. (I seconded among a smaller
group but I want to make sure it is seconded formally.)
On the non-754 issue. I believe John's interpretation is
correct. A machine that is based on another arithmetic
system altogether is unlikely to meet many of the other
normative statements we will likely make in the course
of the development of this standard.
However in our other discussions the case of the Cell
came up. The format the Cell uses is 754 single & double.
The arithmetic is odd for single but correct for double.
And there is some lack of flag support which is OK because
flags are optional under 754 anyway. The oddity for
single arithmetic is that the Cell designers felt themselves
justified to cut corners by truncating the arithmetic on the
grounds that it was intended to be used for video game
graphics.
Unfortunately for this decision, since the Cell is so cheap
& so fast it is being used for things much more important
than killing the ogre & saving the princess. For example
it is used in the PetaFlops machine at Los Alamos for full
3-D simulations of the next generation of nuclear weapons.
I believe that the Cell could be made to be compliant to
1788 with some effort. And it would be nice if the Los
Alamos people could be ASSURED that their designs will
work.
At least, *I* would feel better about it. :-)
Dan
> To: stds-1788@xxxxxxxxxxxxxxxxx
> From: John Pryce <j.d.pryce@xxxxxxxxxxxx>
> Subject: Motion 4
> Date: Thu, 9 Apr 2009 08:28:19 +0100
>
> <<Sorry, forgot to adjust the title of the motion!. Try this.>>
>
> Dear members of the working group.
> I propose the following motion.
> John Pryce
>
>
> ===Motion P1788/M0004.01_Keep_to_754===
> Proposer: John Pryce
> Seconder: Dan Zuras
>
> ===Motion text===
> P1788 says nothing about interval arithmetic on computer
> architectures that
> are not 754-compliant.
>
> ===Rationale===
> This motion is in support of KISS: "Keep It Simple, Stupid". It
> delimits 1788's responsibilities. It limits and simplifies our work
> in that our design decisions may freely assume the existence of NaN
> (and possible payload), the sign-exponent-significand layout, etc.
>
> The question has been asked: If 1788 says nothing about non-754
> architectures, does that mean that any interval implementation on
> such an architecture is automatically 1788-compliant?
>
> This IMO is not a question about the motion but about what standard-
> compliance means. My understanding is that
> an implementation is compliant with a standard iff it obeys every
> normative clause in that standard.
>
> Therefore it is *conceivable* that an implementation on a non-754
> machine can be made 1788-compliant -- in the unlikely case that 1788
> actually makes no 754-specific decisions!
>
> In practice, presumably, that will not be so. Then, someone may
> consider that 1788 should define a "lower tier" of compliance for
> intervals on non-754 machines. They may wish to move a suitable
> amendment to this motion.