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

Rounded operations discussion...

BTW, I changed the thread title for the sanity of our voting tabulator.  :)



Dan Zuras wrote:

If you believe conforming systems must provide the 8 distinct rounded
operations, though not necessarily in hardware, then it seems the language 
is an improvement... though if we wish to split hairs it may still be
necessary to give specific definitions for the terms "operation codes" and 
"instruction". Usually these terms are associated with hardware
implementations.

Nate


There are many things here.

First, all 754 implementations that I am aware of
today DO provide for directed rounding instructions
in hardware.

It is just that the 8087 legacy has most of them
providing it with the inefficient 1980s method of
rounding modes.  Some more modern machines do as
Ulrich suggests but most don't.


I agree.




Thus things like distinct operation codes, single
instructions & an integral part of the arithmetic
operation are not splitting hairs.  They are a
fundamental difference in approach.  A difference
that most machine architectures are unequipped to
handle.

Finally, even if there WERE machines out there
that did the directed rounding instructions (or
some part of them) in software I would STILL think
this is a bad idea for us as a standards body to
demand of our implementers.

It is not our place to make hardware decisions for
our implementers.  They are far better trained to
do that than we are.  It is rather our place to
decide how intervals shall appear (the shall is
intentional here) to the user of intervals.  This
is where we (present company excepted) are far
better trained then they are.
I don't think you need to convince anyone of this, because no one is arguing this point (that I am aware of). 







>
> I think the intent is clear & it does not change my mind.
> All my arguments apply to this motion as well as the other.
>
> I guess I vote NO whichever is the correct motion.
>
>
> Dan
>


Don't get me wrong:  I think what Prof Kulisch has
in mind is a good idea.  The last machine I worked
on before I retired had opcodes for this.  It is
just not something that we have the right to demand
(shall) or even ask (should) of our implementers.

So I vote no.
The current motion does not demand (shall). So I take it you believe it is too much even to ask (should).
IEEE 754-2008 asks for last-bit accurate transcendental functions. No hardware does this, and even software implementations can be slow. But was it wrong to set the bar so high? 

Nate