On 7/30/07, Andre Pouliot <[EMAIL PROTECTED]> wrote:
> In that case I'm going to modify the program to make the result go to
> infinity when we receive an infinite/NAN or force the output at zero when we
> receive an denormalized/zero value. So no more denormalized value.

Yeah.  Now, the way you did it before wasn't bad.  Basically, if the
exponent isn't min, then make 1.mantissa the intermediate operand,
while if the exponent is min, make it 0.mantissa.  That's the simplest
way to deal with numbers on the way in.  But then if you find that
neither of the top two bits of the product is 1, then slam the output
all to zero.  Cheat as much as you possibly can!

> All case
> that make the pipeline behave strangely will be covered. The case infinite
> multiplied by zero will result in zero unless someone have an objection.

No objection.  In fact, minimize what you do with infinity.  Nowhere
do we actually perform a divide.  The one place involves a reciprocal
and a multiply, and we can cheat there too.  Keep in mind that for
colors, anything greater than 1.0 gets clamped.  For coordinates,
larger values are accepted, but they're still much smaller than
infinity.

Since OGA1 will be implemented first for FPGA, we can actually take a
"try it an see" approach.  Obviously, we want people taking the
analytical approach to determine if there are any places where we
would care about an infinity.  But we can also experiment with lots of
apps and see if anything actually works with really large numbers.
Maybe we can not bother to clamp the exponent (let it wrap), maybe we
have to clamp it in some places.  The question we want to consider is
what to do with ill-behaved apps that hand us numbers that are beyond
our ability to handle.

-- 
Timothy Normand Miller
http://www.cse.ohio-state.edu/~millerti
Open Graphics Project
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to