isNaN function

[pxqr]

Sometimes we need to test if some values are NaNs, for instance we might want to filter NaNs out from an array. The following function works with both CUDA and Interpreter backends:

isNaN :: (IsScalar a, Elt a) => Exp a -> Exp Bool
isNaN x = x /=* x 

Should it work with every IEEE 754 compliant backend or it must be defined as primitive operation?

[robeverest]

I believe that definition of isNaN is part of the IEEE 754 spec, or at least the rule that says NaN is not equal to anything, including itself, is.

Whether it should be defined as a primitive operation really depends on whether backends could provide a more efficient implementation.

[tmcdonell]

It might be a good opportunity to add a few similar functions from the RealFloat class, like isInfinite. Currently there is no way to test/represent exceptional values (ping #136).

[andygill]

Have a look at @conal’s boolean package on hackage. Data.Boolean.Numbers, written by @jbracker, gives a deep embedding overloading for the Haskell number classes. We used it in Sunroof. It may or may not work for you, but no sense in reinventing the wheel.

[tmcdonell]

There are still missing functions from RealFrac and RealFloat, but isNan exists now (thanks @mikusp!) so I'll close this ticket.