bikeshed: use "subnormal" instead of "denormal"?

[StefanKarpinski]

"Subnormal" seems to be the preferred newer term for "denormal" and certainly is more intuitive. Should we change our terminology to be a bit more modern or is "denormal" still the standard nomenclature?

[JeffBezanson]

Is there a reference for this?
I see APIs out there using both issubnormal and isdenormal. We can change if there is some evidence subnormal is preferred.

[pao]

The SSE floating point control flag DZ uses (you guessed it) "denormal". But SSE isn't exactly a recent instruction set, either.

[bsxfan]

Can one enable flushing of denormals to zero in Julia?

[bsxfan]

Just found the answer here: #789.

ccall(:jl_zero_denormals, Bool, (Bool,), true)

[ViralBShah]

We should probably have a better interface and document it as well.

[pao]

Yeah, I never got that far. Agree on both points; should that be opened as a new issue?

[ViralBShah]

I guess this issue will get closed after the bikeshed, but we still need a better way to provide this capability, and document it - so perhaps a new issue is a good idea.

[simonbyrne]

The IEEE 754-2008 standard uses the term subnormal.

[StefanKarpinski]

That kind of clinches it for me – "subnormal" is both more intuitive and what the IEEE 754 spec uses, so it seems to be the term for this going forward. Let's be forward-looking and use "subnormal". The rest of the Internet, however – including Wikipedia – doesn't quite seem to have caught up with this new and improved nomenclature though.

[JeffBezanson]

Ok, if the new IEEE standard uses it then let's use it.

[bsxfan]

Is 0.0 subnormal? A salient feature of subnormals is that they make computation slow. Zeros do not.

Wikipedia says: "In computer science, denormal numbers or denormalized numbers (now often called subnormal numbers) fill the underflow gap around zero in floating-point arithmetic. Any non-zero number with magnitude smaller than the smallest normal number is 'subnormal'."

But later it says: "The denormal floats are a linearly-spaced set of values which span the gap between the negative and positive normal floats."

julia> isdenormal(0.0)
true 

[StefanKarpinski]

I'm pretty sure that's a bug – 0.0 and -0.0 should not be considered subnormal.

[JeffBezanson]

Zero is "non-normal" in the sense that it does not use the usual 1.m normalized representation. However, I'm pretty sure subnormals are identified with gradual underflow, making zero its own special case and not a subnormal number. A slow implementation is not part of the definition :)