add bswap method for Float16

[JeffBezanson]

Float16 seems to be the only primitive base numeric type that doesn't support bswap. Seems to be an oversight.

[musm]

Could also just simplify all these definitions to

bswap(x::IEEEFloat) = bswap_int(x)

[Keno]

No particular objection to this change, but in general do we really want bswap on Float? I wouldn't be surprised if that didn't work on a number of systems because of NaN/subnormal canonicalization.

[stevengj]

@Keno, I don't follow you. My understanding is that a standard IEEE 754 fp value can be stored with the bytes in bigendian or littleendian order; it doesn't change the logical interpretation.

(The terminology "bigendian" and "littleendian" doesn't make as much sense when applied to a floating-point value that is a mashup of significand and exponent bits, but it seems pretty standard.)

[Keno]

@Keno, I don't follow you. My understanding is that a standard IEEE 754 fp value can be stored with the bytes in bigendian or littleendian order; it doesn't change the logical interpretation.

Yes, but once you load it into registers, you've committed to a byte order and certain architectures canonicalize values on load, so if you first load it into an fp register and then bswap it, you may not get the same as bswapping it in int registers and then loading it into an fp register.

[JeffBezanson]

Good point; maybe we should only have it for Float16, since hardware doesn't support it 😛

[Keno]

Good point; maybe we should only have it for Float16, since hardware doesn't support it 😛

GPUs do

[simonbyrne]

Doesn't that argument imply that we shouldn't allow reinterpret between floats and ints at all?

[Keno]

Reinterpret is generally fine though of course you can't expect it to round trip ok in all cases (e.g. Naan payloads, subnormals). That's not really too much of a problem in practice because you don't tend to use those values a lot. But for the bswap case, having these methods encourages getting it wrong by loading first and then bswapping, which is likely to give incorrect answers even for reasonable values

[StefanKarpinski]

This seems like a fairly silly debate. There's an obvious thing this should mean, it should do that.

[Keno]

Well, the point is it can't do the obvious thing on some architectures because fp values get normalized upon loading into fp registers. Anyway, I'm not objecting to this change in particular.