setrounding/round interaction

[jmkuhn]

julia> setrounding(Float64, RoundUp)
0

julia> round(1.2)
2.0

julia> round(Int, 1.2)
2

julia> round(1.2, RoundNearest)
2.0

julia> round(Int, 1.2, RoundNearest)
2

I expect the last 2 to be 1.0/1.

[StefanKarpinski]

The rounding mode affects rounding direction in case of ties. If you want to round down, use the floor function instead of the round function.

[jmkuhn]

I'm not sure what you mean by ties in this case. We currently have

julia> setrounding(Float64, RoundUp)
0

julia> round(1.2)
2.0

julia> round(1.2, RoundDown)
1.0

julia> round(1.2, RoundNearest)
2.0

It seems inconsistent to honor RoundDown but not RoundNearest in this case.

[simonbyrne]

I know, it's annoying, but there's not much we can do.

The problem is that LLVM doesn't expose a rounding intrinsic which matches round(x, RoundNearest): the closest they have is llvm.rint which depends on the global rounding mode. Presumably at some point they will need one since there is a proposal to introduce roundeven function into C, but until that happens, this seems like the least worst option.

One thing we may want to do is not export setrounding, since it is quite unreliable (as LLVM doesn't really yet support rounding mode control), and just mark it off as experimental. Also I think the whole idea for global rounding control is bad, and that we should use something like Cassette.jl, but that's a matter for another day.

[jmkuhn]

I agree that the idea for global rounding control is bad. Since it is exported I was just tying to make it consistent, but without support from llvm I understand your choice.

[vtjnash]

The problem is that LLVM doesn't expose a rounding intrinsic

It was added recently (http://llvm.org/docs/LangRef.html#llvm-experimental-constrained-rint-intrinsic), although support is likely still considered experimental and limited

[simonbyrne]

Not quite: the run time behaviour still depends on the global rounding mode, it just controls the compile time behaviour.

[simonbyrne]

setrounding has now been deprecated: #27166