Float16 setrounding error

[AnderGray]

Does anyone know why this is happening?

using IntervalArithmetic

a = interval(Float16(1.0), Float16(3.0))
a + a

gives:

ERROR: MethodError: no method matching rounding_raw(::Type{Float16})
Closest candidates are:
  rounding_raw(::Type{BigFloat}) at mpfr.jl:121
  rounding_raw(::Type{var"#s79"} where var"#s79"<:Union{Float32, Float64}) at rounding.jl:155
Stacktrace:
 [1] setrounding(f::IntervalArithmetic.var"#11#12"{Float16, Float16}, #unused#::Type{Float16}, rounding::RoundingMode{:Down})
   @ Base.Rounding ./rounding.jl:173
 [2] +(#unused#::IntervalArithmetic.IntervalRounding{:tight}, a::Float16, b::Float16, #unused#::RoundingMode{:Down})
   @ IntervalArithmetic ~/.julia/packages/IntervalArithmetic/UR6Qe/src/intervals/rounding.jl:177
 [3] +(a::Float16, b::Float16, r::RoundingMode{:Down})
   @ IntervalArithmetic ~/.julia/packages/IntervalArithmetic/UR6Qe/src/intervals/rounding.jl:278
 [4] +(a::Interval{Float16}, b::Interval{Float16})
   @ IntervalArithmetic ~/.julia/packages/IntervalArithmetic/UR6Qe/src/intervals/arithmetic.jl:86
 [5] top-level scope
   @ REPL[5]:1

I get the same error when:

lo = Float16(1)
hi = Float16(3)

out = setrounding(Float16, RoundDown) do
	collect(range(lo, hi, length = 10))
end

giving:

ERROR: MethodError: no method matching rounding_raw(::Type{Float16})
Closest candidates are:
  rounding_raw(::Type{BigFloat}) at mpfr.jl:121
  rounding_raw(::Type{var"#s79"} where var"#s79"<:Union{Float32, Float64}) at rounding.jl:155
Stacktrace:
 [1] setrounding(f::var"#1#2", #unused#::Type{Float16}, rounding::RoundingMode{:Down})
   @ Base.Rounding ./rounding.jl:173
 [2] top-level scope
   @ REPL[8]:1

[lucaferranti]

I think that's because the setrounding function does not support 16 bits floating points. A work around could be to use the accurate rounding mode, which uses prevfloat and nextfloat internally after each operation. This would not necessarily be tight though.

julia> setrounding(Interval, :accurate)
:accurate

julia> a + a
[1.99902, 6.00391]

[lucaferranti]

On the other hand, the default rounding mode :tight should not be using setrounding at all, but the RoundingEmulator.jl package which uses error-free transformations. @dpsanders isn't it so, or am I missing something?