Make duals' operations through steprange use linrange instead

[tom-plaa]

Trying to solve https://github.com/JuliaDiff/ForwardDiff.jl/issues/380. Does not fix the "show" ambiguity. A bit naively done, please review and advise if there's anything that needs to be done differently.

[mcabbott]

Some other ways to make StepRangeLen with Dual, one of which makes a TwicePrecision{Dual}:

julia> (1:4) * Dual(1,2) |> typeof  # UnitRange, not tested
StepRangeLen{Dual{Nothing, Int64, 1}, Dual{Nothing, Int64, 1}, Dual{Nothing, Int64, 1}, Int64}

julia> (1:2:4) * Dual(1.0,2) |> typeof
StepRangeLen{Dual{Nothing, Float64, 1}, Dual{Nothing, Float64, 1}, Dual{Nothing, Float64, 1}, Int64}

julia> (0:0.1:1) ./ Dual(1.0,2) |> typeof
StepRangeLen{Dual{Nothing, Float64, 1}, Base.TwicePrecision{Dual{Nothing, Float64, 1}}, Base.TwicePrecision{Dual{Nothing, Float64, 1}}, Int64}

These don't seem to cause problems, so perhaps it's simplest to leave them alone.

[mcabbott]

Failure on 1.6 seems to be this, arguably a bug in Base, but can it be worked around here?

julia> using ForwardDiff: Dual

julia> range(1, step = Dual(1,1), length = 5)
ERROR: InexactError: Int(Int64, Dual{Nothing}(5,4))
Stacktrace:
 [1] Int64
   @ ~/.julia/packages/ForwardDiff/pDtsf/src/dual.jl:362 [inlined]
 [2] convert
   @ ./number.jl:7 [inlined]
 [3] _rangestyle(#unused#::Base.Ordered, #unused#::Base.ArithmeticWraps, a::Int64, step::Dual{Nothing, Int64, 1}, len::Int64)
   @ Base ./range.jl:118
...

julia> VERSION
v"1.6.0"

julia> range(1.0, step = Dual(1,1), length = 5)
ERROR: StackOverflowError:
Stacktrace:
 [1] _range(a::Float64, st::Dual{Nothing, Float64, 1}, #unused#::Nothing, len::Int64) (repeats 79984 times)
   @ Base ./range.jl:111

[tom-plaa]

On 1.8, it seems the behavior is

julia> collect(range(1,1,length=5))
5-element Vector{Float64}:
 1.0
 1.0
 1.0
 1.0
 1.0

julia> collect(range(1, Dual(1,1), length=5))
5-element Vector{Dual{Nothing, Float64, 1}}:
 Dual{Nothing}(1.0,0.0)
 Dual{Nothing}(1.0,0.25)
 Dual{Nothing}(1.0,0.5)
 Dual{Nothing}(1.0,0.75)
 Dual{Nothing}(1.0,1.0)

Which operation is causing the errors? The one on the dual component? For the Int one, we could force it to convert to Float, I guess, since it seems that range on Base also does this (at least in this anecdotal example).

[lupemba]

I just ran into this issue. Is the PR still active?

[tom-plaa]

If there's anything that needs to be changed for the PR to merged, I don't know exactly what is. Or if it should be redone. @lupemba , if in the meantime it's a blocking issue for you, you can do some type piracy in your script by including the changes that are here. I did that in a project I have and it solved my problem.

[lupemba]

@tom-plaa, Thanks for the advice. I have already solved the problem by just converting my small range to a vector.

I believe that the PR was not merge because of the failing tests on Julia 1.6. The CI tests are now so old that I cannot see the logs anymore. I therefore don't know which test that failed. The PR is probably also a bit behind master.

I would suggest.

• Pull the latest changes from master into the branch.
• Use julia up to run the tests with julia 1.6 on you own computer.
• Fix the failed test.
• Update this PR
• All checks should now be green.

I think they will merge this PR as when all the CI checks pass :)

[lupemba]

@tom-plaa I might try to make a PR that solves this myself. I would.

• Copy the tests from this PR (Hope it is fine with you)
• Try to add dual support directly for TwinPrecision .

[tom-plaa]

Feel free to reuse my code. The tests were mostly written by @mcabbott, not myself.