Why is seeding ydual necessary?

[gdalle]

There are some lines in the code that I don't understand, essentially all the ones that show up here:
https://github.com/search?q=repo%3AJuliaDiff%2FForwardDiff.jl%20seed!(y&type=code

Why would we need to set the dual numbers inside ydual to a specific value, before overwriting them with f!(ydual, xdual)? Can we just get rid of these steps?

This is causing issues when y has non-bitstype elements, because they may not be initialized the first time that jacobian is called.

Related:

• Safely initialize the inplace Jacobian cache JuliaNLSolvers/NLSolversBase.jl#160
• Add a precompilation workload JuliaNLSolvers/LsqFit.jl#263

[gdalle]

MWE of the error:

julia> using ForwardDiff

julia> x = BigFloat[1, 2]
2-element Vector{BigFloat}:
 1.0
 2.0

julia> y = similar(x)
2-element Vector{BigFloat}:
 #undef
 #undef

julia> ForwardDiff.jacobian(copyto!, y, x)
ERROR: UndefRefError: access to undefined reference
Stacktrace:
  [1] getindex
    @ ./essentials.jl:917 [inlined]
  [2] _broadcast_getindex
    @ ./broadcast.jl:644 [inlined]
  [3] _getindex
    @ ./broadcast.jl:674 [inlined]
  [4] _broadcast_getindex
    @ ./broadcast.jl:650 [inlined]
  [5] getindex
    @ ./broadcast.jl:610 [inlined]
  [6] macro expansion
    @ ./broadcast.jl:973 [inlined]
  [7] macro expansion
    @ ./simdloop.jl:77 [inlined]
  [8] copyto!
    @ ./broadcast.jl:972 [inlined]
  [9] copyto!
    @ ./broadcast.jl:925 [inlined]
 [10] materialize!
    @ ./broadcast.jl:883 [inlined]
 [11] materialize!
    @ ./broadcast.jl:880 [inlined]
 [12] seed!(duals::Vector{ForwardDiff.Dual{…}}, x::Vector{BigFloat}, seed::ForwardDiff.Partials{2, BigFloat})
    @ ForwardDiff ~/.julia/packages/ForwardDiff/UBbGT/src/apiutils.jl:45
 [13] seed!
    @ ~/.julia/packages/ForwardDiff/UBbGT/src/apiutils.jl:45 [inlined]
 [14] vector_mode_dual_eval!
    @ ~/.julia/packages/ForwardDiff/UBbGT/src/apiutils.jl:30 [inlined]
 [15] vector_mode_jacobian(f!::typeof(copyto!), y::Vector{…}, x::Vector{…}, cfg::ForwardDiff.JacobianConfig{…})
    @ ForwardDiff ~/.julia/packages/ForwardDiff/UBbGT/src/jacobian.jl:139
 [16] jacobian
    @ ~/.julia/packages/ForwardDiff/UBbGT/src/jacobian.jl:40 [inlined]
 [17] jacobian(f!::typeof(copyto!), y::Vector{…}, x::Vector{…}, cfg::ForwardDiff.JacobianConfig{…})
    @ ForwardDiff ~/.julia/packages/ForwardDiff/UBbGT/src/jacobian.jl:37
 [18] jacobian(f!::typeof(copyto!), y::Vector{BigFloat}, x::Vector{BigFloat})
    @ ForwardDiff ~/.julia/packages/ForwardDiff/UBbGT/src/jacobian.jl:37
 [19] top-level scope
    @ REPL[5]:1
Some type information was truncated. Use `show(err)` to see complete types.

[devmotion]

Maybe to zero out partials from possible previous invocations (yduals is taken from the gradient config)?

[gdalle]

Oh right, because if f! doesn't touch some parts of y we want the partials to be zero, and the primals to match whatever was given.

Any idea how to stop it from erroring in the MWE above?

[devmotion]

Closing as a duplicate of #436.

[devmotion]

#743 seems to fix it.