`TracerSparsityDetector` + in-place RHS cache fails

[SouthEndMusic]

I would expect DifferentiationInterface.jl to internally make a copy of p with the right type for SCT.

MWE:

using ADTypes: AutoForwardDiff
using DifferentiationInterface: AutoSparse, Cache, prepare_jacobian
using SparseConnectivityTracer: TracerSparsityDetector
using SparseMatrixColorings: GreedyColoringAlgorithm
using ForwardDiff

function f!(du, u, p)
    p[1] = u[1]
    du[1] = -3.0 * p[1]
end

backend = AutoSparse(
    AutoForwardDiff();
    sparsity_detector=TracerSparsityDetector(),
    coloring_algorithm=GreedyColoringAlgorithm(),
)

du = zeros(1)
u = ones(1)
p = [3.0]

prep = prepare_jacobian(f!, du, backend, u, Cache(p))

Stacktrace:

ERROR: TypeError: in typeassert, expected Float64, got a value of type SparseConnectivityTracer.GradientTracer{SparseConnectivityTracer.IndexSetGradientPattern{Int64, BitSet}}
Stacktrace:
 [1] setindex!(A::Vector{…}, x::SparseConnectivityTracer.GradientTracer{…}, i::Int64)
   @ Base .\array.jl:987
 [2] f!(du::Vector{…}, u::Vector{…}, p::Vector{…})
...

[adrhill]

Interesting, this looks like a situation in which "bring your own cache" doesn't work, since it has to be of SCT's internal GradientTracer eltype. How is this handled for ForwardDiff @gdalle?

[gdalle]

For ForwardDiff, translation to Dual is handled explicitly with the DI.Cache contexts:

DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfaceForwardDiffExt/utils.jl

Lines 90 to 93 in d29659e

	function _translate(::Type{T}, ::Val{B}, c::DI.Cache) where {T,B}
	c0 = DI.unwrap(c)
	return make_dual(T, c0, ntuple(_ -> similar(c0), Val(B))) # TODO: optimize
	end

The problem here is that the ADTypes interface for sparsity detection is ADTypes.jacobian_sparsity(f, x, detector), so I have to shove all of the contexts inside a closure for f. In order to fix this, I'd probably have to add an SCT extension in DI, and SCT would need to export SCT.trace_input (the equivalent of make_dual above) so that I build upon it.

[adrhill]

SCT would need to export SCT.trace_input

I'm on it!

[adrhill]

Do you really need trace_input, which assigns input indices to tracers? Or do you just need a stable way to allocate a buffer à la similar(x, T)?

[gdalle]

The latter, I don't care about specific indices. Even the tracer eltype would be enough

[adrhill]

To cover Dual tracers, which require a primal type, the similar approach is probably easier than the eltype approach.

[adrhill]

SCT v0.6.14 now has a stable API for the allocation of buffers: https://adrianhill.de/SparseConnectivityTracer.jl/dev/user/api/#Memory-allocation

[gdalle]

And #739 is the corresponding DI patch