PolyesterForwardDiff automatic chunksize selection?

[avik-pal]

Currently PolyesterForwardDiff forwards a chunksize of nothing. I think it would make sense to reuse the automatic chuunksize selection of ForwardDiff here as well

[gdalle]

I think this already happens here?

DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfacePolyesterForwardDiffExt/DifferentiationInterfacePolyesterForwardDiffExt.jl

Lines 25 to 33 in 88c48c1

	function single_threaded(backend::AutoPolyesterForwardDiff{C,T}) where {C,T}
	return AutoForwardDiff{C,T}(backend.tag)
	end
	DI.check_available(::AutoPolyesterForwardDiff) = true
	function DI.pick_batchsize(backend::AutoPolyesterForwardDiff, dimension::Integer)
	return DI.pick_batchsize(single_threaded(backend), dimension)
	end

[avik-pal]

Interesting. I got the error:

[OOP] u0: SVector{2, Float64}: Error During Test at REPL[87]:7
  Got exception outside of a @test
  MethodError: no method matching rem(::Int64, ::Type{Union{Nothing, Int64}})
  
  Closest candidates are:
    rem(::Any, ::Any, ::RoundingMode{:FromZero})
     @ Base div.jl:103
    rem(::Any, ::Any, ::RoundingMode{:Nearest})
     @ Base div.jl:100
    rem(::Any, ::Any, ::RoundingMode{:Up})
     @ Base div.jl:99
    ...
  
  Stacktrace:
    [1] cld_fast(a::Int64, b::Nothing)
      @ PolyesterForwardDiff /mnt/.julia/packages/PolyesterForwardDiff/oqvIt/src/PolyesterForwardDiff.jl:10
    [2] threaded_jacobian!(f::DifferentiationInterface.FixTail{NonlinearFunction{false, SciMLBase.FullSpecialize, typeof(quadratic_f), UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED_NO_TIME), Nothing, Nothing, Nothing}, Tuple{Float64}}, Δx::Matrix{Float64}, x::SVector{2, Float64}, ::ForwardDiff.Chunk{nothing}, check::Val{false})
      @ PolyesterForwardDiff /mnt/.julia/packages/PolyesterForwardDiff/oqvIt/src/PolyesterForwardDiff.jl:113
    [3] threaded_jacobian!(f::DifferentiationInterface.FixTail{NonlinearFunction{false, SciMLBase.FullSpecialize, typeof(quadratic_f), UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED_NO_TIME), Nothing, Nothing, Nothing}, Tuple{Float64}}, Δx::Matrix{Float64}, x::SVector{2, Float64}, ::ForwardDiff.Chunk{nothing})
      @ PolyesterForwardDiff /mnt/.julia/packages/PolyesterForwardDiff/oqvIt/src/PolyesterForwardDiff.jl:112
    [4] jacobian!(f::Function, jac::Matrix{Float64}, ::DifferentiationInterface.NoJacobianPrep, ::AutoPolyesterForwardDiff{nothing, Nothing}, x::SVector{2, Float64}, contexts::DifferentiationInterface.Constant{Float64})
      @ DifferentiationInterfacePolyesterForwardDiffExt /mnt/.julia/packages/DifferentiationInterface/gPcSq/ext/DifferentiationInterfacePolyesterForwardDiffExt/onearg.jl:205
    [5] jacobian(f::Function, prep::DifferentiationInterface.NoJacobianPrep, backend::AutoPolyesterForwardDiff{nothing, Nothing}, x::SVector{2, Float64}, contexts::DifferentiationInterface.Constant{Float64})
      @ DifferentiationInterfacePolyesterForwardDiffExt /mnt/.julia/packages/DifferentiationInterface/gPcSq/ext/DifferentiationInterfacePolyesterForwardDiffExt/onearg.jl:229
    [6] jacobian(f::NonlinearFunction{false, SciMLBase.FullSpecialize, typeof(quadratic_f), UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED_NO_TIME), Nothing, Nothing, Nothing}, backend::AutoPolyesterForwardDiff{nothing, Nothing}, x::SVector{2, Float64}, contexts::DifferentiationInterface.Constant{Float64})
      @ DifferentiationInterface /mnt/.julia/packages/DifferentiationInterface/gPcSq/src/fallbacks/no_prep.jl:49
    [7] (::SimpleNonlinearSolve.Utils.var"#4#6"{ImmutableNonlinearProblem{SVector{2, Float64}, false, Float64, NonlinearFunction{false, SciMLBase.FullSpecialize, typeof(quadratic_f), UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED_NO_TIME), Nothing, Nothing, Nothing}, @Kwargs{}, SciMLBase.StandardNonlinearProblem}, AutoPolyesterForwardDiff{nothing, Nothing}})(u::SVector{2, Float64}, p::Float64)
      @ SimpleNonlinearSolve.Utils /mnt/software/nonlinearsolve/NonlinearSolve.jl/lib/SimpleNonlinearSolve/src/utils.jl:228
    [8] value_and_jacobian(f::Function, prep::DifferentiationInterface.NoJacobianPrep, backend::AutoPolyesterForwardDiff{nothing, Nothing}, x::SVector{2, Float64}, contexts::DifferentiationInterface.Constant{Float64})
      @ DifferentiationInterfacePolyesterForwardDiffExt /mnt/.julia/packages/DifferentiationInterface/gPcSq/ext/DifferentiationInterfacePolyesterForwardDiffExt/onearg.jl:215
    [9] value_and_jacobian(f::SimpleNonlinearSolve.Utils.var"#4#6"{ImmutableNonlinearProblem{SVector{2, Float64}, false, Float64, NonlinearFunction{false, SciMLBase.FullSpecialize, typeof(quadratic_f), UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, typeof(SciMLBase.DEFAULT_OBSERVED_NO_TIME), Nothing, Nothing, Nothing}, @Kwargs{}, SciMLBase.StandardNonlinearProblem}, AutoPolyesterForwardDiff{nothing, Nothing}}, backend::AutoPolyesterForwardDiff{nothing, Nothing}, x::SVector{2, Float64}, contexts::DifferentiationInterface.Constant{Float64})
      @ DifferentiationInterface /mnt/.julia/packages/DifferentiationInterface/gPcSq/src/fallbacks/no_prep.jl:61

I can give you a better MWE once SciML/NonlinearSolve.jl#458 this massive PR is merged

[gdalle]

My bad, I see what you mean. My custom batch size selection is forwarded, but in methods like this one

DifferentiationInterface.jl/DifferentiationInterface/ext/DifferentiationInterfacePolyesterForwardDiffExt/onearg.jl

Lines 129 to 140 in 88c48c1

	function DI.gradient!(
	f,
	grad,
	::GradientPrep,
	::AutoPolyesterForwardDiff{K},
	x::AbstractVector,
	contexts::Vararg{Context,C},
	) where {K,C}
	fc = with_contexts(f, contexts...)
	threaded_gradient!(fc, grad, x, Chunk{K}())
	return grad
	end

I should distinguish on whether the chunksize in the backend is nothing

[gdalle]

It's an easy fix, I'll do it tomorrow