Support for SparseDiffTools v2

[gaurav-arya]

No description provided.

[vpuri3]

can you also handle

OrdinaryDiffEq.jl/src/derivative_utils.jl

Lines 856 to 857 in 3885e70

	jacvec = SparseDiffTools.JacVec(UJacobianWrapper(_f, t, p), copy(u),
	OrdinaryDiffEqTag(), autodiff = alg_autodiff(alg))

OrdinaryDiffEq.jl/src/derivative_utils.jl

Lines 872 to 873 in 3885e70

	jacvec = SparseDiffTools.JacVec(UJacobianWrapper(_f, t, p), copy(u),
	OrdinaryDiffEqTag(), autodiff = alg_autodiff(alg))

alg_autodiff would need to be updated to return an AbstractADType https://github.com/SciML/ADTypes.jl/blob/main/src/ADTypes.jl

[vpuri3]

@ChrisRackauckas can you explain why we are using the OOP update_coefficients(L,u,p,t) here?

OrdinaryDiffEq.jl/src/perform_step/linear_perform_step.jl

Line 781 in cd2f814

L = update_coefficients(A, uprev, p, t)

That is something we have not fully supported in SciMLOps.

EDIT
OOP update_coeffs is fully supported in SciMLOps 0.2.2

[gaurav-arya]

@ChrisRackauckas are there any convergence tests of implicit methods with a time- or parameter-dependent f? I believe the test suite passed even though I think such a test should have failed prior to c7e778e (update_coefficients for a WOperator would not have properly updated the Jacobian operator if it depended on p or t)

If there aren't, where would you recommend adding one?

(I mean an explicit dependence on p or t, not an implicit one through u, that should indeed work even previously)

[ChrisRackauckas]

There are but not with operators I think. It would be good to explicitly test that

[vpuri3]

@gaurav-arya , check out test/regression/time_derivative_test. Looks like what youre looking for is being tested there?

[vpuri3]

looks like the time dependence test passed @gaurav-arya .

@ChrisRackauckas is there anything left to do here?

[gaurav-arya]

@vpuri3 do you know the cause of the speedup in the numbers you posted for Krylov solves? Was there a type instability before or has something changed algorithmically?

[ChrisRackauckas]

what is the effect here? Can you time a static array ROBER and make sure that doesn't regress?

[ChrisRackauckas]

? I don't understand.

[ChrisRackauckas]

Is that a regression?

[vpuri3]

@ChrisRackauckas StaticWOperator regression

using OrdinaryDiffEq, StaticArrays, BenchmarkTools

function rober(u, p, t)
    y₁, y₂, y₃ = u
    k₁, k₂, k₃ = p
    dy₁ = -k₁ * y₁ + k₃ * y₂ * y₃
    dy₂ = k₁ * y₁ - k₂ * y₂^2 - k₃ * y₂ * y₃
    dy₃ = k₂ * y₂^2
    SA[dy₁, dy₂, dy₃]
end

prob = ODEProblem{false}(rober, SA[1.0, 0.0, 0.0], (0.0, 1e5), SA[0.04, 3e7, 1e4])

for Alg in (Rodas4, Rosenbrock23)
    @info "$Alg"
    int = init(prob, Alg())
    @assert int.cache.W isa OrdinaryDiffEq.StaticWOperator

    @btime solve(prob, $Alg(), save_everystep = false)
end

on this branch

julia> include("b.jl")
[ Info: Rodas4
  59.250 μs (900 allocations: 38.14 KiB)
[ Info: Rosenbrock23
  46.375 μs (744 allocations: 30.97 KiB)

on master

julia> include("b.jl")
WARNING: using StaticArrays.push in module Main conflicts with an existing identifier.
[ Info: Rodas4
  60.042 μs (900 allocations: 38.14 KiB)
[ Info: Rosenbrock23
  48.542 μs (744 allocations: 31.08 KiB)

[vpuri3]

weird the accuracy error isn't showing up on my local tests (MacOS, Linux w. JL 1.9, ). rerunning CI hoping that its just a fluke

[ChrisRackauckas]

It went away?

The last thing then is to run Downstream locally and make sure it all works on master branches. If so, then this is good to merge.

[vpuri3]

it did go away. The only errors now are from downstream.

I'll run the downstream tests locally with the PR branches this afternoon.

[vpuri3]

@ChrisRackauckas, I ran downstream, ODEInterfaceRegression tests locally with this branch of stochasticdiffeq. Everything went smoothly except for:

1. an accuracy error in test/downstream/delaydiffeq.jl (should be a local problem as delaydiffeq tests pass on CI)
2. this error in test/odeinterface/init_dt_vs_dopri_tests.jl line 9, test/odeinterface/odeinterface_regression line 41.

Init dt vs dorpri tests: Error During Test at /Users/vp/.julia/packages/SafeTestsets/A8
3XK/src/SafeTestsets.jl:25                                                             
  Got exception outside of a @test                                                     
  LoadError: Cannot find method(s) for dopri5! I've tried to loadODESolvers(), but it d
idn't work. Please check ODEInterface.help_solversupport() and call loadODESolvers and 
check also this output. For further information see also ODEInterface.help_install. 

I also updated the compats in downstream/Project.toml. With that and the StochasticDiffEq PR branch, the compatibility error disappeared. Can also confirm that tests pass locally on ProbNumDiffEq PR branch with this ODE.jl branch.

[vpuri3]

Can also confirm that all StochasticDiffEq tests pass locally on the PR branch with this ODE branch

[vpuri3]

@ChrisRackauckas good to merge?

[ChrisRackauckas]

Alright, here it goes.

[ChrisRackauckas]

Where is the PR to DelayDiffEq?

[vpuri3]

Looks like delaydiffeq doesn't need a PR. The only reference to alg_autodiff (which is the API change) is below, and works fine with this branch.

function OrdinaryDiffEq.alg_autodiff(alg::AbstractMethodOfStepsAlgorithm)
    OrdinaryDiffEq.alg_autodiff(alg.alg)
end

I also ran DelayDiffEq tests locally. Everything passed.