Merge branch 'master' into upstream

lib/OrdinaryDiffEqFIRK/Project.toml

@@ -6,12 +6,15 @@ version = "1.1.0"
 [deps]
 DiffEqBase = "2b5f629d-d688-5b77-993f-72d75c75574e"
 FastBroadcast = "7034ab61-46d4-4ed7-9d0f-46aef9175898"
+GenericLinearAlgebra = "14197337-ba66-59df-a3e3-ca00e7dcff7a"
+GenericSchur = "c145ed77-6b09-5dd9-b285-bf645a82121e"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 LinearSolve = "7ed4a6bd-45f5-4d41-b270-4a48e9bafcae"
 MuladdMacro = "46d2c3a1-f734-5fdb-9937-b9b9aeba4221"
 OrdinaryDiffEqCore = "bbf590c4-e513-4bbe-9b18-05decba2e5d8"
 OrdinaryDiffEqDifferentiation = "4302a76b-040a-498a-8c04-15b101fed76b"
 OrdinaryDiffEqNonlinearSolve = "127b3ac7-2247-4354-8eb6-78cf4e7c58e8"
+Polynomials = "f27b6e38-b328-58d1-80ce-0feddd5e7a45"
 RecursiveArrayTools = "731186ca-8d62-57ce-b412-fbd966d074cd"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 SciMLOperators = "c0aeaf25-5076-4817-a8d5-81caf7dfa961"

lib/OrdinaryDiffEqFIRK/src/OrdinaryDiffEqFIRK.jl

@@ -18,6 +18,7 @@ import OrdinaryDiffEqCore: alg_order, calculate_residuals!,
                            get_current_adaptive_order, get_fsalfirstlast,
                            isfirk
 using MuladdMacro, DiffEqBase, RecursiveArrayTools
+using Polynomials, GenericLinearAlgebra, GenericSchur
 using SciMLOperators: AbstractSciMLOperator
 using LinearAlgebra: I, UniformScaling, mul!, lu
 import LinearSolve
@@ -42,6 +43,6 @@ include("firk_tableaus.jl")
 include("firk_perform_step.jl")
 include("integrator_interface.jl")
 
-export RadauIIA3, RadauIIA5, RadauIIA9
+export RadauIIA3, RadauIIA5, RadauIIA9, AdaptiveRadau
 
 end

lib/OrdinaryDiffEqFIRK/src/alg_utils.jl

@@ -3,10 +3,12 @@ qmax_default(alg::Union{RadauIIA3, RadauIIA5, RadauIIA9}) = 8
 alg_order(alg::RadauIIA3) = 3
 alg_order(alg::RadauIIA5) = 5
 alg_order(alg::RadauIIA9) = 9
+alg_order(alg::AdaptiveRadau) = 9
 
 isfirk(alg::RadauIIA3) = true
 isfirk(alg::RadauIIA5) = true
 isfirk(alg::RadauIIA9) = true
+isfirk(alg::AdaptiveRadau) = true
 
 alg_adaptive_order(alg::RadauIIA3) = 1
 alg_adaptive_order(alg::RadauIIA5) = 3

lib/OrdinaryDiffEqFIRK/src/algorithms.jl

@@ -153,3 +153,42 @@ function RadauIIA9(; chunk_size = Val{0}(), autodiff = Val{true}(),
         controller,
         step_limiter!)
 end
+
+struct AdaptiveRadau{CS, AD, F, P, FDT, ST, CJ, Tol, C1, C2, StepLimiter} <:
+    OrdinaryDiffEqNewtonAdaptiveAlgorithm{CS, AD, FDT, ST, CJ}
+ linsolve::F
+ precs::P
+ smooth_est::Bool
+ extrapolant::Symbol
+ κ::Tol
+ maxiters::Int
+ fast_convergence_cutoff::C1
+ new_W_γdt_cutoff::C2
+ controller::Symbol
+ step_limiter!::StepLimiter
+ num_stages::Int
+end
+
+function AdaptiveRadau(; chunk_size = Val{0}(), autodiff = Val{true}(),
+     standardtag = Val{true}(), concrete_jac = nothing,
+     diff_type = Val{:forward}, num_stages = 3, 
+     linsolve = nothing, precs = DEFAULT_PRECS,
+     extrapolant = :dense, fast_convergence_cutoff = 1 // 5,
+     new_W_γdt_cutoff = 1 // 5,
+     controller = :Predictive, κ = nothing, maxiters = 10, smooth_est = true,
+     step_limiter! = trivial_limiter!)
+ AdaptiveRadau{_unwrap_val(chunk_size), _unwrap_val(autodiff), typeof(linsolve),
+     typeof(precs), diff_type, _unwrap_val(standardtag), _unwrap_val(concrete_jac),
+     typeof(κ), typeof(fast_convergence_cutoff),
+     typeof(new_W_γdt_cutoff), typeof(step_limiter!)}(linsolve,
+     precs,
+     smooth_est,
+     extrapolant,
+     κ,
+     maxiters,
+     fast_convergence_cutoff,
+     new_W_γdt_cutoff,
+     controller,
+     step_limiter!, num_stages)
+end
+

lib/OrdinaryDiffEqFIRK/src/firk_caches.jl

@@ -470,3 +470,164 @@ function alg_cache(alg::RadauIIA9, u, rate_prototype, ::Type{uEltypeNoUnits},
         linsolve1, linsolve2, linsolve3, rtol, atol, dt, dt,
         Convergence, alg.step_limiter!)
 end
+
+mutable struct AdaptiveRadauConstantCache{F, Tab, Tol, Dt, U, JType} <:
+    OrdinaryDiffEqConstantCache
+    uf::F
+    tab::Tab
+    κ::Tol
+    ηold::Tol
+    iter::Int
+    cont::AbstractVector{U}
+    dtprev::Dt
+    W_γdt::Dt
+    status::NLStatus
+    J::JType
+end
+
+function alg_cache(alg::AdaptiveRadau, u, rate_prototype, ::Type{uEltypeNoUnits},
+        ::Type{uBottomEltypeNoUnits},
+        ::Type{tTypeNoUnits}, uprev, uprev2, f, t, dt, reltol, p, calck,
+        ::Val{false}) where {uEltypeNoUnits, uBottomEltypeNoUnits, tTypeNoUnits}
+    uf = UDerivativeWrapper(f, t, p)
+    uToltype = constvalue(uBottomEltypeNoUnits)
+    num_stages = alg.num_stages
+    tab = adaptiveRadauTableau(uToltype, constvalue(tTypeNoUnits), num_stages)
+
+    cont = Vector{typeof(u)}(undef, num_stages - 1)
+    for i in 1: (num_stages - 1)
+        cont[i] = zero(u)
+    end
+
+    κ = alg.κ !== nothing ? convert(uToltype, alg.κ) : convert(uToltype, 1 // 100)
+    J = false .* _vec(rate_prototype) .* _vec(rate_prototype)'
+
+    AdaptiveRadauConstantCache(uf, tab, κ, one(uToltype), 10000, cont, dt, dt,
+        Convergence, J)
+end
+
+mutable struct AdaptiveRadauCache{uType, cuType, uNoUnitsType, rateType, JType, W1Type, W2Type,
+    UF, JC, F1, F2, Tab, Tol, Dt, rTol, aTol, StepLimiter} <:
+               OrdinaryDiffEqMutableCache
+    u::uType
+    uprev::uType
+    z::AbstractVector{uType}
+    w::AbstractVector{uType}
+    dw1::uType
+    ubuff::uType
+    dw2::AbstractVector{cuType}
+    cubuff::AbstractVector{cuType}
+    cont::AbstractVector{uType}
+    du1::rateType
+    fsalfirst::rateType
+    k::AbstractVector{rateType}
+    fw::AbstractVector{rateType}
+    J::JType
+    W1::W1Type #real
+    W2::AbstractVector{W2Type} #complex
+    uf::UF
+    tab::Tab
+    κ::Tol
+    ηold::Tol
+    iter::Int
+    tmp::AbstractVector{uType}
+    atmp::uNoUnitsType
+    jac_config::JC
+    linsolve1::F1 #real
+    linsolve2::AbstractVector{F2} #complex
+    rtol::rTol
+    atol::aTol
+    dtprev::Dt
+    W_γdt::Dt
+    status::NLStatus
+    step_limiter!::StepLimiter
+end
+
+function alg_cache(alg::AdaptiveRadau, u, rate_prototype, ::Type{uEltypeNoUnits},
+        ::Type{uBottomEltypeNoUnits},
+        ::Type{tTypeNoUnits}, uprev, uprev2, f, t, dt, reltol, p, calck,
+        ::Val{true}) where {uEltypeNoUnits, uBottomEltypeNoUnits, tTypeNoUnits}
+    uf = UJacobianWrapper(f, t, p)
+    uToltype = constvalue(uBottomEltypeNoUnits)
+    alg.num_stages = num_stages
+    tab = adaptiveRadauTableau(uToltype, constvalue(tTypeNoUnits), num_stages)
+
+    κ = alg.κ !== nothing ? convert(uToltype, alg.κ) : convert(uToltype, 1 // 100)
+
+    z = Vector{typeof(u)}(undef, num_stages)
+    w = Vector{typeof(u)}(undef, num_stages)
+    for i in 1:s
+        z[i] = w[i] = zero(u)
+    end
+
+    dw1 = zero(u)
+    ubuff = zero(u)
+    dw2 = Vector{typeof(u)}(undef, floor(Int, num_stages/2))
+    for i in 1 : floor(Int, num_stages/2)
+        dw2[i] = similar(u, Complex{eltype(u)})
+        recursivefill!(dw[i], false)
+    end
+    cubuff = Vector{typeof(u)}(undef, floor(Int, num_stages/2))
+    for i in 1 :floor(Int, num_stages/2)
+        cubuff[i] = similar(u, Complex{eltype(u)})
+        recursivefill!(cubuff[i], false)
+    end
+
+    cont = Vector{typeof(u)}(undef, num_stages - 1)
+    for i in 1: (num_stages - 1)
+        cont[i] = zero(u)
+    end
+
+    fsalfirst = zero(rate_prototype)
+    fw = Vector{typeof(rate_prototype)}(undef, num_stages)
+    k = Vector{typeof(rate_prototype)}(undef, num_stages)
+    for i in 1: num_stages
+        k[i] = fw[i] = zero(rate_prototype)
+    end
+
+    J, W1 = build_J_W(alg, u, uprev, p, t, dt, f, uEltypeNoUnits, Val(true))
+    if J isa AbstractSciMLOperator
+        error("Non-concrete Jacobian not yet supported by RadauIIA5.")
+    end
+    W2 = vector{typeof(Complex{W1})}(undef, floor(Int, num_stages/2))
+    for i in 1 : floor(Int, num_stages/2)
+        W2[i] = similar(J, Complex{eltype(W1)})
+        recursivefill!(w2[i], false)
+    end
+
+    du1 = zero(rate_prototype)
+
+    tmp = Vector{typeof(u)}(undef, binomial(num_stages,2))
+    for i in 1 : binomial(num_stages,2)
+        tmp[i] = zero(u)
+    end
+
+    atmp = similar(u, uEltypeNoUnits)
+    recursivefill!(atmp, false)
+
+    jac_config = build_jac_config(alg, f, uf, du1, uprev, u, tmp, dw1)
+
+    linprob = LinearProblem(W1, _vec(ubuff); u0 = _vec(dw1))
+    linsolve1 = init(linprob, alg.linsolve, alias_A = true, alias_b = true,
+        assumptions = LinearSolve.OperatorAssumptions(true))
+
+    linsolve2 = Vector{typeof(linsolve1)}(undef, floor(Int, num_stages/2))
+    for i in 1 : floor(int, num_stages/2)
+        linprob = LinearProblem(W2[i], _vec(cubuff[i]); u0 = _vec(dw2[i]))
+        linsolve2 = init(linprob, alg.linsolve, alias_A = true, alias_b = true,
+        assumptions = LinearSolve.OperatorAssumptions(true))
+    end
+
+    rtol = reltol isa Number ? reltol : zero(reltol)
+    atol = reltol isa Number ? reltol : zero(reltol)
+
+    AdaptiveRadauCache(u, uprev,
+        z, w, dw1, ubuff, dw2, cubuff, cont,
+        du1, fsalfirst, k, fw,
+        J, W1, W2,
+        uf, tab, κ, one(uToltype), 10000,
+        tmp, atmp, jac_config,
+        linsolve1, linsolve2, rtol, atol, dt, dt,
+        Convergence, alg.step_limiter!)
+end
+