Merge pull request #621 from jClugstor/forwarddiff_overloads

Overloads for LinearProblems with ForwardDiff Dual numbers

Author: ChrisRackauckas

Project.toml

@@ -35,6 +35,7 @@ CUDSS = "45b445bb-4962-46a0-9369-b4df9d0f772e"
 EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
 FastAlmostBandedMatrices = "9d29842c-ecb8-4973-b1e9-a27b1157504e"
 FastLapackInterface = "29a986be-02c6-4525-aec4-84b980013641"
+ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 HYPRE = "b5ffcf37-a2bd-41ab-a3da-4bd9bc8ad771"
 IterativeSolvers = "42fd0dbc-a981-5370-80f2-aaf504508153"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
@@ -53,6 +54,7 @@ LinearSolveCUDSSExt = "CUDSS"
 LinearSolveEnzymeExt = "EnzymeCore"
 LinearSolveFastAlmostBandedMatricesExt = "FastAlmostBandedMatrices"
 LinearSolveFastLapackInterfaceExt = "FastLapackInterface"
+LinearSolveForwardDiffExt = "ForwardDiff"
 LinearSolveHYPREExt = "HYPRE"
 LinearSolveIterativeSolversExt = "IterativeSolvers"
 LinearSolveKernelAbstractionsExt = "KernelAbstractions"

ext/LinearSolveForwardDiffExt.jl

@@ -0,0 +1,241 @@
+module LinearSolveForwardDiffExt
+
+using LinearSolve
+using LinearAlgebra
+using ForwardDiff
+using ForwardDiff: Dual, Partials
+using SciMLBase
+using RecursiveArrayTools
+
+const DualLinearProblem = LinearProblem{
+    <:Union{Number, <:AbstractArray, Nothing}, iip,
+    <:Union{<:Dual{T, V, P}, <:AbstractArray{<:Dual{T, V, P}}},
+    <:Union{<:Dual{T, V, P}, <:AbstractArray{<:Dual{T, V, P}}},
+    <:Any
+} where {iip, T, V, P}
+
+const DualALinearProblem = LinearProblem{
+    <:Union{Number, <:AbstractArray, Nothing},
+    iip,
+    <:Union{<:Dual{T, V, P}, <:AbstractArray{<:Dual{T, V, P}}},
+    <:Union{Number, <:AbstractArray},
+    <:Any
+} where {iip, T, V, P}
+
+const DualBLinearProblem = LinearProblem{
+    <:Union{Number, <:AbstractArray, Nothing},
+    iip,
+    <:Union{Number, <:AbstractArray},
+    <:Union{<:Dual{T, V, P}, <:AbstractArray{<:Dual{T, V, P}}},
+    <:Any
+} where {iip, T, V, P}
+
+const DualAbstractLinearProblem = Union{
+    DualLinearProblem, DualALinearProblem, DualBLinearProblem}
+
+LinearSolve.@concrete mutable struct DualLinearCache
+    linear_cache
+    dual_type
+    partials_A
+    partials_b
+end
+
+function linearsolve_forwarddiff_solve(cache::DualLinearCache, alg, args...; kwargs...)
+    # Solve the primal problem
+    dual_u0 = copy(cache.linear_cache.u)
+    sol = solve!(cache.linear_cache, alg, args...; kwargs...)
+    primal_b = copy(cache.linear_cache.b)
+    uu = sol.u
+
+    primal_sol = deepcopy(sol)
+
+    # Solves Dual partials separately 
+    ∂_A = cache.partials_A
+    ∂_b = cache.partials_b
+
+    rhs_list = xp_linsolve_rhs(uu, ∂_A, ∂_b)
+
+    partial_cache = cache.linear_cache
+    partial_cache.u = dual_u0
+
+    for i in eachindex(rhs_list)
+        partial_cache.b = rhs_list[i]
+        rhs_list[i] = copy(solve!(partial_cache, alg, args...; kwargs...).u)
+    end
+
+    # Reset to the original `b`, users will expect that `b` doesn't change if they don't tell it to
+    partial_cache.b = primal_b
+
+    partial_sols = rhs_list
+
+    primal_sol, partial_sols
+end
+
+function xp_linsolve_rhs(uu, ∂_A::Union{<:Partials, <:AbstractArray{<:Partials}},
+        ∂_b::Union{<:Partials, <:AbstractArray{<:Partials}})
+    A_list = partials_to_list(∂_A)
+    b_list = partials_to_list(∂_b)
+
+    Auu = [A * uu for A in A_list]
+
+    return b_list .- Auu
+end
+
+function xp_linsolve_rhs(
+        uu, ∂_A::Union{<:Partials, <:AbstractArray{<:Partials}}, ∂_b::Nothing)
+    A_list = partials_to_list(∂_A)
+
+    Auu = [A * uu for A in A_list]
+
+    return -Auu
+end
+
+function xp_linsolve_rhs(
+        uu, ∂_A::Nothing, ∂_b::Union{<:Partials, <:AbstractArray{<:Partials}})
+    b_list = partials_to_list(∂_b)
+    b_list
+end
+
+function SciMLBase.solve(prob::DualAbstractLinearProblem, args...; kwargs...)
+    return solve(prob, nothing, args...; kwargs...)
+end
+
+function SciMLBase.solve(prob::DualAbstractLinearProblem, ::Nothing, args...;
+        assump = OperatorAssumptions(issquare(prob.A)), kwargs...)
+    return solve(prob, LinearSolve.defaultalg(prob.A, prob.b, assump), args...; kwargs...)
+end
+
+function SciMLBase.solve(prob::DualAbstractLinearProblem,
+        alg::LinearSolve.SciMLLinearSolveAlgorithm, args...; kwargs...)
+    solve!(init(prob, alg, args...; kwargs...))
+end
+
+function linearsolve_dual_solution(
+        u::Number, partials, dual_type)
+    return dual_type(u, partials)
+end
+
+function linearsolve_dual_solution(
+        u::AbstractArray, partials, dual_type)
+    partials_list = RecursiveArrayTools.VectorOfArray(partials)
+    return map(((uᵢ, pᵢ),) -> dual_type(uᵢ, Partials(Tuple(pᵢ))),
+        zip(u, partials_list[i, :] for i in 1:length(partials_list[1])))
+end
+
+function SciMLBase.init(
+        prob::DualAbstractLinearProblem, alg::LinearSolve.SciMLLinearSolveAlgorithm,
+        args...;
+        alias = LinearAliasSpecifier(),
+        abstol = LinearSolve.default_tol(real(eltype(prob.b))),
+        reltol = LinearSolve.default_tol(real(eltype(prob.b))),
+        maxiters::Int = length(prob.b),
+        verbose::Bool = false,
+        Pl = nothing,
+        Pr = nothing,
+        assumptions = OperatorAssumptions(issquare(prob.A)),
+        sensealg = LinearSolveAdjoint(),
+        kwargs...)
+
+    (; A, b, u0, p) = prob
+    new_A = nodual_value(A)
+    new_b = nodual_value(b)
+    new_u0 = nodual_value(u0)
+
+    ∂_A = partial_vals(A)
+    ∂_b = partial_vals(b)
+
+    #primal_prob = LinearProblem(new_A, new_b, u0 = new_u0)
+    primal_prob = remake(prob; A = new_A, b = new_b, u0 = new_u0)
+
+    if get_dual_type(prob.A) !== nothing
+        dual_type = get_dual_type(prob.A)
+    elseif get_dual_type(prob.b) !== nothing
+        dual_type = get_dual_type(prob.b)
+    end
+
+    non_partial_cache = init(
+        primal_prob, alg, args...; alias = alias, abstol = abstol, reltol = reltol,
+        maxiters = maxiters, verbose = verbose, Pl = Pl, Pr = Pr, assumptions = assumptions,
+        sensealg = sensealg, u0 = new_u0, kwargs...)
+    return DualLinearCache(non_partial_cache, dual_type, ∂_A, ∂_b)
+end
+
+function SciMLBase.solve!(cache::DualLinearCache, args...; kwargs...)
+    sol,
+    partials = linearsolve_forwarddiff_solve(
+        cache::DualLinearCache, cache.alg, args...; kwargs...)
+
+    dual_sol = linearsolve_dual_solution(sol.u, partials, cache.dual_type)
+    return SciMLBase.build_linear_solution(
+        cache.alg, dual_sol, sol.resid, cache; sol.retcode, sol.iters, sol.stats
+    )
+end
+
+# If setting A or b for DualLinearCache, put the Dual-stripped versions in the LinearCache
+# Also "forwards" setproperty so that 
+function Base.setproperty!(dc::DualLinearCache, sym::Symbol, val)
+    # If the property is A or b, also update it in the LinearCache
+    if sym === :A || sym === :b || sym === :u
+        setproperty!(dc.linear_cache, sym, nodual_value(val))
+    elseif hasfield(LinearSolve.LinearCache, sym)
+        setproperty!(dc.linear_cache, sym, val)
+    end
+
+    # Update the partials if setting A or b
+    if sym === :A
+        setfield!(dc, :partials_A, partial_vals(val))
+    elseif  sym === :b
+        setfield!(dc, :partials_b, partial_vals(val))
+    else
+        setfield!(dc, sym, val)
+    end
+end
+
+# "Forwards" getproperty to LinearCache if necessary
+function Base.getproperty(dc::DualLinearCache, sym::Symbol)
+    if hasfield(LinearSolve.LinearCache, sym)
+        return getproperty(dc.linear_cache, sym)
+    else
+        return getfield(dc, sym)
+    end
+end
+
+
+
+# Helper functions for Dual numbers
+get_dual_type(x::Dual) = typeof(x)
+get_dual_type(x::AbstractArray{<:Dual}) = eltype(x)
+get_dual_type(x) = nothing
+
+partial_vals(x::Dual) = ForwardDiff.partials(x)
+partial_vals(x::AbstractArray{<:Dual}) = map(ForwardDiff.partials, x)
+partial_vals(x) = nothing
+
+nodual_value(x) = x
+nodual_value(x::Dual) = ForwardDiff.value(x)
+nodual_value(x::AbstractArray{<:Dual}) = map(ForwardDiff.value, x)
+
+
+function partials_to_list(partial_matrix::Vector)
+    p = eachindex(first(partial_matrix))
+    [[partial[i] for partial in partial_matrix] for i in p]
+end
+
+function partials_to_list(partial_matrix)
+    p = length(first(partial_matrix))
+    m, n = size(partial_matrix)
+    res_list = fill(zeros(m, n), p)
+    for k in 1:p
+        res = zeros(m, n)
+        for i in 1:m
+            for j in 1:n
+                res[i, j] = partial_matrix[i, j][k]
+            end
+        end
+        res_list[k] = res
+    end
+    return res_list
+end
+
+
+end

test/forwarddiff_overloads.jl

@@ -0,0 +1,82 @@
+using LinearSolve
+using ForwardDiff
+using Test
+
+function h(p)
+    (A = [p[1] p[2]+1 p[2]^3;
+          3*p[1] p[1]+5 p[2] * p[1]-4;
+          p[2]^2 9*p[1] p[2]],
+        b = [p[1] + 1, p[2] * 2, p[1]^2])
+end
+
+A, b = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+
+prob = LinearProblem(A, b)
+overload_x_p = solve(prob)
+backslash_x_p = A \ b
+krylov_overload_x_p = solve(prob, KrylovJL_GMRES())
+@test ≈(overload_x_p, backslash_x_p, rtol = 1e-9)
+@test ≈(krylov_overload_x_p, backslash_x_p, rtol = 1e-9)
+
+krylov_prob = LinearProblem(A, b, u0 = rand(3))
+krylov_u0_sol = solve(krylov_prob, KrylovJL_GMRES())
+
+@test ≈(krylov_u0_sol, backslash_x_p, rtol = 1e-9)
+
+
+A, _ = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+backslash_x_p = A \ [6.0, 10.0, 25.0]
+prob = LinearProblem(A, [6.0, 10.0, 25.0])
+
+@test ≈(solve(prob).u, backslash_x_p, rtol = 1e-9) 
+@test ≈(solve(prob, KrylovJL_GMRES()).u, backslash_x_p, rtol = 1e-9)
+
+_, b = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+A = [5.0 6.0 125.0; 15.0 10.0 21.0; 25.0 45.0 5.0]
+backslash_x_p = A \ b
+prob = LinearProblem(A, b)
+
+@test ≈(solve(prob).u, backslash_x_p, rtol = 1e-9)
+@test ≈(solve(prob, KrylovJL_GMRES()).u, backslash_x_p, rtol = 1e-9)
+
+A, b = h([ForwardDiff.Dual(10.0, 1.0, 0.0), ForwardDiff.Dual(10.0, 0.0, 1.0)])
+
+prob = LinearProblem(A, b)
+cache = init(prob)
+
+new_A, new_b = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+cache.A = new_A
+cache.b = new_b
+
+x_p = solve!(cache)
+backslash_x_p = new_A \ new_b
+
+@test ≈(x_p, backslash_x_p, rtol = 1e-9)
+
+# Just update A
+A, b = h([ForwardDiff.Dual(10.0, 1.0, 0.0), ForwardDiff.Dual(10.0, 0.0, 1.0)])
+
+prob = LinearProblem(A, b)
+cache = init(prob)
+
+new_A, _ = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+cache.A = new_A
+
+x_p = solve!(cache)
+backslash_x_p = new_A \ b
+
+@test ≈(x_p, backslash_x_p, rtol = 1e-9)
+
+# Just update b
+A, b = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+
+prob = LinearProblem(A, b)
+cache = init(prob)
+
+_, new_b = h([ForwardDiff.Dual(5.0, 1.0, 0.0), ForwardDiff.Dual(5.0, 0.0, 1.0)])
+cache.b = new_b
+
+x_p = solve!(cache)
+backslash_x_p = A \ new_b
+
+@test ≈(x_p, backslash_x_p, rtol = 1e-9)

test/runtests.jl

@@ -16,6 +16,7 @@ if GROUP == "All" || GROUP == "Core"
     @time @safetestset "SparseVector b Tests" include("sparse_vector.jl")
     @time @safetestset "Default Alg Tests" include("default_algs.jl")
     @time @safetestset "Adjoint Sensitivity" include("adjoint.jl")
+    @time @safetestset "ForwardDiff Overloads" include("forwarddiff_overloads.jl")
     @time @safetestset "Traits" include("traits.jl")
     @time @safetestset "BandedMatrices" include("banded.jl")
 end