Merge pull request #449 from SciML/ap/adjoint

Adjoints for Linear Solve

Author: ChrisRackauckas

Project.toml

@@ -1,10 +1,11 @@
 name = "LinearSolve"
 uuid = "7ed4a6bd-45f5-4d41-b270-4a48e9bafcae"
 authors = ["SciML"]
-version = "2.24.0"
+version = "2.25.0"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 ConcreteStructs = "2569d6c7-a4a2-43d3-a901-331e8e4be471"
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
 EnumX = "4e289a0a-7415-4d19-859d-a7e5c4648b56"
@@ -16,6 +17,7 @@ Krylov = "ba0b0d4f-ebba-5204-a429-3ac8c609bfb7"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MKL_jll = "856f044c-d86e-5d09-b602-aeab76dc8ba7"
+Markdown = "d6f4376e-aef5-505a-96c1-9c027394607a"
 PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 RecursiveFactorization = "f2c3362d-daeb-58d1-803e-2bc74f2840b4"
@@ -64,6 +66,7 @@ ArrayInterface = "7.7"
 BandedMatrices = "1.5"
 BlockDiagonals = "0.1.42"
 CUDA = "5"
+ChainRulesCore = "1.22"
 ConcreteStructs = "0.2.3"
 DocStringExtensions = "0.9.3"
 EnumX = "1.0.4"
@@ -85,6 +88,7 @@ KrylovKit = "0.6"
 Libdl = "1.10"
 LinearAlgebra = "1.10"
 MPI = "0.20"
+Markdown = "1.10"
 Metal = "0.5"
 MultiFloats = "1"
 Pardiso = "0.5"
@@ -96,7 +100,7 @@ RecursiveArrayTools = "3.8"
 RecursiveFactorization = "0.2.14"
 Reexport = "1"
 SafeTestsets = "0.1"
-SciMLBase = "2.23.0"
+SciMLBase = "2.26.3"
 SciMLOperators = "0.3.7"
 Setfield = "1"
 SparseArrays = "1.10"
@@ -106,6 +110,7 @@ StaticArrays = "1.5"
 StaticArraysCore = "1.4.2"
 Test = "1"
 UnPack = "1"
+Zygote = "0.6.69"
 julia = "1.10"
 
 [extras]
@@ -133,6 +138,7 @@ SafeTestsets = "1bc83da4-3b8d-516f-aca4-4fe02f6d838f"
 StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["Aqua", "Test", "IterativeSolvers", "InteractiveUtils", "JET", "KrylovKit", "Pkg", "Random", "SafeTestsets", "MultiFloats", "ForwardDiff", "HYPRE", "MPI", "BlockDiagonals", "Enzyme", "FiniteDiff", "BandedMatrices", "FastAlmostBandedMatrices", "StaticArrays", "AllocCheck", "StableRNGs"]
+test = ["Aqua", "Test", "IterativeSolvers", "InteractiveUtils", "JET", "KrylovKit", "Pkg", "Random", "SafeTestsets", "MultiFloats", "ForwardDiff", "HYPRE", "MPI", "BlockDiagonals", "Enzyme", "FiniteDiff", "BandedMatrices", "FastAlmostBandedMatrices", "StaticArrays", "AllocCheck", "StableRNGs", "Zygote"]

ext/LinearSolveHYPREExt.jl

@@ -5,7 +5,7 @@ using HYPRE.LibHYPRE: HYPRE_Complex
 using HYPRE: HYPRE, HYPREMatrix, HYPRESolver, HYPREVector
 using LinearSolve: HYPREAlgorithm, LinearCache, LinearProblem, LinearSolve,
                    OperatorAssumptions, default_tol, init_cacheval, __issquare,
-                   __conditioning
+                   __conditioning, LinearSolveAdjoint
 using SciMLBase: LinearProblem, SciMLBase
 using UnPack: @unpack
 using Setfield: @set!
@@ -68,6 +68,7 @@ function SciMLBase.init(prob::LinearProblem, alg::HYPREAlgorithm,
         Pl = LinearAlgebra.I,
         Pr = LinearAlgebra.I,
         assumptions = OperatorAssumptions(),
+        sensealg = LinearSolveAdjoint(),
         kwargs...)
     @unpack A, b, u0, p = prob
 
@@ -89,10 +90,9 @@ function SciMLBase.init(prob::LinearProblem, alg::HYPREAlgorithm,
     cache = LinearCache{
         typeof(A), typeof(b), typeof(u0), typeof(p), typeof(alg), Tc,
         typeof(Pl), typeof(Pr), typeof(reltol),
-        typeof(__issquare(assumptions))
+        typeof(__issquare(assumptions)), typeof(sensealg)
     }(A, b, u0, p, alg, cacheval, isfresh, Pl, Pr, abstol, reltol,
-        maxiters,
-        verbose, assumptions)
+        maxiters, verbose, assumptions, sensealg)
     return cache
 end
 

src/LinearSolve.jl

@@ -23,6 +23,8 @@ PrecompileTools.@recompile_invalidations begin
     using FastLapackInterface
     using DocStringExtensions
     using EnumX
+    using Markdown
+    using ChainRulesCore
     import InteractiveUtils
 
     import StaticArraysCore: StaticArray, SVector, MVector, SMatrix, MMatrix
@@ -42,6 +44,8 @@ PrecompileTools.@recompile_invalidations begin
     import Preferences
 end
 
+const CRC = ChainRulesCore
+
 if Preferences.@load_preference("LoadMKL_JLL", true)
     using MKL_jll
     const usemkl = MKL_jll.is_available()
@@ -125,6 +129,7 @@ include("solve_function.jl")
 include("default.jl")
 include("init.jl")
 include("extension_algs.jl")
+include("adjoint.jl")
 include("deprecated.jl")
 
 @generated function SciMLBase.solve!(cache::LinearCache, alg::AbstractFactorization;
@@ -240,4 +245,6 @@ export MetalLUFactorization
 
 export OperatorAssumptions, OperatorCondition
 
+export LinearSolveAdjoint
+
 end

src/adjoint.jl

@@ -0,0 +1,93 @@
+# TODO: Preconditioners? Should Pl be transposed and made Pr and similar for Pr.
+
+@doc doc"""
+    LinearSolveAdjoint(; linsolve = missing)
+
+Given a Linear Problem ``A x = b`` computes the sensitivities for ``A`` and ``b`` as:
+
+```math
+\begin{align}
+A^T \lambda &= \partial x   \\
+\partial A  &= -\lambda x^T \\
+\partial b  &= \lambda
+\end{align}
+```
+
+For more details, check [these notes](https://math.mit.edu/~stevenj/18.336/adjoint.pdf).
+
+## Choice of Linear Solver
+
+Note that in most cases, it makes sense to use the same linear solver for the adjoint as the
+forward solve (this is done by keeping the linsolve as `missing`). For example, if the
+forward solve was performed via a Factorization, then we can reuse the factorization for the
+adjoint solve. However, for specific structured matrices if ``A^T`` is known to have a
+specific structure distinct from ``A`` then passing in a `linsolve` will be more efficient.
+"""
+@kwdef struct LinearSolveAdjoint{L} <:
+              SciMLBase.AbstractSensitivityAlgorithm{0, false, :central}
+    linsolve::L = missing
+end
+
+function CRC.rrule(::typeof(SciMLBase.solve), prob::LinearProblem,
+        alg::SciMLLinearSolveAlgorithm, args...; alias_A = default_alias_A(
+            alg, prob.A, prob.b), kwargs...)
+    # sol = solve(prob, alg, args...; kwargs...)
+    cache = init(prob, alg, args...; kwargs...)
+    (; A, sensealg) = cache
+
+    @assert sensealg isa LinearSolveAdjoint "Currently only `LinearSolveAdjoint` is supported for adjoint sensitivity analysis."
+
+    # Decide if we need to cache `A` and `b` for the reverse pass
+    if sensealg.linsolve === missing
+        # We can reuse the factorization so no copy is needed
+        # Krylov Methods don't modify `A`, so it's safe to just reuse it
+        # No Copy is needed even for the default case
+        if !(alg isa AbstractFactorization || alg isa AbstractKrylovSubspaceMethod ||
+             alg isa DefaultLinearSolver)
+            A_ = alias_A ? deepcopy(A) : A
+        end
+    else
+        A_ = deepcopy(A)
+    end
+
+    sol = solve!(cache)
+
+    function ∇linear_solve(∂sol)
+        ∂∅ = NoTangent()
+
+        ∂u = ∂sol.u
+        if sensealg.linsolve === missing
+            λ = if cache.cacheval isa Factorization
+                cache.cacheval' \ ∂u
+            elseif cache.cacheval isa Tuple && cache.cacheval[1] isa Factorization
+                first(cache.cacheval)' \ ∂u
+            elseif alg isa AbstractKrylovSubspaceMethod
+                invprob = LinearProblem(transpose(cache.A), ∂u)
+                solve(invprob, alg; cache.abstol, cache.reltol, cache.verbose).u
+            elseif alg isa DefaultLinearSolver
+                LinearSolve.defaultalg_adjoint_eval(cache, ∂u)
+            else
+                invprob = LinearProblem(transpose(A_), ∂u) # We cached `A`
+                solve(invprob, alg; cache.abstol, cache.reltol, cache.verbose).u
+            end
+        else
+            invprob = LinearProblem(transpose(A_), ∂u) # We cached `A`
+            λ = solve(
+                invprob, sensealg.linsolve; cache.abstol, cache.reltol, cache.verbose).u
+        end
+
+        ∂A = -λ * transpose(sol.u)
+        ∂b = λ
+        ∂prob = LinearProblem(∂A, ∂b, ∂∅)
+
+        return (∂∅, ∂prob, ∂∅, ntuple(_ -> ∂∅, length(args))...)
+    end
+
+    return sol, ∇linear_solve
+end
+
+function CRC.rrule(::Type{<:LinearProblem}, A, b, p; kwargs...)
+    prob = LinearProblem(A, b, p)
+    ∇prob(∂prob) = (NoTangent(), ∂prob.A, ∂prob.b, ∂prob.p)
+    return prob, ∇prob
+end

src/common.jl

@@ -65,7 +65,7 @@ end
 __issquare(assump::OperatorAssumptions) = assump.issq
 __conditioning(assump::OperatorAssumptions) = assump.condition
 
-mutable struct LinearCache{TA, Tb, Tu, Tp, Talg, Tc, Tl, Tr, Ttol, issq}
+mutable struct LinearCache{TA, Tb, Tu, Tp, Talg, Tc, Tl, Tr, Ttol, issq, S}
     A::TA
     b::Tb
     u::Tu
@@ -80,6 +80,7 @@ mutable struct LinearCache{TA, Tb, Tu, Tp, Talg, Tc, Tl, Tr, Ttol, issq}
     maxiters::Int
     verbose::Bool
     assumptions::OperatorAssumptions{issq}
+    sensealg::S
 end
 
 function Base.setproperty!(cache::LinearCache, name::Symbol, x)
@@ -138,6 +139,7 @@ function SciMLBase.init(prob::LinearProblem, alg::SciMLLinearSolveAlgorithm,
         Pl = IdentityOperator(size(prob.A)[1]),
         Pr = IdentityOperator(size(prob.A)[2]),
         assumptions = OperatorAssumptions(issquare(prob.A)),
+        sensealg = LinearSolveAdjoint(),
         kwargs...)
     @unpack A, b, u0, p = prob
 
@@ -171,17 +173,22 @@ function SciMLBase.init(prob::LinearProblem, alg::SciMLLinearSolveAlgorithm,
     Tc = typeof(cacheval)
 
     cache = LinearCache{typeof(A), typeof(b), typeof(u0_), typeof(p), typeof(alg), Tc,
-        typeof(Pl), typeof(Pr), typeof(reltol), typeof(assumptions.issq)}(A, b, u0_,
-        p, alg, cacheval, isfresh, Pl, Pr, abstol, reltol, maxiters, verbose, assumptions)
+        typeof(Pl), typeof(Pr), typeof(reltol), typeof(assumptions.issq),
+        typeof(sensealg)}(A, b, u0_, p, alg, cacheval, isfresh, Pl, Pr, abstol, reltol,
+        maxiters, verbose, assumptions, sensealg)
     return cache
 end
 
 function SciMLBase.solve(prob::LinearProblem, args...; kwargs...)
-    solve!(init(prob, nothing, args...; kwargs...))
+    return solve(prob, nothing, args...; kwargs...)
 end
 
-function SciMLBase.solve(prob::LinearProblem,
-        alg::Union{SciMLLinearSolveAlgorithm, Nothing},
+function SciMLBase.solve(prob::LinearProblem, ::Nothing, args...;
+        assump = OperatorAssumptions(issquare(prob.A)), kwargs...)
+    return solve(prob, defaultalg(prob.A, prob.b, assump), args...; kwargs...)
+end
+
+function SciMLBase.solve(prob::LinearProblem, alg::SciMLLinearSolveAlgorithm,
         args...; kwargs...)
     solve!(init(prob, alg, args...; kwargs...))
 end

src/factorization.jl

@@ -779,26 +779,30 @@ function SciMLBase.solve!(cache::LinearCache, alg::UMFPACKFactorization; kwargs.
                  cacheval.colptr &&
                  SparseArrays.decrement(SparseArrays.getrowval(A)) ==
                  cacheval.rowval)
-                fact = lu(SparseMatrixCSC(size(A)..., getcolptr(A), rowvals(A),
-                    nonzeros(A)), check=false)
+                fact = lu(
+                    SparseMatrixCSC(size(A)..., getcolptr(A), rowvals(A),
+                        nonzeros(A)),
+                    check = false)
             else
                 fact = lu!(cacheval,
                     SparseMatrixCSC(size(A)..., getcolptr(A), rowvals(A),
-                        nonzeros(A)), check=false)
+                        nonzeros(A)), check = false)
             end
         else
-            fact = lu(SparseMatrixCSC(size(A)..., getcolptr(A), rowvals(A), nonzeros(A)), check=false)
+            fact = lu(SparseMatrixCSC(size(A)..., getcolptr(A), rowvals(A), nonzeros(A)),
+                check = false)
         end
         cache.cacheval = fact
         cache.isfresh = false
     end
 
-    F =  @get_cacheval(cache, :UMFPACKFactorization)
+    F = @get_cacheval(cache, :UMFPACKFactorization)
     if F.status == SparseArrays.UMFPACK.UMFPACK_OK
         y = ldiv!(cache.u, F, cache.b)
         SciMLBase.build_linear_solution(alg, y, nothing, cache)
     else
-        SciMLBase.build_linear_solution(alg, cache.u, nothing, cache; retcode=ReturnCode.Infeasible)
+        SciMLBase.build_linear_solution(
+            alg, cache.u, nothing, cache; retcode = ReturnCode.Infeasible)
     end
 end