Merge pull request #454 from SciML/operator_defaults

Fix and test solvers for non-square operators

Author: ChrisRackauckas

src/LinearSolve.jl

@@ -103,6 +103,8 @@ EnumX.@enumx DefaultAlgorithmChoice begin
     AppleAccelerateLUFactorization
     MKLLUFactorization
     QRFactorizationPivoted
+    KrylovJL_CRAIGMR
+    KrylovJL_LSMR
 end
 
 struct DefaultLinearSolver <: SciMLLinearSolveAlgorithm

src/default.jl

@@ -1,6 +1,6 @@
 needs_concrete_A(alg::DefaultLinearSolver) = true
 mutable struct DefaultLinearSolverInit{T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
-    T13, T14, T15, T16, T17, T18, T19}
+    T13, T14, T15, T16, T17, T18, T19, T20, T21}
     LUFactorization::T1
     QRFactorization::T2
     DiagonalFactorization::T3
@@ -20,6 +20,8 @@ mutable struct DefaultLinearSolverInit{T1, T2, T3, T4, T5, T6, T7, T8, T9, T10,
     AppleAccelerateLUFactorization::T17
     MKLLUFactorization::T18
     QRFactorizationPivoted::T19
+    KrylovJL_CRAIGMR::T20
+    KrylovJL_LSMR::T21
 end
 
 # Legacy fallback
@@ -254,11 +256,11 @@ function algchoice_to_alg(alg::Symbol)
     elseif alg === :AppleAccelerateLUFactorization
         AppleAccelerateLUFactorization()
     elseif alg === :QRFactorizationPivoted
-        @static if VERSION ≥ v"1.7beta"
-            QRFactorization(ColumnNorm())
-        else
-            QRFactorization(Val(true))
-        end
+        QRFactorization(ColumnNorm())
+    elseif alg === :KrylovJL_CRAIGMR
+        KrylovJL_CRAIGMR()
+    elseif alg === :KrylovJL_LSMR
+        KrylovJL_LSMR()
     else
         error("Algorithm choice symbol $alg not allowed in the default")
     end
@@ -387,7 +389,7 @@ end
             quote
                 getproperty(cache.cacheval,$(Meta.quot(alg)))' \ dy
             end
-        elseif alg in Symbol.((DefaultAlgorithmChoice.KrylovJL_GMRES,))
+        elseif alg in Symbol.((DefaultAlgorithmChoice.KrylovJL_GMRES,DefaultAlgorithmChoice.KrylovJL_LSMR, DefaultAlgorithmChoice.KrylovJL_CRAIGMR))
             quote
                 invprob = LinearSolve.LinearProblem(transpose(cache.A), dy)
                 solve(invprob, cache.alg;

src/iterative_wrappers.jl

@@ -243,7 +243,21 @@ function SciMLBase.solve!(cache::LinearCache, alg::KrylovJL; kwargs...)
     itmax = cache.maxiters
     verbose = cache.verbose ? 1 : 0
 
-    args = (@get_cacheval(cache, :KrylovJL_GMRES), cache.A, cache.b)
+    cacheval = if cache.alg isa DefaultLinearSolver
+        if alg.KrylovAlg === Krylov.gmres!
+            @get_cacheval(cache, :KrylovJL_GMRES)
+        elseif alg.KrylovAlg === Krylov.craigmr!
+            @get_cacheval(cache, :KrylovJL_CRAIGMR)
+        elseif alg.KrylovAlg === Krylov.lsmr!
+            @get_cacheval(cache, :KrylovJL_LSMR)
+        else
+            error("Default linear solver can only be these three choices! Report this bug!")
+        end
+    else
+        cache.cacheval
+    end
+
+    args = (cacheval, cache.A, cache.b)
     kwargs = (atol = atol, rtol = rtol, itmax = itmax, verbose = verbose,
         ldiv = true, history = true, alg.kwargs...)
 
@@ -268,7 +282,7 @@ function SciMLBase.solve!(cache::LinearCache, alg::KrylovJL; kwargs...)
     end
 
     stats = @get_cacheval(cache, :KrylovJL_GMRES).stats
-    resid = stats.residuals |> last
+    resid = !isempty(stats.residuals) ? last(stats.residuals) : zero(eltype(stats.residuals))
 
     retcode = if !stats.solved
         if stats.status == "maximum number of iterations exceeded"

test/default_algs.jl

@@ -67,3 +67,42 @@ prob = LinearProblem(sparse(A), b)
 
 prob = LinearProblem(big.(rand(10, 10)), big.(zeros(10)))
 solve(prob)
+
+## Operator defaults
+## https://github.com/SciML/LinearSolve.jl/issues/414
+
+m, n = 2, 2
+A = rand(m, n)
+b = rand(m)
+x = rand(n)
+f = (du, u, p, t) -> mul!(du, A, u)
+fadj = (du, u, p, t) -> mul!(du, A', u)
+fo = FunctionOperator(f, x, b; op_adjoint = fadj)
+prob = LinearProblem(fo, b)
+sol1 = solve(prob)
+sol2 = solve(prob, LinearSolve.KrylovJL_GMRES())
+@test sol1.u == sol2.u
+
+m, n = 3, 2
+A = rand(m, n)
+b = rand(m)
+x = rand(n)
+f = (du, u, p, t) -> mul!(du, A, u)
+fadj = (du, u, p, t) -> mul!(du, A', u)
+fo = FunctionOperator(f, x, b; op_adjoint = fadj)
+prob = LinearProblem(fo, b)
+sol1 = solve(prob)
+sol2 = solve(prob, LinearSolve.KrylovJL_LSMR())
+@test sol1.u == sol2.u
+
+m, n = 2, 3
+A = rand(m, n)
+b = rand(m)
+x = rand(n)
+f = (du, u, p, t) -> mul!(du, A, u)
+fadj = (du, u, p, t) -> mul!(du, A', u)
+fo = FunctionOperator(f, x, b; op_adjoint = fadj)
+prob = LinearProblem(fo, b)
+sol1 = solve(prob)
+sol2 = solve(prob, LinearSolve.KrylovJL_CRAIGMR())
+@test sol1.u == sol2.u