Setup NonlinearSolveAlg with jacobian reuse

Author: ChrisRackauckas

lib/OrdinaryDiffEqCore/src/misc_utils.jl

@@ -133,6 +133,7 @@ function get_differential_vars(f, u)
 end
 
 isnewton(::Any) = false
+isnonlinearsolve(::Any) = false
 
 function _bool_to_ADType(::Val{true}, ::Val{CS}, _) where {CS}
     Base.depwarn(

lib/OrdinaryDiffEqDifferentiation/src/OrdinaryDiffEqDifferentiation.jl

@@ -35,7 +35,7 @@ using OrdinaryDiffEqCore: OrdinaryDiffEqAlgorithm, OrdinaryDiffEqAdaptiveImplici
                           OrdinaryDiffEqAdaptiveExponentialAlgorithm, @unpack,
                           AbstractNLSolver, nlsolve_f, issplit,
                           concrete_jac, unwrap_alg, OrdinaryDiffEqCache, _vec, standardtag,
-                          isnewton, _unwrap_val,
+                          isnewton, isnonlinearsolve, _unwrap_val,
                           set_new_W!, set_W_γdt!, alg_difftype, unwrap_cache, diffdir,
                           get_W, isfirstcall, isfirststage, isJcurrent,
                           get_new_W_γdt_cutoff,

lib/OrdinaryDiffEqDifferentiation/src/derivative_utils.jl

@@ -458,7 +458,7 @@ function do_newJW(integrator, alg, nlsolver, repeat_step)::NTuple{2, Bool}
         return true, true
     end
     # TODO: add `isJcurrent` support for Rosenbrock solvers
-    if !isnewton(nlsolver)
+    if !isnewton(nlsolver) && !isnonlinearsolve(nlsolver)
         isfreshJ = !(integrator.alg isa CompositeAlgorithm) &&
                    (integrator.iter > 1 && errorfail && !integrator.u_modified)
         return !isfreshJ, true

lib/OrdinaryDiffEqNonlinearSolve/src/OrdinaryDiffEqNonlinearSolve.jl

@@ -48,12 +48,12 @@ using OrdinaryDiffEqCore: resize_nlsolver!, _initialize_dae!,
 
 import OrdinaryDiffEqCore: _initialize_dae!, isnewton, get_W, isfirstcall, isfirststage,
                            isJcurrent, get_new_W_γdt_cutoff, resize_nlsolver!, apply_step!,
-                           postamble!
+                           postamble!, isnonlinearsolve
 
 import OrdinaryDiffEqDifferentiation: update_W!, is_always_new, build_uf, build_J_W,
                                       WOperator, StaticWOperator, wrapprecs,
                                       build_jac_config, dolinsolve, alg_autodiff,
-                                      resize_jac_config!
+                                      resize_jac_config!, do_newJW
 
 import StaticArrays: SArray, MVector, SVector, @SVector, StaticArray, MMatrix, SA,
                      StaticMatrix

lib/OrdinaryDiffEqNonlinearSolve/src/newton.jl

@@ -81,7 +81,14 @@ end
     @unpack tstep, invγdt = cache
 
     nlcache = nlsolver.cache.cache
-    step!(nlcache)
+
+    if is_always_new(nlsolver) || new_jac || new_W
+        recompute_jacobian = true
+    else
+        recompute_jacobian = false
+    end
+
+    step!(nlcache; recompute_jacobian)
     nlsolver.ztmp = nlcache.u
 
     ustep = compute_ustep(tmp, γ, z, method)
@@ -103,7 +110,15 @@ end
     @unpack tstep, invγdt, atmp, ustep = cache
 
     nlcache = nlsolver.cache.cache
-    step!(nlcache)
+    new_jac, new_W = do_newJW(integrator, integrator.alg, nlsolver, false)
+
+    if is_always_new(nlsolver) || new_jac || new_W
+        recompute_jacobian = true
+    else
+        recompute_jacobian = false
+    end
+
+    step!(nlcache; recompute_jacobian)
     @.. broadcast=false ztmp=nlcache.u
 
     ustep = compute_ustep!(ustep, tmp, γ, z, method)

lib/OrdinaryDiffEqNonlinearSolve/src/nlsolve.jl

@@ -65,7 +65,7 @@ function nlsolve!(nlsolver::NL, integrator::DiffEqBase.DEIntegrator,
         # check divergence (not in initial step)
         if iter > 1
             θ = prev_θ = has_prev_θ ? max(0.3 * prev_θ, ndz / ndzprev) : ndz / ndzprev
-
+            
             # When one Newton iteration basically does nothing, it's likely that we
             # are at the precision limit of floating point number. Thus, we just call
             # it convergence/divergence according to `ndz` directly.
@@ -105,7 +105,7 @@ function nlsolve!(nlsolver::NL, integrator::DiffEqBase.DEIntegrator,
         # don't trust θ for non-adaptive on first iter because the solver doesn't provide feedback
         # for us to know whether our previous nlsolve converged sufficiently well
         check_η_convergance = (iter > 1 ||
-                               (isnewton(nlsolver) && isadaptive(integrator.alg)))
+                               ((isnewton(nlsolver) || isnonlinearsolve(nlsolver)) && isadaptive(integrator.alg)))
         if (iter == 1 && ndz < 1e-5) ||
            (check_η_convergance && η >= zero(η) && η * ndz < κ)
             nlsolver.status = Convergence
@@ -114,7 +114,7 @@ function nlsolve!(nlsolver::NL, integrator::DiffEqBase.DEIntegrator,
         end
     end
 
-    if isnewton(nlsolver) && nlsolver.status == Divergence &&
+    if (isnewton(nlsolver) || isnonlinearsolve(nlsolver)) && nlsolver.status == Divergence &&
        !isJcurrent(nlsolver, integrator)
         nlsolver.status = TryAgain
         nlsolver.nfails += 1

lib/OrdinaryDiffEqNonlinearSolve/src/type.jl

@@ -218,4 +218,5 @@ mutable struct NonlinearSolveCache{uType, tType, rateType, tType2, P, C} <:
     invγdt::tType2
     prob::P
     cache::C
+    new_W::Bool
 end

lib/OrdinaryDiffEqNonlinearSolve/src/utils.jl

@@ -14,6 +14,10 @@ isnewton(nlsolver::AbstractNLSolver) = isnewton(nlsolver.cache)
 isnewton(::AbstractNLSolverCache) = false
 isnewton(::Union{NLNewtonCache, NLNewtonConstantCache}) = true
 
+isnonlinearsolve(nlsolver::AbstractNLSolver) = isnonlinearsolve(nlsolver.cache)
+isnonlinearsolve(::AbstractNLSolverCache) = false
+isnonlinearsolve(::NonlinearSolveCache) = true
+
 is_always_new(nlsolver::AbstractNLSolver) = is_always_new(nlsolver.alg)
 check_div(nlsolver::AbstractNLSolver) = check_div(nlsolver.alg)
 check_div(alg) = isdefined(alg, :check_div) ? alg.check_div : true
@@ -32,9 +36,9 @@ getnfails(_) = 0
 getnfails(nlsolver::AbstractNLSolver) = nlsolver.nfails
 
 set_new_W!(nlsolver::AbstractNLSolver, val::Bool)::Bool = set_new_W!(nlsolver.cache, val)
-set_new_W!(nlcache::Union{NLNewtonCache, NLNewtonConstantCache}, val::Bool)::Bool = nlcache.new_W = val
+set_new_W!(nlcache::Union{NLNewtonCache, NLNewtonConstantCache, NonlinearSolveCache}, val::Bool)::Bool = nlcache.new_W = val
 get_new_W!(nlsolver::AbstractNLSolver)::Bool = get_new_W!(nlsolver.cache)
-get_new_W!(nlcache::Union{NLNewtonCache, NLNewtonConstantCache})::Bool = nlcache.new_W
+get_new_W!(nlcache::Union{NLNewtonCache, NLNewtonConstantCache, NonlinearSolveCache})::Bool = nlcache.new_W
 get_new_W!(::AbstractNLSolverCache)::Bool = true
 
 get_W(nlsolver::AbstractNLSolver) = get_W(nlsolver.cache)
@@ -231,7 +235,7 @@ function build_nlsolver(
             end
             prob = NonlinearProblem(NonlinearFunction{true}(nlf), ztmp, nlp_params)
             cache = init(prob, nlalg.alg)
-            nlcache = NonlinearSolveCache(ustep, tstep, k, atmp, invγdt, prob, cache)
+            nlcache = NonlinearSolveCache(ustep, tstep, k, atmp, invγdt, prob, cache, true)
         else
             nlcache = NLNewtonCache(ustep, tstep, k, atmp, dz, J, W, true,
                 true, true, tType(dt), du1, uf, jac_config,
@@ -316,7 +320,7 @@ function build_nlsolver(
             prob = NonlinearProblem(NonlinearFunction{false}(nlf), copy(ztmp), nlp_params)
             cache = init(prob, nlalg.alg)
             nlcache = NonlinearSolveCache(
-                nothing, tstep, nothing, nothing, invγdt, prob, cache)
+                nothing, tstep, nothing, nothing, invγdt, prob, cache, true)
         else
             nlcache = NLNewtonConstantCache(tstep, J, W, true, true, true, tType(dt), uf,
                 invγdt, tType(nlalg.new_W_dt_cutoff), t)