test(assemblers): add incomplete Cholesky preconditioner tests

This commit is contained in:
Jukka Aho
2026-05-09 18:37:30 +03:00
parent a9d643bc28
commit 2a3afadd39
@@ -0,0 +1,361 @@
# This file is a part of JuliaFEM.
# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
"""
`ICholPreconditioner` (IC(0)) tests.
Locks in the contract of the no-fill incomplete-Cholesky preconditioner
added in B+++:
1. **Algebraic correctness** on small dense / tridiagonal SPD matrices:
`L * L' ≈ K` exactly when the IC(0) sparsity pattern is full
(dense lower triangle, tridiagonals, etc.); when fill is dropped,
it merely approximates `K`.
2. **Constructor robustness**: the diagonal-shift retry path produces
a valid factor on a near-indefinite matrix without corrupting the
input `K`.
3. **`ldiv!(y, P, x)` is the exact inverse of `L L'`** — i.e. the
two triangular solves are applied correctly.
4. **PCG with IC(0) converges in ≤ scalar-Jacobi iterations** on
stiffer elasticity problems (the canonical benchmark for IC(0)).
5. **End-to-end matrix-free PenaltyDirichlet pipeline** through the
`(cache, asm, kernel, mesh; dirichlet)` factory: IC(0) of the
constrained `K` plus `matrix_free_op` reproduces the assembled
direct solve.
6. **Zero allocations on `ldiv!`** (the hot path).
"""
using Test
using JuliaFEM
using JuliaFEM: ContinuumFormulation, FullThreeD, Vertex
using JuliaFEM: @DOFSet, DOF
using JuliaFEM: LinearElastic, Displacement, ContinuumKernel
using JuliaFEM: DOFBasedCOOAssembler, DOFBasedCOOCache
using JuliaFEM: extract_system, apply_K!
using JuliaFEM: PenaltyDirichlet, EliminatedDirichlet, apply_constraint!
using JuliaFEM: matrix_free_op
using JuliaFEM: JacobiPreconditioner, ICholPreconditioner
using JuliaFEM: UniformBodyForce, apply_load!
using JuliaFEM: create_elements!
using LinearAlgebra
using SparseArrays
using Tensors
using Random
# ---------------------------------------------------------------------------
# Mesh + setup helpers (same shape as test_block_jacobi.jl)
# ---------------------------------------------------------------------------
function _hex8_box(nx::Int, ny::Int, nz::Int;
Lx::Float64 = 1.0, Ly::Float64 = 1.0, Lz::Float64 = 1.0)
nodes = Vec{3,Float64}[]
nidx(i, j, k) = (i - 1) + (j - 1) * (nx + 1) + (k - 1) * (nx + 1) * (ny + 1) + 1
for k in 1:(nz + 1), j in 1:(ny + 1), i in 1:(nx + 1)
push!(nodes, Vec{3}((Lx * Float64(i - 1) / nx,
Ly * Float64(j - 1) / ny,
Lz * Float64(k - 1) / nz)))
end
conns = NTuple{8,UInt32}[]
for k in 1:nz, j in 1:ny, i in 1:nx
n1 = nidx(i, j, k)
n2 = nidx(i + 1, j, k)
n3 = nidx(i + 1, j + 1, k)
n4 = nidx(i, j + 1, k)
n5 = nidx(i, j, k + 1)
n6 = nidx(i + 1, j, k + 1)
n7 = nidx(i + 1, j + 1, k + 1)
n8 = nidx(i, j + 1, k + 1)
push!(conns, (UInt32(n1), UInt32(n2), UInt32(n3), UInt32(n4),
UInt32(n5), UInt32(n6), UInt32(n7), UInt32(n8)))
end
return Mesh{8,Hexahedron{8}}(nodes, conns)
end
function _setup_elasticity(mesh)
material = LinearElastic(E = 210e9, ν = 0.3)
kernel = ContinuumKernel(ContinuumFormulation{FullThreeD}(),
material, Displacement{3}())
S = @DOFSet{u::DOF{Displacement{3}, Vertex}}
elements, dof_mgr = create_elements!(mesh, Element{Hexahedron{8}, Lagrange{1}, S})
asm = DOFBasedCOOAssembler()
cache = DOFBasedCOOCache(elements, dof_mgr, mesh, kernel)
return cache, asm, kernel, mesh
end
# ---------------------------------------------------------------------------
# 1. Algebraic correctness: L * L' ≈ K on representative SPD matrices
# ---------------------------------------------------------------------------
@testset "ICholPreconditioner: L * L' ≈ K (algebraic correctness)" begin
println("\n" * "=" ^ 70)
println("IC(0) — algebraic correctness")
println("=" ^ 70)
@testset "1D Laplacian (tridiagonal — IC(0) is exact)" begin
for n in (5, 10, 50)
K = spdiagm(-1 => -ones(n - 1),
0 => 2.0 * ones(n),
1 => -ones(n - 1))
P = ICholPreconditioner(K)
Kr = Matrix(P.L) * Matrix(P.L)'
err = norm(Kr - Matrix(K)) / norm(Matrix(K))
@test err < 1e-12
println(" n=$n Laplacian ‖L Lᵀ - K‖/‖K‖ = $(round(err; sigdigits=3))")
end
end
@testset "Random dense SPD (full lower-triangle pattern)" begin
Random.seed!(20260508)
for n in (10, 20, 50)
M = randn(n, n)
K = sparse(M * M' + 5.0 * LinearAlgebra.I)
P = ICholPreconditioner(K)
Kr = Matrix(P.L) * Matrix(P.L)'
err = norm(Kr - Matrix(K)) / norm(Matrix(K))
@test err < 1e-10
println(" n=$n dense SPD ‖L Lᵀ - K‖/‖K‖ = $(round(err; sigdigits=3))")
end
end
@testset "Sparse SPD with dropped fill (IC(0) is approximate)" begin
Random.seed!(20260508)
n = 60
# Banded SPD with bandwidth ~5; off-band entries inside the
# band trigger fill in exact Cholesky → IC(0) drops it.
K = spdiagm(0 => 4.0 * ones(n))
for d in 1:5
v = -randn(n - d) ./ d
K += spdiagm(d => v, -d => v)
end
# Make safely SPD.
K += (2.0 * abs(minimum(eigvals(Matrix(K))))) * sparse(LinearAlgebra.I, n, n)
@test issymmetric(K)
@test minimum(eigvals(Matrix(K))) > 0
P = ICholPreconditioner(K)
Kr = Matrix(P.L) * Matrix(P.L)'
err = norm(Kr - Matrix(K)) / norm(Matrix(K))
# Approximate, so just bound modestly.
@test err < 5e-1
# Sparsity preserved.
@test nnz(P.L) == nnz(LowerTriangular(K))
println(" banded SPD n=$n err=$(round(err; sigdigits=3)) " *
"nnz(L)=$(nnz(P.L)) (= nnz(tril(K)))")
end
end
# ---------------------------------------------------------------------------
# 2. Constructor robustness: diagonal-shift retry, non-aliasing of input
# ---------------------------------------------------------------------------
@testset "ICholPreconditioner: diagonal-shift retry + input preservation" begin
println("\n" * "=" ^ 70)
println("IC(0) — diagonal-shift retry + non-aliasing")
println("=" ^ 70)
Random.seed!(20260508)
n = 30
# Construct a matrix that is SPD but with a near-zero eigenvalue
# — IC(0) on its lower triangle may produce a non-positive
# diagonal mid-factorisation. The shift retry should rescue it.
M = randn(n, n)
K0 = M * M'
# Slightly perturb diagonal downwards to provoke breakdown.
K0 -= (0.95 * minimum(eigvals(K0))) * Matrix(LinearAlgebra.I, n, n)
K = sparse(K0)
Korig = copy(K) # snapshot for non-aliasing test
P = ICholPreconditioner(K) # must NOT throw
# Input preserved (no aliasing leaked through):
@test K.nzval == Korig.nzval
@test K.rowval == Korig.rowval
@test K.colptr == Korig.colptr
# Factor is at least defined on every diagonal (no NaN / Inf).
L = P.L
@test all(isfinite, nonzeros(L))
diagL = [L[i, i] for i in 1:n]
@test all(>(0), diagL)
println(" n=$n retried IC(0) succeeded min(diag(L))=" *
"$(round(minimum(diagL); sigdigits=3))")
end
# ---------------------------------------------------------------------------
# 3. ldiv!(y, P, x) is the exact (L * L')^{-1} action
# ---------------------------------------------------------------------------
@testset "ICholPreconditioner: ldiv! agrees with (L * L')^{-1} * x" begin
Random.seed!(20260508)
n = 25
M = randn(n, n)
K = sparse(M * M' + 2.0 * LinearAlgebra.I)
P = ICholPreconditioner(K)
L = Matrix(P.L)
for trial in 1:3
x = randn(n)
y = zeros(n)
ldiv!(y, P, x)
ref = (L * L') \ x
rel = norm(y - ref) / max(norm(ref), 1.0)
@test rel < 1e-9
end
end
# ---------------------------------------------------------------------------
# 4. PCG with IC(0) ≤ scalar-Jacobi iterations on a stiff elasticity
# cantilever (the canonical benchmark for IC(0)).
# ---------------------------------------------------------------------------
@testset "IC(0): ≤ scalar-Jacobi iterations on elasticity cantilever" begin
using IterativeSolvers: cg!
using LinearOperators: LinearOperator
println("\n" * "=" ^ 70)
println("IC(0) vs SCALAR JACOBI — CG iteration count")
println("=" ^ 70)
nx, ny, nz = 12, 3, 3
mesh = _hex8_box(nx, ny, nz; Lx = 3.0, Ly = 0.5, Lz = 0.5)
cache, asm, kernel, m = _setup_elasticity(mesh)
n = cache.ndofs
nodes = m.nodes
tol = 1e-9
# Cantilever fixed at x=0, gravity body force.
fixed_dofs = Int[]; fixed_vals = Float64[]
for i in 1:length(nodes)
base = 3 * (i - 1)
if nodes[i][1] < tol
for α in 1:3
push!(fixed_dofs, base + α); push!(fixed_vals, 0.0)
end
end
end
bc = EliminatedDirichlet(fixed_dofs, fixed_vals)
b = Vec{3,Float64}((0.0, 0.0, -7850.0 * 9.81))
rhs = zeros(n); apply_load!(rhs, UniformBodyForce(b), cache, asm, kernel, m)
apply_constraint!(rhs, bc)
op = matrix_free_op(cache, asm, kernel, m; dirichlet = bc)
linop = LinearOperator(Float64, n, n, true, true, op)
# Scalar Jacobi
P_jac = JacobiPreconditioner(cache, asm, kernel, m; dirichlet = bc)
u_j = zeros(n)
h_j = cg!(u_j, linop, rhs; Pl = P_jac, abstol = 0.0, reltol = 1e-10,
maxiter = 4 * n, log = true)
# IC(0)
P_ic = ICholPreconditioner(cache, asm, kernel, m; dirichlet = bc)
u_i = zeros(n)
h_i = cg!(u_i, linop, rhs; Pl = P_ic, abstol = 0.0, reltol = 1e-10,
maxiter = 4 * n, log = true)
iters_j = h_j[2].iters
iters_i = h_i[2].iters
@test isapprox(u_j, u_i; atol = 1e-6, rtol = 1e-6)
@test iters_i <= iters_j # IC(0) ≤ scalar-Jacobi (often much smaller)
@test maximum(abs, u_i) > 1e-15
println(" cantilever nx=$nx ny=$ny nz=$nz ndof=$n " *
"scalar-Jac iters=$iters_j IC(0) iters=$iters_i")
end
# ---------------------------------------------------------------------------
# 5. End-to-end PenaltyDirichlet matrix-free PCG with IC(0) factory
# ---------------------------------------------------------------------------
@testset "IC(0): PenaltyDirichlet inhomogeneous CG matches direct solve" begin
using IterativeSolvers: cg!
using LinearOperators: LinearOperator
println("\n" * "=" ^ 70)
println("IC(0) — PenaltyDirichlet end-to-end CG")
println("=" ^ 70)
nx, ny, nz = 4, 1, 1
mesh = _hex8_box(nx, ny, nz; Lx = 1.0, Ly = 0.1, Lz = 0.1)
cache, asm, kernel, m = _setup_elasticity(mesh)
n = cache.ndofs
nodes = m.nodes
tol = 1e-9
# Same prescribed displacement as test_block_jacobi.jl (penalty form):
# u(x=0) = 0, u_x(x=1) = 0.01.
fixed_dofs = Int[]; fixed_vals = Float64[]
for i in 1:length(nodes)
x = nodes[i][1]
base = 3 * (i - 1)
if x < tol
for α in 1:3
push!(fixed_dofs, base + α); push!(fixed_vals, 0.0)
end
elseif x > 1.0 - tol
push!(fixed_dofs, base + 1); push!(fixed_vals, 0.01)
end
end
bc_pen = PenaltyDirichlet(fixed_dofs, fixed_vals; penalty = 1e10)
# Direct reference
assemble!(cache, asm, kernel, m)
K, _ = extract_system(cache)
Kbc = Matrix(K)
apply_constraint!(Kbc, bc_pen)
bbc = zeros(n); apply_constraint!(bbc, bc_pen)
u_dir = Kbc \ bbc
# Matrix-free CG with IC(0)
op = matrix_free_op(cache, asm, kernel, m; dirichlet = bc_pen)
linop = LinearOperator(Float64, n, n, true, true, op)
P = ICholPreconditioner(cache, asm, kernel, m; dirichlet = bc_pen)
u_mf = zeros(n)
h = cg!(u_mf, linop, bbc; Pl = P, abstol = 1e-12, reltol = 1e-12,
maxiter = 4 * n, log = true)
rel = norm(u_mf - u_dir) / max(norm(u_dir), 1.0)
@test rel < 1e-6
@test isapprox(u_mf[3 * nx + 1], 0.01; atol = 5e-3)
println(" Elast bar nx=$nx ndof=$n fixed=$(length(fixed_dofs)) " *
"iters=$(h[2].iters) rel(u_mf vs u_dir)=" *
"$(round(rel; sigdigits = 3)) u_x(x=1)=" *
"$(round(u_mf[3 * nx + 1]; sigdigits = 4)) (penalty offset expected)")
end
# ---------------------------------------------------------------------------
# 6. Zero allocations on the hot path (ldiv!)
# ---------------------------------------------------------------------------
@testset "IC(0): zero allocations on ldiv!" begin
println("\n" * "=" ^ 70)
println("IC(0) — ZERO-ALLOC on ldiv!")
println("=" ^ 70)
@testset "ldiv!(y, P, x) on $(nx)×$(ny)×$(nz)" for (nx, ny, nz) in
[(1, 1, 1), (2, 1, 1), (3, 2, 2)]
mesh = _hex8_box(nx, ny, nz)
cache, asm, kernel, m = _setup_elasticity(mesh)
n_dof = cache.ndofs
P = ICholPreconditioner(cache, asm, kernel, m)
x = randn(n_dof); y = zeros(n_dof)
ldiv!(y, P, x)
GC.gc()
a = @allocated ldiv!(y, P, x)
@test a == 0
println(" $(nx)×$(ny)×$(nz) ndof=$n_dof ldiv!=$a")
end
end