Support hermitian types for KernelMatrix

src/kernelmatrix.jl

@@ -48,6 +48,12 @@ function KernelMatrix(f, X, Y)
     return KernelMatrix{typeof(f),typeof(X),typeof(Y),T}(f, X, Y)
 end
 
+# light wrapper for Hermitian
+const HermitianKernelMatrix = Hermitian{<:Any,<:KernelMatrix}
+
+rowelements(K::HermitianKernelMatrix) = K |> parent |> rowelements
+colelements(K::HermitianKernelMatrix) = K |> parent |> colelements
+
 """
     assemble_hmatrix(K::AbstractKernelMatrix[; atol, rank, rtol, kwargs...])
 
@@ -57,7 +63,8 @@ arguments are passed to the [`PartialACA`](@ref) constructor, and the remaining
 keyword arguments are forwarded to the main `assemble_hmatrix` function.
 """
 function assemble_hmatrix(
-    K::AbstractKernelMatrix;
+    K::Union{KernelMatrix,HermitianKernelMatrix};
+    # K::KernelMatrix;
     atol = 0,
     rank = typemax(Int),
     rtol = atol > 0 || rank < typemax(Int) ? 0 : sqrt(eps(Float64)),

src/multiplication.jl

@@ -221,26 +221,26 @@
 function _mul_dense!(C::Base.Matrix, A, B, a)
     Adata = isleaf(A) ? data(A) : A
     Bdata = isleaf(B) ? data(B) : B
-    if Adata isa HMatrix
-        if Bdata isa Matrix
+    if parent(Adata) isa HMatrix
+        if parent(Bdata) isa Matrix
             _mul131!(C, Adata, Bdata, a)
-        elseif Bdata isa RkMatrix
+        elseif parent(Bdata) isa RkMatrix
             _mul132!(C, Adata, Bdata, a)
         end
-    elseif Adata isa AdjOrMat
-        if Bdata isa Matrix
+    elseif parent(Adata) isa AdjOrMat
+        if parent(Bdata) isa Matrix
             _mul111!(C, Adata, Bdata, a)
-        elseif Bdata isa RkMatrix
+        elseif parent(Bdata) isa RkMatrix
             _mul112!(C, Adata, Bdata, a)
-        elseif Bdata isa HMatrix
+        elseif parent(Bdata) isa HMatrix
             _mul113!(C, Adata, Bdata, a)
         end
-    elseif Adata isa RkMatrix
-        if Bdata isa Matrix
+    elseif parent(Adata) isa RkMatrix
+        if parent(Bdata) isa Matrix
             _mul121!(C, Adata, Bdata, a)
-        elseif Bdata isa RkMatrix
+        elseif parent(Bdata) isa RkMatrix
             _mul122!(C, Adata, Bdata, a)
-        elseif Bdata isa HMatrix
+        elseif parent(Bdata) isa HMatrix
             _mul123!(C, Adata, Bdata, a)
         end
     else

test/cholesky_test.jl

@@ -10,26 +10,49 @@ include(joinpath(HMatrices.PROJECT_ROOT, "test", "testutils.jl"))
 
 Random.seed!(1)
 
-m = 5000
-T = Float64
-X = points_on_sphere(m)
-Y = X
+@testset "Assemble and solve" begin
+    m = 5000
+    T = Float64
+    X = points_on_sphere(m)
+    Y = X
 
-K = laplace_matrix(X, X)
+    K = laplace_matrix(X, X)
 
-splitter = CardinalitySplitter(; nmax = 50)
-Xclt = ClusterTree(X, splitter)
-Yclt = ClusterTree(Y, splitter)
-adm = StrongAdmissibilityStd(3)
-comp = PartialACA(; atol = 1e-10)
-for threads in (false, true)
-    H = assemble_hmatrix(Hermitian(K), Xclt, Yclt; adm, comp, threads, distributed = false)
-    hchol = cholesky(H; atol = 1e-10)
-    y = rand(m)
-    M = Matrix(K)
-    exact = M \ y
-    approx = hchol \ y
-    @test norm(exact - approx, Inf) < 1e-10
-    # test multiplication by checking if the solution is correct
-    @test hchol.L * (hchol.U * approx) ≈ y
+    splitter = CardinalitySplitter(; nmax = 50)
+    Xclt = ClusterTree(X, splitter)
+    Yclt = ClusterTree(Y, splitter)
+    adm = StrongAdmissibilityStd(3)
+    comp = PartialACA(; atol = 1e-10)
+    for threads in (false, true)
+        H = assemble_hmatrix(
+            Hermitian(K),
+            Xclt,
+            Yclt;
+            adm,
+            comp,
+            threads,
+            distributed = false,
+        )
+        hchol = cholesky(H; atol = 1e-10)
+        y = rand(m)
+        M = Matrix(K)
+        exact = M \ y
+        approx = hchol \ y
+        @test norm(exact - approx, Inf) < 1e-10
+        # test multiplication by checking if the solution is correct
+        @test hchol.L * (hchol.U * approx) ≈ y
+    end
+end
+
+@testset "Issue 80" begin
+    m        = 1000
+    pts      = rand(SVector{2,Float64}, m)
+    km       = KernelMatrix((x, y) -> exp(-norm(x - y)), pts, pts) |> Hermitian
+    hk       = assemble_hmatrix(km; atol = 1e-10)
+    hkf      = cholesky(hk; atol = 1e-10)
+    full     = Matrix(hk)
+    y        = rand(m)
+    x_exact  = full \ y
+    x_approx = hkf \ y
+    norm(x_exact - x_approx, Inf) / norm(x_exact, Inf) < 1e-8
 end