Don't default to MKL on EPYC

Project.toml

@@ -7,6 +7,7 @@ version = "2.30.2"
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 ConcreteStructs = "2569d6c7-a4a2-43d3-a901-331e8e4be471"
+CpuId = "adafc99b-e345-5852-983c-f28acb93d879"
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
 EnumX = "4e289a0a-7415-4d19-859d-a7e5c4648b56"
 FastLapackInterface = "29a986be-02c6-4525-aec4-84b980013641"
@@ -72,6 +73,7 @@ CUDA = "5"
 CUDSS = "0.1, 0.2, 0.3"
 ChainRulesCore = "1.22"
 ConcreteStructs = "0.2.3"
+CpuId = "0.3.1"
 DocStringExtensions = "0.9.3"
 EnumX = "1.0.4"
 Enzyme = "0.11.15, 0.12"

src/LinearSolve.jl

@@ -41,10 +41,11 @@ import PrecompileTools
  import Krylov
  using SciMLBase
  import Preferences
+ import CpuId
 
 const CRC = ChainRulesCore
 
-if Preferences.@load_preference("LoadMKL_JLL", true)
+if Preferences.@load_preference("LoadMKL_JLL", !occursin("EPYC", CpuId.cpubrand()))
     using MKL_jll
     const usemkl = MKL_jll.is_available()
 else

test/adjoint.jl

@@ -62,7 +62,7 @@ db12 = FiniteDiff.finite_difference_gradient(
 db22 = FiniteDiff.finite_difference_gradient(
     x -> f3(eltype(x).(A), eltype(x).(b1), x), copy(b1))
 
-@test dA≈dA2 atol=1e-4
+@test dA≈dA2 rtol=1e-3
 @test db1 ≈ db12
 @test db2 ≈ db22
 

test/qa.jl

@@ -6,7 +6,7 @@ using LinearSolve, Aqua
     Aqua.test_piracies(LinearSolve,
         treat_as_own = [LinearProblem])
     Aqua.test_project_extras(LinearSolve)
-    Aqua.test_stale_deps(LinearSolve)
+    Aqua.test_stale_deps(LinearSolve, ignore = [:MKL_jll])
     Aqua.test_unbound_args(LinearSolve)
     Aqua.test_undefined_exports(LinearSolve)
 end

test/static_arrays.jl

@@ -17,13 +17,11 @@ for alg in (nothing, LUFactorization(), SVDFactorization(), CholeskyFactorizatio
     sol = solve(LinearProblem(A, b), alg)
     @inferred solve(LinearProblem(A, b), alg)
     @test norm(A * sol .- b) < 1e-10
-
-    if __non_native_static_array_alg(alg)
-        @test_broken __solve_no_alloc(A, b, alg)
+    if alg isa KrylovJL{typeof(LinearSolve.Krylov.gmres!)}
+        @test_broken __solve_no_alloc(A, b, alg) isa SciMLBase.LinearSolution
     else
-        @test_nowarn __solve_no_alloc(A, b, alg)
+        @test_nowarn __solve_no_alloc(A, b, alg) isa SciMLBase.LinearSolution
     end
-
     cache = init(LinearProblem(A, b), alg)
     sol = solve!(cache)
     @test norm(A * sol .- b) < 1e-10