Possible race condition problem when using eigen with @threads in windows?

[atbug]

using LinearAlgebra

function test(H)
    E = zeros(N, Threads.nthreads())
    Threads.@threads for ik = 1:484
        E[:, Threads.threadid()] += eigen(H).values
    end
    Esum=zeros(N)
    for i=1:Threads.nthreads()
        Esum += E[:,i]
    end
    return Esum
end

N = 4
H = rand(N, N)+im*rand(N, N)
H = H+H'

reference = test(H)
for i in 1:20
    if maximum(abs.(test(H)-reference))>1.0e-7
        println("race?")
    end
end

Expected no output. However, with N=2 we get a small chance(~10%) to get the "race?" message. With N=4 almost every execution of test() returns different values.

Note that the problem only occurs on Windows. On linux no output would show up. Here's the test environment:

julia> versioninfo()
Julia Version 1.0.3
Commit 099e826241 (2018-12-18 01:34 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i5-7200U CPU @ 2.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.0 (ORCJIT, skylake)
Environment:
  JULIA_NUM_THREADS = 4

[andreasnoack]

This seems to be a problem with the thread safety of OpenBLAS's xsymv. I could reduce the problem to

function test2(H, f)
    nn = 484
    E = Vector{Any}(undef, nn)
    Threads.@threads for ik = 1:nn
        E[ik] = f(H)
    end
    return E
end

julia> sum(t -> abs.(t), diff(test2(H, t -> BLAS.symv!('U', 1.0, copy(t), ones(4), 0.0, zeros(4)))))
4-element Array{Float64,1}:
  0.0
  0.04083050218188866
  2.7073001729582646
 17.051752655369707

I'll report it upstream.

[andreasnoack]

Actually, it looks like the OpenBLAS developers have been working a lot on thread safety lately and indeed this issue fixed with the version of OpenBLAS we use on nightly.