`wait` some times fails

When you call wait on both CPU and GPU. The synchronization can be unstable. The error message is

any: Test Failed at /home/leo/jcode/lab/wait_fail.jl:34
  Expression: Array(c1) ≈ ones(100)
   Evaluated: Float32[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0  …  0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] ≈ [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0  …  1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
Stacktrace:
 [1] top-level scope at /home/leo/jcode/lab/wait_fail.jl:34
 [2] top-level scope at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.4/Test/src/Test.jl:1113
 [3] top-level scope at /home/leo/jcode/lab/wait_fail.jl:31
Test Summary: | Pass  Fail  Total
any           |    1     1      2

MWE: you need to try the test for about 10 times, or more

using CuArrays, KernelAbstractions
using Test

@kernel function ba_kernel(x)
    i = @index(Global)
    x[i] += 1
end

function f1()
    nthread=4
    a = zeros(100)
    event1 = ba_kernel(CPU(), nthread)(a; ndrange=100)
    wait(event1)
    a
end

function f2()
    blockdim=256
    x = CuArrays.zeros(100)
    event1 = ba_kernel(CUDA(), blockdim)(x; ndrange=100)
    wait(event1)
    x
end

@testset "fail sometime" begin
    r1 = f1()
    c1 = f2()
    @test r1 ≈ ones(100)
    @test Array(c1) ≈ ones(100)
end

@testset "always work" begin
    c1 = f2()
    @test Array(c1) ≈ ones(100)
end

Start by running the supplied Julia MWE repeatedly, focusing on the CPU and GPU calls to ba_kernel and their wait(event1) usage. Compare the intermittent results from the “fail sometime” testset with the consistently passing GPU-only testset. Done means synchronization is stable and both result arrays match ones(100) across repeated runs.