Throw an error in checktask when a task failed

[efaulhaber]

Solves JuliaSIMD/Polyester.jl#153.

julia> @batch for j in axes(x, 2)
                  for i in 1:5
                      if i > 3 && j in (2, 4, 9)
                          error("some error")
                      end
                      x[i, j] = 1
                  end
              end
ERROR: TaskFailedException
Stacktrace:
 [1] checktask(tid::UInt32)
   @ ThreadingUtilities ~/.julia/dev/ThreadingUtilities/src/threadtasks.jl:74
 [2] wait
   @ ~/.julia/packages/ThreadingUtilities/3z3g0/src/threadtasks.jl:88 [inlined]
 [3] wait
   @ ~/.julia/packages/ThreadingUtilities/3z3g0/src/threadtasks.jl:81 [inlined]
 [4] macro expansion
   @ ~/.julia/packages/Polyester/eqrC9/src/batch.jl:257 [inlined]
 [5] _batch_no_reserve
   @ ~/.julia/packages/Polyester/eqrC9/src/batch.jl:168 [inlined]
 [6] batch(::var"#56#57", ::Val{…}, ::Tuple{}, ::Tuple{}, ::Tuple{…}, ::Static.StaticInt{…}, ::Static.StaticInt{…}, ::Polyester.NoLoop, ::Polyester.CombineIndices, ::Matrix{…})
   @ Polyester ~/.julia/packages/Polyester/eqrC9/src/batch.jl:334
 [7] top-level scope
   @ ~/.julia/packages/Polyester/eqrC9/src/closure.jl:456

    nested task error: some error
    Stacktrace:
     [1] error(s::String)
       @ Base ./error.jl:35
     [2] macro expansion
       @ ./REPL[25]:4 [inlined]
     [3] #56
       @ ~/.julia/packages/Polyester/eqrC9/src/closure.jl:309 [inlined]
     [4] (::Polyester.BatchClosure{var"#56#57", ManualMemory.Reference{Tuple{…}}, false, Tuple{}})(p::Ptr{UInt64})
       @ Polyester ~/.julia/packages/Polyester/eqrC9/src/batch.jl:11
     [5] _call
       @ ~/.julia/packages/ThreadingUtilities/3z3g0/src/threadtasks.jl:11 [inlined]
     [6] (::ThreadingUtilities.ThreadTask)()
       @ ThreadingUtilities ~/.julia/dev/ThreadingUtilities/src/threadtasks.jl:29
Some type information was truncated. Use `show(err)` to see complete types.

Note that JuliaSIMD/Polyester.jl#154 should probably get merged first.

[efaulhaber]

@oscardssmith ?

[oscardssmith]

this is a package in not especially familiar with, so I wanted to see green CI, but CI seems somewhat broken

[efaulhaber]

Note that this is what @chriselrod suggested in JuliaSIMD/Polyester.jl#153.

[efaulhaber]

So, how do we proceed here?

[efaulhaber]

Bump

[oscardssmith]

Sorry for this taking so long. CI for this repo was broken (fixed by #55)

[oscardssmith]

Looks like CI is fixed! (although it still is requiring manual approval for some reason)

[oscardssmith]

@efaulhaber looks like this PR is broken.

[efaulhaber]

I spent the last half hour trying to understand this test, but I still don't know what it does:

struct Copy{P} end
function (::Copy{P})(p::Ptr{UInt}) where {P}
  _, (ptry, ptrx, N) = ThreadingUtilities.load(p, P, 2 * sizeof(UInt))
  N > 0 || throw("This function throws if N == 0 for testing purposes.")
  @simd ivdep for n ∈ 1:N
    unsafe_store!(ptry, unsafe_load(ptrx, n), n)
  end
  ThreadingUtilities._atomic_store!(p, ThreadingUtilities.SPIN)
end
@generated function copy_ptr(::A, ::B) where {A,B}
  c = Copy{Tuple{A,B,Int}}()
  quote
    @cfunction($c, Cvoid, (Ptr{UInt},))
  end
end
function setup_copy!(p, y, x)
  N = length(y)
  @assert length(x) == N
  py = pointer(y)
  px = pointer(x)
  fptr = copy_ptr(py, px)
  offset = ThreadingUtilities.store!(p, fptr, sizeof(UInt))
  ThreadingUtilities.store!(p, (py, px, N), offset)
end

@inline launch_thread_copy!(tid, y, x) = ThreadingUtilities.launch(setup_copy!, tid, y, x)

function test_copy(tid, N = 100_000)
  a = rand(N)
  b = rand(N)
  c = rand(N)
  x = similar(a) .= NaN
  y = similar(b) .= NaN
  z = similar(c) .= NaN
  GC.@preserve a b c x y z begin
    launch_thread_copy!(tid, x, a)
    yield()
    @assert !ThreadingUtilities.wait(tid)
    launch_thread_copy!(tid, y, b)
    yield()
    @assert !ThreadingUtilities.wait(tid)
    launch_thread_copy!(tid, z, c)
    yield()
    @assert !ThreadingUtilities.wait(ThreadingUtilities.taskpointer(tid))
  end
  @test a == x
  @test b == y
  @test c == z
end

The yield() calls wait, which throws the error throw("This function throws if N == 0 for testing purposes.").
On main, this error is just ignored, which is why this test worked before.

[oscardssmith]

@chriselrod do you know what's should happen here?

[chriselrod]

ThreadingUtilities.jl/test/threadingutilities.jl

Lines 80 to 86 in e867205

	for tid ∈ eachindex(ThreadingUtilities.TASKS)
	launch_thread_copy!(tid, Float64[], Float64[])
	end
	sleep(1)
	@test all(istaskfailed, ThreadingUtilities.TASKS)
	@test all(ThreadingUtilities.wait, eachindex(ThreadingUtilities.TASKS))
	@test !any(istaskfailed, ThreadingUtilities.TASKS)

This is supposed to

1. Make all tasks error.
2. sleep
3. check that all tasks have errored
4. ThreadingUtilities.wait on them to reset them, which returns true to indicate that a reset had occured
5. test that none of them are in a failed state.

test_copy doesn't get called until after this, once all tests are in a non-failed state. It gets called with non-0 N, so it won't fail with an N==0 error.

This PR changes whose responsibility it is to do what. Before, wait is supposed to check for success/failure, and forward any errors.

[efaulhaber]

That makes sense. Thanks! I fixed the tests now and added a few comments.
I now test that wait throws an error for each thread.

[efaulhaber]

@oscardssmith Most tests are passing now. The nightly builds and docs failures seem to be unrelated.

[efaulhaber]

@oscardssmith would you mind also adding a new release?