guvectorize in-place operations work only in some cases

The following works, which is good (numba 0.51.2, llvmlite 0.34.0):

guvectorized function writing to the output, on GPU

@guvectorize(["(f8[:], f8[:])"], "() -> ()", target='cuda')
def test(inp, out):
    out[0] = inp[0]

feeding in device arrays: ✔️

ones = np.ones(10, dtype=np.float64)
twos = np.ones_like(ones) * 2
ones_d = numba.cuda.to_device(ones)
twos_d = numba.cuda.to_device(twos)

test(ones_d, out=twos_d)
twos_d.copy_to_host()

>>> array([1., 1., 1., 1., 1., 1., 1., 1., 1., 1.])

feeding in host arrays: ✔️

ones = np.ones(10, dtype=np.float64)
twos = np.ones_like(ones) * 2

test(ones, out=twos)
twos

>>> array([1., 1., 1., 1., 1., 1., 1., 1., 1., 1.])

guvectorized function modifying the output, on GPU

@guvectorize(["(f8[:], f8[:])"], "() -> ()", target='cuda')
def inplace_test(inp, out):
    out[0] += inp[0]    

feeding in device arrays: ✔️

ones = np.ones(10, dtype=np.float64)
twos = np.ones_like(ones) * 2
ones_d = numba.cuda.to_device(ones)
twos_d = numba.cuda.to_device(twos)

inplace_test(ones_d, out=twos_d)
twos_d.copy_to_host()

>>> array([3., 3., 3., 3., 3., 3., 3., 3., 3., 3.])

But this does not work:

feeding in host arrays: ❌

ones = np.ones(10, dtype=np.float64)
twos = np.ones_like(ones) * 2

inplace_test(ones, out=twos)
twos

>>> array([2., 2., 2., 2., 2., 2., 2., 2., 2., 2.])

the output host array contains just something… 🤷‍♂️

[philippeller]

Still convinced that this is a bug. The outputs are not being assigned. If the outputs of a function cannot be relied upon, then in my universe this is considered a bug. What was references as the same issue by @gmarkall is talking about writing to the inputs.

[gmarkall]

Moving to 0.56 as I’ll not get the time to look into this for 0.55.