`dpnp.allclose` issues an error with scalars

[abagusetty]

Checks with allclose should have worked even with scalars using dpnp. Works for cupy and numpy

Error:

Traceback (most recent call last):
  File "/home/abagusetty/gpu4pyscf-testing/test_dpnp_allclose.py", line 9, in <module>
    dp.allclose(a, b)
  File "/home/abagusetty/gpu4pyscf-testing/dpnp/dpnp/dpnp_iface_logic.py", line 259, in allclose
    return all(isclose(a, b, rtol=rtol, atol=atol, equal_nan=equal_nan))
  File "/home/abagusetty/gpu4pyscf-testing/dpnp/dpnp/dpnp_iface_logic.py", line 868, in isclose
    dpnp.check_supported_arrays_type(a, b, scalar_type=True)
  File "/home/abagusetty/gpu4pyscf-testing/dpnp/dpnp/dpnp_iface.py", line 375, in check_supported_arrays_type
    raise TypeError(
TypeError: At least one input must be of supported array type, but got all scalars.

Reproducer:

import numpy as np
import dpnp as dp

a = np.float64(1.0)
b = np.float64(1.0)

np.allclose(a,b)
dp.allclose(a,b)

[antonwolfy]

@abagusetty this is intended behavior. We would like to avoid implicit memory allocation on default device, since it might lead to unexpected error in the user code, when, for instance, all handling data was explicitly allocated on some device but dp.allclose with scalar input produces the result allocated on different device and then it might cause unexpected ExecutionPlacementError exception:

ones = dpnp.ones(10, device='cpu')

a = np.float64(1.0)
b = np.float64(1.0)

# let say the default device is L0 GPU and then res in would be allocated on it
res = dpnp.allclose(a, b)

# that would produce ExecutionPlacementError exception
ones + res

Only array constructor functions, like asarray, ones, zeros and similar ones can be called without device keyword, which means the result will be allocated on the default device. In other cases dpnp assumes to return a result which follows the same placement as an input array. Which can't be achieved when inputs don't contain any dpnp or dpctl array.

[abagusetty]

@antonwolfy Thanks for pointing the info. This helps!