dpctl usm_ndarray with `copy=false`

[abagusetty]

I came across a case where the input array is C-contiguous but was trying to create a output array with order=F & copy=False. There are some differences between cupy & dpnp/dpctl. I am not sure where (i.e., dpnp or dpctl) to post this issue.

I am not sure what is the guidance on this. The implementation in dpctl/tensor seems to be doing the right thing but dont understand the cupy's behavior here. Attached has some backtrace info and was wondering if someone has any pointers on this.

import dpnp as cp

# Create a C-contiguous array
a = cp.arange(12).reshape(3, 4)  # shape (3,4) — row-major layout

print("a.flags.c_contiguous =", a.flags.c_contiguous)  # True
print("a.flags.f_contiguous =", a.flags.f_contiguous)  # False

# Now try to call cupy.array with order='F', copy=False
b = cp.array(a, order='F', copy=False)

print("b.flags.c_contiguous =", b.flags.c_contiguous)
print("b.flags.f_contiguous =", b.flags.f_contiguous)
print("Did copy occur? a is b:", a is b)

With cupy this is the output:

a.flags.c_contiguous = True
a.flags.f_contiguous = False
b.flags.c_contiguous = False
b.flags.f_contiguous = True
Did copy occur? a is b: False

With dpnp/dpctl:

a.flags.c_contiguous = True
a.flags.f_contiguous = False
Traceback (most recent call last):
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/test_cupy_array_copyfalse.py", line 10, in <module>
    b = cp.array(a, order='F', copy=False)
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpnp/dpnp/dpnp_iface_arraycreation.py", line 429, in array
    result = dpnp_container.asarray(
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpnp/dpnp/dpnp_container.py", line 121, in asarray
    array_obj = dpt.asarray(
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpctl/dpctl/tensor/_ctors.py", line 573, in asarray
    return _asarray_from_usm_ndarray(
  File "/lus/flare/projects/NWChemEx_aesp_CNDA/abagusetty/gpu4pyscf/dpctl/dpctl/tensor/_ctors.py", line 159, in _asarray_from_usm_ndarray
    raise ValueError("asarray(..., copy=False) is not possible")
ValueError: asarray(..., copy=False) is not possible

[ndgrigorian]

Hi @abagusetty ,

sorry for the delay in this issue, it got a bit lost in the shuffle. I've tested it out and I can confirm that indeed, CuPy provides a new array (and new memory pointer) in this case, meaning copy keyword is not honored.

I've also tested with NumPy: it seems the copy keyword was ignored in cases where a copy was deemed necessary prior to NumPy 2, I would guess CuPy was simply imitating that behavior.

[abagusetty]

@ndgrigorian No worries. Thanks for taking time. What is your inclination for dpnp/dpctl behavior to be. Should it also align with numpy & cupy

[ndgrigorian]

@ndgrigorian No worries. Thanks for taking time. What is your inclination for dpnp/dpctl behavior to be. Should it also align with numpy & cupy

I think dpnp/dpctl should do what NumPy 2+ is doing, which is throw an error for this case, which is currently true

I'd recommend filing an issue with CuPy pointing out the change in NumPy 2

[abagusetty]

Closing this since dpctl/dpnp behavior is as expected