BUG: StridedMemoryView fails to wrap NumPy array

[carterbox]

I'm using numpy 2.0.2 and cuda-core 0.1.1. I cannot successfully convert a numpy array into a StridedMemoryView. My understanding is that this should be possible because numpy arrays implement the dlpack interface.

import nvmath  # used to patch cupy imports of nvidia-wheels
import numpy as np
import cuda.core.experimental as ccx
import cupy as cp
import torch

host_array = np.ones(50)
device_array = cp.ones(50)

device = ccx.Device()
device.set_current()

torch_array = torch.ones(50)

stream = device.create_stream()

# StridedMemoryView cannot wrap NumPy arrays because of impossible constraints on the
# stream_ptr parameter.

# Try 1: Provide a valid stream pointer
print(host_array)
wrapped_host_array = ccx.utils.StridedMemoryView(host_array, stream_ptr=stream.handle)
# RuntimeError: NumPy only supports stream=None.
print(wrapped_host_array)

# Try 2: Provide no stream pointer
print(host_array)
wrapped_host_array = ccx.utils.StridedMemoryView(host_array, stream_ptr=None)
# BufferError: stream=None is ambiguous with view()
print(wrapped_host_array)

print(torch_array)
wrapped_torch_array = ccx.utils.StridedMemoryView(torch_array, stream_ptr=None)
# BufferError: stream=None is ambiguous with view()
print(wrapped_torch_array)

# The following work

print(torch_array)
wrapped_torch_array = ccx.utils.StridedMemoryView(torch_array, stream_ptr=stream.handle)
print(wrapped_torch_array)

torch_array = torch_array.to('cuda')

print(torch_array)
wrapped_torch_array = ccx.utils.StridedMemoryView(torch_array, stream_ptr=stream.handle)
print(wrapped_torch_array)

print(device_array)
wrapped_device_array = ccx.utils.StridedMemoryView(device_array, stream_ptr=stream.handle)
print(wrapped_device_array)

[leofang]

Could you try passing stream_ptr=-1?

cuda-python/cuda_core/tests/test_utils.py

Lines 50 to 53 in 5d86ebb

	def test_strided_memory_view_cpu(self, in_arr):
	# stream_ptr=-1 means "the consumer does not care"
	view = StridedMemoryView(in_arr, stream_ptr=-1)
	self._check_view(view, in_arr)

It is a documented deviation from the __dlpack__ interface:

cuda-python/cuda_core/cuda/core/experimental/_memoryview.pyx

Lines 40 to 45 in 5d86ebb

	To opt-out of the stream ordering operation in either DLPack or CAI,
	please pass ``stream_ptr=-1``. Note that this deviates (on purpose)
	from the semantics of ``obj.__dlpack__(stream=None, ...)`` since ``cuda.core``
	does not encourage using the (legacy) default/null stream, but is
	consistent with the CAI's semantics. For DLPack, ``stream=-1`` will be
	internally passed to ``obj.__dlpack__()`` instead.

[carterbox]

That's right. I dunno where I was getting "RuntimeError: NumPy only supports stream=None.", but apparently it wasn't from CUDA core.