BUG,ENH: DLPack will refuse to export certain strided arrays that could be exported

[seberg]

The following array cannot be exported by DLPack currently, but would be valid. This is very obscure though:

# Not contiguous array (and not size 1) which has strides that along a dimension of size 1, so that they do not matter:
arr = np.ndarray(dtype="f8", strides=(4, 8), shape=(1, 10), buffer=np.ones(1000, dtype="uint8"))
arr.__dlpack__()  # could be exported, but fails.

I expected it to be easier to achieve, but arr.view only works with C-contiguous arrays.

EDIT: Honestly, if someone ever is tempted to just close this, fine with me (although fixing is probably also easy).

[tirthasheshpatel]

Your example works for me. The strides % itemsize check is only done for non-c-contiguous arrays. So, for example, this fails:

>>> x = np.ndarray(dtype='f8', shape=(10,5,1), strides=(8,80,4), buffer=np.ones(1000, dtype=np.uint8), order='F')
>>> y = np._from_dlpack(x)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: DLPack only supports strides which are a multiple of itemsize.

But it'd be nice to be able to export arrays irrespective of strides if they only have one element in some dimension.