Fix strides for 0-sized empty arrays

[alexfikl]

For 0-sized arrays, this returned

>> c_contiguous_strides(8, (18, 0))
(0, 8)

which messes up cl.array.empty(queue, (18, 0), np.float64).strides. On the other hand, numpy gives

>>> np.empty((18, 0)).strides
(8, 8)

Not sure this is the right general fix, but it does the trick for this little example.

xref: inducer/arraycontext#91.

[inducer]

Thanks! I think this is fine, but I'd like better comments! :) And code duplication is bad (sorry!).

[inducer]

LGTM!

[inducer]

Can you pull this into the pyopencl change (inducer/pyopencl#514)? I'll pull it into pycuda and loopy.

[inducer]

• Update compyte for strides on zero-sized arrays loopy#496
• https://gitlab.tiker.net/inducer/pycuda/-/merge_requests/57

[seberg]

Super random, but NumPy indeed "ignores" size 1 dimensions when computing strides. However, NumPy will also when deciding whether strides are f or c-contiguous. So internally, numpy would consider the strides you had as valid f-contiguous strides (and will only sanitize this when exporting using the buffer protocol, which was necessary for cython at some point).

EDIT: To be clear, checking for contiguity, numpy completely ignores the stride value, and any stride value if the array is empty.

In that sense is_f_contiguous_strides is not the same as arr.flags.f_contiguous (it is not possible to do this by comparing strides only).

[inducer]

Thanks for the heads-up! I've recorded this in #37.