[FabianIsensee]

Hi Francesc,

A small one, found while benchmarking partial overwrites for Blosc/c-blosc2#847.

Problem. Assigning a NumPy value that is not C-contiguous raises (on main and in 4.14.1):

import blosc2, numpy as np
y = np.arange(20 * 30 * 40, dtype=np.float32).reshape(20, 30, 40)
a = blosc2.zeros((20, 30, 40), dtype=np.float32)
a[2:9, 3:17] = y[2:9, 3:17]  # ValueError: ndarray is not C-contiguous

NDArray.set_slice() gets the value through the buffer protocol as one dense C-order buffer (PyBUF_SIMPLE), which NumPy cannot provide for a view with gaps (a box of a larger array), with negative strides or with permuted axes, nor for a Fortran-ordered array. A value of another dtype only worked when astype() happened to return a C-ordered copy, which it does not for Fortran-ordered values.

Fix. __setitem__ makes the value C-contiguous right before set_slice(): np.ascontiguousarray(value), a no-op for C-contiguous arrays and the copy users make by hand today otherwise. The non-unit-step and fancy-index paths already copy, so they are unchanged.

Tests. test_setitem_non_contiguous_value: strided, reversed, Fortran-ordered, axis-permuted and 1-D strided values, of the array's dtype and of another one, in memory and on disk, against NumPy. 14 of the 20 cases fail on main. The full suite passes (2 b2view HTTP tests need aiohttp, which my environment lacks).

🤖 Generated with Claude Code

[FrancescAlted]

Sure thing!