DataArrayXxx: InterpKernelException raised for arrays with unused but non-standard strides

[j-demersseman]

Description:

MEDCoupling raises an exception when constructing a DataArray from a numpy array that is logically contiguous but whose unused stride does not match the item size. This happens in at least two cases:

• Column vector of shape (N, 1) — the stride along axis 1 is irrelevant since there is only one column, but MEDCoupling validates it anyway.
• Single-element 1-D array — the stride is irrelevant since there is only one element.

Version: MEDCoupling 9.14.0

Steps to reproduce:

import medcoupling as mc
import numpy as np

# Case 1: column vector with non-standard (but unused) column stride
base = np.arange(10, dtype=np.float64)
col = np.lib.stride_tricks.as_strided(base, shape=(10, 1), strides=(8, 1928))
da = mc.DataArrayDouble(col)  # raises InterpKernelException

# Case 2: single-element array with non-standard stride
base = np.arange(1, dtype=np.float64)
single = np.lib.stride_tricks.as_strided(base, shape=(1,), strides=(4,))
da = mc.DataArrayDouble(single)  # raises InterpKernelException

Workaround

NPT = TypeVar("NPT", bound=np.generic)

def as_contiguous_for_mc(arr: npt.NDArray[np.generic], dtype: type[NPT]) -> npt.NDArray[NPT]:
    """Ensure the numpy array is C-contiguous and of the specified dtype for MEDCoupling."""
    array = np.ascontiguousarray(arr, dtype=dtype)
    if array.ndim == 2 and array.strides[1] != array.itemsize and array.shape[1] == 1:
        # For column (n, 1), Numpy don't check for column contiguity (because this stride is useless)
        array = array.reshape(-1, 1)  # force to redo the column stride
    elif array.ndim == 1 and array.size == 1:
        # For single element array, Numpy don't check for contiguity
        array = array.copy()
    return array

Root cause analysis:

• For a (N, 1) array, the stride of the second dimension is never used during data traversal; only the first-dimension stride matters.
• For a single-element array, no stride is ever applied.

In both cases, NumPy may assign an arbitrary stride value to the unused dimension, which MEDCoupling's validator incorrectly rejects.

[j-demersseman]

I'm not sure if it's still an issue after the resolition of the issue #11. I couldn't test it.