ValueError continuous memory view and 0-size numpy array

[paugier]

A ValueError is raised if one gives a numpy array with shape (2, 0, 1) to such Cython functions with a 3d memoryview as argument:

cimport numpy as np
import numpy as np
np.import_array()

cpdef myfunc3d(np.float64_t[:, :, ::1] arr):
    if arr.size > 0:
        print(arr[0, 0, 0])
    else:
        print('size == 0')

The code to get the error:

import numpy as np
arr = np.ones((2, 0, 1))
print(arr.flags)
myfunc3d(arr)

which gives:

C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : True
  WRITEABLE : True
  ALIGNED : True
  UPDATEIFCOPY : False
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-7-81a397764d36> in <module>()
      1 arr = np.ones((2, 0, 1))
      2 print(arr.flags)
----> 3 myfunc3d(arr)

_cython_magic_5dfab2d7f2a0e837d79c56566c9147b4.pyx in _cython_magic_5dfab2d7f2a0e837d79c56566c9147b4.myfunc3d()

ValueError: Buffer and memoryview are not contiguous in the same dimension.
In [ ]:

Strangely, no error is raised for a shape (1, 0, 1)!

[ashwinvis]

Strangely enough, by changing the function definition as shown below does not raise any error.

from cython cimport view

def myfunc3d(np.float_t[:,:,::view.contiguous] arr):

Possible explanation from Cython docs

The difference between the contiguous flag and the ::1 specifier is that the former specifies contiguity for only one dimension, whereas the latter specifies contiguity for all following (Fortran) or preceding (C) dimensions:

However this is not a solution, the issue still persists, IMHO.

[ashwinvis]

To add to the previous comment, although no error is raised... the output size == 0 is unexpected!

>>> import numpy as np; arr = np.ascontiguousarray(np.ones((2, 0, 1))); print(arr.flags); myfunc3d(arr)

  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : True
  WRITEABLE : True
  ALIGNED : True
  WRITEBACKIFCOPY : False
  UPDATEIFCOPY : False
size == 0

>>> arr.size

0

EDIT: Correction, I was not thinking clearly.

[c-f-h]

I ran into this too. Here's another minimal example to reproduce this:

cdef void foo(double[:, :, ::1] X):
    pass

import numpy as np
X = np.zeros((4,4,4))

# this works
foo(np.ascontiguousarray(X[:, 2:3, :]))

# this fails with 'ValueError: Buffer and memoryview are not contiguous in the
# same dimension.'
foo(np.ascontiguousarray(X[:, 2:2, :]))

The problem is that __Pyx_ValidateAndInit_memviewslice does not correctly realize that a 0-size array is always C-contiguous.

This crops up in the wild, too. A search for the error message reveals many instances where people inadvertently passed 0-size arrays to cython-implemented functions and got confusing error messages. E.g.: https://stackoverflow.com/questions/46470492/buffer-and-memoryview-are-not-contiguous-in-the-same-dimension-in-python

[c-f-h]

For reference, here is the corresponding check for contiguity in numpy:

https://github.com/numpy/numpy/blob/3873ec46816f3fe0f705394242f83a764d4ce90e/numpy/core/src/multiarray/flagsobject.c#L123

Note that it is contingent on the NPY_RELAXED_STRIDES_CHECKING compile-time environment variable, however this seems to be always on by default.

The Cython function __Pyx_ValidateAndInit_memviewslice does not have the same check for 0-length dimensions.

[scoder]

Yes, that's a bug. It should behave like NumPy.
PR welcome. Tests can go into tests/memoryview/memslice.pyx, there are several contiguity tests already.