memory leak with typed ndarray and memory views

[jhauschild]

Hi,
I found a memory leak when using the np.ndarray[np.intp_t, ndim=2] syntax or memory views of dimension at least 2.
Strangely, the memory leak only appears if the second dimension is 0, irregardless of the other dimensions. I checked that it also appears with ndim=3 (also if the second dimension is zero).

Minimal example:
# a.pyx
# compile: cythonize -bi a.pyx
# run: python -c 'import a; a.main()'

cimport numpy as cnp
import numpy as np
import psutil

cnp.import_array()

def check_memleak(var):
    cdef cnp.ndarray[cnp.intp_t, ndim=2] c_var = var  # commenting this out prevents the leak
    #cdef cnp.intp_t[:, :] c_var = var  # memory views have the same problem
    return True

def mem():
    print(str(round(psutil.Process().memory_info().rss/1024./1024., 2)) + ' MB')

def main():
    for shape in [(2, 2), (2, 0), (0, 2), (0, 0)]:
        print("shape = ", shape)
        var = np.zeros(shape, dtype=np.intp)
        mem()
        for i in range(100000):
            check_memleak(var)
        mem()

Sample output:

$python -c 'import a; a.main()'
('shape = ', (2, 2))
27.95 MB
27.95 MB
('shape = ', (2, 0))
27.95 MB
43.81 MB
('shape = ', (0, 2))
40.21 MB
40.21 MB
('shape = ', (0, 0))
40.21 MB
43.82 MB

Any ideas what might cause this strange behaviour?

[scoder]

I looked into this a bit and could reproduce it. Not sure where the problem is, though. It would help to see if moving either NumPy or Cython out of the way would show the same issue, so that we can see if it's due to Cython or NumPy (or a mix of both).

Also, arrays with a zero dimension, while certainly not useless, do not seem like a particularly critical use case to fix. That might also be the reason why it's not been found and/or reported so far.