ndfilters.convolve wrap mode not working when the array is chunked

What happened: Used ndfilters.convolve using the wrap mode and the output had the wrong values on the edges What you expected to happen: The function to output the correct values on the edges. Minimal Complete Verifiable Example:

import numpy as np
import dask.array as da
from dask_image.ndfilters import convolve


def test_correct(F,axis = 0):
    if axis !=0:
        F = da.swapaxes(F, 0, axis)
        result = test_correct(F,axis = 0)
        return result
    
    kernel = np.array([1,-1,0,1,-1]).reshape((5,) + (1,) * (F.ndim - 1))
    # The output using ndfilters.convolve with the wrap mode
    O1 = convolve(F,kernel, mode = 'wrap')
    # Output created by forcing the wrapping using dask.array.overlap.overlap then calculating the convolution. 
    # taking care of the boundary
    F_1 = da.overlap.overlap(F, depth = {axis : 2}, boundary = 'periodic')
    # calculate the convolution
    O2_int = convolve(F_1,kernel, mode = 'constant')
    # removing the useless ghosts zones. 
    O2 = da.overlap.trim_overlap(O2_int, depth = {axis : 2})

    return O1.compute(), O2.compute()

test_1D_no_chunks = da.from_array([1,2,3,4,5,6,7,8,9],chunks = (-1,))
test_1D_chunked = da.from_array([1,2,3,4,5,6,7,8,9],chunks = (5,))

print('1D no chunks :',*test_correct(test_1D_no_chunks), sep = '\n', end = '\n\n')
print('1D chunked :', *test_correct(test_1D_chunked), sep = '\n', end = '\n\n')

test_2D_no_chunks = da.from_array(np.arange(25).reshape(5,5),chunks =(-1,-1))
test_2D_chunked = da.from_array(np.arange(25).reshape(5,5),chunks =(-1,2))

print('2D no chunks axis 0  :',*test_correct(test_2D_no_chunks,axis = 0), sep = '\n\n', end = '\n\n\n')
print('2D no chunks axis 1  :',*test_correct(test_2D_no_chunks,axis = 1), sep = '\n\n', end = '\n\n\n')

print('2D chunked axis 0  :',*test_correct(test_2D_chunked,axis = 0), sep = '\n\n', end = '\n\n\n')
print('2D chunked axis 1  :',*test_correct(test_2D_chunked,axis = 1), sep = '\n\n', end = '\n\n\n')

1D no chunks :
[ 2 -7  2  2  2  2  2 -7  2] # ndfilters.convolve
[ 2 -7  2  2  2  2  2 -7  2] # expected result

1D chunked :
[ 2 -5  2  2  2  2  2 -4  2] # ndfilters.convolve
[ 2 -7  2  2  2  2  2 -7  2] # expected result

2D no chunks axis 0  :

[[ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]] # ndfilters.convolve

[[ 10  10  10  10  10] 
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]] # expected result


2D no chunks axis 1  :

[[ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]] # ndfilters.convolve

[[ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]] # expected result


2D chunked axis 0  :

[[ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]] # ndfilters.convolve

[[ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]
 [-15 -15 -15 -15 -15]
 [ 10  10  10  10  10]] # expected result


2D chunked axis 1  :

[[ 2  2  2  2  2]
 [-2 -2 -2 -2 -2]
 [ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]] # ndfilters.convolve

[[ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]
 [-3 -3 -3 -3 -3]
 [ 2  2  2  2  2]] # expected result

Anything else we need to know?:

The workaround is used there to get the correct value might be a good way to fix this bug.

Environment:

• Dask version: dask 2021.4.1; dask-core 2021.4.1; dask-image 0.6.0
• Python version: 3.9.5
• Operating System: Ubuntu 20
• Install method (conda, pip, source): dask and dask-core: pip; dask-image: conda-forge

[anlavandier]

The issue is now closed with the merging of #243

[GenevieveBuckley]

Thank you for the report!