[BUG] distance_transform_edt gives bad output for 2D images > 1024 in size on either axis

[grlee77]

Describe the bug
There is currently a bug in the 2D implementation of distance_transform_edt where certain lines of the output have incorrect values due to a bad default choice of internal CUDA grid/block parameters at this size. Fortunately, the 3D distance transform does not suffer from the same issue.

A temporary workaround in release 22.08.00 is to manually specify the block_params argument and set the first parameter to 1. This is smaller than needed for optimal performance, but will ensure a correct result.

In general for 2D images where one axis is larger than 1024, please set block_params=(1, 32, 2). Other than as a workaround for this bug or in case of advanced performance profiling use cases, we recommend leaveling block_params=None to allow internal choice of these kernel launch parameters.

Steps/Code to reproduce bug
a minimal example

import math
import cupy as cp

import matplotlib.pyplot as plt

from cucim.core.operations.morphology import distance_transform_edt

shape = (1200, 1200)
size = math.prod(shape)
ntrue = .0005* size
p_true =  ntrue / size
p_false = 1 - p_true

cp.random.seed(123)
image = cp.random.choice([0, 1], size=shape, p=(p_false, p_true))

distances, coords = distance_transform_edt(
    image == 0, return_distances=True, return_indices=True
)
plt.figure()
plt.imshow(distances.get(), vmin=0, vmax=116)
plt.show()

gives this distances output where the bright horizontal lines are invalid distances

whereas calling with

distances, coords = distance_transform_edt(
    image == 0, return_distances=True, return_indices=True, block_params(1, 32, 2)
)

gives the correct result:

Expected behavior
no artifacts with the default choice of block_params

Environment details (please complete the following information):
cuCIM v22.08.,00

Additional context
Add any other context about the problem here.