Compare array wise of two 3D matrices in Numpy without for loops

Leverage broadcasting after extending - the input to two 4D versions such that we can compare the pairiwise-array-blocks against each other along the first axis while keeping the last two axes aligned -

result = (M[:,None] == M).all((2,3))

We can extend this to generic n-dim array case using last two axes as input to .all() -

(M[:,None] == M).all((-2,-1))

Leverage views to have a more memory efficient and hence performant one -

# https://stackoverflow.com/a/44999009/ @Divakar
def view1D(a): # a is array
    a = np.ascontiguousarray(a)
    void_dt = np.dtype((np.void, a.dtype.itemsize * a.shape[1]))
    return a.view(void_dt).ravel()

M1D = view1D(M.reshape(M.shape[0],-1))
result = M1D[:,None] == M1D

Timings on large array -

In [48]: np.random.seed(0)
    ...: M = np.random.randint(0,10,(100,100,100))

In [49]: %timeit (M[:,None] == M).all((-2,-1))
10 loops, best of 3: 92.2 ms per loop

In [50]: %%timeit 
    ...: M1D = view1D(M.reshape(M.shape[0],-1))
    ...: M1D[:,None] == M1D
1000 loops, best of 3: 627 µs per loop

Original one -

In [54]: %%timeit
    ...: result = np.zeros((M.shape[0], M.shape[0]))
    ...: for i in range(M.shape[0]):
    ...:     for j in range(M.shape[0]):
    ...:         result[i,j] = numpy.array_equal(M[i], M[j])
10 loops, best of 3: 125 ms per loop

The conclusion would be - Remove loops, but watch out for memory usage. If possible, find other ways that keeps it vectorized and memory efficient.