While the proposed duplicate What does the Python Ellipsis object do? answers the question in a general python context, its use in an nditer loop requires, I think, added information.
https://docs.scipy.org/doc/numpy/reference/arrays.nditer.html#modifying-array-values
Regular assignment in Python simply changes a reference in the local or global variable dictionary instead of modifying an existing variable in place. This means that simply assigning to x will not place the value into the element of the array, but rather switch x from being an array element reference to being a reference to the value you assigned. To actually modify the element of the array, x should be indexed with the ellipsis.
That section includes your code example.
So in my words, the x[...] = ... modifies x in-place; x = ... would have broken the link to the nditer variable, and not changed it. It's like x[:] = ... but works with arrays of any dimension (including 0d). In this context x isn't just a number, it's an array.
Perhaps the closest thing to this nditer iteration, without nditer is:
In [667]: for i, x in np.ndenumerate(a):
...: print(i, x)
...: a[i] = 2 * x
...:
(0, 0) 0
(0, 1) 1
...
(1, 2) 5
In [668]: a
Out[668]:
array([[ 0, 2, 4],
[ 6, 8, 10]])
Notice that I had to index and modify a[i] directly. I could not have used, x = 2*x. In this iteration x is a scalar, and thus not mutable
In [669]: for i,x in np.ndenumerate(a):
...: x[...] = 2 * x
...
TypeError: 'numpy.int32' object does not support item assignment
But in the nditer case x is a 0d array, and mutable.
In [671]: for x in np.nditer(a, op_flags=['readwrite']):
...: print(x, type(x), x.shape)
...: x[...] = 2 * x
...:
0 <class 'numpy.ndarray'> ()
4 <class 'numpy.ndarray'> ()
...
And because it is 0d, x[:] cannot be used instead of x[...]
----> 3 x[:] = 2 * x
IndexError: too many indices for array
A simpler array iteration might also give insight:
In [675]: for x in a:
...: print(x, x.shape)
...: x[:] = 2 * x
...:
[ 0 8 16] (3,)
[24 32 40] (3,)
this iterates on the rows (1st dim) of a. x is then a 1d array, and can be modified with either x[:]=... or x[...]=....
And if I add the external_loop flag from the next section, x is now a 1d array, and x[:] = would work. But x[...] = still works and is more general. x[...] is used all the other nditer examples.
In [677]: for x in np.nditer(a, op_flags=['readwrite'], flags=['external_loop']):
...: print(x, type(x), x.shape)
...: x[...] = 2 * x
[ 0 16 32 48 64 80] <class 'numpy.ndarray'> (6,)
Compare this simple row iteration (on a 2d array):
In [675]: for x in a:
...: print(x, x.shape)
...: x[:] = 2 * x
...:
[ 0 8 16] (3,)
[24 32 40] (3,)
this iterates on the rows (1st dim) of a. x is then a 1d array, and can be modified with either x[:] = ... or x[...] = ....
Read and experiment with this nditer page all the way through to the end. By itself, nditer is not that useful in python. It does not speed up iteration - not until you port your code to cython.np.ndindex is one of the few non-compiled numpy functions that uses nditer.
