My reading of it is that it's essentially one-to-one (up to minor constant factors), they took the binary circuit for xorshift128 and replace the xor gates by f

My reading of it is that it's essentially one-to-one (up to minor constant factors), they took the binary circuit for xorshift128 and replace the xor gates by functionally equivalent NN blocks, and then trained the weights.