Why do I care again? Array operations vs loops.

I use vectorized operations in (NumPy/R/Matlab/...) for speed! I don't need to write loops.

The message in Julia is not that you shouldn't write vectorized code. You can write vectorized code when it is natural, but you don't have to write vectorized code just for the sake of speed. Sometimes it's more natural to write a loop. For example:

cumulative operations
moving window operations
conditional operations on a vector
iterative algorithms

And occationally, you might want to write loops for optimization...

Using Array operations



In [0]:

    
# two 200 x 200 matricies
n = 200
A = rand(n, n)
B = rand(n, n);



In [1]:

    
f(A, B) = 2A + 3B + 4A.*A









    Out[1]:





f (generic function with 1 method)



In [2]:

    
using TimeIt



In [2]:

    
@timeit f(A, B);









    



1000 loops, best of 3: 925.69 µs per loop

This is easy to read, but allocates many temporary arrays.

Explicit loops



In [3]:

    
function f2(A, B)
    length(A) == length(B) || error("array length mismatch")
    C = similar(A, promote_type(eltype(A),eltype(B)))
    for i=1:length(C)
        @inbounds a = A[i]
        @inbounds C[i] = 2a + 3B[i] + 4a*a
    end
    return C
end









    Out[3]:





f2 (generic function with 1 method)



In [4]:

    
@timeit f2(A, B)









    



1000 loops, best of 3: 154.49 µs per loop

Pre-allocate output



In [7]:

    
function f3!(A, B, C)
    length(A) == length(B) == length(C) || error("array length mismatch")
    for i=1:length(C)
        a = A[i]
        C[i] = 2a + 3B[i] + 4a*a
    end
end









    Out[7]:





f3! (generic function with 1 method)



In [8]:

    
C = similar(A, promote_type(eltype(A),eltype(B)))
@timeit f3!(A, B, C)









    



10000 loops, best of 3: 60.97 µs per loop



In [ ]: