append an array as a column to an pre initialized matrix - julia

Question

I wish to use indexing to populate a pre initialized matrix with the results of my array for loop output:
A = Float64.(reshape(1.0:81.0,9,9))
# initialize output
B = zeros(Int64, 2, 9)
j = 1
for j in 1:size(A,2) # loop cols
out = [sum(A[:,j]),j]
out = reshape(out,2,1) # make column to append
# append out to the B
global B = hcat(B,out) # this grows...
end
I initialized B = zeros(Int64, 2, 9)
same dims as the expected output of the sum operation.
in my real world example - I am iterating through j, columns and, i rows - the output is then an array... rather than use hcat() to append the array to my output can I do it with indexing?
In the above it uses hcat() which will then append to the existing B so it grows. I have since tried initializg with rows 2 and cols 0 so hcat() builds to correct output dim:
B = zeros(Int64, 2, 0)
I am doubting if hcat() will be memory efficient (excuse using global for example sakes) - if I couldn't do it with indexing I can populate it for another inner loop at my [i,j]. But perhaps someone has a way I can append an array as a column to existing pre initialized output?

Answer 1

The recommendation is to pre-allocate B and fill it afterwards. I wrap the code in a function as it simplifies benchmarking:
function f2()
A = reshape(1:81,9,9)
B = zeros(Int64, 2, 9 + size(A,2))
for j in 1:size(A,2) # loop cols
B[:, j + 9] .= (sum(view(A, :, j)), j)
end
B
end
Your old code is:
function f1()
A = Float64.(reshape(1.0:81.0,9,9))
B = zeros(Int64, 2, 9)
j = 1
for j in 1:size(A,2) # loop cols
out = [sum(A[:,j]),j]
out = reshape(out,2,1) # make column to append
# append out to the B
B = hcat(B,out)
end
B
end
And here is a comparison:
julia> #btime f1()
8.567 μs (83 allocations: 7.72 KiB)
2×18 Array{Float64,2}:
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 45.0 126.0 207.0 288.0 369.0 450.0 531.0 612.0 693.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0
julia> #btime f2()
73.662 ns (1 allocation: 368 bytes)
2×18 Array{Int64,2}:
0 0 0 0 0 0 0 0 0 45 126 207 288 369 450 531 612 693
0 0 0 0 0 0 0 0 0 1 2 3 4 5 6 7 8 9
And you can see that the difference is very significant.
Some more minor comments to your original code:
there is no need to call Float64. on reshape(1.0:81.0,9,9), the object already holds elements that have Float64 values
in your code there was an inconsistency that initilally B held Int64 and A held Float64 - i have made this consistent (I chose Int64, but equally well you could use Float64)
sum(A[:,j]) unnecessarily allocated a new object; it is faster to use a view
You did not have to call reshape(out,2,1) on out before hcat as vectors are already treated as columnar objects