Compilation time not vanishing on subsequent runs

New to Julia
1

Hi!

I’ve got a relatively simple piece of code that takes an 2D array (“correlation”), and gives you a radial average working outwards from the centre.

xs = range(1, 100, 100)
ys = range(1, 100, 100)
R(x, y) = (sqrt((x-51)^2 + (y-51)^2))
autocorr_distances = R.(xs, ys')

@time begin
for r in range(1, 50, 50) 
    indexes = findall(x->(x<(r+0.5) && x>(r-0.5)), autocorr_distances)
    sum = 0
    for index in indexes
        sum += correlation[index[1], index[2]]
    end
    radial_sum[trunc(Int,r)] = sum/length(indexes)
end 
end

The code itself works fine and computes what I want it to, but when I time it, it yields:

0.171478 seconds (565.92 k allocations: 26.139 MiB, 98.28% compilation time)

The roughly 98% compilation time remains even when I run the code multiple times. I am a computer science and Julia beginner, so apologies if the answer is obvious!

1 Like
2

My guess would be the redefinition of this anonymous function, each time you run it you have to compile it again.

4 Likes
3

You’re correct, I swapped out the findall for something explicit, namely;

radial_sum[r] = Statistics.mean(correlation[r-0.5 .< autocorr_distances .< r + 0.5])

And now I get:

0.000932 seconds (651 allocations: 360.500 KiB)

Thank you!

4

Just for fun, I put your original code in a function:

function testit()
    xs = range(1, 100, 100)
    ys = range(1, 100, 100)
    R(x, y) = (sqrt((x-51)^2 + (y-51)^2))
    autocorr_distances = R.(xs, ys')
    correlation = rand(100,100)
    radial_sum = zeros(50)

    @time begin
        for r in range(1, 50, 50) 
            indexes = findall(x->(x<(r+0.5) && x>(r-0.5)), autocorr_distances)
            sum = 0
            for index in indexes
                sum += correlation[index[1], index[2]]
            end
            radial_sum[trunc(Int,r)] = sum/length(indexes)
        end 
    end
end

I had to add initialization for correlation and radial_sum which weren’t provided in your code. With this function definition I get the following output:

julia> testit()
  0.072925 seconds (12.87 k allocations: 833.984 KiB, 99.13% compilation time)

julia> testit()
  0.000796 seconds (350 allocations: 412.312 KiB)

julia> testit()
  0.000629 seconds (350 allocations: 412.312 KiB)

Roughly the same as your results with Statistics.mean. The moral is to beware of nonconstant globals.

I would also recommend replacing the range(1,50,50) with the simpler 1:50 which, when iterated, takes integer values, so that radial_sum[trunc(Int,r)] can be replaced by radial_sum[r]. Also, your variable name sum is not ideal, since there is a built-in sum function in Julia. Finally, since correlation is presumably a matrix of floating-point values, you should initialize summ (new name) to 0.0 (or even better to zero(eltype(correlation))) to avoid type instability.

4 Likes
5

Hi @PeterSimon,

Thank you so much for all of the advice, I really appreciate it! Your cautioning about nonconstant globals was key to solving a similar issue elsewhere in my code.

Thanks again!