Benchmarking array closures

General Usage Performance
,
1

Want to verify that I’ve done this in a way that gives valid results.

In the code below dostuff! updates 2 arrays in place. clos! creates a closure that encloses the arrays and accepts arguments for the change factors.

Function doitem! is closer to what I do in my code: one element of one or more arrays is modified in place. clositem! encloses the arrays and accepts arguments for the changes.

The good news, if I’ve done this correctly, there appears to be no performance penalty for enclosing the arrays. That’s great because the code where I modify arrays is in a very hot loop.

The slight surprise is that there is a higher % penalty (though quite small in absolute time) when modifying a single element of each array.

Comments, corrections, suggestions welcome!

Code:

function buildit()
    print("\nEnter n: "); n = parse(Int64, (chomp(readline())))
    return rand(n), collect(1:n)
end

function dostuff!(xfl, xint, arr1, arr2)
    arr1 .*= xfl
    arr2 .+= xint
end

function doitem!(xfl, xint, item, arr1, arr2)
    arr1[item] *= xfl
    arr2[item] += xint
end

arrfl, arrint = buildit()

clos!(xfl, xint) = dostuff!(xfl, xint, arrfl, arrint)
clositem!(xfl, xint, item) = doitem!(xfl, xint, item, arrfl, arrint)

Benchmarks for arrays with 100_000 elements:

julia> @btime dostuff!(1.0, 1, arrfl, arrint);
  2.903 μs (0 allocations: 0 bytes)

julia> @btime clos!(1.0, 1);
  2.907 μs (1 allocation: 16 bytes)

julia> @btime doitem!(1.0, 1, 25, arrfl, arrint)
  13.847 ns (1 allocation: 16 bytes)

julia> @btime clositem!(1.0, 1, 25)
  19.121 ns (2 allocations: 32 bytes)
2

You’re benchmarking with non-const global variables, in fact capturing non-const global variables in closures.

(The parent post was written two years ago but somehow edited 5 days ago, so I didn’t need to scroll much to find it.)

5

Thanks.

Are you sure you’ve enclosed the arrays? Seems so. Easy to verify. Add a statement to the outer function that prints an element…

6

Please don’t comment (“necropost”) on years-old threads. The original issue in this thread is due to using global variables.

If you have specific new performance issue with closures that you want to discuss, please start a new thread with example code.

1 Like