Dispatching by values is very slow - Feature or Problem?

[viraltux]

@Sukera

@benchmark already loops for you and you’re also counting creation time of the random matrix, so that’s going to skew your benchmark significantly. I recommend using the setup keyword for @benchmark (e.g. setup=(x=rand(n)) and passing x to sort! without interpolation), as well as evals=1 to make sure each matrix is only ever used once.

BenchmarkTools is not very well suited to compare benchmarks as discussed in Identical functions benchmark differently but that’s another story. Actually benchmarks with those settings only makes differences wider, and using a dispatching function instead dispatching by value makes sortn! ~300% faster than default Base.sort!

using SortingNetworks
n = 4

#206.279 μs ±  10.895 μs
@benchmark for i in 1:10000 sort!(x)  end  setup=(x=rand(n)) evals=1

#64.406 μs ±   5.169 μs
@benchmark for i in 1:10000 sortn!(x) end  setup=(x=rand(n)) evals=1

and for sort vs swapsort

# 565.565 μs ± 142.188 μs
@benchmark for i in 1:10000 sort(x)  end  setup=(x=rand(n)) evals=1

#1.449 ms ± 205.297 μs
@benchmark for i in 1:10000 swapsort(x) end  setup=(x=rand(n)) evals=1