Performance

[henry2004y]

I am a little bit surprised about the performance of argmaxima compared with a direct implementation:

using Test, BenchmarkTools
using Peaks

function localmaxmin!(y, maxs::Vector{Int}, mins::Vector{Int})
   empty!(maxs)
   empty!(mins)
   for i in 2:length(y)-1
       if y[i+1] < y[i] > y[i-1]
           push!(maxs, i)
       elseif y[i+1] > y[i] < y[i-1]
           push!(mins, i)
       end
   end
end

function localmaxmin2!(x, maxs, mins)
   append!(maxs2, argmaxima(x))
   append!(mins2, argminima(x))
   maxs2, mins2
end

function localmax(y)
   maxs = Int[]
   for i in 2:length(y)-1
       if y[i+1] < y[i] > y[i-1]
           push!(maxs, i)
       end
   end
   maxs
end

function localmax2(x)
   maxs2 = argmaxima(x)
end

maxs = Int[]
mins  = Int[]
maxs2= Int[]
mins2 = Int[]
x = zeros(100000)
x_max = 2:100:100000
x[x_max] .= 1
x_min = 5:100:100000
x[x_min] .= -1
localmaxmin!(x, maxs, mins)
localmaxmin2!(x, maxs2, mins2)

@test maxs == maxs2
@test mins == mins2

@btime localmax(x)
@btime localmax2(x)
@test localmax(x) == localmax2(x)

On my laptop, this returns

102.428 μs (10 allocations: 16.39 KiB)
380.387 μs (6 allocations: 398.92 KiB)
Test Passed

which is 3 times slower.

[halleysfifthinc]

Actually its almost 4x slower! I too am surprised by the size of the difference in performance! I'm also surprised about the difference in allocation size.

The argmaxima function has more features which explains at least some of the decreased performance compared to your localmax function: argmaxima supports larger/custom window sizes, identification of plateaus, and handles the presence of NaNs and missings. The latter two features likely contribute the bulk of the increased runtime.

In any case, if the localmax function meets your needs, it doesn't appear to that switching to Peaks.jl would be beneficial, at the moment.