perf.jl

function add(A,B,reps)
    for i in 1:reps
        C = A + B
    end
end

function add_bc(A,B,reps)
    for i in 1:reps
        C = A .+ B
    end
end

function add_bc_prealloc(A,B,reps)
    C = similar(A)
    for i in 1:reps
        broadcast!(.+, C, A, B)
    end
end

for op in (:add, :add_bc, :add_bc_prealloc)
    for N in (10, 100, 1000)
        A = float(rand(N, N))
        B = float(rand(N, N))
        reps = int(1e7 / N^2)

        if op == :add
            add(A,B,1)
            t = minimum([ @elapsed add(A,B,reps) for i in 1:5])
        elseif op == :add_bc
            add_bc(A,B,1)
            t = minimum([ @elapsed add_bc(A,B,reps) for i in 1:5])
        elseif op == :add_bc_prealloc
            add_bc_prealloc(A,B,1)
            t = minimum([ @elapsed add_bc_prealloc(A,B,reps) for i in 1:5])
        end
        println("$op, $N x $N: $t s")
    end
end