Testing per-function-call allocation against re-using buffer in type

buftest.jl

type BufContainer
    inc_amt::Float32
    buf::Array{Float32}

    BufContainer(inc_amt) = new(inc_amt, Array(Float32, 1024))
end

function with_alloc(cont::BufContainer)
    buf = Array(Float32, 1024)
    for i in 1:length(buf)
        buf[i] += cont.inc_amt
    end
    return buf
end

function without_alloc(cont::BufContainer)
    for i in 1:length(cont.buf)
        cont.buf[i] += cont.inc_amt
    end
    return cont.buf
end

function without_alloc_with_temp(cont::BufContainer)
    temp = cont.buf
    for i in 1:length(temp)
        temp[i] += cont.inc_amt
    end
    return temp
end

function using_parameter(cont::BufContainer, buf::Array{Float32})
    for i in 1:length(buf)
        buf[i] += cont.inc_amt
    end
    return buf
end

function buftest()
    cont = BufContainer(1)
    buf = Array(Float32, 1024)
    # call once for JITting
    _ = with_alloc(cont)
    _ = without_alloc(cont)
    _ = without_alloc_with_temp(cont)
    _ = using_parameter(cont, buf)

    N_TRIALS=100000
    gc()
    println("Doing $N_TRIALS Interations of each...\n-------")
    println("Timing with allocation each call")
    @time begin
        for _ in 1:N_TRIALS
            res = with_alloc(cont)
        end
        gc()
    end

    println("Timing without allocation each call")
    @time begin
        for _ in 1:N_TRIALS
            res = without_alloc(cont)
        end
        gc()
    end

    println("Timing without allocation using a temp buffer each call")
    @time begin
        for _ in 1:N_TRIALS
            res = without_alloc_with_temp(cont)
        end
        gc()
    end

    println("Timing passing array as a parameter")
    @time begin
        for _ in 1:N_TRIALS
            res = using_parameter(cont, cont.buf)
        end
        gc()
    end
end

buftest()

# run on 1.6GHz i5

# Doing 100000 Interations of each...
# -------
# Timing with allocation each call
# elapsed time: 0.474479476 seconds (417591824 bytes allocated, 42.90% gc time)
# Timing without allocation each call
# elapsed time: 0.35568662 seconds (1591824 bytes allocated, 2.88% gc time)
# Timing without allocation using a temp buffer each call
# elapsed time: 0.18984845 seconds (1591824 bytes allocated, 5.64% gc time)
# Timing passing array as a parameter
# elapsed time: 0.173606393 seconds (1591824 bytes allocated, 6.03% gc time)

With this code I get:

Timing with allocation each call
elapsed time: 0.001781621 seconds (4167824 bytes allocated)
Timing without allocation each call
elapsed time: 0.001545691 seconds (7824 bytes allocated)
Timing without allocation using a temp buffer each call
elapsed time: 0.00088557 seconds (7824 bytes allocated)
Timing passing array as a parameter
elapsed time: 0.000885131 seconds (7824 bytes allocated)

The difference between 2 & 3 is because of getting the field of cont (I think)

type BufContainer
    inc_amt::Float32
    buf::Array{Float32,1}

    BufContainer(inc_amt) = new(inc_amt, Array(Float32, 1024))
end

function with_alloc(cont::BufContainer)
    buf = Array(Float32, 1024)
    for i in 1:length(buf)
        buf[i] += cont.inc_amt
    end
    return buf
end

function without_alloc!(cont::BufContainer)
    for i in 1:length(cont.buf)
        cont.buf[i] += cont.inc_amt
    end
end

function without_alloc_with_temp!(cont::BufContainer)
    temp = cont.buf
    for i in 1:length(temp)
        temp[i] += cont.inc_amt
    end
end

function using_parameter!(cont::BufContainer, buf)
    for i in 1:length(buf)
        buf[i] += cont.inc_amt
    end
end

cont = BufContainer(1)
buf = Array(Float32, 1024)
# call once for JITting
_ = with_alloc(cont)
_ = without_alloc!(cont)
_ = without_alloc_with_temp!(cont)
_ = using_parameter!(cont, buf)

gc()
println("Timing with allocation each call")
@time for _ in 1:1000
    cont.buf = with_alloc(cont)
end
gc()
println("Timing without allocation each call")
@time for _ in 1:1000
    without_alloc!(cont)
end
gc()
println("Timing without allocation using a temp buffer each call")
@time for _ in 1:1000
    res = without_alloc_with_temp!(cont)
end
gc()
println("Timing passing array as a parameter")
@time for _ in 1:1000
    using_parameter!(cont, buf)
end