Is this the maximum perf i can obtain?

[lrnv]

@djholiver

Why do you need to allocate to slack (or even provide it as a variable) in the two methods?

because I actually need the result for later computations. Look again at my runtime_test! function:

function runtime_test!(rez,slack,D)
    n = length(rez)
    N = length(D)
    rez[1] = exp_and_mean!(slack,D,N)
    for i in 1:n-1
        rez[i+1] = prod_and_mean!(slack,D,N)
    end
    return rez
end

You see that the slack is passed along from one function to the other. Actually, if you denote by X the rndom variable that corresponds to the data D, my function computes \mathbb E\left(X^k e^{-X}\right), for k \in \{0,...,n-1\}.

@djholiver

is there also potentially an error in “prod_and_mean” ? do you mean to sum zz (I change to multiple above).

From your input setup vars, you are passing in slack as a vector of zeros so doesnt this yield zero for all cases in “prod_and_mean” ? Obviously they may be intended.

No there is no error in “prod_and_mean”, I indeed mean to sum zz. Yes this is intended, since slack is not expected to be zeros when calling the prod_and_mean function, look again at the runtime_test! function

@djholiver

Do you have any opportunity to modify all “Ds” before being passing into the functions at all or are they always generated before being called? You might benefit from creating the entire “D” matrix space in exp and not exp forms before executing your functions (you’d also be able to combine them if this is applicable).

I do not understand what you mean. This computation is done only once for a given dataset D, I am afraid. The dataset D changes from one execution to the next, albeit not completely randomly: I compute it from another dataset data as follows:

data=.... #(sized (10,10000)
for i in 1:n_iterations 
    e = rand(10) # A new one is picked at each iteration 
    D = data'e # now a vector of size 10000
    runtime_test!(rez,slack,D)
    do_something_with!(rez)
end

data is fixed from one iteration to the other, but e is not (and therefore neither D, slack and rez). rez is needed for the rest of the computations, but D and slack are not.