Optimisation: Alma built-in vs Pine

//@version=4
// The built-in is ~15x faster than the Pine version.

study("PP alma built-in vs usr function")
s = close, l = 4, o = 0.85, g = 6, d = 1

pine_alma(series, windowsize, offset, sigma) =>
    m = floor(offset * (windowsize - 1))
    s = windowsize / sigma
    norm = 0.0
    sum = 0.0
    for i = 0 to windowsize - 1
        weight = exp(-1 * pow(i - m, 2) / (2 * pow(s, 2)))
        norm := norm + weight
        sum := sum + series[windowsize - i - 1] * weight
    sum / norm

a = 0.
for i=1 to 9000
    // a := alma(s+i,l,o,g) //K iterations working: 5 15 25 50 100 125 140, Not working: 200 150 145 140
    a := pine_alma(s+i,l,o,g) //K iterations working: 5 7.5 8.5 9 // Not working: 125 25 15 10
plot(a)