Slow implementation of convolve in chapel

conv.chpl

use Random;
use LinearAlgebra;

proc rand(D: int, seed : int, mag: real):[1..D] real {
  var X: [1..D] real;
  fillRandom(X, seed, algorithm=RNG.NPB);
  return mag*(2*X-1);
}

proc conv(A: [?D1] real, K: [?D2] real) {
  var B: [1..(A.size - K.size)] real;

  /*[i in 1..B.size] B[i] = dot(A[i..(i+K.size-1)], K);*/
  /*12.029s*/

  /*[i in 1..B.size] B[i] = + reduce (A[i..(i+K.size-1)] * K);*/
  /*8.923s*/

  /*for i in 1..B.size {
    var s = 0.0;
    for j in 1..K.size {
      s += A[i+j-1] * K[j];
    }
    B[i] = s;
  }*/
  /*5.778s*/

  /*Check first 5 elements are equivalent*/
  /*writeln(B[1..5]);*/
  return B;
}
var X = rand(1000000, 1, 1);
var K : [1..100] real;
K = 1/100.0;
conv(X,K);