A very interesting s/w performance optimiztion case.

Here is a simple pseudocode:

function f(array_A)
{
  SoA = transform1(A);
  for i=0 to 4
    array_B=load_array_from_SoA(SoA, i);
    array_C=transform2(array_B);
    array_D+=transform3(array_C);
  return transform4(array_D);
}

SoA is a structure of arrays.
transform functions take an array, make a lot of calculations, return transformed array.
Baseline performance: let's say 200 seconds.
( Read more... )