Some advices about vectorization

  In order to make full advantage of vector computers it is necessary that the code is highly vectorizable. Below we shall give some advices of how to improve the vectorizability of a code without reducing the scalar performance.

1. Use only clean loops without IF, GOTO or CALL statements. For example:

                   do 100i = 1,n
                        if (var > 0) then
                              a(i) = 0d0
                        else
                              a(i) = b(i)
                        endif
             100   continue
   

   should be replaced by:

                   if (var > 0) then
                         do 100i = 1,n
                            a(i) = 0d0
             100         continue
                   else
                         do 110i = 1,n
                            a(i) = b(i)
             110         continue
                   endif
   

2. Use if possible and meaningful BLAS subroutines to perform simple linear algebraic operations. In most vector and parallel computers efficient implementations of these routines are available.
3. Make sure that the inner loop is the longest one of a series of nested do-loops. In that case the vector speed can be best utilized.
4. Use the inner loop for the first index of more-dimensional arrays in case of nested loops. This may increase the efficiency, provided the inner loop is long enough.
5. Avoid indirect addressing if possible.
6. Avoid division in a do-loop.
7. Avoid (seemingly) recursions in a loop.