Parallelising Loops

• Loop1 (source):
  all examples in Fortran 90
  portable data types with modul kinds (source)
  routines init and check provide some checking
  function f(x, work):

  • lots of computation (controlled by parameter work)
  • no side effects
  • returns x (approximately)
  function call inside loop
• Loop2 (source):
  all directives start with !$OMP
  basic construction: parallel region

  • PARALLEL / END PARALLEL
  • fork/join, barrier at exit
  • 
  distribution of work with DO (optional END DO)
  • index range 1..n partitioned into blocks and distributed to threads
  loop counters i and j automatically private

  • all loop indices inside of PARALLEL region (lexically = without subroutines)
  other variables shared (a, v, work)

  • no problem with a, work (read only)
  • ok with v (all threads write at different indices)
  function call works (f has no side effects),

  • local variables of f are private
  compiler doesn't check independance of loop iterations!
• Same example in C: loop2_c (source):
  directives in C start with #pragma omp
  C block instead of END directive:

  • #pragma omp parallel
  • {
  •   ...
  • }
  distribution of loops with for instead of do directive