OpenMP and Multithreading

Tutorials

• From LLNL

C++

simple parallelization

• brute force loop parallelization with direct access to elements using [] (for instance with int[], vectors<T>, valarray<T>...)

<source lang="cpp"> Container cont;

1. pragma omp parallel for

for(int i=0; i < cont.size(); i++)

  foo(cont[i]);

</source>

• with stl iterators on containers, provided foo() does independent processes (not efficient):

<source lang="cpp"> Container cont; Container::iterator It;

1. pragma omp parallel private(It)

{

   for(It = cont->begin(); It != cont->end(); It++)
   {

1. pragma omp single nowait

    foo(It);
   }

} </source>

• calculating a sum

<source lang="cpp"> Type count = 0; Container<Type> a;

1. pragma omp parallel for

for (int i = 0; i < a.size(); ++i)

 {

1. pragma omp atomic

   count += a[i];
 }

</source> or better <source lang="cpp"> Type count = 0; Container<Type> a;

1. pragma omp parallel for reduction(+:count)

for (int i = 0; i < a.size(); ++i)

   count += a[i];

</source> you can use the reduction sentence for a list of several variables (cannot be arrays or structured data type) <source lang="cpp"> double c = 0.0; double c2 = 1.0;

1. pragma omp parallel for shared(a) reduction(+:c,c2)

for (int i = 0; i < a.size(); ++i)

  {
    c += a[i];
    c2 += a[i]+1.0;
  }

</source>

Python

one may preferably use multiprocessing