CODEWITHSUNDEEP
copenmp
Escadin
Escadin

Reputation: 15

openmp parallel for - How does it handle previously private index?

I have a parallel block containing 2 parallel for-loops:

    int i, j;
    #pragma omp parallel
    {
      #pragma omp for
      for(i=0; i < foo; i++)
        work();


      #pragma omp for private(j)
      for(i=0; i < foo; i++)
        for(j=0, j < foo; j++)
          work();
    }

If I were to make it private like this:

int i, j;
    #pragma omp parallel private(i)
    {
      #pragma omp for
      for(i=0; i < foo; i++)
        work();


      #pragma omp for private(j)
      for(i=0; i < foo; i++)
        for(j=0, j < foo; j++)
          work();
    }

Then I have NUM_THREADS amount of i copies. Will openMP still be able to schedule my threads based on i in parallel for loops and how? If I don't make it private that way (see first code example) then what behaviour can I expect from i between the two for-loops?

It is not a duplicate because I know you can usually let the parralel loop implicitly create a private version of i but I am more concerned about whether that works as expected with a variable that has been private before or has even been worked on before while being temporarily private.

Upvotes: 1

Views: 1419

Answers (2)

Z boson
Z boson

Reputation: 33659

To answer your question

what behaviour can I expect from i between the two for-loops?

In your first example i is only private in the work sharing region (in the for loops). Between the two work sharing regions it's still shared. In the second example since you declared i private for the parallel region it's private everywhere.

The is easy to show.

i = 1;
#pragma omp parallel
{
    #pragma omp for
    for(i=0; i<10; i++);
    i = 10;
}
printf("%d\n", i);

This prints 10 because i is shared except in the work sharing region and every thread writes 10 to the shared i.

However, 

i = 1;
#pragma omp parallel private(i)
{
    #pragma omp for
    for(i=0; i<10; i++);
    i = 10;
}
printf("%d\n", i);

prints 1 because i is private in the entire parallel region and does not modify the i outside of the parallel region.

In your current code since you only use i in the work sharing region then it makes no difference but if used i not in a work sharing region between the loops it could make a difference. This could lead to a subtle bug. Since you only use i in the work sharing region then I would suggest declaring i in the work sharing region using for(int i=0; ... or declare it private for the whole region. The same goes for j.

Upvotes: 1

kangshiyin
kangshiyin

Reputation: 9781

You second for-loop will not compile unless you change your code as

#pragma omp parallel
{
  #pragma omp for
  for(i=0; i < foo; i++)
    work();

  #pragma omp for private(j)
  for(i=0; i < foo; i++)
    for(j=0, j < foo; j++)
      work();
}

and 

int i, j;
#pragma omp parallel private(i)
{
  #pragma omp for
  for(i=0; i < foo; i++)
    do();

  #pragma omp for private(j)
  for(i=0; i < foo; i++)
    for(j=0, j < foo; j++)
      work();
}

It is now clear what scope the private i is.

After the modification, all of the 4 for-loops will work as expected.

Will openMP still be able to schedule my threads based on i in parallel for loops and how?

Yes, at each #pragma omp for, the private is will be initialized properly and used in the parallel for.

If I don't make it private that way (see first code example) then what behavior can I expect from i between the two for-loops?

The index in parallel for defined by #pragma omp for will always be private even if you declare it as shared. So your first code example still work.

Actually it doesn't matter whether you declare i as private or not according to this answer.

https://stackoverflow.com/a/37845938/1957265

Upvotes: 0

Related Questions