The OpenMP multi-threading API

  • OpenMP:

    • OpenMP (Open Multi-Processing) is an application programming interface (API) that supports multi-platform shared-memory multiprocessing programming

  • How OpenMP works:

    • OpenMP is implemented using multithreading

    • A primary thread creates a specified number of sub-threads that execute in parallel

The pragma compiler directive

  • The #pragme directive:

    • In C: 

         #pragma

      is a special purpose compiler directive that is used to turn on or turn off some feature(s)

  • Example: 

      #pragma GCC poison printf: disallow the use of printf

  • The Open Multi-Processing (OpenMP) is a directive-based API for developing parallel programs on shared-memory SIMD processors

DEMO: demo/OpenMP/pragma.c

Defining parallel regions in a C program

  • Parallel region:

    • Parallel region = code in a C program that is executed by multiple threads

  • Syntax to define a parallel region in OpenMP: 

     #pragma omp parallel [Options...] 
 { ... 
   ... Parallel region
   ...
   ... Program statements between the braces
   ... are executed in parallel by multiple threads      
   ...
 }

  • Controlling the number of threads in the parallel region:

    1. Set the environment variable OMP_NUM_THREADS
      (e.g.: export OMP_NUM_THREADS=4)

    2. Set number of threads with library function omp_set_num_threads(n)

Your first OpenMP program

  • The Hello World program in OpenMP: 

    int main(int argc, char *argv[])
{
   #pragma omp parallel 
   {
      printf("Hello World!\n");
   }
}

  • Compiling an OpenMP C program: 

        gcc   -fopenmp    C-program

  • Output: printed 4 times ( by default: # threads = # CPU cores) 

    Hello World!
Hello World!
Hello World!
Hello World!

DEMO: demo/OpenMP/hello1.c

Finding the ID of a thread and the number of threads in the parallel region

  • Some useful OpenMP library functions: 

      omp_get_thread_num( ):  returns the ID of a thread

  omp_get_num_threads( ): returns the # threads in the parallel region

  • Example: 

    int main(int argc, char *argv[])
{
   #pragma omp parallel 
   {
      int nthreads;
      int myid;

      myid = omp_get_thread_num(); // Get my ID in thread group
      printf("Hello, I am thread %d...\n", myid);

      if (myid == 0)
      {
         nthreads = omp_get_num_threads();
         printf("Thread 0: Number of threads used = %d\n", nthreads);
      }
   }
}

DEMO: demo/OpenMP/hello2.c

What is happening inside a OpenMP program

  • Graphical representation of the execution of the OpenMP "Hello World" program:

    • Before the compiler directive #pragma omp parallel, the program is single-threaded

    • The compiler directive #pragma omp parallel, tells the C compiler to create a thread team to execute the parallel region

    • At the end of the parallel region, the threads are implicitly joined