Введение в оптимизацию приложений с помощью инструментов Intel

OpenMP fundamentials

Показывать лекцию целиком

OpenMP is

The presentation can be downloaded here.

  • An application programming interface (API) that supports shared-memory programming for C/C++ and Fortran
  • Pros:
  • Simple
  • Cross-platform
  • Small overhead
  • Data parallelism support
    •

    Usage

    Compiler directives:

    C/C++

    #pragma omp directive [clause, …]

    Fortran

    !$OMP directive [clause, …]
C$OMP directive [clause, …]
*$OMP directive [clause, …]

    Parallel execution

    Parallel Regions

    Main OpenMP directive

    #pragma omp parallel

    #pragma omp parallel
{
  printf( "hello world from thread %d of %d\n",
    omp_get_thread_num(),
    omp_get_num_threads() );
}

    Параллельное исполнение

  • Most of the OpenMP instructions are preprocessor directives
  • Main construction is "omp parallel [smth]"
    •  (рис 7.1)

    OpenMP parallel model

  • Memory is shared
  • Task is divided into the threads
  • Variables can be
  • shared by the threads
  • private, available only for one thread
    •
  • Uncareful or wrong variable usage can lead to wrong execution results.

    • Fork-join model

  • Program execution starts from the master thread
  • With OpenMP directive master thread creates the additional threads
  • After the parallel region is finished all threads are synchronized
  • Main thread continues to execute the sequential part
    •  (рис 7.2)

    Основные конструкции OpenMP

  • #pragma omp for Each thread gets its own amount of data – data parallelism
  • #pragma omp section Each section will be executed in a separate thread – functional parallelism
  • #pragma omp single Sequential execution. Only one thread will execute this code
    •

    OpenMP sections

    #pragma omp sections [ clause [ clause ] ... ] new-line
{
[#pragma omp section new-line ]
structured-block1
[#pragma omp section new-line
structured-block2 ]
}

    Functional Parallelism

    #pragma omp parallel
#pragma omp sections nowait 
{
  thread1_work();
#pragma omp section
  thread2_work();
#pragma omp section
  thread3_work();
#pragma omp section
  thread4_work(); 
}

    OpenMP for directive

    #pragma omp for [ clause [ clause ] ...

    Following loop will be executed in parallel (the iterations will be divided by the execution threads)

    #pragma omp parallel private(f)
{
  f=7;
#pragma omp for
   for (i=0; i<20; i++)
   a[i] = b[i] + f * (i+1);
} /* omp end parallel */
    (рис 7.3)

    Available definitions:

      private( list ) 
  reduction( operator: list ) 
  schedule( type [ , chunk ] )
  nowait (для #pragma omp for)

    At the end of the loop all threads will be synchronized unless"nowait" directive is mentioned

    schedule defines iteration space scattering method (default behaviour depend on OpenMP version)

    OpenMP variables

  • private ( list ) Each of the listed variables will have the local copy for each exection thread
  • shared ( list ) All the thread will share the same instance of the variable
  • firstprivate ( list ) All the local copies will be initialized by master thread value
  • lastprivate ( list ) The resulting master thread value will be taken from the last thread executed
  • …
    •  (рис 7.4)

    All the variables are shared by default, except the local variables inside a function calls and the loop iterators

    Example

    int x;
x = 0;                           // Initialize x to zero
#pragma omp parallel for firstprivate(x)   // Copy value
       // of x
                                 // from master
for (i = 0; i < 10000; i++) {
     x = x + i;
}
printf( "x is %d\n", x );     // Print out value of x
/* Actually needs lastprivate(x) to copy value back out to master */

    OpenMP schedule clause

    schedule( type [ , chunk ] )

  • static: Every thread gets fixed amount of data
  • dynamic: Amount of data will depend on the thread execution speed
  • guided: Threads will get decreased amounts of data dynamically
  • runtime: Schedule type will be defined at runtime
    •  (рис 7.5) Loop scheduling

    Main OpenMP functions

    int omp_get_num_threads(void);
int omp_get_thread_num(void); 
…

    http://www.openmp.org/

    OpenMP synchronization

    Implicit sunchrionization is performed at the end of any parallel section (unless nowait option is mentioned)

  • сritical – can be executed only by one thread at a time.
  • atomic – Special critical section version for the atomic operations
  • barrier – synchronization point
  • ordered – sequential execution
  • master – only the main thread will execute the following code
  • …
    •

    OpenMP critical

    cnt = 0;
f=7;
#pragma omp parallel
{
#pragma omp for
    for (i=0; i<20; i++) {
      if (b[i] == 0) {
          #pragma omp critical
         cnt ++;
     } /* endif */
  a[i] = b[i] + f * (i+1);
  } /* end for */
} /*omp end parallel */
    (рис 7.6)

    More information

    OpenMP Homepage: http://www.openmp.org/

    Introduction to OpenMP - tutorial from WOMPEI 2000 http://www.llnl.gov/computing/tutorials/workshops/workshop/openMP

    Writing and Tuning OpenMP Programs on Distributed Shared Memory Machines http://www.sao.nrc.ca/~gabriel/openmp

    R.Chandra, L. Dagum, D. Kohr, D. Maydan, J. McDonald, R. Menon:

    Parallel programming in OpenMP.

    Academic Press, San Diego, USA, 2000, ISBN 1-55860-671-8

    R. Eigenmann, Michael J. Voss (Eds):

    OpenMP Shared Memory Parallel Programming.

    Springer LNCS 2104, Berlin, 2001, ISBN 3-540-42346-X

    Предыдущее занятие Следующее занятие
    Вернуться к учебному плану
    ^