Directiva for

Anterior

Siguiente

Los bucles dentro de una región paralela pueden beneficiarse de la existencia de múltiples hilos. Cada hilo se puede encargar de determinadas iteraciones del bucle, y de manera mancomunada con los demás hilos, acelerar la ejecución de estos bloques de código. Los hilos se sincronizan al finalizar el bucle. En efecto, la manera más obvia de acelerar la ejecución de un programa es paralelizar sus bucles.

Sintaxis básica

#pragma omp parallel
{
    <parallel region code>
    
    #pragma omp for 
    for(...){
        <loop body code>
    } // implicit barrier here

    <parallel region code>
} // implicit barrier here

Ejemplo 1

Vamos a distribuir las iteraciones del bucle entre los hilos disponibles:

#include <stdio.h>
#include <omp.h>

int main(){
    int m = omp_get_thread_num();
    printf("master thread=%d in serial region\n", m);

    #pragma omp parallel
    {
        int t = omp_get_thread_num();
        printf("thread=%d in parallel region\n", t);
        
        #pragma omp for
        for(int i=0; i<16 ; i++){
            int f = omp_get_thread_num();
            printf("id=%d processed by thread=%d\n", i, f);
        } // implicit barrier here
    
        printf("thread=%d says goodbye\n", t);
    } // implicit barrier here
    
    printf("master thread=%d in serial region\n", m);
}

Descarga el código aquí.

¿Cómo compilar?

clang -fopenmp -I/home/user/llvm/llvm-build/projects/openmp/runtime/src/ 
     -o for-000 for-000.c

¿Cómo ejecutar?

./for-000 

Salida

master thread=0 in serial region
thread=2 in parallel region
thread=1 in parallel region
thread=3 in parallel region
id=12 processed by thread=3
id=13 processed by thread=3
id=14 processed by thread=3
id=15 processed by thread=3
id=4 processed by thread=1
id=5 processed by thread=1
id=6 processed by thread=1
id=7 processed by thread=1
thread=0 in parallel region
id=0 processed by thread=0
id=1 processed by thread=0
id=2 processed by thread=0
id=3 processed by thread=0
id=8 processed by thread=2
id=9 processed by thread=2
id=10 processed by thread=2
id=11 processed by thread=2
thread=1 says goodbye
thread=3 says goodbye
thread=2 says goodbye
thread=0 says goodbye
master thread=0 in serial region

Cada hilo se hizo cargo de 4 de las 16 iteraciones. Las iteraciones de cada hilo fueron consecutivas.

Ejemplo 2

¿Qué pasa si no usamos la directiva for?:

#include <stdio.h>
#include <omp.h>

int main(){
    int m = omp_get_thread_num();
    printf("master thread=%d in serial region\n", m);

    #pragma omp parallel
    {
        int t = omp_get_thread_num();
        printf("thread=%d in parallel region\n", t);
        
        for(int i=0; i<16 ; i++){
            int f = omp_get_thread_num();
            printf("id=%d processed by thread=%d\n", i, f);
        }
    
        printf("thread=%d says goodbye\n", t);
    } // implicit barrier here
    
    printf("master thread=%d in serial region\n", m);
}

Descarga el código aquí.

¿Cómo compilar?

clang -fopenmp -I/home/user/llvm/llvm-build/projects/openmp/runtime/src/ 
     -o for-001 for-001.c

¿Cómo ejecutar?

./for-001

Salida

master thread=0 in serial region
thread=0 in parallel region
id=0 processed by thread=0
id=1 processed by thread=0
id=2 processed by thread=0
id=3 processed by thread=0
id=4 processed by thread=0
id=5 processed by thread=0
thread=2 in parallel region
id=0 processed by thread=2
id=1 processed by thread=2
id=2 processed by thread=2
id=3 processed by thread=2
id=4 processed by thread=2
id=5 processed by thread=2
id=6 processed by thread=2
id=7 processed by thread=2
id=8 processed by thread=2
id=9 processed by thread=2
id=10 processed by thread=2
id=11 processed by thread=2
id=12 processed by thread=2
id=13 processed by thread=2
id=14 processed by thread=2
id=15 processed by thread=2
thread=2 says goodbye
thread=1 in parallel region
id=0 processed by thread=1
id=1 processed by thread=1
id=2 processed by thread=1
id=3 processed by thread=1
id=4 processed by thread=1
id=5 processed by thread=1
id=6 processed by thread=1
id=7 processed by thread=1
id=8 processed by thread=1
id=9 processed by thread=1
id=10 processed by thread=1
id=11 processed by thread=1
id=12 processed by thread=1
id=13 processed by thread=1
id=14 processed by thread=1
id=15 processed by thread=1
thread=1 says goodbye
id=6 processed by thread=0
id=7 processed by thread=0
id=8 processed by thread=0
id=9 processed by thread=0
id=10 processed by thread=0
id=11 processed by thread=0
id=12 processed by thread=0
id=13 processed by thread=0
id=14 processed by thread=0
id=15 processed by thread=0
thread=0 says goodbye
thread=3 in parallel region
id=0 processed by thread=3
id=1 processed by thread=3
id=2 processed by thread=3
id=3 processed by thread=3
id=4 processed by thread=3
id=5 processed by thread=3
id=6 processed by thread=3
id=7 processed by thread=3
id=8 processed by thread=3
id=9 processed by thread=3
id=10 processed by thread=3
id=11 processed by thread=3
id=12 processed by thread=3
id=13 processed by thread=3
id=14 processed by thread=3
id=15 processed by thread=3
thread=3 says goodbye
master thread=0 in serial region

En este caso, todos los hilos ejecutaron el bucle de manera redundante.

Preguntas

• Modifica el programa de tal manera que se ejecuten dos bucles en lugar de uno.
• ¿Cómo harías para paralelizar el bucle interior también?

Anterior

Siguiente

Estás en el Nivel 2: Bucles paralelos y tareas concurrentes con OpenMP. ¿Deseas volver al inicio?