Directiva Taskloop

Anterior

Siguiente

Esta directiva se usa para emitir tareas basadas en un bucle. El cuerpo del bucle constituye el cuerpo de cada tarea emitida. Por defecto, cada iteración es una tarea.

Sintaxis básica

#pragma omp parallel
{ 
    <parallel region code>

    #pragma omp single
    {
        #pragma omp taskloop
        for(...)
        {
            <loop body>
        } // implicit barrier here

        #pragma omp taskloop nogroup
        for(...)
        {
            <loop body>
        } // no barrier here
    } // implicit barrier here

    <parallel region code>
} // implicit barrier here

Ejemplo 1

El siguiente código va a crear tareas de manera secuencial:

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

int main(){
    int m = omp_get_thread_num();
    printf("thread=%d in serial region\n", m);
    
    #pragma omp parallel 
    {
        #pragma omp single
        {
            int s = omp_get_thread_num();
            printf("thread=%d in single region is creating tasks\n", s);
            
            #pragma omp taskloop
            for(int i=0; i<16; i++)
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed iteration %d\n", t, i);
                
            } // implicit barrier here

            #pragma omp task
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed final task\n", t);
            }
        } // implicit barrier here
    } // implicit barrier here
    
    m = omp_get_thread_num();
    printf("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 taskloop-000 taskloop-000.c

¿Cómo ejecutar?

./taskloop-000

Salida

thread=0 in serial region
thread=0 in single region is creating tasks
thread=1 executed iteration 0
thread=0 executed iteration 15
thread=0 executed iteration 14
thread=2 executed iteration 1
thread=2 executed iteration 4
thread=2 executed iteration 5
thread=2 executed iteration 6
thread=2 executed iteration 7
thread=2 executed iteration 8
thread=1 executed iteration 3
thread=3 executed iteration 2
thread=3 executed iteration 11
thread=3 executed iteration 12
thread=0 executed iteration 13
thread=2 executed iteration 9
thread=1 executed iteration 10
thread=3 executed final task
thread=0 in serial region

Se crean 16 tareas, una por cada iteración, asignadas al primer hilo disponible en ese momento. La tarea final siempre se ejecuta después del taskloop, dado la barrera implícita.

Ejemplo 2

El siguiente código va a crear tareas de manera secuencial:

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

int main(){
    int m = omp_get_thread_num();
    printf("thread=%d in serial region\n", m);
    
    #pragma omp parallel 
    {
        #pragma omp single
        {
            int s = omp_get_thread_num();
            printf("thread=%d in single region is creating tasks\n", s);
            
            #pragma omp taskloop nogroup
            for(int i=0; i<16; i++)
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed iteration %d\n", t, i);
                
            } // no barrier here anymore

            #pragma omp task
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed final task\n", t);
            }
        } // implicit barrier here
    } // implicit barrier here
    
    m = omp_get_thread_num();
    printf("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 taskloop-001 taskloop-001.c

¿Cómo ejecutar?

./taskloop-001

Salida

thread=0 in serial region
thread=0 in single region is creating tasks
thread=3 executed iteration 0
thread=2 executed iteration 1
thread=3 executed iteration 3
thread=0 executed final task
thread=2 executed iteration 4
thread=3 executed iteration 5
thread=3 executed iteration 7
thread=1 executed iteration 2
thread=1 executed iteration 9
thread=1 executed iteration 10
thread=1 executed iteration 11
thread=1 executed iteration 12
thread=3 executed iteration 8
thread=3 executed iteration 14
thread=2 executed iteration 6
thread=1 executed iteration 13
thread=0 executed iteration 15
thread=0 in serial region

Se crean 16 tareas, una por cada iteración, asignadas al primer hilo disponible en ese momento. La tarea final se ejecuta asíncronamente junto con las del taskloop, dado la cláusula nogroup.

Ejemplo 3

El siguiente código va a crear tareas de manera secuencial:

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

int main(){
    int m = omp_get_thread_num();
    printf("thread=%d in serial region\n", m);
    
    #pragma omp parallel 
    {
        #pragma omp single
        {
            int s = omp_get_thread_num();
            printf("thread=%d in single region is creating tasks\n", s);
            
            #pragma omp taskloop num_tasks(3)
            for(int i=0; i<16; i++)
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed iteration %d\n", t, i);
                
            } // implicit barrier here
        } // implicit barrier here
    } // implicit barrier here
    
    m = omp_get_thread_num();
    printf("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 taskloop-002 taskloop-002.c

¿Cómo ejecutar?

./taskloop-002

Salida

thread=0 in serial region
thread=2 in single region is creating tasks
thread=0 executed iteration 0
thread=1 executed iteration 6
thread=1 executed iteration 7
thread=1 executed iteration 8
thread=1 executed iteration 9
thread=1 executed iteration 10
thread=2 executed iteration 11
thread=2 executed iteration 12
thread=2 executed iteration 13
thread=2 executed iteration 14
thread=2 executed iteration 15
thread=0 executed iteration 1
thread=0 executed iteration 2
thread=0 executed iteration 3
thread=0 executed iteration 4
thread=0 executed iteration 5
thread=0 in serial region

Se crean solo 3 tareas al tiempo, una por cada iteración, asignadas al primer hilo disponible en ese momento.

Ejemplo 4

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

int main(){
    int m = omp_get_thread_num();
    printf("thread=%d in serial region\n", m);
    
    #pragma omp parallel 
    {
        #pragma omp single
        {
            int s = omp_get_thread_num();
            printf("thread=%d in single region is creating tasks\n", s);
            
            #pragma omp taskloop grainsize(4)
            for(int i=0; i<16; i++)
            {
                int t = omp_get_thread_num();
                printf("thread=%d executed iteration %d\n", t, i);
                
            } // implicit barrier here
        } // implicit barrier here
    } // implicit barrier here
    
    m = omp_get_thread_num();
    printf("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 taskloop-003 taskloop-003.c

¿Cómo ejecutar?

./taskloop-003

Salida

thread=0 in serial region
thread=0 in single region is creating tasks
thread=1 executed iteration 0
thread=1 executed iteration 1
thread=1 executed iteration 2
thread=1 executed iteration 3
thread=2 executed iteration 4
thread=2 executed iteration 5
thread=2 executed iteration 6
thread=2 executed iteration 7
thread=0 executed iteration 12
thread=0 executed iteration 13
thread=0 executed iteration 14
thread=0 executed iteration 15
thread=3 executed iteration 8
thread=3 executed iteration 9
thread=3 executed iteration 10
thread=3 executed iteration 11
thread=0 in serial region

Se crean 4 tareas de 4 iteraciones cada una, asignadas al primer hilo disponible en ese momento.

Preguntas

• Experimenta con diferentes valores de grainsize y num_tasks

Anterior

Siguiente

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