atomic_math.c

MPI stands for the Message Passing Interface. Written by the MPI Forum (a large committee comprising

/*
 * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
 *                         University Research and Technology
 *                         Corporation.  All rights reserved.
 * Copyright (c) 2004-2005 The University of Tennessee and The University
 *                         of Tennessee Research Foundation.  All rights
 *                         reserved.
 * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
 *                         University of Stuttgart.  All rights reserved.
 * Copyright (c) 2004-2005 The Regents of the University of California.
 *                         All rights reserved.
 * $COPYRIGHT$
 * 
 * Additional copyrights may follow
 * 
 * $HEADER$
 */

#define OMPI_BUILDING 0
#include "ompi_config.h"

#ifdef HAVE_PTHREAD_H
#include <pthread.h>
#endif
#include <stdlib.h>
#include <stdio.h>

#include "opal/sys/atomic.h"

#define TEST_REPS 500

int32_t val32 = 0;
#if OPAL_HAVE_ATOMIC_MATH_64
int64_t val64 = 0;
#endif
int valint = 0;

static void* atomic_math_test(void* arg)
{
    int count = *((int*) arg);
    int i;

    for (i = 0 ; i < count ; ++i) {
        opal_atomic_add_32(&val32, 5);
#if OPAL_HAVE_ATOMIC_MATH_64
        opal_atomic_add_64(&val64, 6);
#endif
        opal_atomic_add(&valint, 4);
    }

    return NULL;
}


static int
atomic_math_test_th(int count, int thr_count)
{
    int value;
#if OMPI_HAVE_POSIX_THREADS
    pthread_t *th;
    int tid, ret = 0;

    th = (pthread_t *) malloc(thr_count * sizeof(pthread_t));
    if (!th) { 
        perror("malloc"); 
        exit(EXIT_FAILURE); 
    }

    /* Ok to use a single instance of ip_union_t from the stack here
       because a) we're passing the same count to all threads, and b)
       we're waiting for all the threads to finish before leaving this
       function, so there's no race condition of the instance
       disappearing before the threads start. */
    value = count;
    for (tid = 0; tid < thr_count; tid++) {
        if (pthread_create(&th[tid], NULL, atomic_math_test, 
                           (void*) &value) != 0) {
            perror("pthread_create");
            exit(EXIT_FAILURE);
        }
    }

    /* -- wait for the thread set to finish -- */
    for (tid = 0; tid < thr_count; tid++) {
        void *thread_return;

        if (pthread_join(th[tid], &thread_return) != 0) {
            perror("pthread_join");
            exit(EXIT_FAILURE);
        }
    }
    free(th);

    return ret;
#else
    value = count;
    if (thr_count == 1) {
        atomic_math_test((void*) &value);
    } else {
        return 77;
    }
#endif

    return 0;
}


int
main(int argc, char *argv[])
{
    int ret = 77;
    int num_threads = 1;

    if (argc != 2) {
        printf("*** Incorrect number of arguments.  Skipping test\n");
        return 77;
    }
    num_threads = atoi(argv[1]);

    ret = atomic_math_test_th(TEST_REPS, num_threads);
    if (ret == 77) return ret;
    opal_atomic_mb();
    if (val32 != TEST_REPS * num_threads * 5) {
        printf("opal_atomic_add32 failed.  Expected %d, got %d.\n",
               TEST_REPS * num_threads * 5, val32);
        ret = 1;
    }
#if OPAL_HAVE_ATOMIC_MATH_64
    if (val64 != TEST_REPS * num_threads * 6) {
        /* Safe to case to (int) here because we know it's going to be
           a small value */
        printf("opal_atomic_add32 failed.  Expected %d, got %d.\n",
               TEST_REPS * num_threads * 6, (int) val64);
        ret = 1;
    }
#else
    printf("      * skipping 64 bit tests\n");
#endif
    if (valint != TEST_REPS * num_threads * 4) {
        printf("opal_atomic_add32 failed.  Expected %d, got %d.\n",
               TEST_REPS * num_threads * 4, valint);
        ret = 1;
    }

    return ret;
}