/*
 * Multi-core evaluation.  Takes number of theads as an
 * optional argument.  On a 4-core device, run with "1" and
 * again with "4".  If the times are about the same, all
 * cores were used.  If the 4-core run takes twice as long,
 * only two cores were used.
 *
 * The threshold at which the kernel migrates a thread to
 * a new core is system-specific.  You can try to force it
 * with sched_getaffinity().
 *
 * Not useful as a benchmark.
 *
 * desktop: gcc -g -Wall spinner.c -o spinner -lpthread
 */
#include <stdlib.h>
#include <stdint.h>
#include <stdio.h>
#include <pthread.h>
#include <sys/time.h>

#define _GNU_SOURCE
#include <sched.h>

#define ONE_MILLION 1000000LL

// This needs to be long enough for the kernel to decide
// that the thread should be on its own core.
#define SPIN_COUNT 20000

void* spin(void* vtid)
{
    int tid = (int)(uintptr_t) vtid;
    int reps = SPIN_COUNT + tid;
    int ret = 0;
    int i, j;

    for (i = 0; i < reps; i++) {
        for (j = 0; j < 100000; j++) {
            ret += i * j;
        }
    }

    // Print the number of the core we're currently on.
    // This isn't all that useful, because threads can
    // move between cores, so even if all 4 cores were
    // used you might not see all 4 represented in the
    // output.  OTOH, if you do see 4 different core
    // numbers you know they were all used.
    printf("Thread %d done, was on core %d\n",
        tid, sched_getcpu());

    return (void*)(uintptr_t) ret;
}

int main(int argc, char** argv)
{
    int numThreads = 4;

    if (argc == 2) {
        numThreads = atoi(argv[1]);
        if (numThreads < 0 || numThreads > 256) {
            fprintf(stderr,
                "ERROR: invalid numThreads=%d\n", numThreads);
            return 2;
        }
    }

    printf("Starting %d threads\n", numThreads);
    struct timeval tv;
    gettimeofday(&tv, NULL);
    long long startWhen = (tv.tv_sec * ONE_MILLION) + tv.tv_usec;

    pthread_t threads[numThreads];
    int i;
    for (i = 0; i < numThreads; i++) {
        int cc = pthread_create(&threads[i], NULL, spin,
            (void*)(uintptr_t) i);
        if (cc != 0) {
            fprintf(stderr,
                "ERROR: pthread_create %d failed: %d\n", i, cc);
            return 1;
        }
    }

    for (i = 0; i < numThreads; i++) {
        int cc = pthread_join(threads[i], NULL);
        if (cc != 0) {
            fprintf(stderr,
                "ERROR: pthread_join %d failed: %d\n", i, cc);
            return 1;
        }
    }

    gettimeofday(&tv, NULL);
    long long endWhen = (tv.tv_sec * ONE_MILLION) + tv.tv_usec;

    printf("All threads finished in %lldms\n",
        (endWhen - startWhen) / 1000);

    return 0;
}