//#define _BSD_SOURCE
#pragma offload_attribute(push,target(mic))
#include <stdio.h>
#pragma offload_attribute(pop)
#include <stdlib.h>
#include <unistd.h>
#include <malloc.h>
#include <pthread.h>
#include <time.h>

#define Nwords   (2*4096)

struct ctx {
	int *data;
	int tid;
	int ntid;
	int sleep_us;
};

static void *offload_thread(void *data)
{
	struct ctx *c = (struct ctx *)data;
	int tid = c->tid;
	int ntid = c->ntid;
	int *mybit = c->data;
	int Nsleep_us = c->sleep_us;

	long a = 0;

        while (a != 7)  {
		_Pragma("offload target(mic:0) in ( mybit : length(0) alloc_if(0) free_if(0))")
		{
			a = 0;
			for (int j=0; j<Nwords; j++) {
				mybit[j] = rand( )%Nwords;
				a += (long)mybit[j];
			}
		}
                usleep(Nsleep_us);
	}

	_Pragma("offload target(mic:0) out ( mybit : length(Nwords) alloc_if(0) free_if(1))")
	{ }

	printf("tid %d my bit = %d", tid, mybit[Nwords/2]);
        return NULL;
}


int main(int nargs, char **argv)
{
        int NTHREADS = atoi(argv[1]);
	int sleep_us = atoi(argv[2]);
        int i;
        pthread_t off_thread[NTHREADS];
	struct ctx *c = malloc(NTHREADS*sizeof(struct ctx));

	for (int i=0; i < NTHREADS; i++ ) {
		c[i].tid = i;
		c[i].ntid = NTHREADS;
		c[i].sleep_us = sleep_us;
		int *data = memalign(sysconf(_SC_PAGESIZE), Nwords*sizeof(int));
		_Pragma("offload_transfer target(mic:0) in ( data : length(Nwords) alloc_if(1) free_if(0))");
		c[i].data = data;

		pthread_create ( &off_thread[i], NULL, offload_thread, &c[i] );
	}

	for (int i=0; i < NTHREADS; i++ )
		pthread_join(off_thread[i], NULL);
}