#include <stdio.h>
#include <stdlib.h>

int main(int argc, char** argv) {

    /***********************  FPGA  *************************/
    int workers, gangs, sizeF;

    workers = 16;
    gangs = 256;
    sizeF = workers * gangs;

    float *D, *E;
    int j;
    int sum = 0;
    D = (float*) malloc(sizeF * sizeof(float));
    E = (float*) malloc(sizeF * sizeof(float));

    for (j = 0; j < sizeF; j++) {
        D[j] = (float) j;
    }

// #pragma acc data copyin(D[0:sizeF]) copyout(sum)
#pragma acc data copyin(D[0:sizeF]) copyout(E[0:sizeF])
{
#pragma acc kernels
{
#pragma acc loop independent
    for (j = 0; j < sizeF; j++) {
        E[j] = D[j];
    }
}
}

//    for (j = 0; j < sizeF; j++) {
//        sum += E[j];
//    }

    for (j = 0; j < sizeF; j++) {
        if (D[j] != E[j]) sum++;
    }
    printf("errorFPGA:%d\n", sum);

    // printf("sum:%f\n", sum);
    //printf("sum:%f\n", sum);
    /********************************************************/

    return 0;
}