Replacing 3 syncs with one predicated sync.

synchronization.cu

__global__ void three_syncs(int *data, int size, int thresh)
{
    __shared__ int flag;
    int tid = threaIdx.x;
    do {
        __syncthreads();
        if (!tid)
            flag = 0;
        __syncthreads();
        if (tid < size) {
            data[tid]--;
            // ...stuff
            if (data[tid] > thresh)
                flag = 1;
        }
        __syncthreads();
    } while (flag);
}

__global__ void one_sync_or(int *data, int size, int thresh)
{
    int flag;
    int tid = threadIdx.x;
    do {
        flag = 0;
        if (tid < size) {
            data[tid]--;
            // ..stuff
            if (data[tid] > thresh)
                flag = 1
        }
        flag = __syncthreads_or(flag);
    } while (flag);
};