for(int stride = 1; stride <= BLOCK_SIZE; stride *= 2) {
        int i = (t + 1) * stride * 2 - 1;

        if(i < BLOCK_SIZE << 1)
            scan_array[i] += scan_array[i - stride];

        __syncthreads();
    }

    if(t == 0)
        out[blockIdx.x] = scan_array[2 * BLOCK_SIZE - 1];