void tile() {
    uint32_t dim = 2;
    uint32_t M = dim;
    uint32_t N = dim;
    uint32_t K = 3;
    uint32_t group_size = 1;

    PackedArrayImpl a("arrA", 32, M*K);
    PackedArrayImpl b("arrB", 32, K*N);

    for (size_t i = 0; i < M; i++) {
        for (size_t j = 0; j < K; j++) {
            a.set(i*M + j, (i*M + j));
        }
    }

    for (size_t i = 0; i < K; i++) {
        for (size_t j = 0; j < N; j++) {
            b.set(i*M + j, (20 + i*K + j));
        }
    }


    ifstream infile{ "../samples/fydp-tune/multi_tile.cl" };
    string kernel{ istreambuf_iterator<char>(infile), istreambuf_iterator<char>() };
    string headerA = a.getConfig().generateOpenCLCode(true, group_size);
    string headerB = b.getConfig().generateOpenCLCode(true, group_size);
    string full_kernel = headerA + headerB + kernel;

    cout<<full_kernel<<endl;

    int32_t *cellA = a.getCells();
    int32_t *cellB = b.getCells();
    std::vector<int32_t> arrA(cellA, cellA+a.physical_capacity());
    std::vector<int32_t> arrB(cellB, cellB+b.physical_capacity());
    std::vector<int32_t> target(M*N);
    std::vector<int32_t> arrC(M*N);

    for (size_t i = 0; i < M; i++) {
        for (size_t j = 0; j < N; j++) {
            target[i*N + j] = 0;
            for (size_t k = 0; k < K; k++) {
                target[i*N+ j] += a.get(i*K+k) * b.get(k*N+j);
                //cout << "a[" << i << "][" << k << "] = " << a.get(i*K+k);
                //cout << " * b[" << k << "][" << j << "] = " << b.get(k*N+j) << endl;
            }
        }
    }

    for (auto x : target) cout << x << " ";
    cout << endl;
    cout << endl;
    for (auto x : arrA) cout << x << " ";
    cout << endl;
    for (auto x : arrB) cout << x << " ";
    cout << endl;

    cltune::Tuner tuner(size_t{0}, size_t{0});
     tuner.AddKernelFromString(full_kernel, "multi_tile", {M, N}, {group_size, group_size});
    tuner.SetReference({"../samples/fydp-tune/ref.cl"}, "ref", {M, N}, {group_size, group_size});

    tuner.AddArgumentScalar((int)M);
    tuner.AddArgumentScalar((int)N);
    tuner.AddArgumentScalar((int)K);
    tuner.AddArgumentInput(arrA);
    tuner.AddArgumentInput(arrB);
    tuner.AddArgumentInput(target);
    tuner.AddArgumentOutput(arrC);
    tuner.SetNumRuns(1);


    // Starts the tuner
    tuner.Tune();

}