auto to_dense_column_major_naive(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    for (size_t i = 0; i < n_col; ++i)
        for (size_t j = 0; j < n_row; ++j)
            out_vec[i * n_row + j] = vec[i][j];
    return out_vec;
}

auto to_dense_row_major_blocking(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    size_t block_side = 8;

    for (size_t l = 0; l < n_col; l += block_side) {
        for (size_t k = 0; k < n_row; k += block_side) {
            for (size_t j = l; j < l + block_side && j < n_col; ++j) {
                auto const &column = vec[j];
                for (size_t i = k; i < k + block_side && i < n_row; ++i)
                    out_vec[i * n_col + j] = column[i];
            }
        }
    }
    return out_vec;
}

auto to_dense_column_major_naive(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    for (size_t i = 0; i < n_col; ++i)
        for (size_t j = 0; j < n_row; ++j)
            out_vec[i * n_row + j] = vec[i][j];
    return out_vec;
}

auto to_dense_row_major_naive(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    for (size_t i = 0; i < n_col; ++i)
        for (size_t j = 0; j < n_row; ++j)
            out_vec[j * n_col + i] = vec[i][j];
    return out_vec;
}

auto to_dense_row_major_blocking(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    size_t block_side = 8;

    for (size_t l = 0; l < n_col; l += block_side) {
        for (size_t k = 0; k < n_row; k += block_side) {
            for (size_t j = l; j < l + block_side && j < n_col; ++j) {
                auto const &column = vec[j];
                for (size_t i = k; i < k + block_side && i < n_row; ++i)
                    out_vec[i * n_col + j] = column[i];
            }
        }
    }
    return out_vec;
}

auto to_dense_column_major_blocking(std::vector<std::vector<double>> const & vec)
    -> std::vector<double>
{
    auto n_col = vec.size();
    auto n_row = vec[0].size();
    std::vector<double> out_vec(n_col * n_row);
    size_t block_side = 8;

    for (size_t l = 0; l < n_col; l += block_side) {
        for (size_t k = 0; k < n_row; k += block_side) {
            for (size_t j = l; j < l + block_side && j < n_col; ++j) {
                auto const &column = vec[j];
                for (size_t i = k; i < k + block_side && i < n_row; ++i)
                    out_vec[j * n_row + i] = column[i];
            }
        }
    }
    return out_vec;
}

auto make_vecvec() -> std::vector<std::vector<double>>
{
  std::vector<std::vector<double>> vecvec(50, std::vector<double>(4000));
  std::mt19937 mersenne {2019};
  std::uniform_real_distribution<double> dist(-1000, 1000);
  for (auto &vec: vecvec)
   for (auto &val: vec)
       val = dist(mersenne);
  return vecvec;
}

static void NaiveColumnMajor(benchmark::State& state) {
  // Code before the loop is not measured

  auto vecvec = make_vecvec();
  for (auto _ : state) {
    benchmark::DoNotOptimize(to_dense_column_major_naive(vecvec));
  }
}
BENCHMARK(NaiveColumnMajor);

static void NaiveRowMajor(benchmark::State& state) {
  // Code before the loop is not measured

  auto vecvec = make_vecvec();
  for (auto _ : state) {
    benchmark::DoNotOptimize(to_dense_row_major_naive(vecvec));
  }
}
BENCHMARK(NaiveRowMajor);

static void BlockingRowMajor(benchmark::State& state) {
  // Code before the loop is not measured

  auto vecvec = make_vecvec();
  for (auto _ : state) {
    benchmark::DoNotOptimize(to_dense_row_major_blocking(vecvec));
  }
}
BENCHMARK(BlockingRowMajor);

static void BlockingColumnMajor(benchmark::State& state) {
  // Code before the loop is not measured

  auto vecvec = make_vecvec();
  for (auto _ : state) {
    benchmark::DoNotOptimize(to_dense_column_major_blocking(vecvec));
  }
}
BENCHMARK(BlockingColumnMajor);