categorical_type::categorical_type(const nd::array& categories, bool presorted)
    : base_type(categorical_type_id, custom_kind, 4, 4, type_flag_scalar, 0, 0, 0)
{
    intptr_t category_count;
    if (presorted) {
        // This is construction shortcut, for the case when the categories are already
        // sorted. No validation of this is done, the caller should have ensured it
        // was correct already, typically by construction.
        m_categories = categories.eval_immutable();
        m_category_tp = m_categories.get_type().at(0);

        category_count = categories.get_dim_size();
        m_value_to_category_index.resize(category_count);
        m_category_index_to_value.resize(category_count);
        for (size_t i = 0; i != (size_t)category_count; ++i) {
            m_value_to_category_index[i] = i;
            m_category_index_to_value[i] = i;
        }

    } else {
        // Process the categories array to make sure it's valid
        const ndt::type& cdt = categories.get_type();
        if (cdt.get_type_id() != strided_dim_type_id) {
            throw dynd::type_error("categorical_type only supports construction from a strided array of categories");
        }
        m_category_tp = categories.get_type().at(0);
        if (!m_category_tp.is_scalar()) {
            throw dynd::type_error("categorical_type only supports construction from a 1-dimensional strided array of categories");
        }

        category_count = categories.get_dim_size();
        intptr_t categories_stride = reinterpret_cast<const strided_dim_type_arrmeta *>(categories.get_arrmeta())->stride;

        const char *categories_element_arrmeta = categories.get_arrmeta() + sizeof(strided_dim_type_arrmeta);
        comparison_ckernel_builder k;
        ::make_comparison_kernel(&k, 0,
                        m_category_tp, categories_element_arrmeta,
                        m_category_tp, categories_element_arrmeta,
                        comparison_type_sorting_less, &eval::default_eval_context);

        cmp less(k.get_function(), k.get());
        set<const char *, cmp> uniques(less);

        m_value_to_category_index.resize(category_count);
        m_category_index_to_value.resize(category_count);

        // create the mapping from indices of (to be lexicographically sorted) categories to values
        for (size_t i = 0; i != (size_t)category_count; ++i) {
            m_category_index_to_value[i] = i;
            const char *category_value = categories.get_readonly_originptr() +
                            i * categories_stride;

            if (uniques.find(category_value) == uniques.end()) {
                uniques.insert(category_value);
            } else {
                stringstream ss;
                ss << "categories must be unique: category value ";
                m_category_tp.print_data(ss, categories_element_arrmeta, category_value);
                ss << " appears more than once";
                throw std::runtime_error(ss.str());
            }
        }
        // TODO: Putting everything in a set already caused a sort operation to occur,
        //       there's no reason we should need a second sort.
        std::sort(m_category_index_to_value.begin(), m_category_index_to_value.end(),
                        sorter(categories.get_readonly_originptr(), categories_stride,
                            k.get_function(), k.get()));

        // invert the m_category_index_to_value permutation
        for (uint32_t i = 0; i < m_category_index_to_value.size(); ++i) {
            m_value_to_category_index[m_category_index_to_value[i]] = i;
        }

        m_categories = make_sorted_categories(uniques, m_category_tp,
                        categories_element_arrmeta);
    }

    // Use the number of categories to set which underlying integer storage to use
    if (category_count <= 256) {
        m_storage_type = ndt::make_type<uint8_t>();
    } else if (category_count <= 65536) {
        m_storage_type = ndt::make_type<uint16_t>();
    } else {
        m_storage_type = ndt::make_type<uint32_t>();
    }
    m_members.data_size = m_storage_type.get_data_size();
    m_members.data_alignment = (uint8_t)m_storage_type.get_data_alignment();
}
Exemple #2
0
ndt::categorical_type::categorical_type(const nd::array &categories, bool presorted)
    : base_type(categorical_id, 4, 4, type_flag_none, 0, 0, 0)
{
  intptr_t category_count;
  if (presorted) {
    // This is construction shortcut, for the case when the categories are
    // already
    // sorted. No validation of this is done, the caller should have ensured it
    // was correct already, typically by construction.
    m_categories = categories.eval_immutable();
    m_category_tp = m_categories.get_type().at(0);

    category_count = categories.get_dim_size();
    m_value_to_category_index = nd::range(category_count);
    m_value_to_category_index.flag_as_immutable();
    m_category_index_to_value = m_value_to_category_index;
  }
  else {
    // Process the categories array to make sure it's valid
    const type &cdt = categories.get_type();
    if (cdt.get_id() != fixed_dim_id) {
      throw dynd::type_error("categorical_type only supports construction from "
                             "a fixed-dim array of categories");
    }
    m_category_tp = categories.get_type().at(0);
    if (!m_category_tp.is_scalar()) {
      throw dynd::type_error("categorical_type only supports construction from "
                             "a 1-dimensional strided array of categories");
    }

    category_count = categories.get_dim_size();
    intptr_t categories_stride = reinterpret_cast<const fixed_dim_type_arrmeta *>(categories.get()->metadata())->stride;

    const char *categories_element_arrmeta = categories.get()->metadata() + sizeof(fixed_dim_type_arrmeta);
    nd::kernel_builder k;
    kernel_single_t fn = k.get()->get_function<kernel_single_t>();

    cmp less(fn, k.get());
    set<const char *, cmp> uniques(less);

    m_value_to_category_index = nd::empty(category_count, make_type<intptr_t>());
    m_category_index_to_value = nd::empty(category_count, make_type<intptr_t>());

    // create the mapping from indices of (to be lexicographically sorted)
    // categories to values
    for (size_t i = 0; i != (size_t)category_count; ++i) {
      unchecked_fixed_dim_get_rw<intptr_t>(m_category_index_to_value, i) = i;
      const char *category_value = categories.cdata() + i * categories_stride;

      if (uniques.find(category_value) == uniques.end()) {
        uniques.insert(category_value);
      }
      else {
        stringstream ss;
        ss << "categories must be unique: category value ";
        m_category_tp.print_data(ss, categories_element_arrmeta, category_value);
        ss << " appears more than once";
        throw std::runtime_error(ss.str());
      }
    }
    // TODO: Putting everything in a set already caused a sort operation to
    // occur,
    //       there's no reason we should need a second sort.
    std::sort(&unchecked_fixed_dim_get_rw<intptr_t>(m_category_index_to_value, 0),
              &unchecked_fixed_dim_get_rw<intptr_t>(m_category_index_to_value, category_count),
              sorter(categories.cdata(), categories_stride, fn, k.get()));

    // invert the m_category_index_to_value permutation
    for (intptr_t i = 0; i < category_count; ++i) {
      unchecked_fixed_dim_get_rw<intptr_t>(m_value_to_category_index,
                                           unchecked_fixed_dim_get<intptr_t>(m_category_index_to_value, i)) = i;
    }

    m_categories = make_sorted_categories(uniques, m_category_tp, categories_element_arrmeta);
  }

  // Use the number of categories to set which underlying integer storage to use
  if (category_count <= 256) {
    m_storage_type = make_type<uint8_t>();
  }
  else if (category_count <= 65536) {
    m_storage_type = make_type<uint16_t>();
  }
  else {
    m_storage_type = make_type<uint32_t>();
  }
  this->data_size = m_storage_type.get_data_size();
  this->data_alignment = (uint8_t)m_storage_type.get_data_alignment();
}