C++ array::get_arrmeta方法代码示例

static nd::array view_concrete(const nd::array &arr, const ndt::type &tp)
{
  // Allocate a result array to attempt the view in it
  nd::array result(make_array_memory_block(tp.get_arrmeta_size()));
  // Copy the fields
  result.get_ndo()->data.ptr = arr.get_ndo()->data.ptr;
  if (arr.get_ndo()->data.ref == NULL) {
    // Embedded data, need reference to the array
    result.get_ndo()->data.ref = arr.get_memblock().release();
  } else {
    // Use the same data reference, avoid producing a chain
    result.get_ndo()->data.ref = arr.get_data_memblock().release();
  }
  result.get_ndo()->m_type = ndt::type(tp).release();
  result.get_ndo()->m_flags = arr.get_ndo()->m_flags;
  // First handle a special case of viewing outermost "var" as "fixed[#]"
  if (arr.get_type().get_type_id() == var_dim_type_id && tp.get_type_id() == fixed_dim_type_id) {
    const var_dim_type_arrmeta *in_am = reinterpret_cast<const var_dim_type_arrmeta *>(arr.get_arrmeta());
    const var_dim_type_data *in_dat = reinterpret_cast<const var_dim_type_data *>(arr.get_readonly_originptr());
    fixed_dim_type_arrmeta *out_am = reinterpret_cast<fixed_dim_type_arrmeta *>(result.get_arrmeta());
    out_am->dim_size = tp.extended<ndt::fixed_dim_type>()->get_fixed_dim_size();
    out_am->stride = in_am->stride;
    if ((intptr_t)in_dat->size == out_am->dim_size) {
      // Use the more specific data reference from the var arrmeta if possible
      if (in_am->blockref != NULL) {
        memory_block_decref(result.get_ndo()->data.ref);
        memory_block_incref(in_am->blockref);
        result.get_ndo()->data.ref = in_am->blockref;
      }
      result.get_ndo()->data.ptr = in_dat->begin + in_am->offset;
      // Try to copy the rest of the arrmeta as a view
      if (try_view(arr.get_type().extended<ndt::base_dim_type>()->get_element_type(),
                   arr.get_arrmeta() + sizeof(var_dim_type_arrmeta),
                   tp.extended<ndt::base_dim_type>()->get_element_type(),
                   result.get_arrmeta() + sizeof(fixed_dim_type_arrmeta), arr.get_memblock().get())) {
        return result;
      }
    }
  }
  // Otherwise try to copy the arrmeta as a view
  else if (try_view(arr.get_type(), arr.get_arrmeta(), tp, result.get_arrmeta(), arr.get_memblock().get())) {
    // If it succeeded, return it
    return result;
  }

  stringstream ss;
  ss << "Unable to view nd::array of type " << arr.get_type();
  ss << " as type " << tp;
  throw type_error(ss.str());
}

nd::array dynd::format_json(const nd::array &n, bool struct_as_list)
{
  // Create a UTF-8 string
  nd::array result = nd::empty(ndt::string_type::make());

  // Initialize the output with some memory
  output_data out;
  out.out_string.resize(1024);
  out.out_begin = out.out_string.begin();
  out.out_capacity_end = out.out_string.end();
  out.out_end = out.out_begin;
  out.struct_as_list = struct_as_list;

  if (!n.get_type().is_expression()) {
    ::format_json(out, n.get_type(), n.get_arrmeta(), n.get_readonly_originptr());
  } else {
    nd::array tmp = n.eval();
    ::format_json(out, tmp.get_type(), tmp.get_arrmeta(), tmp.get_readonly_originptr());
  }

  // Shrink the memory to fit, and set the pointers in the output
  string *d = reinterpret_cast<string *>(result.get_readwrite_originptr());
  d->assign(out.out_string.data(), out.out_end - out.out_begin);

  // Finalize processing and mark the result as immutable
  result.get_type().extended()->arrmeta_finalize_buffers(result.get_arrmeta());
  result.flag_as_immutable();

  return result;
}

static nd::array array_function_dereference(const nd::array &self)
{
  // Follow the pointers to eliminate them
  ndt::type dt = self.get_type();
  const char *arrmeta = self.get_arrmeta();
  char *data = self.get_ndo()->m_data_pointer;
  memory_block_data *dataref = self.get_ndo()->m_data_reference;
  if (dataref == NULL) {
    dataref = self.get_memblock().get();
  }
  uint64_t flags = self.get_ndo()->m_flags;

  while (dt.get_type_id() == pointer_type_id) {
    const pointer_type_arrmeta *md =
        reinterpret_cast<const pointer_type_arrmeta *>(arrmeta);
    dt = dt.extended<ndt::pointer_type>()->get_target_type();
    arrmeta += sizeof(pointer_type_arrmeta);
    data = *reinterpret_cast<char **>(data) + md->offset;
    dataref = md->blockref;
  }

  // Create an array without the pointers
  nd::array result(make_array_memory_block(dt.get_arrmeta_size()));
  if (!dt.is_builtin()) {
    dt.extended()->arrmeta_copy_construct(result.get_arrmeta(), arrmeta,
                                          &self.get_ndo()->m_memblockdata);
  }
  result.get_ndo()->m_type = dt.release();
  result.get_ndo()->m_data_pointer = data;
  result.get_ndo()->m_data_reference = dataref;
  memory_block_incref(result.get_ndo()->m_data_reference);
  result.get_ndo()->m_flags = flags;
  return result;
}

static void json_as_buffer(const nd::array &json, nd::array &out_tmp_ref,
                           const char *&begin, const char *&end)
{
  // Check the type of 'json', and get pointers to the begin/end of a UTF-8
  // buffer
  ndt::type json_type = json.get_type().value_type();
  switch (json_type.get_kind()) {
  case string_kind: {
    const ndt::base_string_type *sdt =
        json_type.extended<ndt::base_string_type>();
    switch (sdt->get_encoding()) {
    case string_encoding_ascii:
    case string_encoding_utf_8:
      out_tmp_ref = json.eval();
      // The data is already UTF-8, so use the buffer directly
      sdt->get_string_range(&begin, &end, out_tmp_ref.get_arrmeta(),
                            out_tmp_ref.get_readonly_originptr());
      break;
    default: {
      // The data needs to be converted to UTF-8 before parsing
      ndt::type utf8_tp = ndt::string_type::make(string_encoding_utf_8);
      out_tmp_ref = json.ucast(utf8_tp).eval();
      sdt = static_cast<const ndt::base_string_type *>(utf8_tp.extended());
      sdt->get_string_range(&begin, &end, out_tmp_ref.get_arrmeta(),
                            out_tmp_ref.get_readonly_originptr());
      break;
    }
    }
    break;
  }
  case bytes_kind: {
    out_tmp_ref = json.eval();
    const ndt::base_bytes_type *bdt =
        json_type.extended<ndt::base_bytes_type>();
    bdt->get_bytes_range(&begin, &end, out_tmp_ref.get_arrmeta(),
                         out_tmp_ref.get_readonly_originptr());
    break;
  }
  default: {
    stringstream ss;
    ss << "Input for JSON parsing must be either bytes (interpreted as UTF-8) "
          "or a string, not \"" << json_type << "\"";
    throw runtime_error(ss.str());
    break;
  }
  }
}

void dynd::typed_data_assign(const ndt::type &dst_tp, const char *dst_arrmeta,
                             char *dst_data, const nd::array &src_arr,
                             const eval::eval_context *ectx)
{
  typed_data_assign(dst_tp, dst_arrmeta, dst_data, src_arr.get_type(),
                    src_arr.get_arrmeta(), src_arr.get_readonly_originptr(),
                    ectx);
}

 const char *get_category_data_from_value(size_t value) const {
     if (value >= get_category_count()) {
         throw std::runtime_error("category value is out of bounds");
     }
     return m_categories.get_readonly_originptr() +
            m_value_to_category_index[value] *
                reinterpret_cast<const strided_dim_type_arrmeta *>(
                    m_categories.get_arrmeta())->stride;
 }

 const char *get_category_data_from_value(uint32_t value) const
 {
   if (value >= get_category_count()) {
     throw std::runtime_error("category value is out of bounds");
   }
   return m_categories.get_readonly_originptr() +
          unchecked_fixed_dim_get<intptr_t>(m_value_to_category_index,
                                            value) *
              reinterpret_cast<const fixed_dim_type_arrmeta *>(
                  m_categories.get_arrmeta())->stride;
 }

uint32_t categorical_type::get_value_from_category(const nd::array& category) const
{
    if (category.get_type() == m_category_tp) {
        // If the type is right, get the category value directly
        return get_value_from_category(category.get_arrmeta(), category.get_readonly_originptr());
    } else {
        // Otherwise convert to the correct type, then get the category value
        nd::array c = nd::empty(m_category_tp);
        c.val_assign(category);
        return get_value_from_category(c.get_arrmeta(), c.get_readonly_originptr());
    }
}

static nd::array view_as_bytes(const nd::array &arr, const ndt::type &tp)
{
  if (arr.get_type().get_flags() & type_flag_destructor) {
    // Can't view arrays of object type
    return nd::array();
  }

  // Get the essential components of the array to analyze
  memory_block_ptr data_ref = arr.get_data_memblock();
  char *data_ptr = arr.get_ndo()->data.ptr;
  ndt::type data_tp = arr.get_type();
  const char *data_meta = arr.get_arrmeta();
  intptr_t data_dim_size = -1, data_stride = 0;
  // Repeatedly refine the data
  while (data_tp.get_type_id() != uninitialized_type_id) {
    refine_bytes_view(data_ref, data_ptr, data_tp, data_meta, data_dim_size, data_stride);
  }
  // Check that it worked, and that the resulting data pointer is aligned
  if (data_dim_size < 0 ||
      !offset_is_aligned(reinterpret_cast<size_t>(data_ptr), tp.extended<ndt::bytes_type>()->get_target_alignment())) {
    // This signals we could not view the data as a
    // contiguous chunk of bytes
    return nd::array();
  }

  char *result_data_ptr = NULL;
  nd::array result(make_array_memory_block(tp.extended()->get_arrmeta_size(), tp.get_data_size(),
                                           tp.get_data_alignment(), &result_data_ptr));
  // Set the bytes extents
  ((char **)result_data_ptr)[0] = data_ptr;
  ((char **)result_data_ptr)[1] = data_ptr + data_dim_size;
  // Set the array arrmeta
  array_preamble *ndo = result.get_ndo();
  ndo->m_type = ndt::type(tp).release();
  ndo->data.ptr = result_data_ptr;
  ndo->data.ref = NULL;
  ndo->m_flags = arr.get_flags();
  // Set the bytes arrmeta
  bytes_type_arrmeta *ndo_meta = reinterpret_cast<bytes_type_arrmeta *>(result.get_arrmeta());
  ndo_meta->blockref = data_ref.release();
  return result;
}

void dynd::parse_json(nd::array &out, const char *json_begin,
                      const char *json_end, const eval::eval_context *ectx)
{
  try {
    const char *begin = json_begin, *end = json_end;
    ndt::type tp = out.get_type();
    ::parse_json(tp, out.get_arrmeta(), out.get_readwrite_originptr(), begin,
                 end, ectx);
    begin = skip_whitespace(begin, end);
    if (begin != end) {
      throw json_parse_error(begin, "unexpected trailing JSON text", tp);
    }
  }
  catch (const json_parse_error &e) {
    stringstream ss;
    string line_prev, line_cur;
    int line, column;
    get_error_line_column(json_begin, json_end, e.get_position(), line_prev,
                          line_cur, line, column);
    ss << "Error parsing JSON at line " << line << ", column " << column
       << "\n";
    ss << "DyND Type: " << e.get_type() << "\n";
    ss << "Message: " << e.what() << "\n";
    print_json_parse_error_marker(ss, line_prev, line_cur, line, column);
    throw invalid_argument(ss.str());
  }
  catch (const parse::parse_error &e) {
    stringstream ss;
    string line_prev, line_cur;
    int line, column;
    get_error_line_column(json_begin, json_end, e.get_position(), line_prev,
                          line_cur, line, column);
    ss << "Error parsing JSON at line " << line << ", column " << column
       << "\n";
    ss << "Message: " << e.what() << "\n";
    print_json_parse_error_marker(ss, line_prev, line_cur, line, column);
    throw invalid_argument(ss.str());
  }
}

nd::array dynd::format_json(const nd::array& n, bool struct_as_list)
{
  // Create a UTF-8 string
  nd::array result = nd::empty(ndt::make_string());

  // Initialize the output with some memory
  output_data out;
  out.blockref = reinterpret_cast<const string_type_arrmeta *>(
                     result.get_arrmeta())->blockref;
  out.api = get_memory_block_pod_allocator_api(out.blockref);
  out.api->allocate(out.blockref, 1024, 1, &out.out_begin,
                    &out.out_capacity_end);
  out.out_end = out.out_begin;
  out.struct_as_list = struct_as_list;

  if (!n.get_type().is_expression()) {
    ::format_json(out, n.get_type(), n.get_arrmeta(),
                  n.get_readonly_originptr());
  } else {
    nd::array tmp = n.eval();
    ::format_json(out, tmp.get_type(), tmp.get_arrmeta(),
                  tmp.get_readonly_originptr());
  }

  // Shrink the memory to fit, and set the pointers in the output
  string_type_data *d =
      reinterpret_cast<string_type_data *>(result.get_readwrite_originptr());
  d->begin = out.out_begin;
  d->end = out.out_capacity_end;
  out.api->resize(out.blockref, out.out_end - out.out_begin, &d->begin,
                  &d->end);

  // Finalize processing and mark the result as immutable
  result.get_type().extended()->arrmeta_finalize_buffers(result.get_arrmeta());
  result.flag_as_immutable();

  return result;
}

nd::array dynd::struct_concat(nd::array lhs, nd::array rhs)
{
  nd::array res;
  if (lhs.is_null()) {
    res = rhs;
    return res;
  }
  if (rhs.is_null()) {
    res = lhs;
    return res;
  }
  const ndt::type &lhs_tp = lhs.get_type(), &rhs_tp = rhs.get_type();
  if (lhs_tp.get_kind() != struct_kind) {
    stringstream ss;
    ss << "Cannot concatenate array with type " << lhs_tp << " as a struct";
    throw invalid_argument(ss.str());
  }
  if (rhs_tp.get_kind() != struct_kind) {
    stringstream ss;
    ss << "Cannot concatenate array with type " << rhs_tp << " as a struct";
    throw invalid_argument(ss.str());
  }

  // Make an empty shell struct by concatenating the fields together
  intptr_t lhs_n = lhs_tp.extended<ndt::base_struct_type>()->get_field_count();
  intptr_t rhs_n = rhs_tp.extended<ndt::base_struct_type>()->get_field_count();
  intptr_t res_n = lhs_n + rhs_n;
  nd::array res_field_names = nd::empty(res_n, ndt::string_type::make());
  nd::array res_field_types = nd::empty(res_n, ndt::make_type());
  res_field_names(irange(0, lhs_n)).vals() = lhs_tp.extended<ndt::base_struct_type>()->get_field_names();
  res_field_names(irange(lhs_n, res_n)).vals() = rhs_tp.extended<ndt::base_struct_type>()->get_field_names();
  res_field_types(irange(0, lhs_n)).vals() = lhs_tp.extended<ndt::base_struct_type>()->get_field_types();
  res_field_types(irange(lhs_n, res_n)).vals() = rhs_tp.extended<ndt::base_struct_type>()->get_field_types();
  ndt::type res_tp = ndt::struct_type::make(res_field_names, res_field_types);
  const ndt::type *res_field_tps = res_tp.extended<ndt::base_struct_type>()->get_field_types_raw();
  res = nd::empty_shell(res_tp);

  // Initialize the default data offsets for the struct arrmeta
  ndt::struct_type::fill_default_data_offsets(res_n, res_tp.extended<ndt::base_struct_type>()->get_field_types_raw(),
                                              reinterpret_cast<uintptr_t *>(res.get_arrmeta()));
  // Get information about the arrmeta layout of the input and res
  const uintptr_t *lhs_arrmeta_offsets = lhs_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
  const uintptr_t *rhs_arrmeta_offsets = rhs_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
  const uintptr_t *res_arrmeta_offsets = res_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
  const char *lhs_arrmeta = lhs.get_arrmeta();
  const char *rhs_arrmeta = rhs.get_arrmeta();
  char *res_arrmeta = res.get_arrmeta();
  // Copy the arrmeta from the input arrays
  for (intptr_t i = 0; i < lhs_n; ++i) {
    const ndt::type &tp = res_field_tps[i];
    if (!tp.is_builtin()) {
      tp.extended()->arrmeta_copy_construct(res_arrmeta + res_arrmeta_offsets[i], lhs_arrmeta + lhs_arrmeta_offsets[i],
                                            lhs.get_data_memblock().get());
    }
  }
  for (intptr_t i = 0; i < rhs_n; ++i) {
    const ndt::type &tp = res_field_tps[i + lhs_n];
    if (!tp.is_builtin()) {
      tp.extended()->arrmeta_copy_construct(res_arrmeta + res_arrmeta_offsets[i + lhs_n],
                                            rhs_arrmeta + rhs_arrmeta_offsets[i], rhs.get_data_memblock().get());
    }
  }

  // Get information about the data layout of the input and res
  const uintptr_t *lhs_data_offsets = lhs_tp.extended<ndt::base_struct_type>()->get_data_offsets(lhs.get_arrmeta());
  const uintptr_t *rhs_data_offsets = rhs_tp.extended<ndt::base_struct_type>()->get_data_offsets(rhs.get_arrmeta());
  const uintptr_t *res_data_offsets = res_tp.extended<ndt::base_struct_type>()->get_data_offsets(res.get_arrmeta());
  const char *lhs_data = lhs.get_readonly_originptr();
  const char *rhs_data = rhs.get_readonly_originptr();
  char *res_data = res.get_readwrite_originptr();
  // Copy the data from the input arrays
  for (intptr_t i = 0; i < lhs_n; ++i) {
    const ndt::type &tp = res_field_tps[i];
    typed_data_copy(tp, res_arrmeta + res_arrmeta_offsets[i], res_data + res_data_offsets[i],
                    lhs_arrmeta + lhs_arrmeta_offsets[i], lhs_data + lhs_data_offsets[i]);
  }

  for (intptr_t i = 0; i < rhs_n; ++i) {
    const ndt::type &tp = res_field_tps[i + lhs_n];
    typed_data_copy(tp, res_arrmeta + res_arrmeta_offsets[i + lhs_n], res_data + res_data_offsets[i + lhs_n],
                    rhs_arrmeta + rhs_arrmeta_offsets[i], rhs_data + rhs_data_offsets[i]);
  }

  return res;
}

//.........这里部分代码省略.........
  } else {
    // Get the function pointer for the first_call, for the case with
    // a reduction identity
    if (kernreq == kernel_request_single) {
      e->set_first_call_function(
          &strided_inner_broadcast_kernel_extra::single_first_with_ident);
    } else if (kernreq == kernel_request_strided) {
      e->set_first_call_function(
          &strided_inner_broadcast_kernel_extra::strided_first_with_ident);
    } else {
      stringstream ss;
      ss << "make_lifted_reduction_ckernel: unrecognized request "
         << (int)kernreq;
      throw runtime_error(ss.str());
    }
    if (reduction_identity.get_type() != dst_tp) {
      stringstream ss;
      ss << "make_lifted_reduction_ckernel: reduction identity type ";
      ss << reduction_identity.get_type() << " does not match dst type ";
      ss << dst_tp;
      throw runtime_error(ss.str());
    }
    e->ident_data = reduction_identity.get_readonly_originptr();
    e->ident_ref = reduction_identity.get_memblock().release();
  }
  // The function pointer for followup accumulation calls
  e->set_followup_call_function(
      &strided_inner_broadcast_kernel_extra::strided_followup);
  // The striding parameters
  e->dst_stride = dst_stride;
  e->src_stride = src_stride;
  e->size = src_size;
  // Validate that the provided callables are unary operations,
  // and have the correct types
  if (elwise_reduction_tp->get_npos() != 1 &&
      elwise_reduction_tp->get_npos() != 2) {
    stringstream ss;
    ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
    ss << "funcproto must be unary or a binary expr with all equal types";
    throw runtime_error(ss.str());
  }
  if (elwise_reduction_tp->get_return_type() != dst_tp) {
    stringstream ss;
    ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
    ss << "dst type is " << elwise_reduction_tp->get_return_type();
    ss << ", expected " << dst_tp;
    throw type_error(ss.str());
  }
  if (elwise_reduction_tp->get_pos_type(0) != src_tp) {
    stringstream ss;
    ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
    ss << "src type is " << elwise_reduction_tp->get_return_type();
    ss << ", expected " << src_tp;
    throw type_error(ss.str());
  }
  if (dst_initialization != NULL) {
    check_dst_initialization(dst_initialization_tp, dst_tp, src_tp);
  }
  if (elwise_reduction_tp->get_npos() == 2) {
    ckb_offset = kernels::wrap_binary_as_unary_reduction_ckernel(
        ckb, ckb_offset, right_associative, kernel_request_strided);
    ndt::type src_tp_doubled[2] = {src_tp, src_tp};
    const char *src_arrmeta_doubled[2] = {src_arrmeta, src_arrmeta};
    ckb_offset = elwise_reduction->instantiate(
        elwise_reduction->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
        dst_arrmeta, elwise_reduction_tp->get_npos(), src_tp_doubled,
        src_arrmeta_doubled, kernel_request_strided, ectx, nd::array(),
        std::map<nd::string, ndt::type>());
  } else {
    ckb_offset = elwise_reduction->instantiate(
        elwise_reduction->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
        dst_arrmeta, elwise_reduction_tp->get_npos(), &src_tp, &src_arrmeta,
        kernel_request_strided, ectx, nd::array(),
        std::map<nd::string, ndt::type>());
  }
  // Make sure there's capacity for the next ckernel
  reinterpret_cast<ckernel_builder<kernel_request_host> *>(ckb)
      ->reserve(ckb_offset + sizeof(ckernel_prefix));
  // Need to retrieve 'e' again because it may have moved
  e = reinterpret_cast<ckernel_builder<kernel_request_host> *>(ckb)
          ->get_at<strided_inner_broadcast_kernel_extra>(root_ckb_offset);
  e->dst_init_kernel_offset = ckb_offset - root_ckb_offset;
  if (dst_initialization != NULL) {
    ckb_offset = dst_initialization->instantiate(
        dst_initialization->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
        dst_arrmeta, elwise_reduction_tp->get_npos(), &src_tp, &src_arrmeta,
        kernel_request_strided, ectx, nd::array(),
        std::map<nd::string, ndt::type>());
  } else if (reduction_identity.is_null()) {
    ckb_offset =
        make_assignment_kernel(ckb, ckb_offset, dst_tp, dst_arrmeta, src_tp,
                               src_arrmeta, kernel_request_strided, ectx);
  } else {
    ckb_offset = make_assignment_kernel(
        ckb, ckb_offset, dst_tp, dst_arrmeta, reduction_identity.get_type(),
        reduction_identity.get_arrmeta(), kernel_request_strided, ectx);
  }

  return ckb_offset;
}

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();
}

 size_t get_category_count() const {
     return (size_t) reinterpret_cast<const strided_dim_type_arrmeta *>(
                m_categories.get_arrmeta())->dim_size;
 }