utils.h

Go to the documentation of this file.

// Copyright (C) 2019-2026 Garth N. Wells and Jørgen S. Dokken
//
// This file is part of DOLFINx (https://www.fenicsproject.org)
//
// SPDX-License-Identifier:    LGPL-3.0-or-later
 
#pragma once
 
#include "EntityMap.h"
#include "Mesh.h"
#include "MeshTags.h"
#include "Topology.h"
#include "graphbuild.h"
#include <algorithm>
#include <basix/mdspan.hpp>
#include <concepts>
#include <cstdint>
#include <dolfinx/graph/AdjacencyList.h>
#include <dolfinx/graph/ordering.h>
#include <dolfinx/graph/partition.h>
#include <format>
#include <functional>
#include <mpi.h>
#include <numeric>
#include <optional>
#include <ranges>
#include <span>
#include <vector>

 
namespace dolfinx::fem
{
class ElementDofLayout;
}
 
namespace dolfinx::mesh
{
enum class CellType : std::int8_t;

enum class GhostMode : std::uint8_t
{
  none,
  shared_facet
};
 
namespace impl
{
template <typename T>
void reorder_list(std::span<T> list, std::span<const std::int32_t> nodemap)
{
  if (nodemap.empty())
    return;
 
  assert(list.size() % nodemap.size() == 0);
  std::size_t degree = list.size() / nodemap.size();
  const std::vector<T> orig(list.begin(), list.end());
  for (std::size_t n = 0; n < nodemap.size(); ++n)
  {
    std::span links_old(orig.data() + n * degree, degree);
    auto links_new = list.subspan(nodemap[n] * degree, degree);
    std::ranges::copy(links_old, links_new.begin());
  }
}

template <std::floating_point T>
std::tuple<std::vector<std::int32_t>, std::vector<T>, std::vector<std::int32_t>>
compute_vertex_coords_boundary(const mesh::Mesh<T>& mesh, int dim,
                               std::span<const std::int32_t> facets)
{
  auto topology = mesh.topology();
  assert(topology);
  const int tdim = topology->dim();
  if (dim == tdim)
  {
    throw std::runtime_error(
        "Cannot use mesh::locate_entities_boundary (boundary) for cells.");
  }
 
  // Build set of vertices on boundary and set of boundary entities
  mesh.topology_mutable()->create_connectivity(tdim - 1, 0);
  mesh.topology_mutable()->create_connectivity(tdim - 1, dim);
  std::vector<std::int32_t> vertices, entities;
  {
    auto f_to_v = topology->connectivity(tdim - 1, 0);
    assert(f_to_v);
    auto f_to_e = topology->connectivity(tdim - 1, dim);
    assert(f_to_e);
    for (auto f : facets)
    {
      auto v = f_to_v->links(f);
      vertices.insert(vertices.end(), v.begin(), v.end());
      auto e = f_to_e->links(f);
      entities.insert(entities.end(), e.begin(), e.end());
    }
 
    // Build vector of boundary vertices
    {
      std::ranges::sort(vertices);
      auto [unique_end, range_end] = std::ranges::unique(vertices);
      vertices.erase(unique_end, range_end);
    }
 
    {
      std::ranges::sort(entities);
      auto [unique_end, range_end] = std::ranges::unique(entities);
      entities.erase(unique_end, range_end);
    }
  }
 
  // Get geometry data
  auto x_dofmap = mesh.geometry().dofmaps().front();
  std::span<const T> x_nodes = mesh.geometry().x();
 
  // Get all vertex 'node' indices
  mesh.topology_mutable()->create_connectivity(0, tdim);
  mesh.topology_mutable()->create_connectivity(tdim, 0);
  auto v_to_c = topology->connectivity(0, tdim);
  assert(v_to_c);
  auto c_to_v = topology->connectivity(tdim, 0);
  assert(c_to_v);
  std::vector<T> x_vertices(3 * vertices.size(), -1.0);
  std::vector<std::int32_t> vertex_to_pos(v_to_c->num_nodes(), -1);
  for (std::size_t i = 0; i < vertices.size(); ++i)
  {
    const std::int32_t v = vertices[i];
 
    // Get first cell and find position
    const std::int32_t c = v_to_c->links(v).front();
    auto cell_vertices = c_to_v->links(c);
    auto it = std::ranges::find(cell_vertices, v);
    assert(it != cell_vertices.end());
    const std::size_t local_pos
        = std::ranges::distance(cell_vertices.begin(), it);
 
    auto dofs = md::submdspan(x_dofmap, c, md::full_extent);
    for (std::size_t j = 0; j < 3; ++j)
      x_vertices[j * vertices.size() + i] = x_nodes[3 * dofs[local_pos] + j];
    vertex_to_pos[v] = i;
  }
 
  return {std::move(entities), std::move(x_vertices), std::move(vertex_to_pos)};
}
 
} // namespace impl

std::vector<std::int32_t> exterior_facet_indices(const Topology& topology,
                                                 int facet_type_idx);

std::vector<std::int32_t> exterior_facet_indices(const Topology& topology);

using CellPartitionFunction = std::function<graph::AdjacencyList<std::int32_t>(
    MPI_Comm comm, int nparts, const std::vector<CellType>& cell_types,
    const std::vector<std::span<const std::int64_t>>& cells)>;

using CellReorderFunction = std::function<std::vector<std::int32_t>(
    const graph::AdjacencyList<std::int32_t>&)>;

inline auto
create_boundary_vertices_fn(const CellReorderFunction& reorder_fn,
                            std::optional<std::int32_t> max_facet_to_cell_links)
{
  return [&, max_facet_to_cell_links](
             const std::vector<CellType>& celltypes,
             const std::vector<fem::ElementDofLayout>& doflayouts,
             const std::vector<std::vector<int>>& ghost_owners,
             std::vector<std::vector<std::int64_t>>& cells,
             std::vector<std::vector<std::int64_t>>& cells_v,
             std::vector<std::vector<std::int64_t>>& original_idx,
             int num_threads) -> std::vector<std::int64_t>
  {
    // Build local dual graph for owned cells to (i) get list of vertices
    // on the process boundary and (ii) apply re-ordering to cells for
    // locality
 
    spdlog::info("Build local dual graphs, re-order cells, and compute process "
                 "boundary vertices.");
 
    std::vector<std::pair<std::vector<std::int64_t>, int>> facets;
 
    // Build lists of cells (by cell type) that excludes ghosts
    std::vector<std::span<const std::int64_t>> cells1_v_local;
    for (std::size_t i = 0; i < celltypes.size(); ++i)
    {
      int num_cell_vertices = mesh::num_cell_vertices(celltypes[i]);
      std::size_t num_owned_cells
          = cells_v[i].size() / num_cell_vertices - ghost_owners[i].size();
      cells1_v_local.emplace_back(cells_v[i].data(),
                                  num_owned_cells * num_cell_vertices);
 
      // Build local dual graph for cell type
      auto [graph, unmatched_facets, max_v, _facet_attached_cells]
          = build_local_dual_graph(std::vector{celltypes[i]},
                                   std::vector{cells1_v_local.back()},
                                   max_facet_to_cell_links, num_threads);
 
      // Store unmatched_facets for current cell type
      facets.emplace_back(std::move(unmatched_facets), max_v);
 
      // Compute re-ordering of graph
      const std::vector<std::int32_t> remap = reorder_fn(graph);
 
      // Update 'original' indices
      const std::vector<std::int64_t>& orig_idx = original_idx[i];
      std::vector<std::int64_t> _original_idx(orig_idx.size());
      std::copy_n(orig_idx.rbegin(), ghost_owners[i].size(),
                  _original_idx.rbegin());
      {
        for (std::size_t j = 0; j < remap.size(); ++j)
          _original_idx[remap[j]] = orig_idx[j];
      }
      original_idx[i] = _original_idx;
 
      // Reorder cells
      impl::reorder_list(
          std::span(cells_v[i].data(), remap.size() * num_cell_vertices),
          remap);
      impl::reorder_list(
          std::span(cells[i].data(), remap.size() * doflayouts[i].num_dofs()),
          remap);
    }
 
    if (facets.size() == 1) // Optimisation for single cell type
    {
      std::vector<std::int64_t>& vertices = facets.front().first;
 
      // Remove duplicated vertex indices
      std::ranges::sort(vertices);
      auto [unique_end, range_end] = std::ranges::unique(vertices);
      vertices.erase(unique_end, range_end);
 
      // Remove -1 if it appears as first entity. This can happen in
      // mixed topology meshes where '-1' is used to pad facet data when
      // cells facets have differing numbers of vertices.
      if (!vertices.empty() and vertices.front() == -1)
        vertices.erase(vertices.begin());
 
      return vertices;
    }
    else
    {
      // Pack 'unmatched' facets for all cell types into a single
      // column-major array (facets0): column j holds vertex j across
      // all facets, so the multi-column sort_by_perm() overload can
      // operate directly on contiguous per-column data.
      std::size_t num_facets = std::accumulate(
          facets.begin(), facets.end(), std::size_t(0),
          [](std::size_t x, auto& y)
          { return x + (y.second > 0 ? y.first.size() / y.second : 0); });
      int max_v = std::ranges::max_element(facets, [](auto& a, auto& b)
                                           { return a.second < b.second; })
                      ->second;
 
      std::vector<std::int64_t> facets0_b(max_v * num_facets, -1);
      std::vector<std::span<std::int64_t>> facets0(max_v);
      for (int j = 0; j < max_v; ++j)
        facets0[j] = std::span(facets0_b.data() + j * num_facets, num_facets);
 
      {
        std::size_t row = 0;
        for (const auto& [v_data, num_v] : facets)
        {
          for (auto it = v_data.begin(); it != v_data.end(); it += num_v, ++row)
            for (int j = 0; j < num_v; ++j)
              facets0[j][row] = *std::next(it, j);
        }
      }
 
      // Compute row permutation
      std::vector<std::span<const std::int64_t>> facets0_view(facets0.begin(),
                                                              facets0.end());
      const std::vector<std::int32_t> perm = dolfinx::sort_by_perm(
          std::span<std::span<const std::int64_t>>(facets0_view));
 
      // For facets in facets0 that appear only once, store the facet
      // vertices
      std::vector<std::int64_t> vertices;
      // TODO: allocate memory for vertices
 
      // Number of leading valid (non -1 padding) vertices in row
      auto trim_len = [&facets0, max_v](std::int32_t row)
      {
        int n = max_v;
        while (n > 0 and facets0[n - 1][row] < 0)
          --n;
        return n;
      };
 
      auto it = perm.begin();
      while (it != perm.end())
      {
        std::int32_t row0 = *it;
        int n = trim_len(row0);
 
        // Find iterator to next facet whose leading n vertices differ
        // from row0
        auto it1 = std::find_if_not(it, perm.end(),
                                    [&facets0, row0, n](std::int32_t row)
                                    {
                                      for (int j = 0; j < n; ++j)
                                        if (facets0[j][row] != facets0[j][row0])
                                          return false;
                                      return true;
                                    });
 
        // If no repeated facet found, insert row0 vertices
        if (std::ranges::distance(it, it1) == 1)
        {
          for (int j = 0; j < n; ++j)
            vertices.push_back(facets0[j][row0]);
        }
        else if (std::ranges::distance(it, it1) > 2)
          throw std::runtime_error("More than two matching facets found.");
 
        // Advance iterator
        it = it1;
      }
 
      // Remove duplicate indices
      std::ranges::sort(vertices);
      auto [unique_end, range_end] = std::ranges::unique(vertices);
      vertices.erase(unique_end, range_end);
 
      return vertices;
    }
  };
}

std::vector<std::int64_t> extract_topology(CellType cell_type,
                                           const fem::ElementDofLayout& layout,
                                           std::span<const std::int64_t> cells);

template <std::floating_point T>
std::vector<T> h(const Mesh<T>& mesh, std::span<const std::int32_t> entities,
                 int dim)
{
  if (entities.empty())
    return std::vector<T>();
  if (dim == 0)
    return std::vector<T>(entities.size(), 0);
 
  // Get the geometry dofs for the vertices of each entity
  const auto [vertex_xdofs, xdof_shape]
      = entities_to_geometry(mesh, dim, entities, false);
 
  // Get the  geometry coordinate
  std::span<const T> x = mesh.geometry().x();
 
  // Function to compute the length of (p0 - p1)
  auto delta_norm = [](auto&& p0, auto&& p1)
  {
    T norm = 0;
    for (std::size_t i = 0; i < 3; ++i)
      norm += (p0[i] - p1[i]) * (p0[i] - p1[i]);
    return std::sqrt(norm);
  };
 
  // Compute greatest distance between any to vertices
  assert(dim > 0);
  std::vector<T> h(entities.size(), 0);
  for (std::size_t e = 0; e < entities.size(); ++e)
  {
    // Get geometry 'dof' for each vertex of entity e
    std::span<const std::int32_t> e_vertices(
        vertex_xdofs.data() + e * xdof_shape[1], xdof_shape[1]);
 
    // Compute maximum distance between any two vertices
    for (std::size_t i = 0; i < e_vertices.size(); ++i)
    {
      std::span<const T, 3> p0(x.data() + 3 * e_vertices[i], 3);
      for (std::size_t j = i + 1; j < e_vertices.size(); ++j)
      {
        std::span<const T, 3> p1(x.data() + 3 * e_vertices[j], 3);
        h[e] = std::max(h[e], delta_norm(p0, p1));
      }
    }
  }
 
  return h;
}

template <std::floating_point T>
std::vector<T> cell_normals(const Mesh<T>& mesh, int dim,
                            std::span<const std::int32_t> entities)
{
  if (entities.empty())
    return std::vector<T>();
 
  auto topology = mesh.topology();
  assert(topology);
  if (topology->cell_type() == CellType::prism and dim == 2)
  {
    throw std::runtime_error(
        "Cell normal computation for prism cells not yet supported.");
  }
 
  const int gdim = mesh.geometry().dim();
  const CellType type = cell_entity_type(topology->cell_type(), dim, 0);
 
  // Find geometry nodes for topology entities
  std::span<const T> x = mesh.geometry().x();
  const auto [geometry_entities, eshape]
      = entities_to_geometry(mesh, dim, entities, false);
 
  std::vector<T> n(entities.size() * 3);
  switch (type)
  {
  case CellType::interval:
  {
    if (gdim > 2)
      throw std::invalid_argument("Interval cell normal undefined in 3D.");
    for (std::size_t i = 0; i < entities.size(); ++i)
    {
      // Get the two vertices as points
      std::array vertices{geometry_entities[i * eshape[1]],
                          geometry_entities[i * eshape[1] + 1]};
      std::array p = {std::span<const T, 3>(x.data() + 3 * vertices[0], 3),
                      std::span<const T, 3>(x.data() + 3 * vertices[1], 3)};
 
      // Define normal by rotating tangent counter-clockwise
      std::array<T, 3> t;
      std::ranges::transform(p[1], p[0], t.begin(),
                             [](auto x, auto y) { return x - y; });
 
      T norm = std::sqrt(t[0] * t[0] + t[1] * t[1]);
      std::span<T, 3> ni(n.data() + 3 * i, 3);
      ni[0] = -t[1] / norm;
      ni[1] = t[0] / norm;
      ni[2] = 0.0;
    }
    return n;
  }
  case CellType::triangle:
  {
    for (std::size_t i = 0; i < entities.size(); ++i)
    {
      // Get the three vertices as points
      std::array vertices = {geometry_entities[i * eshape[1] + 0],
                             geometry_entities[i * eshape[1] + 1],
                             geometry_entities[i * eshape[1] + 2]};
      std::array p = {std::span<const T, 3>(x.data() + 3 * vertices[0], 3),
                      std::span<const T, 3>(x.data() + 3 * vertices[1], 3),
                      std::span<const T, 3>(x.data() + 3 * vertices[2], 3)};
 
      // Compute (p1 - p0) and (p2 - p0)
      std::array<T, 3> dp1, dp2;
      std::ranges::transform(p[1], p[0], dp1.begin(),
                             [](auto x, auto y) { return x - y; });
      std::ranges::transform(p[2], p[0], dp2.begin(),
                             [](auto x, auto y) { return x - y; });
 
      // Define cell normal via cross product of first two edges
      std::array<T, 3> ni = math::cross(dp1, dp2);
      T norm = std::sqrt(ni[0] * ni[0] + ni[1] * ni[1] + ni[2] * ni[2]);
      std::ranges::transform(ni, std::next(n.begin(), 3 * i),
                             [norm](auto x) { return x / norm; });
    }
 
    return n;
  }
  case CellType::quadrilateral:
  {
    // TODO: check
    for (std::size_t i = 0; i < entities.size(); ++i)
    {
      // Get the three vertices as points
      std::array vertices = {geometry_entities[i * eshape[1] + 0],
                             geometry_entities[i * eshape[1] + 1],
                             geometry_entities[i * eshape[1] + 2]};
      std::array p = {std::span<const T, 3>(x.data() + 3 * vertices[0], 3),
                      std::span<const T, 3>(x.data() + 3 * vertices[1], 3),
                      std::span<const T, 3>(x.data() + 3 * vertices[2], 3)};
 
      // Compute (p1 - p0) and (p2 - p0)
      std::array<T, 3> dp1, dp2;
      std::ranges::transform(p[1], p[0], dp1.begin(),
                             [](auto x, auto y) { return x - y; });
      std::ranges::transform(p[2], p[0], dp2.begin(),
                             [](auto x, auto y) { return x - y; });
 
      // Define cell normal via cross product of first two edges
      std::array<T, 3> ni = math::cross(dp1, dp2);
      T norm = std::sqrt(ni[0] * ni[0] + ni[1] * ni[1] + ni[2] * ni[2]);
      std::ranges::transform(ni, std::next(n.begin(), 3 * i),
                             [norm](auto x) { return x / norm; });
    }
 
    return n;
  }
  default:
    throw std::invalid_argument(
        "cell_normal not supported for this cell type.");
  }
}

template <std::floating_point T>
std::vector<T> compute_midpoints(const Mesh<T>& mesh, int dim,
                                 std::span<const std::int32_t> entities)
{
  if (entities.empty())
    return std::vector<T>();
 
  std::span<const T> x = mesh.geometry().x();
 
  // Build map from entity -> geometry dof
  const auto [e_to_g, eshape]
      = entities_to_geometry(mesh, dim, entities, false);
 
  std::vector<T> x_mid(entities.size() * 3, 0);
  for (std::size_t e = 0; e < entities.size(); ++e)
  {
    std::span<T, 3> p(x_mid.data() + 3 * e, 3);
    std::span<const std::int32_t> rows(e_to_g.data() + e * eshape[1],
                                       eshape[1]);
    for (auto row : rows)
    {
      std::span<const T, 3> xg(x.data() + 3 * row, 3);
      std::ranges::transform(p, xg, p.begin(),
                             [size = rows.size()](auto x, auto y)
                             { return x + y / size; });
    }
  }
 
  return x_mid;
}
 
namespace impl
{
template <std::floating_point T>
std::pair<std::vector<T>, std::array<std::size_t, 2>>
compute_vertex_coords(const mesh::Mesh<T>& mesh)
{
  auto topology = mesh.topology();
  assert(topology);
  const int tdim = topology->dim();
 
  // Create entities and connectivities
 
  // Get all vertex 'node' indices
  const std::int32_t num_vertices = topology->index_map(0)->size_local()
                                    + topology->index_map(0)->num_ghosts();
 
  std::vector<std::int32_t> vertex_to_node(num_vertices);
  for (int cell_type_idx = 0,
           num_cell_types = topology->entity_types(tdim).size();
       cell_type_idx < num_cell_types; ++cell_type_idx)
  {
    auto x_dofmap = mesh.geometry().dofmaps().at(cell_type_idx);
    auto c_to_v = topology->connectivity({tdim, cell_type_idx}, {0, 0});
    assert(c_to_v);
    for (int c = 0; c < c_to_v->num_nodes(); ++c)
    {
      auto x_dofs = md::submdspan(x_dofmap, c, md::full_extent);
      auto vertices = c_to_v->links(c);
      for (std::size_t i = 0; i < vertices.size(); ++i)
        vertex_to_node[vertices[i]] = x_dofs[i];
    }
  }
 
  // Pack coordinates of vertices
  std::span<const T> x_nodes = mesh.geometry().x();
  std::vector<T> x_vertices(3 * vertex_to_node.size(), 0.0);
  for (std::size_t i = 0; i < vertex_to_node.size(); ++i)
  {
    std::int32_t pos = 3 * vertex_to_node[i];
    for (std::size_t j = 0; j < 3; ++j)
      x_vertices[j * vertex_to_node.size() + i] = x_nodes[pos + j];
  }
 
  return {std::move(x_vertices), {3, vertex_to_node.size()}};
}
 
} // namespace impl

template <typename Fn, typename T>
concept MarkerFn = std::is_invocable_r<
    std::vector<std::int8_t>, Fn,
    md::mdspan<const T,
               md::extents<std::size_t, 3, md::dynamic_extent>>>::value;

template <std::floating_point T, MarkerFn<T> U>
std::vector<std::int32_t> locate_entities(const Mesh<T>& mesh, int dim,
                                          U marker, int entity_type_idx)
{
 
  using cmdspan3x_t
      = md::mdspan<const T, md::extents<std::size_t, 3, md::dynamic_extent>>;
 
  // Run marker function on vertex coordinates
  const auto [xdata, xshape] = impl::compute_vertex_coords(mesh);
 
  cmdspan3x_t x(xdata.data(), xshape);
  const std::vector<std::int8_t> marked = marker(x);
  if (marked.size() != x.extent(1))
    throw std::runtime_error("Length of array of markers is wrong.");
 
  auto topology = mesh.topology();
  assert(topology);
  const int tdim = topology->dim();
 
  mesh.topology_mutable()->create_entities(dim);
  if (dim < tdim)
    mesh.topology_mutable()->create_connectivity(dim, 0);
 
  // Iterate over entities of dimension 'dim' to build vector of marked
  // entities
  auto e_to_v = topology->connectivity({dim, entity_type_idx}, {0, 0});
  assert(e_to_v);
  std::vector<std::int32_t> entities;
  for (int e = 0; e < e_to_v->num_nodes(); ++e)
  {
    // Iterate over entity vertices
    bool all_vertices_marked = true;
    for (std::int32_t v : e_to_v->links(e))
    {
      if (!marked[v])
      {
        all_vertices_marked = false;
        break;
      }
    }
 
    if (all_vertices_marked)
      entities.push_back(e);
  }
 
  return entities;
}

template <std::floating_point T, MarkerFn<T> U>
std::vector<std::int32_t> locate_entities(const Mesh<T>& mesh, int dim,
                                          U marker)
{
  const int num_entity_types = mesh.topology()->entity_types(dim).size();
  if (num_entity_types > 1)
  {
    throw std::runtime_error(
        "Multiple entity types of this dimension. Specify entity type index");
  }
  return locate_entities(mesh, dim, marker, 0);
}

template <std::floating_point T, MarkerFn<T> U>
std::vector<std::int32_t> locate_entities_boundary(const Mesh<T>& mesh, int dim,
                                                   U marker)
{
  // TODO Rewrite this function, it should be possible to simplify considerably
  auto topology = mesh.topology();
  assert(topology);
  int tdim = topology->dim();
  if (dim == tdim)
  {
    throw std::runtime_error(
        "Cannot use mesh::locate_entities_boundary (boundary) for cells.");
  }
 
  // Compute list of boundary facets
  mesh.topology_mutable()->create_entities(tdim - 1);
  mesh.topology_mutable()->create_connectivity(tdim - 1, tdim);
  std::vector<std::int32_t> boundary_facets = exterior_facet_indices(*topology);
 
  using cmdspan3x_t
      = md::mdspan<const T, md::extents<std::size_t, 3, md::dynamic_extent>>;
 
  // Run marker function on the vertex coordinates
  auto [facet_entities, xdata, vertex_to_pos]
      = impl::compute_vertex_coords_boundary(mesh, dim, boundary_facets);
  cmdspan3x_t x(xdata.data(), 3, xdata.size() / 3);
  std::vector<std::int8_t> marked = marker(x);
  if (marked.size() != x.extent(1))
    throw std::runtime_error("Length of array of markers is wrong.");
 
  // Loop over entities and check vertex markers
  mesh.topology_mutable()->create_entities(dim);
  auto e_to_v = topology->connectivity(dim, 0);
  assert(e_to_v);
  std::vector<std::int32_t> entities;
  for (auto e : facet_entities)
  {
    // Iterate over entity vertices
    bool all_vertices_marked = true;
    for (auto v : e_to_v->links(e))
    {
      const std::int32_t pos = vertex_to_pos[v];
      if (!marked[pos])
      {
        all_vertices_marked = false;
        break;
      }
    }
 
    // Mark facet with all vertices marked
    if (all_vertices_marked)
      entities.push_back(e);
  }
 
  return entities;
}

template <std::floating_point T>
std::pair<std::vector<std::int32_t>, std::array<std::size_t, 2>>
entities_to_geometry(const Mesh<T>& mesh, int dim,
                     std::span<const std::int32_t> entities,
                     bool permute = false)
{
  auto topology = mesh.topology();
  assert(topology);
  CellType cell_type = topology->cell_type();
  if ((cell_type == CellType::prism or cell_type == CellType::pyramid)
      and dim == 2)
  {
    throw std::runtime_error("mesh::entities_to_geometry for prism/pyramid "
                             "cell facets not yet supported.");
  }
 
  const int tdim = topology->dim();
  const Geometry<T>& geometry = mesh.geometry();
  auto xdofs = geometry.dofmaps().front();
 
  // Get the DOF layout and the number of DOFs per entity
  const fem::CoordinateElement<T>& coord_ele = geometry.cmaps().front();
  const fem::ElementDofLayout layout = coord_ele.create_dof_layout();
  const std::size_t num_entity_dofs = layout.entity_closure_dofs(dim, 0).size();
  std::vector<std::int32_t> entity_xdofs;
  entity_xdofs.reserve(entities.size() * num_entity_dofs);
  std::array<std::size_t, 2> eshape{entities.size(), num_entity_dofs};
 
  // Get the element's closure DOFs
  const std::vector<std::vector<std::vector<int>>>& closure_dofs_all
      = layout.entity_closure_dofs_all();
 
  // Special case when dim == tdim (cells)
  if (dim == tdim)
  {
    for (std::int32_t c : entities)
    {
      // Extract degrees of freedom
      auto x_c = md::submdspan(xdofs, c, md::full_extent);
      for (std::int32_t entity_dof : closure_dofs_all[tdim][0])
        entity_xdofs.push_back(x_c[entity_dof]);
    }
 
    return {std::move(entity_xdofs), eshape};
  }
 
  assert(dim != tdim);
 
  auto e_to_c = topology->connectivity(dim, tdim);
  if (!e_to_c)
  {
    throw std::runtime_error(std::format(
        "Entity-to-cell connectivity has not been computed. Missing dims "
        "{}->{}",
        dim, tdim));
  }
 
  auto c_to_e = topology->connectivity(tdim, dim);
  if (!c_to_e)
  {
    throw std::runtime_error(std::format(
        "Cell-to-entity connectivity has not been computed. Missing dims "
        "{}->{}",
        tdim, dim));
  }
 
  // Get the cell info, which is needed to permute the closure dofs
  std::span<const std::uint32_t> cell_info;
  if (permute)
    cell_info = std::span(mesh.topology()->get_cell_permutation_info());
 
  for (std::int32_t e : entities)
  {
    // Get a cell connected to the entity
    assert(!e_to_c->links(e).empty());
    std::int32_t c = e_to_c->links(e).front();
 
    // Get the local index of the entity
    std::span<const std::int32_t> cell_entities = c_to_e->links(c);
    auto it = std::find(cell_entities.begin(), cell_entities.end(), e);
    assert(it != cell_entities.end());
    std::size_t local_entity = std::ranges::distance(cell_entities.begin(), it);
 
    // Cell sub-entities must be permuted so that their local
    // orientation agrees with their global orientation
    std::vector<std::int32_t> closure_dofs(closure_dofs_all[dim][local_entity]);
    if (permute)
    {
      mesh::CellType entity_type
          = mesh::cell_entity_type(cell_type, dim, local_entity);
      coord_ele.permute_subentity_closure(closure_dofs, cell_info[c],
                                          entity_type, local_entity);
    }
 
    // Extract degrees of freedom
    auto x_c = md::submdspan(xdofs, c, md::full_extent);
    for (std::int32_t entity_dof : closure_dofs)
      entity_xdofs.push_back(x_c[entity_dof]);
  }
 
  return {std::move(entity_xdofs), eshape};
}

CellPartitionFunction
create_cell_partitioner(mesh::GhostMode ghost_mode, graph::partition_fn partfn,
                        std::optional<std::int32_t> max_facet_to_cell_links);

CellPartitionFunction
create_cell_partitioner(mesh::GhostMode ghost_mode,
                        std::optional<std::int32_t> max_facet_to_cell_links);

std::vector<std::int32_t>
compute_incident_entities(const Topology& topology,
                          std::span<const std::int32_t> entities, int d0,
                          int d1);

template <typename U>
Mesh<typename std::remove_reference_t<typename U::value_type>> create_mesh(
    MPI_Comm comm, MPI_Comm commt,
    std::vector<std::span<const std::int64_t>> cells,
    const std::vector<fem::CoordinateElement<
        typename std::remove_reference_t<typename U::value_type>>>& elements,
    MPI_Comm commg, const U& x, std::array<std::size_t, 2> xshape,
    const CellPartitionFunction& partitioner,
    std::optional<std::int32_t> max_facet_to_cell_links, int num_threads,
    const CellReorderFunction& reorder_fn = graph::reorder_rcm)
{
  if (cells.size() != elements.size())
    throw std::runtime_error("Number of cell arrays and elements must match.");
  std::vector<CellType> celltypes;
  std::ranges::transform(elements, std::back_inserter(celltypes),
                         [](auto& e) { return e.cell_shape(); });
  std::vector<fem::ElementDofLayout> doflayouts;
  std::ranges::transform(elements, std::back_inserter(doflayouts),
                         [](auto& e) { return e.create_dof_layout(); });
 
  // Note: `extract_topology` extracts topology data, i.e. just the
  // vertices. For P1 geometry this should just be the identity
  // operator. For other elements the filtered lists may have 'gaps',
  // i.e. the indices might not be contiguous.
  //
  // `extract_topology` could be skipped for 'P1 geometry' elements
 
  std::int32_t num_cell_types = cells.size();
 
  // -- Partition topology across ranks of comm
  std::vector<std::vector<std::int64_t>> cells1(num_cell_types);
  std::vector<std::vector<std::int64_t>> original_idx1(num_cell_types);
  std::vector<std::vector<int>> ghost_owners(num_cell_types);
  if (partitioner)
  {
    spdlog::info("Using partitioner with cell data ({} cell types)",
                 num_cell_types);
    graph::AdjacencyList<std::int32_t> dest(0);
    if (commt != MPI_COMM_NULL)
    {
      int size = dolfinx::MPI::size(comm);
      std::vector<std::vector<std::int64_t>> t(num_cell_types);
      std::vector<std::span<const std::int64_t>> tspan(num_cell_types);
      for (std::int32_t i = 0; i < num_cell_types; ++i)
      {
        t[i] = extract_topology(celltypes[i], doflayouts[i], cells[i]);
        tspan[i] = std::span(t[i]);
      }
      dest = partitioner(commt, size, celltypes, tspan);
    }
 
    std::int32_t cell_offset = 0;
    for (std::int32_t i = 0; i < num_cell_types; ++i)
    {
      std::size_t num_cell_nodes = doflayouts[i].num_dofs();
      if (cells[i].size() % num_cell_nodes != 0)
      {
        throw std::runtime_error("Cell array size is not a multiple of the "
                                 "number of nodes per cell.");
      }
      std::size_t num_cells = cells[i].size() / num_cell_nodes;
 
      // Extract destination AdjacencyList for this cell type
      std::vector<std::int32_t> offsets_i(
          std::next(dest.offsets().begin(), cell_offset),
          std::next(dest.offsets().begin(), cell_offset + num_cells + 1));
      std::vector<std::int32_t> data_i(
          std::next(dest.array().begin(), offsets_i.front()),
          std::next(dest.array().begin(), offsets_i.back()));
      std::int32_t offset_0 = offsets_i.front();
      std::ranges::for_each(offsets_i,
                            [&offset_0](std::int32_t& j) { j -= offset_0; });
      graph::AdjacencyList<std::int32_t> dest_i(data_i, offsets_i);
      cell_offset += num_cells;
 
      // Distribute cells (topology, includes higher-order 'nodes') to
      // destination rank
      std::vector<int> src_ranks;
      std::tie(cells1[i], src_ranks, original_idx1[i], ghost_owners[i])
          = graph::build::distribute(comm, cells[i],
                                     {num_cells, num_cell_nodes}, dest_i);
      spdlog::debug("Got {} cells from distribution", cells1[i].size());
    }
  }
  else
  {
    // No partitioning, construct a global index
    std::int64_t num_owned = 0;
    for (std::int32_t i = 0; i < num_cell_types; ++i)
    {
      cells1[i] = std::vector<std::int64_t>(cells[i].begin(), cells[i].end());
      std::int32_t num_cell_nodes = doflayouts[i].num_dofs();
      if (cells1[i].size() % num_cell_nodes != 0)
      {
        throw std::runtime_error("Cell array size is not a multiple of the "
                                 "number of nodes per cell.");
      }
      original_idx1[i].resize(cells1[i].size() / num_cell_nodes);
      num_owned += original_idx1[i].size();
    }
 
    // Add on global offset
    std::int64_t global_offset = 0;
    MPI_Exscan(&num_owned, &global_offset, 1, MPI_INT64_T, MPI_SUM, comm);
    for (std::int32_t i = 0; i < num_cell_types; ++i)
    {
      std::iota(original_idx1[i].begin(), original_idx1[i].end(),
                global_offset);
      global_offset += original_idx1[i].size();
    }
  }
 
  // Extract cell 'topology', i.e. extract the vertices for each cell
  // and discard any 'higher-order' nodes
  std::vector<std::vector<std::int64_t>> cells1_v(num_cell_types);
  for (std::int32_t i = 0; i < num_cell_types; ++i)
  {
    cells1_v[i] = extract_topology(celltypes[i], doflayouts[i], cells1[i]);
    spdlog::info("Extract basic topology: {}->{}", cells1[i].size(),
                 cells1_v[i].size());
  }
 
  auto boundary_v_fn
      = create_boundary_vertices_fn(reorder_fn, max_facet_to_cell_links);
  const std::vector<std::int64_t> boundary_v
      = boundary_v_fn(celltypes, doflayouts, ghost_owners, cells1, cells1_v,
                      original_idx1, num_threads);
 
  spdlog::debug("Got {} boundary vertices", boundary_v.size());
 
  // Create Topology
  std::vector<std::span<const std::int64_t>> cells1_v_span;
  std::ranges::transform(cells1_v, std::back_inserter(cells1_v_span),
                         [](auto& c) { return std::span(c); });
  std::vector<std::span<const std::int64_t>> original_idx1_span;
  std::ranges::transform(original_idx1, std::back_inserter(original_idx1_span),
                         [](auto& c) { return std::span(c); });
  std::vector<std::span<const int>> ghost_owners_span;
  std::ranges::transform(ghost_owners, std::back_inserter(ghost_owners_span),
                         [](auto& c) { return std::span(c); });
  Topology topology
      = create_topology(comm, celltypes, cells1_v_span, original_idx1_span,
                        ghost_owners_span, boundary_v, num_threads);
 
  // Create connectivities required higher-order geometries for creating
  // a Geometry object
  for (int i = 0; i < num_cell_types; ++i)
  {
    const auto& entity_dofs = doflayouts[i].entity_dofs_all();
    for (int dim = 1; dim < topology.dim(); ++dim)
    {
      // Accumulate count of all dofs on this dimension
      int dim_sum
          = std::accumulate(entity_dofs[dim].begin(), entity_dofs[dim].end(), 0,
                            [](int c, auto v) { return c + v.size(); });
 
      spdlog::debug("Counting entity dofs, dim={}: {}", dim, dim_sum);
      if (dim_sum > 0)
        topology.create_entities(dim);
    }
 
    if (elements[i].needs_dof_permutations())
      topology.create_entity_permutations();
  }
 
  // Build list of unique (global) node indices from cells1 and
  // distribute coordinate data
  std::vector<std::int64_t> nodes1, nodes2;
  for (std::vector<std::int64_t>& c : cells1)
    nodes1.insert(nodes1.end(), c.begin(), c.end());
  for (std::vector<std::int64_t>& c : cells1)
    nodes2.insert(nodes2.end(), c.begin(), c.end());
 
  dolfinx::radix_sort(nodes1);
  auto [unique_end, range_end] = std::ranges::unique(nodes1);
  nodes1.erase(unique_end, range_end);
 
  std::vector coords
      = dolfinx::MPI::distribute_data(comm, nodes1, commg, x, xshape[1]);
 
  // Create geometry object
  Geometry geometry
      = create_geometry(topology, elements, nodes1, nodes2, coords, xshape[1]);
 
  return Mesh(comm, std::make_shared<Topology>(std::move(topology)),
              std::move(geometry));
}

template <typename U>
Mesh<typename std::remove_reference_t<typename U::value_type>> create_mesh(
    MPI_Comm comm, MPI_Comm commt, std::span<const std::int64_t> cells,
    const fem::CoordinateElement<
        typename std::remove_reference_t<typename U::value_type>>& element,
    MPI_Comm commg, const U& x, std::array<std::size_t, 2> xshape,
    const CellPartitionFunction& partitioner,
    std::optional<std::int32_t> max_facet_to_cell_links, int num_threads,
    const CellReorderFunction& reorder_fn = graph::reorder_rcm)
{
  return create_mesh(comm, commt, std::vector{cells}, std::vector{element},
                     commg, x, xshape, partitioner, max_facet_to_cell_links,
                     num_threads, reorder_fn);
}

template <typename U>
Mesh<typename std::remove_reference_t<typename U::value_type>>
create_mesh(MPI_Comm comm, std::span<const std::int64_t> cells,
            const fem::CoordinateElement<
                std::remove_reference_t<typename U::value_type>>& elements,
            const U& x, std::array<std::size_t, 2> xshape, GhostMode ghost_mode,
            std::optional<std::int32_t> max_facet_to_cell_links = 2)
{
  if (dolfinx::MPI::size(comm) == 1)
  {
    return create_mesh(comm, comm, std::vector{cells}, std::vector{elements},
                       comm, x, xshape, nullptr, max_facet_to_cell_links, 1);
  }
  else
  {
    return create_mesh(
        comm, comm, std::vector{cells}, std::vector{elements}, comm, x, xshape,
        create_cell_partitioner(ghost_mode, max_facet_to_cell_links),
        max_facet_to_cell_links, 1);
  }
}

template <std::floating_point T>
std::pair<Geometry<T>, std::vector<int32_t>>
create_subgeometry(const Mesh<T>& mesh, int dim,
                   std::span<const std::int32_t> subentity_to_entity)
{
  const Geometry<T>& geometry = mesh.geometry();
 
  // Get the geometry dofs in the sub-geometry based on the entities in
  // sub-geometry
  const fem::ElementDofLayout layout
      = geometry.cmaps().front().create_dof_layout();
 
  const std::vector<std::int32_t> x_indices
      = entities_to_geometry(mesh, dim, subentity_to_entity, true).first;
 
  std::vector<std::int32_t> sub_x_dofs = x_indices;
  std::ranges::sort(sub_x_dofs);
  auto [unique_end, range_end] = std::ranges::unique(sub_x_dofs);
  sub_x_dofs.erase(unique_end, range_end);
 
  // Get the sub-geometry dofs owned by this process
  auto x_index_map = geometry.index_map();
  assert(x_index_map);
 
  std::shared_ptr<common::IndexMap> sub_x_dof_index_map;
  std::vector<std::int32_t> subx_to_x_dofmap;
  {
    auto [map, new_to_old] = common::create_sub_index_map(
        *x_index_map, sub_x_dofs, common::IndexMapOrder::any, true);
    sub_x_dof_index_map = std::make_shared<common::IndexMap>(std::move(map));
    subx_to_x_dofmap = std::move(new_to_old);
  }
 
  // Create sub-geometry coordinates
  std::span<const T> x = geometry.x();
  std::int32_t sub_num_x_dofs = subx_to_x_dofmap.size();
  std::vector<T> sub_x(3 * sub_num_x_dofs);
  for (std::int32_t i = 0; i < sub_num_x_dofs; ++i)
  {
    std::copy_n(std::next(x.begin(), 3 * subx_to_x_dofmap[i]), 3,
                std::next(sub_x.begin(), 3 * i));
  }
 
  // Create geometry to sub-geometry  map
  std::vector<std::int32_t> x_to_subx_dof_map(
      x_index_map->size_local() + x_index_map->num_ghosts(), -1);
  for (std::size_t i = 0; i < subx_to_x_dofmap.size(); ++i)
    x_to_subx_dof_map[subx_to_x_dofmap[i]] = i;
 
  // Create sub-geometry dofmap
  std::vector<std::int32_t> sub_x_dofmap;
  sub_x_dofmap.reserve(x_indices.size());
  std::ranges::transform(x_indices, std::back_inserter(sub_x_dofmap),
                         [&x_to_subx_dof_map](auto x_dof)
                         {
                           assert(x_to_subx_dof_map[x_dof] != -1);
                           return x_to_subx_dof_map[x_dof];
                         });
 
  // Sub-geometry coordinate element
  CellType sub_xcell
      = cell_entity_type(geometry.cmaps().front().cell_shape(), dim, 0);
 
  // Special handling of point meshes, as they only support constant
  // basis functions
  int degree
      = (sub_xcell == CellType::point) ? 0 : geometry.cmaps().front().degree();
  fem::CoordinateElement<T> sub_cmap(sub_xcell, degree,
                                     geometry.cmaps().front().variant());
 
  // Sub-geometry input_global_indices
  const std::vector<std::int64_t>& igi = geometry.input_global_indices();
  std::vector<std::int64_t> sub_igi;
  sub_igi.reserve(subx_to_x_dofmap.size());
  std::ranges::transform(subx_to_x_dofmap, std::back_inserter(sub_igi),
                         [&igi](auto sub_x_dof) { return igi[sub_x_dof]; });
 
  // Create geometry
  return {Geometry(
              sub_x_dof_index_map,
              std::vector<std::vector<std::int32_t>>{std::move(sub_x_dofmap)},
              {sub_cmap}, std::move(sub_x), geometry.dim(), std::move(sub_igi)),
          std::move(subx_to_x_dofmap)};
}

template <std::floating_point T>
std::tuple<Mesh<T>, EntityMap, EntityMap, std::vector<std::int32_t>>
create_submesh(const Mesh<T>& mesh, int dim,
               std::span<const std::int32_t> entities)
{
  // Create sub-topology
  mesh.topology_mutable()->create_connectivity(dim, 0);
  auto [topology, subentity_to_entity, subvertex_to_vertex]
      = mesh::create_subtopology(*mesh.topology(), dim, entities);
 
  // Create sub-geometry
  const int tdim = mesh.topology()->dim();
  mesh.topology_mutable()->create_entities(dim);
  mesh.topology_mutable()->create_connectivity(dim, tdim);
  mesh.topology_mutable()->create_connectivity(tdim, dim);
  mesh.topology_mutable()->create_entity_permutations();
  auto [geometry, subx_to_x_dofmap]
      = mesh::create_subgeometry(mesh, dim, subentity_to_entity);
 
  Mesh<T> submesh
      = Mesh(mesh.comm(), std::make_shared<Topology>(std::move(topology)),
             std::move(geometry));
  EntityMap entity_map(mesh.topology(), submesh.topology(), dim,
                       subentity_to_entity);
  EntityMap vertex_map(mesh.topology(), submesh.topology(), 0,
                       subvertex_to_vertex);
  return {std::move(submesh), std::move(entity_map), std::move(vertex_map),
          std::move(subx_to_x_dofmap)};
}

template <typename T>
MeshTags<T> transfer_meshtags_to_submesh(
    const MeshTags<T>& tags,
    std::shared_ptr<const dolfinx::mesh::Topology> submesh_topology,
    const EntityMap& vertex_map, const EntityMap& cell_map)
{
  int tag_dim = tags.dim();
  int submesh_tdim = submesh_topology->dim();
  auto topology = tags.topology();
  if (tag_dim > submesh_tdim)
  {
    throw std::runtime_error("Tag dimension must be less than or equal to "
                             "submesh dimension");
  }
  std::shared_ptr<const dolfinx::common::IndexMap> sub_cell_imap
      = submesh_topology->index_map(submesh_tdim);
  if (!sub_cell_imap)
  {
    throw std::runtime_error(
        std::format("Entities of dimension {} does not exist in mesh topology.",
                    submesh_tdim));
  }
 
  // Create a map from parent entity to submesh cell
  std::int32_t submesh_num_cells
      = sub_cell_imap->size_local() + sub_cell_imap->num_ghosts();
  auto sub_cells = std::ranges::views::iota(0, submesh_num_cells);
  std::vector<std::int32_t> sub_cell_to_parent_entity
      = cell_map.sub_topology_to_topology(sub_cells, false);
 
  // Create a full lookup for all cells on the parent mesh, as the tag can have
  // entities that are not in the submesh
  auto parent_entity_imap = topology->index_map(submesh_tdim);
  if (!parent_entity_imap)
  {
    throw std::runtime_error(std::format(
        "Entities of dimension {} does not exist in parent mesh topology.",
        submesh_tdim));
  }
  std::size_t num_parent_entities
      = parent_entity_imap->size_local() + parent_entity_imap->num_ghosts();
  std::vector<std::int32_t> parent_entity_to_sub_cell(num_parent_entities, -1);
  for (std::size_t i = 0; i < sub_cell_to_parent_entity.size(); ++i)
    parent_entity_to_sub_cell[sub_cell_to_parent_entity[i]]
        = static_cast<std::int32_t>(i);
 
  // Get map from submesh vertex to parent vertex
  std::vector<std::int32_t> sub_to_parent_vertex;
  {
    auto sub_vertex_map = submesh_topology->index_map(0);
    std::int32_t num_sub_vertices
        = sub_vertex_map->size_local() + sub_vertex_map->num_ghosts();
    auto sub_vertices = std::ranges::views::iota(0, num_sub_vertices);
 
    sub_to_parent_vertex
        = vertex_map.sub_topology_to_topology(sub_vertices, false);
  }
  // Access various connectivity maps
  auto sub_e_to_v = submesh_topology->connectivity(tag_dim, 0);
  auto sub_c_to_e = submesh_topology->connectivity(submesh_tdim, tag_dim);
  auto sub_entity_imap = submesh_topology->index_map(tag_dim);
  auto e_to_v = topology->connectivity(tag_dim, 0);
  std::shared_ptr<const dolfinx::graph::AdjacencyList<std::int32_t>>
      e_to_sub_cell = nullptr;
  if (tag_dim != submesh_tdim)
  {
    e_to_sub_cell = topology->connectivity(tag_dim, submesh_tdim);
    if (!e_to_sub_cell)
    {
      throw std::runtime_error(
          std::format("Missing connectivity between {} and {} in parent mesh",
                      tag_dim, submesh_tdim));
    }
  }
 
  if (!sub_e_to_v)
  {
    throw std::runtime_error(std::format(
        "Missing connectivity between {} and {} in submesh", tag_dim, 0));
  }
  if (!sub_c_to_e)
  {
    throw std::runtime_error(
        std::format("Missing connectivity between {} and {} in submesh",
                    submesh_tdim, tag_dim));
  }
  if (!sub_entity_imap)
  {
    throw std::runtime_error(std::format(
        "Entities of dimension {} does not exist in submesh topology.",
        tag_dim));
  }
  if (!e_to_v)
  {
    throw std::runtime_error(
        std::format("Missing connectivity between {} and 0", tag_dim));
  }
 
  // Prepare sub entity to parent map
  std::size_t num_sub_entities
      = sub_entity_imap->size_local() + sub_entity_imap->num_ghosts();
  constexpr T max_val = std::numeric_limits<T>::max();
  std::vector<T> submesh_values(num_sub_entities, max_val);
  std::vector<std::int32_t> submesh_indices(num_sub_entities);
  std::iota(submesh_indices.begin(), submesh_indices.end(), 0);
 
  std::span<const std::int32_t> tagged_entities = tags.indices();
  std::span<const T> tagged_values = tags.values();
  // For each entity in the tag, find all cells of the submesh connected to this
  // entity
  for (std::size_t i = 0; i < tagged_entities.size(); ++i)
  {
    auto find_and_map_sub_entity
        = [tag_dim, submesh_tdim, &e_to_v, &parent_entity_to_sub_cell,
           &sub_to_parent_vertex, &sub_e_to_v, &sub_c_to_e,
           &e_to_sub_cell](std::int32_t entity)
    {
      // Fast exit if the tag dimension is the same as the submesh dimension,
      // as we can directly map the parent entity to the submesh cell
      if (tag_dim == submesh_tdim)
        return parent_entity_to_sub_cell[entity];
 
      // Given an entity in the parent meshtag, find all submesh-cells that are
      // entities in parent mesh that contain this entity.
      auto entity_vertices = e_to_v->links(entity);
      auto parent_sub_cells = e_to_sub_cell->links(entity);
      auto submesh_cells
          = parent_sub_cells
            | std::views::transform([&parent_entity_to_sub_cell](auto c)
                                    { return parent_entity_to_sub_cell[c]; })
            | std::views::filter([](auto sub_cell) { return sub_cell != -1; });
      for (auto sub_cell : submesh_cells)
      {
        for (auto sub_entity : sub_c_to_e->links(sub_cell))
        {
          // Convert submesh entity vertices to parent vertices
          auto parent_vertices
              = sub_e_to_v->links(sub_entity)
                | std::views::transform([&sub_to_parent_vertex](auto v)
                                        { return sub_to_parent_vertex[v]; });
 
          // Check if all parent vertices of the submesh entity are in the
          // parent entity
          bool entity_matches = std::ranges::all_of(
              parent_vertices,
              [&entity_vertices](auto p_v)
              {
                // With C++23 this can use std::ranges::contains
                return std::ranges::find(entity_vertices, p_v)
                       != std::ranges::end(entity_vertices);
              });
 
          // If a match is found, apply values and exit the lambda immediately
          if (entity_matches)
            return sub_entity;
        }
      }
      return -1;
    };
 
    // Execute the search for the current entity
    std::int32_t sub_entity = find_and_map_sub_entity(tagged_entities[i]);
    if (sub_entity != -1)
      submesh_values[sub_entity] = tagged_values[i];
  }
 
  // Filter out the entities that were never mapped (values still equal max)
  std::vector<std::int32_t> filtered_indices;
  std::vector<T> filtered_values;
  filtered_indices.reserve(num_sub_entities);
  filtered_values.reserve(num_sub_entities);
  for (std::size_t i = 0; i < submesh_values.size(); ++i)
  {
    if (submesh_values[i] != max_val)
    {
      filtered_indices.push_back(submesh_indices[i]);
      filtered_values.push_back(submesh_values[i]);
    }
  }
  filtered_indices.shrink_to_fit();
  filtered_values.shrink_to_fit();
  MeshTags<T> new_meshtag(submesh_topology, tag_dim, filtered_indices,
                          filtered_values, tags.name());
  return new_meshtag;
}
 
} // namespace dolfinx::mesh