Multigrid: skip re-evaluation of unrefined nodes

firedrake/cython/mgimpl.pyx

@@ -353,6 +353,99 @@ def adaptive_parent_child_cell_maps(PETSc.DM coarse_dm,
     return np.asarray(coarse_to_fine), np.asarray(fine_to_coarse)
 
 
+@cython.boundscheck(False)
+@cython.wraparound(False)
+def preserved_points(PETSc.DM coarse_dm,
+                     PETSc.Section coarse_cell_numbering,
+                     PETSc.DM fine_dm,
+                     PETSc.Section fine_cell_numbering,
+                     np.ndarray coarse_to_fine_cells):
+    """Pair unrefined fine points with their coarse originals.
+
+    Adaptive refinement copies an untouched coarse cell into the fine mesh
+    without change. It therefore preserves the cone of every point in that
+    cell, and so preserves the whole plex closure, point for point. Such a
+    cell has exactly one child. The right-padding of ``coarse_to_fine_cells``
+    with -1 identifies which cells these are.
+
+    Parameters
+    ----------
+    coarse_dm : PETSc.DM
+        The coarse mesh DMPlex.
+    coarse_cell_numbering : PETSc.Section
+        The cell numbering section of the coarse mesh.
+    fine_dm : PETSc.DM
+        The adaptively refined DMPlex.
+    fine_cell_numbering : PETSc.Section
+        The cell numbering section of the fine mesh.
+    coarse_to_fine_cells : numpy.ndarray
+        The Firedrake-numbered coarse-to-fine cell map.
+
+    Returns
+    -------
+    numpy.ndarray
+        An array over the chart of ``fine_dm``. For each fine point, it
+        holds the coarse point it was copied from, or -1 if refinement
+        changed that point.
+
+    """
+    cdef:
+        PetscInt ncoarse, nfine, max_children, c, i, off, child
+        PetscInt cStart, cEnd, pStart, pEnd, coarse_size, fine_size
+        PetscInt *coarse_closure = NULL
+        PetscInt *fine_closure = NULL
+        PetscInt[::1] coarse_point, fine_point, fine_to_coarse
+        PetscInt[:, ::1] coarse_to_fine
+
+    coarse_to_fine = coarse_to_fine_cells
+    ncoarse = num_owned_cells(coarse_dm)
+    assert ncoarse == coarse_to_fine.shape[0]
+    max_children = coarse_to_fine.shape[1]
+    nfine = num_owned_cells(fine_dm)
+
+    # Both cell maps are in Firedrake numbering, so invert each mesh's cell
+    # numbering section to get back to the plex points the closures live on.
+    coarse_point = np.full(ncoarse, -1, dtype=IntType)
+    cStart, cEnd = coarse_dm.getHeightStratum(0)
+    for c in range(cStart, cEnd):
+        CHKERR(PetscSectionGetOffset(coarse_cell_numbering.sec, c, &off))
+        if 0 <= off < ncoarse:
+            coarse_point[off] = c
+    fine_point = np.full(nfine, -1, dtype=IntType)
+    cStart, cEnd = fine_dm.getHeightStratum(0)
+    for c in range(cStart, cEnd):
+        CHKERR(PetscSectionGetOffset(fine_cell_numbering.sec, c, &off))
+        if 0 <= off < nfine:
+            fine_point[off] = c
+
+    pStart, pEnd = fine_dm.getChart()
+    fine_to_coarse = np.full(pEnd - pStart, -1, dtype=IntType)
+    for c in range(ncoarse):
+        child = coarse_to_fine[c, 0]
+        if child < 0 or (max_children > 1 and coarse_to_fine[c, 1] >= 0):
+            continue
+        if coarse_point[c] < 0 or fine_point[child] < 0:
+            continue
+        CHKERR(DMPlexGetTransitiveClosure(coarse_dm.dm, coarse_point[c], PETSC_TRUE,
+                                          &coarse_size, &coarse_closure))
+        CHKERR(DMPlexGetTransitiveClosure(fine_dm.dm, fine_point[child], PETSC_TRUE,
+                                          &fine_size, &fine_closure))
+        # A one-child cell that refinement did change would have a closure
+        # of a different size. Skip it and let the transfer kernel handle it.
+        if coarse_size == fine_size:
+            for i in range(coarse_size):
+                # Each closure interleaves a point with its orientation. Copy
+                # a point only when its orientation matches in both meshes:
+                # only then do the two cells order their nodes the same way.
+                if coarse_closure[2*i + 1] == fine_closure[2*i + 1]:
+                    fine_to_coarse[fine_closure[2*i] - pStart] = coarse_closure[2*i]
+        CHKERR(DMPlexRestoreTransitiveClosure(coarse_dm.dm, coarse_point[c], PETSC_TRUE,
+                                              &coarse_size, &coarse_closure))
+        CHKERR(DMPlexRestoreTransitiveClosure(fine_dm.dm, fine_point[child], PETSC_TRUE,
+                                              &fine_size, &fine_closure))
+    return np.asarray(fine_to_coarse)
+
+
 # Exposition:
 #
 # These next functions compute maps from coarse mesh cells to fine

firedrake/mg/interface.py

@@ -102,7 +102,11 @@ def prolong(coarse, fine):
         for d in [coarse, coarse_coords]:
             d.dat.global_to_local_begin(op2.READ)
             d.dat.global_to_local_end(op2.READ)
-        op2.par_loop(kernel, fine.node_set, *kernel_args)
+        # Adaptive refinement leaves most of the mesh unchanged. Copy the
+        # value at those nodes instead of evaluating it there.
+        node_subset = utils.transfer_node_subset(Vc, Vf)
+        op2.par_loop(kernel, node_subset, *kernel_args)
+        utils.prolong_preserved_nodes(coarse, fine)
 
         if needs_quadrature:
             # Transfer to the actual target space
@@ -184,7 +188,11 @@ def restrict(fine_dual, coarse_dual):
         for d in [coarse_coords]:
             d.dat.global_to_local_begin(op2.READ)
             d.dat.global_to_local_end(op2.READ)
-        op2.par_loop(kernel, fine_dual.node_set, *kernel_args)
+        # Restriction is the transpose of prolongation. It skips the same
+        # fine nodes and adds their value to the matching coarse node.
+        node_subset = utils.transfer_node_subset(Vc, Vf)
+        op2.par_loop(kernel, node_subset, *kernel_args)
+        utils.restrict_preserved_nodes(fine_dual, coarse_dual)
         fine_dual = coarse_dual
     return coarse_dual
 

firedrake/mg/utils.py

@@ -1,11 +1,14 @@
 import numpy
 from fractions import Fraction
+from mpi4py import MPI
 from pyop2 import op2
+from firedrake.petsc import PETSc
 from firedrake.utils import IntType
 from firedrake.functionspacedata import entity_dofs_key
 import finat.ufl
 import firedrake
 from firedrake.cython import mgimpl as impl
+from firedrake.halo import _get_mtype
 
 
 def fine_node_to_coarse_node_map(Vf, Vc):
@@ -307,6 +310,213 @@ def coarse_cell_child_count(Vc, Vf):
         return cache.setdefault(key, op2.Dat(dset, counts, dtype=IntType))
 
 
+def _preserved_point_sf(coarse_mesh, fine_mesh, coarse_to_fine):
+    """Create the SF that pairs unrefined points with their coarse originals.
+
+    Adaptive refinement leaves some cells untouched. This SF maps each
+    unrefined point in ``fine_mesh`` back to the coarse point it came from.
+
+    Parameters
+    ----------
+    coarse_mesh : firedrake.mesh.AbstractMeshTopology
+        The mesh before refinement.
+    fine_mesh : firedrake.mesh.AbstractMeshTopology
+        The mesh after refinement.
+    coarse_to_fine : numpy.ndarray
+        The coarse-to-fine cell map that relates the two meshes.
+
+    Returns
+    -------
+    PETSc.SF
+        An SF with roots on the points of ``coarse_mesh`` and leaves on the
+        unrefined points of ``fine_mesh``. Returns `None` if refinement
+        changed every cell, as a uniform refinement does.
+
+    """
+    coarse_plex = coarse_mesh.topology_dm
+    fine_plex = fine_mesh.topology_dm
+    fine_to_coarse_points = impl.preserved_points(
+        coarse_plex, coarse_mesh._cell_numbering,
+        fine_plex, fine_mesh._cell_numbering,
+        coarse_to_fine,
+    )
+    leaves, = numpy.nonzero(fine_to_coarse_points >= 0)
+    # A uniform refinement preserves no points. Every rank must agree on
+    # whether to build the SF at all, not just the ranks with no leaves.
+    if not fine_plex.comm.tompi4py().allreduce(len(leaves) > 0, op=MPI.LOR):
+        return None
+    leaves = leaves.astype(IntType)
+    # Refinement acts on each rank's own plex. A fine point and the coarse
+    # point it was copied from always live on the same rank.
+    remote = numpy.empty((len(leaves), 2), dtype=IntType)
+    remote[:, 0] = coarse_plex.comm.rank
+    remote[:, 1] = fine_to_coarse_points[leaves]
+    pStart, pEnd = coarse_plex.getChart()
+    point_sf = PETSc.SF().create(comm=coarse_plex.comm)
+    point_sf.setGraph(pEnd - pStart, leaves, remote)
+    return point_sf
+
+
+def preserved_node_sf(Vc, Vf):
+    """Find the nodes that adaptive refinement leaves unchanged.
+
+    An unrefined cell has the same nodes in both spaces. The transfer
+    operators can then copy values between them instead of evaluating them.
+    This is cheaper, and exact.
+
+    Parameters
+    ----------
+    Vc : firedrake.functionspaceimpl.WithGeometry
+        The coarse function space.
+    Vf : firedrake.functionspaceimpl.WithGeometry
+        The fine function space, on the next level of the same hierarchy.
+
+    Returns
+    -------
+    PETSc.SF
+        An SF with roots on the nodes of ``Vc`` and leaves on the matching
+        nodes of ``Vf``. Returns `None` if no nodes match.
+
+    """
+    if Vc.ufl_element() != Vf.ufl_element() or Vc.boundary_set != Vf.boundary_set:
+        # A space and its counterpart on the refined mesh use the same node
+        # layout on an unrefined cell only when the element and the boundary
+        # set both match.
+        return None
+    if Vc.extruded or Vf.extruded:
+        # The DMPlex of an extruded mesh stores only the 2D base mesh. Each
+        # point there represents a whole vertical column of nodes, and a
+        # Section cannot address one node within that column. Give up here
+        # and let the transfer kernel evaluate every node instead.
+        return None
+    hierarchy, levelc = get_level(Vc.mesh())
+    _, levelf = get_level(Vf.mesh())
+    if hierarchy is None or levelc + Fraction(1, hierarchy.refinements_per_level) != levelf:
+        return None
+    cache = Vf.mesh().topology._shared_data_cache["hierarchy_preserved_node_sf"]
+    key = _cache_key(Vc, Vf)
+    try:
+        return cache[key]
+    except KeyError:
+        coarse_to_fine = hierarchy.coarse_to_fine_cells[levelc]
+        point_sf = _preserved_point_sf(Vc.mesh().topology, Vf.mesh().topology,
+                                       coarse_to_fine)
+        if point_sf is None:
+            return cache.setdefault(key, None)
+        root_section = Vc.dm.getSection()
+        leaf_section = Vf.dm.getSection()
+        # `distributeSection` builds its own section over the range of points
+        # that the SF touches. Only the broadcast root offsets are needed
+        # here. Pad them back out to the full chart that `createSectionSF`
+        # expects.
+        remote_offsets, distributed_section = point_sf.distributeSection(root_section)
+        pStart, pEnd = leaf_section.getChart()
+        lpStart, lpEnd = distributed_section.getChart()
+        offsets = numpy.zeros(pEnd - pStart, dtype=IntType)
+        offsets[lpStart - pStart:lpEnd - pStart] = remote_offsets
+        section_sf = point_sf.createSectionSF(root_section, offsets, leaf_section)
+        # The transfer kernels compute only the owned fine nodes and leave
+        # the halo to a later exchange. Keep only the owned leaves here too:
+        # a ghost fine node reduced onto its coarse node would count twice.
+        nroots, ilocal, iremote = section_sf.getGraph()
+        owned = ilocal < Vf.node_set.size
+        trimmed = PETSc.SF().create(comm=section_sf.comm)
+        trimmed.setGraph(nroots, ilocal[owned], iremote[owned])
+        return cache.setdefault(key, trimmed)
+
+
+def transfer_node_subset(Vc, Vf):
+    """Find the fine nodes that the transfer kernels must evaluate.
+
+    These are the nodes of ``Vf`` that :func:`preserved_node_sf` does not
+    already account for. Prolongation and restriction can copy the rest.
+
+    Parameters
+    ----------
+    Vc : firedrake.functionspaceimpl.WithGeometry
+        The coarse function space.
+    Vf : firedrake.functionspaceimpl.WithGeometry
+        The fine function space, on the next level of the same hierarchy.
+
+    Returns
+    -------
+    pyop2.types.set.Set or pyop2.types.set.Subset
+        A subset of the nodes of ``Vf``, or ``Vf.node_set`` itself if
+        :func:`preserved_node_sf` found no preserved nodes.
+
+    """
+    section_sf = preserved_node_sf(Vc, Vf)
+    if section_sf is None:
+        return Vf.node_set
+    cache = Vf.mesh().topology._shared_data_cache["hierarchy_transfer_node_subset"]
+    key = _cache_key(Vc, Vf)
+    try:
+        return cache[key]
+    except KeyError:
+        _, preserved, _ = section_sf.getGraph()
+        nodes = numpy.setdiff1d(numpy.arange(Vf.node_set.size, dtype=IntType),
+                                preserved)
+        return cache.setdefault(key, op2.Subset(Vf.node_set, nodes))
+
+
+def prolong_preserved_nodes(coarse, fine):
+    """Copy coarse values onto the fine nodes that adaptive refinement preserved.
+
+    Parameters
+    ----------
+    coarse : firedrake.function.Function
+        The function on the coarse mesh.
+    fine : firedrake.function.Function
+        The function on the refined mesh. The transfer kernel has already
+        computed its other nodes.
+
+    """
+
+    section_sf = preserved_node_sf(coarse.function_space(), fine.function_space())
+    if section_sf is None:
+        return
+    mtype, _ = _get_mtype(fine.dat)
+    # The source coarse node can be a ghost node. Only owned fine nodes are
+    # written here, the same as the transfer kernel writes.
+    source = coarse.dat.data_ro_with_halos
+    target = fine.dat.data_wo
+    section_sf.bcastBegin(mtype, source, target, MPI.REPLACE)
+    section_sf.bcastEnd(mtype, source, target, MPI.REPLACE)
+
+
+def restrict_preserved_nodes(fine_dual, coarse_dual):
+    """Add the contribution of preserved nodes to the coarse dual.
+
+    Prolongation copies a preserved node's value without change. Restriction
+    is its transpose, so it adds the fine value to the coarse node unchanged.
+
+    Parameters
+    ----------
+    fine_dual : firedrake.cofunction.Cofunction
+        The cofunction on the refined mesh.
+    coarse_dual : firedrake.cofunction.Cofunction
+        The cofunction on the coarse mesh. It already holds the contribution
+        that the transfer kernel accumulated from the other fine nodes.
+
+    """
+
+    coarse_V = coarse_dual.function_space()
+    section_sf = preserved_node_sf(coarse_V, fine_dual.function_space())
+    if section_sf is None:
+        return
+    buffer = firedrake.Function(coarse_V)
+    mtype, _ = _get_mtype(buffer.dat)
+    source = fine_dual.dat.data_ro
+    target = buffer.dat.data_wo_with_halos
+    section_sf.reduceBegin(mtype, source, target, MPI.SUM)
+    section_sf.reduceEnd(mtype, source, target, MPI.SUM)
+    # A preserved coarse node can be a ghost on the rank that owns the
+    # matching fine node. Reduce the contributions onto the owning rank.
+    buffer.dat.local_to_global_begin(op2.INC)
+    buffer.dat.local_to_global_end(op2.INC)
+    coarse_dual.dat.data[...] += buffer.dat.data_ro
+
+
 def physical_node_locations(V):
     element = V.ufl_element()
     if V.value_shape: