Source code for firedrake.preconditioners.pmg

from functools import partial
from itertools import chain
from firedrake.dmhooks import (attach_hooks, get_appctx, push_appctx, pop_appctx,
                               add_hook, get_parent, push_parent, pop_parent,
                               get_function_space, set_function_space)
from firedrake.petsc import PETSc
from firedrake.preconditioners.base import PCBase, SNESBase, PCSNESBase
from firedrake.nullspace import VectorSpaceBasis, MixedVectorSpaceBasis
from firedrake.solving_utils import _SNESContext
from firedrake.tsfc_interface import extract_numbered_coefficients
from firedrake.utils import ScalarType_c, IntType_c, cached_property
from tsfc.finatinterface import create_element
from tsfc import compile_expression_dual_evaluation
from pyop2 import op2
from pyop2.caching import cached
from pyop2.utils import as_tuple

import firedrake
import finat
import FIAT
import ufl
import finat.ufl
import loopy
import numpy
import os
import tempfile
import weakref

__all__ = ("PMGPC", "PMGSNES")


class PMGBase(PCSNESBase):
    """A class for implementing p-multigrid.

    Internally, this sets up a DM with a custom coarsen routine
    that p-coarsens the problem. This DM is passed to an internal
    PETSc PC of type MG and with options prefix ``pmg_``. The
    relaxation to apply on every p-level is described by ``pmg_mg_levels_``,
    and the coarse solve by ``pmg_mg_coarse_``. Geometric multigrid
    or any other solver in firedrake may be applied to the coarse problem.

    Other PETSc options inspected by this class are:
    - 'pmg_mg_coarse_degree': polynomial degree of the coarse level
    - 'pmg_mg_coarse_mat_type': can be either a `PETSc.Mat.Type`, or 'matfree'
    - 'pmg_mg_coarse_pmat_type': can be either a `PETSc.Mat.Type`, or 'matfree'
    - 'pmg_mg_coarse_form_compiler_mode': can be 'spectral' (default), 'vanilla', 'coffee', or 'tensor'
    - 'pmg_mg_levels_transfer_mat_type': can be either 'aij' or 'matfree'

    The p-coarsening is implemented in the `coarsen_element` routine.
    This takes in a :class:`finat.ufl.finiteelement.FiniteElement` and either returns a
    new, coarser element, or raises a `ValueError` (if the supplied element
    should be the coarsest one of the hierarchy).

    The default coarsen_element is to perform power-of-2 reduction
    of the polynomial degree.
    It is expected that some applications of this preconditioner
    will subclass :class:`PMGBase` to override `coarsen_element` and
    `coarsen_form`.
    """

    _prefix = "pmg_"
    # This is parallel-safe because the keys are ids of a collective objects
    _coarsen_cache = weakref.WeakKeyDictionary()
    _transfer_cache = weakref.WeakKeyDictionary()

    def coarsen_element(self, ele):
        """Coarsen a given element to form the next problem down in the p-hierarchy.

        If the supplied element should form the coarsest level of the p-hierarchy,
        raise `ValueError`. Otherwise, return a new :class:`finat.ufl.finiteelement.FiniteElement`.

        By default, this does power-of-2 coarsening in polynomial degree until
        we reach the coarse degree specified through PETSc options (1 by default).

        Parameters
        ----------
        ele :
            A :class:`finat.ufl.finiteelement.FiniteElement` to coarsen.
        """
        degree = PMGBase.max_degree(ele)
        if degree <= self.coarse_degree:
            raise ValueError
        return PMGBase.reconstruct_degree(ele, max(degree//2, self.coarse_degree))

    def coarsen_form(self, form, fine_to_coarse_map):
        """Coarsen a form, by replacing the solution, test and trial functions.
        """
        return ufl.replace(form, fine_to_coarse_map)

    def initialize(self, obj):
        # Make a new DM.
        # Hook up a (new) coarsen routine on that DM.
        # Make a new PC, of type MG (or SNES of type FAS).
        # Assign the DM to that PC (or SNES).

        odm = obj.getDM()
        ctx = get_appctx(odm)
        if ctx is None:
            raise ValueError("No context found.")
        if not isinstance(ctx, _SNESContext):
            raise ValueError("Don't know how to get form from %r" % ctx)
        fcp = ctx._problem.form_compiler_parameters
        mode = fcp.get("mode", "spectral") if fcp is not None else "spectral"

        test, trial = ctx.J.arguments()
        if test.function_space() != trial.function_space():
            raise NotImplementedError("test and trial spaces must be the same")

        prefix = obj.getOptionsPrefix()
        options_prefix = prefix + self._prefix
        pdm = PETSc.DMShell().create(comm=obj.comm)
        pdm.setOptionsPrefix(options_prefix)

        ppc = self.configure_pmg(obj, pdm)
        self.is_snes = isinstance(obj, PETSc.SNES)

        # Get the coarse degree from PETSc options
        copts = PETSc.Options(ppc.getOptionsPrefix() + ppc.getType() + "_coarse_")
        self.coarse_degree = copts.getInt("degree", default=1)
        self.coarse_mat_type = copts.getString("mat_type", default=ctx.mat_type)
        self.coarse_pmat_type = copts.getString("pmat_type", default=self.coarse_mat_type)
        self.coarse_form_compiler_mode = copts.getString("form_compiler_mode", default=mode)

        # Construct a list with the elements we'll be using
        V = test.function_space()
        ele = V.ufl_element()
        elements = [ele]
        while True:
            try:
                ele_ = self.coarsen_element(ele)
                assert ele_.value_shape == ele.value_shape
                ele = ele_
            except ValueError:
                break
            elements.append(ele)

        sf = odm.getPointSF()
        section = odm.getDefaultSection()
        attach_hooks(pdm, level=len(elements)-1, sf=sf, section=section)
        # Now overwrite some routines on the DM
        pdm.setRefine(None)
        pdm.setCoarsen(self.coarsen)
        if self.is_snes:
            pdm.setSNESFunction(_SNESContext.form_function)
            pdm.setSNESJacobian(_SNESContext.form_jacobian)
            pdm.setKSPComputeOperators(_SNESContext.compute_operators)

        set_function_space(pdm, get_function_space(odm))

        parent = get_parent(odm)
        assert parent is not None
        add_hook(parent, setup=partial(push_parent, pdm, parent), teardown=partial(pop_parent, pdm, parent), call_setup=True)
        add_hook(parent, setup=partial(push_appctx, pdm, ctx), teardown=partial(pop_appctx, pdm, ctx), call_setup=True)

        ppc.incrementTabLevel(1, parent=obj)
        ppc.setFromOptions()
        ppc.setUp()
        self.ppc = ppc

    def update(self, obj):
        self.ppc.setUp()

    def view(self, obj, viewer=None):
        if viewer is None:
            viewer = PETSc.Viewer.STDOUT
        viewer.printfASCII("p-multigrid PC\n")
        if hasattr(self, "ppc"):
            self.ppc.view(viewer=viewer)

    def destroy(self, obj):
        if hasattr(self, "ppc"):
            self.ppc.destroy()

    def coarsen(self, fdm, comm):
        # Coarsen the _SNESContext of a DM fdm
        # return the coarse DM cdm of the coarse _SNESContext
        fctx = get_appctx(fdm)
        parent = get_parent(fdm)
        assert parent is not None

        test, trial = fctx.J.arguments()
        fV = test.function_space()
        cele = self.coarsen_element(fV.ufl_element())

        # Have we already done this?
        cctx = fctx._coarse
        if cctx is not None:
            cV = cctx.J.arguments()[0].function_space()
            if (cV.ufl_element() == cele) and (cV.mesh() == fV.mesh()):
                return cV.dm

        cV = firedrake.FunctionSpace(fV.mesh(), cele)
        cdm = cV.dm

        fproblem = fctx._problem
        fu = fproblem.u
        cu = firedrake.Function(cV)

        fdeg = PMGBase.max_degree(fV.ufl_element())
        cdeg = PMGBase.max_degree(cV.ufl_element())

        fine_to_coarse_map = {fu: cu,
                              test: test.reconstruct(function_space=cV),
                              trial: trial.reconstruct(function_space=cV)}

        def _coarsen_form(a):
            if isinstance(a, ufl.Form):
                a = self.coarsen_form(a, fine_to_coarse_map)
                a = type(a)([f.reconstruct(metadata=self.coarsen_quadrature(f.metadata(), fdeg, cdeg))
                             for f in a.integrals()])
            return a

        cJ = _coarsen_form(fctx.J)
        cJp = cJ if fctx.Jp is fctx.J else _coarsen_form(fctx.Jp)
        # This fixes a subtle bug where you are applying PMGPC on a mixed
        # problem with geometric multigrid only on one block and an non-Lagrange element
        # on the other block (gmg breaks for non-Lagrange elements)
        cF = _coarsen_form(fctx.F) if self.is_snes else ufl.action(cJ, cu)

        fcp = self.coarsen_quadrature(fproblem.form_compiler_parameters, fdeg, cdeg)
        cbcs = self.coarsen_bcs(fproblem.bcs, cV)

        # Coarsen the appctx: the user might want to provide solution-dependant expressions and forms
        cappctx = dict(fctx.appctx)
        for key in cappctx:
            val = cappctx[key]
            if isinstance(val, dict):
                cappctx[key] = self.coarsen_quadrature(val, fdeg, cdeg)
            elif isinstance(val, ufl.Form):
                cappctx[key] = _coarsen_form(val)
            elif isinstance(val, ufl.classes.Expr):
                cappctx[key] = ufl.replace(val, fine_to_coarse_map)

        # If we're the coarsest grid of the p-hierarchy, don't
        # overwrite the coarsen routine; this is so that you can
        # use geometric multigrid for the p-coarse problem
        try:
            self.coarsen_element(cele)
            cdm.setCoarsen(self.coarsen)
            mat_type = fctx.mat_type
            pmat_type = fctx.pmat_type
        except ValueError:
            mat_type = self.coarse_mat_type
            pmat_type = self.coarse_pmat_type
            fcp = dict(fcp or {}, mode=self.coarse_form_compiler_mode)

        # Coarsen the problem and the _SNESContext
        cproblem = firedrake.NonlinearVariationalProblem(cF, cu, bcs=cbcs, J=cJ, Jp=cJp,
                                                         form_compiler_parameters=fcp,
                                                         is_linear=fproblem.is_linear)

        cctx = type(fctx)(cproblem, mat_type, pmat_type,
                          appctx=cappctx,
                          pre_jacobian_callback=fctx._pre_jacobian_callback,
                          pre_function_callback=fctx._pre_function_callback,
                          post_jacobian_callback=fctx._post_jacobian_callback,
                          post_function_callback=fctx._post_function_callback,
                          options_prefix=fctx.options_prefix,
                          transfer_manager=fctx.transfer_manager)

        # FIXME setting up the _fine attribute triggers gmg injection.
        # cctx._fine = fctx
        fctx._coarse = cctx

        add_hook(parent, setup=partial(push_parent, cdm, parent), teardown=partial(pop_parent, cdm, parent), call_setup=True)
        add_hook(parent, setup=partial(push_appctx, cdm, cctx), teardown=partial(pop_appctx, cdm, cctx), call_setup=True)

        cdm.setOptionsPrefix(fdm.getOptionsPrefix())
        cdm.setKSPComputeOperators(_SNESContext.compute_operators)
        cdm.setCreateInterpolation(self.create_interpolation)
        cdm.setCreateInjection(self.create_injection)

        if cu in cJ.coefficients():
            # Only inject state if the coarse state is a dependency of the coarse Jacobian.
            inject = cdm.createInjection(fdm)

            def inject_state():
                with cu.dat.vec_wo as xc, fu.dat.vec_ro as xf:
                    inject.mult(xf, xc)

            add_hook(parent, setup=inject_state, call_setup=True)

        interpolate = None
        if fctx._nullspace or fctx._nullspace_T or fctx._near_nullspace:
            interpolate, _ = cdm.createInterpolation(fdm)
        cctx._nullspace = self.coarsen_nullspace(fctx._nullspace, cV, interpolate)
        cctx._nullspace_T = self.coarsen_nullspace(fctx._nullspace_T, cV, interpolate)
        cctx._near_nullspace = self.coarsen_nullspace(fctx._near_nullspace, cV, interpolate)
        cctx.set_nullspace(cctx._nullspace, cV._ises, transpose=False, near=False)
        cctx.set_nullspace(cctx._nullspace_T, cV._ises, transpose=True, near=False)
        cctx.set_nullspace(cctx._near_nullspace, cV._ises, transpose=False, near=True)
        return cdm

    def coarsen_quadrature(self, metadata, fdeg, cdeg):
        """Coarsen the quadrature degree in a dictionary preserving the ratio of
           quadrature nodes to interpolation nodes (qdeg+1)//(fdeg+1)."""
        try:
            qdeg = metadata["quadrature_degree"]
            coarse_qdeg = max(2*cdeg+1, ((qdeg+1)*(cdeg+1)+fdeg)//(fdeg+1)-1)
            return dict(metadata, quadrature_degree=coarse_qdeg)
        except (KeyError, TypeError):
            return metadata

    def coarsen_bcs(self, fbcs, cV):
        """Coarsen a list of bcs"""
        cbcs = []
        for bc in fbcs:
            cache = self._coarsen_cache.setdefault(bc, {})
            key = (cV.ufl_element(), self.is_snes)
            try:
                coarse_bc = cache[key]
            except KeyError:
                cV_ = cV
                for index in bc._indices:
                    cV_ = cV_.sub(index)
                cbc_value = self.coarsen_bc_value(bc, cV_)
                if isinstance(bc, firedrake.DirichletBC):
                    coarse_bc = cache.setdefault(key, bc.reconstruct(V=cV_, g=cbc_value))
                else:
                    raise NotImplementedError("Unsupported BC type, please get in touch if you need this")
            cbcs.append(coarse_bc)
        return cbcs

    def coarsen_nullspace(self, fine_nullspace, cV, interpolate):
        """Coarsen a nullspace"""
        if fine_nullspace is None:
            return fine_nullspace
        cache = self._coarsen_cache.setdefault(fine_nullspace, {})
        key = cV.ufl_element()
        try:
            return cache[key]
        except KeyError:
            if isinstance(fine_nullspace, MixedVectorSpaceBasis):
                if interpolate.getType() == "python":
                    interpolate = interpolate.getPythonContext()
                submats = [interpolate.getNestSubMatrix(i, i) for i in range(len(cV))]
                coarse_bases = []
                for fs, submat, basis in zip(cV, submats, fine_nullspace._bases):
                    if isinstance(basis, VectorSpaceBasis):
                        coarse_bases.append(self.coarsen_nullspace(basis, fs, submat))
                    else:
                        coarse_bases.append(cV.sub(basis.index))
                coarse_nullspace = MixedVectorSpaceBasis(cV, coarse_bases)
            elif isinstance(fine_nullspace, VectorSpaceBasis):
                coarse_vecs = []
                for xf in fine_nullspace._petsc_vecs:
                    wc = firedrake.Function(cV)
                    with wc.dat.vec_wo as xc:
                        # the nullspace basis is in the dual of V
                        interpolate.multTranspose(xf, xc)
                    coarse_vecs.append(wc)
                coarse_nullspace = VectorSpaceBasis(coarse_vecs, constant=fine_nullspace._constant, comm=fine_nullspace.comm)
                coarse_nullspace.orthonormalize()
            else:
                return fine_nullspace
            return cache.setdefault(key, coarse_nullspace)

    def create_transfer(self, mat_type, cctx, fctx, cbcs, fbcs):
        """Create a transfer operator"""
        cache = self._transfer_cache.setdefault(fctx, {})
        key = (mat_type, cctx, cbcs, fbcs)
        try:
            return cache[key]
        except KeyError:
            if mat_type == "matfree":
                construct_mat = prolongation_matrix_matfree
            elif mat_type == "aij":
                construct_mat = prolongation_matrix_aij
            else:
                raise ValueError("Unknown matrix type")
            cV = cctx.J.arguments()[0].function_space()
            fV = fctx.J.arguments()[0].function_space()
            cbcs = tuple(cctx._problem.bcs) if cbcs else tuple()
            fbcs = tuple(fctx._problem.bcs) if fbcs else tuple()
            return cache.setdefault(key, construct_mat(cV, fV, cbcs, fbcs))

    def create_interpolation(self, dmc, dmf):
        prefix = dmc.getOptionsPrefix()
        mat_type = PETSc.Options(prefix).getString("mg_levels_transfer_mat_type", default="matfree")
        return self.create_transfer(mat_type, get_appctx(dmc), get_appctx(dmf), True, False), None

    def create_injection(self, dmc, dmf):
        prefix = dmc.getOptionsPrefix()
        mat_type = PETSc.Options(prefix).getString("mg_levels_transfer_mat_type", default="matfree")
        return self.create_transfer(mat_type, get_appctx(dmf), get_appctx(dmc), False, False)

    @staticmethod
    def max_degree(ele):
        """Return the maximum degree of a :class:`finat.ufl.finiteelement.FiniteElement`"""
        return max(as_tuple(ele.degree()))

    @staticmethod
    def reconstruct_degree(ele, degree):
        """Reconstruct an element, modifying its polynomial degree.

        By default, reconstructed EnrichedElements, TensorProductElements,
        and MixedElements will have the degree of the sub-elements shifted
        by the same amount so that the maximum degree is `degree`.
        This is useful to coarsen spaces like NCF(k) x DQ(k-1).

        Parameters
        ----------
        ele :
            A :class:`finat.ufl.finiteelement.FiniteElement` to reconstruct.
        degree :
            An integer degree.

        Returns
        -------
        ele :
            The reconstructed element.
        """
        if isinstance(ele, finat.ufl.VectorElement):
            return type(ele)(PMGBase.reconstruct_degree(ele._sub_element, degree), dim=ele.num_sub_elements)
        elif isinstance(ele, finat.ufl.TensorElement):
            return type(ele)(PMGBase.reconstruct_degree(ele._sub_element, degree), shape=ele._shape, symmetry=ele.symmetry())
        elif isinstance(ele, finat.ufl.EnrichedElement):
            shift = degree - PMGBase.max_degree(ele)
            return type(ele)(*(PMGBase.reconstruct_degree(e, PMGBase.max_degree(e) + shift) for e in ele._elements))
        elif isinstance(ele, finat.ufl.TensorProductElement):
            shift = degree - PMGBase.max_degree(ele)
            return type(ele)(*(PMGBase.reconstruct_degree(e, PMGBase.max_degree(e) + shift) for e in ele.sub_elements), cell=ele.cell)
        elif isinstance(ele, finat.ufl.MixedElement):
            shift = degree - PMGBase.max_degree(ele)
            return type(ele)(*(PMGBase.reconstruct_degree(e, PMGBase.max_degree(e) + shift) for e in ele.sub_elements))
        elif isinstance(ele, finat.ufl.WithMapping):
            return type(ele)(PMGBase.reconstruct_degree(ele.wrapee, degree), ele.mapping())
        elif isinstance(ele, (finat.ufl.HDivElement, finat.ufl.HCurlElement, finat.ufl.BrokenElement)):
            return type(ele)(PMGBase.reconstruct_degree(ele._element, degree))
        elif isinstance(ele, finat.ufl.RestrictedElement):
            return type(ele)(PMGBase.reconstruct_degree(ele._element, degree), restriction_domain=ele._restriction_domain)
        else:
            return ele.reconstruct(degree=degree)


[docs] class PMGPC(PCBase, PMGBase): _prefix = "pmg_"
[docs] def configure_pmg(self, pc, pdm): odm = pc.getDM() ppc = PETSc.PC().create(comm=pc.comm) ppc.setOptionsPrefix(pc.getOptionsPrefix() + "pmg_") ppc.setType("mg") ppc.setOperators(*pc.getOperators()) ppc.setDM(pdm) # PETSc unfortunately requires us to make an ugly hack. # We would like to use GMG for the coarse solve, at least # sometimes. But PETSc will use this p-DM's getRefineLevels() # instead of the getRefineLevels() of the MeshHierarchy to # decide how many levels it should use for PCMG applied to # the p-MG's coarse problem. So we need to set an option # for the user, if they haven't already; I don't know any # other way to get PETSc to know this at the right time. opts = PETSc.Options(pc.getOptionsPrefix() + "pmg_") opts["mg_coarse_pc_mg_levels"] = odm.getRefineLevel() + 1 return ppc
[docs] def apply(self, pc, x, y): return self.ppc.apply(x, y)
[docs] def applyTranspose(self, pc, x, y): return self.ppc.applyTranspose(x, y)
[docs] def coarsen_bc_value(self, bc, cV): return 0
[docs] class PMGSNES(SNESBase, PMGBase): _prefix = "pfas_"
[docs] def configure_pmg(self, snes, pdm): odm = snes.getDM() psnes = PETSc.SNES().create(comm=snes.comm) psnes.setOptionsPrefix(snes.getOptionsPrefix() + "pfas_") psnes.setType("fas") psnes.setDM(pdm) psnes.setTolerances(max_it=1) psnes.setConvergenceTest("skip") (f, residual) = snes.getFunction() assert residual is not None (fun, args, kargs) = residual psnes.setFunction(fun, f.duplicate(), args=args, kargs=kargs) pdm.setGlobalVector(f.duplicate()) psnes.setSolution(snes.getSolution()) # PETSc unfortunately requires us to make an ugly hack. # We would like to use GMG for the coarse solve, at least # sometimes. But PETSc will use this p-DM's getRefineLevels() # instead of the getRefineLevels() of the MeshHierarchy to # decide how many levels it should use for PCMG applied to # the p-MG's coarse problem. So we need to set an option # for the user, if they haven't already; I don't know any # other way to get PETSc to know this at the right time. opts = PETSc.Options(snes.getOptionsPrefix() + "pfas_") opts["fas_coarse_pc_mg_levels"] = odm.getRefineLevel() + 1 opts["fas_coarse_snes_fas_levels"] = odm.getRefineLevel() + 1 return psnes
[docs] def step(self, snes, x, f, y): ctx = get_appctx(snes.dm) push_appctx(self.ppc.dm, ctx) x.copy(y) self.ppc.solve(snes.vec_rhs or None, y) y.aypx(-1, x) snes.setConvergedReason(self.ppc.getConvergedReason()) pop_appctx(self.ppc.dm)
[docs] def coarsen_bc_value(self, bc, cV): if not isinstance(bc._original_arg, firedrake.Function): return bc._original_arg coarse = firedrake.Function(cV) coarse.interpolate(bc._original_arg) return coarse
def prolongation_transfer_kernel_action(Vf, expr): to_element = create_element(Vf.ufl_element()) kernel = compile_expression_dual_evaluation(expr, to_element, Vf.ufl_element(), log=PETSc.Log.isActive()) coefficients = extract_numbered_coefficients(expr, kernel.coefficient_numbers) if kernel.needs_external_coords: coefficients = [Vf.mesh().coordinates] + coefficients return op2.Kernel(kernel.ast, kernel.name, requires_zeroed_output_arguments=True, flop_count=kernel.flop_count, events=(kernel.event,)), coefficients def expand_element(ele): """Expand a FiniteElement as an EnrichedElement of TensorProductElements, discarding modifiers.""" if isinstance(ele, finat.FlattenedDimensions): return expand_element(ele.product) elif isinstance(ele, (finat.HDivElement, finat.HCurlElement)): return expand_element(ele.wrappee) elif isinstance(ele, finat.DiscontinuousElement): return expand_element(ele.element) elif isinstance(ele, finat.EnrichedElement): terms = list(map(expand_element, ele.elements)) return finat.EnrichedElement(terms) elif isinstance(ele, finat.TensorProductElement): factors = list(map(expand_element, ele.factors)) terms = [tuple()] for e in factors: new_terms = [] for f in e.elements if isinstance(e, finat.EnrichedElement) else [e]: f_factors = tuple(f.factors) if isinstance(f, finat.TensorProductElement) else (f,) new_terms.extend(t_factors + f_factors for t_factors in terms) terms = new_terms terms = list(map(finat.TensorProductElement, terms)) return finat.EnrichedElement(terms) else: return ele def hash_fiat_element(element): """FIAT elements are not hashable, this is not the best way to create a hash""" restriction = None e = element if isinstance(e, FIAT.DiscontinuousElement): # this hash does not care about inter-element continuity e = e._element if isinstance(e, FIAT.RestrictedElement): restriction = tuple(e._indices) e = e._element if len(restriction) == e.space_dimension(): restriction = None family = e.__class__.__name__ degree = e.order return (family, element.ref_el, degree, restriction) def generate_key_evaluate_dual(source, target, derivative=None): return hash_fiat_element(source) + hash_fiat_element(target) + (derivative,) def get_readonly_view(arr): result = arr.view() result.flags.writeable = False return result @cached({}, key=generate_key_evaluate_dual) def evaluate_dual(source, target, derivative=None): """Evaluate the action of a set of dual functionals of the target element on the (derivative of the) basis functions of the source element. Parameters ---------- source : A :class:`FIAT.CiarletElement` to interpolate. target : A :class:`FIAT.CiarletElement` defining the interpolation space. derivative : ``str`` or ``None`` An optional differential operator to apply on the source expression, (currently only "grad" is supported). Returns ------- A read-only :class:`numpy.ndarray` with the evaluation of the target dual basis on the (derivative of the) source primal basis. """ primal = source.get_nodal_basis() dual = target.get_dual_set() A = dual.to_riesz(primal) B = numpy.transpose(primal.get_coeffs()) if derivative == "grad": dmats = primal.get_dmats() assert len(dmats) == 1 alpha = (1,) for i in range(len(alpha)): for j in range(alpha[i]): B = numpy.dot(dmats[i], B) elif derivative in ("curl", "div"): raise NotImplementedError(f"Dual evaluation of {derivative} is not yet implemented.") elif derivative is not None: raise ValueError(f"Invalid derivaitve type {derivative}.") return get_readonly_view(numpy.dot(A, B)) @cached({}, key=generate_key_evaluate_dual) def compare_element(e1, e2): """Numerically compare two :class:`FIAT.elements`. Equality is satisfied if e2.dual_basis(e1.primal_basis) == identity.""" if e1 is e2: return True if e1.space_dimension() != e2.space_dimension(): return False B = evaluate_dual(e1, e2) return numpy.allclose(B, numpy.eye(B.shape[0]), rtol=1E-14, atol=1E-14) @cached({}, key=lambda V: V.ufl_element()) @PETSc.Log.EventDecorator("GetLineElements") def get_permutation_to_line_elements(V): """Find DOF permutation to factor out the EnrichedElement expansion into common TensorProductElements. This routine exposes structure to e.g vectorize prolongation of NCE or NCF accross vector components, by permuting all components into a common TensorProductElement. This is temporary while we wait for dual evaluation of :class:`finat.EnrichedElement`. Parameters ---------- V : A :class:`.FunctionSpace`. Returns ------- A 3-tuple of the DOF permutation, the unique terms in expansion as a list of tuples of :class:`FIAT.FiniteElements`, and the cyclic permutations of the axes to form the element given by their shifts in list of `int` tuples """ finat_element = V.finat_element expansion = expand_element(finat_element) if expansion.space_dimension() != finat_element.space_dimension(): raise ValueError("Failed to decompose %s into tensor products" % V.ufl_element()) line_elements = [] terms = expansion.elements if hasattr(expansion, "elements") else [expansion] for term in terms: factors = term.factors if hasattr(term, "factors") else (term,) fiat_factors = tuple(e.fiat_equivalent for e in reversed(factors)) if any(e.get_reference_element().get_spatial_dimension() != 1 for e in fiat_factors): raise ValueError("Failed to decompose %s into line elements" % V.ufl_element()) line_elements.append(fiat_factors) shapes = [tuple(e.space_dimension() for e in factors) for factors in line_elements] sizes = list(map(numpy.prod, shapes)) dof_ranges = numpy.cumsum([0] + sizes) dof_perm = [] unique_line_elements = [] shifts = [] visit = [False for e in line_elements] while False in visit: base = line_elements[visit.index(False)] tdim = len(base) pshape = tuple(e.space_dimension() for e in base) unique_line_elements.append(base) axes_shifts = tuple() for shift in range(tdim): if finat_element.formdegree != 2: shift = (tdim - shift) % tdim perm = base[shift:] + base[:shift] for i, term in enumerate(line_elements): if not visit[i]: is_perm = all(e1.space_dimension() == e2.space_dimension() for e1, e2 in zip(perm, term)) if is_perm: is_perm = all(compare_element(e1, e2) for e1, e2 in zip(perm, term)) if is_perm: axes_shifts += ((tdim - shift) % tdim, ) dofs = numpy.arange(*dof_ranges[i:i+2], dtype=PETSc.IntType).reshape(pshape) dofs = numpy.transpose(dofs, axes=numpy.roll(numpy.arange(tdim), -shift)) assert dofs.shape == shapes[i] dof_perm.append(dofs.flat) visit[i] = True shifts.append(axes_shifts) dof_perm = get_readonly_view(numpy.concatenate(dof_perm)) return dof_perm, unique_line_elements, shifts def get_permuted_map(V): """ Return a PermutedMap with the same tensor product shape for every component of H(div) or H(curl) tensor product elements """ indices, _, _ = get_permutation_to_line_elements(V) if numpy.all(indices[:-1] < indices[1:]): return V.cell_node_map() return op2.PermutedMap(V.cell_node_map(), indices) # Common kernel to compute y = kron(A3, kron(A2, A1)) * x # Vector and tensor field generalization from Deville, Fischer, and Mund section 8.3.1. kronmxv_code = """ #include <petscsys.h> #include <petscblaslapack.h> static inline void kronmxv_inplace(PetscBLASInt tflag, PetscBLASInt mx, PetscBLASInt my, PetscBLASInt mz, PetscBLASInt nx, PetscBLASInt ny, PetscBLASInt nz, PetscBLASInt nel, PetscScalar *A1, PetscScalar *A2, PetscScalar *A3, PetscScalar **x, PetscScalar **y){ /* Kronecker matrix-vector product y = op(A) * x, A = kron(A3, kron(A2, A1)) where: op(A) = transpose(A) if tflag>0 else A op(A1) is mx-by-nx, op(A2) is my-by-ny, op(A3) is mz-by-nz, x is (nx*ny*nz)-by-nel, y is (mx*my*mz)-by-nel. Important notes: This routine is in-place: the input data in x and y are destroyed in the process. Need to allocate nel*max(mx, nx)*max(my, ny)*max(mz, nz) memory for both x and y. */ PetscScalar *ptr[2] = {*x, *y}; PetscScalar zero = 0.0E0, one = 1.0E0; PetscBLASInt m, n, k, s, p, lda; PetscBLASInt ires = 0; char tran = 'T', notr = 'N'; char TA1 = tflag ? tran : notr; char TA2 = tflag ? notr : tran; if(A1){ m = mx; k = nx; n = ny*nz*nel; lda = tflag ? nx : mx; BLASgemm_(&TA1, &notr, &m, &n, &k, &one, A1, &lda, ptr[ires], &k, &zero, ptr[!ires], &m); ires = !ires; } if(A2){ p = 0; s = 0; m = mx; k = ny; n = my; lda = tflag ? ny : my; for(PetscBLASInt i=0; i<nz*nel; i++){ BLASgemm_(&notr, &TA2, &m, &n, &k, &one, ptr[ires]+p, &m, A2, &lda, &zero, ptr[!ires]+s, &m); p += m*k; s += m*n; } ires = !ires; } if(A3){ p = 0; s = 0; m = mx*my; k = nz; n = mz; lda = tflag ? nz : mz; for(PetscBLASInt i=0; i<nel; i++){ BLASgemm_(&notr, &TA2, &m, &n, &k, &one, ptr[ires]+p, &m, A3, &lda, &zero, ptr[!ires]+s, &m); p += m*k; s += m*n; } ires = !ires; } // Reassign pointers such that y always points to the result *x = ptr[!ires]; *y = ptr[ires]; return; } static inline void kronmxv(PetscBLASInt tflag, PetscBLASInt mx, PetscBLASInt my, PetscBLASInt mz, PetscBLASInt nx, PetscBLASInt ny, PetscBLASInt nz, PetscBLASInt nel, PetscScalar *A1, PetscScalar *A2, PetscScalar *A3, PetscScalar *x, PetscScalar *y, PetscScalar *xwork, PetscScalar *ywork){ /* Same as kronmxv_inplace, but the work buffers allow the input data in x to be kept untouched. */ PetscScalar *ptr[2] = {xwork, ywork}; if(ptr[0] != x) for(PetscBLASInt j=0; j<nx*ny*nz*nel; j++) ptr[0][j] = x[j]; kronmxv_inplace(tflag, mx, my, mz, nx, ny, nz, nel, A1, A2, A3, &ptr[0], &ptr[1]); if(ptr[1] != y) for(PetscBLASInt j=0; j<mx*my*mz*nel; j++) y[j] = ptr[1][j]; return; } static inline void permute_axis(PetscBLASInt axis, PetscBLASInt n0, PetscBLASInt n1, PetscBLASInt n2, PetscBLASInt n3, PetscScalar *x, PetscScalar *y){ /* Apply a cyclic permutation to a n0 x n1 x n2 x n3 array x, exponsing axis as the fast direction. Write the result on y. */ PetscBLASInt p = 0; PetscBLASInt s0, s1, s2, s3; if (axis == 0){ s0 = 1; s1 = s0*n0; s2 = s1*n1; s3 = s2*n2; }else if(axis == 1){ s1 = 1; s2 = s1*n1; s0 = s2*n2; s3 = s0*n0; }else if(axis == 2){ s2 = 1; s0 = s2*n2; s1 = s0*n0; s3 = s1*n1; } for(PetscBLASInt i3=0; i3<n3; i3++) for(PetscBLASInt i2=0; i2<n2; i2++) for(PetscBLASInt i1=0; i1<n1; i1++) for(PetscBLASInt i0=0; i0<n0; i0++) y[s0*i0 + s1*i1 + s2*i2 + s3*i3] = x[p++]; return; } static inline void ipermute_axis(PetscBLASInt axis, PetscBLASInt n0, PetscBLASInt n1, PetscBLASInt n2, PetscBLASInt n3, PetscScalar *x, PetscScalar *y){ /* Apply the transpose of permute_axis, reading from y and adding to x. */ PetscBLASInt p = 0; PetscBLASInt s0, s1, s2, s3; if (axis == 0){ s0 = 1; s1 = s0*n0; s2 = s1*n1; s3 = s2*n2; }else if(axis == 1){ s1 = 1; s2 = s1*n1; s0 = s2*n2; s3 = s0*n0; }else if(axis == 2){ s2 = 1; s0 = s2*n2; s1 = s0*n0; s3 = s1*n1; } for(PetscBLASInt i3=0; i3<n3; i3++) for(PetscBLASInt i2=0; i2<n2; i2++) for(PetscBLASInt i1=0; i1<n1; i1++) for(PetscBLASInt i0=0; i0<n0; i0++) x[p++] += y[s0*i0 + s1*i1 + s2*i2 + s3*i3]; return; } """ @PETSc.Log.EventDecorator("MakeKronCode") def make_kron_code(Vc, Vf, t_in, t_out, mat_name, scratch): """ Return interpolation and restriction kernels between enriched tensor product elements """ operator_decl = [] prolong_code = [] restrict_code = [] _, celems, cshifts = get_permutation_to_line_elements(Vc) _, felems, fshifts = get_permutation_to_line_elements(Vf) shifts = fshifts in_place = False if len(felems) == len(celems): in_place = all((len(fs)*Vf.value_size == len(cs)*Vc.value_size) for fs, cs in zip(fshifts, cshifts)) psize = Vf.value_size if not in_place: if len(celems) == 1: psize = Vc.value_size pelem = celems[0] perm_name = "perm_%s" % t_in celems = celems*len(felems) elif len(felems) == 1: shifts = cshifts psize = Vf.value_size pelem = felems[0] perm_name = "perm_%s" % t_out felems = felems*len(celems) else: raise ValueError("Cannot assign fine to coarse DOFs") if set(cshifts) == set(fshifts): csize = Vc.value_size * Vc.finat_element.space_dimension() prolong_code.append(f""" for({IntType_c} j=1; j<{len(fshifts)}; j++) for({IntType_c} i=0; i<{csize}; i++) {t_in}[j*{csize} + i] = {t_in}[i]; """) restrict_code.append(f""" for({IntType_c} j=1; j<{len(fshifts)}; j++) for({IntType_c} i=0; i<{csize}; i++) {t_in}[i] += {t_in}[j*{csize} + i]; """) elif pelem == celems[0]: for k in range(len(shifts)): if Vc.value_size*len(shifts[k]) < Vf.value_size: shifts[k] = shifts[k]*(Vf.value_size//Vc.value_size) pshape = [e.space_dimension() for e in pelem] pargs = ", ".join(map(str, pshape+[1]*(3-len(pshape)))) pstride = psize * numpy.prod(pshape) perm = sum(shifts, tuple()) perm_data = ", ".join(map(str, perm)) operator_decl.append(f""" PetscBLASInt {perm_name}[{len(perm)}] = {{ {perm_data} }}; """) prolong_code.append(f""" for({IntType_c} j=1; j<{len(perm)}; j++) permute_axis({perm_name}[j], {pargs}, {psize}, {t_in}, {t_in}+j*{pstride}); """) restrict_code.append(f""" for({IntType_c} j=1; j<{len(perm)}; j++) ipermute_axis({perm_name}[j], {pargs}, {psize}, {t_in}, {t_in}+j*{pstride}); """) fskip = 0 cskip = 0 Jlen = 0 Jmats = [] fshapes = [] cshapes = [] has_code = False identity_filter = lambda A: numpy.array([]) if A.shape[0] == A.shape[1] and numpy.allclose(A, numpy.eye(A.shape[0])) else A for celem, felem, shift in zip(celems, felems, shifts): if len(felem) != len(celem): raise ValueError("Fine and coarse elements do not have the same number of factors") if len(felem) > 3: raise ValueError("More than three factors are not supported") # Declare array shapes to be used as literals inside the kernels nscal = psize*len(shift) fshape = [e.space_dimension() for e in felem] cshape = [e.space_dimension() for e in celem] fshapes.append((nscal,) + tuple(fshape)) cshapes.append((nscal,) + tuple(cshape)) J = [identity_filter(evaluate_dual(ce, fe)).T for ce, fe in zip(celem, felem)] if any(Jk.size and numpy.isclose(Jk, 0.0E0).all() for Jk in J): prolong_code.append(f""" for({IntType_c} i=0; i<{nscal*numpy.prod(fshape)}; i++) {t_out}[i+{fskip}] = 0.0E0; """) restrict_code.append(f""" for({IntType_c} i=0; i<{nscal*numpy.prod(cshape)}; i++) {t_in}[i+{cskip}] = 0.0E0; """) else: Jsize = numpy.cumsum([Jlen] + [Jk.size for Jk in J]) Jptrs = ["%s+%d" % (mat_name, Jsize[k]) if J[k].size else "NULL" for k in range(len(J))] Jmats.extend(J) Jlen = Jsize[-1] # The Kronecker product routines assume 3D shapes, so in 1D and 2D we pass NULL instead of J Jargs = ", ".join(Jptrs+["NULL"]*(3-len(Jptrs))) fargs = ", ".join(map(str, fshape+[1]*(3-len(fshape)))) cargs = ", ".join(map(str, cshape+[1]*(3-len(cshape)))) if in_place: prolong_code.append(f""" kronmxv_inplace(0, {fargs}, {cargs}, {nscal}, {Jargs}, &{t_in}, &{t_out}); """) restrict_code.append(f""" kronmxv_inplace(1, {cargs}, {fargs}, {nscal}, {Jargs}, &{t_out}, &{t_in}); """) elif shifts == fshifts: if has_code and psize > 1: raise ValueError("Single tensor product to many tensor products not implemented for vectors") # Single tensor product to many prolong_code.append(f""" kronmxv(0, {fargs}, {cargs}, {nscal}, {Jargs}, {t_in}+{cskip}, {t_out}+{fskip}, {scratch}, {t_out}+{fskip}); """) restrict_code.append(f""" kronmxv(1, {cargs}, {fargs}, {nscal}, {Jargs}, {t_out}+{fskip}, {t_in}+{cskip}, {t_out}+{fskip}, {scratch}); """) else: # Many tensor products to single tensor product if has_code: raise ValueError("Many tensor products to single tensor product not implemented") fskip = 0 prolong_code.append(f""" kronmxv(0, {fargs}, {cargs}, {nscal}, {Jargs}, {t_in}+{cskip}, {t_out}+{fskip}, {t_in}+{cskip}, {t_out}+{fskip}); """) restrict_code.append(f""" kronmxv(1, {cargs}, {fargs}, {nscal}, {Jargs}, {t_out}+{fskip}, {t_in}+{cskip}, {t_out}+{fskip}, {t_in}+{cskip}); """) has_code = True fskip += nscal*numpy.prod(fshape) cskip += nscal*numpy.prod(cshape) # Pass the 1D interpolators as a hexadecimal string Jdata = ", ".join(map(float.hex, chain.from_iterable(Jk.flat for Jk in Jmats))) operator_decl.append(f""" PetscScalar {mat_name}[{Jlen}] = {{ {Jdata} }}; """) operator_decl = "".join(operator_decl) prolong_code = "".join(prolong_code) restrict_code = "".join(reversed(restrict_code)) shapes = [tuple(map(max, zip(*fshapes))), tuple(map(max, zip(*cshapes)))] if fskip > numpy.prod(shapes[0]): shapes[0] = (fskip, 1, 1, 1) if cskip > numpy.prod(shapes[1]): shapes[1] = (cskip, 1, 1, 1) return operator_decl, prolong_code, restrict_code, shapes def get_piola_tensor(mapping, domain, inverse=False): mapping = mapping.lower() if mapping == "identity": return None elif mapping == "contravariant piola": if inverse: return ufl.JacobianInverse(domain)*ufl.JacobianDeterminant(domain) else: return ufl.Jacobian(domain)/ufl.JacobianDeterminant(domain) elif mapping == "covariant piola": if inverse: return ufl.Jacobian(domain).T else: return ufl.JacobianInverse(domain).T else: raise ValueError("Mapping %s is not supported" % mapping) def cache_generate_code(kernel, comm): _cachedir = os.environ.get('PYOP2_CACHE_DIR', os.path.join(tempfile.gettempdir(), 'pyop2-cache-uid%d' % os.getuid())) key = kernel.cache_key[0] shard, disk_key = key[:2], key[2:] filepath = os.path.join(_cachedir, shard, disk_key) if os.path.exists(filepath): with open(filepath, 'r') as f: code = f.read() else: code = loopy.generate_code_v2(kernel.code).device_code() if comm.rank == 0: os.makedirs(os.path.join(_cachedir, shard), exist_ok=True) with open(filepath, 'w') as f: f.write(code) comm.barrier() return code def make_mapping_code(Q, cmapping, fmapping, t_in, t_out): if fmapping == cmapping: return None A = get_piola_tensor(cmapping, Q.mesh(), inverse=False) B = get_piola_tensor(fmapping, Q.mesh(), inverse=True) tensor = A if B: tensor = ufl.dot(B, tensor) if tensor else B if tensor is None: tensor = ufl.Identity(Q.ufl_element().value_shape[0]) u = ufl.Coefficient(Q) expr = ufl.dot(tensor, u) prolong_map_kernel, coefficients = prolongation_transfer_kernel_action(Q, expr) prolong_map_code = cache_generate_code(prolong_map_kernel, Q._comm) prolong_map_code = prolong_map_code.replace("void expression_kernel", "static void prolongation_mapping") coefficients.remove(u) expr = ufl.dot(u, tensor) restrict_map_kernel, coefficients = prolongation_transfer_kernel_action(Q, expr) restrict_map_code = cache_generate_code(restrict_map_kernel, Q._comm) restrict_map_code = restrict_map_code.replace("void expression_kernel", "static void restriction_mapping") restrict_map_code = restrict_map_code.replace("#include <stdint.h>", "") restrict_map_code = restrict_map_code.replace("#include <complex.h>", "") coefficients.remove(u) coef_args = "".join([", c%d" % i for i in range(len(coefficients))]) coef_decl = "".join([", PetscScalar const *restrict c%d" % i for i in range(len(coefficients))]) qlen = Q.value_size * Q.finat_element.space_dimension() prolong_code = f""" for({IntType_c} i=0; i<{qlen}; i++) {t_out}[i] = 0.0E0; prolongation_mapping({t_out}{coef_args}, {t_in}); """ restrict_code = f""" for({IntType_c} i=0; i<{qlen}; i++) {t_in}[i] = 0.0E0; restriction_mapping({t_in}{coef_args}, {t_out}); """ mapping_code = prolong_map_code + restrict_map_code return coef_decl, prolong_code, restrict_code, mapping_code, coefficients def make_permutation_code(V, vshape, pshape, t_in, t_out, array_name): _, _, shifts = get_permutation_to_line_elements(V) shift = shifts[0] if shift != (0,): ndof = numpy.prod(vshape) permutation = numpy.reshape(numpy.arange(ndof), pshape) axes = numpy.arange(len(shift)) for k in range(permutation.shape[0]): permutation[k] = numpy.reshape(numpy.transpose(permutation[k], axes=numpy.roll(axes, -shift[k])), pshape[1:]) nflip = 0 mapping = V.ufl_element().mapping().lower() if mapping == "contravariant piola": # flip the sign of the first component nflip = ndof//len(shift) elif mapping == "covariant piola": # flip the order of reference components permutation = numpy.flip(permutation, axis=0) permutation = numpy.transpose(numpy.reshape(permutation, vshape)) pdata = ", ".join(map(str, permutation.flat)) decl = f""" PetscInt {array_name}[{ndof}] = {{ {pdata} }}; """ prolong = f""" for({IntType_c} i=0; i<{ndof}; i++) {t_out}[{array_name}[i]] = {t_in}[i]; for({IntType_c} i=0; i<{nflip}; i++) {t_out}[i] = -{t_out}[i]; """ restrict = f""" for({IntType_c} i=0; i<{nflip}; i++) {t_out}[i] = -{t_out}[i]; for({IntType_c} i=0; i<{ndof}; i++) {t_in}[i] = {t_out}[{array_name}[i]]; """ else: decl = "" prolong = f""" for({IntType_c} j=0; j<{vshape[1]}; j++) for({IntType_c} i=0; i<{vshape[0]}; i++) {t_out}[j + {vshape[1]}*i] = {t_in}[i + {vshape[0]}*j]; """ restrict = f""" for({IntType_c} j=0; j<{vshape[1]}; j++) for({IntType_c} i=0; i<{vshape[0]}; i++) {t_in}[i + {vshape[0]}*j] = {t_out}[j + {vshape[1]}*i]; """ return decl, prolong, restrict class StandaloneInterpolationMatrix(object): """ Interpolation matrix for a single standalone space. """ _cache_work = {} def __init__(self, Vc, Vf, Vc_bcs, Vf_bcs): self.uc = self.work_function(Vc) self.uf = self.work_function(Vf) self.Vc = self.uc.function_space() self.Vf = self.uf.function_space() self.Vc_bcs = Vc_bcs self.Vf_bcs = Vf_bcs def work_function(self, V): if isinstance(V, firedrake.Function): return V key = (V.ufl_element(), V.mesh()) try: return self._cache_work[key] except KeyError: return self._cache_work.setdefault(key, firedrake.Function(V)) @cached_property def _weight(self): weight = firedrake.Function(self.Vf) size = self.Vf.finat_element.space_dimension() * self.Vf.value_size kernel_code = f""" void weight(PetscScalar *restrict w){{ for(PetscInt i=0; i<{size}; i++) w[i] += 1.0; return; }} """ kernel = op2.Kernel(kernel_code, "weight", requires_zeroed_output_arguments=True) op2.par_loop(kernel, weight.cell_set, weight.dat(op2.INC, weight.cell_node_map())) with weight.dat.vec as w: w.reciprocal() return weight @cached_property def _kernels(self): try: # We generate custom prolongation and restriction kernels mainly because: # 1. Code generation for the transpose of prolongation is not readily available # 2. Dual evaluation of EnrichedElement is not yet implemented in FInAT uf_map = get_permuted_map(self.Vf) uc_map = get_permuted_map(self.Vc) prolong_kernel, restrict_kernel, coefficients = self.make_blas_kernels(self.Vf, self.Vc) prolong_args = [prolong_kernel, self.uf.cell_set, self.uf.dat(op2.INC, uf_map), self.uc.dat(op2.READ, uc_map), self._weight.dat(op2.READ, uf_map)] except ValueError: # The elements do not have the expected tensor product structure # Fall back to aij kernels uf_map = self.Vf.cell_node_map() uc_map = self.Vc.cell_node_map() prolong_kernel, restrict_kernel, coefficients = self.make_kernels(self.Vf, self.Vc) prolong_args = [prolong_kernel, self.uf.cell_set, self.uf.dat(op2.WRITE, uf_map), self.uc.dat(op2.READ, uc_map)] restrict_args = [restrict_kernel, self.uf.cell_set, self.uc.dat(op2.INC, uc_map), self.uf.dat(op2.READ, uf_map), self._weight.dat(op2.READ, uf_map)] coefficient_args = [c.dat(op2.READ, c.cell_node_map()) for c in coefficients] prolong = op2.ParLoop(*prolong_args, *coefficient_args) restrict = op2.ParLoop(*restrict_args, *coefficient_args) return prolong, restrict def _prolong(self): with self.uf.dat.vec_wo as uf: uf.set(0.0E0) self._kernels[0]() def _restrict(self): with self.uc.dat.vec_wo as uc: uc.set(0.0E0) self._kernels[1]() def view(self, mat, viewer=None): if viewer is None: return typ = viewer.getType() if typ != PETSc.Viewer.Type.ASCII: return viewer.printfASCII("Firedrake matrix-free prolongator %s\n" % type(self).__name__) def getInfo(self, mat, info=None): memory = self.uf.dat.nbytes + self.uc.dat.nbytes if self._weight is not None: memory += self._weight.dat.nbytes if info is None: info = PETSc.Mat.InfoType.GLOBAL_SUM if info == PETSc.Mat.InfoType.LOCAL: return {"memory": memory} elif info == PETSc.Mat.InfoType.GLOBAL_SUM: gmem = mat.comm.tompi4py().allreduce(memory, op=op2.MPI.SUM) return {"memory": gmem} elif info == PETSc.Mat.InfoType.GLOBAL_MAX: gmem = mat.comm.tompi4py().allreduce(memory, op=op2.MPI.MAX) return {"memory": gmem} else: raise ValueError("Unknown info type %s" % info) @staticmethod def make_blas_kernels(Vf, Vc): """ Interpolation and restriction kernels between CG / DG tensor product spaces on quads and hexes. Works by tabulating the coarse 1D basis functions as the (fdegree+1)-by-(cdegree+1) matrix Jhat, and using the fact that the 2D / 3D tabulation is the tensor product J = kron(Jhat, kron(Jhat, Jhat)) """ felem = Vf.ufl_element() celem = Vc.ufl_element() fmapping = felem.mapping().lower() cmapping = celem.mapping().lower() in_place_mapping = False coefficients = [] mapping_code = "" coef_decl = "" if fmapping == cmapping: # interpolate on each direction via Kroncker product operator_decl, prolong_code, restrict_code, shapes = make_kron_code(Vc, Vf, "t0", "t1", "J0", "t2") else: decl = [""]*4 prolong = [""]*5 restrict = [""]*5 # get embedding element for Vf with identity mapping and collocated vector component DOFs try: qelem = felem if qelem.mapping() != "identity": qelem = qelem.reconstruct(mapping="identity") Qf = Vf if qelem == felem else firedrake.FunctionSpace(Vf.mesh(), qelem) mapping_output = make_mapping_code(Qf, cmapping, fmapping, "t0", "t1") in_place_mapping = True except Exception: qelem = finat.ufl.FiniteElement("DQ", cell=felem.cell, degree=PMGBase.max_degree(felem)) if felem.value_shape: qelem = finat.ufl.TensorElement(qelem, shape=felem.value_shape, symmetry=felem.symmetry()) Qf = firedrake.FunctionSpace(Vf.mesh(), qelem) mapping_output = make_mapping_code(Qf, cmapping, fmapping, "t0", "t1") qshape = (Qf.value_size, Qf.finat_element.space_dimension()) # interpolate to embedding fine space decl[0], prolong[0], restrict[0], shapes = make_kron_code(Vc, Qf, "t0", "t1", "J0", "t2") if mapping_output is not None: # permute to FInAT ordering, and apply the mapping decl[1], restrict[1], prolong[1] = make_permutation_code(Vc, qshape, shapes[0], "t0", "t1", "perm0") coef_decl, prolong[2], restrict[2], mapping_code, coefficients = mapping_output if not in_place_mapping: # permute to Kronecker-friendly ordering and interpolate to fine space decl[2], prolong[3], restrict[3] = make_permutation_code(Vf, qshape, shapes[0], "t1", "t0", "perm1") decl[3], prolong[4], restrict[4], _shapes = make_kron_code(Qf, Vf, "t0", "t1", "J1", "t2") shapes.extend(_shapes) operator_decl = "".join(decl) prolong_code = "".join(prolong) restrict_code = "".join(reversed(restrict)) # FInAT elements order the component DOFs related to the same node contiguously. # We transpose before and after the multiplication times J to have each component # stored contiguously as a scalar field, thus reducing the number of dgemm calls. # We could benefit from loop tiling for the transpose, but that makes the code # more complicated. fshape = (Vf.value_size, Vf.finat_element.space_dimension()) cshape = (Vc.value_size, Vc.finat_element.space_dimension()) lwork = numpy.prod([max(*dims) for dims in zip(*shapes)]) lwork = max(lwork, max(numpy.prod(fshape), numpy.prod(cshape))) if cshape[0] == 1: coarse_read = f"""for({IntType_c} i=0; i<{numpy.prod(cshape)}; i++) t0[i] = x[i];""" coarse_write = f"""for({IntType_c} i=0; i<{numpy.prod(cshape)}; i++) x[i] += t0[i];""" else: coarse_read = f""" for({IntType_c} j=0; j<{cshape[1]}; j++) for({IntType_c} i=0; i<{cshape[0]}; i++) t0[j + {cshape[1]}*i] = x[i + {cshape[0]}*j]; """ coarse_write = f""" for({IntType_c} j=0; j<{cshape[1]}; j++) for({IntType_c} i=0; i<{cshape[0]}; i++) x[i + {cshape[0]}*j] += t0[j + {cshape[1]}*i]; """ if (fshape[0] == 1) or in_place_mapping: fine_read = f"""for({IntType_c} i=0; i<{numpy.prod(fshape)}; i++) t1[i] = y[i] * w[i];""" fine_write = f"""for({IntType_c} i=0; i<{numpy.prod(fshape)}; i++) y[i] += t1[i] * w[i];""" else: fine_read = f""" for({IntType_c} j=0; j<{fshape[1]}; j++) for({IntType_c} i=0; i<{fshape[0]}; i++) t1[j + {fshape[1]}*i] = y[i + {fshape[0]}*j] * w[i + {fshape[0]}*j]; """ fine_write = f""" for({IntType_c} j=0; j<{fshape[1]}; j++) for({IntType_c} i=0; i<{fshape[0]}; i++) y[i + {fshape[0]}*j] += t1[j + {fshape[1]}*i] * w[i + {fshape[0]}*j]; """ kernel_code = f""" {mapping_code} {kronmxv_code} void prolongation(PetscScalar *restrict y, const PetscScalar *restrict x, const PetscScalar *restrict w{coef_decl}){{ PetscScalar work[3][{lwork}] = {{0.0E0}}; PetscScalar *t0 = work[0]; PetscScalar *t1 = work[1]; PetscScalar *t2 = work[2]; {operator_decl} {coarse_read} {prolong_code} {fine_write} return; }} void restriction(PetscScalar *restrict x, const PetscScalar *restrict y, const PetscScalar *restrict w{coef_decl}){{ PetscScalar work[3][{lwork}] = {{0.0E0}}; PetscScalar *t0 = work[0]; PetscScalar *t1 = work[1]; PetscScalar *t2 = work[2]; {operator_decl} {fine_read} {restrict_code} {coarse_write} return; }} """ from firedrake.slate.slac.compiler import BLASLAPACK_LIB, BLASLAPACK_INCLUDE prolong_kernel = op2.Kernel(kernel_code, "prolongation", include_dirs=BLASLAPACK_INCLUDE.split(), ldargs=BLASLAPACK_LIB.split(), requires_zeroed_output_arguments=True) restrict_kernel = op2.Kernel(kernel_code, "restriction", include_dirs=BLASLAPACK_INCLUDE.split(), ldargs=BLASLAPACK_LIB.split(), requires_zeroed_output_arguments=True) return prolong_kernel, restrict_kernel, coefficients def make_kernels(self, Vf, Vc): """ Interpolation and restriction kernels between arbitrary elements. This is temporary while we wait for dual evaluation in FInAT. """ prolong_kernel, _ = prolongation_transfer_kernel_action(Vf, self.uc) matrix_kernel, coefficients = prolongation_transfer_kernel_action(Vf, firedrake.TestFunction(Vc)) # The way we transpose the prolongation kernel is suboptimal. # A local matrix is generated each time the kernel is executed. element_kernel = loopy.generate_code_v2(matrix_kernel.code).device_code() element_kernel = element_kernel.replace("void expression_kernel", "static void expression_kernel") coef_args = "".join([", c%d" % i for i in range(len(coefficients))]) coef_decl = "".join([", const %s *restrict c%d" % (ScalarType_c, i) for i in range(len(coefficients))]) dimc = Vc.finat_element.space_dimension() * Vc.value_size dimf = Vf.finat_element.space_dimension() * Vf.value_size restrict_code = f""" {element_kernel} void restriction({ScalarType_c} *restrict Rc, const {ScalarType_c} *restrict Rf, const {ScalarType_c} *restrict w{coef_decl}) {{ {ScalarType_c} Afc[{dimf}*{dimc}] = {{0}}; expression_kernel(Afc{coef_args}); for ({IntType_c} i = 0; i < {dimf}; i++) for ({IntType_c} j = 0; j < {dimc}; j++) Rc[j] += Afc[i*{dimc} + j] * Rf[i] * w[i]; }} """ restrict_kernel = op2.Kernel( restrict_code, "restriction", requires_zeroed_output_arguments=True, events=matrix_kernel.events, ) return prolong_kernel, restrict_kernel, coefficients def multTranspose(self, mat, rf, rc): """ Implement restriction: restrict residual on fine grid rf to coarse grid rc. """ with self.uf.dat.vec_wo as uf: rf.copy(uf) for bc in self.Vf_bcs: bc.zero(self.uf) self._restrict() for bc in self.Vc_bcs: bc.zero(self.uc) with self.uc.dat.vec_ro as uc: uc.copy(rc) def mult(self, mat, xc, xf, inc=False): """ Implement prolongation: prolong correction on coarse grid xc to fine grid xf. """ with self.uc.dat.vec_wo as uc: xc.copy(uc) for bc in self.Vc_bcs: bc.zero(self.uc) self._prolong() for bc in self.Vf_bcs: bc.zero(self.uf) if inc: with self.uf.dat.vec_ro as uf: xf.axpy(1.0, uf) else: with self.uf.dat.vec_ro as uf: uf.copy(xf) def multAdd(self, mat, x, y, w): if y.handle == w.handle: self.mult(mat, x, w, inc=True) else: self.mult(mat, x, w) w.axpy(1.0, y) class MixedInterpolationMatrix(StandaloneInterpolationMatrix): """ Interpolation matrix for a mixed finite element space. """ @cached_property def _weight(self): return None @cached_property def _standalones(self): standalones = [] for i, (uc_sub, uf_sub) in enumerate(zip(self.uc.subfunctions, self.uf.subfunctions)): Vc_sub_bcs = tuple(bc for bc in self.Vc_bcs if bc.function_space().index == i) Vf_sub_bcs = tuple(bc for bc in self.Vf_bcs if bc.function_space().index == i) standalone = StandaloneInterpolationMatrix(uc_sub, uf_sub, Vc_sub_bcs, Vf_sub_bcs) standalones.append(standalone) return standalones @cached_property def _kernels(self): prolong = lambda: [s._prolong() for s in self._standalones] restrict = lambda: [s._restrict() for s in self._standalones] return prolong, restrict def getNestSubMatrix(self, i, j): if i == j: s = self._standalones[i] sizes = (s.uf.dof_dset.layout_vec.getSizes(), s.uc.dof_dset.layout_vec.getSizes()) M_shll = PETSc.Mat().createPython(sizes, s, comm=s.uf._comm) M_shll.setUp() return M_shll else: return None def prolongation_matrix_aij(P1, Pk, P1_bcs=[], Pk_bcs=[]): if isinstance(P1, firedrake.Function): P1 = P1.function_space() if isinstance(Pk, firedrake.Function): Pk = Pk.function_space() sp = op2.Sparsity((Pk.dof_dset, P1.dof_dset), {(i, j): [(rmap, cmap, None)] for i, rmap in enumerate(Pk.cell_node_map()) for j, cmap in enumerate(P1.cell_node_map()) if i == j}) mat = op2.Mat(sp, PETSc.ScalarType) mesh = Pk.mesh() fele = Pk.ufl_element() if type(fele) is finat.ufl.MixedElement: for i in range(fele.num_sub_elements): Pk_bcs_i = [bc for bc in Pk_bcs if bc.function_space().index == i] P1_bcs_i = [bc for bc in P1_bcs if bc.function_space().index == i] rlgmap, clgmap = mat[i, i].local_to_global_maps rlgmap = Pk.sub(i).local_to_global_map(Pk_bcs_i, lgmap=rlgmap) clgmap = P1.sub(i).local_to_global_map(P1_bcs_i, lgmap=clgmap) unroll = any(bc.function_space().component is not None for bc in chain(Pk_bcs_i, P1_bcs_i) if bc is not None) matarg = mat[i, i](op2.WRITE, (Pk.sub(i).cell_node_map(), P1.sub(i).cell_node_map()), lgmaps=((rlgmap, clgmap), ), unroll_map=unroll) expr = firedrake.TestFunction(P1.sub(i)) kernel, coefficients = prolongation_transfer_kernel_action(Pk.sub(i), expr) parloop_args = [kernel, mesh.cell_set, matarg] for coefficient in coefficients: m_ = coefficient.cell_node_map() parloop_args.append(coefficient.dat(op2.READ, m_)) op2.par_loop(*parloop_args) else: rlgmap, clgmap = mat.local_to_global_maps rlgmap = Pk.local_to_global_map(Pk_bcs, lgmap=rlgmap) clgmap = P1.local_to_global_map(P1_bcs, lgmap=clgmap) unroll = any(bc.function_space().component is not None for bc in chain(Pk_bcs, P1_bcs) if bc is not None) matarg = mat(op2.WRITE, (Pk.cell_node_map(), P1.cell_node_map()), lgmaps=((rlgmap, clgmap), ), unroll_map=unroll) expr = firedrake.TestFunction(P1) kernel, coefficients = prolongation_transfer_kernel_action(Pk, expr) parloop_args = [kernel, mesh.cell_set, matarg] for coefficient in coefficients: m_ = coefficient.cell_node_map() parloop_args.append(coefficient.dat(op2.READ, m_)) op2.par_loop(*parloop_args) mat.assemble() return mat.handle def prolongation_matrix_matfree(Vc, Vf, Vc_bcs=[], Vf_bcs=[]): fele = Vf.ufl_element() if type(fele) is finat.ufl.MixedElement: ctx = MixedInterpolationMatrix(Vc, Vf, Vc_bcs, Vf_bcs) else: ctx = StandaloneInterpolationMatrix(Vc, Vf, Vc_bcs, Vf_bcs) sizes = (Vf.dof_dset.layout_vec.getSizes(), Vc.dof_dset.layout_vec.getSizes()) M_shll = PETSc.Mat().createPython(sizes, ctx, comm=Vf._comm) M_shll.setUp() return M_shll