Exemplo n.º 1
0
def inner(expr0, expr1, output_array=None, level=0):
    r"""
    Return weighted discrete inner product of linear or bilinear form

    .. math::

        (f, g)_w^N = \sum_{i\in\mathcal{I}}f(x_i) \overline{g}(x_i) w_i \approx \int_{\Omega} g\, \overline{f}\, w\, dx

    where :math:`\mathcal{I}=0, 1, \ldots, N, N \in \mathbb{Z}^+`, :math:`f`
    is an expression linear in a :class:`.TestFunction`, and :math:`g` is an
    expression that is linear in :class:`.TrialFunction` or :class:`.Function`,
    or it is simply an :class:`.Array` (a solution interpolated on the
    quadrature mesh in physical space). :math:`w` is a weight associated with
    chosen basis, and :math:`w_i` are quadrature weights.

    If the expressions are created in a multidimensional :class:`.TensorProductSpace`,
    then the sum above is over all dimensions. In 2D it becomes:

    .. math::

        (f, g)_w^N = \sum_{i\in\mathcal{I}}\sum_{j\in\mathcal{J}} f(x_i, y_j) \overline{g}(x_i, y_j) w_j w_i

    where :math:`\mathcal{J}=0, 1, \ldots, M, M \in \mathbb{Z}^+`.

    Parameters
    ----------
    expr0, expr1 : :class:`.Expr`, :class:`.BasisFunction`, :class:`.Array`
        or number.
        Either one can be an expression involving a
        BasisFunction (:class:`.TestFunction`, :class:`.TrialFunction` or
        :class:`.Function`) an Array or a number. With expressions (Expr) on a
        BasisFunction we typically mean terms like div(u) or grad(u), where
        u is any one of the different types of BasisFunction.
        One of ``expr0`` or ``expr1`` need to be an expression on a
        TestFunction. If the second then involves a TrialFunction, a matrix is
        returned. If one of ``expr0``/``expr1`` involves a TestFunction and the
        other one is an expression on a Function, or a plain Array, then a
        linear form is assembled and a Function is returned.
        If either expr0 or expr1 is a number (typically 1) or a tuple of
        numbers for a vector, then the inner product represents a non-weighted
        integral over the domain. If a single number, then the other expression
        must be a scalar Array or a Function. If a tuple of numbers, then the
        Array/Function must be a vector.

    output_array:  Function
        Optional return array for linear form.

    level: int
        The level of postprocessing for assembled matrices. Applies only
        to bilinear forms

        - 0 Full postprocessing - diagonal matrices to scale arrays
          and add equal matrices
        - 1 Diagonal matrices to scale arrays, but don't add equal
          matrices
        - 2 No postprocessing, return all assembled matrices

    Returns
    -------
    Depending on dimensionality and the arguments to the forms

        :class:`.Function`
        for linear forms.

        :class:`.SparseMatrix`
        for bilinear 1D forms.

        :class:`.TPMatrix` or list of :class:`.TPMatrix`
        for bilinear multidimensional forms.

        Number, for non-weighted integral where either one of the arguments
        is a number.

    See Also
    --------
    :func:`.project`

    Example
    -------
    Compute mass matrix of Shen's Chebyshev Dirichlet basis:

    >>> from shenfun import Basis
    >>> from shenfun import TestFunction, TrialFunction
    >>> SD = Basis(6, 'Chebyshev', bc=(0, 0))
    >>> u = TrialFunction(SD)
    >>> v = TestFunction(SD)
    >>> B = inner(v, u)
    >>> d = {-2: np.array([-np.pi/2]),
    ...       0: np.array([ 1.5*np.pi, np.pi, np.pi, np.pi]),
    ...       2: np.array([-np.pi/2])}
    >>> [np.all(abs(B[k]-v) < 1e-7) for k, v in d.items()]
    [True, True, True]

    """
    # Wrap a pure numpy array in Array
    if isinstance(expr0,
                  np.ndarray) and not isinstance(expr0, (Array, Function)):
        assert isinstance(expr1, (Expr, BasisFunction))
        if not expr0.flags['C_CONTIGUOUS']:
            expr0 = expr0.copy()
        expr0 = Array(expr1.function_space(), buffer=expr0)
    if isinstance(expr1,
                  np.ndarray) and not isinstance(expr1, (Array, Function)):
        assert isinstance(expr0, (Expr, BasisFunction))
        if not expr1.flags['C_CONTIGUOUS']:
            expr1 = expr1.copy()
        expr1 = Array(expr0.function_space(), buffer=expr1)

    if isinstance(expr0, Number):
        assert isinstance(expr1, (Array, Function))
        space = expr1.function_space()
        if isinstance(space, (TensorProductSpace, MixedTensorProductSpace)):
            df = np.prod(
                np.array([base.domain_factor() for base in space.bases]))
        elif isinstance(space, SpectralBase):
            df = space.domain_factor()
        if isinstance(expr1, Function):
            return (expr0 / df) * dx(expr1.backward())
        return (expr0 / df) * dx(expr1)

    if isinstance(expr1, Number):
        assert isinstance(expr0, (Array, Function))
        space = expr0.function_space()
        if isinstance(space, (TensorProductSpace, MixedTensorProductSpace)):
            df = np.prod(
                np.array([base.domain_factor() for base in space.bases]))
        elif isinstance(space, SpectralBase):
            df = space.domain_factor()
        if isinstance(expr0, Function):
            return (expr1 / df) * dx(expr0.backward())
        return (expr1 / df) * dx(expr0)

    if isinstance(expr0, tuple):
        assert isinstance(expr1, (Array, Function))
        space = expr1.function_space()
        assert isinstance(space, MixedTensorProductSpace)
        assert len(expr0) == len(space)
        result = 0.0
        for e0i, e1i in zip(expr0, expr1):
            result += inner(e0i, e1i)
        return result

    if isinstance(expr1, tuple):
        assert isinstance(expr0, (Array, Function))
        space = expr0.function_space()
        assert isinstance(space, MixedTensorProductSpace)
        assert len(expr1) == len(space)
        result = 0.0
        for e0i, e1i in zip(expr0, expr1):
            result += inner(e0i, e1i)
        return result

    assert np.all([hasattr(e, 'argument') for e in (expr0, expr1)])
    t0 = expr0.argument
    t1 = expr1.argument
    if t0 == 0:
        assert t1 in (1, 2)
        test = expr0
        trial = expr1
    elif t0 in (1, 2):
        assert t1 == 0
        test = expr1
        trial = expr0
    else:
        raise RuntimeError

    if test.rank > 0 and test.expr_rank(
    ) > 0:  # For vector expressions of rank > 0 use recursive algorithm

        if output_array is None and trial.argument == 2:
            output_array = Function(test.function_space())

        if trial.argument == 2:
            # linear form
            for (te, tr, x) in zip(test, trial, output_array):
                x = inner(te, tr, output_array=x)
            return output_array

        result = []
        for te, tr in zip(test, trial):
            l = inner(te, tr, level=level)
            result += l if isinstance(l, list) else [l]
        return result[0] if len(result) == 1 else result

    if output_array is None and trial.argument == 2:
        output_array = Function(test.function_space())

    if trial.argument > 1:
        # Linear form
        assert isinstance(test, (Expr, BasisFunction))
        assert test.argument == 0
        space = test.function_space()
        if isinstance(trial, Array):
            if trial.rank == 0:
                output_array = space.scalar_product(trial, output_array)
                return output_array
            trial = trial.forward()

    # If trial is an Expr with terms, then compute using bilinear form and matvec

    assert isinstance(trial, (Expr, BasisFunction))
    assert isinstance(test, (Expr, BasisFunction))

    if isinstance(trial, BasisFunction):
        trial = Expr(trial)
    if isinstance(test, BasisFunction):
        test = Expr(test)

    assert test.expr_rank() == trial.expr_rank()
    testspace = test.base.function_space()
    trialspace = trial.base.function_space()
    test_scale = test.scales()
    trial_scale = trial.scales()

    uh = None
    if trial.argument == 2:
        uh = trial.base

    A = []
    for vec, (base_test, base_trial, test_ind, trial_ind) in enumerate(
            zip(test.terms(), trial.terms(), test.indices(),
                trial.indices())):  # vector/scalar
        for test_j, b0 in enumerate(base_test):  # second index test
            for trial_j, b1 in enumerate(base_trial):  # second index trial
                sc = test_scale[vec, test_j] * trial_scale[vec, trial_j]
                scb = sc
                M = []
                DM = []
                assert len(b0) == len(b1)
                trial_sp = trialspace
                if isinstance(
                        trialspace, (MixedTensorProductSpace, MixedBasis)
                ):  # could operate on a vector, e.g., div(u), where u is vector
                    trial_sp = trialspace.flatten()[trial_ind[trial_j]]
                test_sp = testspace
                if isinstance(testspace,
                              (MixedTensorProductSpace, MixedBasis)):
                    test_sp = testspace.flatten()[test_ind[test_j]]
                has_bcs = False
                #assert test_sp.compatible_base(trial_sp)
                for i, (a, b) in enumerate(zip(
                        b0, b1)):  # Third index, one inner for each dimension
                    ts = trial_sp[i]
                    sp = test_sp[i]
                    AA = inner_product((sp, a), (ts, b))
                    M.append(AA)
                    # Take care of domains of not standard size
                    if not sp.domain_factor() == 1:
                        sc *= sp.domain_factor()**(a + b)
                        scb *= sp.domain_factor()**(a + b)
                    if not abs(AA.scale - 1.) < 1e-8:
                        sc *= AA.scale
                        AA.scale = 1.0

                    if (ts.boundary_condition() == 'Dirichlet'
                            and not ts.family() in ('laguerre', 'hermite')
                            or (ts.boundary_condition() == 'Biharmonic'
                                and not ts.family() in ('jacobi', ))):
                        if ts.bc.has_nonhomogeneous_bcs():
                            tsc = ts.get_bc_basis()
                            BB = inner_product((sp, a), (tsc, b))
                            if not abs(BB.scale - 1.) < 1e-8:
                                scb *= BB.scale
                                BB.scale = 1.0
                            if BB:
                                DM.append(BB)
                                has_bcs = True
                        else:
                            DM.append(AA)
                    else:
                        DM.append(AA)

                sc = sp.broadcast_to_ndims(np.array([sc]))
                if len(M) == 1:  # 1D case
                    M[0].global_index = (test_ind[test_j], trial_ind[trial_j])
                    M[0].scale = sc[0]
                    M[0].mixedbase = testspace
                    A.append(M[0])
                else:
                    A.append(
                        TPMatrix(M, test_sp, sc,
                                 (test_ind[test_j], trial_ind[trial_j]),
                                 testspace))
                if has_bcs:
                    if len(DM) == 1:  # 1D case
                        DM[0].global_index = (test_ind[test_j],
                                              trial_ind[trial_j])
                        DM[0].scale = scb
                        DM[0].mixedbase = testspace
                        A.append(DM[0])
                    else:
                        A.append(
                            TPMatrix(DM, test_sp, sc,
                                     (test_ind[test_j], trial_ind[trial_j]),
                                     testspace))

    # At this point A contains all matrices of the form. The length of A is
    # the number of inner products. For each index into A there are ndim 1D
    # inner products along, e.g., x, y and z-directions, or just x, y for 2D.
    # The outer product of these matrices is a tensorproduct matrix, and we
    # store the matrices using the TPMatrix class.
    #
    # Diagonal matrices can be eliminated and put in a scale array for the
    # non-diagonal matrices. E.g. for (v, div(grad(u))) in 2D
    #
    # Here A = [TPMatrix([(v[0], u[0]'')_x, (v[1], u[1])_y]),
    #           TPMatrix([(v[0], u[0])_x, (v[1], u[1]'')_y])]
    #
    # where v[0], v[1] are the test functions in x- and y-directions,
    # respectively. For example, v[0] could be a ShenDirichletBasis and v[1]
    # could be a FourierBasis. Same for u.
    #
    # There are now two possibilities, either a linear or a bilinear form.
    # A linear form has trial.argument == 2, whereas a bilinear form has
    # trial.argument == 1. A linear form should assemble to an array and
    # return this array. A bilinear form, on the other hand, should return
    # matrices. Which matrices, and how many, will of course depend on the
    # form and the number of terms.
    #
    # Considering again the tensor product space with ShenDirichlet and Fourier,
    # the list A will contain matrices as shown above. If Fourier is associated
    # with index 1, then (v[1], u[1])_y and (v[1], u[1]'')_y will be diagonal
    # whereas (v[0], u[0]'')_x and (v[0], u[0])_x will in general not. These
    # two matrices are usually termed the stiffness and mass matrices, and they
    # have been implemented in chebyshev/matrices.py or legendre/matrices.py,
    # where they are called ADDmat and BDDmat, respectively.

    if level == 2 and trial.argument == 1:  # No processing of matrices
        return A

    for tpmat in A:
        if isinstance(tpmat, TPMatrix):
            tpmat.simplify_fourier_matrices()

    # Add equal matrices
    B = [A[0]]
    for a in A[1:]:
        found = False
        for b in B:
            if a == b:
                b += a
                found = True
        if not found:
            B.append(a)

    A = B

    if trial.argument == 1:
        return A[0] if len(A) == 1 else A

    wh = np.zeros_like(output_array)
    for b in A:
        if uh.rank > 0:
            wh = b.matvec(uh.v[b.global_index[1]], wh)
        else:
            wh = b.matvec(uh, wh)
        output_array += wh
        wh.fill(0)
    return output_array
Exemplo n.º 2
0
def inner(expr0, expr1, output_array=None, level=0):
    r"""
    Return (weighted or unweighted) discrete inner product of kind

    .. math::

        (f, g)_w^N = \sum_{i\in\mathcal{I}}f(x_i) \overline{g}(x_i) w_i \approx \int_{\Omega} f\, \overline{g}\, w\, dx

    where :math:`\mathcal{I}=0, 1, \ldots, N, N \in \mathbb{Z}^+`, :math:`f`
    is a number or an expression linear in a :class:`.TestFunction`,
    and :math:`g` is an expression that is linear in :class:`.TrialFunction`
    or :class:`.Function`, or it is simply an :class:`.Array` (a solution interpolated on the
    quadrature mesh in physical space). :math:`w` is a weight associated with
    chosen basis, and :math:`w_i` are quadrature weights.

    Note
    ----
        If :math:`f` is a number (typically one) and :math:`g` an :class:`.Array`, then `inner`
        represents an unweighted, regular integral over the domain.

    If the expressions are created in a multidimensional :class:`.TensorProductSpace`,
    then the sum above is over all dimensions. In 2D it becomes:

    .. math::

        (f, g)_w^N = \sum_{i\in\mathcal{I}}\sum_{j\in\mathcal{J}} f(x_i, y_j) \overline{g}(x_i, y_j) w_j w_i

    where :math:`\mathcal{J}=0, 1, \ldots, M, M \in \mathbb{Z}^+`.

    Parameters
    ----------
    expr0, expr1 : :class:`.Expr`, :class:`.BasisFunction`, :class:`.Array`, number
        Either one can be an expression involving a
        BasisFunction (:class:`.TestFunction`, :class:`.TrialFunction` or
        :class:`.Function`) an Array or a number. With expressions (Expr) on a
        BasisFunction we typically mean terms like div(u) or grad(u), where
        u is any one of the different types of BasisFunction.
        If one of ``expr0``/``expr1`` involves a TestFunction and the other one
        involves a TrialFunction, then a tensor product matrix (or a list of
        tensor product matrices) is returned.
        If one of ``expr0``/``expr1`` involves a TestFunction and the other one
        involves a Function, or a plain Array, then a linear form is assembled
        and a Function is returned.
        If one of ``expr0``/``expr1`` is a number (typically 1), or a tuple of
        numbers for a vector, then the inner product represents a non-weighted
        integral over the domain. If a single number, then the other expression
        must be a scalar Array or a Function. If a tuple of numbers, then the
        Array/Function must be a vector.

    output_array:  Function
        Optional return array for linear form.

    level: int
        The level of postprocessing for assembled matrices. Applies only
        to bilinear forms

        - 0 Full postprocessing - diagonal matrices to scale arrays
          and add equal matrices
        - 1 Diagonal matrices to scale arrays, but don't add equal
          matrices
        - 2 No postprocessing, return all assembled matrices

    Returns
    -------
    Depending on dimensionality and the arguments to the forms

        :class:`.Function`
        for linear forms.

        :class:`.SparseMatrix`
        for bilinear 1D forms.

        :class:`.TPMatrix` or list of :class:`.TPMatrix`
        for bilinear multidimensional forms.

        Number, for non-weighted integral where either one of the arguments
        is a number.

    See Also
    --------
    :func:`.project`

    Example
    -------
    Compute mass matrix of Shen's Chebyshev Dirichlet basis:

    >>> from shenfun import FunctionSpace, TensorProductSpace
    >>> from shenfun import TestFunction, TrialFunction, Array
    >>> SD = FunctionSpace(6, 'Chebyshev', bc=(0, 0))
    >>> u = TrialFunction(SD)
    >>> v = TestFunction(SD)
    >>> B = inner(v, u)
    >>> d = {-2: np.array([-np.pi/2]),
    ...       0: np.array([ 1.5*np.pi, np.pi, np.pi, np.pi]),
    ...       2: np.array([-np.pi/2])}
    >>> [np.all(abs(B[k]-v) < 1e-7) for k, v in d.items()]
    [True, True, True]

    # Compute unweighted integral
    >>> F = FunctionSpace(10, 'F', domain=(0, 2*np.pi))
    >>> T = TensorProductSpace(comm, (SD, F))
    >>> area = inner(1, Array(T, val=1))
    >>> print('Area of domain =', area)
    Area of domain = 12.56637061435917

    """
    # Wrap a pure numpy array in Array
    if isinstance(expr0,
                  np.ndarray) and not isinstance(expr0, (Array, Function)):
        assert isinstance(expr1, (Expr, BasisFunction))
        if not expr0.flags['C_CONTIGUOUS']:
            expr0 = expr0.copy()
        expr0 = Array(expr1.function_space(), buffer=expr0)
    if isinstance(expr1,
                  np.ndarray) and not isinstance(expr1, (Array, Function)):
        assert isinstance(expr0, (Expr, BasisFunction))
        if not expr1.flags['C_CONTIGUOUS']:
            expr1 = expr1.copy()
        expr1 = Array(expr0.function_space(), buffer=expr1)

    if isinstance(expr0, Number):
        assert isinstance(expr1, (Array, Function))
        space = expr1.function_space()
        if isinstance(space, (TensorProductSpace, CompositeSpace)):
            df = np.prod(
                np.array([base.domain_factor() for base in space.bases]))
        elif isinstance(space, SpectralBase):
            df = space.domain_factor()
        if isinstance(expr1, Function):
            #return (expr0/df)*dx(expr1.backward())
            expr1 = expr1.backward()
        if hasattr(space, 'hi'):
            if space.hi.prod() != 1:
                expr1 = space.get_measured_array(expr1.copy())

        return (expr0 / df) * dx(expr1)

    if isinstance(expr1, Number):
        assert isinstance(expr0, (Array, Function))
        space = expr0.function_space()
        if isinstance(space, (TensorProductSpace, CompositeSpace)):
            df = np.prod(
                np.array([base.domain_factor() for base in space.bases]))
        elif isinstance(space, SpectralBase):
            df = space.domain_factor()
        if isinstance(expr0, Function):
            #return (expr1/df)*dx(expr0.backward())
            expr0 = expr0.backward()
        if hasattr(space, 'hi'):
            if space.hi.prod() != 1:
                expr0 = space.get_measured_array(expr0.copy())

        return (expr1 / df) * dx(expr0)

    if isinstance(expr0, tuple):
        assert isinstance(expr1, (Array, Function))
        space = expr1.function_space()
        assert isinstance(space, CompositeSpace)
        assert len(expr0) == len(space)
        result = 0.0
        for e0i, e1i in zip(expr0, expr1):
            result += inner(e0i, e1i)
        return result

    if isinstance(expr1, tuple):
        assert isinstance(expr0, (Array, Function))
        space = expr0.function_space()
        assert isinstance(space, CompositeSpace)
        assert len(expr1) == len(space)
        result = 0.0
        for e0i, e1i in zip(expr0, expr1):
            result += inner(e0i, e1i)
        return result

    assert np.all([hasattr(e, 'argument') for e in (expr0, expr1)])
    t0 = expr0.argument
    t1 = expr1.argument
    if t0 == 0:
        assert t1 in (1, 2)
        test = expr0
        trial = expr1
    elif t0 in (1, 2):
        assert t1 == 0
        test = expr1
        trial = expr0
    else:
        raise RuntimeError

    #if isinstance(test, BasisFunction):
    #    recursive = test.tensor_rank > 0
    #elif isinstance(test, Expr):
    recursive = test.function_space().is_composite_space
    if isinstance(trial, Array):
        assert trial.tensor_rank == test.tensor_rank
    elif isinstance(trial, BasisFunction):
        recursive *= (trial.expr_rank() > 0)
    if test.expr_rank() == 0:
        recursive = False

    if recursive:  # Use recursive algorithm for vector expressions of expr_rank > 0, e.g., inner(v, grad(u))
        if output_array is None and trial.argument == 2:
            output_array = Function(test.function_space())

        gij = test.function_space().coors.get_covariant_metric_tensor()
        if trial.argument == 2:
            # linear form

            if test.tensor_rank == 2:
                w0 = np.zeros_like(output_array[0][0])
                for i, (tei, xi) in enumerate(zip(test, output_array)):
                    for j, (teij, xij) in enumerate(zip(tei, xi)):
                        for k, trk in enumerate(trial):
                            if gij[i, k] == 0:
                                continue
                            for l, trkl in enumerate(trk):
                                if gij[j, l] == 0:
                                    continue

                                w0.fill(0)
                                xij += inner(teij * gij[i, k] * gij[j, l],
                                             trkl,
                                             output_array=w0)

            elif test.tensor_rank == 1:
                w0 = np.zeros_like(output_array[0])
                for i, (te, x) in enumerate(zip(test, output_array)):
                    for j, tr in enumerate(trial):
                        if gij[i, j] == 0:
                            continue
                        w0.fill(0)
                        x += inner(te * gij[i, j], tr, output_array=w0)

            return output_array

        result = []

        if test.tensor_rank == 2:
            for i, tei in enumerate(test):
                for j, teij in enumerate(tei):
                    for k, trk in enumerate(trial):
                        if gij[i, k] == 0:
                            continue
                        for l, trkl in enumerate(trk):
                            if gij[j, l] == 0:
                                continue
                            p = inner(teij * gij[i, k] * gij[j, l],
                                      trkl,
                                      level=level)
                            result += p if isinstance(p, list) else [p]

        elif test.tensor_rank == 1:
            for i, te in enumerate(test):
                for j, tr in enumerate(trial):
                    if gij[i, j] == 0:
                        continue
                    l = inner(te, tr * gij[i, j], level=level)
                    result += l if isinstance(l, list) else [l]

        return result[0] if len(result) == 1 else result

    if output_array is None and trial.argument == 2:
        output_array = Function(test.function_space())

    if trial.argument > 1:
        # Linear form
        assert isinstance(test, (Expr, BasisFunction))
        assert test.argument == 0
        space = test.function_space()
        if isinstance(trial, Array):
            if trial.tensor_rank == 0:
                output_array = space.scalar_product(trial, output_array)
                return output_array
            trial = trial.forward()

    assert isinstance(trial, (Expr, BasisFunction))
    assert isinstance(test, (Expr, BasisFunction))

    if isinstance(trial, BasisFunction):
        trial = Expr(trial)
    if isinstance(test, BasisFunction):
        test = Expr(test)

    assert test.expr_rank() == trial.expr_rank()

    testspace = test.base.function_space()
    trialspace = trial.base.function_space()
    test_scale = test.scales()
    trial_scale = trial.scales()

    uh = None
    if trial.argument == 2:
        uh = trial.base

    A = []
    gij = testspace.coors.get_covariant_metric_tensor()
    for vec_i, (base_test,
                test_ind) in enumerate(zip(test.terms(),
                                           test.indices())):  # vector/scalar
        for vec_j, (base_trial,
                    trial_ind) in enumerate(zip(trial.terms(),
                                                trial.indices())):
            g = 1 if len(test.terms()) == 1 else gij[vec_i, vec_j]
            if g == 0:
                continue
            for test_j, b0 in enumerate(base_test):  # second index test
                for trial_j, b1 in enumerate(base_trial):  # second index trial
                    dV = sp.simplify(test_scale[vec_i][test_j] *
                                     trial_scale[vec_j][trial_j] *
                                     testspace.coors.sg * g,
                                     measure=testspace.coors._measure)
                    dV = testspace.coors.refine(dV)

                    assert len(b0) == len(b1)
                    trial_sp = trialspace
                    if isinstance(
                            trialspace, (CompositeSpace, MixedFunctionSpace)
                    ):  # could operate on a vector, e.g., div(u), where u is vector
                        trial_sp = trialspace.flatten()[trial_ind[trial_j]]
                    test_sp = testspace
                    if isinstance(testspace,
                                  (CompositeSpace, MixedFunctionSpace)):
                        test_sp = testspace.flatten()[test_ind[test_j]]
                    has_bcs = False
                    # Check if scale is zero
                    if dV == 0:
                        continue
                    for dv in split(dV):
                        sc = dv['coeff']
                        scb = dv['coeff']
                        M = []
                        DM = []
                        for i, (a, b) in enumerate(
                                zip(b0, b1)
                        ):  # Third index, one inner for each dimension
                            ts = trial_sp[i]
                            tt = test_sp[i]
                            msx = 'xyzrs'[i]
                            msi = dv[msx]

                            # assemble inner product
                            AA = inner_product((tt, a), (ts, b), msi)
                            if len(AA) == 0:
                                AA = Identity(AA.shape, scale=0)

                            M.append(AA)
                            if not abs(AA.scale - 1.) < 1e-8:
                                sc *= AA.scale
                                AA.scale = 1.0

                            if ts.has_nonhomogeneous_bcs:
                                tsc = ts.get_bc_basis()
                                BB = inner_product((tt, a), (tsc, b), msi)
                                if not abs(BB.scale - 1.) < 1e-8:
                                    scb *= BB.scale
                                    BB.scale = 1.0
                                if BB:
                                    DM.append(BB)
                                    has_bcs = True
                                else:
                                    DM.append(0)
                            else:
                                DM.append(AA)

                        sc = tt.broadcast_to_ndims(np.array([sc]))
                        if len(M) == 1:  # 1D case
                            M[0].global_index = (test_ind[test_j],
                                                 trial_ind[trial_j])
                            M[0].scale = sc[0]
                            M[0].mixedbase = testspace
                            A.append(M[0])
                        else:
                            A.append(
                                TPMatrix(
                                    M, test_sp, trial_sp, sc,
                                    (test_ind[test_j], trial_ind[trial_j]),
                                    testspace))
                        if has_bcs:
                            if len(DM) == 1:  # 1D case
                                DM[0].global_index = (test_ind[test_j],
                                                      trial_ind[trial_j])
                                DM[0].scale = scb
                                DM[0].mixedbase = testspace
                                A.append(DM[0])
                            else:
                                if len(trial_sp.get_nonhomogeneous_axes()
                                       ) == 1:
                                    A.append(
                                        TPMatrix(DM, test_sp, trial_sp, sc,
                                                 (test_ind[test_j],
                                                  trial_ind[trial_j]),
                                                 testspace))
                                elif len(trial_sp.get_nonhomogeneous_axes()
                                         ) == 2:
                                    if DM[1] is not 0:
                                        A.append(
                                            TPMatrix([M[0], DM[1]], test_sp,
                                                     trial_sp, sc,
                                                     (test_ind[test_j],
                                                      trial_ind[trial_j]),
                                                     testspace))
                                    if DM[0] is not 0:
                                        A.append(
                                            TPMatrix([DM[0], M[1]], test_sp,
                                                     trial_sp, sc,
                                                     (test_ind[test_j],
                                                      trial_ind[trial_j]),
                                                     testspace))
                                    if DM[0] is not 0 and DM[1] is not 0:
                                        A.append(
                                            TPMatrix(DM, test_sp, trial_sp, sc,
                                                     (test_ind[test_j],
                                                      trial_ind[trial_j]),
                                                     testspace))

    # At this point A contains all matrices of the form. The length of A is
    # the number of inner products. For each index into A there are ndim 1D
    # inner products along, e.g., x, y and z-directions, or just x, y for 2D.
    # The outer product of these matrices is a tensorproduct matrix, and we
    # store the matrices using the TPMatrix class.
    #
    # Diagonal matrices can be eliminated and put in a scale array for the
    # non-diagonal matrices. E.g. for (v, div(grad(u))) in 2D
    #
    # Here A = [TPMatrix([(v[0], u[0]'')_x, (v[1], u[1])_y]),
    #           TPMatrix([(v[0], u[0])_x, (v[1], u[1]'')_y])]
    #
    # where v[0], v[1] are the test functions in x- and y-directions,
    # respectively. For example, v[0] could be a ShenDirichlet and v[1]
    # could be a Fourier space. Same for u.
    #
    # There are now two possibilities, either a linear or a bilinear form.
    # A linear form has trial.argument == 2, whereas a bilinear form has
    # trial.argument == 1. A linear form should assemble to an array and
    # return this array. A bilinear form, on the other hand, should return
    # matrices. Which matrices, and how many, will of course depend on the
    # form and the number of terms.
    #
    # Considering again the tensor product space with ShenDirichlet and Fourier,
    # the list A will contain matrices as shown above. If Fourier is associated
    # with index 1, then (v[1], u[1])_y and (v[1], u[1]'')_y will be diagonal
    # whereas (v[0], u[0]'')_x and (v[0], u[0])_x will in general not.

    if level == 2 and trial.argument == 1:  # No processing of matrices
        return A

    for tpmat in A:
        if isinstance(tpmat, TPMatrix):
            try:
                tpmat.simplify_diagonal_matrices()
            except KeyError:
                continue

    # Add equal matrices
    B = [A[0]]
    for a in A[1:]:
        found = False
        for b in B:
            if a == b:
                b += a
                found = True
        if not found:
            B.append(a)

    A = B

    # Bilinear form, return matrices
    if trial.argument == 1:
        return A[0] if len(A) == 1 else A

    # Linear form, return output_array
    wh = np.zeros_like(output_array)
    for b in A:
        if uh.function_space().is_composite_space:
            wh = b.matvec(uh.v[b.global_index[1]], wh)
        else:
            wh = b.matvec(uh, wh)
        output_array += wh
        wh.fill(0)
    return output_array