""" Sum factorized geometry evaluations """ from dune.perftool.generation import (domain, get_backend, get_counted_variable, iname, instruction, kernel_cached, temporary_variable, ) from dune.perftool.loopy.buffer import get_buffer_temporary from dune.perftool.pdelab.geometry import (local_dimension, world_dimension, ) from dune.perftool.sumfact.symbolic import SumfactKernelInputBase from dune.perftool.sumfact.vectorization import attach_vectorization_info from pytools import ImmutableRecord import pymbolic.primitives as prim @iname def corner_iname(): name = get_counted_variable("corneriname") domain(name, 2 ** local_dimension()) return name class GeoCornersInput(SumfactKernelInputBase, ImmutableRecord): def __init__(self, dir): ImmutableRecord.__init__(self, dir=dir) def realize(self, sf, index, insn_dep): name = get_buffer_temporary(sf.buffer, shape=(2 ** local_dimension(), sf.vector_width), name="input_{}".format(sf.buffer) ) ciname = corner_iname() from dune.perftool.pdelab.geometry import name_geometry geo = name_geometry() # NB: We need to realize this as a C instruction, because the corner # method does return a non-scalar, which does not fit into the current # loopy philosophy for function calls. This problem will be solved once # #11 is resolved. Admittedly, the code looks *really* ugly until that happens. code = "{}[{}*{}+{}] = {}.corner({})[{}];".format(name, sf.vector_width, ciname, index, geo, ciname, self.dir, ) instruction(code=code, within_inames=frozenset({ciname}), assignees=(name,), tags=frozenset({"sumfact_stage{}".format(sf.stage)}), ) @kernel_cached def pymbolic_spatial_coordinate(visitor_indices): assert len(visitor_indices) == 1 # Construct the matrix sequence for the evaluation of the global coordinate. # We need to manually construct this one, because on facets, we want to use the # geometry embedding of the facet into the global space directly without going # through the neighboring cell geometries. That matrix sequence will only have # dim-1 matrices! from dune.perftool.sumfact.tabulation import quadrature_points_per_direction, BasisTabulationMatrix quadrature_size = quadrature_points_per_direction() matrix_sequence = (BasisTabulationMatrix(quadrature_size=quadrature_size, basis_size=2),) * local_dimension() inp = GeoCornersInput(visitor_indices[0]) from dune.perftool.sumfact.symbolic import SumfactKernel sf = SumfactKernel(matrix_sequence=matrix_sequence, input=inp, ) vsf = attach_vectorization_info(sf) # Add a sum factorization kernel that implements the evaluation of # the basis functions at quadrature points (stage 1) from dune.perftool.sumfact.realization import realize_sum_factorization_kernel var, _ = realize_sum_factorization_kernel(vsf) return prim.Subscript(var, vsf.quadrature_index(sf)), None