Speed up (block) matrix assembly by merging SparsityPattern of the blocks

[dokken]

@deepspace

fem.petsc._assemble_matrix_mat(
        mat,
        fem.form(form),
        bcs=bcs,
        diagonal=diagonal,
    )

For this you should be able to just call

fem.petsc.assemble_matrix(
        mat,
        fem.form(form),
        bcs=bcs,
        diagonal=diagonal,
    )

and as shown in:

Modify matrix diagonal -- dolfinx version for A.ident_zeros()

I would suggest the following: import time import dolfinx import numpy as np import ufl from dolfinx import cpp, fem from mpi4py import MPI from petsc4py import PETSc N = 75 degree = 1 mesh = dolfinx.mesh.create_unit_cube(MPI.COMM_WORLD, N, N, N) V = dolfinx.fem.FunctionSpace(mesh, ("CG", degree)) u = ufl.TrialFunction(V) v = ufl.TestFunction(V) print( f"N={N}, Num dofs global {V.dofmap.index_map.size_global*V.dofmap.index_map_bs}") # form is defined on the boundary => matrix wil have man…

you can create your sparsity pattern (and send that into the create_matrix function)

sp = fem.create_sparsity_pattern(form)
sp.insert_diagonal(zero_blocks)
sp.assemble()
end = time.perf_counter()
A = cpp.la.petsc.create_matrix(mesh.comm, sp)

Here I used insert_diagonal as I only wanted diagonal entries, but you can use sp.insert to insert combinations of row and column entries: https://github.com/FEniCS/dolfinx/blob/main/python/dolfinx/wrappers/la.cpp#L254-L263