Repository navigation
MPI
- Elastic, Newmark and Newton solved over the ranks match the serial solve. A solver
other thanpetscraises, as doPhaseFieldBoundConstrain(lsq_linearis serial)
and theSolveof a gathered simulation;Plot_Energyneeds a gathered one. EASYFEA_MPI=offruns one serial simulation per rank, writing one results file per
rank.- The Newton residual norm over the ranks is the root of the summed squares. Under PETSc
no orphan diagonal is written: it reads only the owned rows. Meshelement sums (mass, center) count each element once over the ranks; groups store
isPartitionedandMesh.centeris the rank's.PhaseFieldenergies take their zero shortcut on a reduced value, so a diskSet_Iter
no longer hangs.print(simu)prints on rank 0 only and never communicates: partitioned results are
rank 0's, under a warning.use_mpi.mdlaunches withpython -m mpi4py, so an uncaught exception aborts every
rank.
Visualization
Vizir._Write_element_solution_file: values constant over each element, written as a
degree 0HOSolAt. Every dim-2 group takes the mean of the main-dimension elements
holding all its nodes, so a boundary face gets its element's value; main-dimension
groups too on a 3D mesh.Viziropens a solution file through one context manager writing its header, warp
vector and end.
API
Simu.Get_K_C_M_Freturns the assembled matrices, not copies; the time-schemeAis
summed on the values whenCandMshareK's pattern.- Assembling a complex system raises:
bandxare real ndarrays; the CSR column and
asCsrMatrixare gone. BiLinearFormrows are test functions, columns trial functions.
Development
- CI: the MPI job installs BLAS and LAPACK and builds
petscandpetsc4pywithuv,
cached, so the solves run. ruff formatreplacesblack.
Full Changelog: v7.2.0...v7.3.0