scimate-ai/hpc-skills/hpc-hypre

HPC hypre

Treat hypre as a solver-library family with distinct matrix interfaces and multigrid-heavy preconditioning workflows.

Start

1. Read references/interface-selection.md before creating or repairing a hypre solve path.
2. Read references/matrix-and-ownership-playbook.md when choosing row ownership, indexing, or assembly discipline.
3. Read references/boomeramg-and-krylov-playbook.md when choosing BoomerAMG, Krylov solvers, or a first stable parameter set.
4. Read references/boomeramg-parameter-matrix.md when changing coarsening, interpolation, smoothing, or cycle choices.
5. Read references/krylov-and-preconditioner-matrix.md when the operator class does not fit the default PCG plus BoomerAMG baseline.
6. Read references/structured-and-sstruct-playbook.md when the problem is logically structured or semi-structured.
7. Read references/build-and-integration.md when building hypre directly or consuming it through PETSc or another host code.
8. Read references/error-recovery.md when setup, solve, or parallel behavior fails.

Work sequence

1. Classify the discretization storage first:

• general sparse assembled system -> IJ
• structured grid with regular stencil semantics -> Struct
• mostly structured with coupled variable blocks or irregular pieces -> SStruct

1. Choose a baseline solver pair that matches the operator:

• SPD-like problems -> PCG with BoomerAMG
• general nonsymmetric problems -> GMRES or FlexGMRES with BoomerAMG as the preconditioner

1. Get the baseline solve working before changing coarsening, interpolation, or relaxation details.
2. Keep distributed row ownership and matrix assembly consistent across ranks.

Guardrails

• Do not use the structured interfaces when the data model is really an unstructured sparse matrix.
• Do not start by over-tuning BoomerAMG; defaults are the first checkpoint.
• Do not debug multigrid parameters before confirming the operator, nullspace, and boundary conditions are coherent.
• Do not assume host applications expose every hypre knob unless the integration layer documents it.

Additional References

Load these on demand:

• references/matrix-and-ownership-playbook.md for IJ assembly and distributed indexing rules
• references/boomeramg-parameter-matrix.md for practical AMG knob selection
• references/krylov-and-preconditioner-matrix.md for solver-family selection beyond the default baseline
• references/structured-and-sstruct-playbook.md for structured-grid cases and hybrid layouts
• references/build-and-integration.md for PETSc-backed or direct-link builds
• references/error-recovery.md for setup, memory, and divergence signatures
• references/error-pattern-dictionary.md for quick matching of common hypre failure classes

Reusable Templates

Use assets/templates/ when a concrete starting point is faster than rebuilding the solve path from scratch, especially:

• hypre_ij_pcg_boomeramg.c
• hypre_build_example.sh
• hypre_ij_slurm.sh

Outputs

Summarize:

• chosen hypre interface
• baseline solver and preconditioner path
• multigrid knobs changed from default
• integration path such as direct hypre or PETSc-mediated usage
• the exact failure stage if the workflow is being repaired