PCHYPRE#

Allows you to use the matrix element based preconditioners in the LLNL package hypre as PETSc PC

Options Database Keys#

-pc_hypre_type - One of euclid, pilut, parasails, boomeramg, ams, or ads
-pc_hypre_boomeramg_nodal_coarsen - where n is from 1 to 6 (see HYPRE_BOOMERAMGSetNodal())
-pc_hypre_boomeramg_vec_interp_variant - where v is from 1 to 3 (see HYPRE_BoomerAMGSetInterpVecVariant())
***Many others, run with -*** pc_type hypre -pc_hypre_type XXX -help to see options for the XXX preconditioner

Notes#

Apart from -pc_hypre_type (for which there is PCHYPRESetType()), the many hypre options can ONLY be set via the options database (e.g. the command line or with PetscOptionsSetValue(), there are no functions to set them)

The options -pc_hypre_boomeramg_max_iter and -pc_hypre_boomeramg_tol refer to the number of iterations (V-cycles) and tolerance that boomerAMG does EACH time it is called. So for example, if -pc_hypre_boomeramg_max_iter is set to 2 then 2-V-cycles are being used to define the preconditioner (-pc_hypre_boomeramg_tol should be set to 0.0 - the default - to strictly use a fixed number of iterations per hypre call). -ksp_max_it and -ksp_rtol STILL determine the total number of iterations and tolerance for the Krylov solver. For example, if -pc_hypre_boomeramg_max_iter is 2 and -ksp_max_it is 10 then AT MOST twenty V-cycles of boomeramg will be used.

Note that the option -pc_hypre_boomeramg_relax_type_all defaults to symmetric relaxation (symmetric-SOR/Jacobi), which is required for Krylov solvers like CG that expect symmetry. Otherwise, you may want to use -pc_hypre_boomeramg_relax_type_all SOR/Jacobi.

MatSetNearNullSpace() - if you provide a near null space to your matrix it is ignored by hypre UNLESS you also use the following two options: -pc_hypre_boomeramg_nodal_coarsen <n> -pc_hypre_boomeramg_vec_interp_variant <v>

See PCPFMG, PCSMG, and PCSYSPFMG for access to hypre’s other (nonalgebraic) multigrid solvers

For PCHYPRE type of ams or ads auxiliary data must be provided to the preconditioner with PCHYPRESetDiscreteGradient(), PCHYPRESetDiscreteCurl(), PCHYPRESetInterpolations(), PCHYPRESetAlphaPoissonMatrix(), PCHYPRESetBetaPoissonMatrix(), PCHYPRESetEdgeConstantVectors(), PCHYPREAMSSetInteriorNodes()

Sometimes people want to try algebraic multigrid as a “standalone” solver, that is not accelerating it with a Krylov method. Though we generally do not recommend this since it is usually slower, one should use a KSPType of KSPRICHARDSON (or equivalently -ksp_type richardson) to achieve this. Using KSPPREONLY will not work since it only applies a single cycle of multigrid.

PETSc provides its own geometric and algebraic multigrid solvers PCMG and PCGAMG, also see PCHMG which is useful for certain multicomponent problems

GPU Notes#

To configure hypre BoomerAMG so that it can utilize NVIDIA GPUs run ./configure –download-hypre –with-cuda Then pass VECCUDA vectors and MATAIJCUSPARSE matrices to the solvers and PETSc will automatically utilize hypre’s GPU solvers.

To configure hypre BoomerAMG so that it can utilize AMD GPUs run ./configure –download-hypre –with-hip Then pass VECHIP vectors to the solvers and PETSc will automatically utilize hypre’s GPU solvers.

Level#

intermediate

Location#

src/ksp/pc/impls/hypre/hypre.c

Index of all PC routines
Table of Contents for all manual pages
Index of all manual pages