S = D - C inv(A) B
PCLSC currently doesn't do anything with D, so let's assume it is 0. The idea is that a good approximation to inv(S) is given by
inv(CB) C A B inv(CB)
At some point, we'll be able to form the product CB for you, but for now the application has to provide it (this is usually more efficient anyway). In the case of incompressible flow, CB is a Laplacian, call it L. The current interface is to hang L and a preconditioning matrix Lp on the preconditioning matrix.
If you had called KSPSetOperators(ksp,S,Sp,flg), S should have type MATSCHURCOMPLEMENT and Sp can be any type you like (PCLSC doesn't use it directly) but should have matrices composed with it, under the names "LSC_L" and "LSC_Lp". For example, you might have setup code like this
PetscObjectCompose((PetscObject)Sp,"LSC_L",(PetscObject)L); PetscObjectCompose((PetscObject)Sp,"LSC_Lp",(PetscObject)Lp);
And then your Jacobian assembly would look like
PetscObjectQuery((PetscObject)Sp,"LSC_L",(PetscObject*)&L); PetscObjectQuery((PetscObject)Sp,"LSC_Lp",(PetscObject*)&Lp); if (L) { assembly L } if (Lp) { assemble Lp }
With this, you should be able to choose LSC preconditioning, using e.g. ML's algebraic multigrid to solve with L
-fieldsplit_1_pc_type lsc -fieldsplit_1_lsc_pc_type ml
Since we do not use the values in Sp, you can still put an assembled matrix there to use normal preconditioners.
Level:intermediate
Location:src/ksp/pc/impls/lsc/lsc.c
Index of all PC routines
Table of Contents for all manual pages
Index of all manual pages