static char help[] = "Test file for the PCFactorSetShiftType()\n"; /* * Test file for the PCFactorSetShiftType() routine or -pc_factor_shift_type POSITIVE_DEFINITE option. * The test matrix is the example from Kershaw's paper [J.Comp.Phys 1978] * of a positive definite matrix for which ILU(0) will give a negative pivot. * This means that the CG method will break down; the Manteuffel shift * [Math. Comp. 1980] repairs this. * * Run the executable twice: * 1/ without options: the iterative method diverges because of an * indefinite preconditioner * 2/ with -pc_factor_shift_positive_definite option (or comment in the PCFactorSetShiftType() line below): * the method will now successfully converge. */ #include int main(int argc,char **argv) { KSP ksp; PC pc; Mat A,M; Vec X,B,D; MPI_Comm comm; PetscScalar v; KSPConvergedReason reason; PetscInt i,j,its; PetscErrorCode ierr; PetscFunctionBegin; ierr = PetscInitialize(&argc,&argv,0,help);if (ierr) return ierr; comm = MPI_COMM_SELF; /* * Construct the Kershaw matrix * and a suitable rhs / initial guess */ CHKERRQ(MatCreateSeqAIJ(comm,4,4,4,0,&A)); CHKERRQ(VecCreateSeq(comm,4,&B)); CHKERRQ(VecDuplicate(B,&X)); for (i=0; i<4; i++) { v = 3; CHKERRQ(MatSetValues(A,1,&i,1,&i,&v,INSERT_VALUES)); v = 1; CHKERRQ(VecSetValues(B,1,&i,&v,INSERT_VALUES)); CHKERRQ(VecSetValues(X,1,&i,&v,INSERT_VALUES)); } i =0; v=0; CHKERRQ(VecSetValues(X,1,&i,&v,INSERT_VALUES)); for (i=0; i<3; i++) { v = -2; j=i+1; CHKERRQ(MatSetValues(A,1,&i,1,&j,&v,INSERT_VALUES)); CHKERRQ(MatSetValues(A,1,&j,1,&i,&v,INSERT_VALUES)); } i=0; j=3; v=2; CHKERRQ(MatSetValues(A,1,&i,1,&j,&v,INSERT_VALUES)); CHKERRQ(MatSetValues(A,1,&j,1,&i,&v,INSERT_VALUES)); CHKERRQ(MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY)); CHKERRQ(MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY)); CHKERRQ(VecAssemblyBegin(B)); CHKERRQ(VecAssemblyEnd(B)); CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\nThe Kershaw matrix:\n\n")); CHKERRQ(MatView(A,PETSC_VIEWER_STDOUT_WORLD)); /* * A Conjugate Gradient method * with ILU(0) preconditioning */ CHKERRQ(KSPCreate(comm,&ksp)); CHKERRQ(KSPSetOperators(ksp,A,A)); CHKERRQ(KSPSetType(ksp,KSPCG)); CHKERRQ(KSPSetInitialGuessNonzero(ksp,PETSC_TRUE)); /* * ILU preconditioner; * The iterative method will break down unless you comment in the SetShift * line below, or use the -pc_factor_shift_positive_definite option. * Run the code twice: once as given to see the negative pivot and the * divergence behaviour, then comment in the Shift line, or add the * command line option, and see that the pivots are all positive and * the method converges. */ CHKERRQ(KSPGetPC(ksp,&pc)); CHKERRQ(PCSetType(pc,PCICC)); /* CHKERRQ(PCFactorSetShiftType(prec,MAT_SHIFT_POSITIVE_DEFINITE)); */ CHKERRQ(KSPSetFromOptions(ksp)); CHKERRQ(KSPSetUp(ksp)); /* * Now that the factorisation is done, show the pivots; * note that the last one is negative. This in itself is not an error, * but it will make the iterative method diverge. */ CHKERRQ(PCFactorGetMatrix(pc,&M)); CHKERRQ(VecDuplicate(B,&D)); CHKERRQ(MatGetDiagonal(M,D)); CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\nPivots:\n\n")); CHKERRQ(VecView(D,0)); /* * Solve the system; * without the shift this will diverge with * an indefinite preconditioner */ CHKERRQ(KSPSolve(ksp,B,X)); CHKERRQ(KSPGetConvergedReason(ksp,&reason)); if (reason==KSP_DIVERGED_INDEFINITE_PC) { CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\nDivergence because of indefinite preconditioner;\n")); CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"Run the executable again but with -pc_factor_shift_positive_definite option.\n")); } else if (reason<0) { CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\nOther kind of divergence: this should not happen.\n")); } else { CHKERRQ(KSPGetIterationNumber(ksp,&its)); CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\nConvergence in %d iterations.\n",(int)its)); } CHKERRQ(PetscPrintf(PETSC_COMM_WORLD,"\n")); CHKERRQ(KSPDestroy(&ksp)); CHKERRQ(MatDestroy(&A)); CHKERRQ(VecDestroy(&B)); CHKERRQ(VecDestroy(&X)); CHKERRQ(VecDestroy(&D)); ierr = PetscFinalize(); return ierr; } /*TEST test: filter: sed -e "s/in 5 iterations/in 4 iterations/g" TEST*/