xref: /petsc/src/ksp/pc/impls/bddc/bddcschurs.c (revision 792fecdfe9134cce4d631112660ddd34f063bc17)
15e5bbd0aSStefano Zampini #include <petsc/private/pcbddcimpl.h>
25e5bbd0aSStefano Zampini #include <petsc/private/pcbddcprivateimpl.h>
305709791SSatish Balay #include <../src/mat/impls/dense/seq/dense.h>
408122e43SStefano Zampini #include <petscblaslapack.h>
534a97f8cSStefano Zampini 
69fbee547SJacob Faibussowitsch static inline PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*);
75ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*);
8df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC,Vec,Vec);
9df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC,Vec,Vec);
10d62866d3SStefano Zampini 
11ca92afb2SStefano Zampini /* if v2 is not present, correction is done in-place */
125cbda25cSStefano Zampini PetscErrorCode PCBDDCReuseSolversBenignAdapt(PCBDDCReuseSolvers ctx, Vec v, Vec v2, PetscBool sol, PetscBool full)
13ca92afb2SStefano Zampini {
14ca92afb2SStefano Zampini   PetscScalar    *array;
15ca92afb2SStefano Zampini   PetscScalar    *array2;
16ca92afb2SStefano Zampini 
17ca92afb2SStefano Zampini   PetscFunctionBegin;
18ca92afb2SStefano Zampini   if (!ctx->benign_n) PetscFunctionReturn(0);
195cbda25cSStefano Zampini   if (sol && full) {
205cbda25cSStefano Zampini     PetscInt n_I,size_schur;
215cbda25cSStefano Zampini 
225cbda25cSStefano Zampini     /* get sizes */
239566063dSJacob Faibussowitsch     PetscCall(MatGetSize(ctx->benign_csAIB,&size_schur,NULL));
249566063dSJacob Faibussowitsch     PetscCall(VecGetSize(v,&n_I));
255cbda25cSStefano Zampini     n_I = n_I - size_schur;
265cbda25cSStefano Zampini     /* get schur sol from array */
279566063dSJacob Faibussowitsch     PetscCall(VecGetArray(v,&array));
289566063dSJacob Faibussowitsch     PetscCall(VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I));
299566063dSJacob Faibussowitsch     PetscCall(VecRestoreArray(v,&array));
305cbda25cSStefano Zampini     /* apply interior sol correction */
319566063dSJacob Faibussowitsch     PetscCall(MatMultTranspose(ctx->benign_csAIB,ctx->benign_dummy_schur_vec,ctx->benign_corr_work));
329566063dSJacob Faibussowitsch     PetscCall(VecResetArray(ctx->benign_dummy_schur_vec));
339566063dSJacob Faibussowitsch     PetscCall(MatMultAdd(ctx->benign_AIIm1ones,ctx->benign_corr_work,v,v));
345cbda25cSStefano Zampini   }
35ca92afb2SStefano Zampini   if (v2) {
36ca92afb2SStefano Zampini     PetscInt nl;
37ca92afb2SStefano Zampini 
389566063dSJacob Faibussowitsch     PetscCall(VecGetArrayRead(v,(const PetscScalar**)&array));
399566063dSJacob Faibussowitsch     PetscCall(VecGetLocalSize(v2,&nl));
409566063dSJacob Faibussowitsch     PetscCall(VecGetArray(v2,&array2));
419566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(array2,array,nl));
42ca92afb2SStefano Zampini   } else {
439566063dSJacob Faibussowitsch     PetscCall(VecGetArray(v,&array));
44ca92afb2SStefano Zampini     array2 = array;
45ca92afb2SStefano Zampini   }
46ca92afb2SStefano Zampini   if (!sol) { /* change rhs */
47ca92afb2SStefano Zampini     PetscInt n;
48ca92afb2SStefano Zampini     for (n=0;n<ctx->benign_n;n++) {
49ca92afb2SStefano Zampini       PetscScalar    sum = 0.;
50ca92afb2SStefano Zampini       const PetscInt *cols;
51ca92afb2SStefano Zampini       PetscInt       nz,i;
52ca92afb2SStefano Zampini 
539566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz));
549566063dSJacob Faibussowitsch       PetscCall(ISGetIndices(ctx->benign_zerodiag_subs[n],&cols));
55ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) sum += array[cols[i]];
5622db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX)
5722db5ddcSStefano Zampini       sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz));
5822db5ddcSStefano Zampini #else
59ca92afb2SStefano Zampini       sum = -sum/nz;
6022db5ddcSStefano Zampini #endif
61ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) array2[cols[i]] += sum;
62ca92afb2SStefano Zampini       ctx->benign_save_vals[n] = array2[cols[nz-1]];
63ca92afb2SStefano Zampini       array2[cols[nz-1]] = sum;
649566063dSJacob Faibussowitsch       PetscCall(ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols));
65ca92afb2SStefano Zampini     }
66ca92afb2SStefano Zampini   } else {
67ca92afb2SStefano Zampini     PetscInt n;
68ca92afb2SStefano Zampini     for (n=0;n<ctx->benign_n;n++) {
69ca92afb2SStefano Zampini       PetscScalar    sum = 0.;
70ca92afb2SStefano Zampini       const PetscInt *cols;
71ca92afb2SStefano Zampini       PetscInt       nz,i;
729566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz));
739566063dSJacob Faibussowitsch       PetscCall(ISGetIndices(ctx->benign_zerodiag_subs[n],&cols));
74ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) sum += array[cols[i]];
7522db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX)
7622db5ddcSStefano Zampini       sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz));
7722db5ddcSStefano Zampini #else
78ca92afb2SStefano Zampini       sum = -sum/nz;
7922db5ddcSStefano Zampini #endif
80ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) array2[cols[i]] += sum;
81ca92afb2SStefano Zampini       array2[cols[nz-1]] = ctx->benign_save_vals[n];
829566063dSJacob Faibussowitsch       PetscCall(ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols));
83ca92afb2SStefano Zampini     }
84ca92afb2SStefano Zampini   }
85ca92afb2SStefano Zampini   if (v2) {
869566063dSJacob Faibussowitsch     PetscCall(VecRestoreArrayRead(v,(const PetscScalar**)&array));
879566063dSJacob Faibussowitsch     PetscCall(VecRestoreArray(v2,&array2));
88ca92afb2SStefano Zampini   } else {
899566063dSJacob Faibussowitsch     PetscCall(VecRestoreArray(v,&array));
90ca92afb2SStefano Zampini   }
915cbda25cSStefano Zampini   if (!sol && full) {
925cbda25cSStefano Zampini     Vec      usedv;
935cbda25cSStefano Zampini     PetscInt n_I,size_schur;
945cbda25cSStefano Zampini 
955cbda25cSStefano Zampini     /* get sizes */
969566063dSJacob Faibussowitsch     PetscCall(MatGetSize(ctx->benign_csAIB,&size_schur,NULL));
979566063dSJacob Faibussowitsch     PetscCall(VecGetSize(v,&n_I));
985cbda25cSStefano Zampini     n_I = n_I - size_schur;
995cbda25cSStefano Zampini     /* compute schur rhs correction */
1005cbda25cSStefano Zampini     if (v2) {
1015cbda25cSStefano Zampini       usedv = v2;
1025cbda25cSStefano Zampini     } else {
1035cbda25cSStefano Zampini       usedv = v;
1045cbda25cSStefano Zampini     }
1055cbda25cSStefano Zampini     /* apply schur rhs correction */
1069566063dSJacob Faibussowitsch     PetscCall(MatMultTranspose(ctx->benign_AIIm1ones,usedv,ctx->benign_corr_work));
1079566063dSJacob Faibussowitsch     PetscCall(VecGetArrayRead(usedv,(const PetscScalar**)&array));
1089566063dSJacob Faibussowitsch     PetscCall(VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I));
1099566063dSJacob Faibussowitsch     PetscCall(VecRestoreArrayRead(usedv,(const PetscScalar**)&array));
1109566063dSJacob Faibussowitsch     PetscCall(MatMultAdd(ctx->benign_csAIB,ctx->benign_corr_work,ctx->benign_dummy_schur_vec,ctx->benign_dummy_schur_vec));
1119566063dSJacob Faibussowitsch     PetscCall(VecResetArray(ctx->benign_dummy_schur_vec));
1125cbda25cSStefano Zampini   }
113ca92afb2SStefano Zampini   PetscFunctionReturn(0);
114ca92afb2SStefano Zampini }
115ca92afb2SStefano Zampini 
116df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Solve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose, PetscBool full)
117d62866d3SStefano Zampini {
118df4d28bfSStefano Zampini   PCBDDCReuseSolvers ctx;
119683d3df6SStefano Zampini   PetscBool          copy = PETSC_FALSE;
120d62866d3SStefano Zampini 
121d62866d3SStefano Zampini   PetscFunctionBegin;
1229566063dSJacob Faibussowitsch   PetscCall(PCShellGetContext(pc,&ctx));
123683d3df6SStefano Zampini   if (full) {
124d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
1259566063dSJacob Faibussowitsch     PetscCall(MatMumpsSetIcntl(ctx->F,26,-1));
126d62866d3SStefano Zampini #endif
1275cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
1289566063dSJacob Faibussowitsch     PetscCall(MatMkl_PardisoSetCntl(ctx->F,70,0));
1295cbda25cSStefano Zampini #endif
130683d3df6SStefano Zampini     copy = ctx->has_vertices;
131d4933d67SStefano Zampini   } else { /* interior solver */
1326dba178dSStefano Zampini #if defined(PETSC_HAVE_MUMPS)
1339566063dSJacob Faibussowitsch     PetscCall(MatMumpsSetIcntl(ctx->F,26,0));
1346dba178dSStefano Zampini #endif
135d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
1369566063dSJacob Faibussowitsch     PetscCall(MatMkl_PardisoSetCntl(ctx->F,70,1));
137d4933d67SStefano Zampini #endif
138683d3df6SStefano Zampini     copy = PETSC_TRUE;
139683d3df6SStefano Zampini   }
140683d3df6SStefano Zampini   /* copy rhs into factored matrix workspace */
141683d3df6SStefano Zampini   if (copy) {
142ca92afb2SStefano Zampini     PetscInt    n;
143df4d28bfSStefano Zampini     PetscScalar *array,*array_solver;
144ca92afb2SStefano Zampini 
1459566063dSJacob Faibussowitsch     PetscCall(VecGetLocalSize(rhs,&n));
1469566063dSJacob Faibussowitsch     PetscCall(VecGetArrayRead(rhs,(const PetscScalar**)&array));
1479566063dSJacob Faibussowitsch     PetscCall(VecGetArray(ctx->rhs,&array_solver));
1489566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(array_solver,array,n));
1499566063dSJacob Faibussowitsch     PetscCall(VecRestoreArray(ctx->rhs,&array_solver));
1509566063dSJacob Faibussowitsch     PetscCall(VecRestoreArrayRead(rhs,(const PetscScalar**)&array));
151683d3df6SStefano Zampini 
1529566063dSJacob Faibussowitsch     PetscCall(PCBDDCReuseSolversBenignAdapt(ctx,ctx->rhs,NULL,PETSC_FALSE,full));
153683d3df6SStefano Zampini     if (transpose) {
1549566063dSJacob Faibussowitsch       PetscCall(MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol));
155683d3df6SStefano Zampini     } else {
1569566063dSJacob Faibussowitsch       PetscCall(MatSolve(ctx->F,ctx->rhs,ctx->sol));
157683d3df6SStefano Zampini     }
1589566063dSJacob Faibussowitsch     PetscCall(PCBDDCReuseSolversBenignAdapt(ctx,ctx->sol,NULL,PETSC_TRUE,full));
159683d3df6SStefano Zampini 
160683d3df6SStefano Zampini     /* get back data to caller worskpace */
1619566063dSJacob Faibussowitsch     PetscCall(VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_solver));
1629566063dSJacob Faibussowitsch     PetscCall(VecGetArray(sol,&array));
1639566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(array,array_solver,n));
1649566063dSJacob Faibussowitsch     PetscCall(VecRestoreArray(sol,&array));
1659566063dSJacob Faibussowitsch     PetscCall(VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_solver));
166683d3df6SStefano Zampini   } else {
167ca92afb2SStefano Zampini     if (ctx->benign_n) {
1689566063dSJacob Faibussowitsch       PetscCall(PCBDDCReuseSolversBenignAdapt(ctx,rhs,ctx->rhs,PETSC_FALSE,full));
169ca92afb2SStefano Zampini       if (transpose) {
1709566063dSJacob Faibussowitsch         PetscCall(MatSolveTranspose(ctx->F,ctx->rhs,sol));
171ca92afb2SStefano Zampini       } else {
1729566063dSJacob Faibussowitsch         PetscCall(MatSolve(ctx->F,ctx->rhs,sol));
173ca92afb2SStefano Zampini       }
1749566063dSJacob Faibussowitsch       PetscCall(PCBDDCReuseSolversBenignAdapt(ctx,sol,NULL,PETSC_TRUE,full));
175ca92afb2SStefano Zampini     } else {
176e28d306cSStefano Zampini       if (transpose) {
1779566063dSJacob Faibussowitsch         PetscCall(MatSolveTranspose(ctx->F,rhs,sol));
178e28d306cSStefano Zampini       } else {
1799566063dSJacob Faibussowitsch         PetscCall(MatSolve(ctx->F,rhs,sol));
180e28d306cSStefano Zampini       }
181683d3df6SStefano Zampini     }
182ca92afb2SStefano Zampini   }
1835cbda25cSStefano Zampini   /* restore defaults */
1845cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS)
1859566063dSJacob Faibussowitsch   PetscCall(MatMumpsSetIcntl(ctx->F,26,-1));
1865cbda25cSStefano Zampini #endif
187d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
1889566063dSJacob Faibussowitsch   PetscCall(MatMkl_PardisoSetCntl(ctx->F,70,0));
189d4933d67SStefano Zampini #endif
190d62866d3SStefano Zampini   PetscFunctionReturn(0);
191d62866d3SStefano Zampini }
192d62866d3SStefano Zampini 
193df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC pc, Vec rhs, Vec sol)
194e28d306cSStefano Zampini {
195e28d306cSStefano Zampini   PetscFunctionBegin;
1969566063dSJacob Faibussowitsch   PetscCall(PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_TRUE));
197e28d306cSStefano Zampini   PetscFunctionReturn(0);
198e28d306cSStefano Zampini }
199e28d306cSStefano Zampini 
200df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_CorrectionTranspose(PC pc, Vec rhs, Vec sol)
201e28d306cSStefano Zampini {
202e28d306cSStefano Zampini   PetscFunctionBegin;
2039566063dSJacob Faibussowitsch   PetscCall(PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_TRUE));
204683d3df6SStefano Zampini   PetscFunctionReturn(0);
205683d3df6SStefano Zampini }
206683d3df6SStefano Zampini 
207df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC pc, Vec rhs, Vec sol)
208683d3df6SStefano Zampini {
209683d3df6SStefano Zampini   PetscFunctionBegin;
2109566063dSJacob Faibussowitsch   PetscCall(PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_FALSE));
211683d3df6SStefano Zampini   PetscFunctionReturn(0);
212683d3df6SStefano Zampini }
213683d3df6SStefano Zampini 
214df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_InteriorTranspose(PC pc, Vec rhs, Vec sol)
215683d3df6SStefano Zampini {
216683d3df6SStefano Zampini   PetscFunctionBegin;
2179566063dSJacob Faibussowitsch   PetscCall(PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_FALSE));
218e28d306cSStefano Zampini   PetscFunctionReturn(0);
219e28d306cSStefano Zampini }
220e28d306cSStefano Zampini 
22115579a77SStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_View(PC pc, PetscViewer viewer)
22215579a77SStefano Zampini {
22315579a77SStefano Zampini   PCBDDCReuseSolvers ctx;
22415579a77SStefano Zampini   PetscBool          iascii;
22515579a77SStefano Zampini 
22615579a77SStefano Zampini   PetscFunctionBegin;
2279566063dSJacob Faibussowitsch   PetscCall(PCShellGetContext(pc,&ctx));
2289566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii));
2291baa6e33SBarry Smith   if (iascii) PetscCall(PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_INFO));
2309566063dSJacob Faibussowitsch   PetscCall(MatView(ctx->F,viewer));
2311baa6e33SBarry Smith   if (iascii) PetscCall(PetscViewerPopFormat(viewer));
23215579a77SStefano Zampini   PetscFunctionReturn(0);
23315579a77SStefano Zampini }
23415579a77SStefano Zampini 
235df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolversReset(PCBDDCReuseSolvers reuse)
236d62866d3SStefano Zampini {
237ca92afb2SStefano Zampini   PetscInt       i;
238d62866d3SStefano Zampini 
239d62866d3SStefano Zampini   PetscFunctionBegin;
2409566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&reuse->F));
2419566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->sol));
2429566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->rhs));
2439566063dSJacob Faibussowitsch   PetscCall(PCDestroy(&reuse->interior_solver));
2449566063dSJacob Faibussowitsch   PetscCall(PCDestroy(&reuse->correction_solver));
2459566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&reuse->is_R));
2469566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&reuse->is_B));
2479566063dSJacob Faibussowitsch   PetscCall(VecScatterDestroy(&reuse->correction_scatter_B));
2489566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->sol_B));
2499566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->rhs_B));
250ca92afb2SStefano Zampini   for (i=0;i<reuse->benign_n;i++) {
2519566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&reuse->benign_zerodiag_subs[i]));
252ca92afb2SStefano Zampini   }
2539566063dSJacob Faibussowitsch   PetscCall(PetscFree(reuse->benign_zerodiag_subs));
2549566063dSJacob Faibussowitsch   PetscCall(PetscFree(reuse->benign_save_vals));
2559566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&reuse->benign_csAIB));
2569566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&reuse->benign_AIIm1ones));
2579566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->benign_corr_work));
2589566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&reuse->benign_dummy_schur_vec));
259d62866d3SStefano Zampini   PetscFunctionReturn(0);
260d62866d3SStefano Zampini }
261d5574798SStefano Zampini 
26232fe681dSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Destroy(PC pc)
26332fe681dSStefano Zampini {
26432fe681dSStefano Zampini   PCBDDCReuseSolvers ctx;
26532fe681dSStefano Zampini 
26632fe681dSStefano Zampini   PetscFunctionBegin;
26732fe681dSStefano Zampini   PetscCall(PCShellGetContext(pc,&ctx));
26832fe681dSStefano Zampini   PetscCall(PCBDDCReuseSolversReset(ctx));
26932fe681dSStefano Zampini   PetscCall(PetscFree(ctx));
27032fe681dSStefano Zampini   PetscCall(PCShellSetContext(pc,NULL));
27132fe681dSStefano Zampini   PetscFunctionReturn(0);
27232fe681dSStefano Zampini }
27332fe681dSStefano Zampini 
2745ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S)
2753202ece2SStefano Zampini {
2763202ece2SStefano Zampini   Mat            B, C, D, Bd, Cd, AinvBd;
2773202ece2SStefano Zampini   KSP            ksp;
2783202ece2SStefano Zampini   PC             pc;
2793202ece2SStefano Zampini   PetscBool      isLU, isILU, isCHOL, Bdense, Cdense;
2803202ece2SStefano Zampini   PetscReal      fill = 2.0;
281f11841e3SStefano Zampini   PetscInt       n_I;
2823202ece2SStefano Zampini   PetscMPIInt    size;
2833202ece2SStefano Zampini 
2843202ece2SStefano Zampini   PetscFunctionBegin;
2859566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)M),&size));
2867827d75bSBarry Smith   PetscCheck(size == 1,PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices");
287f11841e3SStefano Zampini   if (reuse == MAT_REUSE_MATRIX) {
288f11841e3SStefano Zampini     PetscBool Sdense;
289f11841e3SStefano Zampini 
2909566063dSJacob Faibussowitsch     PetscCall(PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense));
29128b400f6SJacob Faibussowitsch     PetscCheck(Sdense,PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should dense");
292f11841e3SStefano Zampini   }
2939566063dSJacob Faibussowitsch   PetscCall(MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D));
2949566063dSJacob Faibussowitsch   PetscCall(MatSchurComplementGetKSP(M, &ksp));
2959566063dSJacob Faibussowitsch   PetscCall(KSPGetPC(ksp, &pc));
2969566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU));
2979566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU));
2989566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL));
2999566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense));
3009566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense));
3019566063dSJacob Faibussowitsch   PetscCall(MatGetSize(B,&n_I,NULL));
302f11841e3SStefano Zampini   if (n_I) {
3033202ece2SStefano Zampini     if (!Bdense) {
3049566063dSJacob Faibussowitsch       PetscCall(MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd));
3053202ece2SStefano Zampini     } else {
3063202ece2SStefano Zampini       Bd = B;
3073202ece2SStefano Zampini     }
3083202ece2SStefano Zampini 
3093202ece2SStefano Zampini     if (isLU || isILU || isCHOL) {
3103202ece2SStefano Zampini       Mat fact;
3119566063dSJacob Faibussowitsch       PetscCall(KSPSetUp(ksp));
3129566063dSJacob Faibussowitsch       PetscCall(PCFactorGetMatrix(pc, &fact));
3139566063dSJacob Faibussowitsch       PetscCall(MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd));
3149566063dSJacob Faibussowitsch       PetscCall(MatMatSolve(fact, Bd, AinvBd));
3153202ece2SStefano Zampini     } else {
31607b1e237SStefano Zampini       PetscBool ex = PETSC_TRUE;
31707b1e237SStefano Zampini 
31807b1e237SStefano Zampini       if (ex) {
3193202ece2SStefano Zampini         Mat Ainvd;
3203202ece2SStefano Zampini 
3219566063dSJacob Faibussowitsch         PetscCall(PCComputeOperator(pc, MATDENSE, &Ainvd));
3229566063dSJacob Faibussowitsch         PetscCall(MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd));
3239566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&Ainvd));
32407b1e237SStefano Zampini       } else {
32507b1e237SStefano Zampini         Vec         sol,rhs;
32607b1e237SStefano Zampini         PetscScalar *arrayrhs,*arraysol;
32707b1e237SStefano Zampini         PetscInt    i,nrhs,n;
32807b1e237SStefano Zampini 
3299566063dSJacob Faibussowitsch         PetscCall(MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd));
3309566063dSJacob Faibussowitsch         PetscCall(MatGetSize(Bd,&n,&nrhs));
3319566063dSJacob Faibussowitsch         PetscCall(MatDenseGetArray(Bd,&arrayrhs));
3329566063dSJacob Faibussowitsch         PetscCall(MatDenseGetArray(AinvBd,&arraysol));
3339566063dSJacob Faibussowitsch         PetscCall(KSPGetSolution(ksp,&sol));
3349566063dSJacob Faibussowitsch         PetscCall(KSPGetRhs(ksp,&rhs));
33507b1e237SStefano Zampini         for (i=0;i<nrhs;i++) {
3369566063dSJacob Faibussowitsch           PetscCall(VecPlaceArray(rhs,arrayrhs+i*n));
3379566063dSJacob Faibussowitsch           PetscCall(VecPlaceArray(sol,arraysol+i*n));
3389566063dSJacob Faibussowitsch           PetscCall(KSPSolve(ksp,rhs,sol));
3399566063dSJacob Faibussowitsch           PetscCall(VecResetArray(rhs));
3409566063dSJacob Faibussowitsch           PetscCall(VecResetArray(sol));
34107b1e237SStefano Zampini         }
3429566063dSJacob Faibussowitsch         PetscCall(MatDenseRestoreArray(Bd,&arrayrhs));
3439566063dSJacob Faibussowitsch         PetscCall(MatDenseRestoreArray(AinvBd,&arrayrhs));
34407b1e237SStefano Zampini       }
3453202ece2SStefano Zampini     }
3465ec10c6aSStefano Zampini     if (!Bdense & !issym) {
3479566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&Bd));
3483202ece2SStefano Zampini     }
3495ec10c6aSStefano Zampini 
3505ec10c6aSStefano Zampini     if (!issym) {
3513202ece2SStefano Zampini       if (!Cdense) {
3529566063dSJacob Faibussowitsch         PetscCall(MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd));
3533202ece2SStefano Zampini       } else {
3543202ece2SStefano Zampini         Cd = C;
3553202ece2SStefano Zampini       }
3569566063dSJacob Faibussowitsch       PetscCall(MatMatMult(Cd, AinvBd, reuse, fill, S));
3573202ece2SStefano Zampini       if (!Cdense) {
3589566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&Cd));
3593202ece2SStefano Zampini       }
3605ec10c6aSStefano Zampini     } else {
3619566063dSJacob Faibussowitsch       PetscCall(MatTransposeMatMult(Bd, AinvBd, reuse, fill, S));
3625ec10c6aSStefano Zampini       if (!Bdense) {
3639566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&Bd));
3645ec10c6aSStefano Zampini       }
3655ec10c6aSStefano Zampini     }
3669566063dSJacob Faibussowitsch     PetscCall(MatDestroy(&AinvBd));
367f11841e3SStefano Zampini   }
3683202ece2SStefano Zampini 
3693202ece2SStefano Zampini   if (D) {
3703202ece2SStefano Zampini     Mat       Dd;
3713202ece2SStefano Zampini     PetscBool Ddense;
3723202ece2SStefano Zampini 
3739566063dSJacob Faibussowitsch     PetscCall(PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense));
3743202ece2SStefano Zampini     if (!Ddense) {
3759566063dSJacob Faibussowitsch       PetscCall(MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd));
3763202ece2SStefano Zampini     } else {
3773202ece2SStefano Zampini       Dd = D;
3783202ece2SStefano Zampini     }
379f11841e3SStefano Zampini     if (n_I) {
3809566063dSJacob Faibussowitsch       PetscCall(MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN));
381f11841e3SStefano Zampini     } else {
382f11841e3SStefano Zampini       if (reuse == MAT_INITIAL_MATRIX) {
3839566063dSJacob Faibussowitsch         PetscCall(MatDuplicate(Dd,MAT_COPY_VALUES,S));
384f11841e3SStefano Zampini       } else {
3859566063dSJacob Faibussowitsch         PetscCall(MatCopy(Dd,*S,SAME_NONZERO_PATTERN));
386f11841e3SStefano Zampini       }
387f11841e3SStefano Zampini     }
3883202ece2SStefano Zampini     if (!Ddense) {
3899566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&Dd));
3903202ece2SStefano Zampini     }
3913202ece2SStefano Zampini   } else {
3929566063dSJacob Faibussowitsch     PetscCall(MatScale(*S,-1.0));
3933202ece2SStefano Zampini   }
3943202ece2SStefano Zampini   PetscFunctionReturn(0);
3953202ece2SStefano Zampini }
39634a97f8cSStefano Zampini 
39791af6908SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscBool exact_schur, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, Vec scaling, PetscBool compute_Stilda, PetscBool reuse_solvers, PetscBool benign_trick, PetscInt benign_n, PetscInt benign_p0_lidx[], IS benign_zerodiag_subs[], Mat change, IS change_primal)
398b1b3d7a2SStefano Zampini {
399be83ff47SStefano Zampini   Mat                    F,A_II,A_IB,A_BI,A_BB,AE_II;
400be83ff47SStefano Zampini   Mat                    S_all;
40157a87bf3SStefano Zampini   Vec                    gstash,lstash;
40257a87bf3SStefano Zampini   VecScatter             sstash;
403b7ab4a40SStefano Zampini   IS                     is_I,is_I_layer;
404dc456d91SStefano Zampini   IS                     all_subsets,all_subsets_mult,all_subsets_n;
40557a87bf3SStefano Zampini   PetscScalar            *stasharray,*Bwork;
406dc456d91SStefano Zampini   PetscInt               *nnz,*all_local_idx_N;
407dc456d91SStefano Zampini   PetscInt               *auxnum1,*auxnum2;
4085a95e1ceSStefano Zampini   PetscInt               i,subset_size,max_subset_size;
409683d3df6SStefano Zampini   PetscInt               n_B,extra,local_size,global_size;
41057a87bf3SStefano Zampini   PetscInt               local_stash_size;
41108122e43SStefano Zampini   PetscBLASInt           B_N,B_ierr,B_lwork,*pivots;
4125a95e1ceSStefano Zampini   MPI_Comm               comm_n;
413f4f7d9d6SStefano Zampini   PetscBool              deluxe = PETSC_TRUE;
414f4f7d9d6SStefano Zampini   PetscBool              use_potr = PETSC_FALSE, use_sytr = PETSC_FALSE;
4153b03f7bbSStefano Zampini   PetscViewer            matl_dbg_viewer = NULL;
41635d0533cSStefano Zampini   PetscBool              flg;
417b1b3d7a2SStefano Zampini 
418b1b3d7a2SStefano Zampini   PetscFunctionBegin;
4199566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->A));
4209566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->S));
421e62b6521Sstefano_zampini   if (Ain) {
4229566063dSJacob Faibussowitsch     PetscCall(PetscObjectReference((PetscObject)Ain));
423a64f4aa4SStefano Zampini     sub_schurs->A = Ain;
424a64f4aa4SStefano Zampini   }
4253301b35fSStefano Zampini 
4269566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)Sin));
427a64f4aa4SStefano Zampini   sub_schurs->S = Sin;
428df4d28bfSStefano Zampini   if (sub_schurs->schur_explicit) {
429df4d28bfSStefano Zampini     sub_schurs->schur_explicit = (PetscBool)(!!sub_schurs->A);
430a64f4aa4SStefano Zampini   }
431a64f4aa4SStefano Zampini 
4325a95e1ceSStefano Zampini   /* preliminary checks */
4337827d75bSBarry Smith   PetscCheck(sub_schurs->schur_explicit || !compute_Stilda,PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_PARDISO");
4345a95e1ceSStefano Zampini 
43588113c35SStefano Zampini   if (benign_trick) sub_schurs->is_posdef = PETSC_FALSE;
43688113c35SStefano Zampini 
4373b03f7bbSStefano Zampini   /* debug (MATLAB) */
4387f9db97bSStefano Zampini   if (sub_schurs->debug) {
4397f9db97bSStefano Zampini     PetscMPIInt size,rank;
4407ebab0bbSStefano Zampini     PetscInt    nr,*print_schurs_ranks,print_schurs = PETSC_FALSE;
4417f9db97bSStefano Zampini 
4429566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&size));
4439566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank));
4447f9db97bSStefano Zampini     nr   = size;
4459566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nr,&print_schurs_ranks));
446d0609cedSBarry Smith     PetscOptionsBegin(PetscObjectComm((PetscObject)sub_schurs->l2gmap),sub_schurs->prefix,"BDDC sub_schurs options","PC");
4479566063dSJacob Faibussowitsch     PetscCall(PetscOptionsIntArray("-sub_schurs_debug_ranks","Ranks to debug (all if the option is not used)",NULL,print_schurs_ranks,&nr,&flg));
4487f9db97bSStefano Zampini     if (!flg) print_schurs = PETSC_TRUE;
4497f9db97bSStefano Zampini     else {
4507ebab0bbSStefano Zampini       print_schurs = PETSC_FALSE;
4517f9db97bSStefano Zampini       for (i=0;i<nr;i++) if (print_schurs_ranks[i] == (PetscInt)rank) { print_schurs = PETSC_TRUE; break; }
4527f9db97bSStefano Zampini     }
453d0609cedSBarry Smith     PetscOptionsEnd();
4549566063dSJacob Faibussowitsch     PetscCall(PetscFree(print_schurs_ranks));
4553b03f7bbSStefano Zampini     if (print_schurs) {
4563b03f7bbSStefano Zampini       char filename[256];
4573b03f7bbSStefano Zampini 
4589566063dSJacob Faibussowitsch       PetscCall(PetscSNPrintf(filename,sizeof(filename),"sub_schurs_Schur_r%d.m",PetscGlobalRank));
4599566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&matl_dbg_viewer));
4609566063dSJacob Faibussowitsch       PetscCall(PetscViewerPushFormat(matl_dbg_viewer,PETSC_VIEWER_ASCII_MATLAB));
4613b03f7bbSStefano Zampini     }
4627f9db97bSStefano Zampini   }
4637f9db97bSStefano Zampini 
4645a95e1ceSStefano Zampini   /* restrict work on active processes */
465991c41b4SStefano Zampini   if (sub_schurs->restrict_comm) {
466991c41b4SStefano Zampini     PetscSubcomm subcomm;
467991c41b4SStefano Zampini     PetscMPIInt  color,rank;
468991c41b4SStefano Zampini 
4695a95e1ceSStefano Zampini     color = 0;
4705a95e1ceSStefano Zampini     if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */
4719566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank));
4729566063dSJacob Faibussowitsch     PetscCall(PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm));
4739566063dSJacob Faibussowitsch     PetscCall(PetscSubcommSetNumber(subcomm,2));
4749566063dSJacob Faibussowitsch     PetscCall(PetscSubcommSetTypeGeneral(subcomm,color,rank));
4759566063dSJacob Faibussowitsch     PetscCall(PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL));
4769566063dSJacob Faibussowitsch     PetscCall(PetscSubcommDestroy(&subcomm));
4775a95e1ceSStefano Zampini     if (!sub_schurs->n_subs) {
4789566063dSJacob Faibussowitsch       PetscCall(PetscCommDestroy(&comm_n));
4795a95e1ceSStefano Zampini       PetscFunctionReturn(0);
4805a95e1ceSStefano Zampini     }
481991c41b4SStefano Zampini   } else {
4829566063dSJacob Faibussowitsch     PetscCall(PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL));
483991c41b4SStefano Zampini   }
4845a95e1ceSStefano Zampini 
485b1b3d7a2SStefano Zampini   /* get Schur complement matrices */
486df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) {
487a64f4aa4SStefano Zampini     Mat       tA_IB,tA_BI,tA_BB;
4883301b35fSStefano Zampini     PetscBool isseqsbaij;
4899566063dSJacob Faibussowitsch     PetscCall(MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB));
4909566063dSJacob Faibussowitsch     PetscCall(PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij));
4913301b35fSStefano Zampini     if (isseqsbaij) {
4929566063dSJacob Faibussowitsch       PetscCall(MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB));
4939566063dSJacob Faibussowitsch       PetscCall(MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB));
4949566063dSJacob Faibussowitsch       PetscCall(MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI));
495a64f4aa4SStefano Zampini     } else {
4969566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)tA_BB));
497a64f4aa4SStefano Zampini       A_BB = tA_BB;
4989566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)tA_IB));
499a64f4aa4SStefano Zampini       A_IB = tA_IB;
5009566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)tA_BI));
501a64f4aa4SStefano Zampini       A_BI = tA_BI;
502f11841e3SStefano Zampini     }
503a58a30b4SStefano Zampini   } else {
5045a95e1ceSStefano Zampini     A_II = NULL;
5055a95e1ceSStefano Zampini     A_IB = NULL;
5065a95e1ceSStefano Zampini     A_BI = NULL;
5075a95e1ceSStefano Zampini     A_BB = NULL;
508b1b3d7a2SStefano Zampini   }
5095a95e1ceSStefano Zampini   S_all = NULL;
510b1b3d7a2SStefano Zampini 
511b1b3d7a2SStefano Zampini   /* determine interior problems */
5129566063dSJacob Faibussowitsch   PetscCall(ISGetLocalSize(sub_schurs->is_I,&i));
5133dc780c3SStefano Zampini   if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */
514b1b3d7a2SStefano Zampini     PetscBT                touched;
515b1b3d7a2SStefano Zampini     const PetscInt*        idx_B;
516b1b3d7a2SStefano Zampini     PetscInt               n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering;
517b1b3d7a2SStefano Zampini 
51828b400f6SJacob Faibussowitsch     PetscCheck(xadj,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency");
519b1b3d7a2SStefano Zampini     /* get sizes */
5209566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_I,&n_I));
5219566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_B,&n_B));
522b1b3d7a2SStefano Zampini 
5239566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(n_I+n_B,&local_numbering));
5249566063dSJacob Faibussowitsch     PetscCall(PetscBTCreate(n_I+n_B,&touched));
5259566063dSJacob Faibussowitsch     PetscCall(PetscBTMemzero(n_I+n_B,touched));
526b1b3d7a2SStefano Zampini 
527b1b3d7a2SStefano Zampini     /* all boundary dofs must be skipped when adding layers */
5289566063dSJacob Faibussowitsch     PetscCall(ISGetIndices(sub_schurs->is_B,&idx_B));
529b1b3d7a2SStefano Zampini     for (j=0;j<n_B;j++) {
5309566063dSJacob Faibussowitsch       PetscCall(PetscBTSet(touched,idx_B[j]));
531b1b3d7a2SStefano Zampini     }
5329566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(local_numbering,idx_B,n_B));
5339566063dSJacob Faibussowitsch     PetscCall(ISRestoreIndices(sub_schurs->is_B,&idx_B));
534b1b3d7a2SStefano Zampini 
535b1b3d7a2SStefano Zampini     /* add prescribed number of layers of dofs */
536b1b3d7a2SStefano Zampini     n_local_dofs = n_B;
537b1b3d7a2SStefano Zampini     n_prev_added = n_B;
538b1b3d7a2SStefano Zampini     for (layer=0;layer<nlayers;layer++) {
539b6bace71SJacob Faibussowitsch       PetscInt n_added = 0;
540b1b3d7a2SStefano Zampini       if (n_local_dofs == n_I+n_B) break;
54163a3b9bcSJacob Faibussowitsch       PetscCheck(n_local_dofs <= n_I+n_B,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %" PetscInt_FMT ". Out of bound access (%" PetscInt_FMT " > %" PetscInt_FMT ")",layer,n_local_dofs,n_I+n_B);
5429566063dSJacob Faibussowitsch       PetscCall(PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added));
543b1b3d7a2SStefano Zampini       n_prev_added = n_added;
544b1b3d7a2SStefano Zampini       n_local_dofs += n_added;
545b1b3d7a2SStefano Zampini       if (!n_added) break;
546b1b3d7a2SStefano Zampini     }
5479566063dSJacob Faibussowitsch     PetscCall(PetscBTDestroy(&touched));
548b1b3d7a2SStefano Zampini 
549883469d8SStefano Zampini     /* IS for I layer dofs in original numbering */
5509566063dSJacob Faibussowitsch     PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&is_I_layer));
5519566063dSJacob Faibussowitsch     PetscCall(PetscFree(local_numbering));
5529566063dSJacob Faibussowitsch     PetscCall(ISSort(is_I_layer));
553883469d8SStefano Zampini     /* IS for I layer dofs in I numbering */
554df4d28bfSStefano Zampini     if (!sub_schurs->schur_explicit) {
555b1b3d7a2SStefano Zampini       ISLocalToGlobalMapping ItoNmap;
5569566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap));
5579566063dSJacob Faibussowitsch       PetscCall(ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I));
5589566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingDestroy(&ItoNmap));
559b1b3d7a2SStefano Zampini 
560b1b3d7a2SStefano Zampini       /* II block */
5619566063dSJacob Faibussowitsch       PetscCall(MatCreateSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II));
562b1b3d7a2SStefano Zampini     }
563b1b3d7a2SStefano Zampini   } else {
564b1b3d7a2SStefano Zampini     PetscInt n_I;
565b1b3d7a2SStefano Zampini 
566b1b3d7a2SStefano Zampini     /* IS for I dofs in original numbering */
5679566063dSJacob Faibussowitsch     PetscCall(PetscObjectReference((PetscObject)sub_schurs->is_I));
568a9b99552SStefano Zampini     is_I_layer = sub_schurs->is_I;
569b1b3d7a2SStefano Zampini 
570b1b3d7a2SStefano Zampini     /* IS for I dofs in I numbering (strided 1) */
571df4d28bfSStefano Zampini     if (!sub_schurs->schur_explicit) {
5729566063dSJacob Faibussowitsch       PetscCall(ISGetSize(sub_schurs->is_I,&n_I));
5739566063dSJacob Faibussowitsch       PetscCall(ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I));
574b1b3d7a2SStefano Zampini 
575b1b3d7a2SStefano Zampini       /* II block is the same */
5769566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)A_II));
577b1b3d7a2SStefano Zampini       AE_II = A_II;
578b1b3d7a2SStefano Zampini     }
579b1b3d7a2SStefano Zampini   }
5805a95e1ceSStefano Zampini 
581883469d8SStefano Zampini   /* Get info on subset sizes and sum of all subsets sizes */
5825a95e1ceSStefano Zampini   max_subset_size = 0;
583883469d8SStefano Zampini   local_size = 0;
5845a95e1ceSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
5859566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
5865a95e1ceSStefano Zampini     max_subset_size = PetscMax(subset_size,max_subset_size);
587883469d8SStefano Zampini     local_size += subset_size;
588883469d8SStefano Zampini   }
589883469d8SStefano Zampini 
590883469d8SStefano Zampini   /* Work arrays for local indices */
591883469d8SStefano Zampini   extra = 0;
5929566063dSJacob Faibussowitsch   PetscCall(ISGetLocalSize(sub_schurs->is_B,&n_B));
593df4d28bfSStefano Zampini   if (sub_schurs->schur_explicit && is_I_layer) {
5949566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(is_I_layer,&extra));
595883469d8SStefano Zampini   }
5969566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(n_B+extra,&all_local_idx_N));
597883469d8SStefano Zampini   if (extra) {
598883469d8SStefano Zampini     const PetscInt *idxs;
5999566063dSJacob Faibussowitsch     PetscCall(ISGetIndices(is_I_layer,&idxs));
6009566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(all_local_idx_N,idxs,extra));
6019566063dSJacob Faibussowitsch     PetscCall(ISRestoreIndices(is_I_layer,&idxs));
602883469d8SStefano Zampini   }
6039566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(sub_schurs->n_subs,&auxnum1));
6049566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(sub_schurs->n_subs,&auxnum2));
605883469d8SStefano Zampini 
606883469d8SStefano Zampini   /* Get local indices in local numbering */
607883469d8SStefano Zampini   local_size = 0;
60857a87bf3SStefano Zampini   local_stash_size = 0;
6095a95e1ceSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
610883469d8SStefano Zampini     const PetscInt *idxs;
611883469d8SStefano Zampini 
6129566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
6139566063dSJacob Faibussowitsch     PetscCall(ISGetIndices(sub_schurs->is_subs[i],&idxs));
614eb595f79SStefano Zampini     /* start (smallest in global ordering) and multiplicity */
615eb595f79SStefano Zampini     auxnum1[i] = idxs[0];
61657a87bf3SStefano Zampini     auxnum2[i] = subset_size*subset_size;
617883469d8SStefano Zampini     /* subset indices in local numbering */
6189566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(all_local_idx_N+local_size+extra,idxs,subset_size));
6199566063dSJacob Faibussowitsch     PetscCall(ISRestoreIndices(sub_schurs->is_subs[i],&idxs));
620883469d8SStefano Zampini     local_size += subset_size;
62157a87bf3SStefano Zampini     local_stash_size += subset_size*subset_size;
622883469d8SStefano Zampini   }
623883469d8SStefano Zampini 
624f4f7d9d6SStefano Zampini   /* allocate extra workspace needed only for GETRI or SYTRF */
62511955456SStefano Zampini   use_potr = use_sytr = PETSC_FALSE;
62611955456SStefano Zampini   if (benign_trick || (sub_schurs->is_hermitian && sub_schurs->is_posdef)) {
627f4f7d9d6SStefano Zampini     use_potr = PETSC_TRUE;
62811955456SStefano Zampini   } else if (sub_schurs->is_symmetric) {
62911955456SStefano Zampini     use_sytr = PETSC_TRUE;
63011955456SStefano Zampini   }
63111955456SStefano Zampini   if (local_size && !use_potr) {
63259ac4de7SStefano Zampini     PetscScalar  lwork,dummyscalar = 0.;
63359ac4de7SStefano Zampini     PetscBLASInt dummyint = 0;
634d2627357SStefano Zampini 
635d2627357SStefano Zampini     B_lwork = -1;
6369566063dSJacob Faibussowitsch     PetscCall(PetscBLASIntCast(local_size,&B_N));
6379566063dSJacob Faibussowitsch     PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
638f4f7d9d6SStefano Zampini     if (use_sytr) {
639*792fecdfSBarry Smith       PetscCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,&dummyscalar,&B_N,&dummyint,&lwork,&B_lwork,&B_ierr));
64028b400f6SJacob Faibussowitsch       PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYTRF Lapack routine %d",(int)B_ierr);
641f4f7d9d6SStefano Zampini     } else {
642*792fecdfSBarry Smith       PetscCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,&dummyscalar,&B_N,&dummyint,&lwork,&B_lwork,&B_ierr));
64328b400f6SJacob Faibussowitsch       PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr);
644f4f7d9d6SStefano Zampini     }
6459566063dSJacob Faibussowitsch     PetscCall(PetscFPTrapPop());
6469566063dSJacob Faibussowitsch     PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork));
6479566063dSJacob Faibussowitsch     PetscCall(PetscMalloc2(B_lwork,&Bwork,B_N,&pivots));
648056290a2SStefano Zampini   } else {
649056290a2SStefano Zampini     Bwork = NULL;
650056290a2SStefano Zampini     pivots = NULL;
651d2627357SStefano Zampini   }
652d2627357SStefano Zampini 
65357a87bf3SStefano Zampini   /* prepare data for summing up properly schurs on subsets */
6549566063dSJacob Faibussowitsch   PetscCall(ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n));
6559566063dSJacob Faibussowitsch   PetscCall(ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets));
6569566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&all_subsets_n));
6579566063dSJacob Faibussowitsch   PetscCall(ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult));
6589566063dSJacob Faibussowitsch   PetscCall(ISRenumber(all_subsets,all_subsets_mult,&global_size,&all_subsets_n));
6599566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&all_subsets));
6609566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&all_subsets_mult));
6619566063dSJacob Faibussowitsch   PetscCall(ISGetLocalSize(all_subsets_n,&i));
66263a3b9bcSJacob Faibussowitsch   PetscCheck(i == local_stash_size,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %" PetscInt_FMT " != %" PetscInt_FMT,i,local_stash_size);
6639566063dSJacob Faibussowitsch   PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF,1,local_stash_size,NULL,&lstash));
6649566063dSJacob Faibussowitsch   PetscCall(VecCreateMPI(comm_n,PETSC_DECIDE,global_size,&gstash));
6659566063dSJacob Faibussowitsch   PetscCall(VecScatterCreate(lstash,NULL,gstash,all_subsets_n,&sstash));
6669566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&all_subsets_n));
6672972d61bSStefano Zampini 
6685a95e1ceSStefano Zampini   /* subset indices in local boundary numbering */
6695a95e1ceSStefano Zampini   if (!sub_schurs->is_Ej_all) {
6705a95e1ceSStefano Zampini     PetscInt *all_local_idx_B;
6715a95e1ceSStefano Zampini 
6729566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(local_size,&all_local_idx_B));
6739566063dSJacob Faibussowitsch     PetscCall(ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B));
67463a3b9bcSJacob Faibussowitsch     PetscCheck(subset_size == local_size,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %" PetscInt_FMT " != %" PetscInt_FMT,subset_size,local_size);
6759566063dSJacob Faibussowitsch     PetscCall(ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all));
676b1b3d7a2SStefano Zampini   }
677b1b3d7a2SStefano Zampini 
67872b8c272SStefano Zampini   if (change) {
67972b8c272SStefano Zampini     ISLocalToGlobalMapping BtoS;
68072b8c272SStefano Zampini     IS                     change_primal_B;
68172b8c272SStefano Zampini     IS                     change_primal_all;
68272b8c272SStefano Zampini 
68328b400f6SJacob Faibussowitsch     PetscCheck(!sub_schurs->change_primal_sub,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
68428b400f6SJacob Faibussowitsch     PetscCheck(!sub_schurs->change,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
6859566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change_primal_sub));
68672b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
68772b8c272SStefano Zampini       ISLocalToGlobalMapping NtoS;
6889566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingCreateIS(sub_schurs->is_subs[i],&NtoS));
6899566063dSJacob Faibussowitsch       PetscCall(ISGlobalToLocalMappingApplyIS(NtoS,IS_GTOLM_DROP,change_primal,&sub_schurs->change_primal_sub[i]));
6909566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingDestroy(&NtoS));
69172b8c272SStefano Zampini     }
6929566063dSJacob Faibussowitsch     PetscCall(ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,change_primal,&change_primal_B));
6939566063dSJacob Faibussowitsch     PetscCall(ISLocalToGlobalMappingCreateIS(sub_schurs->is_Ej_all,&BtoS));
6949566063dSJacob Faibussowitsch     PetscCall(ISGlobalToLocalMappingApplyIS(BtoS,IS_GTOLM_DROP,change_primal_B,&change_primal_all));
6959566063dSJacob Faibussowitsch     PetscCall(ISLocalToGlobalMappingDestroy(&BtoS));
6969566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&change_primal_B));
6979566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change));
69872b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
69972b8c272SStefano Zampini       Mat change_sub;
70072b8c272SStefano Zampini 
7019566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
7029566063dSJacob Faibussowitsch       PetscCall(KSPCreate(PETSC_COMM_SELF,&sub_schurs->change[i]));
7039566063dSJacob Faibussowitsch       PetscCall(KSPSetType(sub_schurs->change[i],KSPPREONLY));
70472b8c272SStefano Zampini       if (!sub_schurs->change_with_qr) {
7059566063dSJacob Faibussowitsch         PetscCall(MatCreateSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_sub));
70672b8c272SStefano Zampini       } else {
70772b8c272SStefano Zampini         Mat change_subt;
7089566063dSJacob Faibussowitsch         PetscCall(MatCreateSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_subt));
7099566063dSJacob Faibussowitsch         PetscCall(MatConvert(change_subt,MATSEQDENSE,MAT_INITIAL_MATRIX,&change_sub));
7109566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&change_subt));
71172b8c272SStefano Zampini       }
7129566063dSJacob Faibussowitsch       PetscCall(KSPSetOperators(sub_schurs->change[i],change_sub,change_sub));
7139566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&change_sub));
7149566063dSJacob Faibussowitsch       PetscCall(KSPSetOptionsPrefix(sub_schurs->change[i],sub_schurs->prefix));
7159566063dSJacob Faibussowitsch       PetscCall(KSPAppendOptionsPrefix(sub_schurs->change[i],"sub_schurs_change_"));
71672b8c272SStefano Zampini     }
7179566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&change_primal_all));
71872b8c272SStefano Zampini   }
71972b8c272SStefano Zampini 
7205a95e1ceSStefano Zampini   /* Local matrix of all local Schur on subsets (transposed) */
7215a95e1ceSStefano Zampini   if (!sub_schurs->S_Ej_all) {
72204c5b2e6SStefano Zampini     Mat         T;
72304c5b2e6SStefano Zampini     PetscScalar *v;
72404c5b2e6SStefano Zampini     PetscInt    *ii,*jj;
72504c5b2e6SStefano Zampini     PetscInt    cum,i,j,k;
72604c5b2e6SStefano Zampini 
72704c5b2e6SStefano Zampini     /* MatSeqAIJSetPreallocation + MatSetValues is slow for these kind of matrices (may have large blocks)
72804c5b2e6SStefano Zampini        Allocate properly a representative matrix and duplicate */
7299566063dSJacob Faibussowitsch     PetscCall(PetscMalloc3(local_size+1,&ii,local_stash_size,&jj,local_stash_size,&v));
73004c5b2e6SStefano Zampini     ii[0] = 0;
73104c5b2e6SStefano Zampini     cum   = 0;
73204c5b2e6SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
7339566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
73404c5b2e6SStefano Zampini       for (j=0;j<subset_size;j++) {
73504c5b2e6SStefano Zampini         const PetscInt row = cum+j;
73604c5b2e6SStefano Zampini         PetscInt col = cum;
73704c5b2e6SStefano Zampini 
73804c5b2e6SStefano Zampini         ii[row+1] = ii[row] + subset_size;
73904c5b2e6SStefano Zampini         for (k=ii[row];k<ii[row+1];k++) {
74004c5b2e6SStefano Zampini           jj[k] = col;
74104c5b2e6SStefano Zampini           col++;
74204c5b2e6SStefano Zampini         }
74304c5b2e6SStefano Zampini       }
74404c5b2e6SStefano Zampini       cum += subset_size;
74504c5b2e6SStefano Zampini     }
7469566063dSJacob Faibussowitsch     PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,local_size,local_size,ii,jj,v,&T));
7479566063dSJacob Faibussowitsch     PetscCall(MatDuplicate(T,MAT_DO_NOT_COPY_VALUES,&sub_schurs->S_Ej_all));
7489566063dSJacob Faibussowitsch     PetscCall(MatDestroy(&T));
7499566063dSJacob Faibussowitsch     PetscCall(PetscFree3(ii,jj,v));
75004c5b2e6SStefano Zampini   }
75104c5b2e6SStefano Zampini   /* matrices for deluxe scaling and adaptive selection */
75204c5b2e6SStefano Zampini   if (compute_Stilda) {
75304c5b2e6SStefano Zampini     if (!sub_schurs->sum_S_Ej_tilda_all) {
7549566063dSJacob Faibussowitsch       PetscCall(MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_tilda_all));
75504c5b2e6SStefano Zampini     }
75604c5b2e6SStefano Zampini     if (!sub_schurs->sum_S_Ej_inv_all && deluxe) {
7579566063dSJacob Faibussowitsch       PetscCall(MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_inv_all));
75804c5b2e6SStefano Zampini     }
759aa83b6aeSStefano Zampini   }
760b1b3d7a2SStefano Zampini 
7615a95e1ceSStefano Zampini   /* Compute Schur complements explicitly */
762be83ff47SStefano Zampini   F = NULL;
763d943a642SStefano Zampini   if (!sub_schurs->schur_explicit) {
764d943a642SStefano Zampini     /* this code branch is used when MatFactor with Schur complement support is not present or when explicitly requested;
765d943a642SStefano Zampini        it is not efficient, unless the economic version of the scaling is used */
7665a95e1ceSStefano Zampini     Mat         S_Ej_expl;
7675a95e1ceSStefano Zampini     PetscScalar *work;
7685a95e1ceSStefano Zampini     PetscInt    j,*dummy_idx;
7695a95e1ceSStefano Zampini     PetscBool   Sdense;
7705a95e1ceSStefano Zampini 
7719566063dSJacob Faibussowitsch     PetscCall(PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work));
7725a95e1ceSStefano Zampini     local_size = 0;
773b1b3d7a2SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
7745a95e1ceSStefano Zampini       IS  is_subset_B;
7755a95e1ceSStefano Zampini       Mat AE_EE,AE_IE,AE_EI,S_Ej;
7765a95e1ceSStefano Zampini 
7775a95e1ceSStefano Zampini       /* subsets in original and boundary numbering */
7789566063dSJacob Faibussowitsch       PetscCall(ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B));
7795a95e1ceSStefano Zampini       /* EE block */
7809566063dSJacob Faibussowitsch       PetscCall(MatCreateSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE));
7815a95e1ceSStefano Zampini       /* IE block */
7829566063dSJacob Faibussowitsch       PetscCall(MatCreateSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE));
7835a95e1ceSStefano Zampini       /* EI block */
784d943a642SStefano Zampini       if (sub_schurs->is_symmetric) {
7859566063dSJacob Faibussowitsch         PetscCall(MatCreateTranspose(AE_IE,&AE_EI));
786d943a642SStefano Zampini       } else if (sub_schurs->is_hermitian) {
7879566063dSJacob Faibussowitsch         PetscCall(MatCreateHermitianTranspose(AE_IE,&AE_EI));
7885a95e1ceSStefano Zampini       } else {
7899566063dSJacob Faibussowitsch         PetscCall(MatCreateSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI));
7905a95e1ceSStefano Zampini       }
7919566063dSJacob Faibussowitsch       PetscCall(ISDestroy(&is_subset_B));
7929566063dSJacob Faibussowitsch       PetscCall(MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej));
7939566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&AE_EE));
7949566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&AE_IE));
7959566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&AE_EI));
796b1b3d7a2SStefano Zampini       if (AE_II == A_II) { /* we can reuse the same ksp */
797b1b3d7a2SStefano Zampini         KSP ksp;
7989566063dSJacob Faibussowitsch         PetscCall(MatSchurComplementGetKSP(sub_schurs->S,&ksp));
7999566063dSJacob Faibussowitsch         PetscCall(MatSchurComplementSetKSP(S_Ej,ksp));
800b1b3d7a2SStefano Zampini       } else { /* build new ksp object which inherits ksp and pc types from the original one */
801b1b3d7a2SStefano Zampini         KSP       origksp,schurksp;
802b1b3d7a2SStefano Zampini         PC        origpc,schurpc;
803b1b3d7a2SStefano Zampini         KSPType   ksp_type;
804b1b3d7a2SStefano Zampini         PetscInt  n_internal;
8055a95e1ceSStefano Zampini         PetscBool ispcnone;
806b1b3d7a2SStefano Zampini 
8079566063dSJacob Faibussowitsch         PetscCall(MatSchurComplementGetKSP(sub_schurs->S,&origksp));
8089566063dSJacob Faibussowitsch         PetscCall(MatSchurComplementGetKSP(S_Ej,&schurksp));
8099566063dSJacob Faibussowitsch         PetscCall(KSPGetType(origksp,&ksp_type));
8109566063dSJacob Faibussowitsch         PetscCall(KSPSetType(schurksp,ksp_type));
8119566063dSJacob Faibussowitsch         PetscCall(KSPGetPC(schurksp,&schurpc));
8129566063dSJacob Faibussowitsch         PetscCall(KSPGetPC(origksp,&origpc));
8139566063dSJacob Faibussowitsch         PetscCall(PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone));
8145a95e1ceSStefano Zampini         if (!ispcnone) {
8155a95e1ceSStefano Zampini           PCType pc_type;
8169566063dSJacob Faibussowitsch           PetscCall(PCGetType(origpc,&pc_type));
8179566063dSJacob Faibussowitsch           PetscCall(PCSetType(schurpc,pc_type));
8185a95e1ceSStefano Zampini         } else {
8199566063dSJacob Faibussowitsch           PetscCall(PCSetType(schurpc,PCLU));
8205a95e1ceSStefano Zampini         }
8219566063dSJacob Faibussowitsch         PetscCall(ISGetSize(is_I,&n_internal));
822365a3a41SStefano Zampini         if (!n_internal) { /* UMFPACK gives error with 0 sized problems */
8233ca39a21SBarry Smith           MatSolverType solver = NULL;
8249566063dSJacob Faibussowitsch           PetscCall(PCFactorGetMatSolverType(origpc,(MatSolverType*)&solver));
8251baa6e33SBarry Smith           if (solver) PetscCall(PCFactorSetMatSolverType(schurpc,solver));
826b1b3d7a2SStefano Zampini         }
8279566063dSJacob Faibussowitsch         PetscCall(KSPSetUp(schurksp));
828b1b3d7a2SStefano Zampini       }
8299566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
8309566063dSJacob Faibussowitsch       PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl));
8319566063dSJacob Faibussowitsch       PetscCall(PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_symmetric,MAT_REUSE_MATRIX,&S_Ej_expl));
8329566063dSJacob Faibussowitsch       PetscCall(PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense));
8335a95e1ceSStefano Zampini       if (Sdense) {
8345a95e1ceSStefano Zampini         for (j=0;j<subset_size;j++) {
8355a95e1ceSStefano Zampini           dummy_idx[j]=local_size+j;
836b1b3d7a2SStefano Zampini         }
8379566063dSJacob Faibussowitsch         PetscCall(MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES));
8386c4ed002SBarry Smith       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices");
8399566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&S_Ej));
8409566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&S_Ej_expl));
8415a95e1ceSStefano Zampini       local_size += subset_size;
8425a95e1ceSStefano Zampini     }
8439566063dSJacob Faibussowitsch     PetscCall(PetscFree2(dummy_idx,work));
844b1b3d7a2SStefano Zampini     /* free */
8459566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&is_I));
8469566063dSJacob Faibussowitsch     PetscCall(MatDestroy(&AE_II));
8479566063dSJacob Faibussowitsch     PetscCall(PetscFree(all_local_idx_N));
848883469d8SStefano Zampini   } else {
8495cbda25cSStefano Zampini     Mat               A,cs_AIB_mat = NULL,benign_AIIm1_ones_mat = NULL;
85032fe681dSStefano Zampini     Mat               *gdswA;
8519d54b7f4SStefano Zampini     Vec               Dall = NULL;
852ca92afb2SStefano Zampini     IS                is_A_all,*is_p_r = NULL;
8537ebab0bbSStefano Zampini     MatType           Stype;
8545cbda25cSStefano Zampini     PetscScalar       *work,*S_data,*schur_factor,infty = PETSC_MAX_REAL;
85504c5b2e6SStefano Zampini     PetscScalar       *SEj_arr = NULL,*SEjinv_arr = NULL;
8561683a169SBarry Smith     const PetscScalar *rS_data;
85704c5b2e6SStefano Zampini     PetscInt          n,n_I,size_schur,size_active_schur,cum,cum2;
8583fc34f97SStefano Zampini     PetscBool         economic,solver_S,S_lower_triangular = PETSC_FALSE;
8593fc34f97SStefano Zampini     PetscBool         schur_has_vertices,factor_workaround;
86011955456SStefano Zampini     PetscBool         use_cholesky;
8617ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
8627ebab0bbSStefano Zampini     PetscBool         oldpin;
8637ebab0bbSStefano Zampini #endif
864883469d8SStefano Zampini 
865683d3df6SStefano Zampini     /* get sizes */
86681ea8064SStefano Zampini     n_I = 0;
86781ea8064SStefano Zampini     if (is_I_layer) {
8689566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(is_I_layer,&n_I));
86981ea8064SStefano Zampini     }
870683d3df6SStefano Zampini     economic = PETSC_FALSE;
8719566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_I,&cum));
872683d3df6SStefano Zampini     if (cum != n_I) economic = PETSC_TRUE;
8739566063dSJacob Faibussowitsch     PetscCall(MatGetLocalSize(sub_schurs->A,&n,NULL));
8749d54b7f4SStefano Zampini     size_active_schur = local_size;
8759d54b7f4SStefano Zampini 
876f17d2ae1SStefano Zampini     /* import scaling vector (wrong formulation if we have 3D edges) */
8779d54b7f4SStefano Zampini     if (scaling && compute_Stilda) {
8789d54b7f4SStefano Zampini       const PetscScalar *array;
8799d54b7f4SStefano Zampini       PetscScalar       *array2;
8809d54b7f4SStefano Zampini       const PetscInt    *idxs;
8819d54b7f4SStefano Zampini       PetscInt          i;
8829d54b7f4SStefano Zampini 
8839566063dSJacob Faibussowitsch       PetscCall(ISGetIndices(sub_schurs->is_Ej_all,&idxs));
8849566063dSJacob Faibussowitsch       PetscCall(VecCreateSeq(PETSC_COMM_SELF,size_active_schur,&Dall));
8859566063dSJacob Faibussowitsch       PetscCall(VecGetArrayRead(scaling,&array));
8869566063dSJacob Faibussowitsch       PetscCall(VecGetArray(Dall,&array2));
8879d54b7f4SStefano Zampini       for (i=0;i<size_active_schur;i++) array2[i] = array[idxs[i]];
8889566063dSJacob Faibussowitsch       PetscCall(VecRestoreArray(Dall,&array2));
8899566063dSJacob Faibussowitsch       PetscCall(VecRestoreArrayRead(scaling,&array));
8909566063dSJacob Faibussowitsch       PetscCall(ISRestoreIndices(sub_schurs->is_Ej_all,&idxs));
8919d54b7f4SStefano Zampini       deluxe = PETSC_FALSE;
8929d54b7f4SStefano Zampini     }
893d62866d3SStefano Zampini 
894683d3df6SStefano Zampini     /* size active schurs does not count any dirichlet or vertex dof on the interface */
8953fc34f97SStefano Zampini     factor_workaround = PETSC_FALSE;
8963fc34f97SStefano Zampini     schur_has_vertices = PETSC_FALSE;
897683d3df6SStefano Zampini     cum = n_I+size_active_schur;
898683d3df6SStefano Zampini     if (sub_schurs->is_dir) {
899683d3df6SStefano Zampini       const PetscInt* idxs;
900683d3df6SStefano Zampini       PetscInt        n_dir;
901683d3df6SStefano Zampini 
9029566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_dir,&n_dir));
9039566063dSJacob Faibussowitsch       PetscCall(ISGetIndices(sub_schurs->is_dir,&idxs));
9049566063dSJacob Faibussowitsch       PetscCall(PetscArraycpy(all_local_idx_N+cum,idxs,n_dir));
9059566063dSJacob Faibussowitsch       PetscCall(ISRestoreIndices(sub_schurs->is_dir,&idxs));
906683d3df6SStefano Zampini       cum += n_dir;
90732fe681dSStefano Zampini       if (!sub_schurs->gdsw) factor_workaround = PETSC_TRUE;
908d62866d3SStefano Zampini     }
909683d3df6SStefano Zampini     /* include the primal vertices in the Schur complement */
910367aa537SStefano Zampini     if (exact_schur && sub_schurs->is_vertices && (compute_Stilda || benign_n)) {
911683d3df6SStefano Zampini       PetscInt n_v;
912683d3df6SStefano Zampini 
9139566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_vertices,&n_v));
914683d3df6SStefano Zampini       if (n_v) {
915683d3df6SStefano Zampini         const PetscInt* idxs;
916683d3df6SStefano Zampini 
9179566063dSJacob Faibussowitsch         PetscCall(ISGetIndices(sub_schurs->is_vertices,&idxs));
9189566063dSJacob Faibussowitsch         PetscCall(PetscArraycpy(all_local_idx_N+cum,idxs,n_v));
9199566063dSJacob Faibussowitsch         PetscCall(ISRestoreIndices(sub_schurs->is_vertices,&idxs));
920683d3df6SStefano Zampini         cum += n_v;
92132fe681dSStefano Zampini         if (!sub_schurs->gdsw) factor_workaround = PETSC_TRUE;
9223fc34f97SStefano Zampini         schur_has_vertices = PETSC_TRUE;
923683d3df6SStefano Zampini       }
924683d3df6SStefano Zampini     }
925683d3df6SStefano Zampini     size_schur = cum - n_I;
9269566063dSJacob Faibussowitsch     PetscCall(ISCreateGeneral(PETSC_COMM_SELF,cum,all_local_idx_N,PETSC_OWN_POINTER,&is_A_all));
9277ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
928b470e4b4SRichard Tran Mills     oldpin = sub_schurs->A->boundtocpu;
9299566063dSJacob Faibussowitsch     PetscCall(MatBindToCPU(sub_schurs->A,PETSC_TRUE));
9307ebab0bbSStefano Zampini #endif
931683d3df6SStefano Zampini     if (cum == n) {
9329566063dSJacob Faibussowitsch       PetscCall(ISSetPermutation(is_A_all));
9339566063dSJacob Faibussowitsch       PetscCall(MatPermute(sub_schurs->A,is_A_all,is_A_all,&A));
934683d3df6SStefano Zampini     } else {
9359566063dSJacob Faibussowitsch       PetscCall(MatCreateSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A));
936683d3df6SStefano Zampini     }
9377ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
9389566063dSJacob Faibussowitsch     PetscCall(MatBindToCPU(sub_schurs->A,oldpin));
9397ebab0bbSStefano Zampini #endif
94026cc229bSBarry Smith     PetscCall(MatSetOptionsPrefixFactor(A,sub_schurs->prefix));
94126cc229bSBarry Smith     PetscCall(MatAppendOptionsPrefixFactor(A,"sub_schurs_"));
942ca92afb2SStefano Zampini 
943ca92afb2SStefano Zampini     /* if we actually change the basis for the pressures, LDL^T factors will use a lot of memory
9447ebab0bbSStefano Zampini        this is a workaround */
945ca92afb2SStefano Zampini     if (benign_n) {
9467ebab0bbSStefano Zampini       Vec                    v,benign_AIIm1_ones;
947ca92afb2SStefano Zampini       ISLocalToGlobalMapping N_to_reor;
948ca92afb2SStefano Zampini       IS                     is_p0,is_p0_p;
9495cbda25cSStefano Zampini       PetscScalar            *cs_AIB,*AIIm1_data;
9505cbda25cSStefano Zampini       PetscInt               sizeA;
951ca92afb2SStefano Zampini 
9529566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingCreateIS(is_A_all,&N_to_reor));
9539566063dSJacob Faibussowitsch       PetscCall(ISCreateGeneral(PETSC_COMM_SELF,benign_n,benign_p0_lidx,PETSC_COPY_VALUES,&is_p0));
9549566063dSJacob Faibussowitsch       PetscCall(ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,is_p0,&is_p0_p));
9559566063dSJacob Faibussowitsch       PetscCall(ISDestroy(&is_p0));
9569566063dSJacob Faibussowitsch       PetscCall(MatCreateVecs(A,&v,&benign_AIIm1_ones));
9579566063dSJacob Faibussowitsch       PetscCall(VecGetSize(v,&sizeA));
9589566063dSJacob Faibussowitsch       PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,sizeA,benign_n,NULL,&benign_AIIm1_ones_mat));
9599566063dSJacob Faibussowitsch       PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,size_schur,benign_n,NULL,&cs_AIB_mat));
9609566063dSJacob Faibussowitsch       PetscCall(MatDenseGetArray(cs_AIB_mat,&cs_AIB));
9619566063dSJacob Faibussowitsch       PetscCall(MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data));
9629566063dSJacob Faibussowitsch       PetscCall(PetscMalloc1(benign_n,&is_p_r));
963ca92afb2SStefano Zampini       /* compute colsum of A_IB restricted to pressures */
964ca92afb2SStefano Zampini       for (i=0;i<benign_n;i++) {
9657ebab0bbSStefano Zampini         const PetscScalar *array;
966ca92afb2SStefano Zampini         const PetscInt    *idxs;
967ca92afb2SStefano Zampini         PetscInt          j,nz;
968ca92afb2SStefano Zampini 
9699566063dSJacob Faibussowitsch         PetscCall(ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,benign_zerodiag_subs[i],&is_p_r[i]));
9709566063dSJacob Faibussowitsch         PetscCall(ISGetLocalSize(is_p_r[i],&nz));
9719566063dSJacob Faibussowitsch         PetscCall(ISGetIndices(is_p_r[i],&idxs));
9725cbda25cSStefano Zampini         for (j=0;j<nz;j++) AIIm1_data[idxs[j]+sizeA*i] = 1.;
9739566063dSJacob Faibussowitsch         PetscCall(ISRestoreIndices(is_p_r[i],&idxs));
9749566063dSJacob Faibussowitsch         PetscCall(VecPlaceArray(benign_AIIm1_ones,AIIm1_data+sizeA*i));
9759566063dSJacob Faibussowitsch         PetscCall(MatMult(A,benign_AIIm1_ones,v));
9769566063dSJacob Faibussowitsch         PetscCall(VecResetArray(benign_AIIm1_ones));
9779566063dSJacob Faibussowitsch         PetscCall(VecGetArrayRead(v,&array));
97822db5ddcSStefano Zampini         for (j=0;j<size_schur;j++) {
97922db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX)
98022db5ddcSStefano Zampini           cs_AIB[i*size_schur + j] = (PetscRealPart(array[j+n_I])/nz + PETSC_i*(PetscImaginaryPart(array[j+n_I])/nz));
98122db5ddcSStefano Zampini #else
98222db5ddcSStefano Zampini           cs_AIB[i*size_schur + j] = array[j+n_I]/nz;
98322db5ddcSStefano Zampini #endif
98422db5ddcSStefano Zampini         }
9859566063dSJacob Faibussowitsch         PetscCall(VecRestoreArrayRead(v,&array));
986ca92afb2SStefano Zampini       }
9879566063dSJacob Faibussowitsch       PetscCall(MatDenseRestoreArray(cs_AIB_mat,&cs_AIB));
9889566063dSJacob Faibussowitsch       PetscCall(MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data));
9899566063dSJacob Faibussowitsch       PetscCall(VecDestroy(&v));
9909566063dSJacob Faibussowitsch       PetscCall(VecDestroy(&benign_AIIm1_ones));
9919566063dSJacob Faibussowitsch       PetscCall(MatSetOption(A,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE));
9929566063dSJacob Faibussowitsch       PetscCall(MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE));
9939566063dSJacob Faibussowitsch       PetscCall(MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE));
9949566063dSJacob Faibussowitsch       PetscCall(MatZeroRowsColumnsIS(A,is_p0_p,1.0,NULL,NULL));
9959566063dSJacob Faibussowitsch       PetscCall(ISDestroy(&is_p0_p));
9969566063dSJacob Faibussowitsch       PetscCall(ISLocalToGlobalMappingDestroy(&N_to_reor));
997ca92afb2SStefano Zampini     }
9989566063dSJacob Faibussowitsch     PetscCall(MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_symmetric));
9999566063dSJacob Faibussowitsch     PetscCall(MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian));
10009566063dSJacob Faibussowitsch     PetscCall(MatSetOption(A,MAT_SPD,sub_schurs->is_posdef));
1001883469d8SStefano Zampini 
100211955456SStefano Zampini     /* for complexes, symmetric and hermitian at the same time implies null imaginary part */
100311955456SStefano Zampini     use_cholesky = (PetscBool)((use_potr || use_sytr) && sub_schurs->is_hermitian && sub_schurs->is_symmetric);
100411955456SStefano Zampini 
1005683d3df6SStefano Zampini     /* when using the benign subspace trick, the local Schur complements are SPD */
100635d0533cSStefano Zampini     /* MKL_PARDISO does not handle well the computation of a Schur complement from a symmetric indefinite factorization
100735d0533cSStefano Zampini        Use LU and adapt pivoting perturbation (still, solution is not as accurate as with using MUMPS) */
100835d0533cSStefano Zampini     if (benign_trick) {
100935d0533cSStefano Zampini       sub_schurs->is_posdef = PETSC_TRUE;
10109566063dSJacob Faibussowitsch       PetscCall(PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,&flg));
101135d0533cSStefano Zampini       if (flg) use_cholesky = PETSC_FALSE;
101235d0533cSStefano Zampini     }
1013d47842beSStefano Zampini 
1014f4f7d9d6SStefano Zampini     if (n_I) {
10150aa714b2SStefano Zampini       IS        is_schur;
10167ebab0bbSStefano Zampini       char      stype[64];
10174ba54290SStefano Zampini       PetscBool gpu = PETSC_FALSE;
10185a05ddb0SStefano Zampini 
101911955456SStefano Zampini       if (use_cholesky) {
10209566063dSJacob Faibussowitsch         PetscCall(MatGetFactor(A,sub_schurs->mat_solver_type,MAT_FACTOR_CHOLESKY,&F));
1021883469d8SStefano Zampini       } else {
10229566063dSJacob Faibussowitsch         PetscCall(MatGetFactor(A,sub_schurs->mat_solver_type,MAT_FACTOR_LU,&F));
1023883469d8SStefano Zampini       }
10249566063dSJacob Faibussowitsch       PetscCall(MatSetErrorIfFailure(A,PETSC_TRUE));
102535d0533cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
10269566063dSJacob Faibussowitsch       if (benign_trick) PetscCall(MatMkl_PardisoSetCntl(F,10,10));
102735d0533cSStefano Zampini #endif
1028883469d8SStefano Zampini       /* subsets ordered last */
10299566063dSJacob Faibussowitsch       PetscCall(ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is_schur));
10309566063dSJacob Faibussowitsch       PetscCall(MatFactorSetSchurIS(F,is_schur));
10319566063dSJacob Faibussowitsch       PetscCall(ISDestroy(&is_schur));
1032883469d8SStefano Zampini 
1033883469d8SStefano Zampini       /* factorization step */
103411955456SStefano Zampini       if (use_cholesky) {
10359566063dSJacob Faibussowitsch         PetscCall(MatCholeskyFactorSymbolic(F,A,NULL,NULL));
1036be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */
10379566063dSJacob Faibussowitsch         PetscCall(MatMumpsSetIcntl(F,19,2));
1038be83ff47SStefano Zampini #endif
10399566063dSJacob Faibussowitsch         PetscCall(MatCholeskyFactorNumeric(F,A,NULL));
1040a0b0af32SStefano Zampini         S_lower_triangular = PETSC_TRUE;
1041883469d8SStefano Zampini       } else {
10429566063dSJacob Faibussowitsch         PetscCall(MatLUFactorSymbolic(F,A,NULL,NULL,NULL));
1043be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */
10449566063dSJacob Faibussowitsch         PetscCall(MatMumpsSetIcntl(F,19,3));
1045be83ff47SStefano Zampini #endif
10469566063dSJacob Faibussowitsch         PetscCall(MatLUFactorNumeric(F,A,NULL));
1047883469d8SStefano Zampini       }
10489566063dSJacob Faibussowitsch       PetscCall(MatViewFromOptions(F,(PetscObject)A,"-mat_factor_view"));
1049883469d8SStefano Zampini 
10503b03f7bbSStefano Zampini       if (matl_dbg_viewer) {
105111955456SStefano Zampini         Mat S;
105211955456SStefano Zampini         IS  is;
105311955456SStefano Zampini 
10549566063dSJacob Faibussowitsch         PetscCall(PetscObjectSetName((PetscObject)A,"A"));
10559566063dSJacob Faibussowitsch         PetscCall(MatView(A,matl_dbg_viewer));
10569566063dSJacob Faibussowitsch         PetscCall(MatFactorCreateSchurComplement(F,&S,NULL));
10579566063dSJacob Faibussowitsch         PetscCall(PetscObjectSetName((PetscObject)S,"S"));
10589566063dSJacob Faibussowitsch         PetscCall(MatView(S,matl_dbg_viewer));
10599566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&S));
10609566063dSJacob Faibussowitsch         PetscCall(ISCreateStride(PETSC_COMM_SELF,n_I,0,1,&is));
10619566063dSJacob Faibussowitsch         PetscCall(PetscObjectSetName((PetscObject)is,"I"));
10629566063dSJacob Faibussowitsch         PetscCall(ISView(is,matl_dbg_viewer));
10639566063dSJacob Faibussowitsch         PetscCall(ISDestroy(&is));
10649566063dSJacob Faibussowitsch         PetscCall(ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is));
10659566063dSJacob Faibussowitsch         PetscCall(PetscObjectSetName((PetscObject)is,"B"));
10669566063dSJacob Faibussowitsch         PetscCall(ISView(is,matl_dbg_viewer));
10679566063dSJacob Faibussowitsch         PetscCall(ISDestroy(&is));
10689566063dSJacob Faibussowitsch         PetscCall(PetscObjectSetName((PetscObject)is_A_all,"IA"));
10699566063dSJacob Faibussowitsch         PetscCall(ISView(is_A_all,matl_dbg_viewer));
107032fe681dSStefano Zampini         for (i=0,cum=0;i<sub_schurs->n_subs;i++) {
107132fe681dSStefano Zampini           IS   is;
107232fe681dSStefano Zampini           char name[16];
107332fe681dSStefano Zampini 
107432fe681dSStefano Zampini           PetscCall(PetscSNPrintf(name,sizeof(name),"IE%" PetscInt_FMT,i));
107532fe681dSStefano Zampini           PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
107632fe681dSStefano Zampini           PetscCall(ISCreateStride(PETSC_COMM_SELF,subset_size,cum,1,&is));
107732fe681dSStefano Zampini           PetscCall(PetscObjectSetName((PetscObject)is,name));
107832fe681dSStefano Zampini           PetscCall(ISView(is,matl_dbg_viewer));
107932fe681dSStefano Zampini           PetscCall(ISDestroy(&is));
108032fe681dSStefano Zampini           if (sub_schurs->change) {
108132fe681dSStefano Zampini             Mat T;
108232fe681dSStefano Zampini 
108332fe681dSStefano Zampini             PetscCall(PetscSNPrintf(name,sizeof(name),"TE%" PetscInt_FMT,i));
108432fe681dSStefano Zampini             PetscCall(KSPGetOperators(sub_schurs->change[i],&T,NULL));
108532fe681dSStefano Zampini             PetscCall(PetscObjectSetName((PetscObject)T,name));
108632fe681dSStefano Zampini             PetscCall(MatView(T,matl_dbg_viewer));
108732fe681dSStefano Zampini             PetscCall(PetscSNPrintf(name,sizeof(name),"ITE%" PetscInt_FMT,i));
108832fe681dSStefano Zampini             PetscCall(PetscObjectSetName((PetscObject)sub_schurs->change_primal_sub[i],name));
108932fe681dSStefano Zampini             PetscCall(ISView(sub_schurs->change_primal_sub[i],matl_dbg_viewer));
109032fe681dSStefano Zampini           }
109132fe681dSStefano Zampini           cum += subset_size;
109232fe681dSStefano Zampini         }
109332fe681dSStefano Zampini         PetscCall(PetscViewerFlush(matl_dbg_viewer));
109411955456SStefano Zampini       }
109511955456SStefano Zampini 
1096883469d8SStefano Zampini       /* get explicit Schur Complement computed during numeric factorization */
10979566063dSJacob Faibussowitsch       PetscCall(MatFactorGetSchurComplement(F,&S_all,NULL));
10989566063dSJacob Faibussowitsch       PetscCall(PetscStrncpy(stype,MATSEQDENSE,sizeof(stype)));
10994ba54290SStefano Zampini #if defined(PETSC_HAVE_CUDA)
11009566063dSJacob Faibussowitsch       PetscCall(PetscObjectTypeCompareAny((PetscObject)A,&gpu,MATSEQAIJVIENNACL,MATSEQAIJCUSPARSE,""));
11014ba54290SStefano Zampini #endif
11021baa6e33SBarry Smith       if (gpu) PetscCall(PetscStrncpy(stype,MATSEQDENSECUDA,sizeof(stype)));
11039566063dSJacob Faibussowitsch       PetscCall(PetscOptionsGetString(NULL,sub_schurs->prefix,"-sub_schurs_schur_mat_type",stype,sizeof(stype),NULL));
11049566063dSJacob Faibussowitsch       PetscCall(MatConvert(S_all,stype,MAT_INPLACE_MATRIX,&S_all));
11059566063dSJacob Faibussowitsch       PetscCall(MatSetOption(S_all,MAT_SPD,sub_schurs->is_posdef));
11069566063dSJacob Faibussowitsch       PetscCall(MatSetOption(S_all,MAT_HERMITIAN,sub_schurs->is_hermitian));
11079566063dSJacob Faibussowitsch       PetscCall(MatGetType(S_all,&Stype));
1108b3cb21ddSStefano Zampini 
1109d62866d3SStefano Zampini       /* we can reuse the solvers if we are not using the economic version */
1110683d3df6SStefano Zampini       reuse_solvers = (PetscBool)(reuse_solvers && !economic);
111132fe681dSStefano Zampini       if (!sub_schurs->gdsw) {
1112683d3df6SStefano Zampini         factor_workaround = (PetscBool)(reuse_solvers && factor_workaround);
111303dfb2d7SStefano Zampini         if (!sub_schurs->is_posdef && factor_workaround && compute_Stilda && size_active_schur)
111403dfb2d7SStefano Zampini           reuse_solvers = factor_workaround = PETSC_FALSE;
111532fe681dSStefano Zampini       }
1116df4d28bfSStefano Zampini       solver_S = PETSC_TRUE;
1117ca92afb2SStefano Zampini 
111872b8c272SStefano Zampini       /* update the Schur complement with the change of basis on the pressures */
1119ca92afb2SStefano Zampini       if (benign_n) {
11207ebab0bbSStefano Zampini         const PetscScalar *cs_AIB;
11217ebab0bbSStefano Zampini         PetscScalar       *S_data,*AIIm1_data;
11223b03f7bbSStefano Zampini         Mat               S2 = NULL,S3 = NULL; /* dbg */
11233b03f7bbSStefano Zampini         PetscScalar       *S2_data,*S3_data; /* dbg */
11247ebab0bbSStefano Zampini         Vec               v,benign_AIIm1_ones;
11255cbda25cSStefano Zampini         PetscInt          sizeA;
1126ca92afb2SStefano Zampini 
11279566063dSJacob Faibussowitsch         PetscCall(MatDenseGetArray(S_all,&S_data));
11289566063dSJacob Faibussowitsch         PetscCall(MatCreateVecs(A,&v,&benign_AIIm1_ones));
11299566063dSJacob Faibussowitsch         PetscCall(VecGetSize(v,&sizeA));
1130ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
11319566063dSJacob Faibussowitsch         PetscCall(MatMumpsSetIcntl(F,26,0));
1132ca92afb2SStefano Zampini #endif
1133ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
11349566063dSJacob Faibussowitsch         PetscCall(MatMkl_PardisoSetCntl(F,70,1));
1135ca92afb2SStefano Zampini #endif
11369566063dSJacob Faibussowitsch         PetscCall(MatDenseGetArrayRead(cs_AIB_mat,&cs_AIB));
11379566063dSJacob Faibussowitsch         PetscCall(MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data));
11383b03f7bbSStefano Zampini         if (matl_dbg_viewer) {
11399566063dSJacob Faibussowitsch           PetscCall(MatDuplicate(S_all,MAT_DO_NOT_COPY_VALUES,&S2));
11409566063dSJacob Faibussowitsch           PetscCall(MatDuplicate(S_all,MAT_DO_NOT_COPY_VALUES,&S3));
11419566063dSJacob Faibussowitsch           PetscCall(MatDenseGetArray(S2,&S2_data));
11429566063dSJacob Faibussowitsch           PetscCall(MatDenseGetArray(S3,&S3_data));
11433b03f7bbSStefano Zampini         }
1144ca92afb2SStefano Zampini         for (i=0;i<benign_n;i++) {
11453b03f7bbSStefano Zampini           PetscScalar    *array,sum = 0.,one = 1.,*sums;
1146ca92afb2SStefano Zampini           const PetscInt *idxs;
11473b03f7bbSStefano Zampini           PetscInt       k,j,nz;
114847484b83SStefano Zampini           PetscBLASInt   B_k,B_n;
1149ca92afb2SStefano Zampini 
11509566063dSJacob Faibussowitsch           PetscCall(PetscCalloc1(benign_n,&sums));
11519566063dSJacob Faibussowitsch           PetscCall(VecPlaceArray(benign_AIIm1_ones,AIIm1_data+sizeA*i));
11529566063dSJacob Faibussowitsch           PetscCall(VecCopy(benign_AIIm1_ones,v));
11539566063dSJacob Faibussowitsch           PetscCall(MatSolve(F,v,benign_AIIm1_ones));
11549566063dSJacob Faibussowitsch           PetscCall(MatMult(A,benign_AIIm1_ones,v));
11559566063dSJacob Faibussowitsch           PetscCall(VecResetArray(benign_AIIm1_ones));
11563b03f7bbSStefano Zampini           /* p0 dofs (eliminated) are excluded from the sums */
11573b03f7bbSStefano Zampini           for (k=0;k<benign_n;k++) {
11589566063dSJacob Faibussowitsch             PetscCall(ISGetLocalSize(is_p_r[k],&nz));
11599566063dSJacob Faibussowitsch             PetscCall(ISGetIndices(is_p_r[k],&idxs));
11603b03f7bbSStefano Zampini             for (j=0;j<nz-1;j++) sums[k] -= AIIm1_data[idxs[j]+sizeA*i];
11619566063dSJacob Faibussowitsch             PetscCall(ISRestoreIndices(is_p_r[k],&idxs));
11623b03f7bbSStefano Zampini           }
11639566063dSJacob Faibussowitsch           PetscCall(VecGetArrayRead(v,(const PetscScalar**)&array));
11643b03f7bbSStefano Zampini           if (matl_dbg_viewer) {
11653b03f7bbSStefano Zampini             Vec  vv;
11663b03f7bbSStefano Zampini             char name[16];
11673b03f7bbSStefano Zampini 
11689566063dSJacob Faibussowitsch             PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF,1,size_schur,array+n_I,&vv));
116963a3b9bcSJacob Faibussowitsch             PetscCall(PetscSNPrintf(name,sizeof(name),"Pvs%" PetscInt_FMT,i));
11709566063dSJacob Faibussowitsch             PetscCall(PetscObjectSetName((PetscObject)vv,name));
11719566063dSJacob Faibussowitsch             PetscCall(VecView(vv,matl_dbg_viewer));
11723b03f7bbSStefano Zampini           }
117347484b83SStefano Zampini           /* perform sparse rank updates on symmetric Schur (TODO: move outside of the loop?) */
117447484b83SStefano Zampini           /* cs_AIB already scaled by 1./nz */
117547484b83SStefano Zampini           B_k = 1;
11763b03f7bbSStefano Zampini           for (k=0;k<benign_n;k++) {
11773b03f7bbSStefano Zampini             sum  = sums[k];
11789566063dSJacob Faibussowitsch             PetscCall(PetscBLASIntCast(size_schur,&B_n));
11793b03f7bbSStefano Zampini 
11803b03f7bbSStefano Zampini             if (PetscAbsScalar(sum) == 0.0) continue;
11813b03f7bbSStefano Zampini             if (k == i) {
1182*792fecdfSBarry Smith               PetscCallBLAS("BLASsyrk",BLASsyrk_("L","N",&B_n,&B_k,&sum,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n));
11833b03f7bbSStefano Zampini               if (matl_dbg_viewer) {
1184*792fecdfSBarry Smith                 PetscCallBLAS("BLASsyrk",BLASsyrk_("L","N",&B_n,&B_k,&sum,cs_AIB+i*size_schur,&B_n,&one,S3_data,&B_n));
11853b03f7bbSStefano Zampini               }
11863b03f7bbSStefano Zampini             } else { /* XXX Is it correct to use symmetric rank-2 update with half of the sum? */
11873b03f7bbSStefano Zampini               sum /= 2.0;
1188*792fecdfSBarry Smith               PetscCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,cs_AIB+k*size_schur,&B_n,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n));
11893b03f7bbSStefano Zampini               if (matl_dbg_viewer) {
1190*792fecdfSBarry Smith                 PetscCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,cs_AIB+k*size_schur,&B_n,cs_AIB+i*size_schur,&B_n,&one,S3_data,&B_n));
11913b03f7bbSStefano Zampini               }
11923b03f7bbSStefano Zampini             }
11933b03f7bbSStefano Zampini           }
11945cbda25cSStefano Zampini           sum  = 1.;
1195*792fecdfSBarry Smith           PetscCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,array+n_I,&B_n,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n));
11963b03f7bbSStefano Zampini           if (matl_dbg_viewer) {
1197*792fecdfSBarry Smith             PetscCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,array+n_I,&B_n,cs_AIB+i*size_schur,&B_n,&one,S2_data,&B_n));
11983b03f7bbSStefano Zampini           }
11999566063dSJacob Faibussowitsch           PetscCall(VecRestoreArrayRead(v,(const PetscScalar**)&array));
12005cbda25cSStefano Zampini           /* set p0 entry of AIIm1_ones to zero */
12019566063dSJacob Faibussowitsch           PetscCall(ISGetLocalSize(is_p_r[i],&nz));
12029566063dSJacob Faibussowitsch           PetscCall(ISGetIndices(is_p_r[i],&idxs));
1203282d6408SStefano Zampini           for (j=0;j<benign_n;j++) AIIm1_data[idxs[nz-1]+sizeA*j] = 0.;
12049566063dSJacob Faibussowitsch           PetscCall(ISRestoreIndices(is_p_r[i],&idxs));
12059566063dSJacob Faibussowitsch           PetscCall(PetscFree(sums));
12063b03f7bbSStefano Zampini         }
12079566063dSJacob Faibussowitsch         PetscCall(VecDestroy(&benign_AIIm1_ones));
12083b03f7bbSStefano Zampini         if (matl_dbg_viewer) {
12099566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArray(S2,&S2_data));
12109566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArray(S3,&S3_data));
1211ca92afb2SStefano Zampini         }
1212a7414863SStefano Zampini         if (!S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */
1213a7414863SStefano Zampini           PetscInt k,j;
1214a7414863SStefano Zampini           for (k=0;k<size_schur;k++) {
1215a7414863SStefano Zampini             for (j=k;j<size_schur;j++) {
1216a7414863SStefano Zampini               S_data[j*size_schur+k] = PetscConj(S_data[k*size_schur+j]);
1217a7414863SStefano Zampini             }
1218a7414863SStefano Zampini           }
1219a7414863SStefano Zampini         }
1220a7414863SStefano Zampini 
12215cbda25cSStefano Zampini         /* restore defaults */
12225cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS)
12239566063dSJacob Faibussowitsch         PetscCall(MatMumpsSetIcntl(F,26,-1));
12245cbda25cSStefano Zampini #endif
12255cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
12269566063dSJacob Faibussowitsch         PetscCall(MatMkl_PardisoSetCntl(F,70,0));
12275cbda25cSStefano Zampini #endif
12289566063dSJacob Faibussowitsch         PetscCall(MatDenseRestoreArrayRead(cs_AIB_mat,&cs_AIB));
12299566063dSJacob Faibussowitsch         PetscCall(MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data));
12309566063dSJacob Faibussowitsch         PetscCall(VecDestroy(&v));
12319566063dSJacob Faibussowitsch         PetscCall(MatDenseRestoreArray(S_all,&S_data));
12323b03f7bbSStefano Zampini         if (matl_dbg_viewer) {
12333b03f7bbSStefano Zampini           Mat S;
12343b03f7bbSStefano Zampini 
12359566063dSJacob Faibussowitsch           PetscCall(MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED));
12369566063dSJacob Faibussowitsch           PetscCall(MatFactorCreateSchurComplement(F,&S,NULL));
12379566063dSJacob Faibussowitsch           PetscCall(PetscObjectSetName((PetscObject)S,"Sb"));
12389566063dSJacob Faibussowitsch           PetscCall(MatView(S,matl_dbg_viewer));
12399566063dSJacob Faibussowitsch           PetscCall(MatDestroy(&S));
12409566063dSJacob Faibussowitsch           PetscCall(PetscObjectSetName((PetscObject)S2,"S2P"));
12419566063dSJacob Faibussowitsch           PetscCall(MatView(S2,matl_dbg_viewer));
12429566063dSJacob Faibussowitsch           PetscCall(PetscObjectSetName((PetscObject)S3,"S3P"));
12439566063dSJacob Faibussowitsch           PetscCall(MatView(S3,matl_dbg_viewer));
12449566063dSJacob Faibussowitsch           PetscCall(PetscObjectSetName((PetscObject)cs_AIB_mat,"cs"));
12459566063dSJacob Faibussowitsch           PetscCall(MatView(cs_AIB_mat,matl_dbg_viewer));
12469566063dSJacob Faibussowitsch           PetscCall(MatFactorGetSchurComplement(F,&S_all,NULL));
12473b03f7bbSStefano Zampini         }
12489566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&S2));
12499566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&S3));
1250ca92afb2SStefano Zampini       }
1251a3df083aSStefano Zampini       if (!reuse_solvers) {
1252a3df083aSStefano Zampini         for (i=0;i<benign_n;i++) {
12539566063dSJacob Faibussowitsch           PetscCall(ISDestroy(&is_p_r[i]));
1254a3df083aSStefano Zampini         }
12559566063dSJacob Faibussowitsch         PetscCall(PetscFree(is_p_r));
12569566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&cs_AIB_mat));
12579566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&benign_AIIm1_ones_mat));
1258a3df083aSStefano Zampini       }
1259df4d28bfSStefano Zampini     } else { /* we can't use MatFactor when size_schur == size_of_the_problem */
12609566063dSJacob Faibussowitsch       PetscCall(MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all));
12619566063dSJacob Faibussowitsch       PetscCall(MatGetType(S_all,&Stype));
1262683d3df6SStefano Zampini       reuse_solvers = PETSC_FALSE; /* TODO: why we can't reuse the solvers here? */
1263166598c1SStefano Zampini       factor_workaround = PETSC_FALSE;
1264df4d28bfSStefano Zampini       solver_S = PETSC_FALSE;
1265be83ff47SStefano Zampini     }
1266be83ff47SStefano Zampini 
1267be83ff47SStefano Zampini     if (reuse_solvers) {
126832fe681dSStefano Zampini       Mat                A_II,pA_II,Afake;
126953892102SStefano Zampini       Vec                vec1_B;
1270df4d28bfSStefano Zampini       PCBDDCReuseSolvers msolv_ctx;
12713462e049SStefano Zampini       PetscInt           n_R;
1272d5574798SStefano Zampini 
1273df4d28bfSStefano Zampini       if (sub_schurs->reuse_solver) {
12749566063dSJacob Faibussowitsch         PetscCall(PCBDDCReuseSolversReset(sub_schurs->reuse_solver));
1275e28d306cSStefano Zampini       } else {
12769566063dSJacob Faibussowitsch         PetscCall(PetscNew(&sub_schurs->reuse_solver));
1277d62866d3SStefano Zampini       }
1278df4d28bfSStefano Zampini       msolv_ctx = sub_schurs->reuse_solver;
127932fe681dSStefano Zampini       PetscCall(MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,&pA_II,NULL,NULL,NULL));
12809566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)F));
1281d5574798SStefano Zampini       msolv_ctx->F = F;
12829566063dSJacob Faibussowitsch       PetscCall(MatCreateVecs(F,&msolv_ctx->sol,NULL));
1283683d3df6SStefano Zampini       /* currently PETSc has no support for MatSolve(F,x,x), so cheat and let rhs and sol share the same memory */
1284683d3df6SStefano Zampini       {
1285683d3df6SStefano Zampini         PetscScalar *array;
1286683d3df6SStefano Zampini         PetscInt    n;
1287683d3df6SStefano Zampini 
12889566063dSJacob Faibussowitsch         PetscCall(VecGetLocalSize(msolv_ctx->sol,&n));
12899566063dSJacob Faibussowitsch         PetscCall(VecGetArray(msolv_ctx->sol,&array));
12909566063dSJacob Faibussowitsch         PetscCall(VecCreateSeqWithArray(PetscObjectComm((PetscObject)msolv_ctx->sol),1,n,array,&msolv_ctx->rhs));
12919566063dSJacob Faibussowitsch         PetscCall(VecRestoreArray(msolv_ctx->sol,&array));
1292683d3df6SStefano Zampini       }
12933fc34f97SStefano Zampini       msolv_ctx->has_vertices = schur_has_vertices;
1294d62866d3SStefano Zampini 
1295d62866d3SStefano Zampini       /* interior solver */
12969566063dSJacob Faibussowitsch       PetscCall(PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->interior_solver));
129732fe681dSStefano Zampini       PetscCall(PCSetOperators(msolv_ctx->interior_solver,A_II,pA_II));
12989566063dSJacob Faibussowitsch       PetscCall(PCSetType(msolv_ctx->interior_solver,PCSHELL));
12999566063dSJacob Faibussowitsch       PetscCall(PCShellSetName(msolv_ctx->interior_solver,"Interior solver (w/o Schur factorization)"));
13009566063dSJacob Faibussowitsch       PetscCall(PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx));
13019566063dSJacob Faibussowitsch       PetscCall(PCShellSetView(msolv_ctx->interior_solver,PCBDDCReuseSolvers_View));
13029566063dSJacob Faibussowitsch       PetscCall(PCShellSetApply(msolv_ctx->interior_solver,PCBDDCReuseSolvers_Interior));
13039566063dSJacob Faibussowitsch       PetscCall(PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCReuseSolvers_InteriorTranspose));
130432fe681dSStefano Zampini       if (sub_schurs->gdsw) PetscCall(PCShellSetDestroy(msolv_ctx->interior_solver,PCBDDCReuseSolvers_Destroy));
1305d62866d3SStefano Zampini 
1306d62866d3SStefano Zampini       /* correction solver */
130732fe681dSStefano Zampini       if (!sub_schurs->gdsw) {
13089566063dSJacob Faibussowitsch         PetscCall(PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->correction_solver));
13099566063dSJacob Faibussowitsch         PetscCall(PCSetType(msolv_ctx->correction_solver,PCSHELL));
13109566063dSJacob Faibussowitsch         PetscCall(PCShellSetName(msolv_ctx->correction_solver,"Correction solver (with Schur factorization)"));
13119566063dSJacob Faibussowitsch         PetscCall(PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx));
13129566063dSJacob Faibussowitsch         PetscCall(PCShellSetView(msolv_ctx->interior_solver,PCBDDCReuseSolvers_View));
13139566063dSJacob Faibussowitsch         PetscCall(PCShellSetApply(msolv_ctx->correction_solver,PCBDDCReuseSolvers_Correction));
13149566063dSJacob Faibussowitsch         PetscCall(PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCReuseSolvers_CorrectionTranspose));
131553892102SStefano Zampini 
131653892102SStefano Zampini         /* scatter and vecs for Schur complement solver */
13179566063dSJacob Faibussowitsch         PetscCall(MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B));
13189566063dSJacob Faibussowitsch         PetscCall(MatCreateVecs(sub_schurs->S,&vec1_B,NULL));
13193fc34f97SStefano Zampini         if (!schur_has_vertices) {
13209566063dSJacob Faibussowitsch           PetscCall(ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B));
13219566063dSJacob Faibussowitsch           PetscCall(VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B));
13229566063dSJacob Faibussowitsch           PetscCall(PetscObjectReference((PetscObject)is_A_all));
132353892102SStefano Zampini           msolv_ctx->is_R = is_A_all;
1324683d3df6SStefano Zampini         } else {
1325683d3df6SStefano Zampini           IS              is_B_all;
1326683d3df6SStefano Zampini           const PetscInt* idxs;
1327683d3df6SStefano Zampini           PetscInt        dual,n_v,n;
1328683d3df6SStefano Zampini 
13299566063dSJacob Faibussowitsch           PetscCall(ISGetLocalSize(sub_schurs->is_vertices,&n_v));
1330683d3df6SStefano Zampini           dual = size_schur - n_v;
13319566063dSJacob Faibussowitsch           PetscCall(ISGetLocalSize(is_A_all,&n));
13329566063dSJacob Faibussowitsch           PetscCall(ISGetIndices(is_A_all,&idxs));
13339566063dSJacob Faibussowitsch           PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),dual,idxs+n_I,PETSC_COPY_VALUES,&is_B_all));
13349566063dSJacob Faibussowitsch           PetscCall(ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_B_all,&msolv_ctx->is_B));
13359566063dSJacob Faibussowitsch           PetscCall(ISDestroy(&is_B_all));
13369566063dSJacob Faibussowitsch           PetscCall(ISCreateStride(PetscObjectComm((PetscObject)is_A_all),dual,0,1,&is_B_all));
13379566063dSJacob Faibussowitsch           PetscCall(VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,is_B_all,&msolv_ctx->correction_scatter_B));
13389566063dSJacob Faibussowitsch           PetscCall(ISDestroy(&is_B_all));
13399566063dSJacob Faibussowitsch           PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),n-n_v,idxs,PETSC_COPY_VALUES,&msolv_ctx->is_R));
13409566063dSJacob Faibussowitsch           PetscCall(ISRestoreIndices(is_A_all,&idxs));
1341683d3df6SStefano Zampini         }
13429566063dSJacob Faibussowitsch         PetscCall(ISGetLocalSize(msolv_ctx->is_R,&n_R));
13439566063dSJacob Faibussowitsch         PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF,n_R,n_R,0,NULL,&Afake));
13449566063dSJacob Faibussowitsch         PetscCall(MatAssemblyBegin(Afake,MAT_FINAL_ASSEMBLY));
13459566063dSJacob Faibussowitsch         PetscCall(MatAssemblyEnd(Afake,MAT_FINAL_ASSEMBLY));
13469566063dSJacob Faibussowitsch         PetscCall(PCSetOperators(msolv_ctx->correction_solver,Afake,Afake));
13479566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&Afake));
13489566063dSJacob Faibussowitsch         PetscCall(VecDestroy(&vec1_B));
134932fe681dSStefano Zampini       }
1350ca92afb2SStefano Zampini       /* communicate benign info to solver context */
1351ca92afb2SStefano Zampini       if (benign_n) {
13525cbda25cSStefano Zampini         PetscScalar *array;
13535cbda25cSStefano Zampini 
1354ca92afb2SStefano Zampini         msolv_ctx->benign_n = benign_n;
1355ca92afb2SStefano Zampini         msolv_ctx->benign_zerodiag_subs = is_p_r;
13569566063dSJacob Faibussowitsch         PetscCall(PetscMalloc1(benign_n,&msolv_ctx->benign_save_vals));
13575cbda25cSStefano Zampini         msolv_ctx->benign_csAIB = cs_AIB_mat;
13589566063dSJacob Faibussowitsch         PetscCall(MatCreateVecs(cs_AIB_mat,&msolv_ctx->benign_corr_work,NULL));
13599566063dSJacob Faibussowitsch         PetscCall(VecGetArray(msolv_ctx->benign_corr_work,&array));
13609566063dSJacob Faibussowitsch         PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF,1,size_schur,array,&msolv_ctx->benign_dummy_schur_vec));
13619566063dSJacob Faibussowitsch         PetscCall(VecRestoreArray(msolv_ctx->benign_corr_work,&array));
13625cbda25cSStefano Zampini         msolv_ctx->benign_AIIm1ones = benign_AIIm1_ones_mat;
1363ca92afb2SStefano Zampini       }
1364ada6e2d7SStefano Zampini     } else {
13651baa6e33SBarry Smith       if (sub_schurs->reuse_solver) PetscCall(PCBDDCReuseSolversReset(sub_schurs->reuse_solver));
13669566063dSJacob Faibussowitsch       PetscCall(PetscFree(sub_schurs->reuse_solver));
1367d5574798SStefano Zampini     }
13689566063dSJacob Faibussowitsch     PetscCall(MatDestroy(&A));
13699566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&is_A_all));
13705db18549SStefano Zampini 
1371be83ff47SStefano Zampini     /* Work arrays */
13729566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(max_subset_size*max_subset_size,&work));
1373d2627357SStefano Zampini 
1374be83ff47SStefano Zampini     /* S_Ej_all */
1375be83ff47SStefano Zampini     cum = cum2 = 0;
13769566063dSJacob Faibussowitsch     PetscCall(MatDenseGetArrayRead(S_all,&rS_data));
13779566063dSJacob Faibussowitsch     PetscCall(MatSeqAIJGetArray(sub_schurs->S_Ej_all,&SEj_arr));
137832fe681dSStefano Zampini     if (sub_schurs->sum_S_Ej_inv_all) {
13799566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&SEjinv_arr));
138004c5b2e6SStefano Zampini     }
138132fe681dSStefano Zampini     if (sub_schurs->gdsw) {
138232fe681dSStefano Zampini       PetscCall(MatCreateSubMatrices(sub_schurs->A,sub_schurs->n_subs,sub_schurs->is_subs,sub_schurs->is_subs,MAT_INITIAL_MATRIX,&gdswA));
138332fe681dSStefano Zampini     }
138465d8bf0aSStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
138565d8bf0aSStefano Zampini       PetscInt j;
138665d8bf0aSStefano Zampini 
138732fe681dSStefano Zampini       /* get S_E (or K^i_EE for GDSW) */
13889566063dSJacob Faibussowitsch       PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
138932fe681dSStefano Zampini       if (sub_schurs->gdsw) {
139032fe681dSStefano Zampini         Mat T;
139132fe681dSStefano Zampini 
139232fe681dSStefano Zampini         PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&T));
139332fe681dSStefano Zampini         PetscCall(MatConvert(gdswA[i],MATDENSE,MAT_REUSE_MATRIX,&T));
139432fe681dSStefano Zampini         PetscCall(MatDestroy(&T));
139532fe681dSStefano Zampini       } else {
1396683d3df6SStefano Zampini         if (S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */
1397be83ff47SStefano Zampini           PetscInt k;
1398be83ff47SStefano Zampini           for (k=0;k<subset_size;k++) {
1399be83ff47SStefano Zampini             for (j=k;j<subset_size;j++) {
14001683a169SBarry Smith               work[k*subset_size+j] = rS_data[cum2+k*size_schur+j];
14011683a169SBarry Smith               work[j*subset_size+k] = PetscConj(rS_data[cum2+k*size_schur+j]);
1402be83ff47SStefano Zampini             }
1403be83ff47SStefano Zampini           }
140406a4e24aSStefano Zampini         } else { /* just copy to workspace */
1405be83ff47SStefano Zampini           PetscInt k;
1406be83ff47SStefano Zampini           for (k=0;k<subset_size;k++) {
1407be83ff47SStefano Zampini             for (j=0;j<subset_size;j++) {
14081683a169SBarry Smith               work[k*subset_size+j] = rS_data[cum2+k*size_schur+j];
1409be83ff47SStefano Zampini             }
1410be83ff47SStefano Zampini           }
14119087bf02SStefano Zampini         }
141232fe681dSStefano Zampini       }
14135a95e1ceSStefano Zampini       /* insert S_E values */
1414b7ab4a40SStefano Zampini       if (sub_schurs->change) {
14158760537fSStefano Zampini         Mat change_sub,SEj,T;
141672b8c272SStefano Zampini 
141772b8c272SStefano Zampini         /* change basis */
14189566063dSJacob Faibussowitsch         PetscCall(KSPGetOperators(sub_schurs->change[i],&change_sub,NULL));
14199566063dSJacob Faibussowitsch         PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj));
14208760537fSStefano Zampini         if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */
14218760537fSStefano Zampini           Mat T2;
14229566063dSJacob Faibussowitsch           PetscCall(MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2));
14239566063dSJacob Faibussowitsch           PetscCall(MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T));
14249566063dSJacob Faibussowitsch           PetscCall(MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T));
14259566063dSJacob Faibussowitsch           PetscCall(MatDestroy(&T2));
14268760537fSStefano Zampini         } else {
14279566063dSJacob Faibussowitsch           PetscCall(MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T));
142872b8c272SStefano Zampini         }
14299566063dSJacob Faibussowitsch         PetscCall(MatCopy(T,SEj,SAME_NONZERO_PATTERN));
14309566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&T));
14319566063dSJacob Faibussowitsch         PetscCall(MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],1.0,NULL,NULL));
14329566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&SEj));
143372b8c272SStefano Zampini       }
14349566063dSJacob Faibussowitsch       PetscCall(PetscArraycpy(SEj_arr,work,subset_size*subset_size));
1435862806e4SStefano Zampini       if (compute_Stilda) {
143632fe681dSStefano Zampini         if (deluxe) { /* if adaptivity is requested, invert S_E blocks */
14377ebab0bbSStefano Zampini           Mat               M;
14387ebab0bbSStefano Zampini           const PetscScalar *vals;
14397ebab0bbSStefano Zampini           PetscBool         isdense,isdensecuda;
1440f4f7d9d6SStefano Zampini 
14419566063dSJacob Faibussowitsch           PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&M));
14429566063dSJacob Faibussowitsch           PetscCall(MatSetOption(M,MAT_SPD,sub_schurs->is_posdef));
14439566063dSJacob Faibussowitsch           PetscCall(MatSetOption(M,MAT_HERMITIAN,sub_schurs->is_hermitian));
14447ebab0bbSStefano Zampini           if (!PetscBTLookup(sub_schurs->is_edge,i)) {
14459566063dSJacob Faibussowitsch             PetscCall(MatSetType(M,Stype));
14466c3e6151SStefano Zampini           }
14479566063dSJacob Faibussowitsch           PetscCall(PetscObjectTypeCompare((PetscObject)M,MATSEQDENSE,&isdense));
14489566063dSJacob Faibussowitsch           PetscCall(PetscObjectTypeCompare((PetscObject)M,MATSEQDENSECUDA,&isdensecuda));
14497ebab0bbSStefano Zampini           if (use_cholesky) {
14509566063dSJacob Faibussowitsch             PetscCall(MatCholeskyFactor(M,NULL,NULL));
1451d6462365SStefano Zampini           } else {
14529566063dSJacob Faibussowitsch             PetscCall(MatLUFactor(M,NULL,NULL,NULL));
14532972d61bSStefano Zampini           }
14547ebab0bbSStefano Zampini           if (isdense) {
14559566063dSJacob Faibussowitsch             PetscCall(MatSeqDenseInvertFactors_Private(M));
14567ebab0bbSStefano Zampini #if defined(PETSC_HAVE_CUDA)
14577ebab0bbSStefano Zampini           } else if (isdensecuda) {
14589566063dSJacob Faibussowitsch             PetscCall(MatSeqDenseCUDAInvertFactors_Private(M));
14597ebab0bbSStefano Zampini #endif
146098921bdaSJacob Faibussowitsch           } else SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"Not implemented for type %s",Stype);
14619566063dSJacob Faibussowitsch           PetscCall(MatDenseGetArrayRead(M,&vals));
14629566063dSJacob Faibussowitsch           PetscCall(PetscArraycpy(SEjinv_arr,vals,subset_size*subset_size));
14639566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArrayRead(M,&vals));
14649566063dSJacob Faibussowitsch           PetscCall(MatDestroy(&M));
146532fe681dSStefano Zampini         } else if (scaling) { /* not using deluxe */
14669d54b7f4SStefano Zampini           Mat         SEj;
14679d54b7f4SStefano Zampini           Vec         D;
14689d54b7f4SStefano Zampini           PetscScalar *array;
14699d54b7f4SStefano Zampini 
14709566063dSJacob Faibussowitsch           PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj));
14719566063dSJacob Faibussowitsch           PetscCall(VecGetArray(Dall,&array));
14729566063dSJacob Faibussowitsch           PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF,1,subset_size,array+cum,&D));
14739566063dSJacob Faibussowitsch           PetscCall(VecRestoreArray(Dall,&array));
14749566063dSJacob Faibussowitsch           PetscCall(VecShift(D,-1.));
14759566063dSJacob Faibussowitsch           PetscCall(MatDiagonalScale(SEj,D,D));
14769566063dSJacob Faibussowitsch           PetscCall(MatDestroy(&SEj));
14779566063dSJacob Faibussowitsch           PetscCall(VecDestroy(&D));
14789566063dSJacob Faibussowitsch           PetscCall(PetscArraycpy(SEj_arr,work,subset_size*subset_size));
14799d54b7f4SStefano Zampini         }
148032fe681dSStefano Zampini       }
1481be83ff47SStefano Zampini       cum += subset_size;
1482be83ff47SStefano Zampini       cum2 += subset_size*(size_schur + 1);
148304c5b2e6SStefano Zampini       SEj_arr += subset_size*subset_size;
148404c5b2e6SStefano Zampini       if (SEjinv_arr) SEjinv_arr += subset_size*subset_size;
1485be83ff47SStefano Zampini     }
148632fe681dSStefano Zampini     if (sub_schurs->gdsw) {
148732fe681dSStefano Zampini       PetscCall(MatDestroySubMatrices(sub_schurs->n_subs,&gdswA));
148832fe681dSStefano Zampini     }
14899566063dSJacob Faibussowitsch     PetscCall(MatDenseRestoreArrayRead(S_all,&rS_data));
14909566063dSJacob Faibussowitsch     PetscCall(MatSeqAIJRestoreArray(sub_schurs->S_Ej_all,&SEj_arr));
149132fe681dSStefano Zampini     if (sub_schurs->sum_S_Ej_inv_all) {
14929566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&SEjinv_arr));
149304c5b2e6SStefano Zampini     }
1494df4d28bfSStefano Zampini     if (solver_S) {
14959566063dSJacob Faibussowitsch       PetscCall(MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED));
14964a6c6b0dSStefano Zampini     }
1497683d3df6SStefano Zampini 
14987ebab0bbSStefano Zampini     /* may prevent from unneeded copies, since MUMPS or MKL_Pardiso always use CPU memory
14997ebab0bbSStefano Zampini        however, preliminary tests indicate using GPUs is still faster in the solve phase */
15007ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
15017ebab0bbSStefano Zampini     if (reuse_solvers) {
15027ebab0bbSStefano Zampini       Mat                  St;
15037ebab0bbSStefano Zampini       MatFactorSchurStatus st;
15047ebab0bbSStefano Zampini 
150535d0533cSStefano Zampini       flg  = PETSC_FALSE;
15069566063dSJacob Faibussowitsch       PetscCall(PetscOptionsGetBool(NULL,sub_schurs->prefix,"-sub_schurs_schur_pin_to_cpu",&flg,NULL));
15079566063dSJacob Faibussowitsch       PetscCall(MatFactorGetSchurComplement(F,&St,&st));
15089566063dSJacob Faibussowitsch       PetscCall(MatBindToCPU(St,flg));
15099566063dSJacob Faibussowitsch       PetscCall(MatFactorRestoreSchurComplement(F,&St,st));
15107ebab0bbSStefano Zampini     }
15117ebab0bbSStefano Zampini #endif
15127ebab0bbSStefano Zampini 
1513683d3df6SStefano Zampini     schur_factor = NULL;
151445951f25SStefano Zampini     if (compute_Stilda && size_active_schur) {
1515683d3df6SStefano Zampini 
15169d54b7f4SStefano Zampini       if (sub_schurs->n_subs == 1 && size_schur == size_active_schur && deluxe) { /* we already computed the inverse */
151732fe681dSStefano Zampini         PetscCall(MatSeqAIJGetArrayWrite(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
15189566063dSJacob Faibussowitsch         PetscCall(PetscArraycpy(SEjinv_arr,work,size_schur*size_schur));
151932fe681dSStefano Zampini         PetscCall(MatSeqAIJRestoreArrayWrite(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
15204a6c6b0dSStefano Zampini       } else {
1521683d3df6SStefano Zampini         Mat S_all_inv=NULL;
15227ebab0bbSStefano Zampini 
152332fe681dSStefano Zampini         if (solver_S && !sub_schurs->gdsw) {
1524683d3df6SStefano Zampini           /* for adaptive selection we need S^-1; for solver reusage we need S_\Delta\Delta^-1.
1525683d3df6SStefano Zampini              The latter is not the principal subminor for S^-1. However, the factors can be reused since S_\Delta\Delta is the leading principal submatrix of S */
15263fc34f97SStefano Zampini           if (factor_workaround) {/* invert without calling MatFactorInvertSchurComplement, since we are hacking */
1527683d3df6SStefano Zampini             PetscScalar *data;
1528683d3df6SStefano Zampini             PetscInt     nd = 0;
15296dba178dSStefano Zampini 
1530f4f7d9d6SStefano Zampini             if (!use_potr) {
1531683d3df6SStefano Zampini               SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices");
1532683d3df6SStefano Zampini             }
15339566063dSJacob Faibussowitsch             PetscCall(MatFactorGetSchurComplement(F,&S_all_inv,NULL));
15349566063dSJacob Faibussowitsch             PetscCall(MatDenseGetArray(S_all_inv,&data));
1535683d3df6SStefano Zampini             if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */
15369566063dSJacob Faibussowitsch               PetscCall(ISGetLocalSize(sub_schurs->is_dir,&nd));
1537683d3df6SStefano Zampini             }
15383fc34f97SStefano Zampini 
15393fc34f97SStefano Zampini             /* factor and invert activedofs and vertices (dirichlet dofs does not contribute) */
15403fc34f97SStefano Zampini             if (schur_has_vertices) {
15413fc34f97SStefano Zampini               Mat          M;
15423fc34f97SStefano Zampini               PetscScalar *tdata;
15433fc34f97SStefano Zampini               PetscInt     nv = 0, news;
15443fc34f97SStefano Zampini 
15459566063dSJacob Faibussowitsch               PetscCall(ISGetLocalSize(sub_schurs->is_vertices,&nv));
15463fc34f97SStefano Zampini               news = size_active_schur + nv;
15479566063dSJacob Faibussowitsch               PetscCall(PetscCalloc1(news*news,&tdata));
1548683d3df6SStefano Zampini               for (i=0;i<size_active_schur;i++) {
15499566063dSJacob Faibussowitsch                 PetscCall(PetscArraycpy(tdata+i*(news+1),data+i*(size_schur+1),size_active_schur-i));
15509566063dSJacob Faibussowitsch                 PetscCall(PetscArraycpy(tdata+i*(news+1)+size_active_schur-i,data+i*size_schur+size_active_schur+nd,nv));
15513fc34f97SStefano Zampini               }
15523fc34f97SStefano Zampini               for (i=0;i<nv;i++) {
15533fc34f97SStefano Zampini                 PetscInt k = i+size_active_schur;
15549566063dSJacob Faibussowitsch                 PetscCall(PetscArraycpy(tdata+k*(news+1),data+(k+nd)*(size_schur+1),nv-i));
15553fc34f97SStefano Zampini               }
15563fc34f97SStefano Zampini 
15579566063dSJacob Faibussowitsch               PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,news,news,tdata,&M));
15589566063dSJacob Faibussowitsch               PetscCall(MatSetOption(M,MAT_SPD,PETSC_TRUE));
15599566063dSJacob Faibussowitsch               PetscCall(MatCholeskyFactor(M,NULL,NULL));
15603fc34f97SStefano Zampini               /* save the factors */
15613fc34f97SStefano Zampini               cum = 0;
15629566063dSJacob Faibussowitsch               PetscCall(PetscMalloc1((size_active_schur*(size_active_schur +1))/2+nd,&schur_factor));
15633fc34f97SStefano Zampini               for (i=0;i<size_active_schur;i++) {
15649566063dSJacob Faibussowitsch                 PetscCall(PetscArraycpy(schur_factor+cum,tdata+i*(news+1),size_active_schur-i));
1565683d3df6SStefano Zampini                 cum += size_active_schur - i;
1566683d3df6SStefano Zampini               }
15673fc34f97SStefano Zampini               for (i=0;i<nd;i++) schur_factor[cum+i] = PetscSqrtReal(PetscRealPart(data[(i+size_active_schur)*(size_schur+1)]));
15689566063dSJacob Faibussowitsch               PetscCall(MatSeqDenseInvertFactors_Private(M));
15693fc34f97SStefano Zampini               /* move back just the active dofs to the Schur complement */
15703fc34f97SStefano Zampini               for (i=0;i<size_active_schur;i++) {
15719566063dSJacob Faibussowitsch                 PetscCall(PetscArraycpy(data+i*size_schur,tdata+i*news,size_active_schur));
15723fc34f97SStefano Zampini               }
15739566063dSJacob Faibussowitsch               PetscCall(PetscFree(tdata));
15749566063dSJacob Faibussowitsch               PetscCall(MatDestroy(&M));
15753fc34f97SStefano Zampini             } else { /* we can factorize and invert just the activedofs */
15763fc34f97SStefano Zampini               Mat         M;
15775002105bSStefano Zampini               PetscScalar *aux;
15783fc34f97SStefano Zampini 
15799566063dSJacob Faibussowitsch               PetscCall(PetscMalloc1(nd,&aux));
15805002105bSStefano Zampini               for (i=0;i<nd;i++) aux[i] = 1.0/data[(i+size_active_schur)*(size_schur+1)];
15819566063dSJacob Faibussowitsch               PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,size_active_schur,size_active_schur,data,&M));
15829566063dSJacob Faibussowitsch               PetscCall(MatDenseSetLDA(M,size_schur));
15839566063dSJacob Faibussowitsch               PetscCall(MatSetOption(M,MAT_SPD,PETSC_TRUE));
15849566063dSJacob Faibussowitsch               PetscCall(MatCholeskyFactor(M,NULL,NULL));
15859566063dSJacob Faibussowitsch               PetscCall(MatSeqDenseInvertFactors_Private(M));
15869566063dSJacob Faibussowitsch               PetscCall(MatDestroy(&M));
15879566063dSJacob Faibussowitsch               PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,size_schur,nd,data+size_active_schur*size_schur,&M));
15889566063dSJacob Faibussowitsch               PetscCall(MatZeroEntries(M));
15899566063dSJacob Faibussowitsch               PetscCall(MatDestroy(&M));
15909566063dSJacob Faibussowitsch               PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nd,size_schur,data+size_active_schur,&M));
15919566063dSJacob Faibussowitsch               PetscCall(MatDenseSetLDA(M,size_schur));
15929566063dSJacob Faibussowitsch               PetscCall(MatZeroEntries(M));
15939566063dSJacob Faibussowitsch               PetscCall(MatDestroy(&M));
15945002105bSStefano Zampini               for (i=0;i<nd;i++) data[(i+size_active_schur)*(size_schur+1)] = aux[i];
15959566063dSJacob Faibussowitsch               PetscCall(PetscFree(aux));
15963fc34f97SStefano Zampini             }
15979566063dSJacob Faibussowitsch             PetscCall(MatDenseRestoreArray(S_all_inv,&data));
15983fc34f97SStefano Zampini           } else { /* use MatFactor calls to invert S */
15999566063dSJacob Faibussowitsch             PetscCall(MatFactorInvertSchurComplement(F));
16009566063dSJacob Faibussowitsch             PetscCall(MatFactorGetSchurComplement(F,&S_all_inv,NULL));
1601683d3df6SStefano Zampini           }
160232fe681dSStefano Zampini         } else if (!sub_schurs->gdsw) { /* we need to invert explicitly since we are not using MatFactor for S */
16039566063dSJacob Faibussowitsch           PetscCall(PetscObjectReference((PetscObject)S_all));
1604683d3df6SStefano Zampini           S_all_inv = S_all;
16059566063dSJacob Faibussowitsch           PetscCall(MatDenseGetArray(S_all_inv,&S_data));
16069566063dSJacob Faibussowitsch           PetscCall(PetscBLASIntCast(size_schur,&B_N));
16079566063dSJacob Faibussowitsch           PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
1608f4f7d9d6SStefano Zampini           if (use_potr) {
1609*792fecdfSBarry Smith             PetscCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr));
161028b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr);
1611*792fecdfSBarry Smith             PetscCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr));
161228b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
1613f4f7d9d6SStefano Zampini           } else if (use_sytr) {
1614*792fecdfSBarry Smith             PetscCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
161528b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRF Lapack routine %d",(int)B_ierr);
1616*792fecdfSBarry Smith             PetscCallBLAS("LAPACKsytri",LAPACKsytri_("L",&B_N,S_data,&B_N,pivots,Bwork,&B_ierr));
161728b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRI Lapack routine %d",(int)B_ierr);
1618d6462365SStefano Zampini           } else {
1619*792fecdfSBarry Smith             PetscCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr));
162028b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
1621*792fecdfSBarry Smith             PetscCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
162228b400f6SJacob Faibussowitsch             PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
1623be83ff47SStefano Zampini           }
16249566063dSJacob Faibussowitsch           PetscCall(PetscLogFlops(1.0*size_schur*size_schur*size_schur));
16259566063dSJacob Faibussowitsch           PetscCall(PetscFPTrapPop());
16269566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArray(S_all_inv,&S_data));
162732fe681dSStefano Zampini         } else if (sub_schurs->gdsw) {
162832fe681dSStefano Zampini           Mat      tS, tX, SEj, S_II, S_IE, S_EE;
162932fe681dSStefano Zampini           KSP      pS_II;
163032fe681dSStefano Zampini           PC       pS_II_pc;
163132fe681dSStefano Zampini           IS       EE, II;
163232fe681dSStefano Zampini           PetscInt nS;
163332fe681dSStefano Zampini 
163432fe681dSStefano Zampini           PetscCall(MatFactorCreateSchurComplement(F,&tS,NULL));
163532fe681dSStefano Zampini           PetscCall(MatGetSize(tS,&nS,NULL));
163632fe681dSStefano Zampini           PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
163732fe681dSStefano Zampini           for (i=0,cum=0;i<sub_schurs->n_subs;i++) { /* naive implementation */
163832fe681dSStefano Zampini             PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
163932fe681dSStefano Zampini             PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,SEjinv_arr,&SEj));
164032fe681dSStefano Zampini 
164132fe681dSStefano Zampini             PetscCall(ISCreateStride(PETSC_COMM_SELF,subset_size,cum,1,&EE));
164232fe681dSStefano Zampini             PetscCall(ISComplement(EE,0,nS,&II));
164332fe681dSStefano Zampini             PetscCall(MatCreateSubMatrix(tS,II,II,MAT_INITIAL_MATRIX,&S_II));
164432fe681dSStefano Zampini             PetscCall(MatCreateSubMatrix(tS,II,EE,MAT_INITIAL_MATRIX,&S_IE));
164532fe681dSStefano Zampini             PetscCall(MatCreateSubMatrix(tS,EE,EE,MAT_INITIAL_MATRIX,&S_EE));
164632fe681dSStefano Zampini             PetscCall(ISDestroy(&II));
164732fe681dSStefano Zampini             PetscCall(ISDestroy(&EE));
164832fe681dSStefano Zampini 
164932fe681dSStefano Zampini             PetscCall(KSPCreate(PETSC_COMM_SELF,&pS_II));
165032fe681dSStefano Zampini             PetscCall(KSPSetType(pS_II,KSPPREONLY));
165132fe681dSStefano Zampini             PetscCall(KSPGetPC(pS_II,&pS_II_pc));
165232fe681dSStefano Zampini             PetscCall(PCSetType(pS_II_pc,PCSVD));
165332fe681dSStefano Zampini             PetscCall(KSPSetOptionsPrefix(pS_II,sub_schurs->prefix));
165432fe681dSStefano Zampini             PetscCall(KSPAppendOptionsPrefix(pS_II,"pseudo_"));
165532fe681dSStefano Zampini             PetscCall(KSPSetOperators(pS_II,S_II,S_II));
165632fe681dSStefano Zampini             PetscCall(MatDestroy(&S_II));
165732fe681dSStefano Zampini             PetscCall(KSPSetFromOptions(pS_II));
165832fe681dSStefano Zampini             PetscCall(KSPSetUp(pS_II));
165932fe681dSStefano Zampini             PetscCall(MatDuplicate(S_IE,MAT_DO_NOT_COPY_VALUES,&tX));
166032fe681dSStefano Zampini             PetscCall(KSPMatSolve(pS_II,S_IE,tX));
166132fe681dSStefano Zampini             PetscCall(KSPDestroy(&pS_II));
166232fe681dSStefano Zampini 
166332fe681dSStefano Zampini             PetscCall(MatTransposeMatMult(S_IE,tX,MAT_REUSE_MATRIX,PETSC_DEFAULT,&SEj));
166432fe681dSStefano Zampini             PetscCall(MatDestroy(&S_IE));
166532fe681dSStefano Zampini             PetscCall(MatDestroy(&tX));
166632fe681dSStefano Zampini             PetscCall(MatAYPX(SEj,-1,S_EE,SAME_NONZERO_PATTERN));
166732fe681dSStefano Zampini             PetscCall(MatDestroy(&S_EE));
166832fe681dSStefano Zampini 
166932fe681dSStefano Zampini             PetscCall(MatDestroy(&SEj));
167032fe681dSStefano Zampini             cum += subset_size;
167132fe681dSStefano Zampini             SEjinv_arr += subset_size*subset_size;
167232fe681dSStefano Zampini           }
167332fe681dSStefano Zampini           PetscCall(MatDestroy(&tS));
167432fe681dSStefano Zampini           PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
1675be83ff47SStefano Zampini         }
1676be83ff47SStefano Zampini         /* S_Ej_tilda_all */
1677be83ff47SStefano Zampini         cum = cum2 = 0;
167832fe681dSStefano Zampini         rS_data = NULL;
167932fe681dSStefano Zampini         if (S_all_inv) PetscCall(MatDenseGetArrayRead(S_all_inv,&rS_data));
168032fe681dSStefano Zampini         PetscCall(MatSeqAIJGetArrayWrite(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
1681be83ff47SStefano Zampini         for (i=0;i<sub_schurs->n_subs;i++) {
1682be83ff47SStefano Zampini           PetscInt j;
1683862806e4SStefano Zampini 
16849566063dSJacob Faibussowitsch           PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
1685be83ff47SStefano Zampini           /* get (St^-1)_E */
168672b8c272SStefano Zampini           /* Unless we are changing the variables, I don't need to expand to upper triangular since St^-1
168706a4e24aSStefano Zampini              will be properly accessed later during adaptive selection */
168832fe681dSStefano Zampini           if (rS_data) {
1689a0b0af32SStefano Zampini             if (S_lower_triangular) {
1690be83ff47SStefano Zampini               PetscInt k;
1691b7ab4a40SStefano Zampini               if (sub_schurs->change) {
1692be83ff47SStefano Zampini                 for (k=0;k<subset_size;k++) {
1693be83ff47SStefano Zampini                   for (j=k;j<subset_size;j++) {
16941683a169SBarry Smith                     work[k*subset_size+j] = rS_data[cum2+k*size_schur+j];
16956c3e6151SStefano Zampini                     work[j*subset_size+k] = work[k*subset_size+j];
1696be83ff47SStefano Zampini                   }
1697be83ff47SStefano Zampini                 }
169872b8c272SStefano Zampini               } else {
169972b8c272SStefano Zampini                 for (k=0;k<subset_size;k++) {
170072b8c272SStefano Zampini                   for (j=k;j<subset_size;j++) {
17011683a169SBarry Smith                     work[k*subset_size+j] = rS_data[cum2+k*size_schur+j];
170272b8c272SStefano Zampini                   }
170372b8c272SStefano Zampini                 }
170472b8c272SStefano Zampini               }
170572b8c272SStefano Zampini             } else {
1706be83ff47SStefano Zampini               PetscInt k;
1707be83ff47SStefano Zampini               for (k=0;k<subset_size;k++) {
1708be83ff47SStefano Zampini                 for (j=0;j<subset_size;j++) {
17091683a169SBarry Smith                   work[k*subset_size+j] = rS_data[cum2+k*size_schur+j];
1710be83ff47SStefano Zampini                 }
1711be83ff47SStefano Zampini               }
1712be83ff47SStefano Zampini             }
171332fe681dSStefano Zampini           }
1714b7ab4a40SStefano Zampini           if (sub_schurs->change) {
17158760537fSStefano Zampini             Mat change_sub,SEj,T;
171632fe681dSStefano Zampini             PetscScalar val = sub_schurs->gdsw ? PETSC_SMALL : 1./PETSC_SMALL;
171772b8c272SStefano Zampini 
171872b8c272SStefano Zampini             /* change basis */
17199566063dSJacob Faibussowitsch             PetscCall(KSPGetOperators(sub_schurs->change[i],&change_sub,NULL));
172032fe681dSStefano Zampini             PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,rS_data ? work : SEjinv_arr,&SEj));
17218760537fSStefano Zampini             if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */
17228760537fSStefano Zampini               Mat T2;
17239566063dSJacob Faibussowitsch               PetscCall(MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2));
17249566063dSJacob Faibussowitsch               PetscCall(MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T));
17259566063dSJacob Faibussowitsch               PetscCall(MatDestroy(&T2));
17269566063dSJacob Faibussowitsch               PetscCall(MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T));
17278760537fSStefano Zampini             } else {
17289566063dSJacob Faibussowitsch               PetscCall(MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T));
172972b8c272SStefano Zampini             }
17309566063dSJacob Faibussowitsch             PetscCall(MatCopy(T,SEj,SAME_NONZERO_PATTERN));
17319566063dSJacob Faibussowitsch             PetscCall(MatDestroy(&T));
173232fe681dSStefano Zampini             PetscCall(MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],val,NULL,NULL));
17339566063dSJacob Faibussowitsch             PetscCall(MatDestroy(&SEj));
173472b8c272SStefano Zampini           }
173532fe681dSStefano Zampini           if (rS_data) PetscCall(PetscArraycpy(SEjinv_arr,work,subset_size*subset_size));
1736be83ff47SStefano Zampini           cum += subset_size;
1737be83ff47SStefano Zampini           cum2 += subset_size*(size_schur + 1);
173804c5b2e6SStefano Zampini           SEjinv_arr += subset_size*subset_size;
1739883469d8SStefano Zampini         }
174032fe681dSStefano Zampini         PetscCall(MatSeqAIJRestoreArrayWrite(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr));
174132fe681dSStefano Zampini         if (S_all_inv) {
17429566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArrayRead(S_all_inv,&rS_data));
1743df4d28bfSStefano Zampini           if (solver_S) {
17443fc34f97SStefano Zampini             if (schur_has_vertices) {
17459566063dSJacob Faibussowitsch               PetscCall(MatFactorRestoreSchurComplement(F,&S_all_inv,MAT_FACTOR_SCHUR_FACTORED));
17463fc34f97SStefano Zampini             } else {
17479566063dSJacob Faibussowitsch               PetscCall(MatFactorRestoreSchurComplement(F,&S_all_inv,MAT_FACTOR_SCHUR_INVERTED));
17485db18549SStefano Zampini             }
17493fc34f97SStefano Zampini           }
175032fe681dSStefano Zampini         }
17519566063dSJacob Faibussowitsch         PetscCall(MatDestroy(&S_all_inv));
1752683d3df6SStefano Zampini       }
1753683d3df6SStefano Zampini 
17543fc34f97SStefano Zampini       /* move back factors if needed */
175532fe681dSStefano Zampini       if (schur_has_vertices && factor_workaround && !sub_schurs->gdsw) {
1756683d3df6SStefano Zampini         Mat      S_tmp;
17573fc34f97SStefano Zampini         PetscInt nd = 0;
1758683d3df6SStefano Zampini 
175928b400f6SJacob Faibussowitsch         PetscCheck(solver_S,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
17609566063dSJacob Faibussowitsch         PetscCall(MatFactorGetSchurComplement(F,&S_tmp,NULL));
1761f4f7d9d6SStefano Zampini         if (use_potr) {
1762683d3df6SStefano Zampini           PetscScalar *data;
1763683d3df6SStefano Zampini 
17649566063dSJacob Faibussowitsch           PetscCall(MatDenseGetArray(S_tmp,&data));
17659566063dSJacob Faibussowitsch           PetscCall(PetscArrayzero(data,size_schur*size_schur));
1766683d3df6SStefano Zampini 
1767683d3df6SStefano Zampini           if (S_lower_triangular) {
1768683d3df6SStefano Zampini             cum = 0;
1769683d3df6SStefano Zampini             for (i=0;i<size_active_schur;i++) {
17709566063dSJacob Faibussowitsch               PetscCall(PetscArraycpy(data+i*(size_schur+1),schur_factor+cum,size_active_schur-i));
1771683d3df6SStefano Zampini               cum += size_active_schur-i;
1772683d3df6SStefano Zampini             }
1773683d3df6SStefano Zampini           } else {
17749566063dSJacob Faibussowitsch             PetscCall(PetscArraycpy(data,schur_factor,size_schur*size_schur));
1775683d3df6SStefano Zampini           }
1776683d3df6SStefano Zampini           if (sub_schurs->is_dir) {
17779566063dSJacob Faibussowitsch             PetscCall(ISGetLocalSize(sub_schurs->is_dir,&nd));
1778683d3df6SStefano Zampini             for (i=0;i<nd;i++) {
1779683d3df6SStefano Zampini               data[(i+size_active_schur)*(size_schur+1)] = schur_factor[cum+i];
1780683d3df6SStefano Zampini             }
1781683d3df6SStefano Zampini           }
17826dba178dSStefano Zampini           /* workaround: since I cannot modify the matrices used inside the solvers for the forward and backward substitutions,
1783683d3df6SStefano Zampini              set the diagonal entry of the Schur factor to a very large value */
1784683d3df6SStefano Zampini           for (i=size_active_schur+nd;i<size_schur;i++) {
17856c3e6151SStefano Zampini             data[i*(size_schur+1)] = infty;
1786683d3df6SStefano Zampini           }
17879566063dSJacob Faibussowitsch           PetscCall(MatDenseRestoreArray(S_tmp,&data));
17886542c05fSStefano Zampini         } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices");
17899566063dSJacob Faibussowitsch         PetscCall(MatFactorRestoreSchurComplement(F,&S_tmp,MAT_FACTOR_SCHUR_FACTORED));
17909087bf02SStefano Zampini       }
179132fe681dSStefano Zampini     } else if (factor_workaround && !sub_schurs->gdsw) { /* we need to eliminate any unneeded coupling */
1792367aa537SStefano Zampini       PetscScalar *data;
1793367aa537SStefano Zampini       PetscInt    nd = 0;
1794367aa537SStefano Zampini 
1795367aa537SStefano Zampini       if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */
17969566063dSJacob Faibussowitsch         PetscCall(ISGetLocalSize(sub_schurs->is_dir,&nd));
1797367aa537SStefano Zampini       }
17989566063dSJacob Faibussowitsch       PetscCall(MatFactorGetSchurComplement(F,&S_all,NULL));
17999566063dSJacob Faibussowitsch       PetscCall(MatDenseGetArray(S_all,&data));
1800367aa537SStefano Zampini       for (i=0;i<size_active_schur;i++) {
18019566063dSJacob Faibussowitsch         PetscCall(PetscArrayzero(data+i*size_schur+size_active_schur,size_schur-size_active_schur));
1802367aa537SStefano Zampini       }
1803367aa537SStefano Zampini       for (i=size_active_schur+nd;i<size_schur;i++) {
18049566063dSJacob Faibussowitsch         PetscCall(PetscArrayzero(data+i*size_schur+size_active_schur,size_schur-size_active_schur));
18056c3e6151SStefano Zampini         data[i*(size_schur+1)] = infty;
1806367aa537SStefano Zampini       }
18079566063dSJacob Faibussowitsch       PetscCall(MatDenseRestoreArray(S_all,&data));
18089566063dSJacob Faibussowitsch       PetscCall(MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED));
18094a6c6b0dSStefano Zampini     }
18109566063dSJacob Faibussowitsch     PetscCall(PetscFree(work));
18119566063dSJacob Faibussowitsch     PetscCall(PetscFree(schur_factor));
18129566063dSJacob Faibussowitsch     PetscCall(VecDestroy(&Dall));
18134a6c6b0dSStefano Zampini   }
18149566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&is_I_layer));
18159566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&S_all));
18169566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&A_BB));
18179566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&A_IB));
18189566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&A_BI));
18199566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&F));
18206afe12f5SStefano Zampini 
18219566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(sub_schurs->n_subs,&nnz));
18226afe12f5SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
18239566063dSJacob Faibussowitsch     PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&nnz[i]));
18246afe12f5SStefano Zampini   }
18259566063dSJacob Faibussowitsch   PetscCall(ISCreateGeneral(PETSC_COMM_SELF,sub_schurs->n_subs,nnz,PETSC_OWN_POINTER,&is_I_layer));
18269566063dSJacob Faibussowitsch   PetscCall(MatSetVariableBlockSizes(sub_schurs->S_Ej_all,sub_schurs->n_subs,nnz));
18279566063dSJacob Faibussowitsch   PetscCall(MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY));
18289566063dSJacob Faibussowitsch   PetscCall(MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY));
18295a95e1ceSStefano Zampini   if (compute_Stilda) {
18309566063dSJacob Faibussowitsch     PetscCall(MatSetVariableBlockSizes(sub_schurs->sum_S_Ej_tilda_all,sub_schurs->n_subs,nnz));
18319566063dSJacob Faibussowitsch     PetscCall(MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY));
18329566063dSJacob Faibussowitsch     PetscCall(MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY));
18339d54b7f4SStefano Zampini     if (deluxe) {
18349566063dSJacob Faibussowitsch       PetscCall(MatSetVariableBlockSizes(sub_schurs->sum_S_Ej_inv_all,sub_schurs->n_subs,nnz));
18359566063dSJacob Faibussowitsch       PetscCall(MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY));
18369566063dSJacob Faibussowitsch       PetscCall(MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY));
183708122e43SStefano Zampini     }
18389d54b7f4SStefano Zampini   }
18399566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&is_I_layer));
18406afe12f5SStefano Zampini 
18415db18549SStefano Zampini   /* Get local part of (\sum_j S_Ej) */
184241fd5443SStefano Zampini   if (!sub_schurs->sum_S_Ej_all) {
18439566063dSJacob Faibussowitsch     PetscCall(MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_all));
184441fd5443SStefano Zampini   }
18459566063dSJacob Faibussowitsch   PetscCall(VecSet(gstash,0.0));
18469566063dSJacob Faibussowitsch   PetscCall(MatSeqAIJGetArray(sub_schurs->S_Ej_all,&stasharray));
18479566063dSJacob Faibussowitsch   PetscCall(VecPlaceArray(lstash,stasharray));
18489566063dSJacob Faibussowitsch   PetscCall(VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18499566063dSJacob Faibussowitsch   PetscCall(VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18509566063dSJacob Faibussowitsch   PetscCall(MatSeqAIJRestoreArray(sub_schurs->S_Ej_all,&stasharray));
18519566063dSJacob Faibussowitsch   PetscCall(VecResetArray(lstash));
18529566063dSJacob Faibussowitsch   PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&stasharray));
18539566063dSJacob Faibussowitsch   PetscCall(VecPlaceArray(lstash,stasharray));
18549566063dSJacob Faibussowitsch   PetscCall(VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18559566063dSJacob Faibussowitsch   PetscCall(VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18569566063dSJacob Faibussowitsch   PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&stasharray));
18579566063dSJacob Faibussowitsch   PetscCall(VecResetArray(lstash));
185808122e43SStefano Zampini 
1859f6f667cfSStefano Zampini   /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */
18605a95e1ceSStefano Zampini   if (compute_Stilda) {
18619566063dSJacob Faibussowitsch     PetscCall(VecSet(gstash,0.0));
18629566063dSJacob Faibussowitsch     PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&stasharray));
18639566063dSJacob Faibussowitsch     PetscCall(VecPlaceArray(lstash,stasharray));
18649566063dSJacob Faibussowitsch     PetscCall(VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18659566063dSJacob Faibussowitsch     PetscCall(VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18669566063dSJacob Faibussowitsch     PetscCall(VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18679566063dSJacob Faibussowitsch     PetscCall(VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18689566063dSJacob Faibussowitsch     PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&stasharray));
18699566063dSJacob Faibussowitsch     PetscCall(VecResetArray(lstash));
18709d54b7f4SStefano Zampini     if (deluxe) {
18719566063dSJacob Faibussowitsch       PetscCall(VecSet(gstash,0.0));
18729566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&stasharray));
18739566063dSJacob Faibussowitsch       PetscCall(VecPlaceArray(lstash,stasharray));
18749566063dSJacob Faibussowitsch       PetscCall(VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18759566063dSJacob Faibussowitsch       PetscCall(VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD));
18769566063dSJacob Faibussowitsch       PetscCall(VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18779566063dSJacob Faibussowitsch       PetscCall(VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE));
18789566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&stasharray));
18799566063dSJacob Faibussowitsch       PetscCall(VecResetArray(lstash));
188032fe681dSStefano Zampini     } else if (!sub_schurs->gdsw) {
18819d54b7f4SStefano Zampini       PetscScalar *array;
18829d54b7f4SStefano Zampini       PetscInt    cum;
18839d54b7f4SStefano Zampini 
18849566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array));
18859d54b7f4SStefano Zampini       cum = 0;
18869d54b7f4SStefano Zampini       for (i=0;i<sub_schurs->n_subs;i++) {
18879566063dSJacob Faibussowitsch         PetscCall(ISGetLocalSize(sub_schurs->is_subs[i],&subset_size));
18889566063dSJacob Faibussowitsch         PetscCall(PetscBLASIntCast(subset_size,&B_N));
18899566063dSJacob Faibussowitsch         PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
1890f4f7d9d6SStefano Zampini         if (use_potr) {
1891*792fecdfSBarry Smith           PetscCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr));
189228b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr);
1893*792fecdfSBarry Smith           PetscCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr));
189428b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
1895f4f7d9d6SStefano Zampini         } else if (use_sytr) {
1896*792fecdfSBarry Smith           PetscCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
189728b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRF Lapack routine %d",(int)B_ierr);
1898*792fecdfSBarry Smith           PetscCallBLAS("LAPACKsytri",LAPACKsytri_("L",&B_N,array+cum,&B_N,pivots,Bwork,&B_ierr));
189928b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRI Lapack routine %d",(int)B_ierr);
1900f4f7d9d6SStefano Zampini         } else {
1901*792fecdfSBarry Smith           PetscCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr));
190228b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
1903*792fecdfSBarry Smith           PetscCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
190428b400f6SJacob Faibussowitsch           PetscCheck(!B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
1905f4f7d9d6SStefano Zampini         }
19069566063dSJacob Faibussowitsch         PetscCall(PetscLogFlops(1.0*subset_size*subset_size*subset_size));
19079566063dSJacob Faibussowitsch         PetscCall(PetscFPTrapPop());
19089d54b7f4SStefano Zampini         cum += subset_size*subset_size;
19099d54b7f4SStefano Zampini       }
19109566063dSJacob Faibussowitsch       PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array));
19119566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)sub_schurs->sum_S_Ej_all));
19129566063dSJacob Faibussowitsch       PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_inv_all));
19139d54b7f4SStefano Zampini       sub_schurs->sum_S_Ej_inv_all = sub_schurs->sum_S_Ej_all;
19149d54b7f4SStefano Zampini     }
191508122e43SStefano Zampini   }
19169566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&lstash));
19179566063dSJacob Faibussowitsch   PetscCall(VecDestroy(&gstash));
19189566063dSJacob Faibussowitsch   PetscCall(VecScatterDestroy(&sstash));
191957a87bf3SStefano Zampini 
19203b03f7bbSStefano Zampini   if (matl_dbg_viewer) {
192111955456SStefano Zampini     if (sub_schurs->S_Ej_all) {
19229566063dSJacob Faibussowitsch       PetscCall(PetscObjectSetName((PetscObject)sub_schurs->S_Ej_all,"SE"));
19239566063dSJacob Faibussowitsch       PetscCall(MatView(sub_schurs->S_Ej_all,matl_dbg_viewer));
192411955456SStefano Zampini     }
192511955456SStefano Zampini     if (sub_schurs->sum_S_Ej_all) {
19269566063dSJacob Faibussowitsch       PetscCall(PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_all,"SSE"));
19279566063dSJacob Faibussowitsch       PetscCall(MatView(sub_schurs->sum_S_Ej_all,matl_dbg_viewer));
192811955456SStefano Zampini     }
192911955456SStefano Zampini     if (sub_schurs->sum_S_Ej_inv_all) {
19309566063dSJacob Faibussowitsch       PetscCall(PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_inv_all,"SSEm"));
19319566063dSJacob Faibussowitsch       PetscCall(MatView(sub_schurs->sum_S_Ej_inv_all,matl_dbg_viewer));
193211955456SStefano Zampini     }
193311955456SStefano Zampini     if (sub_schurs->sum_S_Ej_tilda_all) {
19349566063dSJacob Faibussowitsch       PetscCall(PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_tilda_all,"SSEt"));
19359566063dSJacob Faibussowitsch       PetscCall(MatView(sub_schurs->sum_S_Ej_tilda_all,matl_dbg_viewer));
193611955456SStefano Zampini     }
193711955456SStefano Zampini   }
19383202ece2SStefano Zampini 
19395a95e1ceSStefano Zampini   /* free workspace */
194051ab8ad6SStefano Zampini   if (matl_dbg_viewer) PetscCall(PetscViewerFlush(matl_dbg_viewer));
194151ab8ad6SStefano Zampini   if (sub_schurs->debug) PetscCallMPI(MPI_Barrier(comm_n));
19429566063dSJacob Faibussowitsch   PetscCall(PetscViewerDestroy(&matl_dbg_viewer));
19439566063dSJacob Faibussowitsch   PetscCall(PetscFree2(Bwork,pivots));
19449566063dSJacob Faibussowitsch   PetscCall(PetscCommDestroy(&comm_n));
1945b1b3d7a2SStefano Zampini   PetscFunctionReturn(0);
1946b1b3d7a2SStefano Zampini }
1947b1b3d7a2SStefano Zampini 
194832fe681dSStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, const char* prefix, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap, PetscBool copycc, PetscBool gdsw)
1949b1b3d7a2SStefano Zampini {
19509bb4a8caSStefano Zampini   IS              *faces,*edges,*all_cc,vertices;
195132fe681dSStefano Zampini   PetscInt        s,i,n_faces,n_edges,n_all_cc;
1952365a3a41SStefano Zampini   PetscBool       is_sorted,ispardiso,ismumps;
1953b1b3d7a2SStefano Zampini 
1954b1b3d7a2SStefano Zampini   PetscFunctionBegin;
19559566063dSJacob Faibussowitsch   PetscCall(ISSorted(is_I,&is_sorted));
195628b400f6SJacob Faibussowitsch   PetscCheck(is_sorted,PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted");
19579566063dSJacob Faibussowitsch   PetscCall(ISSorted(is_B,&is_sorted));
195828b400f6SJacob Faibussowitsch   PetscCheck(is_sorted,PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted");
1959b1b3d7a2SStefano Zampini 
1960b1b3d7a2SStefano Zampini   /* reset any previous data */
19619566063dSJacob Faibussowitsch   PetscCall(PCBDDCSubSchursReset(sub_schurs));
1962b1b3d7a2SStefano Zampini 
196332fe681dSStefano Zampini   sub_schurs->gdsw = gdsw;
196432fe681dSStefano Zampini 
19655a95e1ceSStefano Zampini   /* get index sets for faces and edges (already sorted by global ordering) */
19669566063dSJacob Faibussowitsch   PetscCall(PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices));
1967b1b3d7a2SStefano Zampini   n_all_cc = n_faces+n_edges;
19689566063dSJacob Faibussowitsch   PetscCall(PetscBTCreate(n_all_cc,&sub_schurs->is_edge));
19699566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(n_all_cc,&all_cc));
197032fe681dSStefano Zampini   n_all_cc = 0;
1971b1b3d7a2SStefano Zampini   for (i=0;i<n_faces;i++) {
197232fe681dSStefano Zampini     PetscCall(ISGetSize(faces[i],&s));
197332fe681dSStefano Zampini     if (!s) continue;
19748b6046baSStefano Zampini     if (copycc) {
197532fe681dSStefano Zampini       PetscCall(ISDuplicate(faces[i],&all_cc[n_all_cc]));
19768b6046baSStefano Zampini     } else {
19779566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)faces[i]));
197832fe681dSStefano Zampini       all_cc[n_all_cc] = faces[i];
1979b1b3d7a2SStefano Zampini     }
198032fe681dSStefano Zampini     n_all_cc++;
19818b6046baSStefano Zampini   }
1982b1b3d7a2SStefano Zampini   for (i=0;i<n_edges;i++) {
198332fe681dSStefano Zampini     PetscCall(ISGetSize(edges[i],&s));
198432fe681dSStefano Zampini     if (!s) continue;
19858b6046baSStefano Zampini     if (copycc) {
198632fe681dSStefano Zampini       PetscCall(ISDuplicate(edges[i],&all_cc[n_all_cc]));
19878b6046baSStefano Zampini     } else {
19889566063dSJacob Faibussowitsch       PetscCall(PetscObjectReference((PetscObject)edges[i]));
198932fe681dSStefano Zampini       all_cc[n_all_cc] = edges[i];
19908b6046baSStefano Zampini     }
199132fe681dSStefano Zampini     PetscCall(PetscBTSet(sub_schurs->is_edge,n_all_cc));
199232fe681dSStefano Zampini     n_all_cc++;
1993b1b3d7a2SStefano Zampini   }
19949566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)vertices));
1995c8272957SStefano Zampini   sub_schurs->is_vertices = vertices;
19969566063dSJacob Faibussowitsch   PetscCall(PCBDDCGraphRestoreCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices));
1997d62866d3SStefano Zampini   sub_schurs->is_dir = NULL;
19989566063dSJacob Faibussowitsch   PetscCall(PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir));
1999b1b3d7a2SStefano Zampini 
2000df4d28bfSStefano Zampini   /* Determine if MatFactor can be used */
20019566063dSJacob Faibussowitsch   PetscCall(PetscStrallocpy(prefix,&sub_schurs->prefix));
2002883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
20039566063dSJacob Faibussowitsch   PetscCall(PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERMUMPS,sizeof(sub_schurs->mat_solver_type)));
200488113c35SStefano Zampini #elif defined(PETSC_HAVE_MKL_PARDISO)
20059566063dSJacob Faibussowitsch   PetscCall(PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,sizeof(sub_schurs->mat_solver_type)));
200688113c35SStefano Zampini #else
20079566063dSJacob Faibussowitsch   PetscCall(PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERPETSC,sizeof(sub_schurs->mat_solver_type)));
2008df4d28bfSStefano Zampini #endif
200988113c35SStefano Zampini #if defined(PETSC_USE_COMPLEX)
201088113c35SStefano Zampini   sub_schurs->is_hermitian  = PETSC_FALSE; /* Hermitian Cholesky is not supported by PETSc and external packages */
201188113c35SStefano Zampini #else
201288113c35SStefano Zampini   sub_schurs->is_hermitian  = PETSC_TRUE;
2013883469d8SStefano Zampini #endif
201488113c35SStefano Zampini   sub_schurs->is_posdef     = PETSC_TRUE;
201511955456SStefano Zampini   sub_schurs->is_symmetric  = PETSC_TRUE;
20167f9db97bSStefano Zampini   sub_schurs->debug         = PETSC_FALSE;
2017991c41b4SStefano Zampini   sub_schurs->restrict_comm = PETSC_FALSE;
2018d0609cedSBarry Smith   PetscOptionsBegin(PetscObjectComm((PetscObject)graph->l2gmap),sub_schurs->prefix,"BDDC sub_schurs options","PC");
20199566063dSJacob Faibussowitsch   PetscCall(PetscOptionsString("-sub_schurs_mat_solver_type","Specific direct solver to use",NULL,sub_schurs->mat_solver_type,sub_schurs->mat_solver_type,sizeof(sub_schurs->mat_solver_type),NULL));
20209566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sub_schurs_symmetric","Symmetric problem",NULL,sub_schurs->is_symmetric,&sub_schurs->is_symmetric,NULL));
20219566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sub_schurs_hermitian","Hermitian problem",NULL,sub_schurs->is_hermitian,&sub_schurs->is_hermitian,NULL));
20229566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sub_schurs_posdef","Positive definite problem",NULL,sub_schurs->is_posdef,&sub_schurs->is_posdef,NULL));
20239566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sub_schurs_restrictcomm","Restrict communicator on active processes only",NULL,sub_schurs->restrict_comm,&sub_schurs->restrict_comm,NULL));
20249566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sub_schurs_debug","Debug output",NULL,sub_schurs->debug,&sub_schurs->debug,NULL));
2025d0609cedSBarry Smith   PetscOptionsEnd();
20269566063dSJacob Faibussowitsch   PetscCall(PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMUMPS,&ismumps));
20279566063dSJacob Faibussowitsch   PetscCall(PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,&ispardiso));
2028365a3a41SStefano Zampini   sub_schurs->schur_explicit = (PetscBool)(ispardiso || ismumps);
2029b1b3d7a2SStefano Zampini 
203011955456SStefano Zampini   /* for reals, symmetric and hermitian are synonims */
203111955456SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
203211955456SStefano Zampini   sub_schurs->is_symmetric = (PetscBool)(sub_schurs->is_symmetric && sub_schurs->is_hermitian);
203311955456SStefano Zampini   sub_schurs->is_hermitian = sub_schurs->is_symmetric;
203411955456SStefano Zampini #endif
203511955456SStefano Zampini 
20369566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)is_I));
2037b1b3d7a2SStefano Zampini   sub_schurs->is_I = is_I;
20389566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)is_B));
2039b1b3d7a2SStefano Zampini   sub_schurs->is_B = is_B;
20409566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)graph->l2gmap));
20415db18549SStefano Zampini   sub_schurs->l2gmap = graph->l2gmap;
20429566063dSJacob Faibussowitsch   PetscCall(PetscObjectReference((PetscObject)BtoNmap));
20435db18549SStefano Zampini   sub_schurs->BtoNmap = BtoNmap;
20445a95e1ceSStefano Zampini   sub_schurs->n_subs = n_all_cc;
2045b1b3d7a2SStefano Zampini   sub_schurs->is_subs = all_cc;
2046b96c3477SStefano Zampini   sub_schurs->S_Ej_all = NULL;
2047b96c3477SStefano Zampini   sub_schurs->sum_S_Ej_all = NULL;
204808122e43SStefano Zampini   sub_schurs->sum_S_Ej_inv_all = NULL;
2049b96c3477SStefano Zampini   sub_schurs->sum_S_Ej_tilda_all = NULL;
2050b96c3477SStefano Zampini   sub_schurs->is_Ej_all = NULL;
2051b1b3d7a2SStefano Zampini   PetscFunctionReturn(0);
2052b1b3d7a2SStefano Zampini }
2053b1b3d7a2SStefano Zampini 
205434a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs)
205534a97f8cSStefano Zampini {
205634a97f8cSStefano Zampini   PCBDDCSubSchurs schurs_ctx;
205734a97f8cSStefano Zampini 
205834a97f8cSStefano Zampini   PetscFunctionBegin;
20599566063dSJacob Faibussowitsch   PetscCall(PetscNew(&schurs_ctx));
20605ff63025SStefano Zampini   schurs_ctx->n_subs = 0;
206134a97f8cSStefano Zampini   *sub_schurs = schurs_ctx;
206234a97f8cSStefano Zampini   PetscFunctionReturn(0);
206334a97f8cSStefano Zampini }
206434a97f8cSStefano Zampini 
206534a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs)
206634a97f8cSStefano Zampini {
206734a97f8cSStefano Zampini   PetscInt       i;
206834a97f8cSStefano Zampini 
206934a97f8cSStefano Zampini   PetscFunctionBegin;
2070aea80f77Sstefano_zampini   if (!sub_schurs) PetscFunctionReturn(0);
20719566063dSJacob Faibussowitsch   PetscCall(PetscFree(sub_schurs->prefix));
20729566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->A));
20739566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->S));
20749566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&sub_schurs->is_I));
20759566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&sub_schurs->is_B));
20769566063dSJacob Faibussowitsch   PetscCall(ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap));
20779566063dSJacob Faibussowitsch   PetscCall(ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap));
20789566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->S_Ej_all));
20799566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_all));
20809566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_inv_all));
20819566063dSJacob Faibussowitsch   PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_tilda_all));
20829566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&sub_schurs->is_Ej_all));
20839566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&sub_schurs->is_vertices));
20849566063dSJacob Faibussowitsch   PetscCall(ISDestroy(&sub_schurs->is_dir));
20859566063dSJacob Faibussowitsch   PetscCall(PetscBTDestroy(&sub_schurs->is_edge));
208634a97f8cSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
20879566063dSJacob Faibussowitsch     PetscCall(ISDestroy(&sub_schurs->is_subs[i]));
208834a97f8cSStefano Zampini   }
20891baa6e33SBarry Smith   if (sub_schurs->n_subs) PetscCall(PetscFree(sub_schurs->is_subs));
20901baa6e33SBarry Smith   if (sub_schurs->reuse_solver) PetscCall(PCBDDCReuseSolversReset(sub_schurs->reuse_solver));
20919566063dSJacob Faibussowitsch   PetscCall(PetscFree(sub_schurs->reuse_solver));
209272b8c272SStefano Zampini   if (sub_schurs->change) {
209372b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
20949566063dSJacob Faibussowitsch       PetscCall(KSPDestroy(&sub_schurs->change[i]));
20959566063dSJacob Faibussowitsch       PetscCall(ISDestroy(&sub_schurs->change_primal_sub[i]));
209672b8c272SStefano Zampini     }
209772b8c272SStefano Zampini   }
20989566063dSJacob Faibussowitsch   PetscCall(PetscFree(sub_schurs->change));
20999566063dSJacob Faibussowitsch   PetscCall(PetscFree(sub_schurs->change_primal_sub));
210034a97f8cSStefano Zampini   sub_schurs->n_subs = 0;
210134a97f8cSStefano Zampini   PetscFunctionReturn(0);
210234a97f8cSStefano Zampini }
210334a97f8cSStefano Zampini 
2104aea80f77Sstefano_zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs* sub_schurs)
2105aea80f77Sstefano_zampini {
2106aea80f77Sstefano_zampini   PetscFunctionBegin;
21079566063dSJacob Faibussowitsch   PetscCall(PCBDDCSubSchursReset(*sub_schurs));
21089566063dSJacob Faibussowitsch   PetscCall(PetscFree(*sub_schurs));
2109aea80f77Sstefano_zampini   PetscFunctionReturn(0);
2110aea80f77Sstefano_zampini }
2111aea80f77Sstefano_zampini 
21129fbee547SJacob Faibussowitsch static inline PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added)
211334a97f8cSStefano Zampini {
211434a97f8cSStefano Zampini   PetscInt       i,j,n;
211534a97f8cSStefano Zampini 
211634a97f8cSStefano Zampini   PetscFunctionBegin;
211734a97f8cSStefano Zampini   n = 0;
211834a97f8cSStefano Zampini   for (i=-n_prev;i<0;i++) {
211934a97f8cSStefano Zampini     PetscInt start_dof = queue_tip[i];
212034a97f8cSStefano Zampini     for (j=xadj[start_dof];j<xadj[start_dof+1];j++) {
212134a97f8cSStefano Zampini       PetscInt dof = adjncy[j];
212234a97f8cSStefano Zampini       if (!PetscBTLookup(touched,dof)) {
21239566063dSJacob Faibussowitsch         PetscCall(PetscBTSet(touched,dof));
212434a97f8cSStefano Zampini         queue_tip[n] = dof;
212534a97f8cSStefano Zampini         n++;
212634a97f8cSStefano Zampini       }
212734a97f8cSStefano Zampini     }
212834a97f8cSStefano Zampini   }
212934a97f8cSStefano Zampini   *n_added = n;
213034a97f8cSStefano Zampini   PetscFunctionReturn(0);
213134a97f8cSStefano Zampini }
2132