xref: /petsc/src/ksp/pc/impls/bddc/bddcschurs.c (revision 72b8c272419574d2ebd5d55592dbcb83cf34cd94)
134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h>
234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h>
308122e43SStefano Zampini #include <petscblaslapack.h>
434a97f8cSStefano Zampini 
53202ece2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*);
65ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*);
7df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC,Vec,Vec);
8df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC,Vec,Vec);
9d62866d3SStefano Zampini 
10367aa537SStefano Zampini /* performs S = S + v^2 * a *v_1^T on the lower-triangular part, n the size of the matrix */
11ca92afb2SStefano Zampini #undef __FUNCT__
12ca92afb2SStefano Zampini #define __FUNCT__ "SparseRankOneUpdate"
13ca92afb2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode SparseRankOneUpdate(PetscScalar *S,PetscInt n,PetscScalar a,PetscScalar *v1,PetscScalar *v2)
14ca92afb2SStefano Zampini {
15ca92afb2SStefano Zampini   PetscInt i;
16ca92afb2SStefano Zampini 
17ca92afb2SStefano Zampini   PetscFunctionBegin;
18ca92afb2SStefano Zampini   for (i=0;i<n;i++) {
19ca92afb2SStefano Zampini     if (PetscUnlikely(PetscAbsScalar(v2[i]) > PETSC_SMALL)) {
20ca92afb2SStefano Zampini       PetscScalar  v2v = a*v2[i];
21ca92afb2SStefano Zampini       PetscBLASInt B_N = n-i,B_one = 1;
22ca92afb2SStefano Zampini       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&v2v,v1+i,&B_one,S+i*(n+1),&B_one));
23ca92afb2SStefano Zampini     }
24ca92afb2SStefano Zampini   }
25ca92afb2SStefano Zampini   PetscFunctionReturn(0);
26ca92afb2SStefano Zampini }
27ca92afb2SStefano Zampini 
28ca92afb2SStefano Zampini /* if v2 is not present, correction is done in-place */
29ca92afb2SStefano Zampini #undef __FUNCT__
30ca92afb2SStefano Zampini #define __FUNCT__ "PCBDDCReuseSolversChangeInterior"
31df4d28bfSStefano Zampini PetscErrorCode PCBDDCReuseSolversChangeInterior(PCBDDCReuseSolvers ctx, Vec v, Vec v2, PetscBool sol)
32ca92afb2SStefano Zampini {
33ca92afb2SStefano Zampini   PetscScalar    *array;
34ca92afb2SStefano Zampini   PetscScalar    *array2;
35ca92afb2SStefano Zampini   PetscErrorCode ierr;
36ca92afb2SStefano Zampini 
37ca92afb2SStefano Zampini   PetscFunctionBegin;
38ca92afb2SStefano Zampini   if (!ctx->benign_n) PetscFunctionReturn(0);
39ca92afb2SStefano Zampini   if (v2) {
40ca92afb2SStefano Zampini     PetscInt nl;
41ca92afb2SStefano Zampini 
42ca92afb2SStefano Zampini     ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr);
43ca92afb2SStefano Zampini     ierr = VecGetLocalSize(v2,&nl);CHKERRQ(ierr);
44ca92afb2SStefano Zampini     ierr = VecGetArray(v2,&array2);CHKERRQ(ierr);
45ca92afb2SStefano Zampini     ierr = PetscMemcpy(array2,array,nl*sizeof(PetscScalar));CHKERRQ(ierr);
46ca92afb2SStefano Zampini   } else {
47ca92afb2SStefano Zampini     ierr = VecGetArray(v,&array);CHKERRQ(ierr);
48ca92afb2SStefano Zampini     array2 = array;
49ca92afb2SStefano Zampini   }
50ca92afb2SStefano Zampini   if (!sol) { /* change rhs */
51ca92afb2SStefano Zampini     PetscInt n;
52ca92afb2SStefano Zampini     for (n=0;n<ctx->benign_n;n++) {
53ca92afb2SStefano Zampini       PetscScalar    sum = 0.;
54ca92afb2SStefano Zampini       const PetscInt *cols;
55ca92afb2SStefano Zampini       PetscInt       nz,i;
56ca92afb2SStefano Zampini 
57ca92afb2SStefano Zampini       ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr);
58ca92afb2SStefano Zampini       ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr);
59ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) sum += array[cols[i]];
60ca92afb2SStefano Zampini       sum = -sum/nz;
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;
64ca92afb2SStefano Zampini       ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr);
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;
72ca92afb2SStefano Zampini       ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr);
73ca92afb2SStefano Zampini       ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr);
74ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) sum += array[cols[i]];
75ca92afb2SStefano Zampini       sum = -sum/nz;
76ca92afb2SStefano Zampini       for (i=0;i<nz-1;i++) array2[cols[i]] += sum;
77ca92afb2SStefano Zampini       array2[cols[nz-1]] = ctx->benign_save_vals[n];
78ca92afb2SStefano Zampini       ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr);
79ca92afb2SStefano Zampini     }
80ca92afb2SStefano Zampini   }
81ca92afb2SStefano Zampini   if (v2) {
82ca92afb2SStefano Zampini     ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr);
83*72b8c272SStefano Zampini     ierr = VecRestoreArray(v2,&array2);CHKERRQ(ierr);
84ca92afb2SStefano Zampini   } else {
85ca92afb2SStefano Zampini     ierr = VecRestoreArray(v,&array);CHKERRQ(ierr);
86ca92afb2SStefano Zampini   }
87ca92afb2SStefano Zampini   PetscFunctionReturn(0);
88ca92afb2SStefano Zampini }
89ca92afb2SStefano Zampini 
90d62866d3SStefano Zampini #undef __FUNCT__
91df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Solve_Private"
92df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Solve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose, PetscBool full)
93d62866d3SStefano Zampini {
94df4d28bfSStefano Zampini   PCBDDCReuseSolvers ctx;
95683d3df6SStefano Zampini   PetscBool          copy = PETSC_FALSE;
96d62866d3SStefano Zampini   PetscErrorCode     ierr;
97d62866d3SStefano Zampini 
98d62866d3SStefano Zampini   PetscFunctionBegin;
99d62866d3SStefano Zampini   ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr);
100683d3df6SStefano Zampini   if (full) {
101d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
102d62866d3SStefano Zampini     ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr);
103d62866d3SStefano Zampini #endif
104683d3df6SStefano Zampini     copy = ctx->has_vertices;
105d4933d67SStefano Zampini   } else { /* interior solver */
1066dba178dSStefano Zampini #if defined(PETSC_HAVE_MUMPS)
107683d3df6SStefano Zampini     ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr);
1086dba178dSStefano Zampini #endif
109d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
110d4933d67SStefano Zampini     ierr = MatMkl_PardisoSetCntl(ctx->F,70,1);CHKERRQ(ierr);
111d4933d67SStefano Zampini #endif
112683d3df6SStefano Zampini     copy = PETSC_TRUE;
113683d3df6SStefano Zampini   }
114683d3df6SStefano Zampini   /* copy rhs into factored matrix workspace */
115683d3df6SStefano Zampini   if (copy) {
116ca92afb2SStefano Zampini     PetscInt    n;
117df4d28bfSStefano Zampini     PetscScalar *array,*array_solver;
118ca92afb2SStefano Zampini 
119ca92afb2SStefano Zampini     ierr = VecGetLocalSize(rhs,&n);CHKERRQ(ierr);
120683d3df6SStefano Zampini     ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr);
121df4d28bfSStefano Zampini     ierr = VecGetArray(ctx->rhs,&array_solver);CHKERRQ(ierr);
122df4d28bfSStefano Zampini     ierr = PetscMemcpy(array_solver,array,n*sizeof(PetscScalar));CHKERRQ(ierr);
123df4d28bfSStefano Zampini     ierr = VecRestoreArray(ctx->rhs,&array_solver);CHKERRQ(ierr);
124683d3df6SStefano Zampini     ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr);
125683d3df6SStefano Zampini 
126ca92afb2SStefano Zampini     ierr = PCBDDCReuseSolversChangeInterior(ctx,ctx->rhs,NULL,PETSC_FALSE);CHKERRQ(ierr);
127683d3df6SStefano Zampini     if (transpose) {
128683d3df6SStefano Zampini       ierr = MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr);
129683d3df6SStefano Zampini     } else {
130683d3df6SStefano Zampini       ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr);
131683d3df6SStefano Zampini     }
132ca92afb2SStefano Zampini     ierr = PCBDDCReuseSolversChangeInterior(ctx,ctx->sol,NULL,PETSC_TRUE);CHKERRQ(ierr);
133683d3df6SStefano Zampini 
134683d3df6SStefano Zampini     /* get back data to caller worskpace */
135df4d28bfSStefano Zampini     ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr);
136683d3df6SStefano Zampini     ierr = VecGetArray(sol,&array);CHKERRQ(ierr);
137df4d28bfSStefano Zampini     ierr = PetscMemcpy(array,array_solver,n*sizeof(PetscScalar));CHKERRQ(ierr);
138683d3df6SStefano Zampini     ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr);
139df4d28bfSStefano Zampini     ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr);
140683d3df6SStefano Zampini   } else {
141ca92afb2SStefano Zampini     if (ctx->benign_n) {
142ca92afb2SStefano Zampini       ierr = PCBDDCReuseSolversChangeInterior(ctx,rhs,ctx->rhs,PETSC_FALSE);CHKERRQ(ierr);
143ca92afb2SStefano Zampini       if (transpose) {
144ca92afb2SStefano Zampini         ierr = MatSolveTranspose(ctx->F,ctx->rhs,sol);CHKERRQ(ierr);
145ca92afb2SStefano Zampini       } else {
146ca92afb2SStefano Zampini         ierr = MatSolve(ctx->F,ctx->rhs,sol);CHKERRQ(ierr);
147ca92afb2SStefano Zampini       }
148ca92afb2SStefano Zampini       ierr = PCBDDCReuseSolversChangeInterior(ctx,sol,NULL,PETSC_TRUE);CHKERRQ(ierr);
149ca92afb2SStefano Zampini     } else {
150e28d306cSStefano Zampini       if (transpose) {
151e28d306cSStefano Zampini         ierr = MatSolveTranspose(ctx->F,rhs,sol);CHKERRQ(ierr);
152e28d306cSStefano Zampini       } else {
1536816873aSStefano Zampini         ierr = MatSolve(ctx->F,rhs,sol);CHKERRQ(ierr);
154e28d306cSStefano Zampini       }
155683d3df6SStefano Zampini     }
156ca92afb2SStefano Zampini   }
157d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
158d4933d67SStefano Zampini   ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr);
159d4933d67SStefano Zampini #endif
160d62866d3SStefano Zampini   PetscFunctionReturn(0);
161d62866d3SStefano Zampini }
162d62866d3SStefano Zampini 
163d62866d3SStefano Zampini #undef __FUNCT__
164df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Correction"
165df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC pc, Vec rhs, Vec sol)
166e28d306cSStefano Zampini {
167e28d306cSStefano Zampini   PetscErrorCode   ierr;
168e28d306cSStefano Zampini 
169e28d306cSStefano Zampini   PetscFunctionBegin;
170df4d28bfSStefano Zampini   ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr);
171e28d306cSStefano Zampini   PetscFunctionReturn(0);
172e28d306cSStefano Zampini }
173e28d306cSStefano Zampini 
174e28d306cSStefano Zampini #undef __FUNCT__
175df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_CorrectionTranspose"
176df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_CorrectionTranspose(PC pc, Vec rhs, Vec sol)
177e28d306cSStefano Zampini {
178e28d306cSStefano Zampini   PetscErrorCode   ierr;
179e28d306cSStefano Zampini 
180e28d306cSStefano Zampini   PetscFunctionBegin;
181df4d28bfSStefano Zampini   ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr);
182683d3df6SStefano Zampini   PetscFunctionReturn(0);
183683d3df6SStefano Zampini }
184683d3df6SStefano Zampini 
185683d3df6SStefano Zampini #undef __FUNCT__
186df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Interior"
187df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC pc, Vec rhs, Vec sol)
188683d3df6SStefano Zampini {
189683d3df6SStefano Zampini   PetscErrorCode   ierr;
190683d3df6SStefano Zampini 
191683d3df6SStefano Zampini   PetscFunctionBegin;
192df4d28bfSStefano Zampini   ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr);
193683d3df6SStefano Zampini   PetscFunctionReturn(0);
194683d3df6SStefano Zampini }
195683d3df6SStefano Zampini 
196683d3df6SStefano Zampini #undef __FUNCT__
197df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_InteriorTranspose"
198df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_InteriorTranspose(PC pc, Vec rhs, Vec sol)
199683d3df6SStefano Zampini {
200683d3df6SStefano Zampini   PetscErrorCode   ierr;
201683d3df6SStefano Zampini 
202683d3df6SStefano Zampini   PetscFunctionBegin;
203df4d28bfSStefano Zampini   ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_FALSE);CHKERRQ(ierr);
204e28d306cSStefano Zampini   PetscFunctionReturn(0);
205e28d306cSStefano Zampini }
206e28d306cSStefano Zampini 
207e28d306cSStefano Zampini #undef __FUNCT__
208df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolversReset"
209df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolversReset(PCBDDCReuseSolvers reuse)
210d62866d3SStefano Zampini {
211ca92afb2SStefano Zampini   PetscInt       i;
212d62866d3SStefano Zampini   PetscErrorCode ierr;
213d62866d3SStefano Zampini 
214d62866d3SStefano Zampini   PetscFunctionBegin;
215d62866d3SStefano Zampini   ierr = MatDestroy(&reuse->F);CHKERRQ(ierr);
216d62866d3SStefano Zampini   ierr = VecDestroy(&reuse->sol);CHKERRQ(ierr);
217d62866d3SStefano Zampini   ierr = VecDestroy(&reuse->rhs);CHKERRQ(ierr);
218d62866d3SStefano Zampini   ierr = PCDestroy(&reuse->interior_solver);CHKERRQ(ierr);
219d62866d3SStefano Zampini   ierr = PCDestroy(&reuse->correction_solver);CHKERRQ(ierr);
22053892102SStefano Zampini   ierr = ISDestroy(&reuse->is_R);CHKERRQ(ierr);
22153892102SStefano Zampini   ierr = ISDestroy(&reuse->is_B);CHKERRQ(ierr);
222d62866d3SStefano Zampini   ierr = VecScatterDestroy(&reuse->correction_scatter_B);CHKERRQ(ierr);
22353892102SStefano Zampini   ierr = VecDestroy(&reuse->sol_B);CHKERRQ(ierr);
22453892102SStefano Zampini   ierr = VecDestroy(&reuse->rhs_B);CHKERRQ(ierr);
225ca92afb2SStefano Zampini   for (i=0;i<reuse->benign_n;i++) {
226ca92afb2SStefano Zampini     ierr = ISDestroy(&reuse->benign_zerodiag_subs[i]);CHKERRQ(ierr);
227ca92afb2SStefano Zampini   }
228ca92afb2SStefano Zampini   ierr = PetscFree(reuse->benign_zerodiag_subs);CHKERRQ(ierr);
229ca92afb2SStefano Zampini   ierr = PetscFree(reuse->benign_save_vals);CHKERRQ(ierr);
230d62866d3SStefano Zampini   PetscFunctionReturn(0);
231d62866d3SStefano Zampini }
232d5574798SStefano Zampini 
233d5574798SStefano Zampini #undef __FUNCT__
2343202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur"
2355ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S)
2363202ece2SStefano Zampini {
2373202ece2SStefano Zampini   Mat            B, C, D, Bd, Cd, AinvBd;
2383202ece2SStefano Zampini   KSP            ksp;
2393202ece2SStefano Zampini   PC             pc;
2403202ece2SStefano Zampini   PetscBool      isLU, isILU, isCHOL, Bdense, Cdense;
2413202ece2SStefano Zampini   PetscReal      fill = 2.0;
242f11841e3SStefano Zampini   PetscInt       n_I;
2433202ece2SStefano Zampini   PetscMPIInt    size;
2443202ece2SStefano Zampini   PetscErrorCode ierr;
2453202ece2SStefano Zampini 
2463202ece2SStefano Zampini   PetscFunctionBegin;
2473202ece2SStefano Zampini   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr);
2483202ece2SStefano Zampini   if (size != 1) {
2493202ece2SStefano Zampini     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices");
2503202ece2SStefano Zampini   }
251f11841e3SStefano Zampini   if (reuse == MAT_REUSE_MATRIX) {
252f11841e3SStefano Zampini     PetscBool Sdense;
253f11841e3SStefano Zampini 
254f11841e3SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr);
255f11841e3SStefano Zampini     if (!Sdense) {
256683d3df6SStefano Zampini       SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should be dense");
257f11841e3SStefano Zampini     }
258f11841e3SStefano Zampini   }
2593202ece2SStefano Zampini   ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr);
2603202ece2SStefano Zampini   ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr);
2613202ece2SStefano Zampini   ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr);
2623202ece2SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr);
2633202ece2SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr);
2643202ece2SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr);
2653202ece2SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr);
2663202ece2SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr);
267f11841e3SStefano Zampini   ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr);
268f11841e3SStefano Zampini   if (n_I) {
2693202ece2SStefano Zampini     if (!Bdense) {
2703202ece2SStefano Zampini       ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr);
2713202ece2SStefano Zampini     } else {
2723202ece2SStefano Zampini       Bd = B;
2733202ece2SStefano Zampini     }
2743202ece2SStefano Zampini 
2753202ece2SStefano Zampini     if (isLU || isILU || isCHOL) {
2763202ece2SStefano Zampini       Mat fact;
2773202ece2SStefano Zampini       ierr = KSPSetUp(ksp);CHKERRQ(ierr);
2783202ece2SStefano Zampini       ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr);
2793202ece2SStefano Zampini       ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr);
2803202ece2SStefano Zampini       ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr);
2813202ece2SStefano Zampini     } else {
28207b1e237SStefano Zampini       PetscBool ex = PETSC_TRUE;
28307b1e237SStefano Zampini 
28407b1e237SStefano Zampini       if (ex) {
2853202ece2SStefano Zampini         Mat Ainvd;
2863202ece2SStefano Zampini 
2873202ece2SStefano Zampini         ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr);
2883202ece2SStefano Zampini         ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr);
2893202ece2SStefano Zampini         ierr = MatDestroy(&Ainvd);CHKERRQ(ierr);
29007b1e237SStefano Zampini       } else {
29107b1e237SStefano Zampini         Vec         sol,rhs;
29207b1e237SStefano Zampini         PetscScalar *arrayrhs,*arraysol;
29307b1e237SStefano Zampini         PetscInt    i,nrhs,n;
29407b1e237SStefano Zampini 
29507b1e237SStefano Zampini         ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr);
29607b1e237SStefano Zampini         ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr);
29707b1e237SStefano Zampini         ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr);
29807b1e237SStefano Zampini         ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr);
29907b1e237SStefano Zampini         ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr);
30007b1e237SStefano Zampini         ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr);
30107b1e237SStefano Zampini         for (i=0;i<nrhs;i++) {
30207b1e237SStefano Zampini           ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr);
30307b1e237SStefano Zampini           ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr);
30407b1e237SStefano Zampini           ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr);
30507b1e237SStefano Zampini           ierr = VecResetArray(rhs);CHKERRQ(ierr);
30607b1e237SStefano Zampini           ierr = VecResetArray(sol);CHKERRQ(ierr);
30707b1e237SStefano Zampini         }
30807b1e237SStefano Zampini         ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr);
30907b1e237SStefano Zampini         ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr);
31007b1e237SStefano Zampini       }
3113202ece2SStefano Zampini     }
3125ec10c6aSStefano Zampini     if (!Bdense & !issym) {
3133202ece2SStefano Zampini       ierr = MatDestroy(&Bd);CHKERRQ(ierr);
3143202ece2SStefano Zampini     }
3155ec10c6aSStefano Zampini 
3165ec10c6aSStefano Zampini     if (!issym) {
3173202ece2SStefano Zampini       if (!Cdense) {
3183202ece2SStefano Zampini         ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr);
3193202ece2SStefano Zampini       } else {
3203202ece2SStefano Zampini         Cd = C;
3213202ece2SStefano Zampini       }
3225ec10c6aSStefano Zampini       ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr);
3233202ece2SStefano Zampini       if (!Cdense) {
3243202ece2SStefano Zampini         ierr = MatDestroy(&Cd);CHKERRQ(ierr);
3253202ece2SStefano Zampini       }
3265ec10c6aSStefano Zampini     } else {
3275ec10c6aSStefano Zampini       ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr);
3285ec10c6aSStefano Zampini       if (!Bdense) {
3295ec10c6aSStefano Zampini         ierr = MatDestroy(&Bd);CHKERRQ(ierr);
3305ec10c6aSStefano Zampini       }
3315ec10c6aSStefano Zampini     }
3325ec10c6aSStefano Zampini     ierr = MatDestroy(&AinvBd);CHKERRQ(ierr);
333f11841e3SStefano Zampini   }
3343202ece2SStefano Zampini 
3353202ece2SStefano Zampini   if (D) {
3363202ece2SStefano Zampini     Mat       Dd;
3373202ece2SStefano Zampini     PetscBool Ddense;
3383202ece2SStefano Zampini 
3393202ece2SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr);
3403202ece2SStefano Zampini     if (!Ddense) {
3413202ece2SStefano Zampini       ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr);
3423202ece2SStefano Zampini     } else {
3433202ece2SStefano Zampini       Dd = D;
3443202ece2SStefano Zampini     }
345f11841e3SStefano Zampini     if (n_I) {
3463202ece2SStefano Zampini       ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
347f11841e3SStefano Zampini     } else {
348f11841e3SStefano Zampini       if (reuse == MAT_INITIAL_MATRIX) {
349f11841e3SStefano Zampini         ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr);
350f11841e3SStefano Zampini       } else {
351f11841e3SStefano Zampini         ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
352f11841e3SStefano Zampini       }
353f11841e3SStefano Zampini     }
3543202ece2SStefano Zampini     if (!Ddense) {
3553202ece2SStefano Zampini       ierr = MatDestroy(&Dd);CHKERRQ(ierr);
3563202ece2SStefano Zampini     }
3573202ece2SStefano Zampini   } else {
3583202ece2SStefano Zampini     ierr = MatScale(*S,-1.0);CHKERRQ(ierr);
3593202ece2SStefano Zampini   }
3603202ece2SStefano Zampini   PetscFunctionReturn(0);
3613202ece2SStefano Zampini }
36234a97f8cSStefano Zampini 
36334a97f8cSStefano Zampini #undef __FUNCT__
3641580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp"
365*72b8c272SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscBool exact_schur, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool faster_deluxe, 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)
366b1b3d7a2SStefano Zampini {
367be83ff47SStefano Zampini   Mat                    F,A_II,A_IB,A_BI,A_BB,AE_II;
368be83ff47SStefano Zampini   Mat                    S_all;
3695a94c880SStefano Zampini   Mat                    global_schur_subsets,work_mat,*submats;
3705db18549SStefano Zampini   ISLocalToGlobalMapping l2gmap_subsets;
371*72b8c272SStefano Zampini   IS                     is_I,is_I_layer,*change_primal_sub;
372dc456d91SStefano Zampini   IS                     all_subsets,all_subsets_mult,all_subsets_n;
373dc456d91SStefano Zampini   PetscInt               *nnz,*all_local_idx_N;
374dc456d91SStefano Zampini   PetscInt               *auxnum1,*auxnum2;
3755a95e1ceSStefano Zampini   PetscInt               i,subset_size,max_subset_size;
376683d3df6SStefano Zampini   PetscInt               n_B,extra,local_size,global_size;
37708122e43SStefano Zampini   PetscBLASInt           B_N,B_ierr,B_lwork,*pivots;
37806a4b1faSStefano Zampini   PetscScalar            *Bwork;
3795a95e1ceSStefano Zampini   PetscSubcomm           subcomm;
3805a95e1ceSStefano Zampini   PetscMPIInt            color,rank;
3815a95e1ceSStefano Zampini   MPI_Comm               comm_n;
38206a4e24aSStefano Zampini   PetscBool              use_getr = PETSC_FALSE;
383b1b3d7a2SStefano Zampini   PetscErrorCode         ierr;
384b1b3d7a2SStefano Zampini 
385b1b3d7a2SStefano Zampini   PetscFunctionBegin;
386a64f4aa4SStefano Zampini   /* update info in sub_schurs */
387a64f4aa4SStefano Zampini   ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr);
388a64f4aa4SStefano Zampini   ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr);
3893301b35fSStefano Zampini   sub_schurs->is_hermitian = PETSC_FALSE;
3903301b35fSStefano Zampini   sub_schurs->is_posdef = PETSC_FALSE;
391a64f4aa4SStefano Zampini   if (Ain) {
392a64f4aa4SStefano Zampini     PetscBool isseqaij;
3933301b35fSStefano Zampini     /* determine if we are dealing with hermitian positive definite problems */
3943301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX)
3953301b35fSStefano Zampini     if (Ain->symmetric_set) {
3963301b35fSStefano Zampini       sub_schurs->is_hermitian = Ain->symmetric;
3973301b35fSStefano Zampini     }
3983301b35fSStefano Zampini #else
3993301b35fSStefano Zampini     if (Ain->hermitian_set) {
4003301b35fSStefano Zampini       sub_schurs->is_hermitian = Ain->hermitian;
4013301b35fSStefano Zampini     }
4023301b35fSStefano Zampini #endif
4033301b35fSStefano Zampini     if (Ain->spd_set) {
4043301b35fSStefano Zampini       sub_schurs->is_posdef = Ain->spd;
4053301b35fSStefano Zampini     }
406a64f4aa4SStefano Zampini 
4073301b35fSStefano Zampini     /* check */
408a64f4aa4SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
4091dac5d76SStefano Zampini     if (compute_Stilda && (!Ain->hermitian_set || !Ain->spd_set)) { /* these are lazy checks, maybe I should throw an error if they are not set */
4103301b35fSStefano Zampini       PetscInt lsize;
4113301b35fSStefano Zampini 
4123301b35fSStefano Zampini       ierr = MatGetSize(Ain,&lsize,NULL);CHKERRQ(ierr);
4133301b35fSStefano Zampini       if (lsize) {
4143301b35fSStefano Zampini         PetscScalar val;
4153301b35fSStefano Zampini         PetscReal   norm;
4163301b35fSStefano Zampini         Vec         vec1,vec2,vec3;
4173301b35fSStefano Zampini 
4183301b35fSStefano Zampini         ierr = MatCreateVecs(Ain,&vec1,&vec2);CHKERRQ(ierr);
4193301b35fSStefano Zampini         ierr = VecDuplicate(vec1,&vec3);CHKERRQ(ierr);
4203301b35fSStefano Zampini         ierr = VecSetRandom(vec1,NULL);CHKERRQ(ierr);
4213301b35fSStefano Zampini         ierr = MatMult(Ain,vec1,vec2);CHKERRQ(ierr);
4223301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX)
4233301b35fSStefano Zampini         ierr = MatMultTranspose(Ain,vec1,vec3);CHKERRQ(ierr);
4243301b35fSStefano Zampini #else
4253301b35fSStefano Zampini         ierr = MatMultHermitianTranspose(Ain,vec1,vec3);CHKERRQ(ierr);
4263301b35fSStefano Zampini #endif
4273301b35fSStefano Zampini         ierr = VecAXPY(vec3,-1.0,vec2);CHKERRQ(ierr);
4283301b35fSStefano Zampini         ierr = VecNorm(vec3,NORM_INFINITY,&norm);CHKERRQ(ierr);
4291dac5d76SStefano Zampini         if (!Ain->hermitian_set) {
430862806e4SStefano Zampini           if (norm > PetscSqrtReal(PETSC_SMALL)) {
4313301b35fSStefano Zampini             sub_schurs->is_hermitian = PETSC_FALSE;
4323301b35fSStefano Zampini           } else {
4333301b35fSStefano Zampini             sub_schurs->is_hermitian = PETSC_TRUE;
4343301b35fSStefano Zampini           }
4351dac5d76SStefano Zampini         }
4361dac5d76SStefano Zampini         if (!Ain->spd_set && !benign_trick) {
4373301b35fSStefano Zampini           ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr);
4383301b35fSStefano Zampini           if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE;
43906a4e24aSStefano Zampini         }
4403301b35fSStefano Zampini         ierr = VecDestroy(&vec1);CHKERRQ(ierr);
4413301b35fSStefano Zampini         ierr = VecDestroy(&vec2);CHKERRQ(ierr);
4423301b35fSStefano Zampini         ierr = VecDestroy(&vec3);CHKERRQ(ierr);
4433301b35fSStefano Zampini       } else {
4443301b35fSStefano Zampini         sub_schurs->is_hermitian = PETSC_TRUE;
4453301b35fSStefano Zampini         sub_schurs->is_posdef = PETSC_TRUE;
4463301b35fSStefano Zampini       }
4473301b35fSStefano Zampini     }
448a64f4aa4SStefano Zampini     if (isseqaij) {
449a64f4aa4SStefano Zampini       ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr);
450a64f4aa4SStefano Zampini       sub_schurs->A = Ain;
4513301b35fSStefano Zampini     } else {
452a64f4aa4SStefano Zampini       ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr);
453a64f4aa4SStefano Zampini     }
454a64f4aa4SStefano Zampini   }
4553301b35fSStefano Zampini 
456a64f4aa4SStefano Zampini   ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr);
457a64f4aa4SStefano Zampini   sub_schurs->S = Sin;
458df4d28bfSStefano Zampini   if (sub_schurs->schur_explicit) {
459df4d28bfSStefano Zampini     sub_schurs->schur_explicit = (PetscBool)(!!sub_schurs->A);
460a64f4aa4SStefano Zampini   }
461a64f4aa4SStefano Zampini 
4625a95e1ceSStefano Zampini   /* preliminary checks */
463df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit && compute_Stilda) {
464df4d28bfSStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_PARDISO");
4655a95e1ceSStefano Zampini   }
4665a95e1ceSStefano Zampini 
4675a95e1ceSStefano Zampini   /* restrict work on active processes */
4685a95e1ceSStefano Zampini   color = 0;
4695a95e1ceSStefano Zampini   if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */
4705a95e1ceSStefano Zampini   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr);
4715a95e1ceSStefano Zampini   ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr);
4725a95e1ceSStefano Zampini   ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr);
4735a95e1ceSStefano Zampini   ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr);
4745a95e1ceSStefano Zampini   ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr);
4755a95e1ceSStefano Zampini   ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr);
4765a95e1ceSStefano Zampini   if (!sub_schurs->n_subs) {
4775a95e1ceSStefano Zampini     ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr);
4785a95e1ceSStefano Zampini     PetscFunctionReturn(0);
4795a95e1ceSStefano Zampini   }
48081ea8064SStefano Zampini   /* ierr = PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL);CHKERRQ(ierr); */
4815a95e1ceSStefano Zampini 
482b1b3d7a2SStefano Zampini   /* get Schur complement matrices */
483df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) {
484a64f4aa4SStefano Zampini     Mat       tA_IB,tA_BI,tA_BB;
4853301b35fSStefano Zampini     PetscBool isseqsbaij;
486a64f4aa4SStefano Zampini     ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr);
4873301b35fSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr);
4883301b35fSStefano Zampini     if (isseqsbaij) {
489a64f4aa4SStefano Zampini       ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr);
490a64f4aa4SStefano Zampini       ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr);
491a64f4aa4SStefano Zampini       ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr);
492a64f4aa4SStefano Zampini     } else {
493a64f4aa4SStefano Zampini       ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr);
494a64f4aa4SStefano Zampini       A_BB = tA_BB;
495a64f4aa4SStefano Zampini       ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr);
496a64f4aa4SStefano Zampini       A_IB = tA_IB;
497a64f4aa4SStefano Zampini       ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr);
498a64f4aa4SStefano Zampini       A_BI = tA_BI;
499f11841e3SStefano Zampini     }
500a58a30b4SStefano Zampini   } else {
5015a95e1ceSStefano Zampini     A_II = NULL;
5025a95e1ceSStefano Zampini     A_IB = NULL;
5035a95e1ceSStefano Zampini     A_BI = NULL;
5045a95e1ceSStefano Zampini     A_BB = NULL;
505b1b3d7a2SStefano Zampini   }
5065a95e1ceSStefano Zampini   S_all = NULL;
507b1b3d7a2SStefano Zampini 
508b1b3d7a2SStefano Zampini   /* determine interior problems */
5093dc780c3SStefano Zampini   ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr);
5103dc780c3SStefano Zampini   if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */
511b1b3d7a2SStefano Zampini     PetscBT                touched;
512b1b3d7a2SStefano Zampini     const PetscInt*        idx_B;
513b1b3d7a2SStefano Zampini     PetscInt               n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering;
514b1b3d7a2SStefano Zampini 
5153dc780c3SStefano Zampini     if (xadj == NULL || adjncy == NULL) {
5163dc780c3SStefano Zampini       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency");
5173dc780c3SStefano Zampini     }
518b1b3d7a2SStefano Zampini     /* get sizes */
519b1b3d7a2SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr);
520b1b3d7a2SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr);
521b1b3d7a2SStefano Zampini 
522b1b3d7a2SStefano Zampini     ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr);
523b1b3d7a2SStefano Zampini     ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr);
524b1b3d7a2SStefano Zampini     ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr);
525b1b3d7a2SStefano Zampini 
526b1b3d7a2SStefano Zampini     /* all boundary dofs must be skipped when adding layers */
527b1b3d7a2SStefano Zampini     ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr);
528b1b3d7a2SStefano Zampini     for (j=0;j<n_B;j++) {
529b1b3d7a2SStefano Zampini       ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr);
530b1b3d7a2SStefano Zampini     }
531b1b3d7a2SStefano Zampini     ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr);
532b1b3d7a2SStefano Zampini     ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr);
533b1b3d7a2SStefano Zampini 
534b1b3d7a2SStefano Zampini     /* add prescribed number of layers of dofs */
535b1b3d7a2SStefano Zampini     n_local_dofs = n_B;
536b1b3d7a2SStefano Zampini     n_prev_added = n_B;
537b1b3d7a2SStefano Zampini     for (layer=0;layer<nlayers;layer++) {
538b1b3d7a2SStefano Zampini       PetscInt n_added;
539b1b3d7a2SStefano Zampini       if (n_local_dofs == n_I+n_B) break;
540b1b3d7a2SStefano Zampini       if (n_local_dofs > n_I+n_B) {
541b1b3d7a2SStefano Zampini         SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+n_B);
542b1b3d7a2SStefano Zampini       }
543b1b3d7a2SStefano Zampini       ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr);
544b1b3d7a2SStefano Zampini       n_prev_added = n_added;
545b1b3d7a2SStefano Zampini       n_local_dofs += n_added;
546b1b3d7a2SStefano Zampini       if (!n_added) break;
547b1b3d7a2SStefano Zampini     }
548b1b3d7a2SStefano Zampini     ierr = PetscBTDestroy(&touched);CHKERRQ(ierr);
549b1b3d7a2SStefano Zampini 
550883469d8SStefano Zampini     /* IS for I layer dofs in original numbering */
551a9b99552SStefano Zampini     ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&is_I_layer);CHKERRQ(ierr);
552b1b3d7a2SStefano Zampini     ierr = PetscFree(local_numbering);CHKERRQ(ierr);
553a9b99552SStefano Zampini     ierr = ISSort(is_I_layer);CHKERRQ(ierr);
554883469d8SStefano Zampini     /* IS for I layer dofs in I numbering */
555df4d28bfSStefano Zampini     if (!sub_schurs->schur_explicit) {
556b1b3d7a2SStefano Zampini       ISLocalToGlobalMapping ItoNmap;
557b1b3d7a2SStefano Zampini       ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr);
558a9b99552SStefano Zampini       ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr);
559b1b3d7a2SStefano Zampini       ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr);
560b1b3d7a2SStefano Zampini 
561b1b3d7a2SStefano Zampini       /* II block */
562b1b3d7a2SStefano Zampini       ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr);
563b1b3d7a2SStefano Zampini     }
564b1b3d7a2SStefano Zampini   } else {
565b1b3d7a2SStefano Zampini     PetscInt n_I;
566b1b3d7a2SStefano Zampini 
567b1b3d7a2SStefano Zampini     /* IS for I dofs in original numbering */
568b1b3d7a2SStefano Zampini     ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr);
569a9b99552SStefano Zampini     is_I_layer = sub_schurs->is_I;
570b1b3d7a2SStefano Zampini 
571b1b3d7a2SStefano Zampini     /* IS for I dofs in I numbering (strided 1) */
572df4d28bfSStefano Zampini     if (!sub_schurs->schur_explicit) {
573b1b3d7a2SStefano Zampini       ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr);
574b1b3d7a2SStefano Zampini       ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr);
575b1b3d7a2SStefano Zampini 
576b1b3d7a2SStefano Zampini       /* II block is the same */
577b1b3d7a2SStefano Zampini       ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr);
578b1b3d7a2SStefano Zampini       AE_II = A_II;
579b1b3d7a2SStefano Zampini     }
580b1b3d7a2SStefano Zampini   }
5815a95e1ceSStefano Zampini 
582883469d8SStefano Zampini   /* Get info on subset sizes and sum of all subsets sizes */
5835a95e1ceSStefano Zampini   max_subset_size = 0;
584883469d8SStefano Zampini   local_size = 0;
5855a95e1ceSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
5865a95e1ceSStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
5875a95e1ceSStefano Zampini     max_subset_size = PetscMax(subset_size,max_subset_size);
588883469d8SStefano Zampini     local_size += subset_size;
589883469d8SStefano Zampini   }
590883469d8SStefano Zampini 
591883469d8SStefano Zampini   /* Work arrays for local indices */
592883469d8SStefano Zampini   extra = 0;
593683d3df6SStefano Zampini   ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr);
594df4d28bfSStefano Zampini   if (sub_schurs->schur_explicit && is_I_layer) {
595a9b99552SStefano Zampini     ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr);
596883469d8SStefano Zampini   }
597683d3df6SStefano Zampini   ierr = PetscMalloc1(n_B+extra,&all_local_idx_N);CHKERRQ(ierr);
598883469d8SStefano Zampini   if (extra) {
599883469d8SStefano Zampini     const PetscInt *idxs;
600a9b99552SStefano Zampini     ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr);
601883469d8SStefano Zampini     ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr);
602a9b99552SStefano Zampini     ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr);
603883469d8SStefano Zampini   }
604883469d8SStefano Zampini   ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr);
605dc456d91SStefano Zampini   ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr);
606dc456d91SStefano Zampini   ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr);
607883469d8SStefano Zampini 
608883469d8SStefano Zampini   /* Get local indices in local numbering */
609883469d8SStefano Zampini   local_size = 0;
6105a95e1ceSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
611883469d8SStefano Zampini     PetscInt j;
612883469d8SStefano Zampini     const    PetscInt *idxs;
613883469d8SStefano Zampini 
6145a95e1ceSStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
6155a95e1ceSStefano Zampini     ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
616eb595f79SStefano Zampini     /* start (smallest in global ordering) and multiplicity */
617eb595f79SStefano Zampini     auxnum1[i] = idxs[0];
618eb595f79SStefano Zampini     auxnum2[i] = subset_size;
619883469d8SStefano Zampini     /* subset indices in local numbering */
620883469d8SStefano Zampini     ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr);
6215a95e1ceSStefano Zampini     ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
622883469d8SStefano Zampini     for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size;
623883469d8SStefano Zampini     local_size += subset_size;
624883469d8SStefano Zampini   }
625883469d8SStefano Zampini 
6265a95e1ceSStefano Zampini   /* allocate extra workspace needed only for GETRI */
627d2627357SStefano Zampini   Bwork = NULL;
628d2627357SStefano Zampini   pivots = NULL;
62906a4e24aSStefano Zampini   if (local_size && !benign_trick && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) {
630d2627357SStefano Zampini     PetscScalar lwork;
631d2627357SStefano Zampini 
63206a4e24aSStefano Zampini     use_getr = PETSC_TRUE;
633d2627357SStefano Zampini     B_lwork = -1;
634d2627357SStefano Zampini     ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr);
635d2627357SStefano Zampini     ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
636d2627357SStefano Zampini     PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr));
637d2627357SStefano Zampini     ierr = PetscFPTrapPop();CHKERRQ(ierr);
638d2627357SStefano Zampini     if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr);
639d2627357SStefano Zampini     ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr);
640d2627357SStefano Zampini     ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr);
641d2627357SStefano Zampini   }
642d2627357SStefano Zampini 
643d2627357SStefano Zampini   /* prepare parallel matrices for summing up properly schurs on subsets */
644dc456d91SStefano Zampini   ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr);
645dc456d91SStefano Zampini   ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr);
646dc456d91SStefano Zampini   ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr);
647dc456d91SStefano Zampini   ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr);
648dc456d91SStefano Zampini   ierr = PCBDDCSubsetNumbering(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr);
649dc456d91SStefano Zampini   ierr = ISDestroy(&all_subsets);CHKERRQ(ierr);
650dc456d91SStefano Zampini   ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr);
651dc456d91SStefano Zampini   ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr);
652dc456d91SStefano Zampini   if (i != local_size) {
653dc456d91SStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %d != %d",i,local_size);
654eb595f79SStefano Zampini   }
655dc456d91SStefano Zampini   ierr = ISLocalToGlobalMappingCreateIS(all_subsets_n,&l2gmap_subsets);CHKERRQ(ierr);
656e176bc59SStefano Zampini   ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,NULL,&work_mat);CHKERRQ(ierr);
657d2627357SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr);
658d2627357SStefano Zampini   ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr);
659d2627357SStefano Zampini   ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr);
660d2627357SStefano Zampini   ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr);
6612972d61bSStefano Zampini 
6625a95e1ceSStefano Zampini   /* subset indices in local boundary numbering */
6635a95e1ceSStefano Zampini   if (!sub_schurs->is_Ej_all) {
6645a95e1ceSStefano Zampini     PetscInt *all_local_idx_B;
6655a95e1ceSStefano Zampini 
6665a95e1ceSStefano Zampini     ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr);
6675a95e1ceSStefano Zampini     ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr);
6685a95e1ceSStefano Zampini     if (subset_size != local_size) {
6695a95e1ceSStefano Zampini       SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size);
6705a95e1ceSStefano Zampini     }
6715a95e1ceSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr);
672b1b3d7a2SStefano Zampini   }
673b1b3d7a2SStefano Zampini 
674*72b8c272SStefano Zampini   change_primal_sub = NULL;
675*72b8c272SStefano Zampini   if (change) {
676*72b8c272SStefano Zampini     ISLocalToGlobalMapping BtoS;
677*72b8c272SStefano Zampini     IS                     change_primal_B;
678*72b8c272SStefano Zampini     IS                     change_primal_all;
679*72b8c272SStefano Zampini 
680*72b8c272SStefano Zampini     ierr = PetscMalloc1(sub_schurs->n_subs,&change_primal_sub);CHKERRQ(ierr);
681*72b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
682*72b8c272SStefano Zampini       ISLocalToGlobalMapping NtoS;
683*72b8c272SStefano Zampini       ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_subs[i],&NtoS);CHKERRQ(ierr);
684*72b8c272SStefano Zampini       ierr = ISGlobalToLocalMappingApplyIS(NtoS,IS_GTOLM_DROP,change_primal,&change_primal_sub[i]);CHKERRQ(ierr);
685*72b8c272SStefano Zampini       ierr = ISLocalToGlobalMappingDestroy(&NtoS);CHKERRQ(ierr);
686*72b8c272SStefano Zampini     }
687*72b8c272SStefano Zampini     ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,change_primal,&change_primal_B);CHKERRQ(ierr);
688*72b8c272SStefano Zampini     ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_Ej_all,&BtoS);CHKERRQ(ierr);
689*72b8c272SStefano Zampini     ierr = ISGlobalToLocalMappingApplyIS(BtoS,IS_GTOLM_DROP,change_primal_B,&change_primal_all);CHKERRQ(ierr);
690*72b8c272SStefano Zampini     ierr = ISLocalToGlobalMappingDestroy(&BtoS);CHKERRQ(ierr);
691*72b8c272SStefano Zampini     ierr = ISDestroy(&change_primal_B);CHKERRQ(ierr);
692*72b8c272SStefano Zampini     if (!sub_schurs->change) {
693*72b8c272SStefano Zampini       ierr = PetscCalloc1(sub_schurs->n_subs,&sub_schurs->change);CHKERRQ(ierr);
694*72b8c272SStefano Zampini     }
695*72b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
696*72b8c272SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
697*72b8c272SStefano Zampini       if (!sub_schurs->change[i]) {
698*72b8c272SStefano Zampini         Mat change_sub;
699*72b8c272SStefano Zampini 
700*72b8c272SStefano Zampini         ierr = KSPCreate(PETSC_COMM_SELF,&sub_schurs->change[i]);CHKERRQ(ierr);
701*72b8c272SStefano Zampini         ierr = KSPSetType(sub_schurs->change[i],KSPPREONLY);CHKERRQ(ierr);
702*72b8c272SStefano Zampini         if (!sub_schurs->change_with_qr) {
703*72b8c272SStefano Zampini           ierr = MatGetSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr);
704*72b8c272SStefano Zampini         } else {
705*72b8c272SStefano Zampini           Mat change_subt;
706*72b8c272SStefano Zampini           ierr = MatGetSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_subt);CHKERRQ(ierr);
707*72b8c272SStefano Zampini           ierr = MatConvert(change_subt,MATSEQDENSE,MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr);
708*72b8c272SStefano Zampini           ierr = MatDestroy(&change_subt);CHKERRQ(ierr);
709*72b8c272SStefano Zampini         }
710*72b8c272SStefano Zampini         ierr = KSPSetOperators(sub_schurs->change[i],change_sub,change_sub);CHKERRQ(ierr);
711*72b8c272SStefano Zampini         ierr = MatDestroy(&change_sub);CHKERRQ(ierr);
712*72b8c272SStefano Zampini         ierr = KSPSetOptionsPrefix(sub_schurs->change[i],"sub_schurs_change_");CHKERRQ(ierr);
713*72b8c272SStefano Zampini       }
714*72b8c272SStefano Zampini     }
715*72b8c272SStefano Zampini     ierr = ISDestroy(&change_primal_all);CHKERRQ(ierr);
716*72b8c272SStefano Zampini   }
717*72b8c272SStefano Zampini 
7185a95e1ceSStefano Zampini   /* Local matrix of all local Schur on subsets (transposed) */
7195a95e1ceSStefano Zampini   if (!sub_schurs->S_Ej_all) {
7205a95e1ceSStefano Zampini     ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr);
7215a95e1ceSStefano Zampini     ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr);
7225a95e1ceSStefano Zampini     ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr);
7235a95e1ceSStefano Zampini     ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr);
724aa83b6aeSStefano Zampini   }
725b1b3d7a2SStefano Zampini 
7265a95e1ceSStefano Zampini   /* Compute Schur complements explicitly */
727be83ff47SStefano Zampini   F = NULL;
728df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) { /* this code branch is used when MatFactor with Schur complemnent support is not present; it is not very efficient, unless the economic version of the scaling is required */
7295a95e1ceSStefano Zampini     Mat         S_Ej_expl;
7305a95e1ceSStefano Zampini     PetscScalar *work;
7315a95e1ceSStefano Zampini     PetscInt    j,*dummy_idx;
7325a95e1ceSStefano Zampini     PetscBool   Sdense;
7335a95e1ceSStefano Zampini 
7345a95e1ceSStefano Zampini     ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr);
7355a95e1ceSStefano Zampini     local_size = 0;
736b1b3d7a2SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
7375a95e1ceSStefano Zampini       IS  is_subset_B;
7385a95e1ceSStefano Zampini       Mat AE_EE,AE_IE,AE_EI,S_Ej;
7395a95e1ceSStefano Zampini 
7405a95e1ceSStefano Zampini       /* subsets in original and boundary numbering */
7415a95e1ceSStefano Zampini       ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr);
7425a95e1ceSStefano Zampini       /* EE block */
7435a95e1ceSStefano Zampini       ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr);
7445a95e1ceSStefano Zampini       /* IE block */
7455a95e1ceSStefano Zampini       ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr);
7465a95e1ceSStefano Zampini       /* EI block */
7475a95e1ceSStefano Zampini       if (sub_schurs->is_hermitian) {
7485a95e1ceSStefano Zampini         ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr);
7495a95e1ceSStefano Zampini       } else {
7505a95e1ceSStefano Zampini         ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr);
7515a95e1ceSStefano Zampini       }
752a64f4aa4SStefano Zampini       ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr);
7535a95e1ceSStefano Zampini       ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr);
7545a95e1ceSStefano Zampini       ierr = MatDestroy(&AE_EE);CHKERRQ(ierr);
7555a95e1ceSStefano Zampini       ierr = MatDestroy(&AE_IE);CHKERRQ(ierr);
7565a95e1ceSStefano Zampini       ierr = MatDestroy(&AE_EI);CHKERRQ(ierr);
757b1b3d7a2SStefano Zampini       if (AE_II == A_II) { /* we can reuse the same ksp */
758b1b3d7a2SStefano Zampini         KSP ksp;
759b1b3d7a2SStefano Zampini         ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr);
7605a95e1ceSStefano Zampini         ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr);
761b1b3d7a2SStefano Zampini       } else { /* build new ksp object which inherits ksp and pc types from the original one */
762b1b3d7a2SStefano Zampini         KSP       origksp,schurksp;
763b1b3d7a2SStefano Zampini         PC        origpc,schurpc;
764b1b3d7a2SStefano Zampini         KSPType   ksp_type;
765b1b3d7a2SStefano Zampini         PetscInt  n_internal;
7665a95e1ceSStefano Zampini         PetscBool ispcnone;
767b1b3d7a2SStefano Zampini 
768b1b3d7a2SStefano Zampini         ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr);
7695a95e1ceSStefano Zampini         ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr);
770b1b3d7a2SStefano Zampini         ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr);
771b1b3d7a2SStefano Zampini         ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr);
772b1b3d7a2SStefano Zampini         ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr);
773b1b3d7a2SStefano Zampini         ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr);
7745a95e1ceSStefano Zampini         ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr);
7755a95e1ceSStefano Zampini         if (!ispcnone) {
7765a95e1ceSStefano Zampini           PCType pc_type;
777b1b3d7a2SStefano Zampini           ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr);
778b1b3d7a2SStefano Zampini           ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr);
7795a95e1ceSStefano Zampini         } else {
7805a95e1ceSStefano Zampini           ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr);
7815a95e1ceSStefano Zampini         }
782b1b3d7a2SStefano Zampini         ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr);
783b1b3d7a2SStefano Zampini         if (n_internal) { /* UMFPACK gives error with 0 sized problems */
784b1b3d7a2SStefano Zampini           MatSolverPackage solver=NULL;
785b1b3d7a2SStefano Zampini           ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr);
786b1b3d7a2SStefano Zampini           if (solver) {
787b1b3d7a2SStefano Zampini             ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr);
788b1b3d7a2SStefano Zampini           }
789b1b3d7a2SStefano Zampini         }
790b1b3d7a2SStefano Zampini         ierr = KSPSetUp(schurksp);CHKERRQ(ierr);
791b1b3d7a2SStefano Zampini       }
7925a95e1ceSStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
7935a95e1ceSStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr);
7945a95e1ceSStefano Zampini       ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr);
7955a95e1ceSStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr);
7965a95e1ceSStefano Zampini       if (Sdense) {
7975a95e1ceSStefano Zampini         for (j=0;j<subset_size;j++) {
7985a95e1ceSStefano Zampini           dummy_idx[j]=local_size+j;
799b1b3d7a2SStefano Zampini         }
8005a95e1ceSStefano Zampini         ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr);
8015a95e1ceSStefano Zampini       } else {
8025a95e1ceSStefano Zampini         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices");
8035a95e1ceSStefano Zampini       }
8045a95e1ceSStefano Zampini       ierr = MatDestroy(&S_Ej);CHKERRQ(ierr);
805a64f4aa4SStefano Zampini       ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr);
8065a95e1ceSStefano Zampini       local_size += subset_size;
8075a95e1ceSStefano Zampini     }
8085a95e1ceSStefano Zampini     ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr);
809b1b3d7a2SStefano Zampini     /* free */
810b1b3d7a2SStefano Zampini     ierr = ISDestroy(&is_I);CHKERRQ(ierr);
811b1b3d7a2SStefano Zampini     ierr = MatDestroy(&AE_II);CHKERRQ(ierr);
8125a95e1ceSStefano Zampini     ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr);
813883469d8SStefano Zampini   } else {
814be83ff47SStefano Zampini     Mat         A;
815ca92afb2SStefano Zampini     Vec         *benign_AIIm1_ones = NULL;
816ca92afb2SStefano Zampini     IS          is_A_all,*is_p_r = NULL;
8176c3e6151SStefano Zampini     PetscScalar *work,*S_data,*schur_factor,*cs_AIB = NULL, infty = PETSC_MAX_REAL;
8186dba178dSStefano Zampini     PetscInt    n,n_I,*dummy_idx,size_schur,size_active_schur,cum,cum2;
819df4d28bfSStefano Zampini     PetscBool   economic,solver_S,S_lower_triangular = PETSC_FALSE,factor_workaround;
820d4933d67SStefano Zampini     char        solver[256];
821883469d8SStefano Zampini 
822683d3df6SStefano Zampini     /* get sizes */
82381ea8064SStefano Zampini     n_I = 0;
82481ea8064SStefano Zampini     if (is_I_layer) {
825a9b99552SStefano Zampini       ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr);
82681ea8064SStefano Zampini     }
827683d3df6SStefano Zampini     economic = PETSC_FALSE;
828683d3df6SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_I,&cum);CHKERRQ(ierr);
829683d3df6SStefano Zampini     if (cum != n_I) economic = PETSC_TRUE;
830683d3df6SStefano Zampini     ierr = MatGetLocalSize(sub_schurs->A,&n,NULL);CHKERRQ(ierr);
831d62866d3SStefano Zampini 
832683d3df6SStefano Zampini     /* size active schurs does not count any dirichlet or vertex dof on the interface */
833683d3df6SStefano Zampini     size_active_schur = local_size;
834683d3df6SStefano Zampini     cum = n_I+size_active_schur;
835683d3df6SStefano Zampini     if (sub_schurs->is_dir) {
836683d3df6SStefano Zampini       const PetscInt* idxs;
837683d3df6SStefano Zampini       PetscInt        n_dir;
838683d3df6SStefano Zampini 
839683d3df6SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_dir,&n_dir);CHKERRQ(ierr);
840683d3df6SStefano Zampini       ierr = ISGetIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr);
841683d3df6SStefano Zampini       ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_dir*sizeof(PetscInt));CHKERRQ(ierr);
842683d3df6SStefano Zampini       ierr = ISRestoreIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr);
843683d3df6SStefano Zampini       cum += n_dir;
844d62866d3SStefano Zampini     }
845683d3df6SStefano Zampini     factor_workaround = PETSC_FALSE;
846683d3df6SStefano Zampini     /* include the primal vertices in the Schur complement */
847367aa537SStefano Zampini     if (exact_schur && sub_schurs->is_vertices && (compute_Stilda || benign_n)) {
848683d3df6SStefano Zampini       PetscInt n_v;
849683d3df6SStefano Zampini 
850683d3df6SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr);
851683d3df6SStefano Zampini       if (n_v) {
852683d3df6SStefano Zampini         const PetscInt* idxs;
853683d3df6SStefano Zampini 
854683d3df6SStefano Zampini         ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr);
855683d3df6SStefano Zampini         ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_v*sizeof(PetscInt));CHKERRQ(ierr);
856683d3df6SStefano Zampini         ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr);
857683d3df6SStefano Zampini         cum += n_v;
858683d3df6SStefano Zampini         factor_workaround = PETSC_TRUE;
859683d3df6SStefano Zampini       }
860683d3df6SStefano Zampini     }
861683d3df6SStefano Zampini     size_schur = cum - n_I;
862683d3df6SStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,all_local_idx_N,PETSC_OWN_POINTER,&is_A_all);CHKERRQ(ierr);
863683d3df6SStefano Zampini     /* get working mat (TODO: factorize without actually permuting)  */
864683d3df6SStefano Zampini     if (cum == n) {
865683d3df6SStefano Zampini       ierr = ISSetPermutation(is_A_all);CHKERRQ(ierr);
866683d3df6SStefano Zampini       ierr = MatPermute(sub_schurs->A,is_A_all,is_A_all,&A);CHKERRQ(ierr);
867683d3df6SStefano Zampini     } else {
8686816873aSStefano Zampini       ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr);
869683d3df6SStefano Zampini     }
870a9b99552SStefano Zampini     ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr);
871ca92afb2SStefano Zampini 
872ca92afb2SStefano Zampini     /* if we actually change the basis for the pressures, LDL^T factors will use a lot of memory
873367aa537SStefano Zampini        n^2 instead of n^1.5 or something. This is a workaround */
874ca92afb2SStefano Zampini     if (benign_n) {
875ca92afb2SStefano Zampini       Vec                    v;
876ca92afb2SStefano Zampini       ISLocalToGlobalMapping N_to_reor;
877ca92afb2SStefano Zampini       IS                     is_p0,is_p0_p;
878ca92afb2SStefano Zampini 
879ca92afb2SStefano Zampini       ierr = ISLocalToGlobalMappingCreateIS(is_A_all,&N_to_reor);CHKERRQ(ierr);
880ca92afb2SStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,benign_n,benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr);
881ca92afb2SStefano Zampini       ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,is_p0,&is_p0_p);CHKERRQ(ierr);
882ca92afb2SStefano Zampini       ierr = ISDestroy(&is_p0);CHKERRQ(ierr);
883ca92afb2SStefano Zampini       ierr = MatCreateVecs(A,&v,NULL);CHKERRQ(ierr);
884ca92afb2SStefano Zampini       ierr = VecDuplicateVecs(v,benign_n,&benign_AIIm1_ones);CHKERRQ(ierr);
885ca92afb2SStefano Zampini       ierr = PetscMalloc1(size_schur*benign_n,&cs_AIB);CHKERRQ(ierr);
886ca92afb2SStefano Zampini       ierr = PetscMalloc1(benign_n,&is_p_r);CHKERRQ(ierr);
887ca92afb2SStefano Zampini       /* compute colsum of A_IB restricted to pressures */
888ca92afb2SStefano Zampini       for (i=0;i<benign_n;i++) {
889ca92afb2SStefano Zampini         PetscScalar    *array;
890ca92afb2SStefano Zampini         const PetscInt *idxs;
891ca92afb2SStefano Zampini         PetscInt       j,nz;
892ca92afb2SStefano Zampini 
893ca92afb2SStefano Zampini         ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,benign_zerodiag_subs[i],&is_p_r[i]);CHKERRQ(ierr);
894ca92afb2SStefano Zampini         ierr = VecGetArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr);
895ca92afb2SStefano Zampini         ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr);
896ca92afb2SStefano Zampini         ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr);
897ca92afb2SStefano Zampini         for (j=0;j<nz;j++) array[idxs[j]] = 1.;
898ca92afb2SStefano Zampini         ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr);
899ca92afb2SStefano Zampini         ierr = VecRestoreArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr);
900ca92afb2SStefano Zampini         ierr = MatMult(A,benign_AIIm1_ones[i],v);CHKERRQ(ierr);
901ca92afb2SStefano Zampini         ierr = VecGetArray(v,&array);CHKERRQ(ierr);
902ca92afb2SStefano Zampini         for (j=0;j<size_schur;j++) cs_AIB[i*size_schur + j] = array[j+n_I];
903ca92afb2SStefano Zampini         ierr = VecRestoreArray(v,&array);CHKERRQ(ierr);
904ca92afb2SStefano Zampini       }
905ca92afb2SStefano Zampini       ierr = VecDestroy(&v);CHKERRQ(ierr);
906ca92afb2SStefano Zampini       ierr = MatSetOption(A,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr);
907ca92afb2SStefano Zampini       ierr = MatZeroRowsColumnsIS(A,is_p0_p,1.0,NULL,NULL);CHKERRQ(ierr);
908ca92afb2SStefano Zampini       ierr = ISDestroy(&is_p0_p);CHKERRQ(ierr);
909ca92afb2SStefano Zampini       ierr = ISLocalToGlobalMappingDestroy(&N_to_reor);CHKERRQ(ierr);
910ca92afb2SStefano Zampini     }
9113301b35fSStefano Zampini     ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr);
9123301b35fSStefano Zampini     ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr);
9133301b35fSStefano Zampini     ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr);
914df4d28bfSStefano Zampini #if defined(PETSC_HAVE_MUMPS)
915d4933d67SStefano Zampini     ierr = PetscStrcpy(solver,MATSOLVERMUMPS);CHKERRQ(ierr);
916df4d28bfSStefano Zampini #else
917df4d28bfSStefano Zampini     ierr = PetscStrcpy(solver,MATSOLVERMKL_PARDISO);CHKERRQ(ierr);
918df4d28bfSStefano Zampini #endif
919d4933d67SStefano Zampini     ierr = PetscOptionsGetString("sub_schurs_","-mat_solver_package",solver,256,NULL);CHKERRQ(ierr);
920883469d8SStefano Zampini 
921683d3df6SStefano Zampini     /* when using the benign subspace trick, the local Schur complements are SPD */
922d47842beSStefano Zampini     if (benign_trick) sub_schurs->is_posdef = PETSC_TRUE;
923d47842beSStefano Zampini 
924683d3df6SStefano Zampini     if (n_I) { /* TODO add ordering from options */
9255a05ddb0SStefano Zampini       IS is_schur;
9265a05ddb0SStefano Zampini 
9279ab7bb16SStefano Zampini       if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
928d4933d67SStefano Zampini         ierr = MatGetFactor(A,solver,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr);
929883469d8SStefano Zampini       } else {
930d4933d67SStefano Zampini         ierr = MatGetFactor(A,solver,MAT_FACTOR_LU,&F);CHKERRQ(ierr);
931883469d8SStefano Zampini       }
932883469d8SStefano Zampini       /* subsets ordered last */
9335a05ddb0SStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is_schur);CHKERRQ(ierr);
9345a05ddb0SStefano Zampini       ierr = MatFactorSetSchurIS(F,is_schur);CHKERRQ(ierr);
9355a05ddb0SStefano Zampini       ierr = ISDestroy(&is_schur);CHKERRQ(ierr);
936883469d8SStefano Zampini 
937883469d8SStefano Zampini       /* factorization step */
9389ab7bb16SStefano Zampini       if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
939883469d8SStefano Zampini         ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr);
940be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */
941be83ff47SStefano Zampini         ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr);
942be83ff47SStefano Zampini #endif
943e8ade678SStefano Zampini         if (sub_schurs->is_posdef) {
944e8ade678SStefano Zampini           ierr = MatFactorSetSchurComplementSolverType(F,1);CHKERRQ(ierr);
945e8ade678SStefano Zampini         } else {
946e8ade678SStefano Zampini           ierr = MatFactorSetSchurComplementSolverType(F,2);CHKERRQ(ierr);
947e8ade678SStefano Zampini         }
948883469d8SStefano Zampini         ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr);
949a0b0af32SStefano Zampini         S_lower_triangular = PETSC_TRUE;
950883469d8SStefano Zampini       } else {
951883469d8SStefano Zampini         ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr);
952be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */
953be83ff47SStefano Zampini         ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr);
954be83ff47SStefano Zampini #endif
955883469d8SStefano Zampini         ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr);
956883469d8SStefano Zampini       }
957883469d8SStefano Zampini 
958883469d8SStefano Zampini       /* get explicit Schur Complement computed during numeric factorization */
9595a05ddb0SStefano Zampini       ierr = MatFactorGetSchurComplement(F,&S_all);CHKERRQ(ierr);
960d62866d3SStefano Zampini       /* we can reuse the solvers if we are not using the economic version */
961683d3df6SStefano Zampini       reuse_solvers = (PetscBool)(reuse_solvers && !economic);
962683d3df6SStefano Zampini       factor_workaround = (PetscBool)(reuse_solvers && factor_workaround);
963df4d28bfSStefano Zampini       solver_S = PETSC_TRUE;
964ca92afb2SStefano Zampini 
965*72b8c272SStefano Zampini       /* update the Schur complement with the change of basis on the pressures */
966ca92afb2SStefano Zampini       if (benign_n) {
967ca92afb2SStefano Zampini         PetscScalar *S_data;
968ca92afb2SStefano Zampini         Vec         v;
969ca92afb2SStefano Zampini 
970ca92afb2SStefano Zampini         ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr);
971ca92afb2SStefano Zampini         ierr = VecDuplicate(benign_AIIm1_ones[0],&v);CHKERRQ(ierr);
972ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
973ca92afb2SStefano Zampini         ierr = MatMumpsSetIcntl(F,26,0);CHKERRQ(ierr);
974ca92afb2SStefano Zampini #endif
975ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
976ca92afb2SStefano Zampini         ierr = MatMkl_PardisoSetCntl(F,70,1);CHKERRQ(ierr);
977ca92afb2SStefano Zampini #endif
978ca92afb2SStefano Zampini         for (i=0;i<benign_n;i++) {
979ca92afb2SStefano Zampini           PetscScalar    *array,sum = 0.;
980ca92afb2SStefano Zampini           const PetscInt *idxs;
981ca92afb2SStefano Zampini           PetscInt       j,nz;
982ca92afb2SStefano Zampini 
983ca92afb2SStefano Zampini           ierr = VecCopy(benign_AIIm1_ones[i],v);CHKERRQ(ierr);
984ca92afb2SStefano Zampini           ierr = MatSolve(F,v,benign_AIIm1_ones[i]);CHKERRQ(ierr);
985ca92afb2SStefano Zampini           ierr = VecGetArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr);
986ca92afb2SStefano Zampini           ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr);
987ca92afb2SStefano Zampini           ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr);
988ca92afb2SStefano Zampini           for (j=0;j<nz;j++) sum += array[idxs[j]];
989ca92afb2SStefano Zampini           sum -= 1.; /* p0 dof (eliminated) is excluded from the sum */
990ca92afb2SStefano Zampini           ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr);
991ca92afb2SStefano Zampini           ierr = VecRestoreArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr);
992ca92afb2SStefano Zampini           ierr = MatMult(A,benign_AIIm1_ones[i],v);CHKERRQ(ierr);
993ca92afb2SStefano Zampini           ierr = VecGetArray(v,&array);CHKERRQ(ierr);
994ca92afb2SStefano Zampini           /* perform sparse rank-1 updates on symmetric Schur */
995ca92afb2SStefano Zampini           sum  = -sum/(1.*nz*nz);
996ca92afb2SStefano Zampini           ierr = SparseRankOneUpdate(S_data,size_schur,sum,cs_AIB+i*size_schur,cs_AIB+i*size_schur);CHKERRQ(ierr);
997ca92afb2SStefano Zampini           sum  = 1./nz;
998ca92afb2SStefano Zampini           ierr = SparseRankOneUpdate(S_data,size_schur,sum,array+n_I,cs_AIB+i*size_schur);CHKERRQ(ierr);
999ca92afb2SStefano Zampini           ierr = SparseRankOneUpdate(S_data,size_schur,sum,cs_AIB+i*size_schur,array+n_I);CHKERRQ(ierr);
1000ca92afb2SStefano Zampini           ierr = VecRestoreArray(v,&array);CHKERRQ(ierr);
1001ca92afb2SStefano Zampini         }
1002ca92afb2SStefano Zampini         ierr = VecDestroy(&v);CHKERRQ(ierr);
1003ca92afb2SStefano Zampini         ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr);
1004ca92afb2SStefano Zampini       }
1005ca92afb2SStefano Zampini       ierr = PetscFree(cs_AIB);CHKERRQ(ierr);
1006ca92afb2SStefano Zampini       ierr = VecDestroyVecs(benign_n,&benign_AIIm1_ones);CHKERRQ(ierr);
1007a3df083aSStefano Zampini       if (!reuse_solvers) {
1008a3df083aSStefano Zampini         for (i=0;i<benign_n;i++) {
1009a3df083aSStefano Zampini           ierr = ISDestroy(&is_p_r[i]);CHKERRQ(ierr);
1010a3df083aSStefano Zampini         }
1011a3df083aSStefano Zampini         ierr = PetscFree(is_p_r);CHKERRQ(ierr);
1012a3df083aSStefano Zampini       }
1013df4d28bfSStefano Zampini     } else { /* we can't use MatFactor when size_schur == size_of_the_problem */
1014be83ff47SStefano Zampini       ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr);
1015683d3df6SStefano Zampini       reuse_solvers = PETSC_FALSE; /* TODO: why we can't reuse the solvers here? */
1016df4d28bfSStefano Zampini       solver_S = PETSC_FALSE;
1017be83ff47SStefano Zampini     }
1018be83ff47SStefano Zampini 
1019be83ff47SStefano Zampini     if (reuse_solvers) {
1020d62866d3SStefano Zampini       Mat                A_II;
102153892102SStefano Zampini       Vec                vec1_B;
1022df4d28bfSStefano Zampini       PCBDDCReuseSolvers msolv_ctx;
1023d5574798SStefano Zampini 
1024df4d28bfSStefano Zampini       if (sub_schurs->reuse_solver) {
1025df4d28bfSStefano Zampini         ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr);
1026e28d306cSStefano Zampini       } else {
1027df4d28bfSStefano Zampini         ierr = PetscNew(&sub_schurs->reuse_solver);CHKERRQ(ierr);
1028d62866d3SStefano Zampini       }
1029df4d28bfSStefano Zampini       msolv_ctx = sub_schurs->reuse_solver;
1030be83ff47SStefano Zampini       ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr);
1031d5574798SStefano Zampini       ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr);
1032d5574798SStefano Zampini       msolv_ctx->F = F;
1033683d3df6SStefano Zampini       ierr = MatCreateVecs(F,&msolv_ctx->sol,NULL);CHKERRQ(ierr);
1034683d3df6SStefano Zampini       /* currently PETSc has no support for MatSolve(F,x,x), so cheat and let rhs and sol share the same memory */
1035683d3df6SStefano Zampini       {
1036683d3df6SStefano Zampini         PetscScalar *array;
1037683d3df6SStefano Zampini         PetscInt    n;
1038683d3df6SStefano Zampini 
1039683d3df6SStefano Zampini         ierr = VecGetLocalSize(msolv_ctx->sol,&n);CHKERRQ(ierr);
1040683d3df6SStefano Zampini         ierr = VecGetArray(msolv_ctx->sol,&array);CHKERRQ(ierr);
1041683d3df6SStefano Zampini         ierr = VecCreateSeqWithArray(PetscObjectComm((PetscObject)msolv_ctx->sol),1,n,array,&msolv_ctx->rhs);CHKERRQ(ierr);
1042683d3df6SStefano Zampini         ierr = VecRestoreArray(msolv_ctx->sol,&array);CHKERRQ(ierr);
1043683d3df6SStefano Zampini       }
1044683d3df6SStefano Zampini       msolv_ctx->has_vertices = factor_workaround;
1045d62866d3SStefano Zampini 
1046d62866d3SStefano Zampini       /* interior solver */
1047683d3df6SStefano Zampini       ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->interior_solver);CHKERRQ(ierr);
1048d62866d3SStefano Zampini       ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr);
1049d62866d3SStefano Zampini       ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr);
1050d62866d3SStefano Zampini       ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr);
1051df4d28bfSStefano Zampini       ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCReuseSolvers_Interior);CHKERRQ(ierr);
1052df4d28bfSStefano Zampini       ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCReuseSolvers_InteriorTranspose);CHKERRQ(ierr);
1053d62866d3SStefano Zampini 
1054d62866d3SStefano Zampini       /* correction solver */
1055683d3df6SStefano Zampini       ierr = PCCreate(PetscObjectComm((PetscObject)A),&msolv_ctx->correction_solver);CHKERRQ(ierr);
1056d62866d3SStefano Zampini       ierr = PCSetOperators(msolv_ctx->correction_solver,A,A);CHKERRQ(ierr);
1057d62866d3SStefano Zampini       ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr);
1058d62866d3SStefano Zampini       ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr);
1059df4d28bfSStefano Zampini       ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCReuseSolvers_Correction);CHKERRQ(ierr);
1060df4d28bfSStefano Zampini       ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCReuseSolvers_CorrectionTranspose);CHKERRQ(ierr);
106153892102SStefano Zampini 
106253892102SStefano Zampini       /* scatter and vecs for Schur complement solver */
106353892102SStefano Zampini       ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr);
106453892102SStefano Zampini       ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr);
1065683d3df6SStefano Zampini       if (!factor_workaround) {
106653892102SStefano Zampini         ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr);
106753892102SStefano Zampini         ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr);
106853892102SStefano Zampini         ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr);
106953892102SStefano Zampini         msolv_ctx->is_R = is_A_all;
1070683d3df6SStefano Zampini       } else {
1071683d3df6SStefano Zampini         IS              is_B_all;
1072683d3df6SStefano Zampini         const PetscInt* idxs;
1073683d3df6SStefano Zampini         PetscInt        dual,n_v,n;
1074683d3df6SStefano Zampini 
1075683d3df6SStefano Zampini         ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr);
1076683d3df6SStefano Zampini         dual = size_schur - n_v;
1077683d3df6SStefano Zampini         ierr = ISGetLocalSize(is_A_all,&n);CHKERRQ(ierr);
1078683d3df6SStefano Zampini         ierr = ISGetIndices(is_A_all,&idxs);CHKERRQ(ierr);
1079683d3df6SStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),dual,idxs+n_I,PETSC_COPY_VALUES,&is_B_all);CHKERRQ(ierr);
1080683d3df6SStefano Zampini         ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_B_all,&msolv_ctx->is_B);CHKERRQ(ierr);
1081683d3df6SStefano Zampini         ierr = ISDestroy(&is_B_all);CHKERRQ(ierr);
1082683d3df6SStefano Zampini         ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_all),dual,0,1,&is_B_all);CHKERRQ(ierr);
1083683d3df6SStefano Zampini         ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,is_B_all,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr);
1084683d3df6SStefano Zampini         ierr = ISDestroy(&is_B_all);CHKERRQ(ierr);
1085683d3df6SStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),n-n_v,idxs,PETSC_COPY_VALUES,&msolv_ctx->is_R);CHKERRQ(ierr);
1086683d3df6SStefano Zampini         ierr = ISRestoreIndices(is_A_all,&idxs);CHKERRQ(ierr);
1087683d3df6SStefano Zampini       }
1088683d3df6SStefano Zampini       ierr = VecDestroy(&vec1_B);CHKERRQ(ierr);
1089ca92afb2SStefano Zampini 
1090ca92afb2SStefano Zampini       /* communicate benign info to solver context */
1091ca92afb2SStefano Zampini       if (benign_n) {
1092ca92afb2SStefano Zampini         msolv_ctx->benign_n = benign_n;
1093ca92afb2SStefano Zampini         msolv_ctx->benign_zerodiag_subs = is_p_r;
1094ca92afb2SStefano Zampini         ierr = PetscMalloc1(benign_n,&msolv_ctx->benign_save_vals);CHKERRQ(ierr);
1095ca92afb2SStefano Zampini       }
1096d5574798SStefano Zampini     }
109708122e43SStefano Zampini     ierr = MatDestroy(&A);CHKERRQ(ierr);
109853892102SStefano Zampini     ierr = ISDestroy(&is_A_all);CHKERRQ(ierr);
10995db18549SStefano Zampini 
1100be83ff47SStefano Zampini     /* Work arrays */
1101be83ff47SStefano Zampini     ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr);
11025a94c880SStefano Zampini 
11035a94c880SStefano Zampini     /* matrices for adaptive selection */
110412d906b1SStefano Zampini     if (compute_Stilda) {
11055a94c880SStefano Zampini       if (!sub_schurs->sum_S_Ej_tilda_all) {
11065a94c880SStefano Zampini         ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr);
11075a94c880SStefano Zampini         ierr = MatSetSizes(sub_schurs->sum_S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr);
11085a94c880SStefano Zampini         ierr = MatSetType(sub_schurs->sum_S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr);
11095a94c880SStefano Zampini         ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_tilda_all,0,nnz);CHKERRQ(ierr);
11105a94c880SStefano Zampini       }
11115a94c880SStefano Zampini       if (!sub_schurs->sum_S_Ej_inv_all) {
11125a94c880SStefano Zampini         ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr);
11135a94c880SStefano Zampini         ierr = MatSetSizes(sub_schurs->sum_S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr);
11145a94c880SStefano Zampini         ierr = MatSetType(sub_schurs->sum_S_Ej_inv_all,MATAIJ);CHKERRQ(ierr);
11155a94c880SStefano Zampini         ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_inv_all,0,nnz);CHKERRQ(ierr);
11165a94c880SStefano Zampini       }
111712d906b1SStefano Zampini     }
1118d2627357SStefano Zampini 
1119be83ff47SStefano Zampini     /* S_Ej_all */
1120be83ff47SStefano Zampini     cum = cum2 = 0;
1121be83ff47SStefano Zampini     ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr);
112265d8bf0aSStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
112365d8bf0aSStefano Zampini       PetscInt j;
112465d8bf0aSStefano Zampini 
11255a95e1ceSStefano Zampini       /* get S_E */
1126b96c3477SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
1127683d3df6SStefano Zampini       if (S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */
1128be83ff47SStefano Zampini         PetscInt k;
1129be83ff47SStefano Zampini         for (k=0;k<subset_size;k++) {
1130be83ff47SStefano Zampini           for (j=k;j<subset_size;j++) {
1131be83ff47SStefano Zampini             work[k*subset_size+j] = S_data[cum2+k*size_schur+j];
1132683d3df6SStefano Zampini             work[j*subset_size+k] = PetscConj(S_data[cum2+k*size_schur+j]);
1133be83ff47SStefano Zampini           }
1134be83ff47SStefano Zampini         }
113506a4e24aSStefano Zampini       } else { /* just copy to workspace */
1136be83ff47SStefano Zampini         PetscInt k;
1137be83ff47SStefano Zampini         for (k=0;k<subset_size;k++) {
1138be83ff47SStefano Zampini           for (j=0;j<subset_size;j++) {
1139be83ff47SStefano Zampini             work[k*subset_size+j] = S_data[cum2+k*size_schur+j];
1140be83ff47SStefano Zampini           }
1141be83ff47SStefano Zampini         }
11429087bf02SStefano Zampini       }
11435a95e1ceSStefano Zampini       /* insert S_E values */
1144be83ff47SStefano Zampini       for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j;
1145*72b8c272SStefano Zampini       if (change) {
1146*72b8c272SStefano Zampini         Mat change_sub,SEj;
1147*72b8c272SStefano Zampini 
1148*72b8c272SStefano Zampini         /* change basis */
1149*72b8c272SStefano Zampini         ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr);
1150*72b8c272SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr);
1151*72b8c272SStefano Zampini         { /* currently there's no support for PtAP with A dense */
1152*72b8c272SStefano Zampini           Mat T,T2,SEj2;
1153*72b8c272SStefano Zampini           ierr = MatConvert(SEj,MATSEQAIJ,MAT_INITIAL_MATRIX,&T);CHKERRQ(ierr);
1154*72b8c272SStefano Zampini           ierr = MatConvert(change_sub,MATSEQAIJ,MAT_INITIAL_MATRIX,&T2);CHKERRQ(ierr);
1155*72b8c272SStefano Zampini           ierr = MatPtAP(T,T2,MAT_INITIAL_MATRIX,1.0,&SEj2);CHKERRQ(ierr);
1156*72b8c272SStefano Zampini           ierr = MatDestroy(&T);CHKERRQ(ierr);
1157*72b8c272SStefano Zampini           ierr = MatDestroy(&T2);CHKERRQ(ierr);
1158*72b8c272SStefano Zampini           ierr = MatCopy(SEj2,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
1159*72b8c272SStefano Zampini           ierr = MatDestroy(&SEj2);CHKERRQ(ierr);
1160*72b8c272SStefano Zampini         }
1161*72b8c272SStefano Zampini         { /* currently there's no support for ZeroRows with A dense */
1162*72b8c272SStefano Zampini           Mat T;
1163*72b8c272SStefano Zampini           ierr = MatConvert(SEj,MATSEQAIJ,MAT_INITIAL_MATRIX,&T);CHKERRQ(ierr);
1164*72b8c272SStefano Zampini           ierr = MatZeroRowsColumnsIS(T,change_primal_sub[i],1.,NULL,NULL);CHKERRQ(ierr);
1165*72b8c272SStefano Zampini           ierr = MatConvert(T,MATSEQDENSE,MAT_REUSE_MATRIX,&T);CHKERRQ(ierr);
1166*72b8c272SStefano Zampini           ierr = MatCopy(T,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
1167*72b8c272SStefano Zampini           ierr = MatDestroy(&T);CHKERRQ(ierr);
1168*72b8c272SStefano Zampini         }
1169*72b8c272SStefano Zampini         ierr = MatDestroy(&SEj);CHKERRQ(ierr);
1170*72b8c272SStefano Zampini       }
11715a95e1ceSStefano Zampini       ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr);
1172a1337663SStefano Zampini 
1173683d3df6SStefano Zampini       /* if adaptivity is requested, invert S_E blocks */
1174862806e4SStefano Zampini       if (compute_Stilda) {
117508122e43SStefano Zampini         ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr);
117608122e43SStefano Zampini         ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
117706a4e24aSStefano Zampini         if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */
11786c3e6151SStefano Zampini           PetscInt k;
11795a95e1ceSStefano Zampini           PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr));
11802972d61bSStefano Zampini           if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr);
11815a95e1ceSStefano Zampini           PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr));
11822972d61bSStefano Zampini           if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
11836c3e6151SStefano Zampini           for (k=0;k<subset_size;k++) {
11846c3e6151SStefano Zampini             for (j=k;j<subset_size;j++) {
11856c3e6151SStefano Zampini               work[j*subset_size+k] = work[k*subset_size+j];
11866c3e6151SStefano Zampini             }
11876c3e6151SStefano Zampini           }
1188d6462365SStefano Zampini         } else {
1189d6462365SStefano Zampini           PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr));
1190d6462365SStefano Zampini           if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
1191d6462365SStefano Zampini           PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
1192d6462365SStefano Zampini           if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
11932972d61bSStefano Zampini         }
119408122e43SStefano Zampini         ierr = PetscFPTrapPop();CHKERRQ(ierr);
11955a94c880SStefano Zampini         ierr = MatSetValues(sub_schurs->sum_S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr);
11969087bf02SStefano Zampini       }
1197be83ff47SStefano Zampini       cum += subset_size;
1198be83ff47SStefano Zampini       cum2 += subset_size*(size_schur + 1);
1199be83ff47SStefano Zampini     }
1200be83ff47SStefano Zampini     ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr);
12014a6c6b0dSStefano Zampini 
1202df4d28bfSStefano Zampini     if (solver_S) {
12035a05ddb0SStefano Zampini       ierr = MatFactorRestoreSchurComplement(F,&S_all);CHKERRQ(ierr);
12044a6c6b0dSStefano Zampini     }
1205683d3df6SStefano Zampini 
1206683d3df6SStefano Zampini     schur_factor = NULL;
120745951f25SStefano Zampini     if (compute_Stilda && size_active_schur) {
1208683d3df6SStefano Zampini 
12094a6c6b0dSStefano Zampini       if (sub_schurs->n_subs == 1 && size_schur == size_active_schur) { /* we already computed the inverse */
12104a6c6b0dSStefano Zampini         PetscInt j;
12114a6c6b0dSStefano Zampini         for (j=0;j<size_schur;j++) dummy_idx[j] = j;
12125a94c880SStefano Zampini         ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr);
12134a6c6b0dSStefano Zampini       } else {
1214683d3df6SStefano Zampini         Mat S_all_inv=NULL;
1215df4d28bfSStefano Zampini         if (solver_S) { /* use MatFactor calls to invert S */
1216*72b8c272SStefano Zampini             /* let MatFactor factor the Schur complement */
12176dba178dSStefano Zampini           ierr = MatFactorFactorizeSchurComplement(F);CHKERRQ(ierr);
1218683d3df6SStefano Zampini           /* for adaptive selection we need S^-1; for solver reusage we need S_\Delta\Delta^-1.
1219683d3df6SStefano 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 */
1220683d3df6SStefano Zampini           if (factor_workaround) {
1221683d3df6SStefano Zampini             PetscScalar *data;
1222683d3df6SStefano Zampini             PetscInt nd = 0;
12236dba178dSStefano Zampini 
1224683d3df6SStefano Zampini             if (!sub_schurs->is_posdef) {
1225683d3df6SStefano Zampini               SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices");
1226683d3df6SStefano Zampini             }
12276dba178dSStefano Zampini             ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr);
1228683d3df6SStefano Zampini             ierr = MatDenseGetArray(S_all_inv,&data);CHKERRQ(ierr);
1229683d3df6SStefano Zampini             if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */
1230683d3df6SStefano Zampini               ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr);
1231683d3df6SStefano Zampini             }
1232683d3df6SStefano Zampini             ierr = PetscMalloc1((size_active_schur*(size_active_schur+1))/2+nd,&schur_factor);CHKERRQ(ierr);
1233683d3df6SStefano Zampini             cum = 0;
1234683d3df6SStefano Zampini             for (i=0;i<size_active_schur;i++) {
1235683d3df6SStefano Zampini               ierr = PetscMemcpy(schur_factor+cum,data+i*(size_schur+1),(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr);
1236683d3df6SStefano Zampini               cum += size_active_schur-i;
1237683d3df6SStefano Zampini             }
1238683d3df6SStefano Zampini             for (i=0;i<nd;i++) schur_factor[cum+i] = data[(i+size_active_schur)*(size_schur+1)];
12396dba178dSStefano Zampini             /* invert without calling MatFactorInvertSchurComplement, since we are hacking */
1240683d3df6SStefano Zampini             ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr);
1241683d3df6SStefano Zampini             ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
1242683d3df6SStefano Zampini             PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,data,&B_N,&B_ierr));
1243683d3df6SStefano Zampini             ierr = PetscFPTrapPop();CHKERRQ(ierr);
1244683d3df6SStefano Zampini             if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
1245683d3df6SStefano Zampini             ierr = MatDenseRestoreArray(S_all_inv,&data);CHKERRQ(ierr);
1246683d3df6SStefano Zampini           } else {
12476dba178dSStefano Zampini             ierr = MatFactorInvertSchurComplement(F);CHKERRQ(ierr);
12486dba178dSStefano Zampini             ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr);
1249683d3df6SStefano Zampini           }
1250df4d28bfSStefano Zampini         } else { /* we need to invert explicitly since we are not using MatFactor for S */
1251683d3df6SStefano Zampini           ierr = PetscObjectReference((PetscObject)S_all);CHKERRQ(ierr);
1252683d3df6SStefano Zampini           S_all_inv = S_all;
1253683d3df6SStefano Zampini           ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr);
1254be83ff47SStefano Zampini           ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr);
1255be83ff47SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
125606a4e24aSStefano Zampini           if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */
1257be83ff47SStefano Zampini             PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr));
1258be83ff47SStefano Zampini             if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr);
1259be83ff47SStefano Zampini             PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr));
1260be83ff47SStefano Zampini             if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
1261d6462365SStefano Zampini           } else {
1262d6462365SStefano Zampini             PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr));
1263d6462365SStefano Zampini             if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
1264d6462365SStefano Zampini             PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
1265d6462365SStefano Zampini             if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
1266be83ff47SStefano Zampini           }
1267be83ff47SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
1268683d3df6SStefano Zampini           ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr);
1269be83ff47SStefano Zampini         }
1270be83ff47SStefano Zampini         /* S_Ej_tilda_all */
1271be83ff47SStefano Zampini         cum = cum2 = 0;
1272683d3df6SStefano Zampini         ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr);
1273be83ff47SStefano Zampini         for (i=0;i<sub_schurs->n_subs;i++) {
1274be83ff47SStefano Zampini           PetscInt j;
1275862806e4SStefano Zampini 
1276862806e4SStefano Zampini           ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
1277be83ff47SStefano Zampini           /* get (St^-1)_E */
1278*72b8c272SStefano Zampini           /* Unless we are changing the variables, I don't need to expand to upper triangular since St^-1
127906a4e24aSStefano Zampini              will be properly accessed later during adaptive selection */
1280a0b0af32SStefano Zampini           if (S_lower_triangular) {
1281be83ff47SStefano Zampini             PetscInt k;
1282*72b8c272SStefano Zampini             if (change) {
1283be83ff47SStefano Zampini               for (k=0;k<subset_size;k++) {
1284be83ff47SStefano Zampini                 for (j=k;j<subset_size;j++) {
1285be83ff47SStefano Zampini                   work[k*subset_size+j] = S_data[cum2+k*size_schur+j];
12866c3e6151SStefano Zampini                   work[j*subset_size+k] = work[k*subset_size+j];
1287be83ff47SStefano Zampini                 }
1288be83ff47SStefano Zampini               }
1289*72b8c272SStefano Zampini             } else {
1290*72b8c272SStefano Zampini               for (k=0;k<subset_size;k++) {
1291*72b8c272SStefano Zampini                 for (j=k;j<subset_size;j++) {
1292*72b8c272SStefano Zampini                   work[k*subset_size+j] = S_data[cum2+k*size_schur+j];
1293*72b8c272SStefano Zampini                 }
1294*72b8c272SStefano Zampini               }
1295*72b8c272SStefano Zampini             }
1296*72b8c272SStefano Zampini           } else {
1297be83ff47SStefano Zampini             PetscInt k;
1298be83ff47SStefano Zampini             for (k=0;k<subset_size;k++) {
1299be83ff47SStefano Zampini               for (j=0;j<subset_size;j++) {
1300be83ff47SStefano Zampini                 work[k*subset_size+j] = S_data[cum2+k*size_schur+j];
1301be83ff47SStefano Zampini               }
1302be83ff47SStefano Zampini             }
1303be83ff47SStefano Zampini           }
1304*72b8c272SStefano Zampini           if (change) {
1305*72b8c272SStefano Zampini             Mat change_sub,SEj;
1306*72b8c272SStefano Zampini 
1307*72b8c272SStefano Zampini             /* change basis */
1308*72b8c272SStefano Zampini             ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr);
1309*72b8c272SStefano Zampini             ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr);
1310*72b8c272SStefano Zampini             {
1311*72b8c272SStefano Zampini               Mat T,T2,SEj2;
1312*72b8c272SStefano Zampini               ierr = MatConvert(SEj,MATSEQAIJ,MAT_INITIAL_MATRIX,&T);CHKERRQ(ierr);
1313*72b8c272SStefano Zampini               ierr = MatConvert(change_sub,MATSEQAIJ,MAT_INITIAL_MATRIX,&T2);CHKERRQ(ierr);
1314*72b8c272SStefano Zampini               ierr = MatPtAP(T,T2,MAT_INITIAL_MATRIX,1.0,&SEj2);CHKERRQ(ierr);
1315*72b8c272SStefano Zampini               ierr = MatDestroy(&T);CHKERRQ(ierr);
1316*72b8c272SStefano Zampini               ierr = MatDestroy(&T2);CHKERRQ(ierr);
1317*72b8c272SStefano Zampini               ierr = MatCopy(SEj2,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
1318*72b8c272SStefano Zampini               ierr = MatDestroy(&SEj2);CHKERRQ(ierr);
1319*72b8c272SStefano Zampini             }
1320*72b8c272SStefano Zampini             /* set diagonal entry to a very large value to pick the basis we are eliminating as the first eigenvectors with adaptive selection */
1321*72b8c272SStefano Zampini             {
1322*72b8c272SStefano Zampini               const PetscInt *idxs;
1323*72b8c272SStefano Zampini               PetscInt n,j;
1324*72b8c272SStefano Zampini               ierr = ISGetLocalSize(change_primal_sub[i],&n);CHKERRQ(ierr);
1325*72b8c272SStefano Zampini               ierr = ISGetIndices(change_primal_sub[i],&idxs);CHKERRQ(ierr);
1326*72b8c272SStefano Zampini               for (j=0;j<n;j++) {
1327*72b8c272SStefano Zampini                 PetscInt k;
1328*72b8c272SStefano Zampini                 for (k=0;k<subset_size;k++) {
1329*72b8c272SStefano Zampini                   work[k+idxs[j]*subset_size] = work[idxs[j]+k*subset_size] = 0.;
1330*72b8c272SStefano Zampini                 }
1331*72b8c272SStefano Zampini                 work[idxs[j] + idxs[j]*subset_size] = 1./PETSC_SMALL;
1332*72b8c272SStefano Zampini               }
1333*72b8c272SStefano Zampini               ierr = ISRestoreIndices(change_primal_sub[i],&idxs);CHKERRQ(ierr);
1334*72b8c272SStefano Zampini             }
1335*72b8c272SStefano Zampini             ierr = MatDestroy(&SEj);CHKERRQ(ierr);
1336*72b8c272SStefano Zampini           }
1337be83ff47SStefano Zampini           for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j;
13385a94c880SStefano Zampini           ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr);
1339be83ff47SStefano Zampini           cum += subset_size;
1340be83ff47SStefano Zampini           cum2 += subset_size*(size_schur + 1);
1341883469d8SStefano Zampini         }
1342683d3df6SStefano Zampini         ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr);
1343df4d28bfSStefano Zampini         if (solver_S) {
13446dba178dSStefano Zampini           ierr = MatFactorRestoreSchurComplement(F,&S_all_inv);CHKERRQ(ierr);
13455db18549SStefano Zampini         }
1346683d3df6SStefano Zampini         ierr = MatDestroy(&S_all_inv);CHKERRQ(ierr);
1347683d3df6SStefano Zampini       }
1348683d3df6SStefano Zampini 
1349683d3df6SStefano Zampini       /* move back factors to Schur data into F */
1350683d3df6SStefano Zampini       if (factor_workaround) {
1351683d3df6SStefano Zampini         Mat S_tmp;
1352683d3df6SStefano Zampini         PetscInt nv,nd=0;
1353683d3df6SStefano Zampini 
1354df4d28bfSStefano Zampini         if (!solver_S) {
1355683d3df6SStefano Zampini           SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
1356683d3df6SStefano Zampini         }
1357683d3df6SStefano Zampini         ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr);
13586dba178dSStefano Zampini         ierr = MatFactorGetSchurComplement(F,&S_tmp);CHKERRQ(ierr);
1359683d3df6SStefano Zampini         if (sub_schurs->is_posdef) {
1360683d3df6SStefano Zampini           PetscScalar *data;
1361683d3df6SStefano Zampini 
1362683d3df6SStefano Zampini           ierr = MatZeroEntries(S_tmp);CHKERRQ(ierr);
1363683d3df6SStefano Zampini           ierr = MatDenseGetArray(S_tmp,&data);CHKERRQ(ierr);
1364683d3df6SStefano Zampini 
1365683d3df6SStefano Zampini           if (S_lower_triangular) {
1366683d3df6SStefano Zampini             cum = 0;
1367683d3df6SStefano Zampini             for (i=0;i<size_active_schur;i++) {
1368683d3df6SStefano Zampini               ierr = PetscMemcpy(data+i*(size_schur+1),schur_factor+cum,(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr);
1369683d3df6SStefano Zampini               cum += size_active_schur-i;
1370683d3df6SStefano Zampini 	    }
1371683d3df6SStefano Zampini           } else {
1372683d3df6SStefano Zampini             ierr = PetscMemcpy(data,schur_factor,size_schur*size_schur*sizeof(PetscScalar));CHKERRQ(ierr);
1373683d3df6SStefano Zampini           }
1374683d3df6SStefano Zampini           if (sub_schurs->is_dir) {
1375683d3df6SStefano Zampini             ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr);
1376683d3df6SStefano Zampini 	    for (i=0;i<nd;i++) {
1377683d3df6SStefano Zampini 	      data[(i+size_active_schur)*(size_schur+1)] = schur_factor[cum+i];
1378683d3df6SStefano Zampini 	    }
1379683d3df6SStefano Zampini           }
13806dba178dSStefano Zampini           /* workaround: since I cannot modify the matrices used inside the solvers for the forward and backward substitutions,
1381683d3df6SStefano Zampini              set the diagonal entry of the Schur factor to a very large value */
1382683d3df6SStefano Zampini           for (i=size_active_schur+nd;i<size_schur;i++) {
13836c3e6151SStefano Zampini             data[i*(size_schur+1)] = infty;
1384683d3df6SStefano Zampini           }
1385683d3df6SStefano Zampini           ierr = MatDenseRestoreArray(S_tmp,&data);CHKERRQ(ierr);
1386683d3df6SStefano Zampini         } else {
1387683d3df6SStefano Zampini           SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices");
1388683d3df6SStefano Zampini         }
13896dba178dSStefano Zampini         ierr = MatFactorRestoreSchurComplement(F,&S_tmp);CHKERRQ(ierr);
13909087bf02SStefano Zampini       }
1391367aa537SStefano Zampini     } else if (factor_workaround) { /* we need to eliminate any unneeded coupling */
1392367aa537SStefano Zampini       PetscScalar *data;
1393367aa537SStefano Zampini       PetscInt    nd = 0;
1394367aa537SStefano Zampini 
1395367aa537SStefano Zampini       if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */
1396367aa537SStefano Zampini         ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr);
1397367aa537SStefano Zampini       }
1398367aa537SStefano Zampini       ierr = MatFactorGetSchurComplement(F,&S_all);CHKERRQ(ierr);
1399367aa537SStefano Zampini       ierr = MatDenseGetArray(S_all,&data);CHKERRQ(ierr);
1400367aa537SStefano Zampini       for (i=0;i<size_active_schur;i++) {
1401367aa537SStefano Zampini         ierr = PetscMemzero(data+i*size_schur+size_active_schur,(size_schur-size_active_schur)*sizeof(PetscScalar));CHKERRQ(ierr);
1402367aa537SStefano Zampini       }
1403367aa537SStefano Zampini       for (i=size_active_schur+nd;i<size_schur;i++) {
1404367aa537SStefano Zampini         ierr = PetscMemzero(data+i*size_schur+size_active_schur,(size_schur-size_active_schur)*sizeof(PetscScalar));CHKERRQ(ierr);
14056c3e6151SStefano Zampini         data[i*(size_schur+1)] = infty;
1406367aa537SStefano Zampini       }
1407367aa537SStefano Zampini       ierr = MatDenseRestoreArray(S_all,&data);CHKERRQ(ierr);
1408367aa537SStefano Zampini       ierr = MatFactorRestoreSchurComplement(F,&S_all);CHKERRQ(ierr);
14094a6c6b0dSStefano Zampini     }
14104a6c6b0dSStefano Zampini     ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr);
1411683d3df6SStefano Zampini     ierr = PetscFree(schur_factor);CHKERRQ(ierr);
14124a6c6b0dSStefano Zampini   }
1413*72b8c272SStefano Zampini   if (change_primal_sub) {
1414*72b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
1415*72b8c272SStefano Zampini       ierr = ISDestroy(&change_primal_sub[i]);CHKERRQ(ierr);
1416*72b8c272SStefano Zampini     }
1417*72b8c272SStefano Zampini   }
1418*72b8c272SStefano Zampini   ierr = PetscFree(change_primal_sub);CHKERRQ(ierr);
14195a94c880SStefano Zampini   ierr = PetscFree(nnz);CHKERRQ(ierr);
1420be83ff47SStefano Zampini   ierr = MatDestroy(&F);CHKERRQ(ierr);
1421a9b99552SStefano Zampini   ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr);
1422a1337663SStefano Zampini   ierr = MatDestroy(&S_all);CHKERRQ(ierr);
1423a64f4aa4SStefano Zampini   ierr = MatDestroy(&A_BB);CHKERRQ(ierr);
1424a64f4aa4SStefano Zampini   ierr = MatDestroy(&A_IB);CHKERRQ(ierr);
1425a64f4aa4SStefano Zampini   ierr = MatDestroy(&A_BI);CHKERRQ(ierr);
14265db18549SStefano Zampini   ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
14275db18549SStefano Zampini   ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
14285a95e1ceSStefano Zampini   if (compute_Stilda) {
14295a94c880SStefano Zampini     ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
14305a94c880SStefano Zampini     ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
14315a94c880SStefano Zampini     ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
14325a94c880SStefano Zampini     ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
143308122e43SStefano Zampini   }
1434a1337663SStefano Zampini 
14355db18549SStefano Zampini   /* Global matrix of all assembled Schur on subsets */
14365db18549SStefano Zampini   ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr);
14373927de2eSStefano Zampini   ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr);
14383927de2eSStefano Zampini   ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr);
14395a95e1ceSStefano Zampini 
14405db18549SStefano Zampini   /* Get local part of (\sum_j S_Ej) */
14415a94c880SStefano Zampini   if (!sub_schurs->sum_S_Ej_all) {
1442dc456d91SStefano Zampini     ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_INITIAL_MATRIX,&submats);CHKERRQ(ierr);
14435a94c880SStefano Zampini     sub_schurs->sum_S_Ej_all = submats[0];
14445a94c880SStefano Zampini   } else {
14455a94c880SStefano Zampini     ierr = PetscMalloc1(1,&submats);CHKERRQ(ierr);
14465a94c880SStefano Zampini     submats[0] = sub_schurs->sum_S_Ej_all;
1447dc456d91SStefano Zampini     ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr);
14485a94c880SStefano Zampini   }
144908122e43SStefano Zampini 
1450ac632422SStefano Zampini   /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */
1451*72b8c272SStefano Zampini #if 0
1452ac632422SStefano Zampini   if (faster_deluxe) {
14535a95e1ceSStefano Zampini     Mat         tmpmat;
14545a95e1ceSStefano Zampini     PetscScalar *array;
14555a95e1ceSStefano Zampini     PetscInt    cum;
14565a95e1ceSStefano Zampini 
14575a95e1ceSStefano Zampini     ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr);
14585a95e1ceSStefano Zampini     cum = 0;
14595a95e1ceSStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
14605a95e1ceSStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
14615a95e1ceSStefano Zampini       ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr);
14625a95e1ceSStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
146306a4e24aSStefano Zampini       if (!use_getr) {
14645a95e1ceSStefano Zampini         PetscInt j,k;
14655a95e1ceSStefano Zampini 
14665a95e1ceSStefano Zampini         PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr));
14675a95e1ceSStefano Zampini         if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr);
14685a95e1ceSStefano Zampini         PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr));
14695a95e1ceSStefano Zampini         if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr);
14705a95e1ceSStefano Zampini         for (j=0;j<B_N;j++) {
14715a95e1ceSStefano Zampini           for (k=j+1;k<B_N;k++) {
14725a95e1ceSStefano Zampini             array[k*B_N+j+cum] = array[j*B_N+k+cum];
14735a95e1ceSStefano Zampini           }
14745a95e1ceSStefano Zampini         }
1475d6462365SStefano Zampini       } else {
1476d6462365SStefano Zampini         PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr));
1477d6462365SStefano Zampini         if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr);
1478d6462365SStefano Zampini         PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr));
1479d6462365SStefano Zampini         if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr);
14805a95e1ceSStefano Zampini       }
14815a95e1ceSStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
14825a95e1ceSStefano Zampini       cum += subset_size*subset_size;
14835a95e1ceSStefano Zampini     }
14845a95e1ceSStefano Zampini     ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr);
14855a95e1ceSStefano Zampini     ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr);
14865a95e1ceSStefano Zampini     ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr);
1487ac632422SStefano Zampini     ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr);
14885a95e1ceSStefano Zampini     sub_schurs->S_Ej_all = tmpmat;
14895a95e1ceSStefano Zampini   }
1490*72b8c272SStefano Zampini #endif
1491f6f667cfSStefano Zampini   /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */
14925a95e1ceSStefano Zampini   if (compute_Stilda) {
14935a94c880SStefano Zampini     ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr);
1494a1337663SStefano Zampini     ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr);
14955a94c880SStefano Zampini     submats[0] = sub_schurs->sum_S_Ej_tilda_all;
1496dc456d91SStefano Zampini     ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr);
14975a94c880SStefano Zampini     ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr);
149808122e43SStefano Zampini     ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr);
14995a94c880SStefano Zampini     submats[0] = sub_schurs->sum_S_Ej_inv_all;
1500dc456d91SStefano Zampini     ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr);
150108122e43SStefano Zampini   }
15023202ece2SStefano Zampini 
15035a95e1ceSStefano Zampini   /* free workspace */
15045a94c880SStefano Zampini   ierr = PetscFree(submats);CHKERRQ(ierr);
150506a4b1faSStefano Zampini   ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr);
1506a1337663SStefano Zampini   ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr);
15073202ece2SStefano Zampini   ierr = MatDestroy(&work_mat);CHKERRQ(ierr);
1508dc456d91SStefano Zampini   ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr);
15095a95e1ceSStefano Zampini   ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr);
1510b1b3d7a2SStefano Zampini   PetscFunctionReturn(0);
1511b1b3d7a2SStefano Zampini }
1512b1b3d7a2SStefano Zampini 
1513b1b3d7a2SStefano Zampini #undef __FUNCT__
1514b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit"
1515a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap)
1516b1b3d7a2SStefano Zampini {
15179bb4a8caSStefano Zampini   IS              *faces,*edges,*all_cc,vertices;
15185a95e1ceSStefano Zampini   PetscInt        i,n_faces,n_edges,n_all_cc;
1519b1b3d7a2SStefano Zampini   PetscBool       is_sorted;
1520b1b3d7a2SStefano Zampini   PetscErrorCode  ierr;
1521b1b3d7a2SStefano Zampini 
1522b1b3d7a2SStefano Zampini   PetscFunctionBegin;
1523b1b3d7a2SStefano Zampini   ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr);
1524b1b3d7a2SStefano Zampini   if (!is_sorted) {
1525b1b3d7a2SStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted");
1526b1b3d7a2SStefano Zampini   }
1527b1b3d7a2SStefano Zampini   ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr);
1528b1b3d7a2SStefano Zampini   if (!is_sorted) {
1529b1b3d7a2SStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted");
1530b1b3d7a2SStefano Zampini   }
1531b1b3d7a2SStefano Zampini 
1532b1b3d7a2SStefano Zampini   /* reset any previous data */
1533b1b3d7a2SStefano Zampini   ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr);
1534b1b3d7a2SStefano Zampini 
15355a95e1ceSStefano Zampini   /* get index sets for faces and edges (already sorted by global ordering) */
15369bb4a8caSStefano Zampini   ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr);
1537b1b3d7a2SStefano Zampini   n_all_cc = n_faces+n_edges;
153808122e43SStefano Zampini   ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr);
1539b1b3d7a2SStefano Zampini   ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr);
1540b1b3d7a2SStefano Zampini   for (i=0;i<n_faces;i++) {
1541b1b3d7a2SStefano Zampini     all_cc[i] = faces[i];
1542b1b3d7a2SStefano Zampini   }
1543b1b3d7a2SStefano Zampini   for (i=0;i<n_edges;i++) {
1544b1b3d7a2SStefano Zampini     all_cc[n_faces+i] = edges[i];
154508122e43SStefano Zampini     ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr);
1546b1b3d7a2SStefano Zampini   }
1547b1b3d7a2SStefano Zampini   ierr = PetscFree(faces);CHKERRQ(ierr);
1548b1b3d7a2SStefano Zampini   ierr = PetscFree(edges);CHKERRQ(ierr);
1549d62866d3SStefano Zampini   sub_schurs->is_dir = NULL;
1550d62866d3SStefano Zampini   ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr);
1551b1b3d7a2SStefano Zampini 
1552df4d28bfSStefano Zampini   /* Determine if MatFactor can be used */
1553df4d28bfSStefano Zampini   sub_schurs->schur_explicit = PETSC_FALSE;
1554883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS)
1555df4d28bfSStefano Zampini   sub_schurs->schur_explicit = PETSC_TRUE;
1556df4d28bfSStefano Zampini #endif
1557df4d28bfSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO)
1558df4d28bfSStefano Zampini   sub_schurs->schur_explicit = PETSC_TRUE;
1559883469d8SStefano Zampini #endif
1560b1b3d7a2SStefano Zampini 
1561b1b3d7a2SStefano Zampini   ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr);
1562b1b3d7a2SStefano Zampini   sub_schurs->is_I = is_I;
1563b1b3d7a2SStefano Zampini   ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr);
1564b1b3d7a2SStefano Zampini   sub_schurs->is_B = is_B;
15655db18549SStefano Zampini   ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr);
15665db18549SStefano Zampini   sub_schurs->l2gmap = graph->l2gmap;
15675db18549SStefano Zampini   ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr);
15685db18549SStefano Zampini   sub_schurs->BtoNmap = BtoNmap;
15695a95e1ceSStefano Zampini   sub_schurs->n_subs = n_all_cc;
1570b1b3d7a2SStefano Zampini   sub_schurs->is_subs = all_cc;
1571df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) { /* sort by local ordering if we are not using MatFactor */
1572b96c3477SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
1573b96c3477SStefano Zampini       ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr);
1574b96c3477SStefano Zampini     }
15759bb4a8caSStefano Zampini   }
1576d62866d3SStefano Zampini   sub_schurs->is_vertices = vertices;
1577b96c3477SStefano Zampini   sub_schurs->S_Ej_all = NULL;
1578b96c3477SStefano Zampini   sub_schurs->sum_S_Ej_all = NULL;
157908122e43SStefano Zampini   sub_schurs->sum_S_Ej_inv_all = NULL;
1580b96c3477SStefano Zampini   sub_schurs->sum_S_Ej_tilda_all = NULL;
1581b96c3477SStefano Zampini   sub_schurs->is_Ej_all = NULL;
1582b1b3d7a2SStefano Zampini   PetscFunctionReturn(0);
1583b1b3d7a2SStefano Zampini }
1584b1b3d7a2SStefano Zampini 
1585b1b3d7a2SStefano Zampini #undef __FUNCT__
158634a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate"
158734a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs)
158834a97f8cSStefano Zampini {
158934a97f8cSStefano Zampini   PCBDDCSubSchurs schurs_ctx;
159034a97f8cSStefano Zampini   PetscErrorCode  ierr;
159134a97f8cSStefano Zampini 
159234a97f8cSStefano Zampini   PetscFunctionBegin;
159334a97f8cSStefano Zampini   ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr);
15945ff63025SStefano Zampini   schurs_ctx->n_subs = 0;
159534a97f8cSStefano Zampini   *sub_schurs = schurs_ctx;
159634a97f8cSStefano Zampini   PetscFunctionReturn(0);
159734a97f8cSStefano Zampini }
159834a97f8cSStefano Zampini 
159934a97f8cSStefano Zampini #undef __FUNCT__
160034a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset"
160134a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs)
160234a97f8cSStefano Zampini {
160334a97f8cSStefano Zampini   PetscInt       i;
160434a97f8cSStefano Zampini   PetscErrorCode ierr;
160534a97f8cSStefano Zampini 
160634a97f8cSStefano Zampini   PetscFunctionBegin;
16071e9c79c2SStefano Zampini   ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr);
1608b1b3d7a2SStefano Zampini   ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr);
1609b1b3d7a2SStefano Zampini   ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr);
1610b1b3d7a2SStefano Zampini   ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr);
16115db18549SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr);
16125db18549SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr);
161341c3ba1bSStefano Zampini   ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr);
161441c3ba1bSStefano Zampini   ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr);
161508122e43SStefano Zampini   ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr);
1616a1337663SStefano Zampini   ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr);
16175db18549SStefano Zampini   ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr);
1618d62866d3SStefano Zampini   ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr);
1619d62866d3SStefano Zampini   ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr);
162008122e43SStefano Zampini   ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr);
162134a97f8cSStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
1622b1b3d7a2SStefano Zampini     ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr);
162334a97f8cSStefano Zampini   }
16245ff63025SStefano Zampini   if (sub_schurs->n_subs) {
1625b1b3d7a2SStefano Zampini     ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr);
16263dc780c3SStefano Zampini   }
1627df4d28bfSStefano Zampini   if (sub_schurs->reuse_solver) {
1628df4d28bfSStefano Zampini     ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr);
1629d62866d3SStefano Zampini   }
1630df4d28bfSStefano Zampini   ierr = PetscFree(sub_schurs->reuse_solver);CHKERRQ(ierr);
1631*72b8c272SStefano Zampini   if (sub_schurs->change) {
1632*72b8c272SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
1633*72b8c272SStefano Zampini       ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr);
1634*72b8c272SStefano Zampini     }
1635*72b8c272SStefano Zampini   }
1636*72b8c272SStefano Zampini   ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr);
163734a97f8cSStefano Zampini   sub_schurs->n_subs = 0;
163834a97f8cSStefano Zampini   PetscFunctionReturn(0);
163934a97f8cSStefano Zampini }
164034a97f8cSStefano Zampini 
164134a97f8cSStefano Zampini #undef __FUNCT__
164234a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private"
16432a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added)
164434a97f8cSStefano Zampini {
164534a97f8cSStefano Zampini   PetscInt       i,j,n;
164634a97f8cSStefano Zampini   PetscErrorCode ierr;
164734a97f8cSStefano Zampini 
164834a97f8cSStefano Zampini   PetscFunctionBegin;
164934a97f8cSStefano Zampini   n = 0;
165034a97f8cSStefano Zampini   for (i=-n_prev;i<0;i++) {
165134a97f8cSStefano Zampini     PetscInt start_dof = queue_tip[i];
165234a97f8cSStefano Zampini     for (j=xadj[start_dof];j<xadj[start_dof+1];j++) {
165334a97f8cSStefano Zampini       PetscInt dof = adjncy[j];
165434a97f8cSStefano Zampini       if (!PetscBTLookup(touched,dof)) {
165534a97f8cSStefano Zampini         ierr = PetscBTSet(touched,dof);CHKERRQ(ierr);
165634a97f8cSStefano Zampini         queue_tip[n] = dof;
165734a97f8cSStefano Zampini         n++;
165834a97f8cSStefano Zampini       }
165934a97f8cSStefano Zampini     }
166034a97f8cSStefano Zampini   }
166134a97f8cSStefano Zampini   *n_added = n;
166234a97f8cSStefano Zampini   PetscFunctionReturn(0);
166334a97f8cSStefano Zampini }
1664