xref: /petsc/src/ksp/pc/impls/bddc/bddcprivate.c (revision 4fee134f50287c79ea8ad6d04a528deca78dfa62)
11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h>
2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h>
3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h>
4674ae819SStefano Zampini #include <petscblaslapack.h>
5674ae819SStefano Zampini 
6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y);
7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y);
8906d46d4SStefano Zampini 
9a3df083aSStefano Zampini #undef __FUNCT__
10a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private"
11a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose)
12a3df083aSStefano Zampini {
13a3df083aSStefano Zampini   PCBDDCBenignMatMult_ctx ctx;
14a3df083aSStefano Zampini   PetscErrorCode          ierr;
15a3df083aSStefano Zampini   PetscBool               apply_right,apply_left,reset_x;
16a3df083aSStefano Zampini 
17a3df083aSStefano Zampini   PetscFunctionBegin;
18a3df083aSStefano Zampini   ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr);
19a3df083aSStefano Zampini   if (transpose) {
20a3df083aSStefano Zampini     apply_right = ctx->apply_left;
21a3df083aSStefano Zampini     apply_left = ctx->apply_right;
22a3df083aSStefano Zampini   } else {
23a3df083aSStefano Zampini     apply_right = ctx->apply_right;
24a3df083aSStefano Zampini     apply_left = ctx->apply_left;
25a3df083aSStefano Zampini   }
26a3df083aSStefano Zampini   reset_x = PETSC_FALSE;
27a3df083aSStefano Zampini   if (apply_right) {
28a3df083aSStefano Zampini     const PetscScalar *ax;
29a3df083aSStefano Zampini     PetscInt          nl,i;
30a3df083aSStefano Zampini 
31a3df083aSStefano Zampini     ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr);
32a3df083aSStefano Zampini     ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr);
33a3df083aSStefano Zampini     ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr);
34a3df083aSStefano Zampini     ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr);
35a3df083aSStefano Zampini     for (i=0;i<ctx->benign_n;i++) {
36a3df083aSStefano Zampini       PetscScalar    sum,val;
37a3df083aSStefano Zampini       const PetscInt *idxs;
38a3df083aSStefano Zampini       PetscInt       nz,j;
39a3df083aSStefano Zampini       ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr);
40a3df083aSStefano Zampini       ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr);
41a3df083aSStefano Zampini       sum = 0.;
42a3df083aSStefano Zampini       if (ctx->apply_p0) {
43a3df083aSStefano Zampini         val = ctx->work[idxs[nz-1]];
44a3df083aSStefano Zampini         for (j=0;j<nz-1;j++) {
45a3df083aSStefano Zampini           sum += ctx->work[idxs[j]];
46a3df083aSStefano Zampini           ctx->work[idxs[j]] += val;
47a3df083aSStefano Zampini         }
48a3df083aSStefano Zampini       } else {
49a3df083aSStefano Zampini         for (j=0;j<nz-1;j++) {
50a3df083aSStefano Zampini           sum += ctx->work[idxs[j]];
51a3df083aSStefano Zampini         }
52a3df083aSStefano Zampini       }
53a3df083aSStefano Zampini       ctx->work[idxs[nz-1]] -= sum;
54a3df083aSStefano Zampini       ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr);
55a3df083aSStefano Zampini     }
56a3df083aSStefano Zampini     ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr);
57a3df083aSStefano Zampini     reset_x = PETSC_TRUE;
58a3df083aSStefano Zampini   }
59a3df083aSStefano Zampini   if (transpose) {
60a3df083aSStefano Zampini     ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr);
61a3df083aSStefano Zampini   } else {
62a3df083aSStefano Zampini     ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr);
63a3df083aSStefano Zampini   }
64a3df083aSStefano Zampini   if (reset_x) {
65a3df083aSStefano Zampini     ierr = VecResetArray(x);CHKERRQ(ierr);
66a3df083aSStefano Zampini   }
67a3df083aSStefano Zampini   if (apply_left) {
68a3df083aSStefano Zampini     PetscScalar *ay;
69a3df083aSStefano Zampini     PetscInt    i;
70a3df083aSStefano Zampini 
71a3df083aSStefano Zampini     ierr = VecGetArray(y,&ay);CHKERRQ(ierr);
72a3df083aSStefano Zampini     for (i=0;i<ctx->benign_n;i++) {
73a3df083aSStefano Zampini       PetscScalar    sum,val;
74a3df083aSStefano Zampini       const PetscInt *idxs;
75a3df083aSStefano Zampini       PetscInt       nz,j;
76a3df083aSStefano Zampini       ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr);
77a3df083aSStefano Zampini       ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr);
78a3df083aSStefano Zampini       val = -ay[idxs[nz-1]];
79a3df083aSStefano Zampini       if (ctx->apply_p0) {
80a3df083aSStefano Zampini         sum = 0.;
81a3df083aSStefano Zampini         for (j=0;j<nz-1;j++) {
82a3df083aSStefano Zampini           sum += ay[idxs[j]];
83a3df083aSStefano Zampini           ay[idxs[j]] += val;
84a3df083aSStefano Zampini         }
85a3df083aSStefano Zampini         ay[idxs[nz-1]] += sum;
86a3df083aSStefano Zampini       } else {
87a3df083aSStefano Zampini         for (j=0;j<nz-1;j++) {
88a3df083aSStefano Zampini           ay[idxs[j]] += val;
89a3df083aSStefano Zampini         }
90a3df083aSStefano Zampini         ay[idxs[nz-1]] = 0.;
91a3df083aSStefano Zampini       }
92a3df083aSStefano Zampini       ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr);
93a3df083aSStefano Zampini     }
94a3df083aSStefano Zampini     ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr);
95a3df083aSStefano Zampini   }
96a3df083aSStefano Zampini   PetscFunctionReturn(0);
97a3df083aSStefano Zampini }
98a3df083aSStefano Zampini 
99a3df083aSStefano Zampini #undef __FUNCT__
100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private"
101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y)
102a3df083aSStefano Zampini {
103a3df083aSStefano Zampini   PetscErrorCode ierr;
104a3df083aSStefano Zampini 
105a3df083aSStefano Zampini   PetscFunctionBegin;
106a3df083aSStefano Zampini   ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr);
107a3df083aSStefano Zampini   PetscFunctionReturn(0);
108a3df083aSStefano Zampini }
109a3df083aSStefano Zampini 
110a3df083aSStefano Zampini #undef __FUNCT__
111a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private"
112a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y)
113a3df083aSStefano Zampini {
114a3df083aSStefano Zampini   PetscErrorCode ierr;
115a3df083aSStefano Zampini 
116a3df083aSStefano Zampini   PetscFunctionBegin;
117a3df083aSStefano Zampini   ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr);
118a3df083aSStefano Zampini   PetscFunctionReturn(0);
119a3df083aSStefano Zampini }
120a3df083aSStefano Zampini 
121a3df083aSStefano Zampini #undef __FUNCT__
122a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat"
123a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore)
124a3df083aSStefano Zampini {
125a3df083aSStefano Zampini   PC_IS                   *pcis = (PC_IS*)pc->data;
126a3df083aSStefano Zampini   PC_BDDC                 *pcbddc = (PC_BDDC*)pc->data;
127a3df083aSStefano Zampini   Mat_IS                  *matis = (Mat_IS*)pc->pmat->data;
128a3df083aSStefano Zampini   PCBDDCBenignMatMult_ctx ctx;
129a3df083aSStefano Zampini   PetscErrorCode          ierr;
130a3df083aSStefano Zampini 
131a3df083aSStefano Zampini   PetscFunctionBegin;
132a3df083aSStefano Zampini   if (!restore) {
133a3df083aSStefano Zampini     Mat                A,A_IB,A_BI;
134a3df083aSStefano Zampini     PetscScalar        *work;
135df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver;
136a3df083aSStefano Zampini 
137a3df083aSStefano Zampini     if (!reuse) {
138a3df083aSStefano Zampini       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot setup shell matrices!");
139a3df083aSStefano Zampini     }
140a3df083aSStefano Zampini 
141a3df083aSStefano Zampini     /* local mat */
142a3df083aSStefano Zampini     ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr);
143a3df083aSStefano Zampini     ierr = MatSetSizes(A,pcis->n,pcis->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
144a3df083aSStefano Zampini     ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr);
145a3df083aSStefano Zampini     ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr);
146a3df083aSStefano Zampini     ierr = MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr);
147a3df083aSStefano Zampini     ierr = PetscNew(&ctx);CHKERRQ(ierr);
148a3df083aSStefano Zampini     ierr = MatShellSetContext(A,ctx);CHKERRQ(ierr);
149a3df083aSStefano Zampini     ctx->apply_left = PETSC_TRUE;
150a3df083aSStefano Zampini     ctx->apply_right = PETSC_TRUE;
151a3df083aSStefano Zampini     ctx->apply_p0 = PETSC_TRUE;
152a3df083aSStefano Zampini     ctx->benign_n = pcbddc->benign_n;
153a3df083aSStefano Zampini     ctx->benign_zerodiag_subs = pcbddc->benign_zerodiag_subs;
154a3df083aSStefano Zampini     ctx->A = matis->A;
155a3df083aSStefano Zampini     ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr);
156a3df083aSStefano Zampini     ctx->work = work;
157a3df083aSStefano Zampini     ierr = MatSetUp(A);CHKERRQ(ierr);
158a3df083aSStefano Zampini     ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
159a3df083aSStefano Zampini     ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
160a3df083aSStefano Zampini     matis->A = A;
161a3df083aSStefano Zampini 
162a3df083aSStefano Zampini     /* A_IB */
163a3df083aSStefano Zampini     ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr);
164a3df083aSStefano Zampini     ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
165a3df083aSStefano Zampini     ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr);
166a3df083aSStefano Zampini     ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr);
167a3df083aSStefano Zampini     ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr);
168a3df083aSStefano Zampini     ierr = PetscNew(&ctx);CHKERRQ(ierr);
169a3df083aSStefano Zampini     ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr);
170a3df083aSStefano Zampini     ctx->apply_left = PETSC_TRUE;
171a3df083aSStefano Zampini     ctx->apply_right = PETSC_FALSE;
172a3df083aSStefano Zampini     ctx->apply_p0 = PETSC_FALSE;
173a3df083aSStefano Zampini     ctx->benign_n = pcbddc->benign_n;
174a3df083aSStefano Zampini     ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs;
175a3df083aSStefano Zampini     ctx->A = pcis->A_IB;
176a3df083aSStefano Zampini     ctx->work = work;
177a3df083aSStefano Zampini     ierr = MatSetUp(A_IB);CHKERRQ(ierr);
178a3df083aSStefano Zampini     ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
179a3df083aSStefano Zampini     ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
180a3df083aSStefano Zampini     pcis->A_IB = A_IB;
181a3df083aSStefano Zampini 
182a3df083aSStefano Zampini     /* A_BI as A_IB^T */
183a3df083aSStefano Zampini     ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr);
184a3df083aSStefano Zampini     pcbddc->benign_original_mat = pcis->A_BI;
185a3df083aSStefano Zampini     pcis->A_BI = A_BI;
186a3df083aSStefano Zampini   } else {
187a3df083aSStefano Zampini     ierr = MatShellGetContext(matis->A,&ctx);CHKERRQ(ierr);
188a3df083aSStefano Zampini     ierr = MatDestroy(&matis->A);CHKERRQ(ierr);
189a3df083aSStefano Zampini     matis->A = ctx->A;
190a3df083aSStefano Zampini     ierr = PetscFree(ctx);CHKERRQ(ierr);
191a3df083aSStefano Zampini     ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr);
192a3df083aSStefano Zampini     ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr);
193a3df083aSStefano Zampini     pcis->A_IB = ctx->A;
194a3df083aSStefano Zampini     ierr = PetscFree(ctx->work);CHKERRQ(ierr);
195a3df083aSStefano Zampini     ierr = PetscFree(ctx);CHKERRQ(ierr);
196a3df083aSStefano Zampini     ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr);
197a3df083aSStefano Zampini     pcis->A_BI = pcbddc->benign_original_mat;
198a3df083aSStefano Zampini     pcbddc->benign_original_mat = NULL;
199a3df083aSStefano Zampini   }
200a3df083aSStefano Zampini   PetscFunctionReturn(0);
201a3df083aSStefano Zampini }
202a3df083aSStefano Zampini 
203a3df083aSStefano Zampini /* used just in bddc debug mode */
204a3df083aSStefano Zampini #undef __FUNCT__
205a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject"
206a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B)
207a3df083aSStefano Zampini {
208a3df083aSStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
209a3df083aSStefano Zampini   Mat_IS         *matis = (Mat_IS*)pc->pmat->data;
210a3df083aSStefano Zampini   Mat            An;
211a3df083aSStefano Zampini   PetscErrorCode ierr;
212a3df083aSStefano Zampini 
213a3df083aSStefano Zampini   PetscFunctionBegin;
214a3df083aSStefano Zampini   ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr);
215a3df083aSStefano Zampini   ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr);
216a3df083aSStefano Zampini   if (is1) {
217a3df083aSStefano Zampini     ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr);
218a3df083aSStefano Zampini     ierr = MatDestroy(&An);CHKERRQ(ierr);
219a3df083aSStefano Zampini   } else {
220a3df083aSStefano Zampini     *B = An;
221a3df083aSStefano Zampini   }
222a3df083aSStefano Zampini   PetscFunctionReturn(0);
223a3df083aSStefano Zampini }
224a3df083aSStefano Zampini 
2251cf9b237SStefano Zampini /* TODO: add reuse flag */
2261cf9b237SStefano Zampini #undef __FUNCT__
2271cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress"
2281cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B)
2291cf9b237SStefano Zampini {
2301cf9b237SStefano Zampini   Mat            Bt;
2311cf9b237SStefano Zampini   PetscScalar    *a,*bdata;
2321cf9b237SStefano Zampini   const PetscInt *ii,*ij;
2331cf9b237SStefano Zampini   PetscInt       m,n,i,nnz,*bii,*bij;
2341cf9b237SStefano Zampini   PetscBool      flg_row;
2351cf9b237SStefano Zampini   PetscErrorCode ierr;
2361cf9b237SStefano Zampini 
2371cf9b237SStefano Zampini   PetscFunctionBegin;
2381cf9b237SStefano Zampini   ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr);
2391cf9b237SStefano Zampini   ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr);
2401cf9b237SStefano Zampini   ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr);
2411cf9b237SStefano Zampini   nnz = n;
2421cf9b237SStefano Zampini   for (i=0;i<ii[n];i++) {
2431cf9b237SStefano Zampini     if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++;
2441cf9b237SStefano Zampini   }
2451cf9b237SStefano Zampini   ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr);
2461cf9b237SStefano Zampini   ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr);
2471cf9b237SStefano Zampini   ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr);
2481cf9b237SStefano Zampini   nnz = 0;
2491cf9b237SStefano Zampini   bii[0] = 0;
2501cf9b237SStefano Zampini   for (i=0;i<n;i++) {
2511cf9b237SStefano Zampini     PetscInt j;
2521cf9b237SStefano Zampini     for (j=ii[i];j<ii[i+1];j++) {
2531cf9b237SStefano Zampini       PetscScalar entry = a[j];
2541cf9b237SStefano Zampini       if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) {
2551cf9b237SStefano Zampini         bij[nnz] = ij[j];
2561cf9b237SStefano Zampini         bdata[nnz] = entry;
2571cf9b237SStefano Zampini         nnz++;
2581cf9b237SStefano Zampini       }
2591cf9b237SStefano Zampini     }
2601cf9b237SStefano Zampini     bii[i+1] = nnz;
2611cf9b237SStefano Zampini   }
2621cf9b237SStefano Zampini   ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr);
2631cf9b237SStefano Zampini   ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr);
2641cf9b237SStefano Zampini   ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr);
2651cf9b237SStefano Zampini   {
2661cf9b237SStefano Zampini     Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data);
2671cf9b237SStefano Zampini     b->free_a = PETSC_TRUE;
2681cf9b237SStefano Zampini     b->free_ij = PETSC_TRUE;
2691cf9b237SStefano Zampini   }
2701cf9b237SStefano Zampini   *B = Bt;
2711cf9b237SStefano Zampini   PetscFunctionReturn(0);
2721cf9b237SStefano Zampini }
2731cf9b237SStefano Zampini 
274674ae819SStefano Zampini #undef __FUNCT__
2754f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents"
2764f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[])
2774f1b2e48SStefano Zampini {
2784f1b2e48SStefano Zampini   Mat                    B;
2794f1b2e48SStefano Zampini   IS                     is_dummy,*cc_n;
2804f1b2e48SStefano Zampini   ISLocalToGlobalMapping l2gmap_dummy;
2814f1b2e48SStefano Zampini   PCBDDCGraph            graph;
2824f1b2e48SStefano Zampini   PetscInt               i,n;
2834f1b2e48SStefano Zampini   PetscInt               *xadj,*adjncy;
2844f1b2e48SStefano Zampini   PetscInt               *xadj_filtered,*adjncy_filtered;
2854f1b2e48SStefano Zampini   PetscBool              flg_row,isseqaij;
2864f1b2e48SStefano Zampini   PetscErrorCode         ierr;
2874f1b2e48SStefano Zampini 
2884f1b2e48SStefano Zampini   PetscFunctionBegin;
28963c961adSStefano Zampini   if (!A->rmap->N || !A->cmap->N) {
29063c961adSStefano Zampini     *ncc = 0;
29163c961adSStefano Zampini     *cc = NULL;
29263c961adSStefano Zampini     PetscFunctionReturn(0);
29363c961adSStefano Zampini   }
2944f1b2e48SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
2954f1b2e48SStefano Zampini   if (!isseqaij && filter) {
2961cf9b237SStefano Zampini     PetscBool isseqdense;
2971cf9b237SStefano Zampini 
2981cf9b237SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr);
2991cf9b237SStefano Zampini     if (!isseqdense) {
3004f1b2e48SStefano Zampini       ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr);
3011cf9b237SStefano Zampini     } else { /* TODO: rectangular case and LDA */
3021cf9b237SStefano Zampini       PetscScalar *array;
3031cf9b237SStefano Zampini       PetscReal   chop=1.e-6;
3041cf9b237SStefano Zampini 
3051cf9b237SStefano Zampini       ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr);
3061cf9b237SStefano Zampini       ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr);
3071cf9b237SStefano Zampini       ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr);
3081cf9b237SStefano Zampini       for (i=0;i<n;i++) {
3091cf9b237SStefano Zampini         PetscInt j;
3101cf9b237SStefano Zampini         for (j=i+1;j<n;j++) {
3111cf9b237SStefano Zampini           PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)]));
3121cf9b237SStefano Zampini           if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.;
3131cf9b237SStefano Zampini           if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.;
3141cf9b237SStefano Zampini         }
3151cf9b237SStefano Zampini       }
3161cf9b237SStefano Zampini       ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr);
3171cf9b237SStefano Zampini       ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr);
3181cf9b237SStefano Zampini     }
3194f1b2e48SStefano Zampini   } else {
3204f1b2e48SStefano Zampini     B = A;
3214f1b2e48SStefano Zampini   }
3224f1b2e48SStefano Zampini   ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
3234f1b2e48SStefano Zampini 
3244f1b2e48SStefano Zampini   /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */
3254f1b2e48SStefano Zampini   if (filter) {
3264f1b2e48SStefano Zampini     PetscScalar *data;
3274f1b2e48SStefano Zampini     PetscInt    j,cum;
3284f1b2e48SStefano Zampini 
3294f1b2e48SStefano Zampini     ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr);
3304f1b2e48SStefano Zampini     ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr);
3314f1b2e48SStefano Zampini     cum = 0;
3324f1b2e48SStefano Zampini     for (i=0;i<n;i++) {
3334f1b2e48SStefano Zampini       PetscInt t;
3344f1b2e48SStefano Zampini 
3354f1b2e48SStefano Zampini       for (j=xadj[i];j<xadj[i+1];j++) {
3364f1b2e48SStefano Zampini         if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) {
3374f1b2e48SStefano Zampini           continue;
3384f1b2e48SStefano Zampini         }
3394f1b2e48SStefano Zampini         adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j];
3404f1b2e48SStefano Zampini       }
3414f1b2e48SStefano Zampini       t = xadj_filtered[i];
3424f1b2e48SStefano Zampini       xadj_filtered[i] = cum;
3434f1b2e48SStefano Zampini       cum += t;
3444f1b2e48SStefano Zampini     }
3454f1b2e48SStefano Zampini     ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr);
3464f1b2e48SStefano Zampini   } else {
3474f1b2e48SStefano Zampini     xadj_filtered = NULL;
3484f1b2e48SStefano Zampini     adjncy_filtered = NULL;
3494f1b2e48SStefano Zampini   }
3504f1b2e48SStefano Zampini 
3514f1b2e48SStefano Zampini   /* compute local connected components using PCBDDCGraph */
3524f1b2e48SStefano Zampini   ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr);
3534f1b2e48SStefano Zampini   ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr);
3544f1b2e48SStefano Zampini   ierr = ISDestroy(&is_dummy);CHKERRQ(ierr);
3554f1b2e48SStefano Zampini   ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr);
3564f1b2e48SStefano Zampini   ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr);
3574f1b2e48SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr);
3584f1b2e48SStefano Zampini   if (xadj_filtered) {
3594f1b2e48SStefano Zampini     graph->xadj = xadj_filtered;
3604f1b2e48SStefano Zampini     graph->adjncy = adjncy_filtered;
3614f1b2e48SStefano Zampini   } else {
3624f1b2e48SStefano Zampini     graph->xadj = xadj;
3634f1b2e48SStefano Zampini     graph->adjncy = adjncy;
3644f1b2e48SStefano Zampini   }
3654f1b2e48SStefano Zampini   ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr);
3664f1b2e48SStefano Zampini   ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr);
3674f1b2e48SStefano Zampini   /* partial clean up */
3684f1b2e48SStefano Zampini   ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr);
3694f1b2e48SStefano Zampini   ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
3701cf9b237SStefano Zampini   if (A != B) {
3714f1b2e48SStefano Zampini     ierr = MatDestroy(&B);CHKERRQ(ierr);
3724f1b2e48SStefano Zampini   }
3734f1b2e48SStefano Zampini 
3744f1b2e48SStefano Zampini   /* get back data */
3751cf9b237SStefano Zampini   if (ncc) *ncc = graph->ncc;
3761cf9b237SStefano Zampini   if (cc) {
3774f1b2e48SStefano Zampini     ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr);
3784f1b2e48SStefano Zampini     for (i=0;i<graph->ncc;i++) {
3794f1b2e48SStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr);
3804f1b2e48SStefano Zampini     }
3814f1b2e48SStefano Zampini     *cc = cc_n;
3821cf9b237SStefano Zampini   }
3834f1b2e48SStefano Zampini   /* clean up graph */
3844f1b2e48SStefano Zampini   graph->xadj = 0;
3854f1b2e48SStefano Zampini   graph->adjncy = 0;
3864f1b2e48SStefano Zampini   ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr);
3874f1b2e48SStefano Zampini   PetscFunctionReturn(0);
3884f1b2e48SStefano Zampini }
3894f1b2e48SStefano Zampini 
3904f1b2e48SStefano Zampini #undef __FUNCT__
3915408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck"
3925408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag)
3935408967cSStefano Zampini {
3945408967cSStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
3955408967cSStefano Zampini   PC_IS*         pcis = (PC_IS*)(pc->data);
396dee84bffSStefano Zampini   IS             dirIS = NULL;
3974f1b2e48SStefano Zampini   PetscInt       i;
3985408967cSStefano Zampini   PetscErrorCode ierr;
3995408967cSStefano Zampini 
4005408967cSStefano Zampini   PetscFunctionBegin;
401dee84bffSStefano Zampini   ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr);
4025408967cSStefano Zampini   if (zerodiag) {
4035408967cSStefano Zampini     Mat            A;
4045408967cSStefano Zampini     Vec            vec3_N;
4055408967cSStefano Zampini     PetscScalar    *vals;
4065408967cSStefano Zampini     const PetscInt *idxs;
407d12d3064SStefano Zampini     PetscInt       nz,*count;
4085408967cSStefano Zampini 
4095408967cSStefano Zampini     /* p0 */
4105408967cSStefano Zampini     ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr);
4115408967cSStefano Zampini     ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr);
4125408967cSStefano Zampini     ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr);
4135408967cSStefano Zampini     ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr);
4144f1b2e48SStefano Zampini     for (i=0;i<nz;i++) vals[i] = 1.;
4155408967cSStefano Zampini     ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
4165408967cSStefano Zampini     ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr);
4175408967cSStefano Zampini     ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr);
4185408967cSStefano Zampini     /* v_I */
4195408967cSStefano Zampini     ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr);
4205408967cSStefano Zampini     for (i=0;i<nz;i++) vals[i] = 0.;
4215408967cSStefano Zampini     ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
4225408967cSStefano Zampini     ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr);
4235408967cSStefano Zampini     ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
4245408967cSStefano Zampini     for (i=0;i<pcis->n_B;i++) vals[i] = 0.;
4255408967cSStefano Zampini     ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
4265408967cSStefano Zampini     ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
4275408967cSStefano Zampini     if (dirIS) {
4285408967cSStefano Zampini       PetscInt n;
4295408967cSStefano Zampini 
4305408967cSStefano Zampini       ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr);
4315408967cSStefano Zampini       ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr);
4325408967cSStefano Zampini       for (i=0;i<n;i++) vals[i] = 0.;
4335408967cSStefano Zampini       ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
4345408967cSStefano Zampini       ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr);
4355408967cSStefano Zampini     }
4365408967cSStefano Zampini     ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr);
4375408967cSStefano Zampini     ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr);
4385408967cSStefano Zampini     ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr);
4395408967cSStefano Zampini     ierr = VecSet(vec3_N,0.);CHKERRQ(ierr);
4405408967cSStefano Zampini     ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr);
4415408967cSStefano Zampini     ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr);
4425408967cSStefano Zampini     ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr);
4434f1b2e48SStefano Zampini     if (PetscAbsScalar(vals[0]) > PETSC_SMALL) {
4445408967cSStefano Zampini       SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0]));
4455408967cSStefano Zampini     }
4465408967cSStefano Zampini     ierr = PetscFree(vals);CHKERRQ(ierr);
4475408967cSStefano Zampini     ierr = VecDestroy(&vec3_N);CHKERRQ(ierr);
448d12d3064SStefano Zampini 
449d12d3064SStefano Zampini     /* there should not be any pressure dofs lying on the interface */
450d12d3064SStefano Zampini     ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr);
451d12d3064SStefano Zampini     ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
452d12d3064SStefano Zampini     for (i=0;i<pcis->n_B;i++) count[idxs[i]]++;
453d12d3064SStefano Zampini     ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
454d12d3064SStefano Zampini     ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr);
455d4d8cf7bSStefano Zampini     for (i=0;i<nz;i++) {
456d12d3064SStefano Zampini       if (count[idxs[i]]) {
457d12d3064SStefano Zampini         SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]);
458d12d3064SStefano Zampini       }
459d4d8cf7bSStefano Zampini     }
460d12d3064SStefano Zampini     ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr);
461d12d3064SStefano Zampini     ierr = PetscFree(count);CHKERRQ(ierr);
4625408967cSStefano Zampini   }
463dee84bffSStefano Zampini   ierr = ISDestroy(&dirIS);CHKERRQ(ierr);
4645408967cSStefano Zampini 
4655408967cSStefano Zampini   /* check PCBDDCBenignGetOrSetP0 */
4665408967cSStefano Zampini   ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr);
4674f1b2e48SStefano Zampini   for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i;
4685408967cSStefano Zampini   ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr);
4694f1b2e48SStefano Zampini   for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1;
4705408967cSStefano Zampini   ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr);
4714f1b2e48SStefano Zampini   for (i=0;i<pcbddc->benign_n;i++) {
4724f1b2e48SStefano Zampini     if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) {
4734f1b2e48SStefano Zampini       SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr);
4744f1b2e48SStefano Zampini     }
4755408967cSStefano Zampini   }
4765408967cSStefano Zampini   PetscFunctionReturn(0);
4775408967cSStefano Zampini }
4785408967cSStefano Zampini 
4795408967cSStefano Zampini #undef __FUNCT__
480339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint"
481339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal)
482339f8db1SStefano Zampini {
483339f8db1SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
4844f1b2e48SStefano Zampini   IS             pressures,zerodiag,*zerodiag_subs;
485b0f5fe93SStefano Zampini   PetscInt       nz,n;
4864f1b2e48SStefano Zampini   PetscBool      sorted,have_null,has_null_pressures,recompute_zerodiag;
487339f8db1SStefano Zampini   PetscErrorCode ierr;
488339f8db1SStefano Zampini 
489339f8db1SStefano Zampini   PetscFunctionBegin;
4909f47a83aSStefano Zampini   ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr);
4919f47a83aSStefano Zampini   ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr);
492339f8db1SStefano Zampini   ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr);
493a3df083aSStefano Zampini   for (n=0;n<pcbddc->benign_n;n++) {
494a3df083aSStefano Zampini     ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr);
495a3df083aSStefano Zampini   }
496a3df083aSStefano Zampini   ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr);
497a3df083aSStefano Zampini   pcbddc->benign_n = 0;
4984f1b2e48SStefano Zampini   /* if a local info on dofs is present, assumes the last field is represented by "pressures"
4994f1b2e48SStefano Zampini      otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not)
5004f1b2e48SStefano Zampini      Checks if all the pressure dofs in each subdomain have a zero diagonal
5014f1b2e48SStefano Zampini      If not, a change of basis on pressures is not needed
5024f1b2e48SStefano Zampini      since the local Schur complements are SPD
5034f1b2e48SStefano Zampini   */
5044f1b2e48SStefano Zampini   has_null_pressures = PETSC_TRUE;
5054f1b2e48SStefano Zampini   have_null = PETSC_TRUE;
50640fa8d13SStefano Zampini   if (pcbddc->n_ISForDofsLocal) {
5074f1b2e48SStefano Zampini     PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1;
5084f1b2e48SStefano Zampini 
5094f1b2e48SStefano Zampini     /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */
5104f1b2e48SStefano Zampini     ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr);
5114f1b2e48SStefano Zampini     ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr);
5124f1b2e48SStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr);
513ba14f8e3SStefano Zampini     ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr);
51440fa8d13SStefano Zampini     ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr);
51540fa8d13SStefano Zampini     if (!sorted) {
51640fa8d13SStefano Zampini       ierr = ISSort(pressures);CHKERRQ(ierr);
51740fa8d13SStefano Zampini     }
51840fa8d13SStefano Zampini   } else {
51940fa8d13SStefano Zampini     pressures = NULL;
52040fa8d13SStefano Zampini   }
52197d764eeSStefano Zampini   /* pcis has not been setup yet, so get the local size from the subdomain matrix */
52297d764eeSStefano Zampini   ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr);
52397d764eeSStefano Zampini   ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr);
524339f8db1SStefano Zampini   ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr);
525339f8db1SStefano Zampini   if (!sorted) {
526339f8db1SStefano Zampini     ierr = ISSort(zerodiag);CHKERRQ(ierr);
527339f8db1SStefano Zampini   }
528339f8db1SStefano Zampini   ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr);
5294f1b2e48SStefano Zampini   if (!nz) {
5304f1b2e48SStefano Zampini     if (n) have_null = PETSC_FALSE;
5314f1b2e48SStefano Zampini     has_null_pressures = PETSC_FALSE;
53240fa8d13SStefano Zampini     ierr = ISDestroy(&zerodiag);CHKERRQ(ierr);
53340fa8d13SStefano Zampini   }
5344f1b2e48SStefano Zampini   recompute_zerodiag = PETSC_FALSE;
5354f1b2e48SStefano Zampini   /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */
5364f1b2e48SStefano Zampini   zerodiag_subs = NULL;
5374f1b2e48SStefano Zampini   pcbddc->benign_n = 0;
5384f1b2e48SStefano Zampini   if (has_null_pressures) {
5394f1b2e48SStefano Zampini     IS       *subs;
5404f1b2e48SStefano Zampini     PetscInt nsubs,i;
5414f1b2e48SStefano Zampini 
5424f1b2e48SStefano Zampini     subs = pcbddc->local_subs;
5434f1b2e48SStefano Zampini     nsubs = pcbddc->n_local_subs;
5444f1b2e48SStefano Zampini     if (nsubs > 1) {
5454f1b2e48SStefano Zampini       ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr);
5464f1b2e48SStefano Zampini       for (i=0;i<nsubs;i++) {
5474f1b2e48SStefano Zampini         ISLocalToGlobalMapping l2g;
5484f1b2e48SStefano Zampini         IS                     t_zerodiag_subs;
5494f1b2e48SStefano Zampini         PetscInt               nl;
5504f1b2e48SStefano Zampini 
5514f1b2e48SStefano Zampini         ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr);
5524f1b2e48SStefano Zampini         ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr);
5534f1b2e48SStefano Zampini         ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr);
5544f1b2e48SStefano Zampini         if (nl) {
5554f1b2e48SStefano Zampini           PetscBool valid = PETSC_TRUE;
5564f1b2e48SStefano Zampini 
5574f1b2e48SStefano Zampini           if (pressures) {
5584f1b2e48SStefano Zampini             IS t_pressure_subs;
5594f1b2e48SStefano Zampini             ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr);
5604f1b2e48SStefano Zampini             ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr);
5614f1b2e48SStefano Zampini             ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr);
5624f1b2e48SStefano Zampini           }
5634f1b2e48SStefano Zampini           if (valid) {
5644f1b2e48SStefano Zampini             ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr);
5654f1b2e48SStefano Zampini             pcbddc->benign_n++;
5664f1b2e48SStefano Zampini           } else {
5674f1b2e48SStefano Zampini             recompute_zerodiag = PETSC_TRUE;
5684f1b2e48SStefano Zampini           }
5694f1b2e48SStefano Zampini         }
5704f1b2e48SStefano Zampini         ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr);
5714f1b2e48SStefano Zampini         ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr);
5724f1b2e48SStefano Zampini       }
5734f1b2e48SStefano Zampini     } else { /* there's just one subdomain (or zero if they have not been detected */
5744f1b2e48SStefano Zampini       PetscBool valid = PETSC_TRUE;
5754f1b2e48SStefano Zampini       if (pressures) {
5764f1b2e48SStefano Zampini         ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr);
5774f1b2e48SStefano Zampini       }
5784f1b2e48SStefano Zampini       if (valid) {
5794f1b2e48SStefano Zampini         pcbddc->benign_n = 1;
580ca92afb2SStefano Zampini         ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr);
5814f1b2e48SStefano Zampini         ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr);
5824f1b2e48SStefano Zampini         zerodiag_subs[0] = zerodiag;
5834f1b2e48SStefano Zampini       }
5844f1b2e48SStefano Zampini     }
5854f1b2e48SStefano Zampini   }
5864f1b2e48SStefano Zampini 
5874f1b2e48SStefano Zampini   if (!pcbddc->benign_n) {
5884f1b2e48SStefano Zampini     ierr = ISDestroy(&zerodiag);CHKERRQ(ierr);
5894f1b2e48SStefano Zampini     recompute_zerodiag = PETSC_FALSE;
5904f1b2e48SStefano Zampini     has_null_pressures = PETSC_FALSE;
5914f1b2e48SStefano Zampini     have_null = PETSC_FALSE;
5924f1b2e48SStefano Zampini   }
5934f1b2e48SStefano Zampini 
5944f1b2e48SStefano Zampini   /* final check for null pressures */
5954f1b2e48SStefano Zampini   if (zerodiag && pressures) {
5964f1b2e48SStefano Zampini     PetscInt nz,np;
5974f1b2e48SStefano Zampini     ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr);
5984f1b2e48SStefano Zampini     ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr);
5994f1b2e48SStefano Zampini     if (nz != np) have_null = PETSC_FALSE;
6004f1b2e48SStefano Zampini   }
6014f1b2e48SStefano Zampini 
6024f1b2e48SStefano Zampini   if (recompute_zerodiag) {
6034f1b2e48SStefano Zampini     ierr = ISDestroy(&zerodiag);CHKERRQ(ierr);
6044f1b2e48SStefano Zampini     if (pcbddc->benign_n == 1) {
6054f1b2e48SStefano Zampini       ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr);
6064f1b2e48SStefano Zampini       zerodiag = zerodiag_subs[0];
6074f1b2e48SStefano Zampini     } else {
6084f1b2e48SStefano Zampini       PetscInt i,nzn,*new_idxs;
6094f1b2e48SStefano Zampini 
6104f1b2e48SStefano Zampini       nzn = 0;
6114f1b2e48SStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) {
6124f1b2e48SStefano Zampini         PetscInt ns;
6134f1b2e48SStefano Zampini         ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr);
6144f1b2e48SStefano Zampini         nzn += ns;
6154f1b2e48SStefano Zampini       }
6164f1b2e48SStefano Zampini       ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr);
6174f1b2e48SStefano Zampini       nzn = 0;
6184f1b2e48SStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) {
6194f1b2e48SStefano Zampini         PetscInt ns,*idxs;
6204f1b2e48SStefano Zampini         ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr);
6214f1b2e48SStefano Zampini         ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr);
6224f1b2e48SStefano Zampini         ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr);
6234f1b2e48SStefano Zampini         ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr);
6244f1b2e48SStefano Zampini         nzn += ns;
6254f1b2e48SStefano Zampini       }
6264f1b2e48SStefano Zampini       ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr);
6274f1b2e48SStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr);
6284f1b2e48SStefano Zampini     }
6294f1b2e48SStefano Zampini     have_null = PETSC_FALSE;
6304f1b2e48SStefano Zampini   }
6314f1b2e48SStefano Zampini 
6324f1b2e48SStefano Zampini   if (has_null_pressures) {
6334f1b2e48SStefano Zampini     IS             zerodiagc;
6344f1b2e48SStefano Zampini     const PetscInt *idxs,*idxsc;
6354f1b2e48SStefano Zampini     PetscInt       i,s,*nnz;
6364f1b2e48SStefano Zampini 
6374f1b2e48SStefano Zampini     ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr);
638339f8db1SStefano Zampini     ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr);
639339f8db1SStefano Zampini     ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr);
640339f8db1SStefano Zampini     /* local change of basis for pressures */
641339f8db1SStefano Zampini     ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr);
64297d764eeSStefano Zampini     ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr);
643339f8db1SStefano Zampini     ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr);
644339f8db1SStefano Zampini     ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
645339f8db1SStefano Zampini     ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr);
6464f1b2e48SStefano Zampini     for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */
6474f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
6484f1b2e48SStefano Zampini       PetscInt nzs,j;
6494f1b2e48SStefano Zampini 
6504f1b2e48SStefano Zampini       ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr);
6514f1b2e48SStefano Zampini       ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr);
6524f1b2e48SStefano Zampini       for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */
6534f1b2e48SStefano Zampini       nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */
6544f1b2e48SStefano Zampini       ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr);
6554f1b2e48SStefano Zampini     }
656339f8db1SStefano Zampini     ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr);
657339f8db1SStefano Zampini     ierr = PetscFree(nnz);CHKERRQ(ierr);
658339f8db1SStefano Zampini     /* set identity on velocities */
659339f8db1SStefano Zampini     for (i=0;i<n-nz;i++) {
660339f8db1SStefano Zampini       ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr);
661339f8db1SStefano Zampini     }
6624f1b2e48SStefano Zampini     ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr);
6634f1b2e48SStefano Zampini     ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr);
6649f47a83aSStefano Zampini     ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr);
6654f1b2e48SStefano Zampini     ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr);
666339f8db1SStefano Zampini     /* set change on pressures */
6674f1b2e48SStefano Zampini     for (s=0;s<pcbddc->benign_n;s++) {
6684f1b2e48SStefano Zampini       PetscScalar *array;
6694f1b2e48SStefano Zampini       PetscInt    nzs;
6704f1b2e48SStefano Zampini 
6714f1b2e48SStefano Zampini       ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr);
6724f1b2e48SStefano Zampini       ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr);
6734f1b2e48SStefano Zampini       for (i=0;i<nzs-1;i++) {
674339f8db1SStefano Zampini         PetscScalar vals[2];
675339f8db1SStefano Zampini         PetscInt    cols[2];
676339f8db1SStefano Zampini 
677339f8db1SStefano Zampini         cols[0] = idxs[i];
6784f1b2e48SStefano Zampini         cols[1] = idxs[nzs-1];
679339f8db1SStefano Zampini         vals[0] = 1.;
680b0f5fe93SStefano Zampini         vals[1] = 1.;
6814f1b2e48SStefano Zampini         ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr);
682339f8db1SStefano Zampini       }
6834f1b2e48SStefano Zampini       ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr);
6844f1b2e48SStefano Zampini       for (i=0;i<nzs-1;i++) array[i] = -1.;
6854f1b2e48SStefano Zampini       array[nzs-1] = 1.;
6864f1b2e48SStefano Zampini       ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr);
6874f1b2e48SStefano Zampini       /* store local idxs for p0 */
6884f1b2e48SStefano Zampini       pcbddc->benign_p0_lidx[s] = idxs[nzs-1];
6894f1b2e48SStefano Zampini       ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr);
690339f8db1SStefano Zampini       ierr = PetscFree(array);CHKERRQ(ierr);
6914f1b2e48SStefano Zampini     }
692339f8db1SStefano Zampini     ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
693339f8db1SStefano Zampini     ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
694a3df083aSStefano Zampini     /* project if needed */
695a3df083aSStefano Zampini     if (pcbddc->benign_change_explicit) {
69697d764eeSStefano Zampini       ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr);
697339f8db1SStefano Zampini       ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
69897d764eeSStefano Zampini       ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr);
69997d764eeSStefano Zampini       ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr);
700a3df083aSStefano Zampini     }
7014f1b2e48SStefano Zampini     /* store global idxs for p0 */
7024f1b2e48SStefano Zampini     ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr);
703339f8db1SStefano Zampini   }
704ca92afb2SStefano Zampini   pcbddc->benign_zerodiag_subs = zerodiag_subs;
7054f1b2e48SStefano Zampini   ierr = ISDestroy(&pressures);CHKERRQ(ierr);
706b0f5fe93SStefano Zampini 
707b0f5fe93SStefano Zampini   /* determines if the coarse solver will be singular or not */
708b0f5fe93SStefano Zampini   ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
709339f8db1SStefano Zampini   *zerodiaglocal = zerodiag;
710339f8db1SStefano Zampini   PetscFunctionReturn(0);
711339f8db1SStefano Zampini }
712339f8db1SStefano Zampini 
713339f8db1SStefano Zampini #undef __FUNCT__
714015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0"
715015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get)
716efc2fbd9SStefano Zampini {
717efc2fbd9SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
718efc2fbd9SStefano Zampini   PetscErrorCode ierr;
719efc2fbd9SStefano Zampini 
720efc2fbd9SStefano Zampini   PetscFunctionBegin;
721efc2fbd9SStefano Zampini   if (!pcbddc->benign_sf) {
722efc2fbd9SStefano Zampini     ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr);
7234f1b2e48SStefano Zampini     ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr);
724efc2fbd9SStefano Zampini   }
725015636ebSStefano Zampini   if (get) { /* use SF to get values */
726efc2fbd9SStefano Zampini     PetscScalar *array;
727efc2fbd9SStefano Zampini 
728efc2fbd9SStefano Zampini     ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr);
7294f1b2e48SStefano Zampini     ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr);
7304f1b2e48SStefano Zampini     ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr);
731efc2fbd9SStefano Zampini     ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr);
7324f1b2e48SStefano Zampini   } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */
7334f1b2e48SStefano Zampini     ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr);
734efc2fbd9SStefano Zampini     ierr = VecAssemblyBegin(v);CHKERRQ(ierr);
735efc2fbd9SStefano Zampini     ierr = VecAssemblyEnd(v);CHKERRQ(ierr);
736efc2fbd9SStefano Zampini   }
737efc2fbd9SStefano Zampini   PetscFunctionReturn(0);
738efc2fbd9SStefano Zampini }
739efc2fbd9SStefano Zampini 
740efc2fbd9SStefano Zampini #undef __FUNCT__
741c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0"
742c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop)
743c263805aSStefano Zampini {
744c263805aSStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
745c263805aSStefano Zampini   PetscErrorCode ierr;
746c263805aSStefano Zampini 
747c263805aSStefano Zampini   PetscFunctionBegin;
748c263805aSStefano Zampini   /* TODO: add error checking
749c263805aSStefano Zampini     - avoid nested pop (or push) calls.
750c263805aSStefano Zampini     - cannot push before pop.
7511c604dc7SStefano Zampini     - cannot call this if pcbddc->local_mat is NULL
752c263805aSStefano Zampini   */
7534f1b2e48SStefano Zampini   if (!pcbddc->benign_n) {
754efc2fbd9SStefano Zampini     PetscFunctionReturn(0);
755efc2fbd9SStefano Zampini   }
756c263805aSStefano Zampini   if (pop) {
757a3df083aSStefano Zampini     if (pcbddc->benign_change_explicit) {
7584f1b2e48SStefano Zampini       IS       is_p0;
7594f1b2e48SStefano Zampini       MatReuse reuse;
760c263805aSStefano Zampini 
761c263805aSStefano Zampini       /* extract B_0 */
7624f1b2e48SStefano Zampini       reuse = MAT_INITIAL_MATRIX;
7634f1b2e48SStefano Zampini       if (pcbddc->benign_B0) {
7644f1b2e48SStefano Zampini         reuse = MAT_REUSE_MATRIX;
7654f1b2e48SStefano Zampini       }
7664f1b2e48SStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr);
7674f1b2e48SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr);
768c263805aSStefano Zampini       /* remove rows and cols from local problem */
769c263805aSStefano Zampini       ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr);
77097d764eeSStefano Zampini       ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
7714f1b2e48SStefano Zampini       ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr);
7724f1b2e48SStefano Zampini       ierr = ISDestroy(&is_p0);CHKERRQ(ierr);
773a3df083aSStefano Zampini     } else {
774a3df083aSStefano Zampini       Mat_IS      *matis = (Mat_IS*)pc->pmat->data;
775a3df083aSStefano Zampini       PetscScalar *vals;
776a3df083aSStefano Zampini       PetscInt    i,n,*idxs_ins;
777a3df083aSStefano Zampini 
778a3df083aSStefano Zampini       ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr);
779a3df083aSStefano Zampini       ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr);
780a3df083aSStefano Zampini       if (!pcbddc->benign_B0) {
781a3df083aSStefano Zampini         ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr);
782a3df083aSStefano Zampini         ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr);
783a3df083aSStefano Zampini         ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
784331e053bSStefano Zampini         if (!pcbddc->current_level) {
785331e053bSStefano Zampini           /* this matrix is actually sparse: the nnz pattern is not known unless we do 2 sweeps of the next loop.
786331e053bSStefano Zampini              Setting nnz=100 should be more than enough for the finest level */
787a3df083aSStefano Zampini           ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,100,NULL);CHKERRQ(ierr);
788331e053bSStefano Zampini         } else {
789331e053bSStefano Zampini           PetscInt *nnz;
790331e053bSStefano Zampini           ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr);
791331e053bSStefano Zampini           for (i=0;i<pcbddc->benign_n;i++) {
792331e053bSStefano Zampini             ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr);
793331e053bSStefano Zampini             nnz[i] = n - nnz[i];
794331e053bSStefano Zampini           }
795331e053bSStefano Zampini           ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr);
796331e053bSStefano Zampini           ierr = PetscFree(nnz);CHKERRQ(ierr);
797331e053bSStefano Zampini         }
798a3df083aSStefano Zampini       }
799a3df083aSStefano Zampini 
800a3df083aSStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) {
801a3df083aSStefano Zampini         PetscScalar *array;
802a3df083aSStefano Zampini         PetscInt    *idxs,j,nz,cum;
803a3df083aSStefano Zampini 
804a3df083aSStefano Zampini         ierr = VecSet(matis->x,0.);CHKERRQ(ierr);
805a3df083aSStefano Zampini         ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr);
806a3df083aSStefano Zampini         ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr);
807a3df083aSStefano Zampini         for (j=0;j<nz;j++) vals[j] = 1.;
808a3df083aSStefano Zampini         ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
809a3df083aSStefano Zampini         ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr);
810a3df083aSStefano Zampini         ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr);
811a3df083aSStefano Zampini         ierr = VecSet(matis->y,0.);CHKERRQ(ierr);
812a3df083aSStefano Zampini         ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr);
813a3df083aSStefano Zampini         ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr);
814a3df083aSStefano Zampini         cum = 0;
815a3df083aSStefano Zampini         for (j=0;j<n;j++) {
816a3df083aSStefano Zampini           if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) {
817a3df083aSStefano Zampini             vals[cum] = array[j];
818a3df083aSStefano Zampini             idxs_ins[cum] = j;
819a3df083aSStefano Zampini             cum++;
820a3df083aSStefano Zampini           }
821a3df083aSStefano Zampini         }
822a3df083aSStefano Zampini         ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr);
823a3df083aSStefano Zampini         ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr);
824a3df083aSStefano Zampini         ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr);
825a3df083aSStefano Zampini       }
826a3df083aSStefano Zampini       ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
827a3df083aSStefano Zampini       ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
828a3df083aSStefano Zampini       ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr);
829a3df083aSStefano Zampini     }
830c263805aSStefano Zampini   } else { /* push */
831a3df083aSStefano Zampini     if (pcbddc->benign_change_explicit) {
8324f1b2e48SStefano Zampini       PetscInt i;
8334f1b2e48SStefano Zampini 
8344f1b2e48SStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) {
8354f1b2e48SStefano Zampini         PetscScalar *B0_vals;
8364f1b2e48SStefano Zampini         PetscInt    *B0_cols,B0_ncol;
8374f1b2e48SStefano Zampini 
8384f1b2e48SStefano Zampini         ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr);
8394f1b2e48SStefano Zampini         ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr);
8407b034428SStefano Zampini         ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr);
8414f1b2e48SStefano Zampini         ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr);
8424f1b2e48SStefano Zampini         ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr);
8434f1b2e48SStefano Zampini       }
844c263805aSStefano Zampini       ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
845c263805aSStefano Zampini       ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
846a3df083aSStefano Zampini     } else {
847a3df083aSStefano Zampini       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n");
848a3df083aSStefano Zampini     }
849c263805aSStefano Zampini   }
850c263805aSStefano Zampini   PetscFunctionReturn(0);
851c263805aSStefano Zampini }
852c263805aSStefano Zampini 
853c263805aSStefano Zampini #undef __FUNCT__
854b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection"
85508122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc)
856b1b3d7a2SStefano Zampini {
857b1b3d7a2SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)pc->data;
85808122e43SStefano Zampini   PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
85908122e43SStefano Zampini   PetscBLASInt    B_dummyint,B_neigs,B_ierr,B_lwork;
86008122e43SStefano Zampini   PetscBLASInt    *B_iwork,*B_ifail;
86108122e43SStefano Zampini   PetscScalar     *work,lwork;
86208122e43SStefano Zampini   PetscScalar     *St,*S,*eigv;
86308122e43SStefano Zampini   PetscScalar     *Sarray,*Starray;
86408122e43SStefano Zampini   PetscReal       *eigs,thresh;
8651b968477SStefano Zampini   PetscInt        i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs;
866f6f667cfSStefano Zampini   PetscBool       allocated_S_St;
86708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
86808122e43SStefano Zampini   PetscReal       *rwork;
86908122e43SStefano Zampini #endif
870b1b3d7a2SStefano Zampini   PetscErrorCode  ierr;
871b1b3d7a2SStefano Zampini 
872b1b3d7a2SStefano Zampini   PetscFunctionBegin;
873df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) {
874df4d28bfSStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO");
87508122e43SStefano Zampini   }
87608122e43SStefano Zampini 
87706a4e24aSStefano Zampini   if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) {
87806a4e24aSStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef);
87906a4e24aSStefano Zampini   }
88006a4e24aSStefano Zampini 
881fd14bc51SStefano Zampini   if (pcbddc->dbg_flag) {
882fd14bc51SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
883fd14bc51SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
884fd14bc51SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr);
8851575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
886fd14bc51SStefano Zampini   }
887fd14bc51SStefano Zampini 
888e496cd5dSStefano Zampini   if (pcbddc->dbg_flag) {
889e496cd5dSStefano Zampini     PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef);
890e496cd5dSStefano Zampini   }
891e496cd5dSStefano Zampini 
89208122e43SStefano Zampini   /* max size of subsets */
89308122e43SStefano Zampini   mss = 0;
89408122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
89508122e43SStefano Zampini     PetscInt subset_size;
896862806e4SStefano Zampini 
89708122e43SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
89808122e43SStefano Zampini     mss = PetscMax(mss,subset_size);
89908122e43SStefano Zampini   }
90008122e43SStefano Zampini 
90108122e43SStefano Zampini   /* min/max and threshold */
90208122e43SStefano Zampini   nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss;
903f6f667cfSStefano Zampini   nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0;
90408122e43SStefano Zampini   nmax = PetscMax(nmin,nmax);
905f6f667cfSStefano Zampini   allocated_S_St = PETSC_FALSE;
906f6f667cfSStefano Zampini   if (nmin) {
907f6f667cfSStefano Zampini     allocated_S_St = PETSC_TRUE;
908f6f667cfSStefano Zampini   }
90908122e43SStefano Zampini 
91008122e43SStefano Zampini   /* allocate lapack workspace */
91108122e43SStefano Zampini   cum = cum2 = 0;
91208122e43SStefano Zampini   maxneigs = 0;
91308122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
91408122e43SStefano Zampini     PetscInt n,subset_size;
915f6f667cfSStefano Zampini 
91608122e43SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
91708122e43SStefano Zampini     n = PetscMin(subset_size,nmax);
9189162d606SStefano Zampini     cum += subset_size;
9199162d606SStefano Zampini     cum2 += subset_size*n;
92008122e43SStefano Zampini     maxneigs = PetscMax(maxneigs,n);
92108122e43SStefano Zampini   }
92208122e43SStefano Zampini   if (mss) {
9239ab7bb16SStefano Zampini     if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
92408122e43SStefano Zampini       PetscBLASInt B_itype = 1;
92508122e43SStefano Zampini       PetscBLASInt B_N = mss;
9264c6709b3SStefano Zampini       PetscReal    zero = 0.0;
9274c6709b3SStefano Zampini       PetscReal    eps = 0.0; /* dlamch? */
92808122e43SStefano Zampini 
92908122e43SStefano Zampini       B_lwork = -1;
93008122e43SStefano Zampini       S = NULL;
93108122e43SStefano Zampini       St = NULL;
932a58a30b4SStefano Zampini       eigs = NULL;
933a58a30b4SStefano Zampini       eigv = NULL;
934a58a30b4SStefano Zampini       B_iwork = NULL;
935a58a30b4SStefano Zampini       B_ifail = NULL;
936d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX)
937d1710679SStefano Zampini       rwork = NULL;
938d1710679SStefano Zampini #endif
9398bec7fa6SStefano Zampini       thresh = 1.0;
94008122e43SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
94108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
94208122e43SStefano Zampini       PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
94308122e43SStefano Zampini #else
94408122e43SStefano Zampini       PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr));
94508122e43SStefano Zampini #endif
94608122e43SStefano Zampini       if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr);
94708122e43SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
94808122e43SStefano Zampini     } else {
94908122e43SStefano Zampini         /* TODO */
95008122e43SStefano Zampini     }
95108122e43SStefano Zampini   } else {
95208122e43SStefano Zampini     lwork = 0;
95308122e43SStefano Zampini   }
95408122e43SStefano Zampini 
95508122e43SStefano Zampini   nv = 0;
956d62866d3SStefano Zampini   if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */
957d62866d3SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr);
95808122e43SStefano Zampini   }
9594c6709b3SStefano Zampini   ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr);
960f6f667cfSStefano Zampini   if (allocated_S_St) {
961f6f667cfSStefano Zampini     ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr);
962f6f667cfSStefano Zampini   }
963f6f667cfSStefano Zampini   ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr);
96408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
96508122e43SStefano Zampini   ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr);
96608122e43SStefano Zampini #endif
9679162d606SStefano Zampini   ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n,
9689162d606SStefano Zampini                       nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr,
9699162d606SStefano Zampini                       nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr,
97008122e43SStefano Zampini                       nv+cum,&pcbddc->adaptive_constraints_idxs,
9719162d606SStefano Zampini                       nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr);
97208122e43SStefano Zampini   ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr);
97308122e43SStefano Zampini 
97408122e43SStefano Zampini   maxneigs = 0;
97508122e43SStefano Zampini   cum = cum2 = cumarray = 0;
9769162d606SStefano Zampini   pcbddc->adaptive_constraints_idxs_ptr[0] = 0;
9779162d606SStefano Zampini   pcbddc->adaptive_constraints_data_ptr[0] = 0;
978d62866d3SStefano Zampini   if (sub_schurs->is_vertices && pcbddc->use_vertices) {
97908122e43SStefano Zampini     const PetscInt *idxs;
98008122e43SStefano Zampini 
981d62866d3SStefano Zampini     ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr);
98208122e43SStefano Zampini     for (cum=0;cum<nv;cum++) {
98308122e43SStefano Zampini       pcbddc->adaptive_constraints_n[cum] = 1;
98408122e43SStefano Zampini       pcbddc->adaptive_constraints_idxs[cum] = idxs[cum];
98508122e43SStefano Zampini       pcbddc->adaptive_constraints_data[cum] = 1.0;
9869162d606SStefano Zampini       pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1;
9879162d606SStefano Zampini       pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1;
98808122e43SStefano Zampini     }
98908122e43SStefano Zampini     cum2 = cum;
990d62866d3SStefano Zampini     ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr);
99108122e43SStefano Zampini   }
99208122e43SStefano Zampini 
99308122e43SStefano Zampini   if (mss) { /* multilevel */
99408122e43SStefano Zampini     ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr);
99508122e43SStefano Zampini     ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr);
99608122e43SStefano Zampini   }
99708122e43SStefano Zampini 
998ffd830a3SStefano Zampini   thresh = pcbddc->adaptive_threshold;
99908122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
100008122e43SStefano Zampini     const PetscInt *idxs;
1001f6f667cfSStefano Zampini     PetscReal      infty = PETSC_MAX_REAL;
1002862806e4SStefano Zampini     PetscInt       j,subset_size,eigs_start = 0;
100308122e43SStefano Zampini     PetscBLASInt   B_N;
1004aff50787SStefano Zampini     PetscBool      same_data = PETSC_FALSE;
100508122e43SStefano Zampini 
1006862806e4SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
1007ffd830a3SStefano Zampini     ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
1008f6f667cfSStefano Zampini     ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr);
1009f6f667cfSStefano Zampini     if (allocated_S_St) { /* S and S_t should be copied since we could need them later */
10109ab7bb16SStefano Zampini       if (sub_schurs->is_hermitian) {
1011aff50787SStefano Zampini         PetscInt j,k;
1012aff50787SStefano Zampini         if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */
1013aff50787SStefano Zampini           ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
1014aff50787SStefano Zampini           ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
101508122e43SStefano Zampini         }
101608122e43SStefano Zampini         for (j=0;j<subset_size;j++) {
1017aff50787SStefano Zampini           for (k=j;k<subset_size;k++) {
1018aff50787SStefano Zampini             S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k];
1019aff50787SStefano Zampini             St[j*subset_size+k] = Starray[cumarray+j*subset_size+k];
1020aff50787SStefano Zampini           }
102108122e43SStefano Zampini         }
102208122e43SStefano Zampini       } else {
102308122e43SStefano Zampini         ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
102408122e43SStefano Zampini         ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
102508122e43SStefano Zampini       }
10268bec7fa6SStefano Zampini     } else {
1027f6f667cfSStefano Zampini       S = Sarray + cumarray;
1028f6f667cfSStefano Zampini       St = Starray + cumarray;
10298bec7fa6SStefano Zampini     }
103008122e43SStefano Zampini 
1031aff50787SStefano Zampini     /* see if we can save some work */
1032aff50787SStefano Zampini     if (sub_schurs->n_subs == 1) {
1033aff50787SStefano Zampini       ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr);
1034aff50787SStefano Zampini     }
1035aff50787SStefano Zampini 
1036aff50787SStefano Zampini     if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */
1037aff50787SStefano Zampini       B_neigs = 0;
1038aff50787SStefano Zampini     } else {
10399ab7bb16SStefano Zampini       if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
104008122e43SStefano Zampini         PetscBLASInt B_itype = 1;
1041f6f667cfSStefano Zampini         PetscBLASInt B_IL, B_IU;
10424c6709b3SStefano Zampini         PetscReal    eps = -1.0; /* dlamch? */
10439552c7c7SStefano Zampini         PetscInt     nmin_s;
104408122e43SStefano Zampini 
1045fd14bc51SStefano Zampini         if (pcbddc->dbg_flag) {
10468bec7fa6SStefano Zampini           PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]);
1047fd14bc51SStefano Zampini         }
1048d16cbb6bSStefano Zampini 
104908122e43SStefano Zampini         ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
1050d16cbb6bSStefano Zampini         if (thresh > 1.+PETSC_SMALL) {
1051d16cbb6bSStefano Zampini 
1052d16cbb6bSStefano Zampini           /* ask for eigenvalues larger than thresh */
105308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
1054f6f667cfSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
105508122e43SStefano Zampini #else
1056f6f667cfSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
105708122e43SStefano Zampini #endif
1058d16cbb6bSStefano Zampini         } else {
1059d16cbb6bSStefano Zampini           B_IU = PetscMax(1,PetscMin(B_N,nmax));
1060d16cbb6bSStefano Zampini           B_IL = 1;
1061d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX)
1062d16cbb6bSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
1063d16cbb6bSStefano Zampini #else
1064d16cbb6bSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
1065d16cbb6bSStefano Zampini #endif
1066d16cbb6bSStefano Zampini         }
106708122e43SStefano Zampini         ierr = PetscFPTrapPop();CHKERRQ(ierr);
106808122e43SStefano Zampini         if (B_ierr) {
106908122e43SStefano Zampini           if (B_ierr < 0 ) {
107008122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
107108122e43SStefano Zampini           } else if (B_ierr <= B_N) {
107208122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
107308122e43SStefano Zampini           } else {
10749552c7c7SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1);
107508122e43SStefano Zampini           }
107608122e43SStefano Zampini         }
107708122e43SStefano Zampini 
107808122e43SStefano Zampini         if (B_neigs > nmax) {
1079fd14bc51SStefano Zampini           if (pcbddc->dbg_flag) {
1080fd14bc51SStefano Zampini             PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   found %d eigs, more than maximum required %d.\n",B_neigs,nmax);
1081fd14bc51SStefano Zampini           }
1082f6f667cfSStefano Zampini           eigs_start = B_neigs -nmax;
108308122e43SStefano Zampini           B_neigs = nmax;
108408122e43SStefano Zampini         }
108508122e43SStefano Zampini 
10869552c7c7SStefano Zampini         nmin_s = PetscMin(nmin,B_N);
10879552c7c7SStefano Zampini         if (B_neigs < nmin_s) {
108808122e43SStefano Zampini           PetscBLASInt B_neigs2;
108908122e43SStefano Zampini 
1090f6f667cfSStefano Zampini           B_IU = B_N - B_neigs;
1091f6f667cfSStefano Zampini           B_IL = B_N - nmin_s + 1;
1092fd14bc51SStefano Zampini           if (pcbddc->dbg_flag) {
1093fd14bc51SStefano Zampini             PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU);
1094fd14bc51SStefano Zampini           }
10959ab7bb16SStefano Zampini           if (sub_schurs->is_hermitian) {
109608122e43SStefano Zampini             PetscInt j;
109708122e43SStefano Zampini             for (j=0;j<subset_size;j++) {
109808122e43SStefano Zampini               ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
109908122e43SStefano Zampini             }
110008122e43SStefano Zampini             for (j=0;j<subset_size;j++) {
110108122e43SStefano Zampini               ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
110208122e43SStefano Zampini             }
110308122e43SStefano Zampini           } else {
110408122e43SStefano Zampini             ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
110508122e43SStefano Zampini             ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
110608122e43SStefano Zampini           }
110708122e43SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
110808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
1109f6f667cfSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
111008122e43SStefano Zampini #else
1111f6f667cfSStefano Zampini           PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
111208122e43SStefano Zampini #endif
111308122e43SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
111408122e43SStefano Zampini           B_neigs += B_neigs2;
111508122e43SStefano Zampini         }
111608122e43SStefano Zampini         if (B_ierr) {
111708122e43SStefano Zampini           if (B_ierr < 0 ) {
111808122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
111908122e43SStefano Zampini           } else if (B_ierr <= B_N) {
112008122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
112108122e43SStefano Zampini           } else {
11229552c7c7SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1);
112308122e43SStefano Zampini           }
112408122e43SStefano Zampini         }
1125fd14bc51SStefano Zampini         if (pcbddc->dbg_flag) {
1126ac47001eSStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   -> Got %d eigs\n",B_neigs);CHKERRQ(ierr);
112708122e43SStefano Zampini           for (j=0;j<B_neigs;j++) {
112808122e43SStefano Zampini             if (eigs[j] == 0.0) {
1129ac47001eSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"     Inf\n");CHKERRQ(ierr);
113008122e43SStefano Zampini             } else {
1131ac47001eSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"     %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr);
1132fd14bc51SStefano Zampini             }
113308122e43SStefano Zampini           }
113408122e43SStefano Zampini         }
113508122e43SStefano Zampini       } else {
113608122e43SStefano Zampini           /* TODO */
113708122e43SStefano Zampini       }
1138aff50787SStefano Zampini     }
11398bec7fa6SStefano Zampini     maxneigs = PetscMax(B_neigs,maxneigs);
11408bec7fa6SStefano Zampini     pcbddc->adaptive_constraints_n[i+nv] = B_neigs;
11419162d606SStefano Zampini     if (B_neigs) {
11429162d606SStefano Zampini       ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
1143fd14bc51SStefano Zampini 
1144fd14bc51SStefano Zampini       if (pcbddc->dbg_flag > 1) {
11459552c7c7SStefano Zampini         PetscInt ii;
11469552c7c7SStefano Zampini         for (ii=0;ii<B_neigs;ii++) {
1147ac47001eSStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr);
11489552c7c7SStefano Zampini           for (j=0;j<B_N;j++) {
1149ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX)
1150ac47001eSStefano Zampini             PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);
1151ac47001eSStefano Zampini             PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);
1152ac47001eSStefano Zampini             ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"       %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr);
1153ac47001eSStefano Zampini #else
1154ac47001eSStefano Zampini             ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"       %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr);
1155ac47001eSStefano Zampini #endif
11569552c7c7SStefano Zampini           }
11579552c7c7SStefano Zampini         }
1158fd14bc51SStefano Zampini       }
115908122e43SStefano Zampini #if 0
11609162d606SStefano Zampini       for (j=0;j<B_neigs;j++) {
116108122e43SStefano Zampini         PetscBLASInt Blas_N,Blas_one = 1.0;
116208122e43SStefano Zampini         PetscScalar norm;
116308122e43SStefano Zampini         ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr);
11649162d606SStefano Zampini         PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,
11659162d606SStefano Zampini                                                    &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one));
116608122e43SStefano Zampini         if (pcbddc->adaptive_constraints_data[cum2] > 0.0) {
116708122e43SStefano Zampini           norm = 1.0/PetscSqrtReal(PetscRealPart(norm));
116808122e43SStefano Zampini         } else {
116908122e43SStefano Zampini           norm = -1.0/PetscSqrtReal(PetscRealPart(norm));
117008122e43SStefano Zampini         }
11719162d606SStefano Zampini         PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one));
1172b1b3d7a2SStefano Zampini       }
1173b1b3d7a2SStefano Zampini #endif
11749162d606SStefano Zampini       ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr);
11759162d606SStefano Zampini       pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size;
11769162d606SStefano Zampini       pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs;
11779162d606SStefano Zampini       cum++;
117808122e43SStefano Zampini     }
117908122e43SStefano Zampini     ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
118008122e43SStefano Zampini     /* shift for next computation */
118108122e43SStefano Zampini     cumarray += subset_size*subset_size;
118208122e43SStefano Zampini   }
1183fd14bc51SStefano Zampini   if (pcbddc->dbg_flag) {
1184fd14bc51SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1185fd14bc51SStefano Zampini   }
118608122e43SStefano Zampini 
118708122e43SStefano Zampini   if (mss) {
118808122e43SStefano Zampini     ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr);
118908122e43SStefano Zampini     ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr);
1190f6f667cfSStefano Zampini     /* destroy matrices (junk) */
1191f6f667cfSStefano Zampini     ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr);
1192f6f667cfSStefano Zampini     ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr);
119308122e43SStefano Zampini   }
1194f6f667cfSStefano Zampini   if (allocated_S_St) {
1195f6f667cfSStefano Zampini     ierr = PetscFree2(S,St);CHKERRQ(ierr);
1196f6f667cfSStefano Zampini   }
1197f6f667cfSStefano Zampini   ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr);
119808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
119908122e43SStefano Zampini   ierr = PetscFree(rwork);CHKERRQ(ierr);
120008122e43SStefano Zampini #endif
120108122e43SStefano Zampini   if (pcbddc->dbg_flag) {
12021b968477SStefano Zampini     PetscInt maxneigs_r;
120308122e43SStefano Zampini     ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
12049b28b941SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr);
120508122e43SStefano Zampini   }
120608122e43SStefano Zampini   PetscFunctionReturn(0);
120708122e43SStefano Zampini }
1208b1b3d7a2SStefano Zampini 
1209674ae819SStefano Zampini #undef __FUNCT__
1210c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers"
1211c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc)
1212c8587f34SStefano Zampini {
1213c8587f34SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
12148629588bSStefano Zampini   PetscScalar    *coarse_submat_vals;
1215c8587f34SStefano Zampini   PetscErrorCode ierr;
1216c8587f34SStefano Zampini 
1217c8587f34SStefano Zampini   PetscFunctionBegin;
1218f4ddd8eeSStefano Zampini   /* Setup local scatters R_to_B and (optionally) R_to_D */
12195e8657edSStefano Zampini   /* PCBDDCSetUpLocalWorkVectors should be called first! */
1220c8587f34SStefano Zampini   ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr);
1221c8587f34SStefano Zampini 
1222684f6988SStefano Zampini   /* Setup local neumann solver ksp_R */
12230fccc4e9SStefano Zampini   /* PCBDDCSetUpLocalScatters should be called first! */
1224684f6988SStefano Zampini   ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr);
1225c8587f34SStefano Zampini 
1226c8587f34SStefano Zampini   /* Change global null space passed in by the user if change of basis has been requested */
1227b9b85e73SStefano Zampini   if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) {
1228c8587f34SStefano Zampini     ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr);
1229c8587f34SStefano Zampini   }
1230c8587f34SStefano Zampini 
12318629588bSStefano Zampini   /*
12328629588bSStefano Zampini      Setup local correction and local part of coarse basis.
12338629588bSStefano Zampini      Gives back the dense local part of the coarse matrix in column major ordering
12348629588bSStefano Zampini   */
123547f4ddc3SStefano Zampini   ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr);
12368629588bSStefano Zampini 
12378629588bSStefano Zampini   /* Compute total number of coarse nodes and setup coarse solver */
12388629588bSStefano Zampini   ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr);
12398629588bSStefano Zampini 
12408629588bSStefano Zampini   /* free */
12418629588bSStefano Zampini   ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr);
1242c8587f34SStefano Zampini   PetscFunctionReturn(0);
1243c8587f34SStefano Zampini }
1244c8587f34SStefano Zampini 
1245c8587f34SStefano Zampini #undef __FUNCT__
1246674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization"
1247674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc)
1248674ae819SStefano Zampini {
1249674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
1250674ae819SStefano Zampini   PetscErrorCode ierr;
1251674ae819SStefano Zampini 
1252674ae819SStefano Zampini   PetscFunctionBegin;
1253674ae819SStefano Zampini   ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr);
1254674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr);
125530368db7SStefano Zampini   ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr);
1256674ae819SStefano Zampini   ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr);
1257674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr);
1258785d1243SStefano Zampini   ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr);
1259674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr);
1260f4ddd8eeSStefano Zampini   ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr);
1261f4ddd8eeSStefano Zampini   ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr);
1262785d1243SStefano Zampini   ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr);
126363602bcaSStefano Zampini   ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr);
126463602bcaSStefano Zampini   ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr);
1265674ae819SStefano Zampini   PetscFunctionReturn(0);
1266674ae819SStefano Zampini }
1267674ae819SStefano Zampini 
1268674ae819SStefano Zampini #undef __FUNCT__
1269674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography"
1270674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc)
1271674ae819SStefano Zampini {
1272674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
12734f1b2e48SStefano Zampini   PetscInt       i;
1274674ae819SStefano Zampini   PetscErrorCode ierr;
1275674ae819SStefano Zampini 
1276674ae819SStefano Zampini   PetscFunctionBegin;
1277b9b85e73SStefano Zampini   ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
1278674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
1279674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
1280674ae819SStefano Zampini   ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr);
12814f1b2e48SStefano Zampini   for (i=0;i<pcbddc->n_local_subs;i++) {
12824f1b2e48SStefano Zampini     ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr);
12834f1b2e48SStefano Zampini   }
12844f1b2e48SStefano Zampini   ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr);
1285b96c3477SStefano Zampini   ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr);
1286674ae819SStefano Zampini   PetscFunctionReturn(0);
1287674ae819SStefano Zampini }
1288674ae819SStefano Zampini 
1289674ae819SStefano Zampini #undef __FUNCT__
1290674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers"
1291674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc)
1292674ae819SStefano Zampini {
1293674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
1294674ae819SStefano Zampini   PetscErrorCode ierr;
1295674ae819SStefano Zampini 
1296674ae819SStefano Zampini   PetscFunctionBegin;
1297674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr);
129858da7f69SStefano Zampini   if (pcbddc->coarse_phi_B) {
1299ca92afb2SStefano Zampini     PetscScalar *array;
130006656605SStefano Zampini     ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr);
130106656605SStefano Zampini     ierr = PetscFree(array);CHKERRQ(ierr);
130258da7f69SStefano Zampini   }
1303674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr);
1304674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr);
130515aaf578SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr);
130615aaf578SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr);
1307674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr);
1308674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr);
1309674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr);
131006656605SStefano Zampini   ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr);
1311674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr);
1312674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr);
13138ce42a96SStefano Zampini   ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr);
1314674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr);
1315674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr);
1316674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
1317f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr);
1318f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr);
1319f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr);
1320f4ddd8eeSStefano Zampini   ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
1321727cdba6SStefano Zampini   ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
13220e6343abSStefano Zampini   ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr);
1323f4ddd8eeSStefano Zampini   ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr);
132470cf5478SStefano Zampini   ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr);
13256e683305SStefano Zampini   ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
132681d14e9dSStefano Zampini   ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr);
13270369aaf7SStefano Zampini   ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr);
13288b9f24d4SStefano Zampini   ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr);
13294f1b2e48SStefano Zampini   ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr);
13308b9f24d4SStefano Zampini   ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr);
1331ca92afb2SStefano Zampini   if (pcbddc->benign_zerodiag_subs) {
1332ca92afb2SStefano Zampini     PetscInt i;
1333ca92afb2SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
1334ca92afb2SStefano Zampini       ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr);
1335ca92afb2SStefano Zampini     }
1336ca92afb2SStefano Zampini     ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr);
1337ca92afb2SStefano Zampini   }
13384f1b2e48SStefano Zampini   ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr);
1339674ae819SStefano Zampini   PetscFunctionReturn(0);
1340674ae819SStefano Zampini }
1341674ae819SStefano Zampini 
1342674ae819SStefano Zampini #undef __FUNCT__
1343f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors"
1344f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc)
13456bfb1811SStefano Zampini {
13466bfb1811SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
13476bfb1811SStefano Zampini   PC_IS          *pcis = (PC_IS*)pc->data;
13486bfb1811SStefano Zampini   VecType        impVecType;
13494f1b2e48SStefano Zampini   PetscInt       n_constraints,n_R,old_size;
13506bfb1811SStefano Zampini   PetscErrorCode ierr;
13516bfb1811SStefano Zampini 
13526bfb1811SStefano Zampini   PetscFunctionBegin;
1353f4ddd8eeSStefano Zampini   if (!pcbddc->ConstraintMatrix) {
1354019a44ceSStefano Zampini     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created");
1355f4ddd8eeSStefano Zampini   }
1356e7b262bdSStefano Zampini   /* get sizes */
13574f1b2e48SStefano Zampini   n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices;
1358b371cd4fSStefano Zampini   n_R = pcis->n - pcbddc->n_vertices;
13596bfb1811SStefano Zampini   ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr);
1360e7b262bdSStefano Zampini   /* local work vectors (try to avoid unneeded work)*/
1361e7b262bdSStefano Zampini   /* R nodes */
1362e7b262bdSStefano Zampini   old_size = -1;
1363e7b262bdSStefano Zampini   if (pcbddc->vec1_R) {
1364e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr);
1365e7b262bdSStefano Zampini   }
1366e7b262bdSStefano Zampini   if (n_R != old_size) {
1367e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr);
1368e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr);
13696bfb1811SStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr);
13706bfb1811SStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr);
13716bfb1811SStefano Zampini     ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr);
13726bfb1811SStefano Zampini     ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr);
1373e7b262bdSStefano Zampini   }
1374e7b262bdSStefano Zampini   /* local primal dofs */
1375e7b262bdSStefano Zampini   old_size = -1;
1376e7b262bdSStefano Zampini   if (pcbddc->vec1_P) {
1377e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr);
1378e7b262bdSStefano Zampini   }
1379e9189074SStefano Zampini   if (pcbddc->local_primal_size != old_size) {
1380e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr);
138183b7ccabSStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr);
1382e9189074SStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr);
13836bfb1811SStefano Zampini     ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr);
1384e7b262bdSStefano Zampini   }
1385e7b262bdSStefano Zampini   /* local explicit constraints */
1386e7b262bdSStefano Zampini   old_size = -1;
1387e7b262bdSStefano Zampini   if (pcbddc->vec1_C) {
1388e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr);
1389e7b262bdSStefano Zampini   }
1390e7b262bdSStefano Zampini   if (n_constraints && n_constraints != old_size) {
1391e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr);
139283b7ccabSStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr);
139383b7ccabSStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr);
139483b7ccabSStefano Zampini     ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr);
139583b7ccabSStefano Zampini   }
13966bfb1811SStefano Zampini   PetscFunctionReturn(0);
13976bfb1811SStefano Zampini }
13986bfb1811SStefano Zampini 
13996bfb1811SStefano Zampini #undef __FUNCT__
140047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection"
140147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n)
140288ebb749SStefano Zampini {
140325084f0cSStefano Zampini   PetscErrorCode  ierr;
140425084f0cSStefano Zampini   /* pointers to pcis and pcbddc */
140588ebb749SStefano Zampini   PC_IS*          pcis = (PC_IS*)pc->data;
140688ebb749SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)pc->data;
1407d62866d3SStefano Zampini   PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
140825084f0cSStefano Zampini   /* submatrices of local problem */
140980677318SStefano Zampini   Mat             A_RV,A_VR,A_VV,local_auxmat2_R;
1410019a44ceSStefano Zampini   /* submatrices of benign trick */
1411d16cbb6bSStefano Zampini   Mat             B0_V = NULL;
141206656605SStefano Zampini   /* submatrices of local coarse problem */
141306656605SStefano Zampini   Mat             S_VV,S_CV,S_VC,S_CC;
141425084f0cSStefano Zampini   /* working matrices */
141506656605SStefano Zampini   Mat             C_CR;
141625084f0cSStefano Zampini   /* additional working stuff */
141706656605SStefano Zampini   PC              pc_R;
14184f1b2e48SStefano Zampini   Mat             F;
1419a3df083aSStefano Zampini   PetscBool       isLU,isCHOL,isILU,need_benign_correction;
142006656605SStefano Zampini 
142125084f0cSStefano Zampini   PetscScalar     *coarse_submat_vals; /* TODO: use a PETSc matrix */
142206656605SStefano Zampini   PetscScalar     *work;
142306656605SStefano Zampini   PetscInt        *idx_V_B;
1424ffd830a3SStefano Zampini   PetscInt        lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I;
142506656605SStefano Zampini   PetscInt        i,n_R,n_D,n_B;
1426ffd830a3SStefano Zampini 
142725084f0cSStefano Zampini   /* some shortcuts to scalars */
142806656605SStefano Zampini   PetscScalar     one=1.0,m_one=-1.0;
142988ebb749SStefano Zampini 
143088ebb749SStefano Zampini   PetscFunctionBegin;
1431ffd830a3SStefano Zampini   if (!pcbddc->symmetric_primal && pcbddc->benign_n) {
1432ffd830a3SStefano Zampini     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented");
1433ffd830a3SStefano Zampini   }
1434ffd830a3SStefano Zampini 
1435ffd830a3SStefano Zampini   /* Set Non-overlapping dimensions */
1436b371cd4fSStefano Zampini   n_vertices = pcbddc->n_vertices;
14374f1b2e48SStefano Zampini   n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices;
1438b371cd4fSStefano Zampini   n_B = pcis->n_B;
1439b371cd4fSStefano Zampini   n_D = pcis->n - n_B;
144088ebb749SStefano Zampini   n_R = pcis->n - n_vertices;
144188ebb749SStefano Zampini 
144288ebb749SStefano Zampini   /* vertices in boundary numbering */
1443785e854fSJed Brown   ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr);
14440e6343abSStefano Zampini   ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr);
144588ebb749SStefano Zampini   if (i != n_vertices) {
144622d5777bSStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i);
144788ebb749SStefano Zampini   }
144888ebb749SStefano Zampini 
144906656605SStefano Zampini   /* Subdomain contribution (Non-overlapping) to coarse matrix  */
1450019a44ceSStefano Zampini   ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr);
145106656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr);
145206656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr);
145306656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr);
145406656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr);
145506656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr);
145606656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr);
145706656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr);
145806656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr);
145906656605SStefano Zampini 
146006656605SStefano Zampini   /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */
146106656605SStefano Zampini   ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr);
146206656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr);
146306656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr);
146406656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr);
1465ffd830a3SStefano Zampini   lda_rhs = n_R;
1466a3df083aSStefano Zampini   need_benign_correction = PETSC_FALSE;
146706656605SStefano Zampini   if (isLU || isILU || isCHOL) {
146806656605SStefano Zampini     ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr);
1469df4d28bfSStefano Zampini   } else if (sub_schurs->reuse_solver) {
1470df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
1471d62866d3SStefano Zampini     MatFactorType      type;
1472d62866d3SStefano Zampini 
1473df4d28bfSStefano Zampini     F = reuse_solver->F;
14746816873aSStefano Zampini     ierr = MatGetFactorType(F,&type);CHKERRQ(ierr);
1475d62866d3SStefano Zampini     if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE;
1476ffd830a3SStefano Zampini     ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr);
1477df4d28bfSStefano Zampini     need_benign_correction = !!reuse_solver->benign_n;
147806656605SStefano Zampini   } else {
147906656605SStefano Zampini     F = NULL;
148006656605SStefano Zampini   }
148106656605SStefano Zampini 
1482ffd830a3SStefano Zampini   /* allocate workspace */
1483ffd830a3SStefano Zampini   n = 0;
1484ffd830a3SStefano Zampini   if (n_constraints) {
1485ffd830a3SStefano Zampini     n += lda_rhs*n_constraints;
1486ffd830a3SStefano Zampini   }
1487ffd830a3SStefano Zampini   if (n_vertices) {
1488ffd830a3SStefano Zampini     n = PetscMax(2*lda_rhs*n_vertices,n);
1489ffd830a3SStefano Zampini     n = PetscMax((lda_rhs+n_B)*n_vertices,n);
1490ffd830a3SStefano Zampini   }
1491ffd830a3SStefano Zampini   ierr = PetscMalloc1(n,&work);CHKERRQ(ierr);
1492ffd830a3SStefano Zampini 
149388ebb749SStefano Zampini   /* Precompute stuffs needed for preprocessing and application of BDDC*/
149488ebb749SStefano Zampini   if (n_constraints) {
149506656605SStefano Zampini     Mat         M1,M2,M3;
149680677318SStefano Zampini     Mat         auxmat;
149706656605SStefano Zampini     IS          is_aux;
149880677318SStefano Zampini     PetscScalar *array,*array2;
149906656605SStefano Zampini 
1500f4ddd8eeSStefano Zampini     ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr);
150180677318SStefano Zampini     ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr);
150288ebb749SStefano Zampini 
150325084f0cSStefano Zampini     /* Extract constraints on R nodes: C_{CR}  */
150425084f0cSStefano Zampini     ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr);
15058ce42a96SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr);
150680677318SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr);
150788ebb749SStefano Zampini 
150880677318SStefano Zampini     /* Assemble         local_auxmat2_R =        (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */
150980677318SStefano Zampini     /* Assemble pcbddc->local_auxmat2   = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */
1510ffd830a3SStefano Zampini     ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr);
151188ebb749SStefano Zampini     for (i=0;i<n_constraints;i++) {
151206656605SStefano Zampini       const PetscScalar *row_cmat_values;
151306656605SStefano Zampini       const PetscInt    *row_cmat_indices;
151406656605SStefano Zampini       PetscInt          size_of_constraint,j;
151588ebb749SStefano Zampini 
151606656605SStefano Zampini       ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr);
151706656605SStefano Zampini       for (j=0;j<size_of_constraint;j++) {
1518ffd830a3SStefano Zampini         work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j];
151906656605SStefano Zampini       }
152006656605SStefano Zampini       ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr);
152106656605SStefano Zampini     }
1522ffd830a3SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr);
152306656605SStefano Zampini     if (F) {
152406656605SStefano Zampini       Mat B;
152506656605SStefano Zampini 
1526ffd830a3SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr);
1527a3df083aSStefano Zampini       if (need_benign_correction) {
1528a3df083aSStefano Zampini         PetscScalar        *marr;
1529df4d28bfSStefano Zampini         PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
1530a3df083aSStefano Zampini 
1531a3df083aSStefano Zampini         ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr);
1532a3df083aSStefano Zampini         for (i=0;i<n_constraints;i++) {
1533a3df083aSStefano Zampini           ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr);
1534df4d28bfSStefano Zampini           ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr);
1535a3df083aSStefano Zampini           ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
1536a3df083aSStefano Zampini         }
1537a3df083aSStefano Zampini         ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr);
1538a3df083aSStefano Zampini       }
153980677318SStefano Zampini       ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr);
1540a3df083aSStefano Zampini       if (need_benign_correction) {
1541a3df083aSStefano Zampini         PetscScalar        *marr;
1542df4d28bfSStefano Zampini         PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
1543a3df083aSStefano Zampini 
1544a3df083aSStefano Zampini         ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr);
1545a3df083aSStefano Zampini         for (i=0;i<n_constraints;i++) {
1546a3df083aSStefano Zampini           ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr);
1547df4d28bfSStefano Zampini           ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr);
1548a3df083aSStefano Zampini           ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
1549a3df083aSStefano Zampini         }
1550a3df083aSStefano Zampini         ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr);
1551a3df083aSStefano Zampini       }
155206656605SStefano Zampini       ierr = MatDestroy(&B);CHKERRQ(ierr);
155306656605SStefano Zampini     } else {
155480677318SStefano Zampini       PetscScalar *marr;
155580677318SStefano Zampini 
155680677318SStefano Zampini       ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr);
155706656605SStefano Zampini       for (i=0;i<n_constraints;i++) {
1558ffd830a3SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr);
1559ffd830a3SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr);
156006656605SStefano Zampini         ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
156106656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
156206656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
156306656605SStefano Zampini       }
156480677318SStefano Zampini       ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr);
156506656605SStefano Zampini     }
156680677318SStefano Zampini     if (!pcbddc->switch_static) {
156780677318SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr);
156880677318SStefano Zampini       ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr);
156980677318SStefano Zampini       ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr);
157080677318SStefano Zampini       for (i=0;i<n_constraints;i++) {
1571ffd830a3SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr);
157280677318SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr);
157380677318SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
157480677318SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
157580677318SStefano Zampini         ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
157680677318SStefano Zampini         ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
157780677318SStefano Zampini       }
157880677318SStefano Zampini       ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr);
157980677318SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr);
158080677318SStefano Zampini       ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr);
158180677318SStefano Zampini     } else {
1582ffd830a3SStefano Zampini       if (lda_rhs != n_R) {
1583ffd830a3SStefano Zampini         IS dummy;
1584ffd830a3SStefano Zampini 
1585ffd830a3SStefano Zampini         ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr);
1586ffd830a3SStefano Zampini         ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr);
1587ffd830a3SStefano Zampini         ierr = ISDestroy(&dummy);CHKERRQ(ierr);
1588ffd830a3SStefano Zampini       } else {
158980677318SStefano Zampini         ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr);
159080677318SStefano Zampini         pcbddc->local_auxmat2 = local_auxmat2_R;
1591ffd830a3SStefano Zampini       }
159225084f0cSStefano Zampini       ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr);
159380677318SStefano Zampini     }
159480677318SStefano Zampini     ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
159580677318SStefano Zampini     /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1}  */
159680677318SStefano Zampini     ierr = MatScale(M3,m_one);CHKERRQ(ierr);
159706656605SStefano Zampini     ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr);
159806656605SStefano Zampini     ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr);
159980677318SStefano Zampini     if (isCHOL) {
160080677318SStefano Zampini       ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr);
160180677318SStefano Zampini     } else {
160225084f0cSStefano Zampini       ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr);
160380677318SStefano Zampini     }
160480677318SStefano Zampini     ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr);
160506656605SStefano Zampini     ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr);
160625084f0cSStefano Zampini     ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr);
160725084f0cSStefano Zampini     ierr = MatDestroy(&M2);CHKERRQ(ierr);
160825084f0cSStefano Zampini     ierr = MatDestroy(&M3);CHKERRQ(ierr);
160980677318SStefano Zampini     /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */
161080677318SStefano Zampini     ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr);
161180677318SStefano Zampini     ierr = MatDestroy(&auxmat);CHKERRQ(ierr);
161206656605SStefano Zampini     ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */
161306656605SStefano Zampini     ierr = MatDestroy(&M1);CHKERRQ(ierr);
1614f4ddd8eeSStefano Zampini   }
161588ebb749SStefano Zampini   /* Get submatrices from subdomain matrix */
16164f1b2e48SStefano Zampini   if (pcbddc->benign_n) {
1617d16cbb6bSStefano Zampini     IS        dummy;
1618d16cbb6bSStefano Zampini     Mat       B0_R;
1619d16cbb6bSStefano Zampini     PetscReal norm;
1620d16cbb6bSStefano Zampini 
16214f1b2e48SStefano Zampini     ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr);
16224f1b2e48SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr);
1623d16cbb6bSStefano Zampini     ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr);
1624d16cbb6bSStefano Zampini     if (norm > PETSC_SMALL) {
1625d16cbb6bSStefano Zampini       SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm);
1626d16cbb6bSStefano Zampini     }
1627d16cbb6bSStefano Zampini     ierr = MatDestroy(&B0_R);CHKERRQ(ierr);
1628d16cbb6bSStefano Zampini     ierr = ISDestroy(&dummy);CHKERRQ(ierr);
1629d16cbb6bSStefano Zampini   }
1630d16cbb6bSStefano Zampini 
163188ebb749SStefano Zampini   if (n_vertices) {
163206656605SStefano Zampini     IS is_aux;
16333a50541eSStefano Zampini 
1634df4d28bfSStefano Zampini     if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */
16356816873aSStefano Zampini       IS tis;
16366816873aSStefano Zampini 
16376816873aSStefano Zampini       ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr);
16386816873aSStefano Zampini       ierr = ISSort(tis);CHKERRQ(ierr);
16396816873aSStefano Zampini       ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr);
16406816873aSStefano Zampini       ierr = ISDestroy(&tis);CHKERRQ(ierr);
16416816873aSStefano Zampini     } else {
16423a50541eSStefano Zampini       ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr);
16436816873aSStefano Zampini     }
16449577ea80SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr);
16459577ea80SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr);
164604708bb6SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr);
16474f1b2e48SStefano Zampini     if (pcbddc->benign_n) {
1648019a44ceSStefano Zampini       IS dummy;
1649019a44ceSStefano Zampini 
16504f1b2e48SStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr);
16514f1b2e48SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr);
1652019a44ceSStefano Zampini       ierr = ISDestroy(&dummy);CHKERRQ(ierr);
1653019a44ceSStefano Zampini     }
165425084f0cSStefano Zampini     ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
165588ebb749SStefano Zampini   }
165688ebb749SStefano Zampini 
165788ebb749SStefano Zampini   /* Matrix of coarse basis functions (local) */
1658f4ddd8eeSStefano Zampini   if (pcbddc->coarse_phi_B) {
165906656605SStefano Zampini     PetscInt on_B,on_primal,on_D=n_D;
166006656605SStefano Zampini     if (pcbddc->coarse_phi_D) {
166106656605SStefano Zampini       ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr);
166206656605SStefano Zampini     }
1663f4ddd8eeSStefano Zampini     ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr);
166406656605SStefano Zampini     if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) {
166506656605SStefano Zampini       PetscScalar *marray;
166606656605SStefano Zampini 
166706656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr);
166806656605SStefano Zampini       ierr = PetscFree(marray);CHKERRQ(ierr);
1669f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr);
1670f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr);
1671f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr);
1672f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr);
1673f4ddd8eeSStefano Zampini     }
1674f4ddd8eeSStefano Zampini   }
167506656605SStefano Zampini 
1676f4ddd8eeSStefano Zampini   if (!pcbddc->coarse_phi_B) {
167706656605SStefano Zampini     PetscScalar *marray;
167888ebb749SStefano Zampini 
167906656605SStefano Zampini     n = n_B*pcbddc->local_primal_size;
16808eeda7d8SStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
168106656605SStefano Zampini       n += n_D*pcbddc->local_primal_size;
168288ebb749SStefano Zampini     }
16833301b35fSStefano Zampini     if (!pcbddc->symmetric_primal) {
168406656605SStefano Zampini       n *= 2;
168588ebb749SStefano Zampini     }
168606656605SStefano Zampini     ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr);
168706656605SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr);
168806656605SStefano Zampini     n = n_B*pcbddc->local_primal_size;
16898eeda7d8SStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
169006656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr);
169106656605SStefano Zampini       n += n_D*pcbddc->local_primal_size;
169288ebb749SStefano Zampini     }
16933301b35fSStefano Zampini     if (!pcbddc->symmetric_primal) {
169406656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr);
16958eeda7d8SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
169606656605SStefano Zampini         n = n_B*pcbddc->local_primal_size;
169706656605SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr);
169888ebb749SStefano Zampini       }
169988ebb749SStefano Zampini     } else {
1700c0553b1fSStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr);
1701c0553b1fSStefano Zampini       pcbddc->coarse_psi_B = pcbddc->coarse_phi_B;
17021b968477SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
1703c0553b1fSStefano Zampini         ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr);
1704c0553b1fSStefano Zampini         pcbddc->coarse_psi_D = pcbddc->coarse_phi_D;
1705c0553b1fSStefano Zampini       }
170688ebb749SStefano Zampini     }
170706656605SStefano Zampini   }
1708019a44ceSStefano Zampini 
170906656605SStefano Zampini   /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */
17104f1b2e48SStefano Zampini   p0_lidx_I = NULL;
17114f1b2e48SStefano Zampini   if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) {
1712d12edf2fSStefano Zampini     const PetscInt *idxs;
1713d12edf2fSStefano Zampini 
1714d12edf2fSStefano Zampini     ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr);
17154f1b2e48SStefano Zampini     ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr);
17164f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
17174f1b2e48SStefano Zampini       ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr);
17184f1b2e48SStefano Zampini     }
1719d12edf2fSStefano Zampini     ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr);
1720d12edf2fSStefano Zampini   }
1721d16cbb6bSStefano Zampini 
172206656605SStefano Zampini   /* vertices */
172306656605SStefano Zampini   if (n_vertices) {
172416f15bc4SStefano Zampini 
1725ffd830a3SStefano Zampini     ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr);
172604708bb6SStefano Zampini 
172716f15bc4SStefano Zampini     if (n_R) {
172814393ed6SStefano Zampini       Mat          A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */
172906656605SStefano Zampini       PetscBLASInt B_N,B_one = 1;
173016f15bc4SStefano Zampini       PetscScalar  *x,*y;
173104708bb6SStefano Zampini       PetscBool    isseqaij;
173206656605SStefano Zampini 
173321eccb56SStefano Zampini       ierr = MatScale(A_RV,m_one);CHKERRQ(ierr);
173414393ed6SStefano Zampini       if (need_benign_correction) {
173514393ed6SStefano Zampini         ISLocalToGlobalMapping RtoN;
173614393ed6SStefano Zampini         IS                     is_p0;
173714393ed6SStefano Zampini         PetscInt               *idxs_p0,n;
173814393ed6SStefano Zampini 
173914393ed6SStefano Zampini         ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr);
174014393ed6SStefano Zampini         ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr);
174114393ed6SStefano Zampini         ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr);
174214393ed6SStefano Zampini         if (n != pcbddc->benign_n) {
174314393ed6SStefano Zampini           SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n);
174414393ed6SStefano Zampini         }
174514393ed6SStefano Zampini         ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr);
174614393ed6SStefano Zampini         ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr);
174714393ed6SStefano Zampini         ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr);
174814393ed6SStefano Zampini         ierr = ISDestroy(&is_p0);CHKERRQ(ierr);
174914393ed6SStefano Zampini       }
175014393ed6SStefano Zampini 
1751ffd830a3SStefano Zampini       if (lda_rhs == n_R) {
1752ffd830a3SStefano Zampini         ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr);
1753ffd830a3SStefano Zampini       } else {
1754ca92afb2SStefano Zampini         PetscScalar    *av,*array;
1755ca92afb2SStefano Zampini         const PetscInt *xadj,*adjncy;
1756ca92afb2SStefano Zampini         PetscInt       n;
1757ca92afb2SStefano Zampini         PetscBool      flg_row;
1758ffd830a3SStefano Zampini 
1759ca92afb2SStefano Zampini         array = work+lda_rhs*n_vertices;
1760ca92afb2SStefano Zampini         ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr);
1761ca92afb2SStefano Zampini         ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr);
1762ca92afb2SStefano Zampini         ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
1763ca92afb2SStefano Zampini         ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr);
1764ca92afb2SStefano Zampini         for (i=0;i<n;i++) {
1765ca92afb2SStefano Zampini           PetscInt j;
1766ca92afb2SStefano Zampini           for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j];
1767ffd830a3SStefano Zampini         }
1768ca92afb2SStefano Zampini         ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
1769ca92afb2SStefano Zampini         ierr = MatDestroy(&A_RV);CHKERRQ(ierr);
1770ca92afb2SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr);
1771ffd830a3SStefano Zampini       }
1772ffd830a3SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr);
177314393ed6SStefano Zampini       if (F) {
1774a3df083aSStefano Zampini         if (need_benign_correction) {
1775df4d28bfSStefano Zampini           PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
1776a3df083aSStefano Zampini           PetscScalar        *marr;
1777a3df083aSStefano Zampini 
1778a3df083aSStefano Zampini           ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr);
177914393ed6SStefano Zampini           /* need \Phi^T A_RV = (I+L)A_RV, L given by
178014393ed6SStefano Zampini 
178114393ed6SStefano Zampini                  | 0 0  0 | (V)
178214393ed6SStefano Zampini              L = | 0 0 -1 | (P-p0)
178314393ed6SStefano Zampini                  | 0 0 -1 | (p0)
178414393ed6SStefano Zampini 
178514393ed6SStefano Zampini           */
1786df4d28bfSStefano Zampini           for (i=0;i<reuse_solver->benign_n;i++) {
178714393ed6SStefano Zampini             const PetscScalar *vals;
178814393ed6SStefano Zampini             const PetscInt    *idxs,*idxs_zero;
178914393ed6SStefano Zampini             PetscInt          n,j,nz;
179014393ed6SStefano Zampini 
1791df4d28bfSStefano Zampini             ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr);
1792df4d28bfSStefano Zampini             ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr);
179314393ed6SStefano Zampini             ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr);
179414393ed6SStefano Zampini             for (j=0;j<n;j++) {
179514393ed6SStefano Zampini               PetscScalar val = vals[j];
179614393ed6SStefano Zampini               PetscInt    k,col = idxs[j];
179714393ed6SStefano Zampini               for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val;
179814393ed6SStefano Zampini             }
179914393ed6SStefano Zampini             ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr);
1800df4d28bfSStefano Zampini             ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr);
180114393ed6SStefano Zampini           }
180214393ed6SStefano Zampini           /* need to correct the rhs */
1803a3df083aSStefano Zampini           for (i=0;i<n_vertices;i++) {
1804a3df083aSStefano Zampini             ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr);
1805df4d28bfSStefano Zampini             ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr);
1806a3df083aSStefano Zampini             ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
1807a3df083aSStefano Zampini           }
1808a3df083aSStefano Zampini           ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr);
1809a3df083aSStefano Zampini         }
181006656605SStefano Zampini         ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr);
181114393ed6SStefano Zampini         /* need to correct the solution */
1812a3df083aSStefano Zampini         if (need_benign_correction) {
1813df4d28bfSStefano Zampini           PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
1814a3df083aSStefano Zampini           PetscScalar        *marr;
1815a3df083aSStefano Zampini 
1816a3df083aSStefano Zampini           ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr);
1817a3df083aSStefano Zampini           for (i=0;i<n_vertices;i++) {
1818a3df083aSStefano Zampini             ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr);
1819df4d28bfSStefano Zampini             ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr);
1820a3df083aSStefano Zampini             ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
1821a3df083aSStefano Zampini           }
1822a3df083aSStefano Zampini           ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr);
1823a3df083aSStefano Zampini         }
182406656605SStefano Zampini       } else {
182506656605SStefano Zampini         ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr);
182606656605SStefano Zampini         for (i=0;i<n_vertices;i++) {
1827ffd830a3SStefano Zampini           ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr);
1828ffd830a3SStefano Zampini           ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr);
182906656605SStefano Zampini           ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
183006656605SStefano Zampini           ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
183106656605SStefano Zampini           ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
183206656605SStefano Zampini         }
183306656605SStefano Zampini         ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr);
183406656605SStefano Zampini       }
183580677318SStefano Zampini       ierr = MatDestroy(&A_RV);CHKERRQ(ierr);
1836ffd830a3SStefano Zampini       /* S_VV and S_CV */
183706656605SStefano Zampini       if (n_constraints) {
183806656605SStefano Zampini         Mat B;
183980677318SStefano Zampini 
1840ffd830a3SStefano Zampini         ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr);
184180677318SStefano Zampini         for (i=0;i<n_vertices;i++) {
1842ffd830a3SStefano Zampini           ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr);
1843ffd830a3SStefano Zampini           ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr);
184480677318SStefano Zampini           ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
184580677318SStefano Zampini           ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
184680677318SStefano Zampini           ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
184780677318SStefano Zampini           ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
184880677318SStefano Zampini         }
1849ffd830a3SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr);
185080677318SStefano Zampini         ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr);
185180677318SStefano Zampini         ierr = MatDestroy(&B);CHKERRQ(ierr);
1852ffd830a3SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr);
185380677318SStefano Zampini         ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr);
185406656605SStefano Zampini         ierr = MatScale(S_CV,m_one);CHKERRQ(ierr);
1855ffd830a3SStefano Zampini         ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr);
1856ffd830a3SStefano Zampini         PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one));
185706656605SStefano Zampini         ierr = MatDestroy(&B);CHKERRQ(ierr);
185806656605SStefano Zampini       }
185904708bb6SStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
186004708bb6SStefano Zampini       if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */
186104708bb6SStefano Zampini         ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr);
186204708bb6SStefano Zampini       }
1863ffd830a3SStefano Zampini       if (lda_rhs != n_R) {
1864ffd830a3SStefano Zampini         ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr);
1865ffd830a3SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr);
1866ffd830a3SStefano Zampini         ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr);
1867ffd830a3SStefano Zampini       }
186806656605SStefano Zampini       ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr);
186914393ed6SStefano Zampini       /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */
187014393ed6SStefano Zampini       if (need_benign_correction) {
1871df4d28bfSStefano Zampini         PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
187214393ed6SStefano Zampini         PetscScalar      *marr,*sums;
187314393ed6SStefano Zampini 
187414393ed6SStefano Zampini         ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr);
187514393ed6SStefano Zampini         ierr = MatDenseGetArray(S_VVt,&marr);
1876df4d28bfSStefano Zampini         for (i=0;i<reuse_solver->benign_n;i++) {
187714393ed6SStefano Zampini           const PetscScalar *vals;
187814393ed6SStefano Zampini           const PetscInt    *idxs,*idxs_zero;
187914393ed6SStefano Zampini           PetscInt          n,j,nz;
188014393ed6SStefano Zampini 
1881df4d28bfSStefano Zampini           ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr);
1882df4d28bfSStefano Zampini           ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr);
188314393ed6SStefano Zampini           for (j=0;j<n_vertices;j++) {
188414393ed6SStefano Zampini             PetscInt k;
188514393ed6SStefano Zampini             sums[j] = 0.;
188614393ed6SStefano Zampini             for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs];
188714393ed6SStefano Zampini           }
188814393ed6SStefano Zampini           ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr);
188914393ed6SStefano Zampini           for (j=0;j<n;j++) {
189014393ed6SStefano Zampini             PetscScalar val = vals[j];
189114393ed6SStefano Zampini             PetscInt k;
189214393ed6SStefano Zampini             for (k=0;k<n_vertices;k++) {
189314393ed6SStefano Zampini               marr[idxs[j]+k*n_vertices] += val*sums[k];
189414393ed6SStefano Zampini             }
189514393ed6SStefano Zampini           }
189614393ed6SStefano Zampini           ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr);
1897df4d28bfSStefano Zampini           ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr);
189814393ed6SStefano Zampini         }
189914393ed6SStefano Zampini         ierr = PetscFree(sums);CHKERRQ(ierr);
190014393ed6SStefano Zampini         ierr = MatDenseRestoreArray(S_VVt,&marr);
190114393ed6SStefano Zampini         ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr);
190214393ed6SStefano Zampini       }
190380677318SStefano Zampini       ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr);
190406656605SStefano Zampini       ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr);
190506656605SStefano Zampini       ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr);
190606656605SStefano Zampini       ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr);
190706656605SStefano Zampini       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one));
190806656605SStefano Zampini       ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr);
190906656605SStefano Zampini       ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr);
191006656605SStefano Zampini       ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
1911d16cbb6bSStefano Zampini       ierr = MatDestroy(&S_VVt);CHKERRQ(ierr);
1912019a44ceSStefano Zampini     } else {
1913d16cbb6bSStefano Zampini       ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
1914d16cbb6bSStefano Zampini     }
19154f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
1916019a44ceSStefano Zampini       const PetscScalar *vals;
1917019a44ceSStefano Zampini       const PetscInt    *idxs;
19184f1b2e48SStefano Zampini       PetscInt          n,j,primal_idx;
1919019a44ceSStefano Zampini 
19204f1b2e48SStefano Zampini       ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr);
19214f1b2e48SStefano Zampini       primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i;
1922d16cbb6bSStefano Zampini       for (j=0;j<n;j++) {
19234f1b2e48SStefano Zampini         coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j];
19244f1b2e48SStefano Zampini         coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j];
1925019a44ceSStefano Zampini       }
19264f1b2e48SStefano Zampini       ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr);
192716f15bc4SStefano Zampini     }
192821eccb56SStefano Zampini     ierr = MatDestroy(&A_VV);CHKERRQ(ierr);
1929d16cbb6bSStefano Zampini 
193006656605SStefano Zampini     /* coarse basis functions */
193106656605SStefano Zampini     for (i=0;i<n_vertices;i++) {
193216f15bc4SStefano Zampini       PetscScalar *y;
193316f15bc4SStefano Zampini 
1934ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr);
193506656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
193606656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr);
193706656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
193806656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
193906656605SStefano Zampini       y[n_B*i+idx_V_B[i]] = 1.0;
194006656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
194106656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
194206656605SStefano Zampini 
194306656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
19444f1b2e48SStefano Zampini         PetscInt j;
19454f1b2e48SStefano Zampini 
194606656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
194706656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr);
194806656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
194906656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
195006656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
19514f1b2e48SStefano Zampini         for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0;
195206656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
195306656605SStefano Zampini       }
195406656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
195506656605SStefano Zampini     }
195604708bb6SStefano Zampini     /* if n_R == 0 the object is not destroyed */
195704708bb6SStefano Zampini     ierr = MatDestroy(&A_RV);CHKERRQ(ierr);
195806656605SStefano Zampini   }
195906656605SStefano Zampini 
196006656605SStefano Zampini   if (n_constraints) {
196106656605SStefano Zampini     Mat B;
196206656605SStefano Zampini 
1963ffd830a3SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr);
196406656605SStefano Zampini     ierr = MatScale(S_CC,m_one);CHKERRQ(ierr);
196580677318SStefano Zampini     ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr);
196606656605SStefano Zampini     ierr = MatScale(S_CC,m_one);CHKERRQ(ierr);
196706656605SStefano Zampini     if (n_vertices) {
196880677318SStefano Zampini       if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */
196980677318SStefano Zampini         ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr);
197080677318SStefano Zampini       } else {
197180677318SStefano Zampini         Mat S_VCt;
197280677318SStefano Zampini 
1973ffd830a3SStefano Zampini         if (lda_rhs != n_R) {
1974ffd830a3SStefano Zampini           ierr = MatDestroy(&B);CHKERRQ(ierr);
1975ffd830a3SStefano Zampini           ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr);
1976ffd830a3SStefano Zampini           ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr);
1977ffd830a3SStefano Zampini         }
197880677318SStefano Zampini         ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr);
197980677318SStefano Zampini         ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
198080677318SStefano Zampini         ierr = MatDestroy(&S_VCt);CHKERRQ(ierr);
198180677318SStefano Zampini       }
198206656605SStefano Zampini     }
198306656605SStefano Zampini     ierr = MatDestroy(&B);CHKERRQ(ierr);
198406656605SStefano Zampini     /* coarse basis functions */
198506656605SStefano Zampini     for (i=0;i<n_constraints;i++) {
198606656605SStefano Zampini       PetscScalar *y;
198706656605SStefano Zampini 
1988ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr);
198906656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
199006656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr);
199106656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
199206656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
199306656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
199406656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
199506656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
19964f1b2e48SStefano Zampini         PetscInt j;
19974f1b2e48SStefano Zampini 
199806656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
199906656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr);
200006656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
200106656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
200206656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
20034f1b2e48SStefano Zampini         for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0;
200406656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
200506656605SStefano Zampini       }
200606656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
200706656605SStefano Zampini     }
200806656605SStefano Zampini   }
200980677318SStefano Zampini   if (n_constraints) {
201080677318SStefano Zampini     ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr);
201180677318SStefano Zampini   }
20124f1b2e48SStefano Zampini   ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr);
2013019a44ceSStefano Zampini   ierr = MatDestroy(&B0_V);CHKERRQ(ierr);
2014019a44ceSStefano Zampini 
201506656605SStefano Zampini   /* compute other basis functions for non-symmetric problems */
20163301b35fSStefano Zampini   if (!pcbddc->symmetric_primal) {
2017ffd830a3SStefano Zampini     Mat         B_V=NULL,B_C=NULL;
2018ffd830a3SStefano Zampini     PetscScalar *marray;
201906656605SStefano Zampini 
202006656605SStefano Zampini     if (n_constraints) {
2021ffd830a3SStefano Zampini       Mat S_CCT,C_CRT;
202206656605SStefano Zampini 
2023ffd830a3SStefano Zampini       ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr);
202406656605SStefano Zampini       ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr);
2025ffd830a3SStefano Zampini       ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr);
202616f15bc4SStefano Zampini       ierr = MatDestroy(&S_CCT);CHKERRQ(ierr);
202706656605SStefano Zampini       if (n_vertices) {
2028ffd830a3SStefano Zampini         Mat S_VCT;
202906656605SStefano Zampini 
203006656605SStefano Zampini         ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr);
2031ffd830a3SStefano Zampini         ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr);
203216f15bc4SStefano Zampini         ierr = MatDestroy(&S_VCT);CHKERRQ(ierr);
203306656605SStefano Zampini       }
2034ffd830a3SStefano Zampini       ierr = MatDestroy(&C_CRT);CHKERRQ(ierr);
203506656605SStefano Zampini     }
203616f15bc4SStefano Zampini     if (n_vertices && n_R) {
2037ffd830a3SStefano Zampini       PetscScalar    *av,*marray;
2038ffd830a3SStefano Zampini       const PetscInt *xadj,*adjncy;
2039ffd830a3SStefano Zampini       PetscInt       n;
2040ffd830a3SStefano Zampini       PetscBool      flg_row;
204106656605SStefano Zampini 
2042ffd830a3SStefano Zampini       /* B_V = B_V - A_VR^T */
2043ffd830a3SStefano Zampini       ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr);
2044ffd830a3SStefano Zampini       ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
2045ffd830a3SStefano Zampini       ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr);
2046ffd830a3SStefano Zampini       ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr);
2047ffd830a3SStefano Zampini       for (i=0;i<n;i++) {
2048ffd830a3SStefano Zampini         PetscInt j;
2049ffd830a3SStefano Zampini         for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j];
2050ffd830a3SStefano Zampini       }
2051ffd830a3SStefano Zampini       ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr);
2052ffd830a3SStefano Zampini       ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
2053ffd830a3SStefano Zampini       ierr = MatDestroy(&A_VR);CHKERRQ(ierr);
205406656605SStefano Zampini     }
205506656605SStefano Zampini 
2056ffd830a3SStefano Zampini     /* currently there's no support for MatTransposeMatSolve(F,B,X) */
2057ffd830a3SStefano Zampini     ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr);
2058ffd830a3SStefano Zampini     for (i=0;i<n_vertices;i++) {
2059ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr);
2060ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr);
206106656605SStefano Zampini       ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
206206656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
206306656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
206406656605SStefano Zampini     }
2065ffd830a3SStefano Zampini     ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr);
2066ffd830a3SStefano Zampini     ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr);
2067ffd830a3SStefano Zampini     for (i=n_vertices;i<n_constraints+n_vertices;i++) {
2068ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr);
2069ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr);
2070ffd830a3SStefano Zampini       ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
2071ffd830a3SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
2072ffd830a3SStefano Zampini       ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
207306656605SStefano Zampini     }
2074ffd830a3SStefano Zampini     ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr);
207506656605SStefano Zampini     /* coarse basis functions */
207606656605SStefano Zampini     for (i=0;i<pcbddc->local_primal_size;i++) {
207706656605SStefano Zampini       PetscScalar *y;
207806656605SStefano Zampini 
2079ffd830a3SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr);
208006656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr);
208106656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr);
208206656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
208306656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
208406656605SStefano Zampini       if (i<n_vertices) {
208506656605SStefano Zampini         y[n_B*i+idx_V_B[i]] = 1.0;
208606656605SStefano Zampini       }
208706656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr);
208806656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
208906656605SStefano Zampini 
209006656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
209106656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr);
209206656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr);
209306656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
209406656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
209506656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
209606656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr);
209706656605SStefano Zampini       }
209806656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
209906656605SStefano Zampini     }
2100ffd830a3SStefano Zampini     ierr = MatDestroy(&B_V);CHKERRQ(ierr);
2101ffd830a3SStefano Zampini     ierr = MatDestroy(&B_C);CHKERRQ(ierr);
210206656605SStefano Zampini   }
2103d62866d3SStefano Zampini   /* free memory */
210488ebb749SStefano Zampini   ierr = PetscFree(idx_V_B);CHKERRQ(ierr);
210506656605SStefano Zampini   ierr = MatDestroy(&S_VV);CHKERRQ(ierr);
210606656605SStefano Zampini   ierr = MatDestroy(&S_CV);CHKERRQ(ierr);
210706656605SStefano Zampini   ierr = MatDestroy(&S_VC);CHKERRQ(ierr);
210806656605SStefano Zampini   ierr = MatDestroy(&S_CC);CHKERRQ(ierr);
2109d62866d3SStefano Zampini   ierr = PetscFree(work);CHKERRQ(ierr);
2110d62866d3SStefano Zampini   if (n_vertices) {
2111d62866d3SStefano Zampini     ierr = MatDestroy(&A_VR);CHKERRQ(ierr);
2112d62866d3SStefano Zampini   }
2113d62866d3SStefano Zampini   if (n_constraints) {
2114d62866d3SStefano Zampini     ierr = MatDestroy(&C_CR);CHKERRQ(ierr);
2115d62866d3SStefano Zampini   }
211688ebb749SStefano Zampini   /* Checking coarse_sub_mat and coarse basis functios */
211788ebb749SStefano Zampini   /* Symmetric case     : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
211888ebb749SStefano Zampini   /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
2119d12edf2fSStefano Zampini   if (pcbddc->dbg_flag) {
212088ebb749SStefano Zampini     Mat         coarse_sub_mat;
212125084f0cSStefano Zampini     Mat         AUXMAT,TM1,TM2,TM3,TM4;
212288ebb749SStefano Zampini     Mat         coarse_phi_D,coarse_phi_B;
212388ebb749SStefano Zampini     Mat         coarse_psi_D,coarse_psi_B;
212488ebb749SStefano Zampini     Mat         A_II,A_BB,A_IB,A_BI;
21258bec7fa6SStefano Zampini     Mat         C_B,CPHI;
21268bec7fa6SStefano Zampini     IS          is_dummy;
21278bec7fa6SStefano Zampini     Vec         mones;
212888ebb749SStefano Zampini     MatType     checkmattype=MATSEQAIJ;
212988ebb749SStefano Zampini     PetscReal   real_value;
213088ebb749SStefano Zampini 
2131a3df083aSStefano Zampini     if (pcbddc->benign_n && !pcbddc->benign_change_explicit) {
2132a3df083aSStefano Zampini       Mat A;
2133a3df083aSStefano Zampini       ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr);
2134a3df083aSStefano Zampini       ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr);
2135a3df083aSStefano Zampini       ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr);
2136a3df083aSStefano Zampini       ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr);
2137a3df083aSStefano Zampini       ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr);
2138a3df083aSStefano Zampini       ierr = MatDestroy(&A);CHKERRQ(ierr);
2139a3df083aSStefano Zampini     } else {
214088ebb749SStefano Zampini       ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr);
214188ebb749SStefano Zampini       ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr);
214288ebb749SStefano Zampini       ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr);
214388ebb749SStefano Zampini       ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr);
2144a3df083aSStefano Zampini     }
214588ebb749SStefano Zampini     ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr);
214688ebb749SStefano Zampini     ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr);
2147ffd830a3SStefano Zampini     if (!pcbddc->symmetric_primal) {
214888ebb749SStefano Zampini       ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr);
214988ebb749SStefano Zampini       ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr);
215088ebb749SStefano Zampini     }
215188ebb749SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr);
215288ebb749SStefano Zampini 
215325084f0cSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
21543301b35fSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr);
215525084f0cSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2156ffd830a3SStefano Zampini     if (!pcbddc->symmetric_primal) {
215788ebb749SStefano Zampini       ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
215888ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr);
215988ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
216088ebb749SStefano Zampini       ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
216188ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr);
216288ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
216388ebb749SStefano Zampini       ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
216488ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr);
216588ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
216688ebb749SStefano Zampini       ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
216788ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr);
216888ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
216988ebb749SStefano Zampini     } else {
217088ebb749SStefano Zampini       ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr);
217188ebb749SStefano Zampini       ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr);
217288ebb749SStefano Zampini       ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
217388ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr);
217488ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
217588ebb749SStefano Zampini       ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
217688ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr);
217788ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
217888ebb749SStefano Zampini     }
217988ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
218088ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
218188ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
218288ebb749SStefano Zampini     ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr);
21834f1b2e48SStefano Zampini     if (pcbddc->benign_n) {
2184d12edf2fSStefano Zampini       Mat         B0_I,B0_B,B0_BPHI,B0_IPHI;
2185d12edf2fSStefano Zampini       PetscScalar *data,*data2;
21864f1b2e48SStefano Zampini       PetscInt    j;
2187d12edf2fSStefano Zampini 
21884f1b2e48SStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr);
21894f1b2e48SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);
21904f1b2e48SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I);
2191d12edf2fSStefano Zampini       ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr);
2192d12edf2fSStefano Zampini       ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr);
2193d12edf2fSStefano Zampini       ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr);
2194d12edf2fSStefano Zampini       ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr);
21954f1b2e48SStefano Zampini       for (j=0;j<pcbddc->benign_n;j++) {
21964f1b2e48SStefano Zampini         PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j;
2197d12edf2fSStefano Zampini         for (i=0;i<pcbddc->local_primal_size;i++) {
21984f1b2e48SStefano Zampini           data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j];
21994f1b2e48SStefano Zampini           data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j];
22004f1b2e48SStefano Zampini         }
2201d12edf2fSStefano Zampini       }
2202d12edf2fSStefano Zampini       ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr);
2203d12edf2fSStefano Zampini       ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr);
2204d12edf2fSStefano Zampini       ierr = MatDestroy(&B0_B);CHKERRQ(ierr);
2205d12edf2fSStefano Zampini       ierr = ISDestroy(&is_dummy);CHKERRQ(ierr);
2206d12edf2fSStefano Zampini       ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr);
2207d12edf2fSStefano Zampini       ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr);
2208d12edf2fSStefano Zampini       ierr = MatDestroy(&B0_I);CHKERRQ(ierr);
2209d12edf2fSStefano Zampini       ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
2210d12edf2fSStefano Zampini       ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr);
2211d12edf2fSStefano Zampini     }
2212d12edf2fSStefano Zampini #if 0
2213d12edf2fSStefano Zampini   {
2214d12edf2fSStefano Zampini     PetscViewer viewer;
2215d12edf2fSStefano Zampini     char filename[256];
2216ffd830a3SStefano Zampini     sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level);
2217d12edf2fSStefano Zampini     ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr);
2218d12edf2fSStefano Zampini     ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
2219ffd830a3SStefano Zampini     ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr);
2220ffd830a3SStefano Zampini     ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr);
2221ffd830a3SStefano Zampini     ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr);
2222d12edf2fSStefano Zampini     ierr = MatView(TM1,viewer);CHKERRQ(ierr);
2223ffd830a3SStefano Zampini     ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr);
2224ffd830a3SStefano Zampini     ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr);
2225ffd830a3SStefano Zampini     if (pcbddc->coarse_phi_D) {
2226ffd830a3SStefano Zampini       ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr);
2227ffd830a3SStefano Zampini       ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr);
2228ffd830a3SStefano Zampini     }
2229ffd830a3SStefano Zampini     if (pcbddc->coarse_psi_B) {
2230ffd830a3SStefano Zampini       ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr);
2231ffd830a3SStefano Zampini       ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr);
2232ffd830a3SStefano Zampini     }
2233ffd830a3SStefano Zampini     if (pcbddc->coarse_phi_B) {
2234ffd830a3SStefano Zampini       ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr);
2235ffd830a3SStefano Zampini       ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr);
2236ffd830a3SStefano Zampini     }
2237d12edf2fSStefano Zampini     ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr);
2238d12edf2fSStefano Zampini   }
2239d12edf2fSStefano Zampini #endif
224081d9aea3SBarry Smith     ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
22418bec7fa6SStefano Zampini     ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
22421575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
224306656605SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d          matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr);
22448bec7fa6SStefano Zampini 
22458bec7fa6SStefano Zampini     /* check constraints */
22464f1b2e48SStefano Zampini     if (!pcbddc->benign_n) { /* TODO: add benign case */
22478bec7fa6SStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr);
22488bec7fa6SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);
22498bec7fa6SStefano Zampini       ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr);
22508bec7fa6SStefano Zampini       ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr);
22518bec7fa6SStefano Zampini       ierr = VecSet(mones,-1.0);CHKERRQ(ierr);
22528bec7fa6SStefano Zampini       ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr);
22538bec7fa6SStefano Zampini       ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
2254bdae7319SStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr);
2255ffd830a3SStefano Zampini       if (!pcbddc->symmetric_primal) {
2256bdae7319SStefano Zampini         ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr);
2257bdae7319SStefano Zampini         ierr = VecSet(mones,-1.0);CHKERRQ(ierr);
2258bdae7319SStefano Zampini         ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr);
2259bdae7319SStefano Zampini         ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
2260bdae7319SStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr);
226188ebb749SStefano Zampini       }
22628bec7fa6SStefano Zampini       ierr = MatDestroy(&C_B);CHKERRQ(ierr);
22638bec7fa6SStefano Zampini       ierr = MatDestroy(&CPHI);CHKERRQ(ierr);
22648bec7fa6SStefano Zampini       ierr = ISDestroy(&is_dummy);CHKERRQ(ierr);
22658bec7fa6SStefano Zampini       ierr = VecDestroy(&mones);CHKERRQ(ierr);
2266d12edf2fSStefano Zampini     }
226725084f0cSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
226888ebb749SStefano Zampini     ierr = MatDestroy(&A_II);CHKERRQ(ierr);
226988ebb749SStefano Zampini     ierr = MatDestroy(&A_BB);CHKERRQ(ierr);
227088ebb749SStefano Zampini     ierr = MatDestroy(&A_IB);CHKERRQ(ierr);
227188ebb749SStefano Zampini     ierr = MatDestroy(&A_BI);CHKERRQ(ierr);
227288ebb749SStefano Zampini     ierr = MatDestroy(&TM1);CHKERRQ(ierr);
227388ebb749SStefano Zampini     ierr = MatDestroy(&TM2);CHKERRQ(ierr);
227488ebb749SStefano Zampini     ierr = MatDestroy(&TM3);CHKERRQ(ierr);
227588ebb749SStefano Zampini     ierr = MatDestroy(&TM4);CHKERRQ(ierr);
227688ebb749SStefano Zampini     ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr);
227788ebb749SStefano Zampini     ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr);
2278ffd830a3SStefano Zampini     if (!pcbddc->symmetric_primal) {
227988ebb749SStefano Zampini       ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr);
228088ebb749SStefano Zampini       ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr);
228188ebb749SStefano Zampini     }
228288ebb749SStefano Zampini     ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr);
228388ebb749SStefano Zampini   }
22848629588bSStefano Zampini   /* get back data */
22858629588bSStefano Zampini   *coarse_submat_vals_n = coarse_submat_vals;
228688ebb749SStefano Zampini   PetscFunctionReturn(0);
228788ebb749SStefano Zampini }
228888ebb749SStefano Zampini 
228988ebb749SStefano Zampini #undef __FUNCT__
2290d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted"
2291d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B)
2292aa0d41d4SStefano Zampini {
2293d65f70fdSStefano Zampini   Mat            *work_mat;
2294d65f70fdSStefano Zampini   IS             isrow_s,iscol_s;
2295d65f70fdSStefano Zampini   PetscBool      rsorted,csorted;
2296d65f70fdSStefano Zampini   PetscInt       rsize,*idxs_perm_r,csize,*idxs_perm_c;
2297aa0d41d4SStefano Zampini   PetscErrorCode ierr;
2298aa0d41d4SStefano Zampini 
2299aa0d41d4SStefano Zampini   PetscFunctionBegin;
2300d65f70fdSStefano Zampini   ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr);
2301d65f70fdSStefano Zampini   ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr);
2302d65f70fdSStefano Zampini   ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr);
2303d65f70fdSStefano Zampini   ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr);
2304aa0d41d4SStefano Zampini 
2305d65f70fdSStefano Zampini   if (!rsorted) {
2306906d46d4SStefano Zampini     const PetscInt *idxs;
2307906d46d4SStefano Zampini     PetscInt *idxs_sorted,i;
2308aa0d41d4SStefano Zampini 
2309d65f70fdSStefano Zampini     ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr);
2310d65f70fdSStefano Zampini     ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr);
2311d65f70fdSStefano Zampini     for (i=0;i<rsize;i++) {
2312d65f70fdSStefano Zampini       idxs_perm_r[i] = i;
2313aa0d41d4SStefano Zampini     }
2314d65f70fdSStefano Zampini     ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr);
2315d65f70fdSStefano Zampini     ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr);
2316d65f70fdSStefano Zampini     for (i=0;i<rsize;i++) {
2317d65f70fdSStefano Zampini       idxs_sorted[i] = idxs[idxs_perm_r[i]];
2318aa0d41d4SStefano Zampini     }
2319d65f70fdSStefano Zampini     ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr);
2320d65f70fdSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr);
2321d65f70fdSStefano Zampini   } else {
2322d65f70fdSStefano Zampini     ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
2323d65f70fdSStefano Zampini     isrow_s = isrow;
2324aa0d41d4SStefano Zampini   }
2325906d46d4SStefano Zampini 
2326d65f70fdSStefano Zampini   if (!csorted) {
2327d65f70fdSStefano Zampini     if (isrow == iscol) {
2328d65f70fdSStefano Zampini       ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr);
2329d65f70fdSStefano Zampini       iscol_s = isrow_s;
2330d65f70fdSStefano Zampini     } else {
2331d65f70fdSStefano Zampini       const PetscInt *idxs;
2332d65f70fdSStefano Zampini       PetscInt *idxs_sorted,i;
2333906d46d4SStefano Zampini 
2334d65f70fdSStefano Zampini       ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr);
2335d65f70fdSStefano Zampini       ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr);
2336d65f70fdSStefano Zampini       for (i=0;i<csize;i++) {
2337d65f70fdSStefano Zampini         idxs_perm_c[i] = i;
2338d65f70fdSStefano Zampini       }
2339d65f70fdSStefano Zampini       ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr);
2340d65f70fdSStefano Zampini       ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr);
2341d65f70fdSStefano Zampini       for (i=0;i<csize;i++) {
2342d65f70fdSStefano Zampini         idxs_sorted[i] = idxs[idxs_perm_c[i]];
2343d65f70fdSStefano Zampini       }
2344d65f70fdSStefano Zampini       ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr);
2345d65f70fdSStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr);
2346d65f70fdSStefano Zampini     }
2347d65f70fdSStefano Zampini   } else {
2348d65f70fdSStefano Zampini     ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
2349d65f70fdSStefano Zampini     iscol_s = iscol;
2350d65f70fdSStefano Zampini   }
2351d65f70fdSStefano Zampini 
2352d648f858SStefano Zampini   ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr);
2353d65f70fdSStefano Zampini 
2354d65f70fdSStefano Zampini   if (!rsorted || !csorted) {
2355906d46d4SStefano Zampini     Mat      new_mat;
2356d65f70fdSStefano Zampini     IS       is_perm_r,is_perm_c;
2357906d46d4SStefano Zampini 
2358d65f70fdSStefano Zampini     if (!rsorted) {
2359d65f70fdSStefano Zampini       PetscInt *idxs_r,i;
2360d65f70fdSStefano Zampini       ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr);
2361d65f70fdSStefano Zampini       for (i=0;i<rsize;i++) {
2362d65f70fdSStefano Zampini         idxs_r[idxs_perm_r[i]] = i;
2363906d46d4SStefano Zampini       }
2364d65f70fdSStefano Zampini       ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr);
2365d65f70fdSStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr);
2366d65f70fdSStefano Zampini     } else {
2367d65f70fdSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr);
2368906d46d4SStefano Zampini     }
2369d65f70fdSStefano Zampini     ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr);
2370d65f70fdSStefano Zampini 
2371d65f70fdSStefano Zampini     if (!csorted) {
2372d65f70fdSStefano Zampini       if (isrow_s == iscol_s) {
2373d65f70fdSStefano Zampini         ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr);
2374d65f70fdSStefano Zampini         is_perm_c = is_perm_r;
2375d65f70fdSStefano Zampini       } else {
2376d65f70fdSStefano Zampini         PetscInt *idxs_c,i;
2377d65f70fdSStefano Zampini         ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr);
2378d65f70fdSStefano Zampini         for (i=0;i<csize;i++) {
2379d65f70fdSStefano Zampini           idxs_c[idxs_perm_c[i]] = i;
2380d65f70fdSStefano Zampini         }
2381d65f70fdSStefano Zampini         ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr);
2382d65f70fdSStefano Zampini         ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr);
2383d65f70fdSStefano Zampini       }
2384d65f70fdSStefano Zampini     } else {
2385d65f70fdSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr);
2386d65f70fdSStefano Zampini     }
2387d65f70fdSStefano Zampini     ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr);
2388d65f70fdSStefano Zampini 
2389d65f70fdSStefano Zampini     ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr);
2390d65f70fdSStefano Zampini     ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr);
2391d65f70fdSStefano Zampini     work_mat[0] = new_mat;
2392d65f70fdSStefano Zampini     ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr);
2393d65f70fdSStefano Zampini     ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr);
2394d65f70fdSStefano Zampini   }
2395d65f70fdSStefano Zampini 
2396d65f70fdSStefano Zampini   ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr);
2397d65f70fdSStefano Zampini   *B = work_mat[0];
2398d65f70fdSStefano Zampini   ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr);
2399d65f70fdSStefano Zampini   ierr = ISDestroy(&isrow_s);CHKERRQ(ierr);
2400d65f70fdSStefano Zampini   ierr = ISDestroy(&iscol_s);CHKERRQ(ierr);
2401d65f70fdSStefano Zampini   PetscFunctionReturn(0);
2402d65f70fdSStefano Zampini }
2403d65f70fdSStefano Zampini 
2404d65f70fdSStefano Zampini #undef __FUNCT__
24055e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix"
24065e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix)
2407aa0d41d4SStefano Zampini {
2408aa0d41d4SStefano Zampini   Mat_IS*        matis = (Mat_IS*)pc->pmat->data;
24095e8657edSStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
2410d65f70fdSStefano Zampini   Mat            new_mat;
24115e8657edSStefano Zampini   IS             is_local,is_global;
2412d65f70fdSStefano Zampini   PetscInt       local_size;
2413d65f70fdSStefano Zampini   PetscBool      isseqaij;
2414aa0d41d4SStefano Zampini   PetscErrorCode ierr;
2415aa0d41d4SStefano Zampini 
2416aa0d41d4SStefano Zampini   PetscFunctionBegin;
2417aa0d41d4SStefano Zampini   ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
24185e8657edSStefano Zampini   ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr);
24195e8657edSStefano Zampini   ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr);
2420b087196eSStefano Zampini   ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr);
2421aa0d41d4SStefano Zampini   ierr = ISDestroy(&is_local);CHKERRQ(ierr);
2422d648f858SStefano Zampini   ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr);
2423aa0d41d4SStefano Zampini   ierr = ISDestroy(&is_global);CHKERRQ(ierr);
2424906d46d4SStefano Zampini 
2425906d46d4SStefano Zampini   /* check */
2426906d46d4SStefano Zampini   if (pcbddc->dbg_flag) {
2427906d46d4SStefano Zampini     Vec       x,x_change;
2428906d46d4SStefano Zampini     PetscReal error;
2429906d46d4SStefano Zampini 
24305e8657edSStefano Zampini     ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr);
2431906d46d4SStefano Zampini     ierr = VecSetRandom(x,NULL);CHKERRQ(ierr);
24325e8657edSStefano Zampini     ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr);
2433e176bc59SStefano Zampini     ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2434e176bc59SStefano Zampini     ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2435d65f70fdSStefano Zampini     ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr);
2436e176bc59SStefano Zampini     ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2437e176bc59SStefano Zampini     ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2438906d46d4SStefano Zampini     ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr);
2439906d46d4SStefano Zampini     ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr);
2440906d46d4SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2441906d46d4SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr);
2442906d46d4SStefano Zampini     ierr = VecDestroy(&x);CHKERRQ(ierr);
2443906d46d4SStefano Zampini     ierr = VecDestroy(&x_change);CHKERRQ(ierr);
2444906d46d4SStefano Zampini   }
2445906d46d4SStefano Zampini 
244622d5777bSStefano Zampini   /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */
24479b28b3ffSStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
244822d5777bSStefano Zampini   if (isseqaij) {
24491cf9b237SStefano Zampini     Mat M;
24501cf9b237SStefano Zampini 
24511cf9b237SStefano Zampini     ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr);
24521cf9b237SStefano Zampini     ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr);
24531cf9b237SStefano Zampini     ierr = MatDestroy(&M);CHKERRQ(ierr);
2454aa0d41d4SStefano Zampini   } else {
24551cf9b237SStefano Zampini     Mat work_mat,M;
24561cf9b237SStefano Zampini 
2457aa0d41d4SStefano Zampini     ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr);
24581cf9b237SStefano Zampini     ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr);
24591cf9b237SStefano Zampini     ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr);
24601cf9b237SStefano Zampini     ierr = MatDestroy(&M);CHKERRQ(ierr);
2461aa0d41d4SStefano Zampini     ierr = MatDestroy(&work_mat);CHKERRQ(ierr);
2462aa0d41d4SStefano Zampini   }
24633301b35fSStefano Zampini   if (matis->A->symmetric_set) {
24643301b35fSStefano Zampini     ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr);
2465e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX)
24663301b35fSStefano Zampini     ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr);
2467e496cd5dSStefano Zampini #endif
24683301b35fSStefano Zampini   }
246945a1bb75SStefano Zampini   /*
247045a1bb75SStefano Zampini   ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
2471d65f70fdSStefano Zampini   ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr);
247245a1bb75SStefano Zampini   */
2473d65f70fdSStefano Zampini   ierr = MatDestroy(&new_mat);CHKERRQ(ierr);
2474aa0d41d4SStefano Zampini   PetscFunctionReturn(0);
2475aa0d41d4SStefano Zampini }
2476aa0d41d4SStefano Zampini 
2477aa0d41d4SStefano Zampini #undef __FUNCT__
2478a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters"
24798ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc)
2480a64d13efSStefano Zampini {
2481a64d13efSStefano Zampini   PC_IS*          pcis = (PC_IS*)(pc->data);
2482a64d13efSStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)pc->data;
2483d62866d3SStefano Zampini   PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
248453892102SStefano Zampini   PetscInt        *idx_R_local=NULL;
24853a50541eSStefano Zampini   PetscInt        n_vertices,i,j,n_R,n_D,n_B;
24863a50541eSStefano Zampini   PetscInt        vbs,bs;
24876816873aSStefano Zampini   PetscBT         bitmask=NULL;
2488a64d13efSStefano Zampini   PetscErrorCode  ierr;
2489a64d13efSStefano Zampini 
2490a64d13efSStefano Zampini   PetscFunctionBegin;
2491b23d619eSStefano Zampini   /*
2492b23d619eSStefano Zampini     No need to setup local scatters if
2493b23d619eSStefano Zampini       - primal space is unchanged
2494b23d619eSStefano Zampini         AND
2495b23d619eSStefano Zampini       - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains)
2496b23d619eSStefano Zampini         AND
2497b23d619eSStefano Zampini       - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine
2498b23d619eSStefano Zampini   */
2499b23d619eSStefano Zampini   if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) {
2500f4ddd8eeSStefano Zampini     PetscFunctionReturn(0);
2501f4ddd8eeSStefano Zampini   }
2502f4ddd8eeSStefano Zampini   /* destroy old objects */
2503f4ddd8eeSStefano Zampini   ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr);
2504f4ddd8eeSStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr);
2505f4ddd8eeSStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr);
2506a64d13efSStefano Zampini   /* Set Non-overlapping dimensions */
2507b371cd4fSStefano Zampini   n_B = pcis->n_B;
2508b371cd4fSStefano Zampini   n_D = pcis->n - n_B;
2509b371cd4fSStefano Zampini   n_vertices = pcbddc->n_vertices;
25103a50541eSStefano Zampini 
2511a64d13efSStefano Zampini   /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */
25126816873aSStefano Zampini 
251353892102SStefano Zampini   /* create auxiliary bitmask and allocate workspace */
2514df4d28bfSStefano Zampini   if (!sub_schurs->reuse_solver) {
2515854ce69bSBarry Smith     ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr);
2516a64d13efSStefano Zampini     ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr);
2517a64d13efSStefano Zampini     for (i=0;i<n_vertices;i++) {
25180e6343abSStefano Zampini       ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr);
2519a64d13efSStefano Zampini     }
2520a64d13efSStefano Zampini 
2521a64d13efSStefano Zampini     for (i=0, n_R=0; i<pcis->n; i++) {
25224641a718SStefano Zampini       if (!PetscBTLookup(bitmask,i)) {
25236816873aSStefano Zampini         idx_R_local[n_R++] = i;
2524a64d13efSStefano Zampini       }
2525a64d13efSStefano Zampini     }
2526df4d28bfSStefano Zampini   } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */
2527df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
25286816873aSStefano Zampini 
2529df4d28bfSStefano Zampini     ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
2530df4d28bfSStefano Zampini     ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr);
25316816873aSStefano Zampini   }
25323a50541eSStefano Zampini 
25333a50541eSStefano Zampini   /* Block code */
25343a50541eSStefano Zampini   vbs = 1;
25353a50541eSStefano Zampini   ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr);
25363a50541eSStefano Zampini   if (bs>1 && !(n_vertices%bs)) {
25373a50541eSStefano Zampini     PetscBool is_blocked = PETSC_TRUE;
25383a50541eSStefano Zampini     PetscInt  *vary;
2539df4d28bfSStefano Zampini     if (!sub_schurs->reuse_solver) {
2540785e854fSJed Brown       ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr);
25413a50541eSStefano Zampini       ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr);
2542d3df7717SStefano Zampini       /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */
2543d3df7717SStefano Zampini       /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */
25440e6343abSStefano Zampini       for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++;
2545d3df7717SStefano Zampini       for (i=0; i<pcis->n/bs; i++) {
25463a50541eSStefano Zampini         if (vary[i]!=0 && vary[i]!=bs) {
25473a50541eSStefano Zampini           is_blocked = PETSC_FALSE;
25483a50541eSStefano Zampini           break;
25493a50541eSStefano Zampini         }
25503a50541eSStefano Zampini       }
2551d3df7717SStefano Zampini       ierr = PetscFree(vary);CHKERRQ(ierr);
2552d3df7717SStefano Zampini     } else {
2553d3df7717SStefano Zampini       /* Verify directly the R set */
2554d3df7717SStefano Zampini       for (i=0; i<n_R/bs; i++) {
2555d3df7717SStefano Zampini         PetscInt j,node=idx_R_local[bs*i];
2556d3df7717SStefano Zampini         for (j=1; j<bs; j++) {
2557d3df7717SStefano Zampini           if (node != idx_R_local[bs*i+j]-j) {
2558d3df7717SStefano Zampini             is_blocked = PETSC_FALSE;
2559d3df7717SStefano Zampini             break;
2560d3df7717SStefano Zampini           }
2561d3df7717SStefano Zampini         }
2562d3df7717SStefano Zampini       }
2563d3df7717SStefano Zampini     }
25643a50541eSStefano Zampini     if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */
25653a50541eSStefano Zampini       vbs = bs;
25663a50541eSStefano Zampini       for (i=0;i<n_R/vbs;i++) {
25673a50541eSStefano Zampini         idx_R_local[i] = idx_R_local[vbs*i]/vbs;
25683a50541eSStefano Zampini       }
25693a50541eSStefano Zampini     }
25703a50541eSStefano Zampini   }
25713a50541eSStefano Zampini   ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr);
2572df4d28bfSStefano Zampini   if (sub_schurs->reuse_solver) {
2573df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
257453892102SStefano Zampini 
2575df4d28bfSStefano Zampini     ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
2576df4d28bfSStefano Zampini     ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr);
257753892102SStefano Zampini     ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr);
2578df4d28bfSStefano Zampini     reuse_solver->is_R = pcbddc->is_R_local;
257953892102SStefano Zampini   } else {
25803a50541eSStefano Zampini     ierr = PetscFree(idx_R_local);CHKERRQ(ierr);
258153892102SStefano Zampini   }
2582a64d13efSStefano Zampini 
2583a64d13efSStefano Zampini   /* print some info if requested */
2584a64d13efSStefano Zampini   if (pcbddc->dbg_flag) {
2585a64d13efSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
2586a64d13efSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
25871575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
2588a64d13efSStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr);
2589a64d13efSStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr);
25904f1b2e48SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %d, v_size = %d, constraints = %d, local_primal_size = %d\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices-pcbddc->benign_n,pcbddc->local_primal_size);CHKERRQ(ierr);
2591a64d13efSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2592a64d13efSStefano Zampini   }
2593a64d13efSStefano Zampini 
2594a64d13efSStefano Zampini   /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */
2595df4d28bfSStefano Zampini   if (!sub_schurs->reuse_solver) {
25966816873aSStefano Zampini     IS       is_aux1,is_aux2;
25976816873aSStefano Zampini     PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local;
25986816873aSStefano Zampini 
25993a50541eSStefano Zampini     ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
2600854ce69bSBarry Smith     ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr);
2601854ce69bSBarry Smith     ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr);
2602a64d13efSStefano Zampini     ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
26034641a718SStefano Zampini     for (i=0; i<n_D; i++) {
26044641a718SStefano Zampini       ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr);
26054641a718SStefano Zampini     }
2606a64d13efSStefano Zampini     ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
2607a64d13efSStefano Zampini     for (i=0, j=0; i<n_R; i++) {
26084641a718SStefano Zampini       if (!PetscBTLookup(bitmask,idx_R_local[i])) {
26094641a718SStefano Zampini         aux_array1[j++] = i;
2610a64d13efSStefano Zampini       }
2611a64d13efSStefano Zampini     }
2612a64d13efSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr);
2613a64d13efSStefano Zampini     ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
2614a64d13efSStefano Zampini     for (i=0, j=0; i<n_B; i++) {
26154641a718SStefano Zampini       if (!PetscBTLookup(bitmask,is_indices[i])) {
26164641a718SStefano Zampini         aux_array2[j++] = i;
2617a64d13efSStefano Zampini       }
2618a64d13efSStefano Zampini     }
2619a64d13efSStefano Zampini     ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
2620a64d13efSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr);
2621a64d13efSStefano Zampini     ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr);
2622a64d13efSStefano Zampini     ierr = ISDestroy(&is_aux1);CHKERRQ(ierr);
2623a64d13efSStefano Zampini     ierr = ISDestroy(&is_aux2);CHKERRQ(ierr);
2624a64d13efSStefano Zampini 
26258eeda7d8SStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
2626785e854fSJed Brown       ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr);
2627a64d13efSStefano Zampini       for (i=0, j=0; i<n_R; i++) {
26284641a718SStefano Zampini         if (PetscBTLookup(bitmask,idx_R_local[i])) {
26294641a718SStefano Zampini           aux_array1[j++] = i;
2630a64d13efSStefano Zampini         }
2631a64d13efSStefano Zampini       }
2632a64d13efSStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr);
2633a64d13efSStefano Zampini       ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr);
2634a64d13efSStefano Zampini       ierr = ISDestroy(&is_aux1);CHKERRQ(ierr);
2635a64d13efSStefano Zampini     }
26364641a718SStefano Zampini     ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr);
26373a50541eSStefano Zampini     ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
2638d62866d3SStefano Zampini   } else {
2639df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
26406816873aSStefano Zampini     IS                 tis;
26416816873aSStefano Zampini     PetscInt           schur_size;
26426816873aSStefano Zampini 
2643df4d28bfSStefano Zampini     ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr);
26446816873aSStefano Zampini     ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr);
2645df4d28bfSStefano Zampini     ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr);
26466816873aSStefano Zampini     ierr = ISDestroy(&tis);CHKERRQ(ierr);
26476816873aSStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
26486816873aSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr);
26496816873aSStefano Zampini       ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr);
26506816873aSStefano Zampini       ierr = ISDestroy(&tis);CHKERRQ(ierr);
2651d62866d3SStefano Zampini     }
2652d62866d3SStefano Zampini   }
2653a64d13efSStefano Zampini   PetscFunctionReturn(0);
2654a64d13efSStefano Zampini }
2655a64d13efSStefano Zampini 
2656304d26faSStefano Zampini 
2657304d26faSStefano Zampini #undef __FUNCT__
2658304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers"
2659684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann)
2660304d26faSStefano Zampini {
2661304d26faSStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
2662304d26faSStefano Zampini   PC_IS          *pcis = (PC_IS*)pc->data;
2663304d26faSStefano Zampini   PC             pc_temp;
2664304d26faSStefano Zampini   Mat            A_RR;
2665f4ddd8eeSStefano Zampini   MatReuse       reuse;
2666304d26faSStefano Zampini   PetscScalar    m_one = -1.0;
2667304d26faSStefano Zampini   PetscReal      value;
266804708bb6SStefano Zampini   PetscInt       n_D,n_R;
26699577ea80SStefano Zampini   PetscBool      use_exact,use_exact_reduced,issbaij;
2670304d26faSStefano Zampini   PetscErrorCode ierr;
2671e604994aSStefano Zampini   /* prefixes stuff */
2672312be037SStefano Zampini   char           dir_prefix[256],neu_prefix[256],str_level[16];
2673e604994aSStefano Zampini   size_t         len;
2674304d26faSStefano Zampini 
2675304d26faSStefano Zampini   PetscFunctionBegin;
2676304d26faSStefano Zampini 
2677e604994aSStefano Zampini   /* compute prefixes */
2678e604994aSStefano Zampini   ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr);
2679e604994aSStefano Zampini   ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr);
2680e604994aSStefano Zampini   if (!pcbddc->current_level) {
2681e604994aSStefano Zampini     ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
2682e604994aSStefano Zampini     ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
2683e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr);
2684e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr);
2685e604994aSStefano Zampini   } else {
2686e604994aSStefano Zampini     ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr);
2687312be037SStefano Zampini     sprintf(str_level,"l%d_",(int)(pcbddc->current_level));
2688e604994aSStefano Zampini     ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr);
2689e604994aSStefano Zampini     len -= 15; /* remove "pc_bddc_coarse_" */
2690312be037SStefano Zampini     if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
2691312be037SStefano Zampini     if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
269234d6797cSStefano Zampini     ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
269334d6797cSStefano Zampini     ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
2694e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr);
2695e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr);
2696e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr);
2697e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr);
2698e604994aSStefano Zampini   }
2699e604994aSStefano Zampini 
2700304d26faSStefano Zampini   /* DIRICHLET PROBLEM */
2701684f6988SStefano Zampini   if (dirichlet) {
2702d5574798SStefano Zampini     PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
2703a3df083aSStefano Zampini     if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) {
2704a3df083aSStefano Zampini       Mat    A_IIn;
2705a3df083aSStefano Zampini 
2706a3df083aSStefano Zampini       ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr);
2707a3df083aSStefano Zampini       ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr);
2708a3df083aSStefano Zampini       pcis->A_II = A_IIn;
2709a3df083aSStefano Zampini     }
27103301b35fSStefano Zampini     if (pcbddc->local_mat->symmetric_set) {
27113301b35fSStefano Zampini       ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr);
2712964fefecSStefano Zampini     }
2713ac78edfcSStefano Zampini     /* Matrix for Dirichlet problem is pcis->A_II */
2714964fefecSStefano Zampini     n_D = pcis->n - pcis->n_B;
2715304d26faSStefano Zampini     if (!pcbddc->ksp_D) { /* create object if not yet build */
2716304d26faSStefano Zampini       ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr);
2717304d26faSStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr);
2718304d26faSStefano Zampini       /* default */
2719304d26faSStefano Zampini       ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr);
2720e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr);
27219577ea80SStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
2722304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr);
27239577ea80SStefano Zampini       if (issbaij) {
27249577ea80SStefano Zampini         ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr);
27259577ea80SStefano Zampini       } else {
2726304d26faSStefano Zampini         ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr);
27279577ea80SStefano Zampini       }
2728304d26faSStefano Zampini       /* Allow user's customization */
2729304d26faSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr);
2730304d26faSStefano Zampini       ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
2731304d26faSStefano Zampini     }
2732d1e9a80fSBarry Smith     ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr);
2733df4d28bfSStefano Zampini     if (sub_schurs->reuse_solver) {
2734df4d28bfSStefano Zampini       PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
2735d62866d3SStefano Zampini 
2736df4d28bfSStefano Zampini       ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr);
2737d5574798SStefano Zampini     }
2738304d26faSStefano Zampini     /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
2739304d26faSStefano Zampini     if (!n_D) {
2740304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr);
2741304d26faSStefano Zampini       ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr);
2742304d26faSStefano Zampini     }
2743304d26faSStefano Zampini     /* Set Up KSP for Dirichlet problem of BDDC */
2744304d26faSStefano Zampini     ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr);
2745304d26faSStefano Zampini     /* set ksp_D into pcis data */
2746304d26faSStefano Zampini     ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr);
2747304d26faSStefano Zampini     ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr);
2748304d26faSStefano Zampini     pcis->ksp_D = pcbddc->ksp_D;
2749684f6988SStefano Zampini   }
2750304d26faSStefano Zampini 
2751304d26faSStefano Zampini   /* NEUMANN PROBLEM */
2752684f6988SStefano Zampini   A_RR = 0;
2753684f6988SStefano Zampini   if (neumann) {
2754d62866d3SStefano Zampini     PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
275504708bb6SStefano Zampini     PetscInt        ibs,mbs;
275604708bb6SStefano Zampini     PetscBool       issbaij;
275704708bb6SStefano Zampini     Mat_IS*         matis = (Mat_IS*)pc->pmat->data;
2758f4ddd8eeSStefano Zampini     /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */
27598ce42a96SStefano Zampini     ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr);
2760f4ddd8eeSStefano Zampini     if (pcbddc->ksp_R) { /* already created ksp */
2761f4ddd8eeSStefano Zampini       PetscInt nn_R;
276281d9aea3SBarry Smith       ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr);
2763f4ddd8eeSStefano Zampini       ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr);
2764f4ddd8eeSStefano Zampini       ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr);
2765f4ddd8eeSStefano Zampini       if (nn_R != n_R) { /* old ksp is not reusable, so reset it */
2766f4ddd8eeSStefano Zampini         ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr);
2767f4ddd8eeSStefano Zampini         ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
2768f4ddd8eeSStefano Zampini         reuse = MAT_INITIAL_MATRIX;
2769f4ddd8eeSStefano Zampini       } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */
2770727cdba6SStefano Zampini         if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */
2771f4ddd8eeSStefano Zampini           ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
2772f4ddd8eeSStefano Zampini           reuse = MAT_INITIAL_MATRIX;
2773f4ddd8eeSStefano Zampini         } else { /* safe to reuse the matrix */
2774f4ddd8eeSStefano Zampini           reuse = MAT_REUSE_MATRIX;
2775f4ddd8eeSStefano Zampini         }
2776f4ddd8eeSStefano Zampini       }
2777f4ddd8eeSStefano Zampini       /* last check */
2778d1e9a80fSBarry Smith       if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
2779f4ddd8eeSStefano Zampini         ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
2780f4ddd8eeSStefano Zampini         reuse = MAT_INITIAL_MATRIX;
2781f4ddd8eeSStefano Zampini       }
2782f4ddd8eeSStefano Zampini     } else { /* first time, so we need to create the matrix */
2783f4ddd8eeSStefano Zampini       reuse = MAT_INITIAL_MATRIX;
2784f4ddd8eeSStefano Zampini     }
2785f4ddd8eeSStefano Zampini     /* extract A_RR */
2786af732b37SStefano Zampini     ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr);
2787af732b37SStefano Zampini     ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr);
278804708bb6SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
278904708bb6SStefano Zampini     if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */
279004708bb6SStefano Zampini       if (matis->A == pcbddc->local_mat) {
279104708bb6SStefano Zampini         ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
279204708bb6SStefano Zampini         ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
2793af732b37SStefano Zampini       } else {
279404708bb6SStefano Zampini         ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
27956816873aSStefano Zampini       }
279604708bb6SStefano Zampini     } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */
279704708bb6SStefano Zampini       if (matis->A == pcbddc->local_mat) {
279804708bb6SStefano Zampini         ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
279904708bb6SStefano Zampini         ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
280004708bb6SStefano Zampini       } else {
280104708bb6SStefano Zampini         ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
280204708bb6SStefano Zampini       }
280304708bb6SStefano Zampini     }
2804f4ddd8eeSStefano Zampini     ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr);
28053301b35fSStefano Zampini     if (pcbddc->local_mat->symmetric_set) {
28063301b35fSStefano Zampini       ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr);
28076816873aSStefano Zampini     }
2808a3df083aSStefano Zampini     if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) {
2809a3df083aSStefano Zampini       Mat A_RRn;
2810a3df083aSStefano Zampini       ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr);
2811a3df083aSStefano Zampini       ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
2812a3df083aSStefano Zampini       A_RR = A_RRn;
2813a3df083aSStefano Zampini     }
2814f4ddd8eeSStefano Zampini     if (!pcbddc->ksp_R) { /* create object if not present */
2815304d26faSStefano Zampini       ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr);
2816304d26faSStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr);
2817304d26faSStefano Zampini       /* default */
2818304d26faSStefano Zampini       ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr);
2819e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr);
2820304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr);
28219577ea80SStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
28229577ea80SStefano Zampini       if (issbaij) {
28239577ea80SStefano Zampini         ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr);
28249577ea80SStefano Zampini       } else {
2825304d26faSStefano Zampini         ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr);
28269577ea80SStefano Zampini       }
2827304d26faSStefano Zampini       /* Allow user's customization */
2828304d26faSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr);
2829304d26faSStefano Zampini       ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
2830304d26faSStefano Zampini     }
2831304d26faSStefano Zampini     /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
2832304d26faSStefano Zampini     if (!n_R) {
2833304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr);
2834304d26faSStefano Zampini       ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr);
2835304d26faSStefano Zampini     }
2836df4d28bfSStefano Zampini     /* Reuse solver if it is present */
2837df4d28bfSStefano Zampini     if (sub_schurs->reuse_solver) {
2838df4d28bfSStefano Zampini       PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
2839d62866d3SStefano Zampini 
2840df4d28bfSStefano Zampini       ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr);
2841d62866d3SStefano Zampini     }
2842ffd830a3SStefano Zampini     ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr);
2843304d26faSStefano Zampini     /* Set Up KSP for Neumann problem of BDDC */
2844304d26faSStefano Zampini     ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr);
2845684f6988SStefano Zampini   }
28466816873aSStefano Zampini   /* free Neumann problem's matrix */
28476816873aSStefano Zampini   ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
2848304d26faSStefano Zampini 
2849304d26faSStefano Zampini   /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */
28500fccc4e9SStefano Zampini   if (pcbddc->NullSpace || pcbddc->dbg_flag) {
2851684f6988SStefano Zampini     if (pcbddc->dbg_flag) {
2852684f6988SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
28531575c14dSBarry Smith       ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
2854684f6988SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
2855684f6988SStefano Zampini     }
2856684f6988SStefano Zampini     if (dirichlet) { /* Dirichlet */
28570fccc4e9SStefano Zampini       ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr);
28580fccc4e9SStefano Zampini       ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr);
28590fccc4e9SStefano Zampini       ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr);
28600fccc4e9SStefano Zampini       ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr);
28610fccc4e9SStefano Zampini       ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr);
2862304d26faSStefano Zampini       /* need to be adapted? */
2863b8ffe317SStefano Zampini       use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE);
2864b8ffe317SStefano Zampini       ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
2865b8ffe317SStefano Zampini       ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr);
2866304d26faSStefano Zampini       /* print info */
2867304d26faSStefano Zampini       if (pcbddc->dbg_flag) {
2868e604994aSStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Dirichlet solve (%s) = % 1.14e \n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_D))->prefix,value);CHKERRQ(ierr);
2869304d26faSStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2870304d26faSStefano Zampini       }
2871b8ffe317SStefano Zampini       if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) {
2872298c0119SStefano Zampini         ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr);
2873304d26faSStefano Zampini       }
2874684f6988SStefano Zampini     }
2875684f6988SStefano Zampini     if (neumann) { /* Neumann */
28766816873aSStefano Zampini       ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr);
28770fccc4e9SStefano Zampini       ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr);
28780fccc4e9SStefano Zampini       ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
28790fccc4e9SStefano Zampini       ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr);
28800fccc4e9SStefano Zampini       ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr);
28810fccc4e9SStefano Zampini       ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr);
2882304d26faSStefano Zampini       /* need to be adapted? */
2883b8ffe317SStefano Zampini       use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE);
2884b8ffe317SStefano Zampini       ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
2885304d26faSStefano Zampini       /* print info */
2886304d26faSStefano Zampini       if (pcbddc->dbg_flag) {
2887e604994aSStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Neumann solve (%s) = % 1.14e\n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_R))->prefix,value);CHKERRQ(ierr);
2888304d26faSStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2889304d26faSStefano Zampini       }
2890b8ffe317SStefano Zampini       if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */
2891298c0119SStefano Zampini         ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr);
2892304d26faSStefano Zampini       }
28930fccc4e9SStefano Zampini     }
2894684f6988SStefano Zampini   }
2895304d26faSStefano Zampini   PetscFunctionReturn(0);
2896304d26faSStefano Zampini }
2897304d26faSStefano Zampini 
2898304d26faSStefano Zampini #undef __FUNCT__
2899ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection"
290080677318SStefano Zampini static PetscErrorCode  PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose)
2901674ae819SStefano Zampini {
2902674ae819SStefano Zampini   PetscErrorCode  ierr;
2903674ae819SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)(pc->data);
2904be83ff47SStefano Zampini   PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
2905674ae819SStefano Zampini 
2906674ae819SStefano Zampini   PetscFunctionBegin;
2907df4d28bfSStefano Zampini   if (!sub_schurs->reuse_solver) {
290880677318SStefano Zampini     ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr);
290920c7b377SStefano Zampini   }
291080677318SStefano Zampini   if (!pcbddc->switch_static) {
291180677318SStefano Zampini     if (applytranspose && pcbddc->local_auxmat1) {
291280677318SStefano Zampini       ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr);
291380677318SStefano Zampini       ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr);
291420c7b377SStefano Zampini     }
2915df4d28bfSStefano Zampini     if (!sub_schurs->reuse_solver) {
291680677318SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
291780677318SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
291820c7b377SStefano Zampini     } else {
2919df4d28bfSStefano Zampini       PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
2920be83ff47SStefano Zampini 
2921df4d28bfSStefano Zampini       ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2922df4d28bfSStefano Zampini       ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
292320c7b377SStefano Zampini     }
2924be83ff47SStefano Zampini   } else {
292580677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
292680677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
292780677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
292880677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
292980677318SStefano Zampini     if (applytranspose && pcbddc->local_auxmat1) {
293080677318SStefano Zampini       ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr);
293180677318SStefano Zampini       ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr);
293280677318SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
293380677318SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2934674ae819SStefano Zampini     }
2935674ae819SStefano Zampini   }
2936df4d28bfSStefano Zampini   if (!sub_schurs->reuse_solver) {
293780677318SStefano Zampini     if (applytranspose) {
293880677318SStefano Zampini       ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr);
293980677318SStefano Zampini     } else {
294080677318SStefano Zampini       ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr);
294180677318SStefano Zampini     }
2942be83ff47SStefano Zampini   } else {
2943df4d28bfSStefano Zampini     PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
2944be83ff47SStefano Zampini 
2945be83ff47SStefano Zampini     if (applytranspose) {
2946df4d28bfSStefano Zampini       ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr);
2947be83ff47SStefano Zampini     } else {
2948df4d28bfSStefano Zampini       ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr);
2949be83ff47SStefano Zampini     }
2950be83ff47SStefano Zampini   }
295180677318SStefano Zampini   ierr = VecSet(inout_B,0.);CHKERRQ(ierr);
295280677318SStefano Zampini   if (!pcbddc->switch_static) {
2953df4d28bfSStefano Zampini     if (!sub_schurs->reuse_solver) {
295480677318SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
295580677318SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2956be83ff47SStefano Zampini     } else {
2957df4d28bfSStefano Zampini       PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
2958be83ff47SStefano Zampini 
2959df4d28bfSStefano Zampini       ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2960df4d28bfSStefano Zampini       ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2961be83ff47SStefano Zampini     }
296280677318SStefano Zampini     if (!applytranspose && pcbddc->local_auxmat1) {
296380677318SStefano Zampini       ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr);
296480677318SStefano Zampini       ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr);
296580677318SStefano Zampini     }
296680677318SStefano Zampini   } else {
296780677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
296880677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
296980677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
297080677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
297180677318SStefano Zampini     if (!applytranspose && pcbddc->local_auxmat1) {
297280677318SStefano Zampini       ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr);
297380677318SStefano Zampini       ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr);
297480677318SStefano Zampini     }
297580677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
297680677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
297780677318SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
297880677318SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2979674ae819SStefano Zampini   }
2980674ae819SStefano Zampini   PetscFunctionReturn(0);
2981674ae819SStefano Zampini }
2982674ae819SStefano Zampini 
2983dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */
2984674ae819SStefano Zampini #undef __FUNCT__
2985674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner"
2986dc359a40SStefano Zampini PetscErrorCode  PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose)
2987674ae819SStefano Zampini {
2988674ae819SStefano Zampini   PetscErrorCode ierr;
2989674ae819SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)(pc->data);
2990674ae819SStefano Zampini   PC_IS*            pcis = (PC_IS*)  (pc->data);
2991674ae819SStefano Zampini   const PetscScalar zero = 0.0;
2992674ae819SStefano Zampini 
2993674ae819SStefano Zampini   PetscFunctionBegin;
2994dc359a40SStefano Zampini   /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */
2995*4fee134fSStefano Zampini   if (!pcbddc->benign_apply_coarse_only) {
2996dc359a40SStefano Zampini     if (applytranspose) {
2997674ae819SStefano Zampini       ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr);
29988eeda7d8SStefano Zampini       if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); }
2999dc359a40SStefano Zampini     } else {
3000674ae819SStefano Zampini       ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr);
3001674ae819SStefano Zampini       if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); }
300215aaf578SStefano Zampini     }
3003*4fee134fSStefano Zampini   } else {
3004*4fee134fSStefano Zampini     ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr);
3005*4fee134fSStefano Zampini   }
3006efc2fbd9SStefano Zampini 
3007efc2fbd9SStefano Zampini   /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */
30084f1b2e48SStefano Zampini   if (pcbddc->benign_n) {
3009efc2fbd9SStefano Zampini     PetscScalar *array;
30104f1b2e48SStefano Zampini     PetscInt    j;
3011efc2fbd9SStefano Zampini 
3012efc2fbd9SStefano Zampini     ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
30134f1b2e48SStefano Zampini     for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j];
3014efc2fbd9SStefano Zampini     ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
3015efc2fbd9SStefano Zampini   }
3016efc2fbd9SStefano Zampini 
301712edc857SStefano Zampini   /* start communications from local primal nodes to rhs of coarse solver */
301812edc857SStefano Zampini   ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr);
301912edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
302012edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
302112edc857SStefano Zampini 
30229f00e9b4SStefano Zampini   /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */
302312edc857SStefano Zampini   /* TODO remove null space when doing multilevel */
302412edc857SStefano Zampini   if (pcbddc->coarse_ksp) {
302551694757SStefano Zampini     Mat coarse_mat;
3026964fefecSStefano Zampini     Vec rhs,sol;
302751694757SStefano Zampini     MatNullSpace nullsp;
3028964fefecSStefano Zampini 
3029964fefecSStefano Zampini     ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr);
3030964fefecSStefano Zampini     ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr);
303151694757SStefano Zampini     ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr);
303251694757SStefano Zampini     ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr);
303351694757SStefano Zampini     if (nullsp) {
303451694757SStefano Zampini       ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr);
303551694757SStefano Zampini     }
303612edc857SStefano Zampini     if (applytranspose) {
3037964fefecSStefano Zampini       ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr);
303812edc857SStefano Zampini     } else {
3039964fefecSStefano Zampini       ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr);
304012edc857SStefano Zampini     }
304151694757SStefano Zampini     if (nullsp) {
304251694757SStefano Zampini       ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr);
304351694757SStefano Zampini     }
304412edc857SStefano Zampini   }
3045674ae819SStefano Zampini 
3046674ae819SStefano Zampini   /* Local solution on R nodes */
3047*4fee134fSStefano Zampini   if (pcis->n && !pcbddc->benign_apply_coarse_only) {
304880677318SStefano Zampini     ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr);
30499f00e9b4SStefano Zampini   }
30509f00e9b4SStefano Zampini   /* communications from coarse sol to local primal nodes */
30519f00e9b4SStefano Zampini   ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
305212edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
3053674ae819SStefano Zampini 
3054*4fee134fSStefano Zampini   /* Sum contributions from the two levels */
3055*4fee134fSStefano Zampini   if (!pcbddc->benign_apply_coarse_only) {
3056dc359a40SStefano Zampini     if (applytranspose) {
3057dc359a40SStefano Zampini       ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr);
3058dc359a40SStefano Zampini       if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); }
3059dc359a40SStefano Zampini     } else {
3060674ae819SStefano Zampini       ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr);
30618eeda7d8SStefano Zampini       if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); }
3062dc359a40SStefano Zampini     }
3063efc2fbd9SStefano Zampini     /* store p0 */
30644f1b2e48SStefano Zampini     if (pcbddc->benign_n) {
3065efc2fbd9SStefano Zampini       PetscScalar *array;
30664f1b2e48SStefano Zampini       PetscInt    j;
3067efc2fbd9SStefano Zampini 
3068efc2fbd9SStefano Zampini       ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
30694f1b2e48SStefano Zampini       for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j];
3070efc2fbd9SStefano Zampini       ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
3071efc2fbd9SStefano Zampini     }
3072*4fee134fSStefano Zampini   } else { /* expand the coarse solution */
3073*4fee134fSStefano Zampini     if (applytranspose) {
3074*4fee134fSStefano Zampini       ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr);
3075*4fee134fSStefano Zampini     } else {
3076*4fee134fSStefano Zampini       ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr);
3077*4fee134fSStefano Zampini     }
3078*4fee134fSStefano Zampini   }
3079674ae819SStefano Zampini   PetscFunctionReturn(0);
3080674ae819SStefano Zampini }
3081674ae819SStefano Zampini 
3082674ae819SStefano Zampini #undef __FUNCT__
3083674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin"
308412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode)
3085674ae819SStefano Zampini {
3086674ae819SStefano Zampini   PetscErrorCode ierr;
3087674ae819SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)(pc->data);
308858da7f69SStefano Zampini   PetscScalar    *array;
308912edc857SStefano Zampini   Vec            from,to;
3090674ae819SStefano Zampini 
3091674ae819SStefano Zampini   PetscFunctionBegin;
309212edc857SStefano Zampini   if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
309312edc857SStefano Zampini     from = pcbddc->coarse_vec;
309412edc857SStefano Zampini     to = pcbddc->vec1_P;
309512edc857SStefano Zampini     if (pcbddc->coarse_ksp) { /* get array from coarse processes */
309612edc857SStefano Zampini       Vec tvec;
309758da7f69SStefano Zampini 
309858da7f69SStefano Zampini       ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
309958da7f69SStefano Zampini       ierr = VecResetArray(tvec);CHKERRQ(ierr);
310012edc857SStefano Zampini       ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
310158da7f69SStefano Zampini       ierr = VecGetArray(tvec,&array);CHKERRQ(ierr);
310258da7f69SStefano Zampini       ierr = VecPlaceArray(from,array);CHKERRQ(ierr);
310358da7f69SStefano Zampini       ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr);
310412edc857SStefano Zampini     }
310512edc857SStefano Zampini   } else { /* from local to global -> put data in coarse right hand side */
310612edc857SStefano Zampini     from = pcbddc->vec1_P;
310712edc857SStefano Zampini     to = pcbddc->coarse_vec;
310812edc857SStefano Zampini   }
310912edc857SStefano Zampini   ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr);
3110674ae819SStefano Zampini   PetscFunctionReturn(0);
3111674ae819SStefano Zampini }
3112674ae819SStefano Zampini 
3113674ae819SStefano Zampini #undef __FUNCT__
3114674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd"
311512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode)
3116674ae819SStefano Zampini {
3117674ae819SStefano Zampini   PetscErrorCode ierr;
3118674ae819SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)(pc->data);
311958da7f69SStefano Zampini   PetscScalar    *array;
312012edc857SStefano Zampini   Vec            from,to;
3121674ae819SStefano Zampini 
3122674ae819SStefano Zampini   PetscFunctionBegin;
312312edc857SStefano Zampini   if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
312412edc857SStefano Zampini     from = pcbddc->coarse_vec;
312512edc857SStefano Zampini     to = pcbddc->vec1_P;
312612edc857SStefano Zampini   } else { /* from local to global -> put data in coarse right hand side */
312712edc857SStefano Zampini     from = pcbddc->vec1_P;
312812edc857SStefano Zampini     to = pcbddc->coarse_vec;
312912edc857SStefano Zampini   }
313012edc857SStefano Zampini   ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr);
313112edc857SStefano Zampini   if (smode == SCATTER_FORWARD) {
313212edc857SStefano Zampini     if (pcbddc->coarse_ksp) { /* get array from coarse processes */
313312edc857SStefano Zampini       Vec tvec;
313458da7f69SStefano Zampini 
313512edc857SStefano Zampini       ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
313658da7f69SStefano Zampini       ierr = VecGetArray(to,&array);CHKERRQ(ierr);
313758da7f69SStefano Zampini       ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr);
313858da7f69SStefano Zampini       ierr = VecRestoreArray(to,&array);CHKERRQ(ierr);
313958da7f69SStefano Zampini     }
314058da7f69SStefano Zampini   } else {
314158da7f69SStefano Zampini     if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */
314258da7f69SStefano Zampini      ierr = VecResetArray(from);CHKERRQ(ierr);
314312edc857SStefano Zampini     }
314412edc857SStefano Zampini   }
3145674ae819SStefano Zampini   PetscFunctionReturn(0);
3146674ae819SStefano Zampini }
3147674ae819SStefano Zampini 
3148984c4197SStefano Zampini /* uncomment for testing purposes */
3149984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */
3150674ae819SStefano Zampini #undef __FUNCT__
3151674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp"
3152674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc)
3153674ae819SStefano Zampini {
3154674ae819SStefano Zampini   PetscErrorCode    ierr;
3155674ae819SStefano Zampini   PC_IS*            pcis = (PC_IS*)(pc->data);
3156674ae819SStefano Zampini   PC_BDDC*          pcbddc = (PC_BDDC*)pc->data;
3157674ae819SStefano Zampini   Mat_IS*           matis = (Mat_IS*)pc->pmat->data;
3158984c4197SStefano Zampini   /* one and zero */
3159984c4197SStefano Zampini   PetscScalar       one=1.0,zero=0.0;
3160984c4197SStefano Zampini   /* space to store constraints and their local indices */
31619162d606SStefano Zampini   PetscScalar       *constraints_data;
31629162d606SStefano Zampini   PetscInt          *constraints_idxs,*constraints_idxs_B;
31639162d606SStefano Zampini   PetscInt          *constraints_idxs_ptr,*constraints_data_ptr;
31649162d606SStefano Zampini   PetscInt          *constraints_n;
3165984c4197SStefano Zampini   /* iterators */
3166b3d85658SStefano Zampini   PetscInt          i,j,k,total_counts,total_counts_cc,cum;
3167984c4197SStefano Zampini   /* BLAS integers */
3168e310c8b4SStefano Zampini   PetscBLASInt      lwork,lierr;
3169e310c8b4SStefano Zampini   PetscBLASInt      Blas_N,Blas_M,Blas_K,Blas_one=1;
3170c4303822SStefano Zampini   PetscBLASInt      Blas_LDA,Blas_LDB,Blas_LDC;
3171727cdba6SStefano Zampini   /* reuse */
31720e6343abSStefano Zampini   PetscInt          olocal_primal_size,olocal_primal_size_cc;
31730e6343abSStefano Zampini   PetscInt          *olocal_primal_ref_node,*olocal_primal_ref_mult;
3174984c4197SStefano Zampini   /* change of basis */
3175b3d85658SStefano Zampini   PetscBool         qr_needed;
31769162d606SStefano Zampini   PetscBT           change_basis,qr_needed_idx;
3177984c4197SStefano Zampini   /* auxiliary stuff */
317864efe560SStefano Zampini   PetscInt          *nnz,*is_indices;
31798a0068c3SStefano Zampini   PetscInt          ncc;
3180984c4197SStefano Zampini   /* some quantities */
318145a1bb75SStefano Zampini   PetscInt          n_vertices,total_primal_vertices,valid_constraints;
3182a58a30b4SStefano Zampini   PetscInt          size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints;
3183984c4197SStefano Zampini 
3184674ae819SStefano Zampini   PetscFunctionBegin;
31858e61c736SStefano Zampini   /* Destroy Mat objects computed previously */
31868e61c736SStefano Zampini   ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
31878e61c736SStefano Zampini   ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
3188088faed8SStefano Zampini   /* save info on constraints from previous setup (if any) */
3189088faed8SStefano Zampini   olocal_primal_size = pcbddc->local_primal_size;
31900e6343abSStefano Zampini   olocal_primal_size_cc = pcbddc->local_primal_size_cc;
31910e6343abSStefano Zampini   ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr);
31920e6343abSStefano Zampini   ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr);
31930e6343abSStefano Zampini   ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr);
31940e6343abSStefano Zampini   ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr);
3195088faed8SStefano Zampini   ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
3196cf5a6209SStefano Zampini 
3197cf5a6209SStefano Zampini   /* print some info */
3198cf5a6209SStefano Zampini   if (pcbddc->dbg_flag) {
3199cf5a6209SStefano Zampini     IS       vertices;
3200cf5a6209SStefano Zampini     PetscInt nv,nedges,nfaces;
3201cf5a6209SStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr);
3202cf5a6209SStefano Zampini     ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr);
3203cf5a6209SStefano Zampini     ierr = ISDestroy(&vertices);CHKERRQ(ierr);
32041575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
3205cf5a6209SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr);
3206cf5a6209SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr);
3207fd14bc51SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges    (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr);
3208fd14bc51SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces    (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr);
3209cf5a6209SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
32101575c14dSBarry Smith     ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
3211cf5a6209SStefano Zampini   }
3212cf5a6209SStefano Zampini 
3213cf5a6209SStefano Zampini   if (!pcbddc->adaptive_selection) {
32149162d606SStefano Zampini     IS           ISForVertices,*ISForFaces,*ISForEdges;
3215cf5a6209SStefano Zampini     MatNullSpace nearnullsp;
3216cf5a6209SStefano Zampini     const Vec    *nearnullvecs;
3217cf5a6209SStefano Zampini     Vec          *localnearnullsp;
3218cf5a6209SStefano Zampini     PetscScalar  *array;
3219cf5a6209SStefano Zampini     PetscInt     n_ISForFaces,n_ISForEdges,nnsp_size;
3220cf5a6209SStefano Zampini     PetscBool    nnsp_has_cnst;
3221674ae819SStefano Zampini     /* LAPACK working arrays for SVD or POD */
3222b3d85658SStefano Zampini     PetscBool    skip_lapack,boolforchange;
3223674ae819SStefano Zampini     PetscScalar  *work;
3224674ae819SStefano Zampini     PetscReal    *singular_vals;
3225674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX)
3226674ae819SStefano Zampini     PetscReal    *rwork;
3227674ae819SStefano Zampini #endif
3228674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
3229674ae819SStefano Zampini     PetscScalar  *temp_basis,*correlation_mat;
3230674ae819SStefano Zampini #else
3231964fefecSStefano Zampini     PetscBLASInt dummy_int=1;
3232964fefecSStefano Zampini     PetscScalar  dummy_scalar=1.;
3233674ae819SStefano Zampini #endif
3234674ae819SStefano Zampini 
3235674ae819SStefano Zampini     /* Get index sets for faces, edges and vertices from graph */
3236d06fc5fdSStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr);
3237d06fc5fdSStefano Zampini     /* free unneeded index sets */
3238d06fc5fdSStefano Zampini     if (!pcbddc->use_vertices) {
3239d06fc5fdSStefano Zampini       ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr);
3240674ae819SStefano Zampini     }
3241d06fc5fdSStefano Zampini     if (!pcbddc->use_edges) {
3242d06fc5fdSStefano Zampini       for (i=0;i<n_ISForEdges;i++) {
3243d06fc5fdSStefano Zampini         ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr);
3244d06fc5fdSStefano Zampini       }
3245d06fc5fdSStefano Zampini       ierr = PetscFree(ISForEdges);CHKERRQ(ierr);
3246d06fc5fdSStefano Zampini       n_ISForEdges = 0;
3247d06fc5fdSStefano Zampini     }
3248d06fc5fdSStefano Zampini     if (!pcbddc->use_faces) {
3249d06fc5fdSStefano Zampini       for (i=0;i<n_ISForFaces;i++) {
3250d06fc5fdSStefano Zampini         ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr);
3251d06fc5fdSStefano Zampini       }
3252d06fc5fdSStefano Zampini       ierr = PetscFree(ISForFaces);CHKERRQ(ierr);
3253d06fc5fdSStefano Zampini       n_ISForFaces = 0;
3254d06fc5fdSStefano Zampini     }
325570022509SStefano Zampini 
325670022509SStefano Zampini #if defined(PETSC_USE_DEBUG)
325770022509SStefano Zampini     /* HACK: when solving singular problems not using vertices, a change of basis is mandatory.
325870022509SStefano Zampini        Also use_change_of_basis should be consistent among processors */
325970022509SStefano Zampini     if (pcbddc->NullSpace) {
326070022509SStefano Zampini       PetscBool tbool[2],gbool[2];
326170022509SStefano Zampini 
326270022509SStefano Zampini       if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) {
3263b8ffe317SStefano Zampini         pcbddc->use_change_of_basis = PETSC_TRUE;
3264d06fc5fdSStefano Zampini         if (!ISForEdges) {
3265d06fc5fdSStefano Zampini           pcbddc->use_change_on_faces = PETSC_TRUE;
3266d06fc5fdSStefano Zampini         }
3267b8ffe317SStefano Zampini       }
3268d06fc5fdSStefano Zampini       tbool[0] = pcbddc->use_change_of_basis;
3269d06fc5fdSStefano Zampini       tbool[1] = pcbddc->use_change_on_faces;
3270d06fc5fdSStefano Zampini       ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
3271d06fc5fdSStefano Zampini       pcbddc->use_change_of_basis = gbool[0];
3272d06fc5fdSStefano Zampini       pcbddc->use_change_on_faces = gbool[1];
327398a51de6SStefano Zampini     }
327470022509SStefano Zampini #endif
327508122e43SStefano Zampini 
3276674ae819SStefano Zampini     /* check if near null space is attached to global mat */
3277674ae819SStefano Zampini     ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr);
3278674ae819SStefano Zampini     if (nearnullsp) {
3279674ae819SStefano Zampini       ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr);
3280f4ddd8eeSStefano Zampini       /* remove any stored info */
3281f4ddd8eeSStefano Zampini       ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr);
3282f4ddd8eeSStefano Zampini       ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr);
3283f4ddd8eeSStefano Zampini       /* store information for BDDC solver reuse */
3284f4ddd8eeSStefano Zampini       ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr);
3285f4ddd8eeSStefano Zampini       pcbddc->onearnullspace = nearnullsp;
3286473ba861SJed Brown       ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr);
3287f4ddd8eeSStefano Zampini       for (i=0;i<nnsp_size;i++) {
3288f4ddd8eeSStefano Zampini         ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr);
3289f4ddd8eeSStefano Zampini       }
3290984c4197SStefano Zampini     } else { /* if near null space is not provided BDDC uses constants by default */
3291984c4197SStefano Zampini       nnsp_size = 0;
3292674ae819SStefano Zampini       nnsp_has_cnst = PETSC_TRUE;
3293674ae819SStefano Zampini     }
3294984c4197SStefano Zampini     /* get max number of constraints on a single cc */
3295984c4197SStefano Zampini     max_constraints = nnsp_size;
3296984c4197SStefano Zampini     if (nnsp_has_cnst) max_constraints++;
3297984c4197SStefano Zampini 
3298674ae819SStefano Zampini     /*
3299674ae819SStefano Zampini          Evaluate maximum storage size needed by the procedure
33009162d606SStefano Zampini          - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]"
33019162d606SStefano Zampini          - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]"
33029162d606SStefano Zampini          There can be multiple constraints per connected component
3303674ae819SStefano Zampini                                                                                                                                                            */
3304674ae819SStefano Zampini     n_vertices = 0;
3305674ae819SStefano Zampini     if (ISForVertices) {
3306674ae819SStefano Zampini       ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr);
3307674ae819SStefano Zampini     }
33089162d606SStefano Zampini     ncc = n_vertices+n_ISForFaces+n_ISForEdges;
33099162d606SStefano Zampini     ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr);
33109162d606SStefano Zampini 
33119162d606SStefano Zampini     total_counts = n_ISForFaces+n_ISForEdges;
33129162d606SStefano Zampini     total_counts *= max_constraints;
3313674ae819SStefano Zampini     total_counts += n_vertices;
33144641a718SStefano Zampini     ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr);
33159162d606SStefano Zampini 
3316674ae819SStefano Zampini     total_counts = 0;
3317674ae819SStefano Zampini     max_size_of_constraint = 0;
3318674ae819SStefano Zampini     for (i=0;i<n_ISForEdges+n_ISForFaces;i++) {
33199162d606SStefano Zampini       IS used_is;
3320674ae819SStefano Zampini       if (i<n_ISForEdges) {
33219162d606SStefano Zampini         used_is = ISForEdges[i];
3322674ae819SStefano Zampini       } else {
33239162d606SStefano Zampini         used_is = ISForFaces[i-n_ISForEdges];
3324674ae819SStefano Zampini       }
33259162d606SStefano Zampini       ierr = ISGetSize(used_is,&j);CHKERRQ(ierr);
3326674ae819SStefano Zampini       total_counts += j;
3327674ae819SStefano Zampini       max_size_of_constraint = PetscMax(j,max_size_of_constraint);
3328674ae819SStefano Zampini     }
33299162d606SStefano Zampini     ierr = PetscMalloc3(total_counts*max_constraints+n_vertices,&constraints_data,total_counts+n_vertices,&constraints_idxs,total_counts+n_vertices,&constraints_idxs_B);CHKERRQ(ierr);
33309162d606SStefano Zampini 
3331984c4197SStefano Zampini     /* get local part of global near null space vectors */
3332785e854fSJed Brown     ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr);
3333984c4197SStefano Zampini     for (k=0;k<nnsp_size;k++) {
3334984c4197SStefano Zampini       ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr);
3335e176bc59SStefano Zampini       ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
3336e176bc59SStefano Zampini       ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
3337984c4197SStefano Zampini     }
3338674ae819SStefano Zampini 
3339242a89d7SStefano Zampini     /* whether or not to skip lapack calls */
3340242a89d7SStefano Zampini     skip_lapack = PETSC_TRUE;
3341a773dcb8SStefano Zampini     if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE;
3342242a89d7SStefano Zampini 
3343984c4197SStefano Zampini     /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */
3344a773dcb8SStefano Zampini     if (!skip_lapack) {
3345674ae819SStefano Zampini       PetscScalar temp_work;
3346911cabfeSStefano Zampini 
3347674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
3348984c4197SStefano Zampini       /* Proper Orthogonal Decomposition (POD) using the snapshot method */
3349785e854fSJed Brown       ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr);
3350785e854fSJed Brown       ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr);
3351785e854fSJed Brown       ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr);
3352674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX)
3353785e854fSJed Brown       ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr);
3354674ae819SStefano Zampini #endif
3355674ae819SStefano Zampini       /* now we evaluate the optimal workspace using query with lwork=-1 */
3356c8244a33SStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr);
3357c8244a33SStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr);
3358674ae819SStefano Zampini       lwork = -1;
3359674ae819SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
3360674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
3361c8244a33SStefano Zampini       PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr));
3362674ae819SStefano Zampini #else
3363c8244a33SStefano Zampini       PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr));
3364674ae819SStefano Zampini #endif
3365674ae819SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
3366984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr);
3367674ae819SStefano Zampini #else /* on missing GESVD */
3368674ae819SStefano Zampini       /* SVD */
3369674ae819SStefano Zampini       PetscInt max_n,min_n;
3370674ae819SStefano Zampini       max_n = max_size_of_constraint;
3371984c4197SStefano Zampini       min_n = max_constraints;
3372984c4197SStefano Zampini       if (max_size_of_constraint < max_constraints) {
3373674ae819SStefano Zampini         min_n = max_size_of_constraint;
3374984c4197SStefano Zampini         max_n = max_constraints;
3375674ae819SStefano Zampini       }
3376785e854fSJed Brown       ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr);
3377674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX)
3378785e854fSJed Brown       ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr);
3379674ae819SStefano Zampini #endif
3380674ae819SStefano Zampini       /* now we evaluate the optimal workspace using query with lwork=-1 */
3381674ae819SStefano Zampini       lwork = -1;
3382e310c8b4SStefano Zampini       ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr);
3383e310c8b4SStefano Zampini       ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr);
3384b7d8b9f8SStefano Zampini       ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr);
3385674ae819SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
3386674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
33879162d606SStefano Zampini       PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr));
3388674ae819SStefano Zampini #else
33899162d606SStefano Zampini       PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr));
3390674ae819SStefano Zampini #endif
3391674ae819SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
3392984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr);
3393984c4197SStefano Zampini #endif /* on missing GESVD */
3394674ae819SStefano Zampini       /* Allocate optimal workspace */
3395674ae819SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr);
3396854ce69bSBarry Smith       ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr);
3397674ae819SStefano Zampini     }
3398674ae819SStefano Zampini     /* Now we can loop on constraining sets */
3399674ae819SStefano Zampini     total_counts = 0;
34009162d606SStefano Zampini     constraints_idxs_ptr[0] = 0;
34019162d606SStefano Zampini     constraints_data_ptr[0] = 0;
3402674ae819SStefano Zampini     /* vertices */
34039162d606SStefano Zampini     if (n_vertices) {
3404674ae819SStefano Zampini       ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
34059162d606SStefano Zampini       ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr);
3406674ae819SStefano Zampini       for (i=0;i<n_vertices;i++) {
34079162d606SStefano Zampini         constraints_n[total_counts] = 1;
34089162d606SStefano Zampini         constraints_data[total_counts] = 1.0;
34099162d606SStefano Zampini         constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1;
34109162d606SStefano Zampini         constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1;
3411674ae819SStefano Zampini         total_counts++;
3412674ae819SStefano Zampini       }
3413674ae819SStefano Zampini       ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
3414674ae819SStefano Zampini       n_vertices = total_counts;
3415674ae819SStefano Zampini     }
3416984c4197SStefano Zampini 
3417674ae819SStefano Zampini     /* edges and faces */
34189162d606SStefano Zampini     total_counts_cc = total_counts;
3419911cabfeSStefano Zampini     for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) {
34209162d606SStefano Zampini       IS        used_is;
34219162d606SStefano Zampini       PetscBool idxs_copied = PETSC_FALSE;
34229162d606SStefano Zampini 
3423911cabfeSStefano Zampini       if (ncc<n_ISForEdges) {
34249162d606SStefano Zampini         used_is = ISForEdges[ncc];
3425984c4197SStefano Zampini         boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */
3426674ae819SStefano Zampini       } else {
34279162d606SStefano Zampini         used_is = ISForFaces[ncc-n_ISForEdges];
3428984c4197SStefano Zampini         boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */
3429674ae819SStefano Zampini       }
3430674ae819SStefano Zampini       temp_constraints = 0;          /* zero the number of constraints I have on this conn comp */
34319162d606SStefano Zampini 
34329162d606SStefano Zampini       ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr);
34339162d606SStefano Zampini       ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
3434984c4197SStefano Zampini       /* change of basis should not be performed on local periodic nodes */
3435984c4197SStefano Zampini       if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE;
3436674ae819SStefano Zampini       if (nnsp_has_cnst) {
34375b08dc53SStefano Zampini         PetscScalar quad_value;
34389162d606SStefano Zampini 
34399162d606SStefano Zampini         ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
34409162d606SStefano Zampini         idxs_copied = PETSC_TRUE;
34419162d606SStefano Zampini 
3442a773dcb8SStefano Zampini         if (!pcbddc->use_nnsp_true) {
3443674ae819SStefano Zampini           quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint));
3444a773dcb8SStefano Zampini         } else {
3445a773dcb8SStefano Zampini           quad_value = 1.0;
3446a773dcb8SStefano Zampini         }
3447674ae819SStefano Zampini         for (j=0;j<size_of_constraint;j++) {
34489162d606SStefano Zampini           constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value;
3449674ae819SStefano Zampini         }
34509162d606SStefano Zampini         temp_constraints++;
3451674ae819SStefano Zampini         total_counts++;
3452674ae819SStefano Zampini       }
3453674ae819SStefano Zampini       for (k=0;k<nnsp_size;k++) {
3454984c4197SStefano Zampini         PetscReal real_value;
34559162d606SStefano Zampini         PetscScalar *ptr_to_data;
34569162d606SStefano Zampini 
3457984c4197SStefano Zampini         ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
34589162d606SStefano Zampini         ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint];
3459674ae819SStefano Zampini         for (j=0;j<size_of_constraint;j++) {
34609162d606SStefano Zampini           ptr_to_data[j] = array[is_indices[j]];
3461674ae819SStefano Zampini         }
3462984c4197SStefano Zampini         ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
3463984c4197SStefano Zampini         /* check if array is null on the connected component */
3464e310c8b4SStefano Zampini         ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
34659162d606SStefano Zampini         PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one));
34665b08dc53SStefano Zampini         if (real_value > 0.0) { /* keep indices and values */
3467674ae819SStefano Zampini           temp_constraints++;
3468674ae819SStefano Zampini           total_counts++;
34699162d606SStefano Zampini           if (!idxs_copied) {
34709162d606SStefano Zampini             ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
34719162d606SStefano Zampini             idxs_copied = PETSC_TRUE;
3472674ae819SStefano Zampini           }
3473674ae819SStefano Zampini         }
34749162d606SStefano Zampini       }
34759162d606SStefano Zampini       ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr);
347645a1bb75SStefano Zampini       valid_constraints = temp_constraints;
3477eb97c9d2SStefano Zampini       if (!pcbddc->use_nnsp_true && temp_constraints) {
3478a773dcb8SStefano Zampini         if (temp_constraints == 1) { /* just normalize the constraint */
34799162d606SStefano Zampini           PetscScalar norm,*ptr_to_data;
34809162d606SStefano Zampini 
34819162d606SStefano Zampini           ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
3482a773dcb8SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
34839162d606SStefano Zampini           PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one));
3484a773dcb8SStefano Zampini           norm = 1.0/PetscSqrtReal(PetscRealPart(norm));
34859162d606SStefano Zampini           PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one));
3486a773dcb8SStefano Zampini         } else { /* perform SVD */
3487984c4197SStefano Zampini           PetscReal   tol = 1.0e-8; /* tolerance for retaining eigenmodes */
34889162d606SStefano Zampini           PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
3489674ae819SStefano Zampini 
3490674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
3491984c4197SStefano Zampini           /* SVD: Y = U*S*V^H                -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag
3492984c4197SStefano Zampini              POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2)
3493984c4197SStefano Zampini              -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined
3494984c4197SStefano Zampini                 the constraints basis will differ (by a complex factor with absolute value equal to 1)
3495984c4197SStefano Zampini                 from that computed using LAPACKgesvd
3496984c4197SStefano Zampini              -> This is due to a different computation of eigenvectors in LAPACKheev
3497984c4197SStefano Zampini              -> The quality of the POD-computed basis will be the same */
3498984c4197SStefano Zampini           ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr);
3499674ae819SStefano Zampini           /* Store upper triangular part of correlation matrix */
3500e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
3501984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
3502674ae819SStefano Zampini           for (j=0;j<temp_constraints;j++) {
3503674ae819SStefano Zampini             for (k=0;k<j+1;k++) {
35049162d606SStefano Zampini               PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k] = BLASdot_(&Blas_N,ptr_to_data+k*size_of_constraint,&Blas_one,ptr_to_data+j*size_of_constraint,&Blas_one));
3505674ae819SStefano Zampini             }
3506674ae819SStefano Zampini           }
3507e310c8b4SStefano Zampini           /* compute eigenvalues and eigenvectors of correlation matrix */
3508e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
3509e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr);
3510674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
3511c8244a33SStefano Zampini           PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr));
3512674ae819SStefano Zampini #else
3513c8244a33SStefano Zampini           PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr));
3514674ae819SStefano Zampini #endif
3515674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
3516984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr);
3517984c4197SStefano Zampini           /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */
3518674ae819SStefano Zampini           j = 0;
3519984c4197SStefano Zampini           while (j < temp_constraints && singular_vals[j] < tol) j++;
3520674ae819SStefano Zampini           total_counts = total_counts-j;
352145a1bb75SStefano Zampini           valid_constraints = temp_constraints-j;
3522e310c8b4SStefano Zampini           /* scale and copy POD basis into used quadrature memory */
3523c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
3524c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
3525c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr);
3526c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3527c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr);
3528c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr);
3529674ae819SStefano Zampini           if (j<temp_constraints) {
3530984c4197SStefano Zampini             PetscInt ii;
3531984c4197SStefano Zampini             for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]);
3532674ae819SStefano Zampini             ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
35339162d606SStefano Zampini             PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,ptr_to_data,&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC));
3534674ae819SStefano Zampini             ierr = PetscFPTrapPop();CHKERRQ(ierr);
3535984c4197SStefano Zampini             for (k=0;k<temp_constraints-j;k++) {
3536674ae819SStefano Zampini               for (ii=0;ii<size_of_constraint;ii++) {
35379162d606SStefano Zampini                 ptr_to_data[k*size_of_constraint+ii] = singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii];
3538674ae819SStefano Zampini               }
3539674ae819SStefano Zampini             }
3540674ae819SStefano Zampini           }
3541674ae819SStefano Zampini #else  /* on missing GESVD */
3542e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
3543e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
3544b7d8b9f8SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3545674ae819SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
3546674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
35479162d606SStefano Zampini           PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr));
3548674ae819SStefano Zampini #else
35499162d606SStefano Zampini           PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr));
3550674ae819SStefano Zampini #endif
3551984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr);
3552674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
3553984c4197SStefano Zampini           /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */
3554e310c8b4SStefano Zampini           k = temp_constraints;
3555e310c8b4SStefano Zampini           if (k > size_of_constraint) k = size_of_constraint;
3556674ae819SStefano Zampini           j = 0;
3557e310c8b4SStefano Zampini           while (j < k && singular_vals[k-j-1] < tol) j++;
355845a1bb75SStefano Zampini           valid_constraints = k-j;
3559911cabfeSStefano Zampini           total_counts = total_counts-temp_constraints+valid_constraints;
3560984c4197SStefano Zampini #endif /* on missing GESVD */
3561674ae819SStefano Zampini         }
3562a773dcb8SStefano Zampini       }
35639162d606SStefano Zampini       /* update pointers information */
35649162d606SStefano Zampini       if (valid_constraints) {
35659162d606SStefano Zampini         constraints_n[total_counts_cc] = valid_constraints;
35669162d606SStefano Zampini         constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint;
35679162d606SStefano Zampini         constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints;
35689162d606SStefano Zampini         /* set change_of_basis flag */
356945a1bb75SStefano Zampini         if (boolforchange) {
3570b3d85658SStefano Zampini           PetscBTSet(change_basis,total_counts_cc);
35719162d606SStefano Zampini         }
3572b3d85658SStefano Zampini         total_counts_cc++;
357345a1bb75SStefano Zampini       }
357445a1bb75SStefano Zampini     }
3575984c4197SStefano Zampini     /* free workspace */
35768f1c130eSStefano Zampini     if (!skip_lapack) {
3577984c4197SStefano Zampini       ierr = PetscFree(work);CHKERRQ(ierr);
3578984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX)
3579984c4197SStefano Zampini       ierr = PetscFree(rwork);CHKERRQ(ierr);
3580984c4197SStefano Zampini #endif
3581984c4197SStefano Zampini       ierr = PetscFree(singular_vals);CHKERRQ(ierr);
3582984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
3583984c4197SStefano Zampini       ierr = PetscFree(correlation_mat);CHKERRQ(ierr);
3584984c4197SStefano Zampini       ierr = PetscFree(temp_basis);CHKERRQ(ierr);
3585984c4197SStefano Zampini #endif
3586984c4197SStefano Zampini     }
3587984c4197SStefano Zampini     for (k=0;k<nnsp_size;k++) {
3588984c4197SStefano Zampini       ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr);
3589984c4197SStefano Zampini     }
3590984c4197SStefano Zampini     ierr = PetscFree(localnearnullsp);CHKERRQ(ierr);
3591cf5a6209SStefano Zampini     /* free index sets of faces, edges and vertices */
3592cf5a6209SStefano Zampini     for (i=0;i<n_ISForFaces;i++) {
3593cf5a6209SStefano Zampini       ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr);
3594cf5a6209SStefano Zampini     }
3595cf5a6209SStefano Zampini     if (n_ISForFaces) {
3596cf5a6209SStefano Zampini       ierr = PetscFree(ISForFaces);CHKERRQ(ierr);
3597cf5a6209SStefano Zampini     }
3598cf5a6209SStefano Zampini     for (i=0;i<n_ISForEdges;i++) {
3599cf5a6209SStefano Zampini       ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr);
3600cf5a6209SStefano Zampini     }
3601cf5a6209SStefano Zampini     if (n_ISForEdges) {
3602cf5a6209SStefano Zampini       ierr = PetscFree(ISForEdges);CHKERRQ(ierr);
3603cf5a6209SStefano Zampini     }
3604cf5a6209SStefano Zampini     ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr);
360508122e43SStefano Zampini   } else {
360608122e43SStefano Zampini     PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
3607984c4197SStefano Zampini 
360808122e43SStefano Zampini     total_counts = 0;
360908122e43SStefano Zampini     n_vertices = 0;
3610d62866d3SStefano Zampini     if (sub_schurs->is_vertices && pcbddc->use_vertices) {
3611d62866d3SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr);
361208122e43SStefano Zampini     }
361308122e43SStefano Zampini     max_constraints = 0;
36149162d606SStefano Zampini     total_counts_cc = 0;
361508122e43SStefano Zampini     for (i=0;i<sub_schurs->n_subs+n_vertices;i++) {
361608122e43SStefano Zampini       total_counts += pcbddc->adaptive_constraints_n[i];
36179162d606SStefano Zampini       if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++;
361808122e43SStefano Zampini       max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]);
361908122e43SStefano Zampini     }
36209162d606SStefano Zampini     constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr;
36219162d606SStefano Zampini     constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr;
36229162d606SStefano Zampini     constraints_idxs = pcbddc->adaptive_constraints_idxs;
36239162d606SStefano Zampini     constraints_data = pcbddc->adaptive_constraints_data;
362474d5cdf7SStefano Zampini     /* constraints_n differs from pcbddc->adaptive_constraints_n */
36259162d606SStefano Zampini     ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr);
36269162d606SStefano Zampini     total_counts_cc = 0;
36279162d606SStefano Zampini     for (i=0;i<sub_schurs->n_subs+n_vertices;i++) {
36289162d606SStefano Zampini       if (pcbddc->adaptive_constraints_n[i]) {
36299162d606SStefano Zampini         constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i];
363008122e43SStefano Zampini       }
363108122e43SStefano Zampini     }
36329162d606SStefano Zampini #if 0
36339162d606SStefano Zampini     printf("Found %d totals (%d)\n",total_counts_cc,total_counts);
36349162d606SStefano Zampini     for (i=0;i<total_counts_cc;i++) {
36359162d606SStefano Zampini       printf("const %d, start %d",i,constraints_idxs_ptr[i]);
36369162d606SStefano Zampini       printf(" end %d:\n",constraints_idxs_ptr[i+1]);
36379162d606SStefano Zampini       for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) {
36389162d606SStefano Zampini         printf(" %d",constraints_idxs[j]);
36399162d606SStefano Zampini       }
36409162d606SStefano Zampini       printf("\n");
36419162d606SStefano Zampini       printf("number of cc: %d\n",constraints_n[i]);
36429162d606SStefano Zampini     }
36431b968477SStefano Zampini     for (i=0;i<n_vertices;i++) {
36448bec7fa6SStefano Zampini       PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]);
36451b968477SStefano Zampini     }
36461b968477SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
36478bec7fa6SStefano Zampini       PetscPrintf(PETSC_COMM_SELF,"[%d] sub %d, edge %d, n %d\n",PetscGlobalRank,i,(PetscBool)PetscBTLookup(sub_schurs->is_edge,i),pcbddc->adaptive_constraints_n[i+n_vertices]);
36481b968477SStefano Zampini     }
364908122e43SStefano Zampini #endif
365008122e43SStefano Zampini 
36518bec7fa6SStefano Zampini     max_size_of_constraint = 0;
36529162d606SStefano Zampini     for (i=0;i<total_counts_cc;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]);
36539162d606SStefano Zampini     ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr);
365408122e43SStefano Zampini     /* Change of basis */
3655b3d85658SStefano Zampini     ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr);
365608122e43SStefano Zampini     if (pcbddc->use_change_of_basis) {
365708122e43SStefano Zampini       for (i=0;i<sub_schurs->n_subs;i++) {
365808122e43SStefano Zampini         if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) {
3659b3d85658SStefano Zampini           ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr);
366008122e43SStefano Zampini         }
366108122e43SStefano Zampini       }
366208122e43SStefano Zampini     }
366308122e43SStefano Zampini   }
3664984c4197SStefano Zampini   pcbddc->local_primal_size = total_counts;
36654f1b2e48SStefano Zampini   ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
366608122e43SStefano Zampini 
36679162d606SStefano Zampini   /* map constraints_idxs in boundary numbering */
36689162d606SStefano Zampini   ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr);
36699162d606SStefano Zampini   if (i != constraints_idxs_ptr[total_counts_cc]) {
36709162d606SStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",constraints_idxs_ptr[total_counts_cc],i);
367108122e43SStefano Zampini   }
3672674ae819SStefano Zampini 
3673674ae819SStefano Zampini   /* Create constraint matrix */
3674674ae819SStefano Zampini   ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
367516f15bc4SStefano Zampini   ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr);
3676984c4197SStefano Zampini   ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr);
3677984c4197SStefano Zampini 
3678984c4197SStefano Zampini   /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */
3679a717540cSStefano Zampini   /* determine if a QR strategy is needed for change of basis */
3680a717540cSStefano Zampini   qr_needed = PETSC_FALSE;
368174d5cdf7SStefano Zampini   ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr);
3682984c4197SStefano Zampini   total_primal_vertices=0;
3683b3d85658SStefano Zampini   pcbddc->local_primal_size_cc = 0;
36849162d606SStefano Zampini   for (i=0;i<total_counts_cc;i++) {
36859162d606SStefano Zampini     size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
3686984c4197SStefano Zampini     if (size_of_constraint == 1) {
36879162d606SStefano Zampini       pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]];
3688b3d85658SStefano Zampini       pcbddc->local_primal_size_cc += 1;
368964efe560SStefano Zampini     } else if (PetscBTLookup(change_basis,i)) {
36909162d606SStefano Zampini       for (k=0;k<constraints_n[i];k++) {
36919162d606SStefano Zampini         pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k];
3692a717540cSStefano Zampini       }
3693b3d85658SStefano Zampini       pcbddc->local_primal_size_cc += constraints_n[i];
369474d5cdf7SStefano Zampini       if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) {
3695a717540cSStefano Zampini         PetscBTSet(qr_needed_idx,i);
3696a717540cSStefano Zampini         qr_needed = PETSC_TRUE;
3697a717540cSStefano Zampini       }
3698fa434743SStefano Zampini     } else {
3699b3d85658SStefano Zampini       pcbddc->local_primal_size_cc += 1;
3700fa434743SStefano Zampini     }
3701a717540cSStefano Zampini   }
3702b371cd4fSStefano Zampini   /* note that the local variable n_vertices used below stores the number of pointwise constraints */
3703b371cd4fSStefano Zampini   pcbddc->n_vertices = total_primal_vertices;
3704674ae819SStefano Zampini   /* permute indices in order to have a sorted set of vertices */
370570022509SStefano Zampini   ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
3706b3d85658SStefano Zampini 
37074f1b2e48SStefano Zampini   ierr = PetscMalloc2(pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr);
37080e6343abSStefano Zampini   ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr);
37090e6343abSStefano Zampini   for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1;
3710984c4197SStefano Zampini 
3711984c4197SStefano Zampini   /* nonzero structure of constraint matrix */
371274d5cdf7SStefano Zampini   /* and get reference dof for local constraints */
3713785e854fSJed Brown   ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr);
3714984c4197SStefano Zampini   for (i=0;i<total_primal_vertices;i++) nnz[i] = 1;
371574d5cdf7SStefano Zampini 
3716984c4197SStefano Zampini   j = total_primal_vertices;
371774d5cdf7SStefano Zampini   total_counts = total_primal_vertices;
3718b3d85658SStefano Zampini   cum = total_primal_vertices;
37199162d606SStefano Zampini   for (i=n_vertices;i<total_counts_cc;i++) {
37204641a718SStefano Zampini     if (!PetscBTLookup(change_basis,i)) {
3721b3d85658SStefano Zampini       pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]];
3722b3d85658SStefano Zampini       pcbddc->local_primal_ref_mult[cum] = constraints_n[i];
3723b3d85658SStefano Zampini       cum++;
37249162d606SStefano Zampini       size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
372574d5cdf7SStefano Zampini       for (k=0;k<constraints_n[i];k++) {
372674d5cdf7SStefano Zampini         pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k];
372774d5cdf7SStefano Zampini         nnz[j+k] = size_of_constraint;
372874d5cdf7SStefano Zampini       }
37299162d606SStefano Zampini       j += constraints_n[i];
3730674ae819SStefano Zampini     }
3731674ae819SStefano Zampini   }
3732674ae819SStefano Zampini   ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr);
3733674ae819SStefano Zampini   ierr = PetscFree(nnz);CHKERRQ(ierr);
3734088faed8SStefano Zampini 
3735674ae819SStefano Zampini   /* set values in constraint matrix */
3736984c4197SStefano Zampini   for (i=0;i<total_primal_vertices;i++) {
37370e6343abSStefano Zampini     ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr);
3738674ae819SStefano Zampini   }
3739984c4197SStefano Zampini   total_counts = total_primal_vertices;
37409162d606SStefano Zampini   for (i=n_vertices;i<total_counts_cc;i++) {
37414641a718SStefano Zampini     if (!PetscBTLookup(change_basis,i)) {
37429162d606SStefano Zampini       PetscInt *cols;
37439162d606SStefano Zampini 
37449162d606SStefano Zampini       size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
37459162d606SStefano Zampini       cols = constraints_idxs+constraints_idxs_ptr[i];
37469162d606SStefano Zampini       for (k=0;k<constraints_n[i];k++) {
37479162d606SStefano Zampini         PetscInt    row = total_counts+k;
37489162d606SStefano Zampini         PetscScalar *vals;
37499162d606SStefano Zampini 
37509162d606SStefano Zampini         vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint;
37519162d606SStefano Zampini         ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr);
37529162d606SStefano Zampini       }
37539162d606SStefano Zampini       total_counts += constraints_n[i];
3754674ae819SStefano Zampini     }
3755674ae819SStefano Zampini   }
3756674ae819SStefano Zampini   /* assembling */
3757674ae819SStefano Zampini   ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
3758674ae819SStefano Zampini   ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
3759088faed8SStefano Zampini 
3760984c4197SStefano Zampini   /*
376145a1bb75SStefano Zampini   ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
3762984c4197SStefano Zampini   ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr);
3763984c4197SStefano Zampini   */
3764674ae819SStefano Zampini   /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */
3765674ae819SStefano Zampini   if (pcbddc->use_change_of_basis) {
3766026de310SStefano Zampini     /* dual and primal dofs on a single cc */
3767984c4197SStefano Zampini     PetscInt     dual_dofs,primal_dofs;
3768984c4197SStefano Zampini     /* working stuff for GEQRF */
376981d9aea3SBarry Smith     PetscScalar  *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t;
3770984c4197SStefano Zampini     PetscBLASInt lqr_work;
3771984c4197SStefano Zampini     /* working stuff for UNGQR */
3772984c4197SStefano Zampini     PetscScalar  *gqr_work,lgqr_work_t;
3773984c4197SStefano Zampini     PetscBLASInt lgqr_work;
3774984c4197SStefano Zampini     /* working stuff for TRTRS */
3775984c4197SStefano Zampini     PetscScalar  *trs_rhs;
37763f08241aSStefano Zampini     PetscBLASInt Blas_NRHS;
3777984c4197SStefano Zampini     /* pointers for values insertion into change of basis matrix */
3778984c4197SStefano Zampini     PetscInt     *start_rows,*start_cols;
3779984c4197SStefano Zampini     PetscScalar  *start_vals;
3780984c4197SStefano Zampini     /* working stuff for values insertion */
37814641a718SStefano Zampini     PetscBT      is_primal;
378264efe560SStefano Zampini     PetscInt     *aux_primal_numbering_B;
3783906d46d4SStefano Zampini     /* matrix sizes */
3784906d46d4SStefano Zampini     PetscInt     global_size,local_size;
3785906d46d4SStefano Zampini     /* temporary change of basis */
3786906d46d4SStefano Zampini     Mat          localChangeOfBasisMatrix;
3787cf5a6209SStefano Zampini     /* extra space for debugging */
3788cf5a6209SStefano Zampini     PetscScalar  *dbg_work;
3789984c4197SStefano Zampini 
3790906d46d4SStefano Zampini     /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */
3791906d46d4SStefano Zampini     ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr);
379216f15bc4SStefano Zampini     ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr);
3793bbb9e6c6SStefano Zampini     ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr);
3794906d46d4SStefano Zampini     /* nonzeros for local mat */
3795bbb9e6c6SStefano Zampini     ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr);
3796bbb9e6c6SStefano Zampini     for (i=0;i<pcis->n;i++) nnz[i]=1;
37979162d606SStefano Zampini     for (i=n_vertices;i<total_counts_cc;i++) {
3798a717540cSStefano Zampini       if (PetscBTLookup(change_basis,i)) {
37999162d606SStefano Zampini         size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
3800a717540cSStefano Zampini         if (PetscBTLookup(qr_needed_idx,i)) {
38019162d606SStefano Zampini           for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint;
3802a717540cSStefano Zampini         } else {
38039162d606SStefano Zampini           nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint;
38049162d606SStefano Zampini           for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2;
3805a717540cSStefano Zampini         }
3806a717540cSStefano Zampini       }
3807a717540cSStefano Zampini     }
3808906d46d4SStefano Zampini     ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr);
3809bbb9e6c6SStefano Zampini     ierr = PetscFree(nnz);CHKERRQ(ierr);
3810a717540cSStefano Zampini     /* Set initial identity in the matrix */
3811bbb9e6c6SStefano Zampini     for (i=0;i<pcis->n;i++) {
3812906d46d4SStefano Zampini       ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr);
3813a717540cSStefano Zampini     }
3814a717540cSStefano Zampini 
3815a717540cSStefano Zampini     if (pcbddc->dbg_flag) {
3816a717540cSStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr);
3817a717540cSStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr);
3818a717540cSStefano Zampini     }
3819a717540cSStefano Zampini 
3820a717540cSStefano Zampini 
3821a717540cSStefano Zampini     /* Now we loop on the constraints which need a change of basis */
3822a717540cSStefano Zampini     /*
3823a717540cSStefano Zampini        Change of basis matrix is evaluated similarly to the FIRST APPROACH in
3824a717540cSStefano Zampini        Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1)
3825a717540cSStefano Zampini 
3826a6b551f4SStefano Zampini        Basic blocks of change of basis matrix T computed by
3827a717540cSStefano Zampini 
3828a6b551f4SStefano Zampini           - Using the following block transformation if there is only a primal dof on the cc (and -pc_bddc_use_qr_single is not specified)
3829a6b551f4SStefano Zampini 
3830a6b551f4SStefano Zampini             | 1        0   ...        0         s_1/S |
3831a6b551f4SStefano Zampini             | 0        1   ...        0         s_2/S |
3832a717540cSStefano Zampini             |              ...                        |
3833a6b551f4SStefano Zampini             | 0        ...            1     s_{n-1}/S |
3834a6b551f4SStefano Zampini             | -s_1/s_n ...    -s_{n-1}/s_n      s_n/S |
3835a717540cSStefano Zampini 
3836a6b551f4SStefano Zampini             with S = \sum_{i=1}^n s_i^2
3837a6b551f4SStefano Zampini             NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering
3838a6b551f4SStefano Zampini                   in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering
3839a6b551f4SStefano Zampini 
3840a6b551f4SStefano Zampini           - QR decomposition of constraints otherwise
3841a717540cSStefano Zampini     */
3842a717540cSStefano Zampini     if (qr_needed) {
3843984c4197SStefano Zampini       /* space to store Q */
3844854ce69bSBarry Smith       ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr);
3845984c4197SStefano Zampini       /* first we issue queries for optimal work */
38463f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr);
38473f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr);
38483f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3849984c4197SStefano Zampini       lqr_work = -1;
38503f08241aSStefano Zampini       PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr));
3851984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr);
3852984c4197SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr);
3853785e854fSJed Brown       ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr);
3854984c4197SStefano Zampini       lgqr_work = -1;
38553f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr);
38563f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr);
38573f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr);
38583f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
38593f08241aSStefano Zampini       if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */
38603f08241aSStefano Zampini       PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr));
3861984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr);
3862984c4197SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr);
3863785e854fSJed Brown       ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr);
3864984c4197SStefano Zampini       /* array to store scaling factors for reflectors */
3865785e854fSJed Brown       ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr);
3866984c4197SStefano Zampini       /* array to store rhs and solution of triangular solver */
3867785e854fSJed Brown       ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr);
3868a717540cSStefano Zampini       /* allocating workspace for check */
3869a717540cSStefano Zampini       if (pcbddc->dbg_flag) {
3870cf5a6209SStefano Zampini         ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr);
3871a717540cSStefano Zampini       }
3872a717540cSStefano Zampini     }
3873984c4197SStefano Zampini     /* array to store whether a node is primal or not */
38744641a718SStefano Zampini     ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr);
3875473ba861SJed Brown     ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr);
38760e6343abSStefano Zampini     ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr);
387739e2fb2aSStefano Zampini     if (i != total_primal_vertices) {
387839e2fb2aSStefano Zampini       SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i);
38794641a718SStefano Zampini     }
388039e2fb2aSStefano Zampini     for (i=0;i<total_primal_vertices;i++) {
388139e2fb2aSStefano Zampini       ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr);
388239e2fb2aSStefano Zampini     }
388339e2fb2aSStefano Zampini     ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr);
3884984c4197SStefano Zampini 
3885a717540cSStefano Zampini     /* loop on constraints and see whether or not they need a change of basis and compute it */
38869162d606SStefano Zampini     for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) {
38879162d606SStefano Zampini       size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts];
38884641a718SStefano Zampini       if (PetscBTLookup(change_basis,total_counts)) {
3889984c4197SStefano Zampini         /* get constraint info */
38909162d606SStefano Zampini         primal_dofs = constraints_n[total_counts];
3891984c4197SStefano Zampini         dual_dofs = size_of_constraint-primal_dofs;
3892984c4197SStefano Zampini 
3893984c4197SStefano Zampini         if (pcbddc->dbg_flag) {
38949162d606SStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d: %d need a change of basis (size %d)\n",total_counts,primal_dofs,size_of_constraint);CHKERRQ(ierr);
3895674ae819SStefano Zampini         }
3896984c4197SStefano Zampini 
3897fa434743SStefano Zampini         if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */
3898a717540cSStefano Zampini 
3899a717540cSStefano Zampini           /* copy quadrature constraints for change of basis check */
3900a717540cSStefano Zampini           if (pcbddc->dbg_flag) {
39019162d606SStefano Zampini             ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
3902a717540cSStefano Zampini           }
3903984c4197SStefano Zampini           /* copy temporary constraints into larger work vector (in order to store all columns of Q) */
39049162d606SStefano Zampini           ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
3905984c4197SStefano Zampini 
3906984c4197SStefano Zampini           /* compute QR decomposition of constraints */
39073f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
39083f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
39093f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3910674ae819SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
39113f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr));
3912984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr);
3913674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
3914984c4197SStefano Zampini 
3915984c4197SStefano Zampini           /* explictly compute R^-T */
3916984c4197SStefano Zampini           ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr);
3917984c4197SStefano Zampini           for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0;
39183f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
39193f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr);
39203f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
39213f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr);
3922984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
39233f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr));
3924984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr);
3925984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
3926984c4197SStefano Zampini 
3927a717540cSStefano Zampini           /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */
39283f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
39293f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
39303f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr);
39313f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3932984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
39333f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr));
3934984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr);
3935984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
3936984c4197SStefano Zampini 
3937984c4197SStefano Zampini           /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints
3938984c4197SStefano Zampini              i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below)
3939984c4197SStefano Zampini              where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */
39403f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
39413f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
39423f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr);
39433f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
39443f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr);
39453f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr);
3946984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
39479162d606SStefano Zampini           PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,constraints_data+constraints_data_ptr[total_counts],&Blas_LDC));
3948984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
39499162d606SStefano Zampini           ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
3950984c4197SStefano Zampini 
3951984c4197SStefano Zampini           /* insert values in change of basis matrix respecting global ordering of new primal dofs */
39529162d606SStefano Zampini           start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]];
3953984c4197SStefano Zampini           /* insert cols for primal dofs */
3954984c4197SStefano Zampini           for (j=0;j<primal_dofs;j++) {
3955984c4197SStefano Zampini             start_vals = &qr_basis[j*size_of_constraint];
39569162d606SStefano Zampini             start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j];
3957906d46d4SStefano Zampini             ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr);
3958984c4197SStefano Zampini           }
3959984c4197SStefano Zampini           /* insert cols for dual dofs */
3960984c4197SStefano Zampini           for (j=0,k=0;j<dual_dofs;k++) {
39619162d606SStefano Zampini             if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) {
3962984c4197SStefano Zampini               start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint];
39639162d606SStefano Zampini               start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k];
3964906d46d4SStefano Zampini               ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr);
3965984c4197SStefano Zampini               j++;
3966674ae819SStefano Zampini             }
3967674ae819SStefano Zampini           }
3968984c4197SStefano Zampini 
3969984c4197SStefano Zampini           /* check change of basis */
3970984c4197SStefano Zampini           if (pcbddc->dbg_flag) {
3971984c4197SStefano Zampini             PetscInt   ii,jj;
3972984c4197SStefano Zampini             PetscBool valid_qr=PETSC_TRUE;
3973c4303822SStefano Zampini             ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr);
3974c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
3975c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr);
3976c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
3977c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr);
3978c4303822SStefano Zampini             ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr);
3979984c4197SStefano Zampini             ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
3980cf5a6209SStefano Zampini             PetscStackCallBLAS("BLASgemm",BLASgemm_("T","N",&Blas_M,&Blas_N,&Blas_K,&one,dbg_work,&Blas_LDA,qr_basis,&Blas_LDB,&zero,&dbg_work[size_of_constraint*primal_dofs],&Blas_LDC));
3981984c4197SStefano Zampini             ierr = PetscFPTrapPop();CHKERRQ(ierr);
3982984c4197SStefano Zampini             for (jj=0;jj<size_of_constraint;jj++) {
3983984c4197SStefano Zampini               for (ii=0;ii<primal_dofs;ii++) {
3984cf5a6209SStefano Zampini                 if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE;
3985cf5a6209SStefano Zampini                 if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE;
3986674ae819SStefano Zampini               }
3987674ae819SStefano Zampini             }
3988984c4197SStefano Zampini             if (!valid_qr) {
398922d5777bSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr);
3990984c4197SStefano Zampini               for (jj=0;jj<size_of_constraint;jj++) {
3991984c4197SStefano Zampini                 for (ii=0;ii<primal_dofs;ii++) {
3992cf5a6209SStefano Zampini                   if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) {
3993cf5a6209SStefano Zampini                     PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not orthogonal to constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]));
3994674ae819SStefano Zampini                   }
3995cf5a6209SStefano Zampini                   if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) {
3996cf5a6209SStefano Zampini                     PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not unitary w.r.t constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]));
3997984c4197SStefano Zampini                   }
3998984c4197SStefano Zampini                 }
3999984c4197SStefano Zampini               }
4000674ae819SStefano Zampini             } else {
400122d5777bSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr);
4002674ae819SStefano Zampini             }
4003674ae819SStefano Zampini           }
4004a717540cSStefano Zampini         } else { /* simple transformation block */
4005a717540cSStefano Zampini           PetscInt    row,col;
4006a6b551f4SStefano Zampini           PetscScalar val,norm;
4007a6b551f4SStefano Zampini 
4008a6b551f4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
40099162d606SStefano Zampini           PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,constraints_data+constraints_data_ptr[total_counts],&Blas_one,constraints_data+constraints_data_ptr[total_counts],&Blas_one));
4010a717540cSStefano Zampini           for (j=0;j<size_of_constraint;j++) {
40119162d606SStefano Zampini             PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j];
40129162d606SStefano Zampini             row = constraints_idxs[constraints_idxs_ptr[total_counts]+j];
4013bbb9e6c6SStefano Zampini             if (!PetscBTLookup(is_primal,row_B)) {
40149162d606SStefano Zampini               col = constraints_idxs[constraints_idxs_ptr[total_counts]];
4015906d46d4SStefano Zampini               ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr);
40169162d606SStefano Zampini               ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr);
4017a717540cSStefano Zampini             } else {
4018a717540cSStefano Zampini               for (k=0;k<size_of_constraint;k++) {
40199162d606SStefano Zampini                 col = constraints_idxs[constraints_idxs_ptr[total_counts]+k];
4020a717540cSStefano Zampini                 if (row != col) {
40219162d606SStefano Zampini                   val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]];
4022a717540cSStefano Zampini                 } else {
40239162d606SStefano Zampini                   val = constraints_data[constraints_data_ptr[total_counts]]/norm;
4024a717540cSStefano Zampini                 }
4025906d46d4SStefano Zampini                 ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr);
4026a717540cSStefano Zampini               }
4027a717540cSStefano Zampini             }
4028a717540cSStefano Zampini           }
402998a51de6SStefano Zampini           if (pcbddc->dbg_flag) {
403022d5777bSStefano Zampini             ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr);
4031a717540cSStefano Zampini           }
4032674ae819SStefano Zampini         }
4033984c4197SStefano Zampini       } else {
4034984c4197SStefano Zampini         if (pcbddc->dbg_flag) {
40359162d606SStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,size_of_constraint);CHKERRQ(ierr);
4036674ae819SStefano Zampini         }
4037674ae819SStefano Zampini       }
4038674ae819SStefano Zampini     }
4039a717540cSStefano Zampini 
4040a717540cSStefano Zampini     /* free workspace */
4041a717540cSStefano Zampini     if (qr_needed) {
4042984c4197SStefano Zampini       if (pcbddc->dbg_flag) {
4043cf5a6209SStefano Zampini         ierr = PetscFree(dbg_work);CHKERRQ(ierr);
4044984c4197SStefano Zampini       }
4045984c4197SStefano Zampini       ierr = PetscFree(trs_rhs);CHKERRQ(ierr);
4046984c4197SStefano Zampini       ierr = PetscFree(qr_tau);CHKERRQ(ierr);
4047984c4197SStefano Zampini       ierr = PetscFree(qr_work);CHKERRQ(ierr);
4048984c4197SStefano Zampini       ierr = PetscFree(gqr_work);CHKERRQ(ierr);
4049984c4197SStefano Zampini       ierr = PetscFree(qr_basis);CHKERRQ(ierr);
4050674ae819SStefano Zampini     }
4051a717540cSStefano Zampini     ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr);
4052906d46d4SStefano Zampini     ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4053906d46d4SStefano Zampini     ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4054906d46d4SStefano Zampini 
4055906d46d4SStefano Zampini     /* assembling of global change of variable */
4056bbb9e6c6SStefano Zampini     {
4057bbb9e6c6SStefano Zampini       Mat      tmat;
405816f15bc4SStefano Zampini       PetscInt bs;
405916f15bc4SStefano Zampini 
4060906d46d4SStefano Zampini       ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr);
4061906d46d4SStefano Zampini       ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr);
4062bbb9e6c6SStefano Zampini       ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr);
4063bbb9e6c6SStefano Zampini       ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr);
4064bbb9e6c6SStefano Zampini       ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
4065bbb9e6c6SStefano Zampini       ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr);
406616f15bc4SStefano Zampini       ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr);
406716f15bc4SStefano Zampini       ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr);
4068906d46d4SStefano Zampini       ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr);
4069bbb9e6c6SStefano Zampini       ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr);
4070bbb9e6c6SStefano Zampini       ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
4071bbb9e6c6SStefano Zampini       ierr = MatDestroy(&tmat);CHKERRQ(ierr);
4072bbb9e6c6SStefano Zampini       ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
4073bbb9e6c6SStefano Zampini       ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr);
4074e176bc59SStefano Zampini       ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4075e176bc59SStefano Zampini       ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4076bbb9e6c6SStefano Zampini       ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr);
4077bbb9e6c6SStefano Zampini       ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr);
4078906d46d4SStefano Zampini     }
4079906d46d4SStefano Zampini     /* check */
4080906d46d4SStefano Zampini     if (pcbddc->dbg_flag) {
4081906d46d4SStefano Zampini       PetscReal error;
4082906d46d4SStefano Zampini       Vec       x,x_change;
4083906d46d4SStefano Zampini 
4084906d46d4SStefano Zampini       ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr);
4085906d46d4SStefano Zampini       ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr);
4086906d46d4SStefano Zampini       ierr = VecSetRandom(x,NULL);CHKERRQ(ierr);
4087906d46d4SStefano Zampini       ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr);
4088e176bc59SStefano Zampini       ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4089e176bc59SStefano Zampini       ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4090bbb9e6c6SStefano Zampini       ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr);
4091e176bc59SStefano Zampini       ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4092e176bc59SStefano Zampini       ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4093906d46d4SStefano Zampini       ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr);
4094906d46d4SStefano Zampini       ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr);
4095906d46d4SStefano Zampini       ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr);
4096906d46d4SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4097bbb9e6c6SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr);
4098906d46d4SStefano Zampini       ierr = VecDestroy(&x);CHKERRQ(ierr);
4099906d46d4SStefano Zampini       ierr = VecDestroy(&x_change);CHKERRQ(ierr);
4100906d46d4SStefano Zampini     }
4101b96c3477SStefano Zampini 
4102b96c3477SStefano Zampini     /* adapt sub_schurs computed (if any) */
4103b96c3477SStefano Zampini     if (pcbddc->use_deluxe_scaling) {
4104b96c3477SStefano Zampini       PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs;
4105b96c3477SStefano Zampini       if (sub_schurs->S_Ej_all) {
4106ac632422SStefano Zampini         Mat                    S_new,tmat;
4107b087196eSStefano Zampini         ISLocalToGlobalMapping NtoSall;
4108b087196eSStefano Zampini         IS                     is_all_N,is_V,is_V_Sall;
4109b087196eSStefano Zampini         const PetscScalar      *array;
4110b087196eSStefano Zampini         const PetscInt         *idxs_V,*idxs_all;
4111b087196eSStefano Zampini         PetscInt               i,n_V;
4112bbb9e6c6SStefano Zampini 
4113bbb9e6c6SStefano Zampini         ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr);
41146816873aSStefano Zampini         ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr);
4115b087196eSStefano Zampini         ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr);
4116b087196eSStefano Zampini         ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr);
4117b087196eSStefano Zampini         ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr);
4118b087196eSStefano Zampini         ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr);
4119bbb9e6c6SStefano Zampini         ierr = ISDestroy(&is_all_N);CHKERRQ(ierr);
4120b087196eSStefano Zampini         ierr = ISDestroy(&is_V);CHKERRQ(ierr);
4121ac632422SStefano Zampini         ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr);
4122b96c3477SStefano Zampini         ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr);
4123ac632422SStefano Zampini         ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr);
4124b087196eSStefano Zampini         ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr);
4125b087196eSStefano Zampini         ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr);
4126b087196eSStefano Zampini         ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr);
4127b087196eSStefano Zampini         ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr);
4128b087196eSStefano Zampini         ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr);
4129b087196eSStefano Zampini         for (i=0;i<n_V;i++) {
4130b087196eSStefano Zampini           PetscScalar val;
4131b087196eSStefano Zampini           PetscInt    idx;
4132b087196eSStefano Zampini 
4133b087196eSStefano Zampini           idx = idxs_V[i];
4134b087196eSStefano Zampini           val = array[idxs_all[idxs_V[i]]];
4135b087196eSStefano Zampini           ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr);
4136b087196eSStefano Zampini         }
4137b087196eSStefano Zampini         ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4138b087196eSStefano Zampini         ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4139ac632422SStefano Zampini         sub_schurs->S_Ej_all = S_new;
4140ac632422SStefano Zampini         ierr = MatDestroy(&S_new);CHKERRQ(ierr);
4141ac632422SStefano Zampini         if (sub_schurs->sum_S_Ej_all) {
4142ac632422SStefano Zampini           ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr);
4143b96c3477SStefano Zampini           ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr);
4144ac632422SStefano Zampini           ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr);
4145b087196eSStefano Zampini           ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr);
4146ac632422SStefano Zampini           sub_schurs->sum_S_Ej_all = S_new;
4147ac632422SStefano Zampini           ierr = MatDestroy(&S_new);CHKERRQ(ierr);
4148ac632422SStefano Zampini         }
4149b087196eSStefano Zampini         ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr);
4150b087196eSStefano Zampini         ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr);
4151b087196eSStefano Zampini         ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr);
4152b96c3477SStefano Zampini         ierr = MatDestroy(&tmat);CHKERRQ(ierr);
4153b087196eSStefano Zampini         ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr);
4154b96c3477SStefano Zampini       }
4155b96c3477SStefano Zampini     }
4156906d46d4SStefano Zampini     ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr);
4157906d46d4SStefano Zampini   } else if (pcbddc->user_ChangeOfBasisMatrix) {
4158b9b85e73SStefano Zampini     ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
4159b9b85e73SStefano Zampini     pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix;
4160b9b85e73SStefano Zampini   }
4161906d46d4SStefano Zampini 
4162906d46d4SStefano Zampini   /* set up change of basis context */
4163906d46d4SStefano Zampini   if (pcbddc->ChangeOfBasisMatrix) {
4164906d46d4SStefano Zampini     PCBDDCChange_ctx change_ctx;
4165906d46d4SStefano Zampini 
4166906d46d4SStefano Zampini     if (!pcbddc->new_global_mat) {
4167906d46d4SStefano Zampini       PetscInt global_size,local_size;
4168906d46d4SStefano Zampini 
4169906d46d4SStefano Zampini       ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr);
4170906d46d4SStefano Zampini       ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr);
4171906d46d4SStefano Zampini       ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr);
4172906d46d4SStefano Zampini       ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr);
4173906d46d4SStefano Zampini       ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr);
4174906d46d4SStefano Zampini       ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr);
4175906d46d4SStefano Zampini       ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr);
4176906d46d4SStefano Zampini       ierr = PetscNew(&change_ctx);CHKERRQ(ierr);
4177906d46d4SStefano Zampini       ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr);
4178906d46d4SStefano Zampini     } else {
4179906d46d4SStefano Zampini       ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr);
4180906d46d4SStefano Zampini       ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr);
4181906d46d4SStefano Zampini       ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr);
4182906d46d4SStefano Zampini     }
4183906d46d4SStefano Zampini     if (!pcbddc->user_ChangeOfBasisMatrix) {
4184906d46d4SStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
4185906d46d4SStefano Zampini       change_ctx->global_change = pcbddc->ChangeOfBasisMatrix;
4186906d46d4SStefano Zampini     } else {
4187906d46d4SStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
4188906d46d4SStefano Zampini       change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix;
4189906d46d4SStefano Zampini     }
4190906d46d4SStefano Zampini     ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr);
4191906d46d4SStefano Zampini     ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr);
4192906d46d4SStefano Zampini     ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4193906d46d4SStefano Zampini     ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4194a8661815SStefano Zampini   } else {
4195a8661815SStefano Zampini     ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr);
4196b9b85e73SStefano Zampini   }
4197a717540cSStefano Zampini 
41984f1b2e48SStefano Zampini   /* add pressure dofs to set of primal nodes for numbering purposes */
41994f1b2e48SStefano Zampini   for (i=0;i<pcbddc->benign_n;i++) {
42004f1b2e48SStefano Zampini     pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i];
42014f1b2e48SStefano Zampini     pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i];
4202019a44ceSStefano Zampini     pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1;
4203019a44ceSStefano Zampini     pcbddc->local_primal_size_cc++;
4204019a44ceSStefano Zampini     pcbddc->local_primal_size++;
4205019a44ceSStefano Zampini   }
4206019a44ceSStefano Zampini 
4207019a44ceSStefano Zampini   /* check if a new primal space has been introduced (also take into account benign trick) */
4208727cdba6SStefano Zampini   pcbddc->new_primal_space_local = PETSC_TRUE;
4209727cdba6SStefano Zampini   if (olocal_primal_size == pcbddc->local_primal_size) {
42109f47a83aSStefano Zampini     ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr);
4211c1c8e736SStefano Zampini     pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
42120e6343abSStefano Zampini     if (!pcbddc->new_primal_space_local) {
42139f47a83aSStefano Zampini       ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr);
4214727cdba6SStefano Zampini       pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
4215727cdba6SStefano Zampini     }
42160e6343abSStefano Zampini   }
42170e6343abSStefano Zampini   ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr);
4218727cdba6SStefano Zampini   /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */
4219727cdba6SStefano Zampini   ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
4220727cdba6SStefano Zampini 
4221a717540cSStefano Zampini   /* flush dbg viewer */
4222b8ffe317SStefano Zampini   if (pcbddc->dbg_flag) {
4223b8ffe317SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4224b8ffe317SStefano Zampini   }
4225a717540cSStefano Zampini 
4226e310c8b4SStefano Zampini   /* free workspace */
4227a717540cSStefano Zampini   ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr);
42284641a718SStefano Zampini   ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr);
422908122e43SStefano Zampini   if (!pcbddc->adaptive_selection) {
42309162d606SStefano Zampini     ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr);
42319162d606SStefano Zampini     ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr);
423208122e43SStefano Zampini   } else {
42339162d606SStefano Zampini     ierr = PetscFree5(pcbddc->adaptive_constraints_n,
42349162d606SStefano Zampini                       pcbddc->adaptive_constraints_idxs_ptr,
42359162d606SStefano Zampini                       pcbddc->adaptive_constraints_data_ptr,
423608122e43SStefano Zampini                       pcbddc->adaptive_constraints_idxs,
423708122e43SStefano Zampini                       pcbddc->adaptive_constraints_data);CHKERRQ(ierr);
42389162d606SStefano Zampini     ierr = PetscFree(constraints_n);CHKERRQ(ierr);
42399162d606SStefano Zampini     ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr);
424008122e43SStefano Zampini   }
4241674ae819SStefano Zampini   PetscFunctionReturn(0);
4242674ae819SStefano Zampini }
4243674ae819SStefano Zampini 
4244674ae819SStefano Zampini #undef __FUNCT__
4245674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface"
4246674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc)
4247674ae819SStefano Zampini {
4248674ae819SStefano Zampini   PC_BDDC     *pcbddc = (PC_BDDC*)pc->data;
4249674ae819SStefano Zampini   PC_IS       *pcis = (PC_IS*)pc->data;
4250674ae819SStefano Zampini   Mat_IS      *matis  = (Mat_IS*)pc->pmat->data;
42517fb0e2dbSStefano Zampini   PetscInt    ierr,i,vertex_size,N;
4252674ae819SStefano Zampini   PetscViewer viewer=pcbddc->dbg_viewer;
4253674ae819SStefano Zampini 
4254674ae819SStefano Zampini   PetscFunctionBegin;
42558e61c736SStefano Zampini   /* Reset previously computed graph */
42568e61c736SStefano Zampini   ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr);
4257674ae819SStefano Zampini   /* Init local Graph struct */
42587fb0e2dbSStefano Zampini   ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr);
42593bbff08aSStefano Zampini   ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr);
4260674ae819SStefano Zampini 
4261575ad6abSStefano Zampini   /* Check validity of the csr graph passed in by the user */
42625099eff2SStefano Zampini   if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) {
42635099eff2SStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs);
4264575ad6abSStefano Zampini   }
42659577ea80SStefano Zampini 
4266674ae819SStefano Zampini   /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */
4267d4d8cf7bSStefano Zampini   if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) {
42684d379d7bSStefano Zampini     PetscInt  *xadj,*adjncy;
42694d379d7bSStefano Zampini     PetscInt  nvtxs;
4270e496cd5dSStefano Zampini     PetscBool flg_row=PETSC_FALSE;
4271674ae819SStefano Zampini 
42722fffb893SStefano Zampini     ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
42732fffb893SStefano Zampini     if (flg_row) {
42744d379d7bSStefano Zampini       ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr);
4275b96c3477SStefano Zampini       pcbddc->computed_rowadj = PETSC_TRUE;
42762fffb893SStefano Zampini     }
42772fffb893SStefano Zampini     ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
4278674ae819SStefano Zampini   }
42799b28b941SStefano Zampini   if (pcbddc->dbg_flag) {
42809b28b941SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4281674ae819SStefano Zampini   }
4282674ae819SStefano Zampini 
428363602bcaSStefano Zampini   /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */
4284674ae819SStefano Zampini   vertex_size = 1;
428563602bcaSStefano Zampini   if (pcbddc->user_provided_isfordofs) {
428663602bcaSStefano Zampini     if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */
428795ecbf38SStefano Zampini       ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr);
428863602bcaSStefano Zampini       for (i=0;i<pcbddc->n_ISForDofs;i++) {
4289e176bc59SStefano Zampini         ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr);
429063602bcaSStefano Zampini         ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr);
4291674ae819SStefano Zampini       }
429263602bcaSStefano Zampini       pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs;
429363602bcaSStefano Zampini       pcbddc->n_ISForDofs = 0;
429463602bcaSStefano Zampini       ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr);
4295674ae819SStefano Zampini     }
429663602bcaSStefano Zampini     /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */
4297674ae819SStefano Zampini     ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr);
429863602bcaSStefano Zampini   } else {
429963602bcaSStefano Zampini     if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */
430063602bcaSStefano Zampini       ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr);
4301854ce69bSBarry Smith       ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr);
430263602bcaSStefano Zampini       for (i=0;i<pcbddc->n_ISForDofsLocal;i++) {
430363602bcaSStefano Zampini         ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr);
430463602bcaSStefano Zampini       }
430563602bcaSStefano Zampini     }
4306674ae819SStefano Zampini   }
4307674ae819SStefano Zampini 
4308674ae819SStefano Zampini   /* Setup of Graph */
4309785d1243SStefano Zampini   if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */
4310e176bc59SStefano Zampini     ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr);
4311785d1243SStefano Zampini   }
4312785d1243SStefano Zampini   if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */
4313e176bc59SStefano Zampini     ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr);
4314785d1243SStefano Zampini   }
431530368db7SStefano Zampini   if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */
431630368db7SStefano Zampini     ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr);
431730368db7SStefano Zampini   }
431830368db7SStefano Zampini   ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr);
4319674ae819SStefano Zampini 
43204f1b2e48SStefano Zampini   /* attach info on disconnected subdomains if present */
43214f1b2e48SStefano Zampini   if (pcbddc->n_local_subs) {
43224f1b2e48SStefano Zampini     PetscInt *local_subs;
43234f1b2e48SStefano Zampini 
43244f1b2e48SStefano Zampini     ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr);
43254f1b2e48SStefano Zampini     for (i=0;i<pcbddc->n_local_subs;i++) {
43264f1b2e48SStefano Zampini       const PetscInt *idxs;
43274f1b2e48SStefano Zampini       PetscInt       nl,j;
43284f1b2e48SStefano Zampini 
43294f1b2e48SStefano Zampini       ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr);
43304f1b2e48SStefano Zampini       ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr);
43314f1b2e48SStefano Zampini       for (j=0;j<nl;j++) {
43324f1b2e48SStefano Zampini         local_subs[idxs[j]] = i;
43334f1b2e48SStefano Zampini       }
43344f1b2e48SStefano Zampini       ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr);
43354f1b2e48SStefano Zampini     }
43364f1b2e48SStefano Zampini     pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs;
43374f1b2e48SStefano Zampini     pcbddc->mat_graph->local_subs = local_subs;
43384f1b2e48SStefano Zampini   }
43394f1b2e48SStefano Zampini 
4340674ae819SStefano Zampini   /* Graph's connected components analysis */
4341674ae819SStefano Zampini   ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr);
4342674ae819SStefano Zampini 
4343674ae819SStefano Zampini   /* print some info to stdout */
4344674ae819SStefano Zampini   if (pcbddc->dbg_flag) {
4345302440fdSBarry Smith     ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr);
4346674ae819SStefano Zampini   }
4347fb180af4SStefano Zampini 
4348fb180af4SStefano Zampini   /* mark topography has done */
4349fb180af4SStefano Zampini   pcbddc->recompute_topography = PETSC_FALSE;
4350674ae819SStefano Zampini   PetscFunctionReturn(0);
4351674ae819SStefano Zampini }
4352674ae819SStefano Zampini 
4353dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */
4354674ae819SStefano Zampini #undef __FUNCT__
4355674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering"
4356dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n)
4357674ae819SStefano Zampini {
4358dc456d91SStefano Zampini   PetscSF        sf;
4359dc456d91SStefano Zampini   PetscLayout    map;
4360dc456d91SStefano Zampini   const PetscInt *idxs;
4361dc456d91SStefano Zampini   PetscInt       *leaf_data,*root_data,*gidxs;
4362dc456d91SStefano Zampini   PetscInt       N,n,i,lbounds[2],gbounds[2],Nl;
4363dc456d91SStefano Zampini   PetscInt       n_n,nlocals,start,first_index;
4364dc456d91SStefano Zampini   PetscMPIInt    commsize;
4365674ae819SStefano Zampini   PetscBool      first_found;
4366674ae819SStefano Zampini   PetscErrorCode ierr;
4367674ae819SStefano Zampini 
4368674ae819SStefano Zampini   PetscFunctionBegin;
4369dc456d91SStefano Zampini   ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr);
4370dc456d91SStefano Zampini   if (subset_mult) {
4371dc456d91SStefano Zampini     PetscCheckSameComm(subset,1,subset_mult,2);
4372dc456d91SStefano Zampini     ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr);
4373dc456d91SStefano Zampini     if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i);
4374674ae819SStefano Zampini   }
4375dc456d91SStefano Zampini   /* create workspace layout for computing global indices of subset */
4376dc456d91SStefano Zampini   ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr);
4377dc456d91SStefano Zampini   lbounds[0] = lbounds[1] = 0;
4378dc456d91SStefano Zampini   for (i=0;i<n;i++) {
4379dc456d91SStefano Zampini     if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i];
4380dc456d91SStefano Zampini     else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i];
4381674ae819SStefano Zampini   }
4382dc456d91SStefano Zampini   lbounds[0] = -lbounds[0];
4383dc456d91SStefano Zampini   ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr);
4384dc456d91SStefano Zampini   gbounds[0] = -gbounds[0];
4385dc456d91SStefano Zampini   N = gbounds[1] - gbounds[0] + 1;
4386dc456d91SStefano Zampini   ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr);
4387dc456d91SStefano Zampini   ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr);
4388dc456d91SStefano Zampini   ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr);
4389dc456d91SStefano Zampini   ierr = PetscLayoutSetUp(map);CHKERRQ(ierr);
4390dc456d91SStefano Zampini   ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr);
4391dc456d91SStefano Zampini 
4392dc456d91SStefano Zampini   /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */
4393dc456d91SStefano Zampini   ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr);
4394dc456d91SStefano Zampini   if (subset_mult) {
4395dc456d91SStefano Zampini     const PetscInt* idxs_mult;
4396dc456d91SStefano Zampini 
4397dc456d91SStefano Zampini     ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr);
4398dc456d91SStefano Zampini     ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr);
4399dc456d91SStefano Zampini     ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr);
4400674ae819SStefano Zampini   } else {
4401dc456d91SStefano Zampini     for (i=0;i<n;i++) leaf_data[i] = 1;
4402674ae819SStefano Zampini   }
4403dc456d91SStefano Zampini   /* local size of new subset */
4404dc456d91SStefano Zampini   n_n = 0;
4405dc456d91SStefano Zampini   for (i=0;i<n;i++) n_n += leaf_data[i];
4406dc456d91SStefano Zampini 
4407dc456d91SStefano Zampini   /* global indexes in layout */
4408dc456d91SStefano Zampini   ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */
4409dc456d91SStefano Zampini   for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0];
4410dc456d91SStefano Zampini   ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr);
4411dc456d91SStefano Zampini   ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr);
4412dc456d91SStefano Zampini   ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr);
4413dc456d91SStefano Zampini   ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr);
4414dc456d91SStefano Zampini 
4415dc456d91SStefano Zampini   /* reduce from leaves to roots */
4416dc456d91SStefano Zampini   ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr);
441764a8e5bcSStefano Zampini   ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr);
441864a8e5bcSStefano Zampini   ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr);
4419dc456d91SStefano Zampini 
4420dc456d91SStefano Zampini   /* count indexes in local part of layout */
4421674ae819SStefano Zampini   nlocals = 0;
4422674ae819SStefano Zampini   first_index = -1;
4423674ae819SStefano Zampini   first_found = PETSC_FALSE;
4424dc456d91SStefano Zampini   for (i=0;i<Nl;i++) {
4425dc456d91SStefano Zampini     if (!first_found && root_data[i]) {
4426674ae819SStefano Zampini       first_found = PETSC_TRUE;
4427674ae819SStefano Zampini       first_index = i;
4428674ae819SStefano Zampini     }
4429dc456d91SStefano Zampini     nlocals += root_data[i];
4430674ae819SStefano Zampini   }
4431dc456d91SStefano Zampini 
4432dc456d91SStefano Zampini   /* cumulative of number of indexes and size of subset without holes */
44335fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN)
4434dc456d91SStefano Zampini   start = 0;
443564a8e5bcSStefano Zampini   ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr);
44365fa240b1SStefano Zampini #else
443764a8e5bcSStefano Zampini   ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr);
44385fa240b1SStefano Zampini   start = start-nlocals;
44395fa240b1SStefano Zampini #endif
44405fa240b1SStefano Zampini 
4441dc456d91SStefano Zampini   if (N_n) { /* compute total size of new subset if requested */
4442dc456d91SStefano Zampini     *N_n = start + nlocals;
4443dc456d91SStefano Zampini     ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr);
4444dc456d91SStefano Zampini     ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr);
4445674ae819SStefano Zampini   }
44465fa240b1SStefano Zampini 
44475fa240b1SStefano Zampini   /* adapt root data with cumulative */
4448674ae819SStefano Zampini   if (first_found) {
4449dc456d91SStefano Zampini     PetscInt old_index;
4450dc456d91SStefano Zampini 
4451dc456d91SStefano Zampini     root_data[first_index] += start;
4452674ae819SStefano Zampini     old_index = first_index;
4453dc456d91SStefano Zampini     for (i=first_index+1;i<Nl;i++) {
4454dc456d91SStefano Zampini       if (root_data[i]) {
4455dc456d91SStefano Zampini         root_data[i] += root_data[old_index];
4456674ae819SStefano Zampini         old_index = i;
4457674ae819SStefano Zampini       }
4458674ae819SStefano Zampini     }
4459674ae819SStefano Zampini   }
4460dc456d91SStefano Zampini 
4461dc456d91SStefano Zampini   /* from roots to leaves */
4462dc456d91SStefano Zampini   ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr);
4463dc456d91SStefano Zampini   ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr);
4464dc456d91SStefano Zampini   ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
4465dc456d91SStefano Zampini 
4466dc456d91SStefano Zampini   /* create new IS with global indexes without holes */
4467dc456d91SStefano Zampini   if (subset_mult) {
4468dc456d91SStefano Zampini     const PetscInt* idxs_mult;
4469dc456d91SStefano Zampini     PetscInt        cum;
4470dc456d91SStefano Zampini 
4471dc456d91SStefano Zampini     cum = 0;
4472dc456d91SStefano Zampini     ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr);
4473dc456d91SStefano Zampini     for (i=0;i<n;i++) {
4474dc456d91SStefano Zampini       PetscInt j;
4475dc456d91SStefano Zampini       for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j;
4476674ae819SStefano Zampini     }
4477dc456d91SStefano Zampini     ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr);
4478674ae819SStefano Zampini   } else {
4479dc456d91SStefano Zampini     for (i=0;i<n;i++) {
4480dc456d91SStefano Zampini       gidxs[i] = leaf_data[i]-1;
4481674ae819SStefano Zampini     }
4482674ae819SStefano Zampini   }
4483dc456d91SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr);
4484dc456d91SStefano Zampini   ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr);
4485674ae819SStefano Zampini   PetscFunctionReturn(0);
4486674ae819SStefano Zampini }
44879a7d3425SStefano Zampini 
44889a7d3425SStefano Zampini #undef __FUNCT__
44899a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs"
44909a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[])
44919a7d3425SStefano Zampini {
44929a7d3425SStefano Zampini   PetscInt       i,j;
44939a7d3425SStefano Zampini   PetscScalar    *alphas;
44949a7d3425SStefano Zampini   PetscErrorCode ierr;
44959a7d3425SStefano Zampini 
44969a7d3425SStefano Zampini   PetscFunctionBegin;
44979a7d3425SStefano Zampini   /* this implements stabilized Gram-Schmidt */
4498785e854fSJed Brown   ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr);
44999a7d3425SStefano Zampini   for (i=0;i<n;i++) {
45009a7d3425SStefano Zampini     ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr);
45019a7d3425SStefano Zampini     if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); }
45029a7d3425SStefano Zampini     for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); }
45039a7d3425SStefano Zampini   }
45049a7d3425SStefano Zampini   ierr = PetscFree(alphas);CHKERRQ(ierr);
45059a7d3425SStefano Zampini   PetscFunctionReturn(0);
45069a7d3425SStefano Zampini }
45079a7d3425SStefano Zampini 
4508e7931f94SStefano Zampini #undef __FUNCT__
450970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern"
4510b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends)
4511e7931f94SStefano Zampini {
451252e5ac9dSStefano Zampini   IS             ranks_send_to;
4513e7931f94SStefano Zampini   PetscInt       n_neighs,*neighs,*n_shared,**shared;
4514e7931f94SStefano Zampini   PetscMPIInt    size,rank,color;
451552e5ac9dSStefano Zampini   PetscInt       *xadj,*adjncy;
451652e5ac9dSStefano Zampini   PetscInt       *adjncy_wgt,*v_wgt,*ranks_send_to_idx;
45173837a79fSStefano Zampini   PetscInt       i,local_size,threshold=0;
45182b510759SStefano Zampini   PetscBool      use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE;
4519e7931f94SStefano Zampini   PetscSubcomm   subcomm;
452052e5ac9dSStefano Zampini   PetscErrorCode ierr;
4521a57a6d2fSStefano Zampini 
4522e7931f94SStefano Zampini   PetscFunctionBegin;
45232b510759SStefano Zampini   ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr);
45242b510759SStefano Zampini   ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr);
45252b510759SStefano Zampini   ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr);
4526e7931f94SStefano Zampini 
4527e7931f94SStefano Zampini   /* Get info on mapping */
45283bbff08aSStefano Zampini   ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr);
45293bbff08aSStefano Zampini   ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr);
4530e7931f94SStefano Zampini 
4531e7931f94SStefano Zampini   /* build local CSR graph of subdomains' connectivity */
4532785e854fSJed Brown   ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr);
4533e7931f94SStefano Zampini   xadj[0] = 0;
4534e7931f94SStefano Zampini   xadj[1] = PetscMax(n_neighs-1,0);
4535785e854fSJed Brown   ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr);
4536785e854fSJed Brown   ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr);
4537e7931f94SStefano Zampini 
45382b510759SStefano Zampini   if (threshold) {
4539d023bfaeSStefano Zampini     PetscInt xadj_count = 0;
45402b510759SStefano Zampini     for (i=1;i<n_neighs;i++) {
4541d023bfaeSStefano Zampini       if (n_shared[i] > threshold) {
4542d023bfaeSStefano Zampini         adjncy[xadj_count] = neighs[i];
4543d023bfaeSStefano Zampini         adjncy_wgt[xadj_count] = n_shared[i];
4544d023bfaeSStefano Zampini         xadj_count++;
4545e7931f94SStefano Zampini       }
4546e7931f94SStefano Zampini     }
4547d023bfaeSStefano Zampini     xadj[1] = xadj_count;
4548c8587f34SStefano Zampini   } else {
4549e7931f94SStefano Zampini     if (xadj[1]) {
4550e7931f94SStefano Zampini       ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr);
4551e7931f94SStefano Zampini       ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr);
4552c8587f34SStefano Zampini     }
4553e7931f94SStefano Zampini   }
45543bbff08aSStefano Zampini   ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr);
4555e7931f94SStefano Zampini   if (use_square) {
4556e7931f94SStefano Zampini     for (i=0;i<xadj[1];i++) {
4557e7931f94SStefano Zampini       adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i];
4558e7931f94SStefano Zampini     }
4559e7931f94SStefano Zampini   }
4560e7931f94SStefano Zampini   ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr);
4561e7931f94SStefano Zampini 
45623837a79fSStefano Zampini   ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr);
4563e7931f94SStefano Zampini 
4564e7931f94SStefano Zampini   /*
4565e7931f94SStefano Zampini     Restrict work on active processes only.
4566e7931f94SStefano Zampini   */
4567e7931f94SStefano Zampini   ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr);
4568e7931f94SStefano Zampini   ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */
4569e7931f94SStefano Zampini   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr);
45702b510759SStefano Zampini   ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr);
4571d3531aaaSJed Brown   ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr);
4572e7931f94SStefano Zampini   if (color) {
4573e7931f94SStefano Zampini     ierr = PetscFree(xadj);CHKERRQ(ierr);
4574e7931f94SStefano Zampini     ierr = PetscFree(adjncy);CHKERRQ(ierr);
4575e7931f94SStefano Zampini     ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr);
4576c8587f34SStefano Zampini   } else {
457752e5ac9dSStefano Zampini     Mat             subdomain_adj;
457852e5ac9dSStefano Zampini     IS              new_ranks,new_ranks_contig;
457952e5ac9dSStefano Zampini     MatPartitioning partitioner;
458052e5ac9dSStefano Zampini     PetscInt        prank,rstart=0,rend=0;
458152e5ac9dSStefano Zampini     PetscInt        *is_indices,*oldranks;
4582b0c7d250SStefano Zampini     PetscBool       aggregate;
4583b0c7d250SStefano Zampini 
4584306c2d5bSBarry Smith     ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr);
4585785e854fSJed Brown     ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr);
4586e7931f94SStefano Zampini     prank = rank;
4587306c2d5bSBarry Smith     ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr);
45888002ef2cSStefano Zampini     /*
4589e7931f94SStefano Zampini     for (i=0;i<size;i++) {
4590e7931f94SStefano Zampini       PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]);
4591c8587f34SStefano Zampini     }
45928002ef2cSStefano Zampini     */
4593e7931f94SStefano Zampini     for (i=0;i<xadj[1];i++) {
4594e7931f94SStefano Zampini       ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr);
4595c8587f34SStefano Zampini     }
4596e7931f94SStefano Zampini     ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr);
4597b0c7d250SStefano Zampini     aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE);
4598b0c7d250SStefano Zampini     if (aggregate) {
4599b0c7d250SStefano Zampini       PetscInt    lrows,row,ncols,*cols;
4600b0c7d250SStefano Zampini       PetscMPIInt nrank;
4601b0c7d250SStefano Zampini       PetscScalar *vals;
4602b0c7d250SStefano Zampini 
4603b0c7d250SStefano Zampini       ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr);
4604b0c7d250SStefano Zampini       lrows = 0;
4605b0c7d250SStefano Zampini       if (nrank<redprocs) {
4606b0c7d250SStefano Zampini         lrows = size/redprocs;
4607b0c7d250SStefano Zampini         if (nrank<size%redprocs) lrows++;
4608b0c7d250SStefano Zampini       }
46095fa240b1SStefano Zampini       ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr);
4610b0c7d250SStefano Zampini       ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr);
4611b0c7d250SStefano Zampini       ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
4612b0c7d250SStefano Zampini       ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
4613b0c7d250SStefano Zampini       row = nrank;
4614b0c7d250SStefano Zampini       ncols = xadj[1]-xadj[0];
4615b0c7d250SStefano Zampini       cols = adjncy;
4616b0c7d250SStefano Zampini       ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr);
4617b0c7d250SStefano Zampini       for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i];
4618b0c7d250SStefano Zampini       ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr);
4619b0c7d250SStefano Zampini       ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4620b0c7d250SStefano Zampini       ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
462152e5ac9dSStefano Zampini       ierr = PetscFree(xadj);CHKERRQ(ierr);
462252e5ac9dSStefano Zampini       ierr = PetscFree(adjncy);CHKERRQ(ierr);
462352e5ac9dSStefano Zampini       ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr);
4624b0c7d250SStefano Zampini       ierr = PetscFree(vals);CHKERRQ(ierr);
4625b0c7d250SStefano Zampini     } else {
4626306c2d5bSBarry Smith       ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr);
4627b0c7d250SStefano Zampini     }
462822b6e8a2SStefano Zampini     /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */
4629e7931f94SStefano Zampini 
4630e7931f94SStefano Zampini     /* Partition */
4631306c2d5bSBarry Smith     ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr);
4632e7931f94SStefano Zampini     ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr);
4633e7931f94SStefano Zampini     if (use_vwgt) {
46343837a79fSStefano Zampini       ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr);
4635e7931f94SStefano Zampini       v_wgt[0] = local_size;
4636e7931f94SStefano Zampini       ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr);
4637c8587f34SStefano Zampini     }
463828143c3dSStefano Zampini     n_subdomains = PetscMin((PetscInt)size,n_subdomains);
463928143c3dSStefano Zampini     ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr);
4640e7931f94SStefano Zampini     ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr);
4641e7931f94SStefano Zampini     ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr);
464222b6e8a2SStefano Zampini     /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */
4643e7931f94SStefano Zampini 
464452e5ac9dSStefano Zampini     /* renumber new_ranks to avoid "holes" in new set of processors */
464552e5ac9dSStefano Zampini     ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr);
464652e5ac9dSStefano Zampini     ierr = ISDestroy(&new_ranks);CHKERRQ(ierr);
464752e5ac9dSStefano Zampini     ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr);
4648b0c7d250SStefano Zampini     if (!redprocs) {
4649b0c7d250SStefano Zampini       ranks_send_to_idx[0] = oldranks[is_indices[0]];
465028143c3dSStefano Zampini     } else {
4651b0c7d250SStefano Zampini       PetscInt    idxs[1];
4652b0c7d250SStefano Zampini       PetscMPIInt tag;
4653b0c7d250SStefano Zampini       MPI_Request *reqs;
4654b0c7d250SStefano Zampini 
4655b0c7d250SStefano Zampini       ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr);
4656b0c7d250SStefano Zampini       ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr);
4657b0c7d250SStefano Zampini       for (i=rstart;i<rend;i++) {
4658b0c7d250SStefano Zampini         ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr);
465928143c3dSStefano Zampini       }
4660b0c7d250SStefano Zampini       ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr);
4661b0c7d250SStefano Zampini       ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
4662b0c7d250SStefano Zampini       ierr = PetscFree(reqs);CHKERRQ(ierr);
4663b0c7d250SStefano Zampini       ranks_send_to_idx[0] = oldranks[idxs[0]];
4664e7931f94SStefano Zampini     }
466552e5ac9dSStefano Zampini     ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr);
4666e7931f94SStefano Zampini     /* clean up */
4667e7931f94SStefano Zampini     ierr = PetscFree(oldranks);CHKERRQ(ierr);
466852e5ac9dSStefano Zampini     ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr);
4669e7931f94SStefano Zampini     ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr);
4670e7931f94SStefano Zampini     ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr);
4671e7931f94SStefano Zampini   }
4672e7931f94SStefano Zampini   ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr);
4673e7931f94SStefano Zampini 
4674e7931f94SStefano Zampini   /* assemble parallel IS for sends */
4675e7931f94SStefano Zampini   i = 1;
4676e7931f94SStefano Zampini   if (color) i=0;
4677e7931f94SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr);
4678e7931f94SStefano Zampini   /* get back IS */
4679e7931f94SStefano Zampini   *is_sends = ranks_send_to;
4680e7931f94SStefano Zampini   PetscFunctionReturn(0);
4681e7931f94SStefano Zampini }
4682e7931f94SStefano Zampini 
4683e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate;
4684e7931f94SStefano Zampini 
4685e7931f94SStefano Zampini #undef __FUNCT__
4686e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble"
468753a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[])
4688e7931f94SStefano Zampini {
468970cf5478SStefano Zampini   Mat                    local_mat;
4690e7931f94SStefano Zampini   IS                     is_sends_internal;
46919d30be91SStefano Zampini   PetscInt               rows,cols,new_local_rows;
469228143c3dSStefano Zampini   PetscInt               i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals;
46939d30be91SStefano Zampini   PetscBool              ismatis,isdense,newisdense,destroy_mat;
4694e7931f94SStefano Zampini   ISLocalToGlobalMapping l2gmap;
4695e7931f94SStefano Zampini   PetscInt*              l2gmap_indices;
4696e7931f94SStefano Zampini   const PetscInt*        is_indices;
4697e7931f94SStefano Zampini   MatType                new_local_type;
4698e7931f94SStefano Zampini   /* buffers */
4699e7931f94SStefano Zampini   PetscInt               *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs;
470028143c3dSStefano Zampini   PetscInt               *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is;
47019d30be91SStefano Zampini   PetscInt               *recv_buffer_idxs_local;
4702e7931f94SStefano Zampini   PetscScalar            *ptr_vals,*send_buffer_vals,*recv_buffer_vals;
4703e7931f94SStefano Zampini   /* MPI */
470428143c3dSStefano Zampini   MPI_Comm               comm,comm_n;
470528143c3dSStefano Zampini   PetscSubcomm           subcomm;
4706e7931f94SStefano Zampini   PetscMPIInt            n_sends,n_recvs,commsize;
470728143c3dSStefano Zampini   PetscMPIInt            *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is;
470828143c3dSStefano Zampini   PetscMPIInt            *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals;
470928143c3dSStefano Zampini   PetscMPIInt            len,tag_idxs,tag_idxs_is,tag_vals,source_dest;
471028143c3dSStefano Zampini   MPI_Request            *send_req_idxs,*send_req_idxs_is,*send_req_vals;
471128143c3dSStefano Zampini   MPI_Request            *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals;
4712e7931f94SStefano Zampini   PetscErrorCode         ierr;
4713e7931f94SStefano Zampini 
4714e7931f94SStefano Zampini   PetscFunctionBegin;
471528143c3dSStefano Zampini   /* TODO: add missing checks */
471628143c3dSStefano Zampini   PetscValidLogicalCollectiveInt(mat,n_subdomains,3);
471728143c3dSStefano Zampini   PetscValidLogicalCollectiveBool(mat,restrict_comm,4);
471828143c3dSStefano Zampini   PetscValidLogicalCollectiveEnum(mat,reuse,5);
471928143c3dSStefano Zampini   PetscValidLogicalCollectiveInt(mat,nis,7);
4720e7931f94SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr);
472128143c3dSStefano Zampini   if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__);
4722e7931f94SStefano Zampini   ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr);
4723e7931f94SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr);
4724e7931f94SStefano Zampini   if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE");
4725e7931f94SStefano Zampini   ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr);
4726e7931f94SStefano Zampini   if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square");
472728143c3dSStefano Zampini   if (reuse == MAT_REUSE_MATRIX && *mat_n) {
472870cf5478SStefano Zampini     PetscInt mrows,mcols,mnrows,mncols;
472970cf5478SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr);
473028143c3dSStefano Zampini     if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS");
473170cf5478SStefano Zampini     ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr);
473270cf5478SStefano Zampini     ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr);
473370cf5478SStefano Zampini     if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows);
473470cf5478SStefano Zampini     if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols);
473570cf5478SStefano Zampini   }
4736e7931f94SStefano Zampini   ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr);
4737e7931f94SStefano Zampini   PetscValidLogicalCollectiveInt(mat,bs,0);
4738e7931f94SStefano Zampini   /* prepare IS for sending if not provided */
4739e7931f94SStefano Zampini   if (!is_sends) {
474028143c3dSStefano Zampini     if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains");
4741b0c7d250SStefano Zampini     ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr);
4742c8587f34SStefano Zampini   } else {
4743e7931f94SStefano Zampini     ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr);
4744e7931f94SStefano Zampini     is_sends_internal = is_sends;
4745c8587f34SStefano Zampini   }
4746e7931f94SStefano Zampini 
4747e7931f94SStefano Zampini   /* get comm */
4748a316fed8SStefano Zampini   ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
4749e7931f94SStefano Zampini 
4750e7931f94SStefano Zampini   /* compute number of sends */
4751e7931f94SStefano Zampini   ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr);
4752e7931f94SStefano Zampini   ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr);
4753e7931f94SStefano Zampini 
4754e7931f94SStefano Zampini   /* compute number of receives */
4755e7931f94SStefano Zampini   ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr);
4756785e854fSJed Brown   ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr);
4757e7931f94SStefano Zampini   ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr);
4758e7931f94SStefano Zampini   ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr);
4759e7931f94SStefano Zampini   for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1;
4760e7931f94SStefano Zampini   ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr);
4761e7931f94SStefano Zampini   ierr = PetscFree(iflags);CHKERRQ(ierr);
4762e7931f94SStefano Zampini 
476328143c3dSStefano Zampini   /* restrict comm if requested */
476428143c3dSStefano Zampini   subcomm = 0;
476528143c3dSStefano Zampini   destroy_mat = PETSC_FALSE;
476628143c3dSStefano Zampini   if (restrict_comm) {
4767779c1cceSStefano Zampini     PetscMPIInt color,subcommsize;
4768779c1cceSStefano Zampini 
476928143c3dSStefano Zampini     color = 0;
477053a05cb3SStefano Zampini     if (restrict_full) {
477153a05cb3SStefano Zampini       if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */
477253a05cb3SStefano Zampini     } else {
477353a05cb3SStefano Zampini       if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */
477453a05cb3SStefano Zampini     }
477528143c3dSStefano Zampini     ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr);
477628143c3dSStefano Zampini     subcommsize = commsize - subcommsize;
477728143c3dSStefano Zampini     /* check if reuse has been requested */
477828143c3dSStefano Zampini     if (reuse == MAT_REUSE_MATRIX) {
477928143c3dSStefano Zampini       if (*mat_n) {
478028143c3dSStefano Zampini         PetscMPIInt subcommsize2;
478128143c3dSStefano Zampini         ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr);
478228143c3dSStefano Zampini         if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2);
478328143c3dSStefano Zampini         comm_n = PetscObjectComm((PetscObject)*mat_n);
478428143c3dSStefano Zampini       } else {
478528143c3dSStefano Zampini         comm_n = PETSC_COMM_SELF;
478628143c3dSStefano Zampini       }
478728143c3dSStefano Zampini     } else { /* MAT_INITIAL_MATRIX */
4788779c1cceSStefano Zampini       PetscMPIInt rank;
4789779c1cceSStefano Zampini 
4790779c1cceSStefano Zampini       ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
479128143c3dSStefano Zampini       ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr);
479228143c3dSStefano Zampini       ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr);
479328143c3dSStefano Zampini       ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr);
4794306c2d5bSBarry Smith       comm_n = PetscSubcommChild(subcomm);
479528143c3dSStefano Zampini     }
479628143c3dSStefano Zampini     /* flag to destroy *mat_n if not significative */
479728143c3dSStefano Zampini     if (color) destroy_mat = PETSC_TRUE;
479828143c3dSStefano Zampini   } else {
479928143c3dSStefano Zampini     comm_n = comm;
480028143c3dSStefano Zampini   }
480128143c3dSStefano Zampini 
4802e7931f94SStefano Zampini   /* prepare send/receive buffers */
4803785e854fSJed Brown   ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr);
4804e7931f94SStefano Zampini   ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr);
4805785e854fSJed Brown   ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr);
4806e7931f94SStefano Zampini   ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr);
480728143c3dSStefano Zampini   if (nis) {
4808854ce69bSBarry Smith     ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr);
480928143c3dSStefano Zampini   }
4810e7931f94SStefano Zampini 
481128143c3dSStefano Zampini   /* Get data from local matrices */
4812e7931f94SStefano Zampini   if (!isdense) {
4813a26c9d0eSStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented");
4814e7931f94SStefano Zampini     /* TODO: See below some guidelines on how to prepare the local buffers */
4815e7931f94SStefano Zampini     /*
4816e7931f94SStefano Zampini        send_buffer_vals should contain the raw values of the local matrix
4817e7931f94SStefano Zampini        send_buffer_idxs should contain:
4818e7931f94SStefano Zampini        - MatType_PRIVATE type
4819e7931f94SStefano Zampini        - PetscInt        size_of_l2gmap
4820e7931f94SStefano Zampini        - PetscInt        global_row_indices[size_of_l2gmap]
4821e7931f94SStefano Zampini        - PetscInt        all_other_info_which_is_needed_to_compute_preallocation_and_set_values
4822e7931f94SStefano Zampini     */
4823e7931f94SStefano Zampini   } else {
4824e7931f94SStefano Zampini     ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr);
48253bbff08aSStefano Zampini     ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr);
4826854ce69bSBarry Smith     ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr);
4827e7931f94SStefano Zampini     send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE;
4828e7931f94SStefano Zampini     send_buffer_idxs[1] = i;
48293bbff08aSStefano Zampini     ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr);
4830e7931f94SStefano Zampini     ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr);
48313bbff08aSStefano Zampini     ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr);
4832e7931f94SStefano Zampini     ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr);
4833e7931f94SStefano Zampini     for (i=0;i<n_sends;i++) {
4834e7931f94SStefano Zampini       ilengths_vals[is_indices[i]] = len*len;
4835e7931f94SStefano Zampini       ilengths_idxs[is_indices[i]] = len+2;
4836c8587f34SStefano Zampini     }
4837c8587f34SStefano Zampini   }
4838e7931f94SStefano Zampini   ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr);
483928143c3dSStefano Zampini   /* additional is (if any) */
484028143c3dSStefano Zampini   if (nis) {
484128143c3dSStefano Zampini     PetscMPIInt psum;
484228143c3dSStefano Zampini     PetscInt j;
484328143c3dSStefano Zampini     for (j=0,psum=0;j<nis;j++) {
484428143c3dSStefano Zampini       PetscInt plen;
484528143c3dSStefano Zampini       ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr);
484628143c3dSStefano Zampini       ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr);
484728143c3dSStefano Zampini       psum += len+1; /* indices + lenght */
484828143c3dSStefano Zampini     }
4849854ce69bSBarry Smith     ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr);
485028143c3dSStefano Zampini     for (j=0,psum=0;j<nis;j++) {
485128143c3dSStefano Zampini       PetscInt plen;
485228143c3dSStefano Zampini       const PetscInt *is_array_idxs;
485328143c3dSStefano Zampini       ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr);
485428143c3dSStefano Zampini       send_buffer_idxs_is[psum] = plen;
485528143c3dSStefano Zampini       ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr);
485628143c3dSStefano Zampini       ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr);
485728143c3dSStefano Zampini       ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr);
485828143c3dSStefano Zampini       psum += plen+1; /* indices + lenght */
485928143c3dSStefano Zampini     }
486028143c3dSStefano Zampini     for (i=0;i<n_sends;i++) {
486128143c3dSStefano Zampini       ilengths_idxs_is[is_indices[i]] = psum;
486228143c3dSStefano Zampini     }
486328143c3dSStefano Zampini     ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr);
486428143c3dSStefano Zampini   }
486528143c3dSStefano Zampini 
4866e7931f94SStefano Zampini   buf_size_idxs = 0;
4867e7931f94SStefano Zampini   buf_size_vals = 0;
486828143c3dSStefano Zampini   buf_size_idxs_is = 0;
4869e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
4870e7931f94SStefano Zampini     buf_size_idxs += (PetscInt)olengths_idxs[i];
4871e7931f94SStefano Zampini     buf_size_vals += (PetscInt)olengths_vals[i];
487228143c3dSStefano Zampini     if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i];
4873e7931f94SStefano Zampini   }
4874785e854fSJed Brown   ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr);
4875785e854fSJed Brown   ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr);
487695ecbf38SStefano Zampini   ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr);
4877e7931f94SStefano Zampini 
4878e7931f94SStefano Zampini   /* get new tags for clean communications */
4879e7931f94SStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr);
4880e7931f94SStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr);
488128143c3dSStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr);
4882e7931f94SStefano Zampini 
4883e7931f94SStefano Zampini   /* allocate for requests */
4884785e854fSJed Brown   ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr);
4885785e854fSJed Brown   ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr);
488695ecbf38SStefano Zampini   ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr);
4887785e854fSJed Brown   ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr);
4888785e854fSJed Brown   ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr);
488995ecbf38SStefano Zampini   ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr);
4890e7931f94SStefano Zampini 
4891e7931f94SStefano Zampini   /* communications */
4892e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
4893e7931f94SStefano Zampini   ptr_vals = recv_buffer_vals;
489428143c3dSStefano Zampini   ptr_idxs_is = recv_buffer_idxs_is;
4895e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
4896e7931f94SStefano Zampini     source_dest = onodes[i];
4897e7931f94SStefano Zampini     ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr);
4898e7931f94SStefano Zampini     ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr);
4899e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
4900e7931f94SStefano Zampini     ptr_vals += olengths_vals[i];
490128143c3dSStefano Zampini     if (nis) {
490228143c3dSStefano Zampini       ierr = MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);CHKERRQ(ierr);
490328143c3dSStefano Zampini       ptr_idxs_is += olengths_idxs_is[i];
490428143c3dSStefano Zampini     }
4905e7931f94SStefano Zampini   }
4906e7931f94SStefano Zampini   for (i=0;i<n_sends;i++) {
4907e7931f94SStefano Zampini     ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr);
4908e7931f94SStefano Zampini     ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr);
4909e7931f94SStefano Zampini     ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr);
491028143c3dSStefano Zampini     if (nis) {
491128143c3dSStefano Zampini       ierr = MPI_Isend(send_buffer_idxs_is,ilengths_idxs_is[source_dest],MPIU_INT,source_dest,tag_idxs_is,comm,&send_req_idxs_is[i]);CHKERRQ(ierr);
491228143c3dSStefano Zampini     }
4913e7931f94SStefano Zampini   }
4914e7931f94SStefano Zampini   ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr);
4915e7931f94SStefano Zampini   ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr);
4916e7931f94SStefano Zampini 
4917e7931f94SStefano Zampini   /* assemble new l2g map */
4918e7931f94SStefano Zampini   ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
4919e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
49209d30be91SStefano Zampini   new_local_rows = 0;
4921e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
49229d30be91SStefano Zampini     new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
4923e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
4924e7931f94SStefano Zampini   }
49259d30be91SStefano Zampini   ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr);
4926e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
49279d30be91SStefano Zampini   new_local_rows = 0;
4928e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
49299d30be91SStefano Zampini     ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr);
49309d30be91SStefano Zampini     new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
4931e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
4932e7931f94SStefano Zampini   }
49339d30be91SStefano Zampini   ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr);
49349d30be91SStefano Zampini   ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr);
4935e7931f94SStefano Zampini   ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr);
4936e7931f94SStefano Zampini 
4937e7931f94SStefano Zampini   /* infer new local matrix type from received local matrices type */
4938e7931f94SStefano Zampini   /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */
4939e7931f94SStefano Zampini   /* it also assumes that if the block size is set, than it is the same among all local matrices (see checks at the beginning of the function) */
4940e7931f94SStefano Zampini   if (n_recvs) {
494128143c3dSStefano Zampini     MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0];
4942e7931f94SStefano Zampini     ptr_idxs = recv_buffer_idxs;
4943e7931f94SStefano Zampini     for (i=0;i<n_recvs;i++) {
4944e7931f94SStefano Zampini       if ((PetscInt)new_local_type_private != *ptr_idxs) {
4945e7931f94SStefano Zampini         new_local_type_private = MATAIJ_PRIVATE;
4946e7931f94SStefano Zampini         break;
4947e7931f94SStefano Zampini       }
4948e7931f94SStefano Zampini       ptr_idxs += olengths_idxs[i];
4949e7931f94SStefano Zampini     }
4950e7931f94SStefano Zampini     switch (new_local_type_private) {
495128143c3dSStefano Zampini       case MATDENSE_PRIVATE:
495228143c3dSStefano Zampini         if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */
4953e7931f94SStefano Zampini           new_local_type = MATSEQAIJ;
4954e7931f94SStefano Zampini           bs = 1;
495528143c3dSStefano Zampini         } else { /* if I receive only 1 dense matrix */
495628143c3dSStefano Zampini           new_local_type = MATSEQDENSE;
495728143c3dSStefano Zampini           bs = 1;
495828143c3dSStefano Zampini         }
4959e7931f94SStefano Zampini         break;
4960e7931f94SStefano Zampini       case MATAIJ_PRIVATE:
4961e7931f94SStefano Zampini         new_local_type = MATSEQAIJ;
4962e7931f94SStefano Zampini         bs = 1;
4963e7931f94SStefano Zampini         break;
4964e7931f94SStefano Zampini       case MATBAIJ_PRIVATE:
4965e7931f94SStefano Zampini         new_local_type = MATSEQBAIJ;
4966e7931f94SStefano Zampini         break;
4967e7931f94SStefano Zampini       case MATSBAIJ_PRIVATE:
4968e7931f94SStefano Zampini         new_local_type = MATSEQSBAIJ;
4969e7931f94SStefano Zampini         break;
4970e7931f94SStefano Zampini       default:
49719d30be91SStefano Zampini         SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__);
4972e7931f94SStefano Zampini         break;
4973e7931f94SStefano Zampini     }
497428143c3dSStefano Zampini   } else { /* by default, new_local_type is seqdense */
497528143c3dSStefano Zampini     new_local_type = MATSEQDENSE;
497628143c3dSStefano Zampini     bs = 1;
4977e7931f94SStefano Zampini   }
4978e7931f94SStefano Zampini 
497970cf5478SStefano Zampini   /* create MATIS object if needed */
498070cf5478SStefano Zampini   if (reuse == MAT_INITIAL_MATRIX) {
4981e7931f94SStefano Zampini     ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr);
4982e176bc59SStefano Zampini     ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr);
498370cf5478SStefano Zampini   } else {
498470cf5478SStefano Zampini     /* it also destroys the local matrices */
498570cf5478SStefano Zampini     ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr);
498670cf5478SStefano Zampini   }
498770cf5478SStefano Zampini   ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr);
4988e7931f94SStefano Zampini   ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr);
49899d30be91SStefano Zampini 
49909d30be91SStefano Zampini   ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
49919d30be91SStefano Zampini 
49929d30be91SStefano Zampini   /* Global to local map of received indices */
49939d30be91SStefano Zampini   ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */
49949d30be91SStefano Zampini   ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr);
49959d30be91SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr);
49969d30be91SStefano Zampini 
49979d30be91SStefano Zampini   /* restore attributes -> type of incoming data and its size */
49989d30be91SStefano Zampini   buf_size_idxs = 0;
49999d30be91SStefano Zampini   for (i=0;i<n_recvs;i++) {
50009d30be91SStefano Zampini     recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs];
50019d30be91SStefano Zampini     recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1];
50029d30be91SStefano Zampini     buf_size_idxs += (PetscInt)olengths_idxs[i];
50039d30be91SStefano Zampini   }
50049d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr);
50059d30be91SStefano Zampini 
50069d30be91SStefano Zampini   /* set preallocation */
50079d30be91SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr);
50089d30be91SStefano Zampini   if (!newisdense) {
50099d30be91SStefano Zampini     PetscInt *new_local_nnz=0;
50109d30be91SStefano Zampini 
50119d30be91SStefano Zampini     ptr_vals = recv_buffer_vals;
50129d30be91SStefano Zampini     ptr_idxs = recv_buffer_idxs_local;
50139d30be91SStefano Zampini     if (n_recvs) {
50149d30be91SStefano Zampini       ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr);
50159d30be91SStefano Zampini     }
50169d30be91SStefano Zampini     for (i=0;i<n_recvs;i++) {
50179d30be91SStefano Zampini       PetscInt j;
50189d30be91SStefano Zampini       if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */
50199d30be91SStefano Zampini         for (j=0;j<*(ptr_idxs+1);j++) {
50209d30be91SStefano Zampini           new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1);
50219d30be91SStefano Zampini         }
50229d30be91SStefano Zampini       } else {
50239d30be91SStefano Zampini         /* TODO */
50249d30be91SStefano Zampini       }
50259d30be91SStefano Zampini       ptr_idxs += olengths_idxs[i];
50269d30be91SStefano Zampini     }
50279d30be91SStefano Zampini     if (new_local_nnz) {
50289d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows);
50299d30be91SStefano Zampini       ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr);
50309d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs;
50319d30be91SStefano Zampini       ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr);
50329d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0);
50339d30be91SStefano Zampini       ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr);
50349d30be91SStefano Zampini     } else {
50359d30be91SStefano Zampini       ierr = MatSetUp(local_mat);CHKERRQ(ierr);
50369d30be91SStefano Zampini     }
50379d30be91SStefano Zampini     ierr = PetscFree(new_local_nnz);CHKERRQ(ierr);
50389d30be91SStefano Zampini   } else {
50399d30be91SStefano Zampini     ierr = MatSetUp(local_mat);CHKERRQ(ierr);
50409d30be91SStefano Zampini   }
5041e7931f94SStefano Zampini 
5042e7931f94SStefano Zampini   /* set values */
5043e7931f94SStefano Zampini   ptr_vals = recv_buffer_vals;
50449d30be91SStefano Zampini   ptr_idxs = recv_buffer_idxs_local;
5045e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
5046e7931f94SStefano Zampini     if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */
5047e7931f94SStefano Zampini       ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr);
50489d30be91SStefano Zampini       ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr);
5049e7931f94SStefano Zampini       ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr);
5050e7931f94SStefano Zampini       ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr);
5051e7931f94SStefano Zampini       ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr);
505228143c3dSStefano Zampini     } else {
505328143c3dSStefano Zampini       /* TODO */
5054e7931f94SStefano Zampini     }
5055e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
5056e7931f94SStefano Zampini     ptr_vals += olengths_vals[i];
5057e7931f94SStefano Zampini   }
5058e7931f94SStefano Zampini   ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
5059e7931f94SStefano Zampini   ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
506070cf5478SStefano Zampini   ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
506170cf5478SStefano Zampini   ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
50629d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr);
50639d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr);
5064e7931f94SStefano Zampini 
5065dfd14d43SStefano Zampini #if 0
506628143c3dSStefano Zampini   if (!restrict_comm) { /* check */
5067e7931f94SStefano Zampini     Vec       lvec,rvec;
5068e7931f94SStefano Zampini     PetscReal infty_error;
5069e7931f94SStefano Zampini 
50702a7a6963SBarry Smith     ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr);
5071e7931f94SStefano Zampini     ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr);
5072e7931f94SStefano Zampini     ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr);
5073e7931f94SStefano Zampini     ierr = VecScale(lvec,-1.0);CHKERRQ(ierr);
507470cf5478SStefano Zampini     ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr);
5075e7931f94SStefano Zampini     ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr);
5076e7931f94SStefano Zampini     ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error);
5077e7931f94SStefano Zampini     ierr = VecDestroy(&rvec);CHKERRQ(ierr);
5078e7931f94SStefano Zampini     ierr = VecDestroy(&lvec);CHKERRQ(ierr);
5079e7931f94SStefano Zampini   }
508028143c3dSStefano Zampini #endif
5081e7931f94SStefano Zampini 
508228143c3dSStefano Zampini   /* assemble new additional is (if any) */
508328143c3dSStefano Zampini   if (nis) {
508428143c3dSStefano Zampini     PetscInt **temp_idxs,*count_is,j,psum;
508528143c3dSStefano Zampini 
508628143c3dSStefano Zampini     ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
5087854ce69bSBarry Smith     ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr);
508828143c3dSStefano Zampini     ptr_idxs = recv_buffer_idxs_is;
508928143c3dSStefano Zampini     psum = 0;
509028143c3dSStefano Zampini     for (i=0;i<n_recvs;i++) {
509128143c3dSStefano Zampini       for (j=0;j<nis;j++) {
509228143c3dSStefano Zampini         PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
509328143c3dSStefano Zampini         count_is[j] += plen; /* increment counting of buffer for j-th IS */
509428143c3dSStefano Zampini         psum += plen;
509528143c3dSStefano Zampini         ptr_idxs += plen+1; /* shift pointer to received data */
509628143c3dSStefano Zampini       }
509728143c3dSStefano Zampini     }
5098854ce69bSBarry Smith     ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr);
5099854ce69bSBarry Smith     ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr);
510028143c3dSStefano Zampini     for (i=1;i<nis;i++) {
510128143c3dSStefano Zampini       temp_idxs[i] = temp_idxs[i-1]+count_is[i-1];
510228143c3dSStefano Zampini     }
510328143c3dSStefano Zampini     ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr);
510428143c3dSStefano Zampini     ptr_idxs = recv_buffer_idxs_is;
510528143c3dSStefano Zampini     for (i=0;i<n_recvs;i++) {
510628143c3dSStefano Zampini       for (j=0;j<nis;j++) {
510728143c3dSStefano Zampini         PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
510828143c3dSStefano Zampini         ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr);
510928143c3dSStefano Zampini         count_is[j] += plen; /* increment starting point of buffer for j-th IS */
511028143c3dSStefano Zampini         ptr_idxs += plen+1; /* shift pointer to received data */
511128143c3dSStefano Zampini       }
511228143c3dSStefano Zampini     }
511328143c3dSStefano Zampini     for (i=0;i<nis;i++) {
511428143c3dSStefano Zampini       ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
511528143c3dSStefano Zampini       ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr);
511628143c3dSStefano Zampini       ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr);
511728143c3dSStefano Zampini     }
511828143c3dSStefano Zampini     ierr = PetscFree(count_is);CHKERRQ(ierr);
511928143c3dSStefano Zampini     ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr);
512028143c3dSStefano Zampini     ierr = PetscFree(temp_idxs);CHKERRQ(ierr);
512128143c3dSStefano Zampini   }
5122e7931f94SStefano Zampini   /* free workspace */
512328143c3dSStefano Zampini   ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr);
5124e7931f94SStefano Zampini   ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
5125e7931f94SStefano Zampini   ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr);
5126e7931f94SStefano Zampini   ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
5127e7931f94SStefano Zampini   if (isdense) {
5128e7931f94SStefano Zampini     ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr);
5129e7931f94SStefano Zampini     ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr);
5130e7931f94SStefano Zampini   } else {
5131e7931f94SStefano Zampini     /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */
5132e7931f94SStefano Zampini   }
513328143c3dSStefano Zampini   if (nis) {
513428143c3dSStefano Zampini     ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
513528143c3dSStefano Zampini     ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr);
513628143c3dSStefano Zampini   }
5137e7931f94SStefano Zampini   ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr);
5138e7931f94SStefano Zampini   ierr = PetscFree(recv_req_vals);CHKERRQ(ierr);
513928143c3dSStefano Zampini   ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr);
5140e7931f94SStefano Zampini   ierr = PetscFree(send_req_idxs);CHKERRQ(ierr);
5141e7931f94SStefano Zampini   ierr = PetscFree(send_req_vals);CHKERRQ(ierr);
514228143c3dSStefano Zampini   ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr);
5143e7931f94SStefano Zampini   ierr = PetscFree(ilengths_vals);CHKERRQ(ierr);
5144e7931f94SStefano Zampini   ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr);
5145e7931f94SStefano Zampini   ierr = PetscFree(olengths_vals);CHKERRQ(ierr);
5146e7931f94SStefano Zampini   ierr = PetscFree(olengths_idxs);CHKERRQ(ierr);
5147e7931f94SStefano Zampini   ierr = PetscFree(onodes);CHKERRQ(ierr);
514828143c3dSStefano Zampini   if (nis) {
514928143c3dSStefano Zampini     ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr);
515028143c3dSStefano Zampini     ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr);
515128143c3dSStefano Zampini     ierr = PetscFree(onodes_is);CHKERRQ(ierr);
515228143c3dSStefano Zampini   }
515328143c3dSStefano Zampini   ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr);
515428143c3dSStefano Zampini   if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */
515528143c3dSStefano Zampini     ierr = MatDestroy(mat_n);CHKERRQ(ierr);
515628143c3dSStefano Zampini     for (i=0;i<nis;i++) {
515728143c3dSStefano Zampini       ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
515828143c3dSStefano Zampini     }
515953a05cb3SStefano Zampini     *mat_n = NULL;
516028143c3dSStefano Zampini   }
5161e7931f94SStefano Zampini   PetscFunctionReturn(0);
5162e7931f94SStefano Zampini }
5163a57a6d2fSStefano Zampini 
516412edc857SStefano Zampini /* temporary hack into ksp private data structure */
5165af0996ceSBarry Smith #include <petsc/private/kspimpl.h>
516612edc857SStefano Zampini 
5167c8587f34SStefano Zampini #undef __FUNCT__
5168c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver"
5169c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals)
5170c8587f34SStefano Zampini {
5171c8587f34SStefano Zampini   PC_BDDC                *pcbddc = (PC_BDDC*)pc->data;
5172c8587f34SStefano Zampini   PC_IS                  *pcis = (PC_IS*)pc->data;
517320a2ab83SStefano Zampini   Mat                    coarse_mat,coarse_mat_is,coarse_submat_dense;
51749881197aSStefano Zampini   MatNullSpace           CoarseNullSpace=NULL;
517520a2ab83SStefano Zampini   ISLocalToGlobalMapping coarse_islg;
51766e683305SStefano Zampini   IS                     coarse_is,*isarray;
51776e683305SStefano Zampini   PetscInt               i,im_active=-1,active_procs=-1;
517830368db7SStefano Zampini   PetscInt               nis,nisdofs,nisneu,nisvert;
5179f9eb5b7dSStefano Zampini   PC                     pc_temp;
5180c8587f34SStefano Zampini   PCType                 coarse_pc_type;
5181c8587f34SStefano Zampini   KSPType                coarse_ksp_type;
5182f9eb5b7dSStefano Zampini   PetscBool              multilevel_requested,multilevel_allowed;
51834f3a063dSStefano Zampini   PetscBool              isredundant,isbddc,isnn,coarse_reuse;
51846e683305SStefano Zampini   Mat                    t_coarse_mat_is;
51856e683305SStefano Zampini   PetscInt               void_procs,ncoarse_ml,ncoarse_ds,ncoarse;
51866e683305SStefano Zampini   PetscMPIInt            all_procs;
518774e2c79eSStefano Zampini   PetscBool              csin_ml,csin_ds,csin,csin_type_simple,redist;
518868457ee5SStefano Zampini   PetscBool              compute_vecs = PETSC_FALSE;
518922bc73bbSStefano Zampini   PetscScalar            *array;
51909881197aSStefano Zampini   PetscErrorCode         ierr;
5191fdc09c96SStefano Zampini 
5192c8587f34SStefano Zampini   PetscFunctionBegin;
5193c8587f34SStefano Zampini   /* Assign global numbering to coarse dofs */
519468457ee5SStefano Zampini   if (pcbddc->new_primal_space || pcbddc->coarse_size == -1) { /* a new primal space is present or it is the first initialization, so recompute global numbering */
5195fa7f1dd8SStefano Zampini     PetscInt ocoarse_size;
51965a75c04eSSatish Balay     compute_vecs = PETSC_TRUE;
5197fa7f1dd8SStefano Zampini     ocoarse_size = pcbddc->coarse_size;
5198f4ddd8eeSStefano Zampini     ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr);
5199f4ddd8eeSStefano Zampini     ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr);
5200f4ddd8eeSStefano Zampini     /* see if we can avoid some work */
5201fa7f1dd8SStefano Zampini     if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */
520251bea450SStefano Zampini       /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */
520351bea450SStefano Zampini       if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) {
5204dc4bcba2SStefano Zampini         PC        pc;
5205dc4bcba2SStefano Zampini         PetscBool isbddc;
5206dc4bcba2SStefano Zampini 
5207dc4bcba2SStefano Zampini         /* temporary workaround since PCBDDC does not have a reset method so far */
5208dc4bcba2SStefano Zampini         ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr);
5209dc4bcba2SStefano Zampini         ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr);
5210dc4bcba2SStefano Zampini         if (isbddc) {
521163c961adSStefano Zampini           ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr);
521263c961adSStefano Zampini         } else {
5213727cdba6SStefano Zampini           ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr);
521463c961adSStefano Zampini         }
5215fa7f1dd8SStefano Zampini         coarse_reuse = PETSC_FALSE;
5216fa7f1dd8SStefano Zampini       } else { /* we can safely reuse already computed coarse matrix */
5217fa7f1dd8SStefano Zampini         coarse_reuse = PETSC_TRUE;
5218f4ddd8eeSStefano Zampini       }
5219fa7f1dd8SStefano Zampini     } else { /* there's no coarse ksp, so we need to create the coarse matrix too */
5220fa7f1dd8SStefano Zampini       coarse_reuse = PETSC_FALSE;
5221f4ddd8eeSStefano Zampini     }
522270cf5478SStefano Zampini     /* reset any subassembling information */
522370cf5478SStefano Zampini     ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr);
52246e683305SStefano Zampini     ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
52256e683305SStefano Zampini   } else { /* primal space is unchanged, so we can reuse coarse matrix */
5226fa7f1dd8SStefano Zampini     coarse_reuse = PETSC_TRUE;
5227f4ddd8eeSStefano Zampini   }
5228c8587f34SStefano Zampini 
52296e683305SStefano Zampini   /* count "active" (i.e. with positive local size) and "void" processes */
52302b510759SStefano Zampini   im_active = !!(pcis->n);
52312b510759SStefano Zampini   ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
52326e683305SStefano Zampini   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr);
52336e683305SStefano Zampini   void_procs = all_procs-active_procs;
52346e683305SStefano Zampini   csin_type_simple = PETSC_TRUE;
523574e2c79eSStefano Zampini   redist = PETSC_FALSE;
523622bc73bbSStefano Zampini   if (pcbddc->current_level && void_procs) {
52376e683305SStefano Zampini     csin_ml = PETSC_TRUE;
52386e683305SStefano Zampini     ncoarse_ml = void_procs;
5239779c1cceSStefano Zampini     /* it has no sense to redistribute on a set of processors larger than the number of active processes */
5240779c1cceSStefano Zampini     if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) {
52416e683305SStefano Zampini       csin_ds = PETSC_TRUE;
524218a45a71SStefano Zampini       ncoarse_ds = pcbddc->redistribute_coarse;
524318a45a71SStefano Zampini       redist = PETSC_TRUE;
524418a45a71SStefano Zampini     } else {
52456e683305SStefano Zampini       csin_ds = PETSC_TRUE;
5246779c1cceSStefano Zampini       ncoarse_ds = active_procs;
5247779c1cceSStefano Zampini       redist = PETSC_TRUE;
524818a45a71SStefano Zampini     }
52496e683305SStefano Zampini   } else {
52506e683305SStefano Zampini     csin_ml = PETSC_FALSE;
52516e683305SStefano Zampini     ncoarse_ml = all_procs;
52526e683305SStefano Zampini     if (void_procs) {
52536e683305SStefano Zampini       csin_ds = PETSC_TRUE;
52546e683305SStefano Zampini       ncoarse_ds = void_procs;
52556e683305SStefano Zampini       csin_type_simple = PETSC_FALSE;
52566e683305SStefano Zampini     } else {
5257779c1cceSStefano Zampini       if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) {
525874e2c79eSStefano Zampini         csin_ds = PETSC_TRUE;
525974e2c79eSStefano Zampini         ncoarse_ds = pcbddc->redistribute_coarse;
526074e2c79eSStefano Zampini         redist = PETSC_TRUE;
526174e2c79eSStefano Zampini       } else {
52626e683305SStefano Zampini         csin_ds = PETSC_FALSE;
52636e683305SStefano Zampini         ncoarse_ds = all_procs;
52646e683305SStefano Zampini       }
52656e683305SStefano Zampini     }
526674e2c79eSStefano Zampini   }
52676e683305SStefano Zampini 
52686e683305SStefano Zampini   /*
52696e683305SStefano Zampini     test if we can go multilevel: three conditions must be satisfied:
52706e683305SStefano Zampini     - we have not exceeded the number of levels requested
52716e683305SStefano Zampini     - we can actually subassemble the active processes
52726e683305SStefano Zampini     - we can find a suitable number of MPI processes where we can place the subassembled problem
52736e683305SStefano Zampini   */
52746e683305SStefano Zampini   multilevel_allowed = PETSC_FALSE;
52756e683305SStefano Zampini   multilevel_requested = PETSC_FALSE;
52766e683305SStefano Zampini   if (pcbddc->current_level < pcbddc->max_levels) {
52776e683305SStefano Zampini     multilevel_requested = PETSC_TRUE;
52786e683305SStefano Zampini     if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) {
5279f9eb5b7dSStefano Zampini       multilevel_allowed = PETSC_FALSE;
52802b510759SStefano Zampini     } else {
5281f9eb5b7dSStefano Zampini       multilevel_allowed = PETSC_TRUE;
5282c8587f34SStefano Zampini     }
5283c8587f34SStefano Zampini   }
52846e683305SStefano Zampini   /* determine number of process partecipating to coarse solver */
52856e683305SStefano Zampini   if (multilevel_allowed) {
52866e683305SStefano Zampini     ncoarse = ncoarse_ml;
52876e683305SStefano Zampini     csin = csin_ml;
528858da7f69SStefano Zampini     redist = PETSC_FALSE;
52896e683305SStefano Zampini   } else {
52906e683305SStefano Zampini     ncoarse = ncoarse_ds;
52916e683305SStefano Zampini     csin = csin_ds;
52926e683305SStefano Zampini   }
5293e7931f94SStefano Zampini 
5294abbbba34SStefano Zampini   /* creates temporary l2gmap and IS for coarse indexes */
5295abbbba34SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr);
5296abbbba34SStefano Zampini   ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr);
5297abbbba34SStefano Zampini 
5298abbbba34SStefano Zampini   /* creates temporary MATIS object for coarse matrix */
529922bc73bbSStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr);
530022bc73bbSStefano Zampini   ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr);
530122bc73bbSStefano Zampini   ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr);
530222bc73bbSStefano Zampini   ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr);
5303e176bc59SStefano Zampini   ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,NULL,&t_coarse_mat_is);CHKERRQ(ierr);
53046e683305SStefano Zampini   ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr);
53056e683305SStefano Zampini   ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
53066e683305SStefano Zampini   ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
5307abbbba34SStefano Zampini   ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr);
5308abbbba34SStefano Zampini 
53096e683305SStefano Zampini   /* compute dofs splitting and neumann boundaries for coarse dofs */
531030368db7SStefano Zampini   if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */
53116e683305SStefano Zampini     PetscInt               *tidxs,*tidxs2,nout,tsize,i;
53126e683305SStefano Zampini     const PetscInt         *idxs;
53136e683305SStefano Zampini     ISLocalToGlobalMapping tmap;
53146e683305SStefano Zampini 
53156e683305SStefano Zampini     /* create map between primal indices (in local representative ordering) and local primal numbering */
53160be93d54SStefano Zampini     ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr);
53176e683305SStefano Zampini     /* allocate space for temporary storage */
5318854ce69bSBarry Smith     ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr);
5319854ce69bSBarry Smith     ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr);
53206e683305SStefano Zampini     /* allocate for IS array */
53216e683305SStefano Zampini     nisdofs = pcbddc->n_ISForDofsLocal;
53226e683305SStefano Zampini     nisneu = !!pcbddc->NeumannBoundariesLocal;
532330368db7SStefano Zampini     nisvert = 0;
532430368db7SStefano Zampini     if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) {
532530368db7SStefano Zampini       nisvert = 1;
532630368db7SStefano Zampini     }
532730368db7SStefano Zampini     nis = nisdofs + nisneu + nisvert;
5328854ce69bSBarry Smith     ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr);
53296e683305SStefano Zampini     /* dofs splitting */
53306e683305SStefano Zampini     for (i=0;i<nisdofs;i++) {
53316e683305SStefano Zampini       /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */
53326e683305SStefano Zampini       ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr);
53336e683305SStefano Zampini       ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr);
53346e683305SStefano Zampini       ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr);
53356e683305SStefano Zampini       ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr);
53366e683305SStefano Zampini       ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr);
533730368db7SStefano Zampini       ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr);
53386e683305SStefano Zampini       /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */
53396e683305SStefano Zampini     }
53406e683305SStefano Zampini     /* neumann boundaries */
53416e683305SStefano Zampini     if (pcbddc->NeumannBoundariesLocal) {
53426e683305SStefano Zampini       /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */
53436e683305SStefano Zampini       ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr);
53446e683305SStefano Zampini       ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr);
53456e683305SStefano Zampini       ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr);
53466e683305SStefano Zampini       ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr);
53476e683305SStefano Zampini       ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr);
534830368db7SStefano Zampini       ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr);
53496e683305SStefano Zampini       /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */
53506e683305SStefano Zampini     }
535130368db7SStefano Zampini     /* primal vertices (benign) */
535230368db7SStefano Zampini     if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) {
535330368db7SStefano Zampini       ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr);
535430368db7SStefano Zampini       ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr);
535530368db7SStefano Zampini       ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr);
535630368db7SStefano Zampini       ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr);
535730368db7SStefano Zampini       ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr);
535830368db7SStefano Zampini       ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr);
535930368db7SStefano Zampini     }
53606e683305SStefano Zampini     /* free memory */
53616e683305SStefano Zampini     ierr = PetscFree(tidxs);CHKERRQ(ierr);
53626e683305SStefano Zampini     ierr = PetscFree(tidxs2);CHKERRQ(ierr);
53636e683305SStefano Zampini     ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr);
53646e683305SStefano Zampini   } else {
53656e683305SStefano Zampini     nis = 0;
53666e683305SStefano Zampini     nisdofs = 0;
53676e683305SStefano Zampini     nisneu = 0;
536830368db7SStefano Zampini     nisvert = 0;
53696e683305SStefano Zampini     isarray = NULL;
53706e683305SStefano Zampini   }
53716e683305SStefano Zampini   /* destroy no longer needed map */
53726e683305SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr);
53736e683305SStefano Zampini 
53746e683305SStefano Zampini   /* restrict on coarse candidates (if needed) */
53756e683305SStefano Zampini   coarse_mat_is = NULL;
53766e683305SStefano Zampini   if (csin) {
53776e683305SStefano Zampini     if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */
537874e2c79eSStefano Zampini       if (redist) {
537974e2c79eSStefano Zampini         PetscMPIInt rank;
5380779c1cceSStefano Zampini         PetscInt    spc,n_spc_p1,dest[1],destsize;
538174e2c79eSStefano Zampini 
538274e2c79eSStefano Zampini         ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr);
538358da7f69SStefano Zampini         spc = active_procs/ncoarse;
538458da7f69SStefano Zampini         n_spc_p1 = active_procs%ncoarse;
5385779c1cceSStefano Zampini         if (im_active) {
5386779c1cceSStefano Zampini           destsize = 1;
538774e2c79eSStefano Zampini           if (rank > n_spc_p1*(spc+1)-1) {
538874e2c79eSStefano Zampini             dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc;
538974e2c79eSStefano Zampini           } else {
539074e2c79eSStefano Zampini             dest[0] = rank/(spc+1);
539174e2c79eSStefano Zampini           }
539274e2c79eSStefano Zampini         } else {
5393779c1cceSStefano Zampini           destsize = 0;
53946e683305SStefano Zampini         }
5395779c1cceSStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
5396779c1cceSStefano Zampini       } else if (csin_type_simple) {
53976e683305SStefano Zampini         PetscMPIInt rank;
53986e683305SStefano Zampini         PetscInt    issize,isidx;
5399779c1cceSStefano Zampini 
54006e683305SStefano Zampini         ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr);
54016e683305SStefano Zampini         if (im_active) {
54026e683305SStefano Zampini           issize = 1;
54036e683305SStefano Zampini           isidx = (PetscInt)rank;
54046e683305SStefano Zampini         } else {
54056e683305SStefano Zampini           issize = 0;
54066e683305SStefano Zampini           isidx = -1;
54076e683305SStefano Zampini         }
54086e683305SStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
5409779c1cceSStefano Zampini       } else { /* get a suitable subassembling pattern from MATIS code */
5410b0c7d250SStefano Zampini         ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
54116e683305SStefano Zampini       }
5412779c1cceSStefano Zampini 
5413779c1cceSStefano Zampini       /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */
5414779c1cceSStefano Zampini       if (!redist || ncoarse <= void_procs) {
5415779c1cceSStefano Zampini         PetscInt ncoarse_cand,tissize,*nisindices;
5416779c1cceSStefano Zampini         PetscInt *coarse_candidates;
5417779c1cceSStefano Zampini         const PetscInt* tisindices;
5418779c1cceSStefano Zampini 
5419779c1cceSStefano Zampini         /* get coarse candidates' ranks in pc communicator */
5420779c1cceSStefano Zampini         ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr);
5421779c1cceSStefano Zampini         ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
5422779c1cceSStefano Zampini         for (i=0,ncoarse_cand=0;i<all_procs;i++) {
5423779c1cceSStefano Zampini           if (!coarse_candidates[i]) {
5424779c1cceSStefano Zampini             coarse_candidates[ncoarse_cand++]=i;
5425779c1cceSStefano Zampini           }
5426779c1cceSStefano Zampini         }
5427779c1cceSStefano Zampini         if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse);
5428779c1cceSStefano Zampini 
5429779c1cceSStefano Zampini 
54306e683305SStefano Zampini         if (pcbddc->dbg_flag) {
54316e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
54326e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr);
54336e683305SStefano Zampini           ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr);
54346e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr);
5435779c1cceSStefano Zampini           for (i=0;i<ncoarse_cand;i++) {
54366e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr);
54376e683305SStefano Zampini           }
54386e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr);
54396e683305SStefano Zampini           ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
54406e683305SStefano Zampini         }
54416e683305SStefano Zampini         /* shift the pattern on coarse candidates */
54426e683305SStefano Zampini         ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr);
54436e683305SStefano Zampini         ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr);
5444854ce69bSBarry Smith         ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr);
54456e683305SStefano Zampini         for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]];
54466e683305SStefano Zampini         ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr);
54476e683305SStefano Zampini         ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr);
54486e683305SStefano Zampini         ierr = PetscFree(coarse_candidates);CHKERRQ(ierr);
54496e683305SStefano Zampini       }
54506e683305SStefano Zampini       if (pcbddc->dbg_flag) {
54516e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
54526e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr);
54536e683305SStefano Zampini         ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr);
54546e683305SStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
54556e683305SStefano Zampini       }
5456779c1cceSStefano Zampini     }
54576e683305SStefano Zampini     /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */
545853a05cb3SStefano Zampini     if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */
545953a05cb3SStefano Zampini       ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr);
546053a05cb3SStefano Zampini     } else { /* this is the last level, so use just receiving processes in subcomm */
546153a05cb3SStefano Zampini       ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr);
546253a05cb3SStefano Zampini     }
54636e683305SStefano Zampini   } else {
54646e683305SStefano Zampini     if (pcbddc->dbg_flag) {
54656e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
54666e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr);
54676e683305SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
54686e683305SStefano Zampini     }
54696e683305SStefano Zampini     ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr);
54706e683305SStefano Zampini     coarse_mat_is = t_coarse_mat_is;
54716e683305SStefano Zampini   }
54726e683305SStefano Zampini 
54736e683305SStefano Zampini   /* create local to global scatters for coarse problem */
547468457ee5SStefano Zampini   if (compute_vecs) {
54756e683305SStefano Zampini     PetscInt lrows;
54766e683305SStefano Zampini     ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr);
54776e683305SStefano Zampini     if (coarse_mat_is) {
54786e683305SStefano Zampini       ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr);
54796e683305SStefano Zampini     } else {
54806e683305SStefano Zampini       lrows = 0;
54816e683305SStefano Zampini     }
54826e683305SStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr);
54836e683305SStefano Zampini     ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr);
54846e683305SStefano Zampini     ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr);
54856e683305SStefano Zampini     ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
54866e683305SStefano Zampini     ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
54876e683305SStefano Zampini   }
54886e683305SStefano Zampini   ierr = ISDestroy(&coarse_is);CHKERRQ(ierr);
54896e683305SStefano Zampini   ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr);
5490c8587f34SStefano Zampini 
5491f9eb5b7dSStefano Zampini   /* set defaults for coarse KSP and PC */
5492f9eb5b7dSStefano Zampini   if (multilevel_allowed) {
5493f9eb5b7dSStefano Zampini     coarse_ksp_type = KSPRICHARDSON;
5494f9eb5b7dSStefano Zampini     coarse_pc_type = PCBDDC;
5495f9eb5b7dSStefano Zampini   } else {
5496f9eb5b7dSStefano Zampini     coarse_ksp_type = KSPPREONLY;
5497f9eb5b7dSStefano Zampini     coarse_pc_type = PCREDUNDANT;
5498c8587f34SStefano Zampini   }
5499c8587f34SStefano Zampini 
55006e683305SStefano Zampini   /* print some info if requested */
55016e683305SStefano Zampini   if (pcbddc->dbg_flag) {
55026e683305SStefano Zampini     if (!multilevel_allowed) {
55036e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
55046e683305SStefano Zampini       if (multilevel_requested) {
55056e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Not enough active processes on level %d (active processes %d, coarsening ratio %d)\n",pcbddc->current_level,active_procs,pcbddc->coarsening_ratio);CHKERRQ(ierr);
55066e683305SStefano Zampini       } else if (pcbddc->max_levels) {
55076e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr);
55086e683305SStefano Zampini       }
55096e683305SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
55106e683305SStefano Zampini     }
55116e683305SStefano Zampini   }
55126e683305SStefano Zampini 
5513f9eb5b7dSStefano Zampini   /* create the coarse KSP object only once with defaults */
55146e683305SStefano Zampini   if (coarse_mat_is) {
55156e683305SStefano Zampini     MatReuse coarse_mat_reuse;
55166a1308c2SStefano Zampini     PetscViewer dbg_viewer = NULL;
55176e683305SStefano Zampini     if (pcbddc->dbg_flag) {
55186e683305SStefano Zampini       dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is));
55196e683305SStefano Zampini       ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr);
55206e683305SStefano Zampini     }
5521f9eb5b7dSStefano Zampini     if (!pcbddc->coarse_ksp) {
5522312be037SStefano Zampini       char prefix[256],str_level[16];
5523e604994aSStefano Zampini       size_t len;
55246e683305SStefano Zampini       ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr);
5525422a814eSBarry Smith       ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr);
5526c8587f34SStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr);
5527f9eb5b7dSStefano Zampini       ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr);
55285f76c7aeSStefano Zampini       ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr);
5529c8587f34SStefano Zampini       ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr);
55306e683305SStefano Zampini       ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr);
5531c8587f34SStefano Zampini       ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr);
5532c8587f34SStefano Zampini       ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr);
5533e604994aSStefano Zampini       /* prefix */
5534e604994aSStefano Zampini       ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr);
5535e604994aSStefano Zampini       ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr);
5536e604994aSStefano Zampini       if (!pcbddc->current_level) {
5537e604994aSStefano Zampini         ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
5538e604994aSStefano Zampini         ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr);
5539c8587f34SStefano Zampini       } else {
5540e604994aSStefano Zampini         ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr);
5541312be037SStefano Zampini         if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
5542312be037SStefano Zampini         if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
554334d6797cSStefano Zampini         ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
5544312be037SStefano Zampini         sprintf(str_level,"l%d_",(int)(pcbddc->current_level));
5545e604994aSStefano Zampini         ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr);
5546e604994aSStefano Zampini       }
5547e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr);
55483e3c6dadSStefano Zampini       /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
55493e3c6dadSStefano Zampini       ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr);
55503e3c6dadSStefano Zampini       ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr);
55513e3c6dadSStefano Zampini       ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr);
5552f9eb5b7dSStefano Zampini       /* allow user customization */
5553f9eb5b7dSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr);
55543e3c6dadSStefano Zampini     }
55553e3c6dadSStefano Zampini     /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
555651bea450SStefano Zampini     ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr);
55573e3c6dadSStefano Zampini     if (nisdofs) {
55583e3c6dadSStefano Zampini       ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr);
55593e3c6dadSStefano Zampini       for (i=0;i<nisdofs;i++) {
55603e3c6dadSStefano Zampini         ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
55613e3c6dadSStefano Zampini       }
55623e3c6dadSStefano Zampini     }
55633e3c6dadSStefano Zampini     if (nisneu) {
55643e3c6dadSStefano Zampini       ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr);
55653e3c6dadSStefano Zampini       ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr);
5566312be037SStefano Zampini     }
556730368db7SStefano Zampini     if (nisvert) {
556830368db7SStefano Zampini       ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr);
556930368db7SStefano Zampini       ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr);
557030368db7SStefano Zampini     }
5571f9eb5b7dSStefano Zampini 
5572f9eb5b7dSStefano Zampini     /* get some info after set from options */
5573f9eb5b7dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr);
5574f9eb5b7dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr);
55754f3a063dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr);
55766e683305SStefano Zampini     if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */
5577f9eb5b7dSStefano Zampini       ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr);
5578f9eb5b7dSStefano Zampini       isbddc = PETSC_FALSE;
5579f9eb5b7dSStefano Zampini     }
558039f0f02cSStefano Zampini     ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
55814f3a063dSStefano Zampini     if (isredundant) {
55824f3a063dSStefano Zampini       KSP inner_ksp;
55834f3a063dSStefano Zampini       PC  inner_pc;
55844f3a063dSStefano Zampini       ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr);
55854f3a063dSStefano Zampini       ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr);
55864f3a063dSStefano Zampini       ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr);
55874f3a063dSStefano Zampini     }
5588f9eb5b7dSStefano Zampini 
5589f9eb5b7dSStefano Zampini     /* assemble coarse matrix */
5590fa7f1dd8SStefano Zampini     if (coarse_reuse) {
559181d9aea3SBarry Smith       ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr);
5592fa7f1dd8SStefano Zampini       ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr);
55936e683305SStefano Zampini       coarse_mat_reuse = MAT_REUSE_MATRIX;
5594fa7f1dd8SStefano Zampini     } else {
55956e683305SStefano Zampini       coarse_mat_reuse = MAT_INITIAL_MATRIX;
5596fa7f1dd8SStefano Zampini     }
5597c8587f34SStefano Zampini     if (isbddc || isnn) {
5598d4d8cf7bSStefano Zampini       if (isbddc) { /* currently there's no API for this */
5599d4d8cf7bSStefano Zampini         PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data;
5600d4d8cf7bSStefano Zampini         pcbddc->detect_disconnected = PETSC_TRUE;
5601d4d8cf7bSStefano Zampini       }
560222bc73bbSStefano Zampini       if (pcbddc->coarsening_ratio > 1) {
560370cf5478SStefano Zampini         if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */
5604b0c7d250SStefano Zampini           ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr);
560522b6e8a2SStefano Zampini           if (pcbddc->dbg_flag) {
56066e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
56076e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr);
56086e683305SStefano Zampini             ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr);
56096e683305SStefano Zampini             ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr);
561022b6e8a2SStefano Zampini           }
561170cf5478SStefano Zampini         }
561253a05cb3SStefano Zampini         ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr);
561370cf5478SStefano Zampini       } else {
561422bc73bbSStefano Zampini         ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr);
561522bc73bbSStefano Zampini         coarse_mat = coarse_mat_is;
561622bc73bbSStefano Zampini       }
561722bc73bbSStefano Zampini     } else {
56182e1e5fa4SStefano Zampini       ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr);
5619c8587f34SStefano Zampini     }
5620c8587f34SStefano Zampini     ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr);
5621c8587f34SStefano Zampini 
56223301b35fSStefano Zampini     /* propagate symmetry info of coarse matrix */
56235a16e3a0SStefano Zampini     ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr);
56243301b35fSStefano Zampini     if (pc->pmat->symmetric_set) {
56253301b35fSStefano Zampini       ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr);
56263301b35fSStefano Zampini     }
56273301b35fSStefano Zampini     if (pc->pmat->hermitian_set) {
56283301b35fSStefano Zampini       ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr);
56293301b35fSStefano Zampini     }
56303301b35fSStefano Zampini     if (pc->pmat->spd_set) {
56313301b35fSStefano Zampini       ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr);
56323301b35fSStefano Zampini     }
56336e683305SStefano Zampini     /* set operators */
56345f76c7aeSStefano Zampini     ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr);
56356e683305SStefano Zampini     if (pcbddc->dbg_flag) {
56366e683305SStefano Zampini       ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr);
56376e683305SStefano Zampini     }
56386e683305SStefano Zampini   } else { /* processes non partecipating to coarse solver (if any) */
56396e683305SStefano Zampini     coarse_mat = 0;
56406e683305SStefano Zampini   }
56416e683305SStefano Zampini   ierr = PetscFree(isarray);CHKERRQ(ierr);
5642b1ecc7b1SStefano Zampini #if 0
5643b9b85e73SStefano Zampini   {
5644b9b85e73SStefano Zampini     PetscViewer viewer;
5645b9b85e73SStefano Zampini     char filename[256];
5646b1ecc7b1SStefano Zampini     sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level);
5647b1ecc7b1SStefano Zampini     ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr);
5648b9b85e73SStefano Zampini     ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
5649b9b85e73SStefano Zampini     ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr);
5650b9b85e73SStefano Zampini     ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr);
5651b9b85e73SStefano Zampini   }
5652b9b85e73SStefano Zampini #endif
5653c8587f34SStefano Zampini 
5654c8587f34SStefano Zampini   /* Compute coarse null space (special handling by BDDC only) */
5655298c0119SStefano Zampini #if 0
5656c8587f34SStefano Zampini   if (pcbddc->NullSpace) {
5657c8587f34SStefano Zampini     ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr);
565898a51de6SStefano Zampini   }
5659298c0119SStefano Zampini #endif
5660b0f5fe93SStefano Zampini   /* hack */
566198a51de6SStefano Zampini   if (pcbddc->coarse_ksp) {
566298a51de6SStefano Zampini     Vec crhs,csol;
566304708bb6SStefano Zampini 
5664f347579bSStefano Zampini     ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr);
5665f347579bSStefano Zampini     ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr);
5666f347579bSStefano Zampini     if (!csol) {
56672a7a6963SBarry Smith       ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr);
5668f9eb5b7dSStefano Zampini     }
5669f347579bSStefano Zampini     if (!crhs) {
56702a7a6963SBarry Smith       ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr);
5671f347579bSStefano Zampini     }
5672b0f5fe93SStefano Zampini   }
5673b0f5fe93SStefano Zampini 
5674b0f5fe93SStefano Zampini   /* compute null space for coarse solver if the benign trick has been requested */
5675b0f5fe93SStefano Zampini   if (pcbddc->benign_null) {
5676b0f5fe93SStefano Zampini 
5677b0f5fe93SStefano Zampini     ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr);
56784f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
56794f1b2e48SStefano Zampini       ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr);
56804f1b2e48SStefano Zampini     }
5681b0f5fe93SStefano Zampini     ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr);
5682b0f5fe93SStefano Zampini     ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr);
5683b0f5fe93SStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5684b0f5fe93SStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5685b0f5fe93SStefano Zampini     if (coarse_mat) {
5686b0f5fe93SStefano Zampini       Vec         nullv;
5687b0f5fe93SStefano Zampini       PetscScalar *array,*array2;
5688b0f5fe93SStefano Zampini       PetscInt    nl;
5689b0f5fe93SStefano Zampini 
5690b0f5fe93SStefano Zampini       ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr);
5691b0f5fe93SStefano Zampini       ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr);
5692b0f5fe93SStefano Zampini       ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr);
5693b0f5fe93SStefano Zampini       ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr);
5694b0f5fe93SStefano Zampini       ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr);
5695b0f5fe93SStefano Zampini       ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr);
5696b0f5fe93SStefano Zampini       ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr);
5697b0f5fe93SStefano Zampini       ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr);
5698b0f5fe93SStefano Zampini       ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr);
5699b0f5fe93SStefano Zampini       ierr = VecDestroy(&nullv);CHKERRQ(ierr);
5700b0f5fe93SStefano Zampini     }
5701b0f5fe93SStefano Zampini   }
5702b0f5fe93SStefano Zampini 
5703b0f5fe93SStefano Zampini   if (pcbddc->coarse_ksp) {
5704b0f5fe93SStefano Zampini     PetscBool ispreonly;
5705b0f5fe93SStefano Zampini 
5706b0f5fe93SStefano Zampini     if (CoarseNullSpace) {
5707b0f5fe93SStefano Zampini       PetscBool isnull;
5708b0f5fe93SStefano Zampini       ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr);
5709d4d8cf7bSStefano Zampini       if (0) {
571030368db7SStefano Zampini         if (isbddc && !pcbddc->benign_saddle_point) {
5711b0f5fe93SStefano Zampini           ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr);
5712b0f5fe93SStefano Zampini         } else {
5713b0f5fe93SStefano Zampini           ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr);
5714b0f5fe93SStefano Zampini         }
5715b0f5fe93SStefano Zampini       } else {
5716b0f5fe93SStefano Zampini         ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr);
5717b0f5fe93SStefano Zampini       }
5718b0f5fe93SStefano Zampini     }
5719b0f5fe93SStefano Zampini     /* setup coarse ksp */
5720b0f5fe93SStefano Zampini     ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr);
5721cbcc2c2aSStefano Zampini     /* Check coarse problem if in debug mode or if solving with an iterative method */
5722cbcc2c2aSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr);
57236e683305SStefano Zampini     if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) {
5724c8587f34SStefano Zampini       KSP       check_ksp;
57252b510759SStefano Zampini       KSPType   check_ksp_type;
5726c8587f34SStefano Zampini       PC        check_pc;
57276e683305SStefano Zampini       Vec       check_vec,coarse_vec;
57286a1308c2SStefano Zampini       PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0;
57292b510759SStefano Zampini       PetscInt  its;
57306e683305SStefano Zampini       PetscBool compute_eigs;
57316e683305SStefano Zampini       PetscReal *eigs_r,*eigs_c;
57326e683305SStefano Zampini       PetscInt  neigs;
57338e185a42SStefano Zampini       const char *prefix;
5734c8587f34SStefano Zampini 
57352b510759SStefano Zampini       /* Create ksp object suitable for estimation of extreme eigenvalues */
57366e683305SStefano Zampini       ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr);
5737422a814eSBarry Smith       ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr);
573823ee1639SBarry Smith       ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr);
5739f4ddd8eeSStefano Zampini       ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr);
57402b510759SStefano Zampini       if (ispreonly) {
57412b510759SStefano Zampini         check_ksp_type = KSPPREONLY;
57426e683305SStefano Zampini         compute_eigs = PETSC_FALSE;
57432b510759SStefano Zampini       } else {
5744cbcc2c2aSStefano Zampini         check_ksp_type = KSPGMRES;
57456e683305SStefano Zampini         compute_eigs = PETSC_TRUE;
5746c8587f34SStefano Zampini       }
5747c8587f34SStefano Zampini       ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr);
57486e683305SStefano Zampini       ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr);
57496e683305SStefano Zampini       ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr);
57506e683305SStefano Zampini       ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr);
5751a7dc3881SStefano Zampini       ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr);
5752a7dc3881SStefano Zampini       ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr);
5753a7dc3881SStefano Zampini       ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr);
5754a7dc3881SStefano Zampini       ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr);
5755c8587f34SStefano Zampini       ierr = KSPSetUp(check_ksp);CHKERRQ(ierr);
5756c8587f34SStefano Zampini       ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr);
5757c8587f34SStefano Zampini       ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr);
5758c8587f34SStefano Zampini       /* create random vec */
57596e683305SStefano Zampini       ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr);
57606e683305SStefano Zampini       ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr);
5761c8587f34SStefano Zampini       ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr);
5762c8587f34SStefano Zampini       if (CoarseNullSpace) {
5763c8587f34SStefano Zampini         ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr);
5764c8587f34SStefano Zampini       }
57656e683305SStefano Zampini       ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr);
5766c8587f34SStefano Zampini       /* solve coarse problem */
57676e683305SStefano Zampini       ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr);
5768c8587f34SStefano Zampini       if (CoarseNullSpace) {
57696e683305SStefano Zampini         ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr);
5770c8587f34SStefano Zampini       }
5771cbcc2c2aSStefano Zampini       /* set eigenvalue estimation if preonly has not been requested */
57726e683305SStefano Zampini       if (compute_eigs) {
5773854ce69bSBarry Smith         ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr);
5774854ce69bSBarry Smith         ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr);
57756e683305SStefano Zampini         ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr);
57766e683305SStefano Zampini         lambda_max = eigs_r[neigs-1];
57776e683305SStefano Zampini         lambda_min = eigs_r[0];
57786e683305SStefano Zampini         if (pcbddc->use_coarse_estimates) {
57796e683305SStefano Zampini           if (lambda_max>lambda_min) {
5780cbcc2c2aSStefano Zampini             ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr);
5781cbcc2c2aSStefano Zampini             ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr);
5782cbcc2c2aSStefano Zampini           }
5783c8587f34SStefano Zampini         }
5784c8587f34SStefano Zampini       }
5785cbcc2c2aSStefano Zampini 
5786c8587f34SStefano Zampini       /* check coarse problem residual error */
57876e683305SStefano Zampini       if (pcbddc->dbg_flag) {
57886e683305SStefano Zampini         PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp));
57896e683305SStefano Zampini         ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr);
57906e683305SStefano Zampini         ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr);
5791c8587f34SStefano Zampini         ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr);
57926e683305SStefano Zampini         ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr);
57936e683305SStefano Zampini         ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr);
5794c8587f34SStefano Zampini         ierr = VecDestroy(&check_vec);CHKERRQ(ierr);
5795779c1cceSStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr);
57966e683305SStefano Zampini         ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr);
57976e683305SStefano Zampini         ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr);
57986e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error   : %1.6e\n",infty_error);CHKERRQ(ierr);
57996e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr);
5800b0f5fe93SStefano Zampini         if (CoarseNullSpace) {
5801b0f5fe93SStefano Zampini           ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr);
5802b0f5fe93SStefano Zampini         }
58036e683305SStefano Zampini         if (compute_eigs) {
58046e683305SStefano Zampini           PetscReal lambda_max_s,lambda_min_s;
5805deec49d1SStefano Zampini           ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr);
5806c8587f34SStefano Zampini           ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr);
58076e683305SStefano Zampini           ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr);
58086e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,lambda_min,lambda_max,lambda_min_s,lambda_max_s);CHKERRQ(ierr);
58096e683305SStefano Zampini           for (i=0;i<neigs;i++) {
58106e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr);
5811c8587f34SStefano Zampini           }
58126e683305SStefano Zampini         }
58136e683305SStefano Zampini         ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr);
58146e683305SStefano Zampini         ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr);
58156e683305SStefano Zampini       }
5816c8587f34SStefano Zampini       ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr);
58176e683305SStefano Zampini       if (compute_eigs) {
58186e683305SStefano Zampini         ierr = PetscFree(eigs_r);CHKERRQ(ierr);
58196e683305SStefano Zampini         ierr = PetscFree(eigs_c);CHKERRQ(ierr);
5820c8587f34SStefano Zampini       }
58216e683305SStefano Zampini     }
58226e683305SStefano Zampini   }
5823cbcc2c2aSStefano Zampini   /* print additional info */
5824cbcc2c2aSStefano Zampini   if (pcbddc->dbg_flag) {
58256e683305SStefano Zampini     /* waits until all processes reaches this point */
58266e683305SStefano Zampini     ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr);
5827cbcc2c2aSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr);
5828cbcc2c2aSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
5829cbcc2c2aSStefano Zampini   }
5830cbcc2c2aSStefano Zampini 
58312b510759SStefano Zampini   /* free memory */
5832c8587f34SStefano Zampini   ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr);
5833fdc635d7SStefano Zampini   ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr);
5834c8587f34SStefano Zampini   PetscFunctionReturn(0);
5835c8587f34SStefano Zampini }
5836674ae819SStefano Zampini 
5837f34684f1SStefano Zampini #undef __FUNCT__
5838f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering"
5839f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n)
5840f34684f1SStefano Zampini {
5841f34684f1SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
5842f34684f1SStefano Zampini   PC_IS*         pcis = (PC_IS*)pc->data;
5843f34684f1SStefano Zampini   Mat_IS*        matis = (Mat_IS*)pc->pmat->data;
5844dc456d91SStefano Zampini   IS             subset,subset_mult,subset_n;
5845dc456d91SStefano Zampini   PetscInt       local_size,coarse_size=0;
584673be2a3aSStefano Zampini   PetscInt       *local_primal_indices=NULL;
5847dc456d91SStefano Zampini   const PetscInt *t_local_primal_indices;
5848f34684f1SStefano Zampini   PetscErrorCode ierr;
5849f34684f1SStefano Zampini 
5850f34684f1SStefano Zampini   PetscFunctionBegin;
5851f34684f1SStefano Zampini   /* Compute global number of coarse dofs */
58520e6343abSStefano Zampini   if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) {
58530e6343abSStefano Zampini     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first");
5854727cdba6SStefano Zampini   }
5855dc456d91SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr);
58563bbff08aSStefano Zampini   ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr);
5857dc456d91SStefano Zampini   ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
5858dc456d91SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr);
5859dc456d91SStefano Zampini   ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr);
5860dc456d91SStefano Zampini   ierr = ISDestroy(&subset);CHKERRQ(ierr);
5861dc456d91SStefano Zampini   ierr = ISDestroy(&subset_mult);CHKERRQ(ierr);
5862dc456d91SStefano Zampini   ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr);
58630e6343abSStefano Zampini   if (local_size != pcbddc->local_primal_size) {
58640e6343abSStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size);
58650e6343abSStefano Zampini   }
5866dc456d91SStefano Zampini   ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr);
5867dc456d91SStefano Zampini   ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr);
5868dc456d91SStefano Zampini   ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr);
5869dc456d91SStefano Zampini   ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr);
5870dc456d91SStefano Zampini   ierr = ISDestroy(&subset_n);CHKERRQ(ierr);
5871f34684f1SStefano Zampini 
5872f34684f1SStefano Zampini   /* check numbering */
5873f34684f1SStefano Zampini   if (pcbddc->dbg_flag) {
5874019a44ceSStefano Zampini     PetscScalar coarsesum,*array,*array2;
5875dc456d91SStefano Zampini     PetscInt    i;
5876b9b85e73SStefano Zampini     PetscBool   set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE;
5877f34684f1SStefano Zampini 
5878f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
5879f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
5880f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr);
58811575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
5882019a44ceSStefano Zampini     /* counter */
5883019a44ceSStefano Zampini     ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
5884019a44ceSStefano Zampini     ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr);
5885019a44ceSStefano Zampini     ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5886019a44ceSStefano Zampini     ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5887019a44ceSStefano Zampini     ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5888019a44ceSStefano Zampini     ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5889f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr);
5890f34684f1SStefano Zampini     for (i=0;i<pcbddc->local_primal_size;i++) {
5891727cdba6SStefano Zampini       ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr);
5892f34684f1SStefano Zampini     }
5893f34684f1SStefano Zampini     ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr);
5894f34684f1SStefano Zampini     ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr);
5895f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
5896e176bc59SStefano Zampini     ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5897e176bc59SStefano Zampini     ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5898e176bc59SStefano Zampini     ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5899e176bc59SStefano Zampini     ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5900f34684f1SStefano Zampini     ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
5901019a44ceSStefano Zampini     ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr);
5902f34684f1SStefano Zampini     for (i=0;i<pcis->n;i++) {
5903019a44ceSStefano Zampini       if (array[i] != 0.0 && array[i] != array2[i]) {
590475c01103SStefano Zampini         PetscInt owned = (PetscInt)PetscRealPart(array[i]);
590575c01103SStefano Zampini         PetscInt neigh = (PetscInt)PetscRealPart(array2[i]);
5906b9b85e73SStefano Zampini         set_error = PETSC_TRUE;
5907019a44ceSStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a %d processes instead of %d!\n",PetscGlobalRank,i,owned,neigh);CHKERRQ(ierr);
5908f34684f1SStefano Zampini       }
5909f34684f1SStefano Zampini     }
5910019a44ceSStefano Zampini     ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr);
5911b9b85e73SStefano Zampini     ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
5912f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
5913f34684f1SStefano Zampini     for (i=0;i<pcis->n;i++) {
5914f34684f1SStefano Zampini       if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]);
5915f34684f1SStefano Zampini     }
5916f34684f1SStefano Zampini     ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
5917f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
5918e176bc59SStefano Zampini     ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5919e176bc59SStefano Zampini     ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
5920f34684f1SStefano Zampini     ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr);
5921f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr);
5922b9b85e73SStefano Zampini     if (pcbddc->dbg_flag > 1 || set_error_reduced) {
5923ca8b9ea9SStefano Zampini       PetscInt *gidxs;
5924ca8b9ea9SStefano Zampini 
5925ca8b9ea9SStefano Zampini       ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr);
59263bbff08aSStefano Zampini       ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr);
5927f34684f1SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr);
5928f34684f1SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
5929f34684f1SStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr);
5930f34684f1SStefano Zampini       for (i=0;i<pcbddc->local_primal_size;i++) {
59314bc2dc4bSStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%d]=%d (%d,%d)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i],gidxs[i]);CHKERRQ(ierr);
5932f34684f1SStefano Zampini       }
5933f34684f1SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
5934ca8b9ea9SStefano Zampini       ierr = PetscFree(gidxs);CHKERRQ(ierr);
5935f34684f1SStefano Zampini     }
5936f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
59371575c14dSBarry Smith     ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr);
5938302440fdSBarry Smith     if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed");
5939f34684f1SStefano Zampini   }
59408bec7fa6SStefano Zampini   /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */
5941f34684f1SStefano Zampini   /* get back data */
5942f34684f1SStefano Zampini   *coarse_size_n = coarse_size;
5943f34684f1SStefano Zampini   *local_primal_indices_n = local_primal_indices;
5944674ae819SStefano Zampini   PetscFunctionReturn(0);
5945674ae819SStefano Zampini }
5946674ae819SStefano Zampini 
5947e456f2a8SStefano Zampini #undef __FUNCT__
5948e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal"
5949a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis)
5950e456f2a8SStefano Zampini {
5951e456f2a8SStefano Zampini   IS             localis_t;
5952a7dc3881SStefano Zampini   PetscInt       i,lsize,*idxs,n;
5953e456f2a8SStefano Zampini   PetscScalar    *vals;
5954e456f2a8SStefano Zampini   PetscErrorCode ierr;
5955e456f2a8SStefano Zampini 
5956e456f2a8SStefano Zampini   PetscFunctionBegin;
5957a7dc3881SStefano Zampini   /* get indices in local ordering exploiting local to global map */
5958e456f2a8SStefano Zampini   ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr);
5959854ce69bSBarry Smith   ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr);
5960e456f2a8SStefano Zampini   for (i=0;i<lsize;i++) vals[i] = 1.0;
5961e456f2a8SStefano Zampini   ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr);
5962a7dc3881SStefano Zampini   ierr = VecSet(gwork,0.0);CHKERRQ(ierr);
5963a7dc3881SStefano Zampini   ierr = VecSet(lwork,0.0);CHKERRQ(ierr);
59641035eff8SStefano Zampini   if (idxs) { /* multilevel guard */
5965a7dc3881SStefano Zampini     ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
59661035eff8SStefano Zampini   }
5967a7dc3881SStefano Zampini   ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr);
5968e456f2a8SStefano Zampini   ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr);
5969e456f2a8SStefano Zampini   ierr = PetscFree(vals);CHKERRQ(ierr);
5970a7dc3881SStefano Zampini   ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr);
5971a7dc3881SStefano Zampini   /* now compute set in local ordering */
5972a7dc3881SStefano Zampini   ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5973a7dc3881SStefano Zampini   ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
5974a7dc3881SStefano Zampini   ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr);
5975a7dc3881SStefano Zampini   ierr = VecGetSize(lwork,&n);CHKERRQ(ierr);
5976a7dc3881SStefano Zampini   for (i=0,lsize=0;i<n;i++) {
5977ff92baa0SMatthew G. Knepley     if (PetscRealPart(vals[i]) > 0.5) {
5978e456f2a8SStefano Zampini       lsize++;
5979e456f2a8SStefano Zampini     }
5980e456f2a8SStefano Zampini   }
5981854ce69bSBarry Smith   ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr);
5982a7dc3881SStefano Zampini   for (i=0,lsize=0;i<n;i++) {
5983ff92baa0SMatthew G. Knepley     if (PetscRealPart(vals[i]) > 0.5) {
5984e456f2a8SStefano Zampini       idxs[lsize++] = i;
5985e456f2a8SStefano Zampini     }
5986e456f2a8SStefano Zampini   }
5987a7dc3881SStefano Zampini   ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr);
5988a7dc3881SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr);
5989e456f2a8SStefano Zampini   *localis = localis_t;
5990e456f2a8SStefano Zampini   PetscFunctionReturn(0);
5991e456f2a8SStefano Zampini }
5992906d46d4SStefano Zampini 
5993906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */
5994906d46d4SStefano Zampini #undef __FUNCT__
5995906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private"
5996906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y)
5997906d46d4SStefano Zampini {
5998906d46d4SStefano Zampini   PCBDDCChange_ctx change_ctx;
5999906d46d4SStefano Zampini   PetscErrorCode   ierr;
6000906d46d4SStefano Zampini 
6001906d46d4SStefano Zampini   PetscFunctionBegin;
6002906d46d4SStefano Zampini   ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr);
6003906d46d4SStefano Zampini   ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr);
6004906d46d4SStefano Zampini   ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr);
6005906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr);
6006906d46d4SStefano Zampini   PetscFunctionReturn(0);
6007906d46d4SStefano Zampini }
6008906d46d4SStefano Zampini 
6009906d46d4SStefano Zampini #undef __FUNCT__
6010906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private"
6011906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y)
6012906d46d4SStefano Zampini {
6013906d46d4SStefano Zampini   PCBDDCChange_ctx change_ctx;
6014906d46d4SStefano Zampini   PetscErrorCode   ierr;
6015906d46d4SStefano Zampini 
6016906d46d4SStefano Zampini   PetscFunctionBegin;
6017906d46d4SStefano Zampini   ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr);
6018906d46d4SStefano Zampini   ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr);
6019906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr);
6020906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr);
6021906d46d4SStefano Zampini   PetscFunctionReturn(0);
6022906d46d4SStefano Zampini }
6023b96c3477SStefano Zampini 
6024b96c3477SStefano Zampini #undef __FUNCT__
6025b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs"
602608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc)
6027b96c3477SStefano Zampini {
6028a64f4aa4SStefano Zampini   PC_IS               *pcis=(PC_IS*)pc->data;
6029b96c3477SStefano Zampini   PC_BDDC             *pcbddc=(PC_BDDC*)pc->data;
6030b96c3477SStefano Zampini   PCBDDCSubSchurs     sub_schurs=pcbddc->sub_schurs;
6031a64f4aa4SStefano Zampini   Mat                 S_j;
6032b96c3477SStefano Zampini   PetscInt            *used_xadj,*used_adjncy;
6033b96c3477SStefano Zampini   PetscBool           free_used_adj;
6034b96c3477SStefano Zampini   PetscErrorCode      ierr;
6035b96c3477SStefano Zampini 
6036b96c3477SStefano Zampini   PetscFunctionBegin;
6037b96c3477SStefano Zampini   /* decide the adjacency to be used for determining internal problems for local schur on subsets */
6038b96c3477SStefano Zampini   free_used_adj = PETSC_FALSE;
603908122e43SStefano Zampini   if (pcbddc->sub_schurs_layers == -1) {
6040b96c3477SStefano Zampini     used_xadj = NULL;
6041b96c3477SStefano Zampini     used_adjncy = NULL;
6042b96c3477SStefano Zampini   } else {
604308122e43SStefano Zampini     if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) {
604408122e43SStefano Zampini       used_xadj = pcbddc->mat_graph->xadj;
604508122e43SStefano Zampini       used_adjncy = pcbddc->mat_graph->adjncy;
604608122e43SStefano Zampini     } else if (pcbddc->computed_rowadj) {
6047b96c3477SStefano Zampini       used_xadj = pcbddc->mat_graph->xadj;
6048b96c3477SStefano Zampini       used_adjncy = pcbddc->mat_graph->adjncy;
6049b96c3477SStefano Zampini     } else {
60502fffb893SStefano Zampini       PetscBool      flg_row=PETSC_FALSE;
6051b96c3477SStefano Zampini       const PetscInt *xadj,*adjncy;
6052b96c3477SStefano Zampini       PetscInt       nvtxs;
6053b96c3477SStefano Zampini 
60542fffb893SStefano Zampini       ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
60552fffb893SStefano Zampini       if (flg_row) {
6056b96c3477SStefano Zampini         ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr);
6057b96c3477SStefano Zampini         ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr);
6058b96c3477SStefano Zampini         ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr);
6059b96c3477SStefano Zampini         free_used_adj = PETSC_TRUE;
60602fffb893SStefano Zampini       } else {
60612fffb893SStefano Zampini         pcbddc->sub_schurs_layers = -1;
60622fffb893SStefano Zampini         used_xadj = NULL;
60632fffb893SStefano Zampini         used_adjncy = NULL;
60642fffb893SStefano Zampini       }
60652fffb893SStefano Zampini       ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
6066b96c3477SStefano Zampini     }
6067b96c3477SStefano Zampini   }
6068d5574798SStefano Zampini 
6069d5574798SStefano Zampini   /* setup sub_schurs data */
6070a64f4aa4SStefano Zampini   ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr);
6071df4d28bfSStefano Zampini   if (!sub_schurs->schur_explicit) {
6072df4d28bfSStefano Zampini     /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */
6073a64f4aa4SStefano Zampini     ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr);
6074ca92afb2SStefano Zampini     ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL);CHKERRQ(ierr);
6075a64f4aa4SStefano Zampini   } else {
60766816873aSStefano Zampini     PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis;
607704708bb6SStefano Zampini     PetscBool isseqaij;
6078a3df083aSStefano Zampini     PetscInt  benign_n;
6079a3df083aSStefano Zampini 
60805feab87aSStefano Zampini     if (!pcbddc->use_vertices && reuse_solvers) {
60815feab87aSStefano Zampini       PetscInt n_vertices;
60825feab87aSStefano Zampini 
60835feab87aSStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr);
60842034aafcSStefano Zampini       reuse_solvers = (PetscBool)!n_vertices;
60855feab87aSStefano Zampini     }
608604708bb6SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
608704708bb6SStefano Zampini     if (!isseqaij) {
608804708bb6SStefano Zampini       Mat_IS* matis = (Mat_IS*)pc->pmat->data;
608904708bb6SStefano Zampini       if (matis->A == pcbddc->local_mat) {
609004708bb6SStefano Zampini         ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
609104708bb6SStefano Zampini         ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
609204708bb6SStefano Zampini       } else {
609304708bb6SStefano Zampini         ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr);
609404708bb6SStefano Zampini       }
609504708bb6SStefano Zampini     }
6096a3df083aSStefano Zampini     if (!pcbddc->benign_change_explicit) {
6097a3df083aSStefano Zampini       benign_n = pcbddc->benign_n;
6098ca92afb2SStefano Zampini     } else {
6099a3df083aSStefano Zampini       benign_n = 0;
6100ca92afb2SStefano Zampini     }
6101a3df083aSStefano Zampini     ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs);CHKERRQ(ierr);
6102ca92afb2SStefano Zampini   }
6103d12d3064SStefano Zampini   ierr = MatDestroy(&S_j);CHKERRQ(ierr);
6104b96c3477SStefano Zampini 
6105b96c3477SStefano Zampini   /* free adjacency */
6106b96c3477SStefano Zampini   if (free_used_adj) {
6107b96c3477SStefano Zampini     ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr);
6108b96c3477SStefano Zampini   }
6109b96c3477SStefano Zampini   PetscFunctionReturn(0);
6110b96c3477SStefano Zampini }
6111b96c3477SStefano Zampini 
6112b96c3477SStefano Zampini #undef __FUNCT__
6113b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs"
611408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc)
6115b96c3477SStefano Zampini {
6116b96c3477SStefano Zampini   PC_IS               *pcis=(PC_IS*)pc->data;
6117b96c3477SStefano Zampini   PC_BDDC             *pcbddc=(PC_BDDC*)pc->data;
6118b96c3477SStefano Zampini   PCBDDCSubSchurs     sub_schurs=pcbddc->sub_schurs;
6119b96c3477SStefano Zampini   PCBDDCGraph         graph;
6120b96c3477SStefano Zampini   PetscErrorCode      ierr;
6121b96c3477SStefano Zampini 
6122b96c3477SStefano Zampini   PetscFunctionBegin;
6123b96c3477SStefano Zampini   /* attach interface graph for determining subsets */
612408122e43SStefano Zampini   if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */
61253301b35fSStefano Zampini     IS       verticesIS,verticescomm;
61263301b35fSStefano Zampini     PetscInt vsize,*idxs;
6127b96c3477SStefano Zampini 
6128b96c3477SStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr);
61293301b35fSStefano Zampini     ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr);
61303301b35fSStefano Zampini     ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr);
61313301b35fSStefano Zampini     ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr);
61323301b35fSStefano Zampini     ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr);
61333301b35fSStefano Zampini     ierr = ISDestroy(&verticesIS);CHKERRQ(ierr);
6134b96c3477SStefano Zampini     ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr);
61357fb0e2dbSStefano Zampini     ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr);
61363301b35fSStefano Zampini     ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr);
61373301b35fSStefano Zampini     ierr = ISDestroy(&verticescomm);CHKERRQ(ierr);
6138b96c3477SStefano Zampini     ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr);
6139b96c3477SStefano Zampini /*
6140b96c3477SStefano Zampini     if (pcbddc->dbg_flag) {
6141b96c3477SStefano Zampini       ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr);
6142b96c3477SStefano Zampini     }
6143b96c3477SStefano Zampini */
6144b96c3477SStefano Zampini   } else {
6145b96c3477SStefano Zampini     graph = pcbddc->mat_graph;
6146b96c3477SStefano Zampini   }
6147b96c3477SStefano Zampini 
6148b96c3477SStefano Zampini   /* sub_schurs init */
6149a64f4aa4SStefano Zampini   ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr);
6150a64f4aa4SStefano Zampini 
6151b96c3477SStefano Zampini   /* free graph struct */
615208122e43SStefano Zampini   if (pcbddc->sub_schurs_rebuild) {
6153b96c3477SStefano Zampini     ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr);
6154b96c3477SStefano Zampini   }
6155b96c3477SStefano Zampini   PetscFunctionReturn(0);
6156b96c3477SStefano Zampini }
6157fa34dd3eSStefano Zampini 
6158fa34dd3eSStefano Zampini #undef __FUNCT__
6159fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator"
6160fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc)
6161fa34dd3eSStefano Zampini {
6162fa34dd3eSStefano Zampini   PC_IS               *pcis=(PC_IS*)pc->data;
6163fa34dd3eSStefano Zampini   PC_BDDC             *pcbddc=(PC_BDDC*)pc->data;
6164fa34dd3eSStefano Zampini   PetscErrorCode      ierr;
6165fa34dd3eSStefano Zampini 
6166fa34dd3eSStefano Zampini   PetscFunctionBegin;
6167fa34dd3eSStefano Zampini   if (pcbddc->n_vertices == pcbddc->local_primal_size) {
6168fa34dd3eSStefano Zampini     IS             zerodiag = NULL;
61694f1b2e48SStefano Zampini     Mat            S_j,B0_B=NULL;
6170fa34dd3eSStefano Zampini     Vec            dummy_vec=NULL,vec_check_B,vec_scale_P;
61714f1b2e48SStefano Zampini     PetscScalar    *p0_check,*array,*array2;
617275c01103SStefano Zampini     PetscReal      norm;
6173fa34dd3eSStefano Zampini     PetscInt       i;
6174fa34dd3eSStefano Zampini 
6175fa34dd3eSStefano Zampini     /* B0 and B0_B */
6176fa34dd3eSStefano Zampini     if (zerodiag) {
6177fa34dd3eSStefano Zampini       IS       dummy;
6178fa34dd3eSStefano Zampini 
61794f1b2e48SStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr);
61804f1b2e48SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr);
6181fa34dd3eSStefano Zampini       ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr);
6182fa34dd3eSStefano Zampini       ierr = ISDestroy(&dummy);CHKERRQ(ierr);
6183fa34dd3eSStefano Zampini     }
6184fa34dd3eSStefano Zampini     /* I need a primal vector to scale primal nodes since BDDC sums contibutions */
6185fa34dd3eSStefano Zampini     ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr);
6186fa34dd3eSStefano Zampini     ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr);
6187fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6188fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6189fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6190fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6191fa34dd3eSStefano Zampini     ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr);
6192fa34dd3eSStefano Zampini     /* S_j */
6193fa34dd3eSStefano Zampini     ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr);
6194fa34dd3eSStefano Zampini     ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr);
6195fa34dd3eSStefano Zampini 
6196fa34dd3eSStefano Zampini     /* mimic vector in \widetilde{W}_\Gamma */
6197fa34dd3eSStefano Zampini     ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr);
6198fa34dd3eSStefano Zampini     /* continuous in primal space */
6199fa34dd3eSStefano Zampini     ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr);
6200fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6201fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6202fa34dd3eSStefano Zampini     ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
62034f1b2e48SStefano Zampini     ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr);
62044f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i];
6205fa34dd3eSStefano Zampini     ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr);
6206fa34dd3eSStefano Zampini     ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
6207fa34dd3eSStefano Zampini     ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr);
6208fa34dd3eSStefano Zampini     ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr);
6209fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6210fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6211fa34dd3eSStefano Zampini     ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr);
6212fa34dd3eSStefano Zampini     ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr);
6213fa34dd3eSStefano Zampini 
6214fa34dd3eSStefano Zampini     /* assemble rhs for coarse problem */
6215fa34dd3eSStefano Zampini     /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */
6216fa34dd3eSStefano Zampini     /* local with Schur */
6217fa34dd3eSStefano Zampini     ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr);
6218fa34dd3eSStefano Zampini     if (zerodiag) {
6219fa34dd3eSStefano Zampini       ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr);
62204f1b2e48SStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i];
6221fa34dd3eSStefano Zampini       ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr);
6222fa34dd3eSStefano Zampini       ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr);
6223fa34dd3eSStefano Zampini     }
6224fa34dd3eSStefano Zampini     /* sum on primal nodes the local contributions */
6225fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6226fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6227fa34dd3eSStefano Zampini     ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
6228fa34dd3eSStefano Zampini     ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr);
6229fa34dd3eSStefano Zampini     for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]];
6230fa34dd3eSStefano Zampini     ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr);
6231fa34dd3eSStefano Zampini     ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
6232fa34dd3eSStefano Zampini     ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr);
6233fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6234fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6235fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6236fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
6237fa34dd3eSStefano Zampini     ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
6238fa34dd3eSStefano Zampini     /* scale primal nodes (BDDC sums contibutions) */
6239fa34dd3eSStefano Zampini     ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr);
6240fa34dd3eSStefano Zampini     ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr);
6241fa34dd3eSStefano Zampini     ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr);
6242fa34dd3eSStefano Zampini     ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr);
6243fa34dd3eSStefano Zampini     ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr);
6244fa34dd3eSStefano Zampini     ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6245fa34dd3eSStefano Zampini     ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
6246fa34dd3eSStefano Zampini     /* global: \widetilde{B0}_B w_\Gamma */
6247fa34dd3eSStefano Zampini     if (zerodiag) {
6248fa34dd3eSStefano Zampini       ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr);
6249fa34dd3eSStefano Zampini       ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr);
62504f1b2e48SStefano Zampini       for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i];
6251fa34dd3eSStefano Zampini       ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr);
6252fa34dd3eSStefano Zampini     }
6253fa34dd3eSStefano Zampini     /* BDDC */
6254fa34dd3eSStefano Zampini     ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr);
6255fa34dd3eSStefano Zampini     ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr);
6256fa34dd3eSStefano Zampini 
6257fa34dd3eSStefano Zampini     ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr);
6258fa34dd3eSStefano Zampini     ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr);
6259fa34dd3eSStefano Zampini     ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr);
6260fa34dd3eSStefano Zampini     PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm);
62614f1b2e48SStefano Zampini     for (i=0;i<pcbddc->benign_n;i++) {
62624f1b2e48SStefano Zampini       PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i]));
6263fa34dd3eSStefano Zampini     }
62644f1b2e48SStefano Zampini     ierr = PetscFree(p0_check);CHKERRQ(ierr);
6265fa34dd3eSStefano Zampini     ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr);
6266fa34dd3eSStefano Zampini     ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr);
6267fa34dd3eSStefano Zampini     ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr);
6268fa34dd3eSStefano Zampini     ierr = MatDestroy(&S_j);CHKERRQ(ierr);
6269fa34dd3eSStefano Zampini     ierr = MatDestroy(&B0_B);CHKERRQ(ierr);
6270fa34dd3eSStefano Zampini   }
6271fa34dd3eSStefano Zampini   PetscFunctionReturn(0);
6272fa34dd3eSStefano Zampini }
6273