xref: /petsc/src/ksp/pc/impls/bddc/bddcprivate.c (revision f6f667cfabdb84fae686043df9fba3b425683839)
1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h>
2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h>
3674ae819SStefano Zampini #include <petscblaslapack.h>
4674ae819SStefano Zampini 
5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y);
6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y);
7906d46d4SStefano Zampini 
8b1b3d7a2SStefano Zampini #undef __FUNCT__
9b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection"
1008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc)
11b1b3d7a2SStefano Zampini {
12b1b3d7a2SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)pc->data;
1308122e43SStefano Zampini   PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
1408122e43SStefano Zampini   PetscBLASInt    B_dummyint,B_neigs,B_ierr,B_lwork;
1508122e43SStefano Zampini   PetscBLASInt    *B_iwork,*B_ifail;
1608122e43SStefano Zampini   PetscScalar     *work,lwork;
1708122e43SStefano Zampini   PetscScalar     *St,*S,*eigv;
1808122e43SStefano Zampini   PetscScalar     *Sarray,*Starray;
1908122e43SStefano Zampini   PetscReal       *eigs,thresh;
201b968477SStefano Zampini   PetscInt        i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs;
21*f6f667cfSStefano Zampini   PetscBool       allocated_S_St;
2208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
2308122e43SStefano Zampini   PetscReal       *rwork;
2408122e43SStefano Zampini #endif
25b1b3d7a2SStefano Zampini   PetscErrorCode  ierr;
26b1b3d7a2SStefano Zampini 
27b1b3d7a2SStefano Zampini   PetscFunctionBegin;
2808122e43SStefano Zampini   if (!sub_schurs->use_mumps) {
2908122e43SStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS");
3008122e43SStefano Zampini   }
3108122e43SStefano Zampini 
32fd14bc51SStefano Zampini   if (pcbddc->dbg_flag) {
33fd14bc51SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
34fd14bc51SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
35fd14bc51SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr);
36fd14bc51SStefano Zampini     ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
37fd14bc51SStefano Zampini   }
38fd14bc51SStefano Zampini 
39e496cd5dSStefano Zampini   if (pcbddc->dbg_flag) {
40e496cd5dSStefano 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);
41e496cd5dSStefano Zampini   }
42e496cd5dSStefano Zampini 
43e496cd5dSStefano Zampini   if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) {
44e496cd5dSStefano 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);
45e496cd5dSStefano Zampini   }
46e496cd5dSStefano Zampini 
4708122e43SStefano Zampini   /* max size of subsets */
4808122e43SStefano Zampini   mss = 0;
4908122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
5008122e43SStefano Zampini     if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) {
5108122e43SStefano Zampini       PetscInt subset_size;
5208122e43SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
5308122e43SStefano Zampini       mss = PetscMax(mss,subset_size);
5408122e43SStefano Zampini     }
5508122e43SStefano Zampini   }
5608122e43SStefano Zampini 
5708122e43SStefano Zampini   /* min/max and threshold */
5808122e43SStefano Zampini   nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss;
59*f6f667cfSStefano Zampini   nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0;
6008122e43SStefano Zampini   nmax = PetscMax(nmin,nmax);
61*f6f667cfSStefano Zampini   allocated_S_St = PETSC_FALSE;
62*f6f667cfSStefano Zampini   if (nmin) {
63*f6f667cfSStefano Zampini     allocated_S_St = PETSC_TRUE;
64*f6f667cfSStefano Zampini   }
6508122e43SStefano Zampini 
6608122e43SStefano Zampini   /* allocate lapack workspace */
6708122e43SStefano Zampini   cum = cum2 = 0;
6808122e43SStefano Zampini   maxneigs = 0;
6908122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
7008122e43SStefano Zampini     if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) {
7108122e43SStefano Zampini       PetscInt n,subset_size;
72*f6f667cfSStefano Zampini 
7308122e43SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
7408122e43SStefano Zampini       n = PetscMin(subset_size,nmax);
7508122e43SStefano Zampini       cum += subset_size*n;
7608122e43SStefano Zampini       cum2 += n;
7708122e43SStefano Zampini       maxneigs = PetscMax(maxneigs,n);
7808122e43SStefano Zampini     }
7908122e43SStefano Zampini   }
8008122e43SStefano Zampini   if (mss) {
819ab7bb16SStefano Zampini     if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
8208122e43SStefano Zampini       PetscBLASInt B_itype = 1;
8308122e43SStefano Zampini       PetscBLASInt B_N = mss;
844c6709b3SStefano Zampini       PetscReal    zero = 0.0;
854c6709b3SStefano Zampini       PetscReal    eps = 0.0; /* dlamch? */
8608122e43SStefano Zampini 
8708122e43SStefano Zampini       B_lwork = -1;
8808122e43SStefano Zampini       S = NULL;
8908122e43SStefano Zampini       St = NULL;
90a58a30b4SStefano Zampini       eigs = NULL;
91a58a30b4SStefano Zampini       eigv = NULL;
92a58a30b4SStefano Zampini       B_iwork = NULL;
93a58a30b4SStefano Zampini       B_ifail = NULL;
948bec7fa6SStefano Zampini       thresh = 1.0;
9508122e43SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
9608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
9708122e43SStefano 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));
9808122e43SStefano Zampini #else
9908122e43SStefano 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));
10008122e43SStefano Zampini #endif
10108122e43SStefano Zampini       if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr);
10208122e43SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
10308122e43SStefano Zampini     } else {
10408122e43SStefano Zampini         /* TODO */
10508122e43SStefano Zampini     }
10608122e43SStefano Zampini   } else {
10708122e43SStefano Zampini     lwork = 0;
10808122e43SStefano Zampini   }
10908122e43SStefano Zampini 
11008122e43SStefano Zampini   nv = 0;
1118bec7fa6SStefano Zampini   if (sub_schurs->is_Ej_com && pcbddc->use_vertices) { /* complement of subsets, each entry is a vertex */
11208122e43SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&nv);CHKERRQ(ierr);
11308122e43SStefano Zampini   }
1144c6709b3SStefano Zampini   ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr);
115*f6f667cfSStefano Zampini   if (allocated_S_St) {
116*f6f667cfSStefano Zampini     ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr);
117*f6f667cfSStefano Zampini   }
118*f6f667cfSStefano Zampini   ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr);
11908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
12008122e43SStefano Zampini   ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr);
12108122e43SStefano Zampini #endif
12208122e43SStefano Zampini   ierr = PetscMalloc4(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n,
12308122e43SStefano Zampini                       nv+cum2+1,&pcbddc->adaptive_constraints_ptrs,
12408122e43SStefano Zampini                       nv+cum,&pcbddc->adaptive_constraints_idxs,
12508122e43SStefano Zampini                       nv+cum,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr);
12608122e43SStefano Zampini   ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr);
12708122e43SStefano Zampini 
12808122e43SStefano Zampini   maxneigs = 0;
12908122e43SStefano Zampini   cum = cum2 = cumarray = 0;
1308bec7fa6SStefano Zampini   if (sub_schurs->is_Ej_com && pcbddc->use_vertices) {
13108122e43SStefano Zampini     const PetscInt *idxs;
13208122e43SStefano Zampini 
13308122e43SStefano Zampini     ierr = ISGetIndices(sub_schurs->is_Ej_com,&idxs);CHKERRQ(ierr);
13408122e43SStefano Zampini     for (cum=0;cum<nv;cum++) {
13508122e43SStefano Zampini       pcbddc->adaptive_constraints_n[cum] = 1;
13608122e43SStefano Zampini       pcbddc->adaptive_constraints_idxs[cum] = idxs[cum];
13708122e43SStefano Zampini       pcbddc->adaptive_constraints_ptrs[cum] = cum;
13808122e43SStefano Zampini       pcbddc->adaptive_constraints_data[cum] = 1.0;
13908122e43SStefano Zampini     }
14008122e43SStefano Zampini     cum2 = cum;
14108122e43SStefano Zampini     ierr = ISRestoreIndices(sub_schurs->is_Ej_com,&idxs);CHKERRQ(ierr);
14208122e43SStefano Zampini   }
14308122e43SStefano Zampini 
14408122e43SStefano Zampini   if (mss) { /* multilevel */
14508122e43SStefano Zampini     ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr);
14608122e43SStefano Zampini     ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr);
14708122e43SStefano Zampini   }
14808122e43SStefano Zampini 
14908122e43SStefano Zampini   for (i=0;i<sub_schurs->n_subs;i++) {
15008122e43SStefano Zampini     PetscInt j,subset_size;
15108122e43SStefano Zampini 
15208122e43SStefano Zampini     ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr);
15308122e43SStefano Zampini     if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) {
15408122e43SStefano Zampini       const PetscInt *idxs;
155*f6f667cfSStefano Zampini       PetscReal      infty = PETSC_MAX_REAL;
156*f6f667cfSStefano Zampini       PetscInt       eigs_start = 0;
15708122e43SStefano Zampini       PetscBLASInt   B_N;
15808122e43SStefano Zampini 
159*f6f667cfSStefano Zampini       ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr);
160*f6f667cfSStefano Zampini       if (allocated_S_St) { /* S and S_t should be copied since we could need them later */
1619ab7bb16SStefano Zampini         if (sub_schurs->is_hermitian) {
16208122e43SStefano Zampini           PetscInt j;
16308122e43SStefano Zampini           for (j=0;j<subset_size;j++) {
16408122e43SStefano Zampini             ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
16508122e43SStefano Zampini           }
16608122e43SStefano Zampini           for (j=0;j<subset_size;j++) {
16708122e43SStefano Zampini             ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
16808122e43SStefano Zampini           }
16908122e43SStefano Zampini         } else {
17008122e43SStefano Zampini           ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
17108122e43SStefano Zampini           ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
17208122e43SStefano Zampini         }
1738bec7fa6SStefano Zampini       } else {
174*f6f667cfSStefano Zampini         S = Sarray + cumarray;
175*f6f667cfSStefano Zampini         St = Starray + cumarray;
1768bec7fa6SStefano Zampini       }
17708122e43SStefano Zampini 
178*f6f667cfSStefano Zampini       /* Threshold: this is an heuristic for edges */
179*f6f667cfSStefano Zampini       ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
180*f6f667cfSStefano Zampini       thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold;
181*f6f667cfSStefano Zampini 
1829ab7bb16SStefano Zampini       if (sub_schurs->is_hermitian && sub_schurs->is_posdef) {
18308122e43SStefano Zampini         PetscBLASInt B_itype = 1;
184*f6f667cfSStefano Zampini         PetscBLASInt B_IL, B_IU;
1854c6709b3SStefano Zampini         PetscReal    eps = -1.0; /* dlamch? */
1869552c7c7SStefano Zampini         PetscInt     nmin_s;
18708122e43SStefano Zampini 
188*f6f667cfSStefano Zampini         /* ask for eigenvalues larger than thresh */
189fd14bc51SStefano Zampini         if (pcbddc->dbg_flag) {
1908bec7fa6SStefano 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]]);
191fd14bc51SStefano Zampini         }
19208122e43SStefano Zampini         ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
19308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
194*f6f667cfSStefano 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));
19508122e43SStefano Zampini #else
196*f6f667cfSStefano 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));
19708122e43SStefano Zampini #endif
19808122e43SStefano Zampini         ierr = PetscFPTrapPop();CHKERRQ(ierr);
19908122e43SStefano Zampini         if (B_ierr) {
20008122e43SStefano Zampini           if (B_ierr < 0 ) {
20108122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
20208122e43SStefano Zampini           } else if (B_ierr <= B_N) {
20308122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
20408122e43SStefano Zampini           } else {
2059552c7c7SStefano 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);
20608122e43SStefano Zampini           }
20708122e43SStefano Zampini         }
20808122e43SStefano Zampini 
20908122e43SStefano Zampini         if (B_neigs > nmax) {
210fd14bc51SStefano Zampini           if (pcbddc->dbg_flag) {
211fd14bc51SStefano Zampini             PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   found %d eigs, more than maximum required %d.\n",B_neigs,nmax);
212fd14bc51SStefano Zampini           }
213*f6f667cfSStefano Zampini           eigs_start = B_neigs -nmax;
21408122e43SStefano Zampini           B_neigs = nmax;
21508122e43SStefano Zampini         }
21608122e43SStefano Zampini 
2179552c7c7SStefano Zampini         nmin_s = PetscMin(nmin,B_N);
2189552c7c7SStefano Zampini         if (B_neigs < nmin_s) {
21908122e43SStefano Zampini           PetscBLASInt B_neigs2;
22008122e43SStefano Zampini 
221*f6f667cfSStefano Zampini           B_IU = B_N - B_neigs;
222*f6f667cfSStefano Zampini           B_IL = B_N - nmin_s + 1;
223fd14bc51SStefano Zampini           if (pcbddc->dbg_flag) {
224fd14bc51SStefano 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);
225fd14bc51SStefano Zampini           }
2269ab7bb16SStefano Zampini           if (sub_schurs->is_hermitian) {
22708122e43SStefano Zampini             PetscInt j;
22808122e43SStefano Zampini             for (j=0;j<subset_size;j++) {
22908122e43SStefano Zampini               ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
23008122e43SStefano Zampini             }
23108122e43SStefano Zampini             for (j=0;j<subset_size;j++) {
23208122e43SStefano Zampini               ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr);
23308122e43SStefano Zampini             }
23408122e43SStefano Zampini           } else {
23508122e43SStefano Zampini             ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
23608122e43SStefano Zampini             ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
23708122e43SStefano Zampini           }
23808122e43SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
23908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
240*f6f667cfSStefano 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));
24108122e43SStefano Zampini #else
242*f6f667cfSStefano 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));
24308122e43SStefano Zampini #endif
24408122e43SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
24508122e43SStefano Zampini           B_neigs += B_neigs2;
24608122e43SStefano Zampini         }
24708122e43SStefano Zampini         if (B_ierr) {
24808122e43SStefano Zampini           if (B_ierr < 0 ) {
24908122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
25008122e43SStefano Zampini           } else if (B_ierr <= B_N) {
25108122e43SStefano Zampini             SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
25208122e43SStefano Zampini           } else {
2539552c7c7SStefano 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);
25408122e43SStefano Zampini           }
25508122e43SStefano Zampini         }
256fd14bc51SStefano Zampini         if (pcbddc->dbg_flag) {
257fd14bc51SStefano Zampini           PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   -> Got %d eigs\n",B_neigs);
25808122e43SStefano Zampini           for (j=0;j<B_neigs;j++) {
25908122e43SStefano Zampini             if (eigs[j] == 0.0) {
260fd14bc51SStefano Zampini               PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"     Inf\n");
26108122e43SStefano Zampini             } else {
262*f6f667cfSStefano Zampini               PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"     %1.6e\n",eigs[j+eigs_start]);
263fd14bc51SStefano Zampini             }
26408122e43SStefano Zampini           }
26508122e43SStefano Zampini         }
26608122e43SStefano Zampini       } else {
26708122e43SStefano Zampini           /* TODO */
26808122e43SStefano Zampini       }
26908122e43SStefano Zampini 
2708bec7fa6SStefano Zampini       maxneigs = PetscMax(B_neigs,maxneigs);
2718bec7fa6SStefano Zampini       pcbddc->adaptive_constraints_n[i+nv] = B_neigs;
272*f6f667cfSStefano Zampini       ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+cum2,eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr);
273fd14bc51SStefano Zampini 
274fd14bc51SStefano Zampini       if (pcbddc->dbg_flag > 1) {
2759552c7c7SStefano Zampini         PetscInt ii;
2769552c7c7SStefano Zampini         for (ii=0;ii<B_neigs;ii++) {
277fd14bc51SStefano Zampini           PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"   -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);
2789552c7c7SStefano Zampini           for (j=0;j<B_N;j++) {
279*f6f667cfSStefano Zampini             PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"       %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+cum2]);
2809552c7c7SStefano Zampini           }
2819552c7c7SStefano Zampini         }
282fd14bc51SStefano Zampini       }
28308122e43SStefano Zampini       for (j=0;j<B_neigs;j++) {
28408122e43SStefano Zampini #if 0
28508122e43SStefano Zampini         {
28608122e43SStefano Zampini           PetscBLASInt Blas_N,Blas_one = 1.0;
28708122e43SStefano Zampini           PetscScalar norm;
28808122e43SStefano Zampini           ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr);
28908122e43SStefano Zampini           PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+cum2,&Blas_one,pcbddc->adaptive_constraints_data+cum2,&Blas_one));
29008122e43SStefano Zampini           if (pcbddc->adaptive_constraints_data[cum2] > 0.0) {
29108122e43SStefano Zampini             norm = 1.0/PetscSqrtReal(PetscRealPart(norm));
29208122e43SStefano Zampini           } else {
29308122e43SStefano Zampini             norm = -1.0/PetscSqrtReal(PetscRealPart(norm));
29408122e43SStefano Zampini           }
29508122e43SStefano Zampini           PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+cum2,&Blas_one));
296b1b3d7a2SStefano Zampini         }
297b1b3d7a2SStefano Zampini #endif
29808122e43SStefano Zampini         ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+cum2,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr);
29908122e43SStefano Zampini         pcbddc->adaptive_constraints_ptrs[cum++] = cum2;
30008122e43SStefano Zampini         cum2 += subset_size;
30108122e43SStefano Zampini       }
30208122e43SStefano Zampini       ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr);
30308122e43SStefano Zampini     }
30408122e43SStefano Zampini     /* shift for next computation */
30508122e43SStefano Zampini     cumarray += subset_size*subset_size;
30608122e43SStefano Zampini   }
307fd14bc51SStefano Zampini   if (pcbddc->dbg_flag) {
308fd14bc51SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
309fd14bc51SStefano Zampini   }
31008122e43SStefano Zampini   pcbddc->adaptive_constraints_ptrs[cum] = cum2;
31108122e43SStefano Zampini 
31208122e43SStefano Zampini   if (mss) {
31308122e43SStefano Zampini     ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr);
31408122e43SStefano Zampini     ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr);
315*f6f667cfSStefano Zampini     /* destroy matrices (junk) */
316*f6f667cfSStefano Zampini     ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr);
317*f6f667cfSStefano Zampini     ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr);
31808122e43SStefano Zampini   }
319*f6f667cfSStefano Zampini   if (allocated_S_St) {
320*f6f667cfSStefano Zampini     ierr = PetscFree2(S,St);CHKERRQ(ierr);
321*f6f667cfSStefano Zampini   }
322*f6f667cfSStefano Zampini   ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr);
32308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX)
32408122e43SStefano Zampini   ierr = PetscFree(rwork);CHKERRQ(ierr);
32508122e43SStefano Zampini #endif
32608122e43SStefano Zampini   if (pcbddc->dbg_flag) {
3271b968477SStefano Zampini     PetscInt maxneigs_r;
32808122e43SStefano Zampini     ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
32908122e43SStefano Zampini     ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr);
33008122e43SStefano Zampini   }
33108122e43SStefano Zampini   PetscFunctionReturn(0);
33208122e43SStefano Zampini }
333b1b3d7a2SStefano Zampini 
334674ae819SStefano Zampini #undef __FUNCT__
335c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers"
336c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc)
337c8587f34SStefano Zampini {
338c8587f34SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
3398629588bSStefano Zampini   PetscScalar    *coarse_submat_vals;
340c8587f34SStefano Zampini   PetscErrorCode ierr;
341c8587f34SStefano Zampini 
342c8587f34SStefano Zampini   PetscFunctionBegin;
343f4ddd8eeSStefano Zampini   /* Setup local scatters R_to_B and (optionally) R_to_D */
3445e8657edSStefano Zampini   /* PCBDDCSetUpLocalWorkVectors should be called first! */
345c8587f34SStefano Zampini   ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr);
346c8587f34SStefano Zampini 
347684f6988SStefano Zampini   /* Setup local neumann solver ksp_R */
3480fccc4e9SStefano Zampini   /* PCBDDCSetUpLocalScatters should be called first! */
349684f6988SStefano Zampini   ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr);
350c8587f34SStefano Zampini 
351c8587f34SStefano Zampini   /* Change global null space passed in by the user if change of basis has been requested */
352b9b85e73SStefano Zampini   if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) {
353c8587f34SStefano Zampini     ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr);
354c8587f34SStefano Zampini   }
355c8587f34SStefano Zampini 
3568629588bSStefano Zampini   /*
3578629588bSStefano Zampini      Setup local correction and local part of coarse basis.
3588629588bSStefano Zampini      Gives back the dense local part of the coarse matrix in column major ordering
3598629588bSStefano Zampini   */
36047f4ddc3SStefano Zampini   ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr);
3618629588bSStefano Zampini 
3628629588bSStefano Zampini   /* Compute total number of coarse nodes and setup coarse solver */
3638629588bSStefano Zampini   ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr);
3648629588bSStefano Zampini 
3658629588bSStefano Zampini   /* free */
3668629588bSStefano Zampini   ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr);
367c8587f34SStefano Zampini   PetscFunctionReturn(0);
368c8587f34SStefano Zampini }
369c8587f34SStefano Zampini 
370c8587f34SStefano Zampini #undef __FUNCT__
371674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization"
372674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc)
373674ae819SStefano Zampini {
374674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
375674ae819SStefano Zampini   PetscErrorCode ierr;
376674ae819SStefano Zampini 
377674ae819SStefano Zampini   PetscFunctionBegin;
378674ae819SStefano Zampini   ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr);
379674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr);
380674ae819SStefano Zampini   ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr);
381674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr);
382785d1243SStefano Zampini   ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr);
383674ae819SStefano Zampini   ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr);
384f4ddd8eeSStefano Zampini   ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr);
385f4ddd8eeSStefano Zampini   ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr);
386785d1243SStefano Zampini   ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr);
38763602bcaSStefano Zampini   ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr);
38863602bcaSStefano Zampini   ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr);
389674ae819SStefano Zampini   PetscFunctionReturn(0);
390674ae819SStefano Zampini }
391674ae819SStefano Zampini 
392674ae819SStefano Zampini #undef __FUNCT__
393674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography"
394674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc)
395674ae819SStefano Zampini {
396674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
397674ae819SStefano Zampini   PetscErrorCode ierr;
398674ae819SStefano Zampini 
399674ae819SStefano Zampini   PetscFunctionBegin;
400b9b85e73SStefano Zampini   ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
401674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
402674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
403674ae819SStefano Zampini   ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr);
404b96c3477SStefano Zampini   ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr);
405674ae819SStefano Zampini   PetscFunctionReturn(0);
406674ae819SStefano Zampini }
407674ae819SStefano Zampini 
408674ae819SStefano Zampini #undef __FUNCT__
409674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers"
410674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc)
411674ae819SStefano Zampini {
412674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
41306656605SStefano Zampini   PetscScalar    *array;
414674ae819SStefano Zampini   PetscErrorCode ierr;
415674ae819SStefano Zampini 
416674ae819SStefano Zampini   PetscFunctionBegin;
417674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr);
41858da7f69SStefano Zampini   if (pcbddc->coarse_phi_B) {
41906656605SStefano Zampini     ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr);
42006656605SStefano Zampini     ierr = PetscFree(array);CHKERRQ(ierr);
42158da7f69SStefano Zampini   }
422674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr);
423674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr);
42415aaf578SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr);
42515aaf578SStefano Zampini   ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr);
426674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr);
427674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr);
42806656605SStefano Zampini   if (pcbddc->local_auxmat2) {
42906656605SStefano Zampini     ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr);
43006656605SStefano Zampini     ierr = PetscFree(array);CHKERRQ(ierr);
43106656605SStefano Zampini   }
432674ae819SStefano Zampini   ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr);
43306656605SStefano Zampini   ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr);
434674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr);
435674ae819SStefano Zampini   ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr);
4368ce42a96SStefano Zampini   ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr);
437674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr);
438674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr);
439674ae819SStefano Zampini   ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
440f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr);
441f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr);
442f4ddd8eeSStefano Zampini   ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr);
443f4ddd8eeSStefano Zampini   ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
444727cdba6SStefano Zampini   ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
445f4ddd8eeSStefano Zampini   ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr);
44670cf5478SStefano Zampini   ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr);
4476e683305SStefano Zampini   ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
448674ae819SStefano Zampini   PetscFunctionReturn(0);
449674ae819SStefano Zampini }
450674ae819SStefano Zampini 
451674ae819SStefano Zampini #undef __FUNCT__
452f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors"
453f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc)
4546bfb1811SStefano Zampini {
4556bfb1811SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
4566bfb1811SStefano Zampini   PC_IS          *pcis = (PC_IS*)pc->data;
4576bfb1811SStefano Zampini   VecType        impVecType;
458e9189074SStefano Zampini   PetscInt       n_constraints,n_R,old_size;
4596bfb1811SStefano Zampini   PetscErrorCode ierr;
4606bfb1811SStefano Zampini 
4616bfb1811SStefano Zampini   PetscFunctionBegin;
462f4ddd8eeSStefano Zampini   if (!pcbddc->ConstraintMatrix) {
463f4ddd8eeSStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created");
464f4ddd8eeSStefano Zampini   }
465e7b262bdSStefano Zampini   /* get sizes */
466e9189074SStefano Zampini   n_constraints = pcbddc->local_primal_size - pcbddc->n_actual_vertices;
467e9189074SStefano Zampini   n_R = pcis->n-pcbddc->n_actual_vertices;
4686bfb1811SStefano Zampini   ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr);
469e7b262bdSStefano Zampini   /* local work vectors (try to avoid unneeded work)*/
470e7b262bdSStefano Zampini   /* R nodes */
471e7b262bdSStefano Zampini   old_size = -1;
472e7b262bdSStefano Zampini   if (pcbddc->vec1_R) {
473e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr);
474e7b262bdSStefano Zampini   }
475e7b262bdSStefano Zampini   if (n_R != old_size) {
476e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr);
477e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr);
4786bfb1811SStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr);
4796bfb1811SStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr);
4806bfb1811SStefano Zampini     ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr);
4816bfb1811SStefano Zampini     ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr);
482e7b262bdSStefano Zampini   }
483e7b262bdSStefano Zampini   /* local primal dofs */
484e7b262bdSStefano Zampini   old_size = -1;
485e7b262bdSStefano Zampini   if (pcbddc->vec1_P) {
486e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr);
487e7b262bdSStefano Zampini   }
488e9189074SStefano Zampini   if (pcbddc->local_primal_size != old_size) {
489e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr);
49083b7ccabSStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr);
491e9189074SStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr);
4926bfb1811SStefano Zampini     ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr);
493e7b262bdSStefano Zampini   }
494e7b262bdSStefano Zampini   /* local explicit constraints */
495e7b262bdSStefano Zampini   old_size = -1;
496e7b262bdSStefano Zampini   if (pcbddc->vec1_C) {
497e7b262bdSStefano Zampini     ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr);
498e7b262bdSStefano Zampini   }
499e7b262bdSStefano Zampini   if (n_constraints && n_constraints != old_size) {
500e7b262bdSStefano Zampini     ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr);
50183b7ccabSStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr);
50283b7ccabSStefano Zampini     ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr);
50383b7ccabSStefano Zampini     ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr);
50483b7ccabSStefano Zampini   }
5056bfb1811SStefano Zampini   PetscFunctionReturn(0);
5066bfb1811SStefano Zampini }
5076bfb1811SStefano Zampini 
5086bfb1811SStefano Zampini #undef __FUNCT__
50947f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection"
51047f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n)
51188ebb749SStefano Zampini {
51225084f0cSStefano Zampini   PetscErrorCode         ierr;
51325084f0cSStefano Zampini   /* pointers to pcis and pcbddc */
51488ebb749SStefano Zampini   PC_IS*                 pcis = (PC_IS*)pc->data;
51588ebb749SStefano Zampini   PC_BDDC*               pcbddc = (PC_BDDC*)pc->data;
51625084f0cSStefano Zampini   /* submatrices of local problem */
51788ebb749SStefano Zampini   Mat                    A_RV,A_VR,A_VV;
51806656605SStefano Zampini   /* submatrices of local coarse problem */
51906656605SStefano Zampini   Mat                    S_VV,S_CV,S_VC,S_CC;
52025084f0cSStefano Zampini   /* working matrices */
52106656605SStefano Zampini   Mat                    C_CR;
52225084f0cSStefano Zampini   /* additional working stuff */
52306656605SStefano Zampini   PC                     pc_R;
52406656605SStefano Zampini   Mat                    F;
52506656605SStefano Zampini   PetscBool              isLU,isCHOL,isILU;
52606656605SStefano Zampini 
52725084f0cSStefano Zampini   PetscScalar            *coarse_submat_vals; /* TODO: use a PETSc matrix */
52806656605SStefano Zampini   PetscScalar            *work;
52906656605SStefano Zampini   PetscInt               *idx_V_B;
53006656605SStefano Zampini   PetscInt               n,n_vertices,n_constraints;
53106656605SStefano Zampini   PetscInt               i,n_R,n_D,n_B;
532b9d89cd5SStefano Zampini   PetscBool              unsymmetric_check;
53345a1bb75SStefano Zampini   /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */
53488ebb749SStefano Zampini   MatType                impMatType;
53525084f0cSStefano Zampini   /* some shortcuts to scalars */
53606656605SStefano Zampini   PetscScalar            one=1.0,m_one=-1.0;
53788ebb749SStefano Zampini 
53888ebb749SStefano Zampini   PetscFunctionBegin;
539e9189074SStefano Zampini   /* get number of vertices (corners plus constraints with change of basis)
540e9189074SStefano Zampini      pcbddc->n_actual_vertices stores the actual number of vertices, pcbddc->n_vertices the number of corners computed */
541e9189074SStefano Zampini   n_vertices = pcbddc->n_actual_vertices;
54288ebb749SStefano Zampini   n_constraints = pcbddc->local_primal_size-n_vertices;
54388ebb749SStefano Zampini   /* Set Non-overlapping dimensions */
54488ebb749SStefano Zampini   n_B = pcis->n_B; n_D = pcis->n - n_B;
54588ebb749SStefano Zampini   n_R = pcis->n-n_vertices;
54688ebb749SStefano Zampini 
54788ebb749SStefano Zampini   /* Set types for local objects needed by BDDC precondtioner */
54888ebb749SStefano Zampini   impMatType = MATSEQDENSE;
54988ebb749SStefano Zampini 
55088ebb749SStefano Zampini   /* vertices in boundary numbering */
551785e854fSJed Brown   ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr);
5525e8657edSStefano Zampini   ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->primal_indices_local_idxs,&i,idx_V_B);CHKERRQ(ierr);
55388ebb749SStefano Zampini   if (i != n_vertices) {
55422d5777bSStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i);
55588ebb749SStefano Zampini   }
55688ebb749SStefano Zampini 
55706656605SStefano Zampini   /* Subdomain contribution (Non-overlapping) to coarse matrix  */
55806656605SStefano Zampini   ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr);
55906656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr);
56006656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr);
56106656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr);
56206656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr);
56306656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr);
56406656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr);
56506656605SStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr);
56606656605SStefano Zampini   ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr);
56706656605SStefano Zampini 
56806656605SStefano Zampini   unsymmetric_check = PETSC_FALSE;
56906656605SStefano Zampini   /* allocate workspace */
57006656605SStefano Zampini   n = 0;
57106656605SStefano Zampini   if (n_constraints) {
57206656605SStefano Zampini     n += n_R*n_constraints;
57306656605SStefano Zampini   }
57406656605SStefano Zampini   if (n_vertices) {
57506656605SStefano Zampini     n = PetscMax(2*n_R*n_vertices,n);
57606656605SStefano Zampini   }
57706656605SStefano Zampini   if (!pcbddc->issym) {
57806656605SStefano Zampini     n = PetscMax(2*n_R*pcbddc->local_primal_size,n);
57906656605SStefano Zampini     unsymmetric_check = PETSC_TRUE;
58006656605SStefano Zampini   }
58106656605SStefano Zampini   ierr = PetscMalloc1(n,&work);CHKERRQ(ierr);
58206656605SStefano Zampini 
58306656605SStefano Zampini   /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */
58406656605SStefano Zampini   ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr);
58506656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr);
58606656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr);
58706656605SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr);
58806656605SStefano Zampini   if (isLU || isILU || isCHOL) {
58906656605SStefano Zampini     ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr);
59006656605SStefano Zampini   } else {
59106656605SStefano Zampini     F = NULL;
59206656605SStefano Zampini   }
59306656605SStefano Zampini 
59488ebb749SStefano Zampini   /* Precompute stuffs needed for preprocessing and application of BDDC*/
59588ebb749SStefano Zampini   if (n_constraints) {
59606656605SStefano Zampini     Mat M1,M2,M3;
59706656605SStefano Zampini     IS  is_aux;
598f4ddd8eeSStefano Zampini     /* see if we can save some allocations */
599f4ddd8eeSStefano Zampini     if (pcbddc->local_auxmat2) {
600f4ddd8eeSStefano Zampini       PetscInt on_R,on_constraints;
601f4ddd8eeSStefano Zampini       ierr = MatGetSize(pcbddc->local_auxmat2,&on_R,&on_constraints);CHKERRQ(ierr);
602f4ddd8eeSStefano Zampini       if (on_R != n_R || on_constraints != n_constraints) {
60306656605SStefano Zampini         PetscScalar *marray;
60406656605SStefano Zampini 
60506656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->local_auxmat2,&marray);CHKERRQ(ierr);
60606656605SStefano Zampini         ierr = PetscFree(marray);CHKERRQ(ierr);
607f4ddd8eeSStefano Zampini         ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr);
608f4ddd8eeSStefano Zampini         ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr);
609f4ddd8eeSStefano Zampini       }
610f4ddd8eeSStefano Zampini     }
61145a1bb75SStefano Zampini     /* auxiliary matrices */
612f4ddd8eeSStefano Zampini     if (!pcbddc->local_auxmat2) {
61306656605SStefano Zampini       PetscScalar *marray;
61406656605SStefano Zampini 
61506656605SStefano Zampini       ierr = PetscMalloc1(2*n_R*n_constraints,&marray);CHKERRQ(ierr);
61606656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,marray,&pcbddc->local_auxmat2);CHKERRQ(ierr);
61706656605SStefano Zampini       marray += n_R*n_constraints;
61806656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_R,marray,&pcbddc->local_auxmat1);CHKERRQ(ierr);
619f4ddd8eeSStefano Zampini     }
62088ebb749SStefano Zampini 
62125084f0cSStefano Zampini     /* Extract constraints on R nodes: C_{CR}  */
62225084f0cSStefano Zampini     ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr);
6238ce42a96SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr);
62425084f0cSStefano Zampini     ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
62588ebb749SStefano Zampini 
62688ebb749SStefano Zampini     /* Assemble local_auxmat2 = - A_{RR}^{-1} C^T_{CR} needed by BDDC application */
62706656605SStefano Zampini     ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr);
62888ebb749SStefano Zampini     for (i=0;i<n_constraints;i++) {
62906656605SStefano Zampini       const PetscScalar *row_cmat_values;
63006656605SStefano Zampini       const PetscInt    *row_cmat_indices;
63106656605SStefano Zampini       PetscInt          size_of_constraint,j;
63288ebb749SStefano Zampini 
63306656605SStefano Zampini       ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr);
63406656605SStefano Zampini       for (j=0;j<size_of_constraint;j++) {
63506656605SStefano Zampini         work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j];
63606656605SStefano Zampini       }
63706656605SStefano Zampini       ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr);
63806656605SStefano Zampini     }
63906656605SStefano Zampini     if (F) {
64006656605SStefano Zampini       Mat B;
64106656605SStefano Zampini 
64206656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr);
64306656605SStefano Zampini       ierr = MatMatSolve(F,B,pcbddc->local_auxmat2);CHKERRQ(ierr);
64406656605SStefano Zampini       ierr = MatDestroy(&B);CHKERRQ(ierr);
64506656605SStefano Zampini     } else {
64606656605SStefano Zampini       PetscScalar *xarray;
64706656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr);
64806656605SStefano Zampini       for (i=0;i<n_constraints;i++) {
64906656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr);
65006656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec2_R,xarray+i*n_R);CHKERRQ(ierr);
65106656605SStefano Zampini         ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
65206656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
65306656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
65406656605SStefano Zampini       }
65506656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&xarray);CHKERRQ(ierr);
65606656605SStefano Zampini     }
65706656605SStefano Zampini 
65806656605SStefano Zampini     /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1}  */
65906656605SStefano Zampini     ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr);
66025084f0cSStefano Zampini     ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr);
66106656605SStefano Zampini     ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr);
66206656605SStefano Zampini     ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr);
66325084f0cSStefano Zampini     ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr);
66406656605SStefano Zampini     ierr = VecSet(pcbddc->vec1_C,m_one);CHKERRQ(ierr);
66506656605SStefano Zampini     ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr);
66625084f0cSStefano Zampini     ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr);
66725084f0cSStefano Zampini     ierr = MatDestroy(&M2);CHKERRQ(ierr);
66825084f0cSStefano Zampini     ierr = MatDestroy(&M3);CHKERRQ(ierr);
66906656605SStefano Zampini     /* Assemble local_auxmat1 = S_CC*C_{CR} needed by BDDC application in KSP and in preproc */
670f4ddd8eeSStefano Zampini     ierr = MatMatMult(M1,C_CR,MAT_REUSE_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr);
67106656605SStefano Zampini     ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */
67206656605SStefano Zampini     ierr = MatDestroy(&M1);CHKERRQ(ierr);
67388ebb749SStefano Zampini   }
67488ebb749SStefano Zampini   /* Get submatrices from subdomain matrix */
67588ebb749SStefano Zampini   if (n_vertices) {
67606656605SStefano Zampini     Mat       newmat;
67706656605SStefano Zampini     IS        is_aux;
6783a50541eSStefano Zampini     PetscInt  ibs,mbs;
6793a50541eSStefano Zampini     PetscBool issbaij;
6803a50541eSStefano Zampini 
6813a50541eSStefano Zampini     ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr);
6829577ea80SStefano Zampini     ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr);
6839577ea80SStefano Zampini     ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr);
6843a50541eSStefano Zampini     if (ibs != mbs) { /* need to convert to SEQAIJ */
6859577ea80SStefano Zampini       ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr);
6869577ea80SStefano Zampini       ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr);
6879577ea80SStefano Zampini       ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr);
6889577ea80SStefano Zampini       ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr);
6899577ea80SStefano Zampini       ierr = MatDestroy(&newmat);CHKERRQ(ierr);
6909577ea80SStefano Zampini     } else {
6913a50541eSStefano Zampini       /* this is safe */
6929577ea80SStefano Zampini       ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr);
6933a50541eSStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
6943a50541eSStefano Zampini       if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */
6953a50541eSStefano Zampini         ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr);
6963a50541eSStefano Zampini         ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr);
69706656605SStefano Zampini         ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr);
6983a50541eSStefano Zampini         ierr = MatDestroy(&newmat);CHKERRQ(ierr);
69903af80d8SStefano Zampini         ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr);
7003a50541eSStefano Zampini       } else {
7019577ea80SStefano Zampini         ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr);
7029577ea80SStefano Zampini         ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr);
7039577ea80SStefano Zampini       }
7043a50541eSStefano Zampini     }
70525084f0cSStefano Zampini     ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
70688ebb749SStefano Zampini   }
70788ebb749SStefano Zampini 
70888ebb749SStefano Zampini   /* Matrix of coarse basis functions (local) */
709f4ddd8eeSStefano Zampini   if (pcbddc->coarse_phi_B) {
71006656605SStefano Zampini     PetscInt on_B,on_primal,on_D=n_D;
71106656605SStefano Zampini     if (pcbddc->coarse_phi_D) {
71206656605SStefano Zampini       ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr);
71306656605SStefano Zampini     }
714f4ddd8eeSStefano Zampini     ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr);
71506656605SStefano Zampini     if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) {
71606656605SStefano Zampini       PetscScalar *marray;
71706656605SStefano Zampini 
71806656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr);
71906656605SStefano Zampini       ierr = PetscFree(marray);CHKERRQ(ierr);
720f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr);
721f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr);
722f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr);
723f4ddd8eeSStefano Zampini       ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr);
724f4ddd8eeSStefano Zampini     }
725f4ddd8eeSStefano Zampini   }
72606656605SStefano Zampini 
727f4ddd8eeSStefano Zampini   if (!pcbddc->coarse_phi_B) {
72806656605SStefano Zampini     PetscScalar *marray;
72988ebb749SStefano Zampini 
73006656605SStefano Zampini     n = n_B*pcbddc->local_primal_size;
7318eeda7d8SStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
73206656605SStefano Zampini       n += n_D*pcbddc->local_primal_size;
73388ebb749SStefano Zampini     }
734b9d89cd5SStefano Zampini     if (!pcbddc->issym) {
73506656605SStefano Zampini       n *= 2;
736f4ddd8eeSStefano Zampini     }
73706656605SStefano Zampini     ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr);
73806656605SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr);
73906656605SStefano Zampini     n = n_B*pcbddc->local_primal_size;
74006656605SStefano Zampini     if (pcbddc->switch_static || pcbddc->dbg_flag) {
74106656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr);
74206656605SStefano Zampini       n += n_D*pcbddc->local_primal_size;
74388ebb749SStefano Zampini     }
74406656605SStefano Zampini     if (!pcbddc->issym) {
74506656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr);
7461b968477SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
74706656605SStefano Zampini         n = n_B*pcbddc->local_primal_size;
74806656605SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr);
74988ebb749SStefano Zampini       }
75088ebb749SStefano Zampini     } else {
751c0553b1fSStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr);
752c0553b1fSStefano Zampini       pcbddc->coarse_psi_B = pcbddc->coarse_phi_B;
7531b968477SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
754c0553b1fSStefano Zampini         ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr);
755c0553b1fSStefano Zampini         pcbddc->coarse_psi_D = pcbddc->coarse_phi_D;
756c0553b1fSStefano Zampini       }
75788ebb749SStefano Zampini     }
75806656605SStefano Zampini   }
75906656605SStefano Zampini   /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */
76006656605SStefano Zampini   /* vertices */
76106656605SStefano Zampini   if (n_vertices) {
76216f15bc4SStefano Zampini 
76316f15bc4SStefano Zampini     if (n_R) {
76406656605SStefano Zampini       Mat          A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */
76506656605SStefano Zampini       PetscBLASInt B_N,B_one = 1;
76616f15bc4SStefano Zampini       PetscScalar  *x,*y;
76706656605SStefano Zampini 
76806656605SStefano Zampini       ierr = PetscMemzero(work,2*n_R*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr);
76906656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr);
77006656605SStefano Zampini       ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr);
77106656605SStefano Zampini       if (F) {
77206656605SStefano Zampini         ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr);
77306656605SStefano Zampini       } else {
77406656605SStefano Zampini         ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr);
77506656605SStefano Zampini         for (i=0;i<n_vertices;i++) {
77606656605SStefano Zampini           ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr);
77706656605SStefano Zampini           ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr);
77806656605SStefano Zampini           ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
77906656605SStefano Zampini           ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
78006656605SStefano Zampini           ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
78106656605SStefano Zampini         }
78206656605SStefano Zampini         ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr);
78306656605SStefano Zampini       }
78406656605SStefano Zampini       ierr = MatScale(A_RRmA_RV,m_one);CHKERRQ(ierr);
78506656605SStefano Zampini       /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */
78606656605SStefano Zampini       if (n_constraints) {
78706656605SStefano Zampini         Mat B;
78806656605SStefano Zampini         ierr = MatMatMult(pcbddc->local_auxmat1,A_RRmA_RV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr);
78906656605SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr);
79006656605SStefano Zampini         ierr = MatMatMult(pcbddc->local_auxmat2,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr);
79106656605SStefano Zampini         ierr = MatScale(S_CV,m_one);CHKERRQ(ierr);
79206656605SStefano Zampini         ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr);
79306656605SStefano Zampini         PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one));
79406656605SStefano Zampini         ierr = MatDestroy(&B);CHKERRQ(ierr);
79506656605SStefano Zampini       }
79606656605SStefano Zampini       ierr = MatConvert(A_VR,impMatType,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr);
79706656605SStefano Zampini       ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr);
79806656605SStefano Zampini       ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr);
79906656605SStefano Zampini       ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr);
80006656605SStefano Zampini       ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr);
80106656605SStefano Zampini       ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr);
80206656605SStefano Zampini       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one));
80306656605SStefano Zampini       ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr);
80406656605SStefano Zampini       ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr);
80506656605SStefano Zampini       ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
80606656605SStefano Zampini       ierr = MatDestroy(&S_VVt);CHKERRQ(ierr);
80716f15bc4SStefano Zampini       ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr);
80816f15bc4SStefano Zampini     } else {
80916f15bc4SStefano Zampini       ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr);
81016f15bc4SStefano Zampini       ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
81116f15bc4SStefano Zampini     }
81206656605SStefano Zampini     /* coarse basis functions */
81306656605SStefano Zampini     for (i=0;i<n_vertices;i++) {
81416f15bc4SStefano Zampini       PetscScalar *y;
81516f15bc4SStefano Zampini 
81606656605SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr);
81706656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
81806656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr);
81906656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
82006656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
82106656605SStefano Zampini       y[n_B*i+idx_V_B[i]] = 1.0;
82206656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
82306656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
82406656605SStefano Zampini 
82506656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
82606656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
82706656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr);
82806656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
82906656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
83006656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
83106656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
83206656605SStefano Zampini       }
83306656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
83406656605SStefano Zampini     }
83516f15bc4SStefano Zampini     ierr = MatDestroy(&A_VV);CHKERRQ(ierr);
83606656605SStefano Zampini     ierr = MatDestroy(&A_RV);CHKERRQ(ierr);
83706656605SStefano Zampini   }
83806656605SStefano Zampini 
83906656605SStefano Zampini   if (n_constraints) {
84006656605SStefano Zampini     Mat B;
84106656605SStefano Zampini 
84206656605SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr);
84306656605SStefano Zampini     ierr = MatScale(S_CC,m_one);CHKERRQ(ierr);
84406656605SStefano Zampini     ierr = MatMatMult(pcbddc->local_auxmat2,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr);
84506656605SStefano Zampini     ierr = MatScale(S_CC,m_one);CHKERRQ(ierr);
84606656605SStefano Zampini     if (n_vertices) {
84706656605SStefano Zampini       ierr = MatMatMult(A_VR,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_VC);CHKERRQ(ierr);
84806656605SStefano Zampini     }
84906656605SStefano Zampini     ierr = MatDestroy(&B);CHKERRQ(ierr);
85006656605SStefano Zampini     /* coarse basis functions */
85106656605SStefano Zampini     for (i=0;i<n_constraints;i++) {
85206656605SStefano Zampini       PetscScalar *y;
85306656605SStefano Zampini 
85406656605SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr);
85506656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
85606656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr);
85706656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
85806656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
85906656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr);
86006656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
86106656605SStefano Zampini 
86206656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
86306656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
86406656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr);
86506656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
86606656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
86706656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
86806656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr);
86906656605SStefano Zampini       }
87006656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
87106656605SStefano Zampini     }
87206656605SStefano Zampini   }
87306656605SStefano Zampini 
87406656605SStefano Zampini   /* compute other basis functions for non-symmetric problems */
87506656605SStefano Zampini   if (!pcbddc->issym) {
87606656605SStefano Zampini     Mat B,X;
87706656605SStefano Zampini 
87806656605SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work,&B);CHKERRQ(ierr);
87906656605SStefano Zampini 
88006656605SStefano Zampini     if (n_constraints) {
88116f15bc4SStefano Zampini       Mat S_CCT,B_C;
88206656605SStefano Zampini 
88306656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr);
88406656605SStefano Zampini       ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr);
88506656605SStefano Zampini       ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr);
88616f15bc4SStefano Zampini       ierr = MatDestroy(&S_CCT);CHKERRQ(ierr);
88706656605SStefano Zampini       if (n_vertices) {
88816f15bc4SStefano Zampini         Mat B_V,S_VCT;
88906656605SStefano Zampini 
89006656605SStefano Zampini         ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr);
89106656605SStefano Zampini         ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr);
89206656605SStefano Zampini         ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr);
89306656605SStefano Zampini         ierr = MatDestroy(&B_V);CHKERRQ(ierr);
89416f15bc4SStefano Zampini         ierr = MatDestroy(&S_VCT);CHKERRQ(ierr);
89506656605SStefano Zampini       }
89606656605SStefano Zampini       ierr = MatDestroy(&B_C);CHKERRQ(ierr);
89706656605SStefano Zampini     }
89816f15bc4SStefano Zampini     if (n_vertices && n_R) {
89906656605SStefano Zampini       Mat          A_VRT;
90006656605SStefano Zampini       PetscBLASInt B_N,B_one = 1;
90106656605SStefano Zampini 
90206656605SStefano Zampini       if (!n_constraints) { /* if there are no constraints, reset work */
90306656605SStefano Zampini         ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr);
90406656605SStefano Zampini       }
90506656605SStefano Zampini       ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+pcbddc->local_primal_size*n_R,&A_VRT);CHKERRQ(ierr);
90606656605SStefano Zampini       ierr = MatTranspose(A_VR,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr);
90706656605SStefano Zampini       ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr);
90806656605SStefano Zampini       PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,work+pcbddc->local_primal_size*n_R,&B_one,work,&B_one));
90916f15bc4SStefano Zampini       ierr = MatDestroy(&A_VRT);CHKERRQ(ierr);
91006656605SStefano Zampini     }
91106656605SStefano Zampini 
91206656605SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,pcbddc->local_primal_size,work+pcbddc->local_primal_size*n_R,&X);CHKERRQ(ierr);
91306656605SStefano Zampini     if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */
91406656605SStefano Zampini       for (i=0;i<pcbddc->local_primal_size;i++) {
91506656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr);
91606656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr);
91706656605SStefano Zampini         ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
91806656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
91906656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
92006656605SStefano Zampini       }
92106656605SStefano Zampini     } else {
92206656605SStefano Zampini       for (i=0;i<pcbddc->local_primal_size;i++) {
92306656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr);
92406656605SStefano Zampini         ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr);
92506656605SStefano Zampini         ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
92606656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
92706656605SStefano Zampini         ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr);
92806656605SStefano Zampini       }
92906656605SStefano Zampini     }
93006656605SStefano Zampini     ierr = MatDestroy(&B);CHKERRQ(ierr);
93106656605SStefano Zampini     /* coarse basis functions */
93206656605SStefano Zampini     for (i=0;i<pcbddc->local_primal_size;i++) {
93306656605SStefano Zampini       PetscScalar *y;
93406656605SStefano Zampini 
93506656605SStefano Zampini       ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr);
93606656605SStefano Zampini       ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr);
93706656605SStefano Zampini       ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr);
93806656605SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
93906656605SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
94006656605SStefano Zampini       if (i<n_vertices) {
94106656605SStefano Zampini         y[n_B*i+idx_V_B[i]] = 1.0;
94206656605SStefano Zampini       }
94306656605SStefano Zampini       ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr);
94406656605SStefano Zampini       ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr);
94506656605SStefano Zampini 
94606656605SStefano Zampini       if (pcbddc->switch_static || pcbddc->dbg_flag) {
94706656605SStefano Zampini         ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr);
94806656605SStefano Zampini         ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr);
94906656605SStefano Zampini         ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
95006656605SStefano Zampini         ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
95106656605SStefano Zampini         ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr);
95206656605SStefano Zampini         ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr);
95306656605SStefano Zampini       }
95406656605SStefano Zampini       ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr);
95506656605SStefano Zampini     }
95606656605SStefano Zampini     ierr = MatDestroy(&X);CHKERRQ(ierr);
95706656605SStefano Zampini   }
95806656605SStefano Zampini 
95988ebb749SStefano Zampini   ierr = PetscFree(idx_V_B);CHKERRQ(ierr);
96006656605SStefano Zampini   ierr = MatDestroy(&S_VV);CHKERRQ(ierr);
96106656605SStefano Zampini   ierr = MatDestroy(&S_CV);CHKERRQ(ierr);
96206656605SStefano Zampini   ierr = MatDestroy(&S_VC);CHKERRQ(ierr);
96306656605SStefano Zampini   ierr = MatDestroy(&S_CC);CHKERRQ(ierr);
96488ebb749SStefano Zampini   /* Checking coarse_sub_mat and coarse basis functios */
96588ebb749SStefano Zampini   /* Symmetric case     : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
96688ebb749SStefano Zampini   /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
96725084f0cSStefano Zampini   if (pcbddc->dbg_flag) {
96888ebb749SStefano Zampini     Mat         coarse_sub_mat;
96925084f0cSStefano Zampini     Mat         AUXMAT,TM1,TM2,TM3,TM4;
97088ebb749SStefano Zampini     Mat         coarse_phi_D,coarse_phi_B;
97188ebb749SStefano Zampini     Mat         coarse_psi_D,coarse_psi_B;
97288ebb749SStefano Zampini     Mat         A_II,A_BB,A_IB,A_BI;
9738bec7fa6SStefano Zampini     Mat         C_B,CPHI;
9748bec7fa6SStefano Zampini     IS          is_dummy;
9758bec7fa6SStefano Zampini     Vec         mones;
97688ebb749SStefano Zampini     MatType     checkmattype=MATSEQAIJ;
97788ebb749SStefano Zampini     PetscReal   real_value;
97888ebb749SStefano Zampini 
97988ebb749SStefano Zampini     ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr);
98088ebb749SStefano Zampini     ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr);
98188ebb749SStefano Zampini     ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr);
98288ebb749SStefano Zampini     ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr);
98388ebb749SStefano Zampini     ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr);
98488ebb749SStefano Zampini     ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr);
985c0553b1fSStefano Zampini     if (unsymmetric_check) {
98688ebb749SStefano Zampini       ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr);
98788ebb749SStefano Zampini       ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr);
98888ebb749SStefano Zampini     }
98988ebb749SStefano Zampini     ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr);
99088ebb749SStefano Zampini 
99125084f0cSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
99206656605SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr);
99325084f0cSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
994c0553b1fSStefano Zampini     if (unsymmetric_check) {
99588ebb749SStefano Zampini       ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
99688ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr);
99788ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
99888ebb749SStefano Zampini       ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
99988ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr);
100088ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
100188ebb749SStefano Zampini       ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
100288ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr);
100388ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
100488ebb749SStefano Zampini       ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
100588ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr);
100688ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
100788ebb749SStefano Zampini     } else {
100888ebb749SStefano Zampini       ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr);
100988ebb749SStefano Zampini       ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr);
101088ebb749SStefano Zampini       ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
101188ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr);
101288ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
101388ebb749SStefano Zampini       ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr);
101488ebb749SStefano Zampini       ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr);
101588ebb749SStefano Zampini       ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr);
101688ebb749SStefano Zampini     }
101788ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
101888ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
101988ebb749SStefano Zampini     ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
102088ebb749SStefano Zampini     ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr);
102181d9aea3SBarry Smith     ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
10228bec7fa6SStefano Zampini     ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
10230fccc4e9SStefano Zampini     ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
102406656605SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d      matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr);
10258bec7fa6SStefano Zampini 
10268bec7fa6SStefano Zampini     /* check constraints */
10278bec7fa6SStefano Zampini     ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr);
10288bec7fa6SStefano Zampini     ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);
10298bec7fa6SStefano Zampini     ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr);
10308bec7fa6SStefano Zampini     ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr);
10318bec7fa6SStefano Zampini     ierr = VecSet(mones,-1.0);CHKERRQ(ierr);
10328bec7fa6SStefano Zampini     ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr);
10338bec7fa6SStefano Zampini     ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr);
10348bec7fa6SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr);
10358bec7fa6SStefano Zampini     ierr = MatDestroy(&C_B);CHKERRQ(ierr);
10368bec7fa6SStefano Zampini     ierr = MatDestroy(&CPHI);CHKERRQ(ierr);
10378bec7fa6SStefano Zampini     ierr = ISDestroy(&is_dummy);CHKERRQ(ierr);
10388bec7fa6SStefano Zampini     ierr = VecDestroy(&mones);CHKERRQ(ierr);
10398bec7fa6SStefano Zampini 
104025084f0cSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
104188ebb749SStefano Zampini     ierr = MatDestroy(&A_II);CHKERRQ(ierr);
104288ebb749SStefano Zampini     ierr = MatDestroy(&A_BB);CHKERRQ(ierr);
104388ebb749SStefano Zampini     ierr = MatDestroy(&A_IB);CHKERRQ(ierr);
104488ebb749SStefano Zampini     ierr = MatDestroy(&A_BI);CHKERRQ(ierr);
104588ebb749SStefano Zampini     ierr = MatDestroy(&TM1);CHKERRQ(ierr);
104688ebb749SStefano Zampini     ierr = MatDestroy(&TM2);CHKERRQ(ierr);
104788ebb749SStefano Zampini     ierr = MatDestroy(&TM3);CHKERRQ(ierr);
104888ebb749SStefano Zampini     ierr = MatDestroy(&TM4);CHKERRQ(ierr);
104988ebb749SStefano Zampini     ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr);
105088ebb749SStefano Zampini     ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr);
1051c0553b1fSStefano Zampini     if (unsymmetric_check) {
105288ebb749SStefano Zampini       ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr);
105388ebb749SStefano Zampini       ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr);
105488ebb749SStefano Zampini     }
105588ebb749SStefano Zampini     ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr);
105688ebb749SStefano Zampini   }
105706656605SStefano Zampini 
105888ebb749SStefano Zampini   /* free memory */
105906656605SStefano Zampini   ierr = PetscFree(work);CHKERRQ(ierr);
106088ebb749SStefano Zampini   if (n_vertices) {
106188ebb749SStefano Zampini     ierr = MatDestroy(&A_VR);CHKERRQ(ierr);
106288ebb749SStefano Zampini   }
106388ebb749SStefano Zampini   if (n_constraints) {
106488ebb749SStefano Zampini     ierr = MatDestroy(&C_CR);CHKERRQ(ierr);
106588ebb749SStefano Zampini   }
10668629588bSStefano Zampini   /* get back data */
10678629588bSStefano Zampini   *coarse_submat_vals_n = coarse_submat_vals;
106888ebb749SStefano Zampini   PetscFunctionReturn(0);
106988ebb749SStefano Zampini }
107088ebb749SStefano Zampini 
107188ebb749SStefano Zampini #undef __FUNCT__
1072d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted"
1073d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B)
1074d65f70fdSStefano Zampini {
1075d65f70fdSStefano Zampini   Mat            *work_mat;
1076d65f70fdSStefano Zampini   IS             isrow_s,iscol_s;
1077d65f70fdSStefano Zampini   PetscBool      rsorted,csorted;
1078d65f70fdSStefano Zampini   PetscInt       rsize,*idxs_perm_r,csize,*idxs_perm_c;
1079d65f70fdSStefano Zampini   PetscErrorCode ierr;
1080d65f70fdSStefano Zampini 
1081d65f70fdSStefano Zampini   PetscFunctionBegin;
1082d65f70fdSStefano Zampini   ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr);
1083d65f70fdSStefano Zampini   ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr);
1084d65f70fdSStefano Zampini   ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr);
1085d65f70fdSStefano Zampini   ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr);
1086d65f70fdSStefano Zampini 
1087d65f70fdSStefano Zampini   if (!rsorted) {
1088d65f70fdSStefano Zampini     const PetscInt *idxs;
1089d65f70fdSStefano Zampini     PetscInt *idxs_sorted,i;
1090d65f70fdSStefano Zampini 
1091d65f70fdSStefano Zampini     ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr);
1092d65f70fdSStefano Zampini     ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr);
1093d65f70fdSStefano Zampini     for (i=0;i<rsize;i++) {
1094d65f70fdSStefano Zampini       idxs_perm_r[i] = i;
1095d65f70fdSStefano Zampini     }
1096d65f70fdSStefano Zampini     ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr);
1097d65f70fdSStefano Zampini     ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr);
1098d65f70fdSStefano Zampini     for (i=0;i<rsize;i++) {
1099d65f70fdSStefano Zampini       idxs_sorted[i] = idxs[idxs_perm_r[i]];
1100d65f70fdSStefano Zampini     }
1101d65f70fdSStefano Zampini     ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr);
1102d65f70fdSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr);
1103d65f70fdSStefano Zampini   } else {
1104d65f70fdSStefano Zampini     ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
1105d65f70fdSStefano Zampini     isrow_s = isrow;
1106d65f70fdSStefano Zampini   }
1107d65f70fdSStefano Zampini 
1108d65f70fdSStefano Zampini   if (!csorted) {
1109d65f70fdSStefano Zampini     if (isrow == iscol) {
1110d65f70fdSStefano Zampini       ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr);
1111d65f70fdSStefano Zampini       iscol_s = isrow_s;
1112d65f70fdSStefano Zampini     } else {
1113d65f70fdSStefano Zampini       const PetscInt *idxs;
1114d65f70fdSStefano Zampini       PetscInt *idxs_sorted,i;
1115d65f70fdSStefano Zampini 
1116d65f70fdSStefano Zampini       ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr);
1117d65f70fdSStefano Zampini       ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr);
1118d65f70fdSStefano Zampini       for (i=0;i<csize;i++) {
1119d65f70fdSStefano Zampini         idxs_perm_c[i] = i;
1120d65f70fdSStefano Zampini       }
1121d65f70fdSStefano Zampini       ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr);
1122d65f70fdSStefano Zampini       ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr);
1123d65f70fdSStefano Zampini       for (i=0;i<csize;i++) {
1124d65f70fdSStefano Zampini         idxs_sorted[i] = idxs[idxs_perm_c[i]];
1125d65f70fdSStefano Zampini       }
1126d65f70fdSStefano Zampini       ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr);
1127d65f70fdSStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr);
1128d65f70fdSStefano Zampini     }
1129d65f70fdSStefano Zampini   } else {
1130d65f70fdSStefano Zampini     ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
1131d65f70fdSStefano Zampini     iscol_s = iscol;
1132d65f70fdSStefano Zampini   }
1133d65f70fdSStefano Zampini 
1134d648f858SStefano Zampini   ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr);
1135d65f70fdSStefano Zampini 
1136d65f70fdSStefano Zampini   if (!rsorted || !csorted) {
1137d65f70fdSStefano Zampini     Mat      new_mat;
1138d65f70fdSStefano Zampini     IS       is_perm_r,is_perm_c;
1139d65f70fdSStefano Zampini 
1140d65f70fdSStefano Zampini     if (!rsorted) {
1141d65f70fdSStefano Zampini       PetscInt *idxs_r,i;
1142d65f70fdSStefano Zampini       ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr);
1143d65f70fdSStefano Zampini       for (i=0;i<rsize;i++) {
1144d65f70fdSStefano Zampini         idxs_r[idxs_perm_r[i]] = i;
1145d65f70fdSStefano Zampini       }
1146d65f70fdSStefano Zampini       ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr);
1147d65f70fdSStefano Zampini       ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr);
1148d65f70fdSStefano Zampini     } else {
1149d65f70fdSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr);
1150d65f70fdSStefano Zampini     }
1151d65f70fdSStefano Zampini     ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr);
1152d65f70fdSStefano Zampini 
1153d65f70fdSStefano Zampini     if (!csorted) {
1154d65f70fdSStefano Zampini       if (isrow_s == iscol_s) {
1155d65f70fdSStefano Zampini         ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr);
1156d65f70fdSStefano Zampini         is_perm_c = is_perm_r;
1157d65f70fdSStefano Zampini       } else {
1158d65f70fdSStefano Zampini         PetscInt *idxs_c,i;
1159d65f70fdSStefano Zampini         ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr);
1160d65f70fdSStefano Zampini         for (i=0;i<csize;i++) {
1161d65f70fdSStefano Zampini           idxs_c[idxs_perm_c[i]] = i;
1162d65f70fdSStefano Zampini         }
1163d65f70fdSStefano Zampini         ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr);
1164d65f70fdSStefano Zampini         ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr);
1165d65f70fdSStefano Zampini       }
1166d65f70fdSStefano Zampini     } else {
1167d65f70fdSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr);
1168d65f70fdSStefano Zampini     }
1169d65f70fdSStefano Zampini     ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr);
1170d65f70fdSStefano Zampini 
1171d65f70fdSStefano Zampini     ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr);
1172d65f70fdSStefano Zampini     ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr);
1173d65f70fdSStefano Zampini     work_mat[0] = new_mat;
1174d65f70fdSStefano Zampini     ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr);
1175d65f70fdSStefano Zampini     ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr);
1176d65f70fdSStefano Zampini   }
1177d65f70fdSStefano Zampini 
1178d65f70fdSStefano Zampini   ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr);
1179d65f70fdSStefano Zampini   *B = work_mat[0];
1180d65f70fdSStefano Zampini   ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr);
1181d65f70fdSStefano Zampini   ierr = ISDestroy(&isrow_s);CHKERRQ(ierr);
1182d65f70fdSStefano Zampini   ierr = ISDestroy(&iscol_s);CHKERRQ(ierr);
1183d65f70fdSStefano Zampini   PetscFunctionReturn(0);
1184d65f70fdSStefano Zampini }
1185d65f70fdSStefano Zampini 
1186d65f70fdSStefano Zampini #undef __FUNCT__
11875e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix"
11885e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix)
1189aa0d41d4SStefano Zampini {
1190aa0d41d4SStefano Zampini   Mat_IS*        matis = (Mat_IS*)pc->pmat->data;
11915e8657edSStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
1192d65f70fdSStefano Zampini   Mat            new_mat;
11935e8657edSStefano Zampini   IS             is_local,is_global;
1194d65f70fdSStefano Zampini   PetscInt       local_size;
1195d65f70fdSStefano Zampini   PetscBool      isseqaij;
1196aa0d41d4SStefano Zampini   PetscErrorCode ierr;
1197aa0d41d4SStefano Zampini 
1198aa0d41d4SStefano Zampini   PetscFunctionBegin;
1199aa0d41d4SStefano Zampini   ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
12005e8657edSStefano Zampini   ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr);
12015e8657edSStefano Zampini   ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr);
1202906d46d4SStefano Zampini   ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr);
1203906d46d4SStefano Zampini   ierr = ISDestroy(&is_local);CHKERRQ(ierr);
1204d648f858SStefano Zampini   ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr);
1205906d46d4SStefano Zampini   ierr = ISDestroy(&is_global);CHKERRQ(ierr);
1206906d46d4SStefano Zampini 
1207906d46d4SStefano Zampini   /* check */
1208906d46d4SStefano Zampini   if (pcbddc->dbg_flag) {
1209906d46d4SStefano Zampini     Vec       x,x_change;
1210906d46d4SStefano Zampini     PetscReal error;
1211906d46d4SStefano Zampini 
12125e8657edSStefano Zampini     ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr);
1213906d46d4SStefano Zampini     ierr = VecSetRandom(x,NULL);CHKERRQ(ierr);
12145e8657edSStefano Zampini     ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr);
12155e8657edSStefano Zampini     ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
12165e8657edSStefano Zampini     ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1217d65f70fdSStefano Zampini     ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr);
12185e8657edSStefano Zampini     ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
12195e8657edSStefano Zampini     ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
1220906d46d4SStefano Zampini     ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr);
1221906d46d4SStefano Zampini     ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr);
1222906d46d4SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1223906d46d4SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr);
1224906d46d4SStefano Zampini     ierr = VecDestroy(&x);CHKERRQ(ierr);
1225906d46d4SStefano Zampini     ierr = VecDestroy(&x_change);CHKERRQ(ierr);
1226906d46d4SStefano Zampini   }
1227906d46d4SStefano Zampini 
122822d5777bSStefano Zampini   /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */
12299b28b3ffSStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
123022d5777bSStefano Zampini   if (isseqaij) {
1231d65f70fdSStefano Zampini     ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr);
1232aa0d41d4SStefano Zampini   } else {
1233aa0d41d4SStefano Zampini     Mat work_mat;
1234aa0d41d4SStefano Zampini     ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr);
1235d65f70fdSStefano Zampini     ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr);
1236aa0d41d4SStefano Zampini     ierr = MatDestroy(&work_mat);CHKERRQ(ierr);
1237aa0d41d4SStefano Zampini   }
1238d65f70fdSStefano Zampini   ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr);
1239e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX)
1240e496cd5dSStefano Zampini   ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr);
1241e496cd5dSStefano Zampini #endif
124245a1bb75SStefano Zampini   /*
124345a1bb75SStefano Zampini   ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
1244d65f70fdSStefano Zampini   ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr);
124545a1bb75SStefano Zampini   */
1246d65f70fdSStefano Zampini   ierr = MatDestroy(&new_mat);CHKERRQ(ierr);
1247aa0d41d4SStefano Zampini   PetscFunctionReturn(0);
1248aa0d41d4SStefano Zampini }
1249aa0d41d4SStefano Zampini 
1250aa0d41d4SStefano Zampini #undef __FUNCT__
1251a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters"
12528ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc)
1253a64d13efSStefano Zampini {
1254a64d13efSStefano Zampini   PC_IS*         pcis = (PC_IS*)(pc->data);
1255a64d13efSStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
12568ce42a96SStefano Zampini   IS             is_aux1,is_aux2;
1257e9189074SStefano Zampini   PetscInt       *aux_array1,*aux_array2,*is_indices,*idx_R_local;
12583a50541eSStefano Zampini   PetscInt       n_vertices,i,j,n_R,n_D,n_B;
12593a50541eSStefano Zampini   PetscInt       vbs,bs;
12604641a718SStefano Zampini   PetscBT        bitmask;
1261a64d13efSStefano Zampini   PetscErrorCode ierr;
1262a64d13efSStefano Zampini 
1263a64d13efSStefano Zampini   PetscFunctionBegin;
1264b23d619eSStefano Zampini   /*
1265b23d619eSStefano Zampini     No need to setup local scatters if
1266b23d619eSStefano Zampini       - primal space is unchanged
1267b23d619eSStefano Zampini         AND
1268b23d619eSStefano Zampini       - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains)
1269b23d619eSStefano Zampini         AND
1270b23d619eSStefano Zampini       - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine
1271b23d619eSStefano Zampini   */
1272b23d619eSStefano Zampini   if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) {
1273f4ddd8eeSStefano Zampini     PetscFunctionReturn(0);
1274f4ddd8eeSStefano Zampini   }
1275f4ddd8eeSStefano Zampini   /* destroy old objects */
1276f4ddd8eeSStefano Zampini   ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr);
1277f4ddd8eeSStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr);
1278f4ddd8eeSStefano Zampini   ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr);
1279a64d13efSStefano Zampini   /* Set Non-overlapping dimensions */
1280a64d13efSStefano Zampini   n_B = pcis->n_B; n_D = pcis->n - n_B;
1281e9189074SStefano Zampini   n_vertices = pcbddc->n_actual_vertices;
12824641a718SStefano Zampini   /* create auxiliary bitmask */
12834641a718SStefano Zampini   ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr);
12844641a718SStefano Zampini   for (i=0;i<n_vertices;i++) {
1285e9189074SStefano Zampini     ierr = PetscBTSet(bitmask,pcbddc->primal_indices_local_idxs[i]);CHKERRQ(ierr);
12864641a718SStefano Zampini   }
12873a50541eSStefano Zampini 
1288a64d13efSStefano Zampini   /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */
1289854ce69bSBarry Smith   ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr);
1290a64d13efSStefano Zampini   for (i=0, n_R=0; i<pcis->n; i++) {
12914641a718SStefano Zampini     if (!PetscBTLookup(bitmask,i)) {
1292a64d13efSStefano Zampini       idx_R_local[n_R] = i;
1293a64d13efSStefano Zampini       n_R++;
1294a64d13efSStefano Zampini     }
1295a64d13efSStefano Zampini   }
12963a50541eSStefano Zampini 
12973a50541eSStefano Zampini   /* Block code */
12983a50541eSStefano Zampini   vbs = 1;
12993a50541eSStefano Zampini   ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr);
13003a50541eSStefano Zampini   if (bs>1 && !(n_vertices%bs)) {
13013a50541eSStefano Zampini     PetscBool is_blocked = PETSC_TRUE;
13023a50541eSStefano Zampini     PetscInt  *vary;
13033a50541eSStefano Zampini     /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */
1304785e854fSJed Brown     ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr);
13053a50541eSStefano Zampini     ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr);
1306e9189074SStefano Zampini     for (i=0; i<n_vertices; i++) vary[pcbddc->primal_indices_local_idxs[i]/bs]++;
1307964fefecSStefano Zampini     for (i=0; i<n_vertices/bs; i++) {
13083a50541eSStefano Zampini       if (vary[i]!=0 && vary[i]!=bs) {
13093a50541eSStefano Zampini         is_blocked = PETSC_FALSE;
13103a50541eSStefano Zampini         break;
13113a50541eSStefano Zampini       }
13123a50541eSStefano Zampini     }
13133a50541eSStefano Zampini     if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */
13143a50541eSStefano Zampini       vbs = bs;
13153a50541eSStefano Zampini       for (i=0;i<n_R/vbs;i++) {
13163a50541eSStefano Zampini         idx_R_local[i] = idx_R_local[vbs*i]/vbs;
13173a50541eSStefano Zampini       }
13183a50541eSStefano Zampini     }
13193a50541eSStefano Zampini     ierr = PetscFree(vary);CHKERRQ(ierr);
13203a50541eSStefano Zampini   }
13213a50541eSStefano Zampini   ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr);
13223a50541eSStefano Zampini   ierr = PetscFree(idx_R_local);CHKERRQ(ierr);
1323a64d13efSStefano Zampini 
1324a64d13efSStefano Zampini   /* print some info if requested */
1325a64d13efSStefano Zampini   if (pcbddc->dbg_flag) {
1326a64d13efSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
1327a64d13efSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
13280fccc4e9SStefano Zampini     ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
1329a64d13efSStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr);
1330a64d13efSStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr);
13313a50541eSStefano 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->local_primal_size);CHKERRQ(ierr);
1332a64d13efSStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"pcbddc->n_vertices = %d, pcbddc->n_constraints = %d\n",pcbddc->n_vertices,pcbddc->n_constraints);CHKERRQ(ierr);
1333a64d13efSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1334a64d13efSStefano Zampini   }
1335a64d13efSStefano Zampini 
1336a64d13efSStefano Zampini   /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */
13373a50541eSStefano Zampini   ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
1338854ce69bSBarry Smith   ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr);
1339854ce69bSBarry Smith   ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr);
1340a64d13efSStefano Zampini   ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
13414641a718SStefano Zampini   for (i=0; i<n_D; i++) {
13424641a718SStefano Zampini     ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr);
13434641a718SStefano Zampini   }
1344a64d13efSStefano Zampini   ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
1345a64d13efSStefano Zampini   for (i=0, j=0; i<n_R; i++) {
13464641a718SStefano Zampini     if (!PetscBTLookup(bitmask,idx_R_local[i])) {
13474641a718SStefano Zampini       aux_array1[j++] = i;
1348a64d13efSStefano Zampini     }
1349a64d13efSStefano Zampini   }
1350a64d13efSStefano Zampini   ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr);
1351a64d13efSStefano Zampini   ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
1352a64d13efSStefano Zampini   for (i=0, j=0; i<n_B; i++) {
13534641a718SStefano Zampini     if (!PetscBTLookup(bitmask,is_indices[i])) {
13544641a718SStefano Zampini       aux_array2[j++] = i;
1355a64d13efSStefano Zampini     }
1356a64d13efSStefano Zampini   }
1357a64d13efSStefano Zampini   ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr);
1358a64d13efSStefano Zampini   ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr);
1359a64d13efSStefano Zampini   ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr);
1360a64d13efSStefano Zampini   ierr = ISDestroy(&is_aux1);CHKERRQ(ierr);
1361a64d13efSStefano Zampini   ierr = ISDestroy(&is_aux2);CHKERRQ(ierr);
1362a64d13efSStefano Zampini 
13638eeda7d8SStefano Zampini   if (pcbddc->switch_static || pcbddc->dbg_flag) {
1364785e854fSJed Brown     ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr);
1365a64d13efSStefano Zampini     for (i=0, j=0; i<n_R; i++) {
13664641a718SStefano Zampini       if (PetscBTLookup(bitmask,idx_R_local[i])) {
13674641a718SStefano Zampini         aux_array1[j++] = i;
1368a64d13efSStefano Zampini       }
1369a64d13efSStefano Zampini     }
1370a64d13efSStefano Zampini     ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr);
1371a64d13efSStefano Zampini     ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr);
1372a64d13efSStefano Zampini     ierr = ISDestroy(&is_aux1);CHKERRQ(ierr);
1373a64d13efSStefano Zampini   }
13744641a718SStefano Zampini   ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr);
13753a50541eSStefano Zampini   ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr);
1376a64d13efSStefano Zampini   PetscFunctionReturn(0);
1377a64d13efSStefano Zampini }
1378a64d13efSStefano Zampini 
1379304d26faSStefano Zampini 
1380304d26faSStefano Zampini #undef __FUNCT__
1381304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers"
1382684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann)
1383304d26faSStefano Zampini {
1384304d26faSStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
1385304d26faSStefano Zampini   PC_IS          *pcis = (PC_IS*)pc->data;
1386304d26faSStefano Zampini   PC             pc_temp;
1387304d26faSStefano Zampini   Mat            A_RR;
1388f4ddd8eeSStefano Zampini   MatReuse       reuse;
1389304d26faSStefano Zampini   PetscScalar    m_one = -1.0;
1390304d26faSStefano Zampini   PetscReal      value;
1391af732b37SStefano Zampini   PetscInt       n_D,n_R,ibs,mbs;
13929577ea80SStefano Zampini   PetscBool      use_exact,use_exact_reduced,issbaij;
1393304d26faSStefano Zampini   PetscErrorCode ierr;
1394e604994aSStefano Zampini   /* prefixes stuff */
1395312be037SStefano Zampini   char           dir_prefix[256],neu_prefix[256],str_level[16];
1396e604994aSStefano Zampini   size_t         len;
1397304d26faSStefano Zampini 
1398304d26faSStefano Zampini   PetscFunctionBegin;
1399304d26faSStefano Zampini 
1400e604994aSStefano Zampini   /* compute prefixes */
1401e604994aSStefano Zampini   ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr);
1402e604994aSStefano Zampini   ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr);
1403e604994aSStefano Zampini   if (!pcbddc->current_level) {
1404e604994aSStefano Zampini     ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
1405e604994aSStefano Zampini     ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
1406e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr);
1407e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr);
1408e604994aSStefano Zampini   } else {
1409e604994aSStefano Zampini     ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr);
1410312be037SStefano Zampini     sprintf(str_level,"l%d_",(int)(pcbddc->current_level));
1411e604994aSStefano Zampini     ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr);
1412e604994aSStefano Zampini     len -= 15; /* remove "pc_bddc_coarse_" */
1413312be037SStefano Zampini     if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
1414312be037SStefano Zampini     if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
141534d6797cSStefano Zampini     ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
141634d6797cSStefano Zampini     ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
1417e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr);
1418e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr);
1419e604994aSStefano Zampini     ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr);
1420e604994aSStefano Zampini     ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr);
1421e604994aSStefano Zampini   }
1422e604994aSStefano Zampini 
1423304d26faSStefano Zampini   /* DIRICHLET PROBLEM */
1424684f6988SStefano Zampini   if (dirichlet) {
1425d5574798SStefano Zampini     PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
1426964fefecSStefano Zampini     if (pcbddc->issym) {
1427964fefecSStefano Zampini       ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr);
1428964fefecSStefano Zampini     }
1429ac78edfcSStefano Zampini     /* Matrix for Dirichlet problem is pcis->A_II */
1430964fefecSStefano Zampini     n_D = pcis->n - pcis->n_B;
1431304d26faSStefano Zampini     if (!pcbddc->ksp_D) { /* create object if not yet build */
1432304d26faSStefano Zampini       ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr);
1433304d26faSStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr);
1434304d26faSStefano Zampini       /* default */
1435304d26faSStefano Zampini       ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr);
1436e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr);
14379577ea80SStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
1438304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr);
14399577ea80SStefano Zampini       if (issbaij) {
14409577ea80SStefano Zampini         ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr);
14419577ea80SStefano Zampini       } else {
1442304d26faSStefano Zampini         ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr);
14439577ea80SStefano Zampini       }
1444304d26faSStefano Zampini       /* Allow user's customization */
1445304d26faSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr);
1446304d26faSStefano Zampini       ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
1447304d26faSStefano Zampini     }
1448d1e9a80fSBarry Smith     ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr);
1449d5574798SStefano Zampini     if (sub_schurs->interior_solver) {
1450d5574798SStefano Zampini       ierr = KSPSetPC(pcbddc->ksp_D,sub_schurs->interior_solver);CHKERRQ(ierr);
1451d5574798SStefano Zampini     }
1452304d26faSStefano Zampini     /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
1453304d26faSStefano Zampini     if (!n_D) {
1454304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr);
1455304d26faSStefano Zampini       ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr);
1456304d26faSStefano Zampini     }
1457304d26faSStefano Zampini     /* Set Up KSP for Dirichlet problem of BDDC */
1458304d26faSStefano Zampini     ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr);
1459304d26faSStefano Zampini     /* set ksp_D into pcis data */
1460304d26faSStefano Zampini     ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr);
1461304d26faSStefano Zampini     ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr);
1462304d26faSStefano Zampini     pcis->ksp_D = pcbddc->ksp_D;
1463684f6988SStefano Zampini   }
1464304d26faSStefano Zampini 
1465304d26faSStefano Zampini   /* NEUMANN PROBLEM */
1466684f6988SStefano Zampini   A_RR = 0;
1467684f6988SStefano Zampini   if (neumann) {
1468f4ddd8eeSStefano Zampini     /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */
14698ce42a96SStefano Zampini     ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr);
1470f4ddd8eeSStefano Zampini     if (pcbddc->ksp_R) { /* already created ksp */
1471f4ddd8eeSStefano Zampini       PetscInt nn_R;
147281d9aea3SBarry Smith       ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr);
1473f4ddd8eeSStefano Zampini       ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr);
1474f4ddd8eeSStefano Zampini       ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr);
1475f4ddd8eeSStefano Zampini       if (nn_R != n_R) { /* old ksp is not reusable, so reset it */
1476f4ddd8eeSStefano Zampini         ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr);
1477f4ddd8eeSStefano Zampini         ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
1478f4ddd8eeSStefano Zampini         reuse = MAT_INITIAL_MATRIX;
1479f4ddd8eeSStefano Zampini       } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */
1480727cdba6SStefano Zampini         if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */
1481f4ddd8eeSStefano Zampini           ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
1482f4ddd8eeSStefano Zampini           reuse = MAT_INITIAL_MATRIX;
1483f4ddd8eeSStefano Zampini         } else { /* safe to reuse the matrix */
1484f4ddd8eeSStefano Zampini           reuse = MAT_REUSE_MATRIX;
1485f4ddd8eeSStefano Zampini         }
1486f4ddd8eeSStefano Zampini       }
1487f4ddd8eeSStefano Zampini       /* last check */
1488d1e9a80fSBarry Smith       if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
1489f4ddd8eeSStefano Zampini         ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
1490f4ddd8eeSStefano Zampini         reuse = MAT_INITIAL_MATRIX;
1491f4ddd8eeSStefano Zampini       }
1492f4ddd8eeSStefano Zampini     } else { /* first time, so we need to create the matrix */
1493f4ddd8eeSStefano Zampini       reuse = MAT_INITIAL_MATRIX;
1494f4ddd8eeSStefano Zampini     }
1495f4ddd8eeSStefano Zampini     /* extract A_RR */
1496af732b37SStefano Zampini     ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr);
1497af732b37SStefano Zampini     ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr);
14983a50541eSStefano Zampini     if (ibs != mbs) {
1499af732b37SStefano Zampini       Mat newmat;
1500af732b37SStefano Zampini       ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr);
1501f4ddd8eeSStefano Zampini       ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr);
1502af732b37SStefano Zampini       ierr = MatDestroy(&newmat);CHKERRQ(ierr);
1503af732b37SStefano Zampini     } else {
1504f4ddd8eeSStefano Zampini       ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr);
1505af732b37SStefano Zampini     }
1506964fefecSStefano Zampini     if (pcbddc->issym) {
1507964fefecSStefano Zampini       ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr);
1508964fefecSStefano Zampini     }
1509f4ddd8eeSStefano Zampini     if (!pcbddc->ksp_R) { /* create object if not present */
1510304d26faSStefano Zampini       ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr);
1511304d26faSStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr);
1512304d26faSStefano Zampini       /* default */
1513304d26faSStefano Zampini       ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr);
1514e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr);
1515304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr);
15169577ea80SStefano Zampini       ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
15179577ea80SStefano Zampini       if (issbaij) {
15189577ea80SStefano Zampini         ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr);
15199577ea80SStefano Zampini       } else {
1520304d26faSStefano Zampini         ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr);
15219577ea80SStefano Zampini       }
1522304d26faSStefano Zampini       /* Allow user's customization */
1523304d26faSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr);
1524304d26faSStefano Zampini       ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
1525304d26faSStefano Zampini     }
1526d1e9a80fSBarry Smith     ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr);
1527304d26faSStefano Zampini     /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
1528304d26faSStefano Zampini     if (!n_R) {
1529304d26faSStefano Zampini       ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr);
1530304d26faSStefano Zampini       ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr);
1531304d26faSStefano Zampini     }
1532304d26faSStefano Zampini     /* Set Up KSP for Neumann problem of BDDC */
1533304d26faSStefano Zampini     ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr);
1534684f6988SStefano Zampini   }
1535304d26faSStefano Zampini 
1536304d26faSStefano Zampini   /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */
15370fccc4e9SStefano Zampini   if (pcbddc->NullSpace || pcbddc->dbg_flag) {
1538684f6988SStefano Zampini     if (pcbddc->dbg_flag) {
1539684f6988SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1540684f6988SStefano Zampini       ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
1541684f6988SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
1542684f6988SStefano Zampini     }
1543684f6988SStefano Zampini     if (dirichlet) { /* Dirichlet */
15440fccc4e9SStefano Zampini       ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr);
15450fccc4e9SStefano Zampini       ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr);
15460fccc4e9SStefano Zampini       ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr);
15470fccc4e9SStefano Zampini       ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr);
15480fccc4e9SStefano Zampini       ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr);
1549304d26faSStefano Zampini       /* need to be adapted? */
1550b8ffe317SStefano Zampini       use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE);
1551b8ffe317SStefano Zampini       ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
1552b8ffe317SStefano Zampini       ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr);
1553304d26faSStefano Zampini       /* print info */
1554304d26faSStefano Zampini       if (pcbddc->dbg_flag) {
1555e604994aSStefano 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);
1556304d26faSStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1557304d26faSStefano Zampini       }
1558b8ffe317SStefano Zampini       if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) {
15598ce42a96SStefano Zampini         ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr);
1560304d26faSStefano Zampini       }
1561684f6988SStefano Zampini     }
1562684f6988SStefano Zampini     if (neumann) { /* Neumann */
15630fccc4e9SStefano Zampini       ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr);
15640fccc4e9SStefano Zampini       ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr);
15650fccc4e9SStefano Zampini       ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr);
15660fccc4e9SStefano Zampini       ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr);
15670fccc4e9SStefano Zampini       ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr);
1568304d26faSStefano Zampini       /* need to be adapted? */
1569b8ffe317SStefano Zampini       use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE);
1570b8ffe317SStefano Zampini       ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
1571304d26faSStefano Zampini       /* print info */
1572304d26faSStefano Zampini       if (pcbddc->dbg_flag) {
1573e604994aSStefano 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);
1574304d26faSStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1575304d26faSStefano Zampini       }
1576b8ffe317SStefano Zampini       if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */
15778ce42a96SStefano Zampini         ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr);
1578304d26faSStefano Zampini       }
15790fccc4e9SStefano Zampini     }
1580684f6988SStefano Zampini   }
1581304d26faSStefano Zampini   /* free Neumann problem's matrix */
1582304d26faSStefano Zampini   ierr = MatDestroy(&A_RR);CHKERRQ(ierr);
1583304d26faSStefano Zampini   PetscFunctionReturn(0);
1584304d26faSStefano Zampini }
1585304d26faSStefano Zampini 
1586304d26faSStefano Zampini #undef __FUNCT__
1587ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection"
158820c7b377SStefano Zampini static PetscErrorCode  PCBDDCSolveSubstructureCorrection(PC pc, Vec rhs, Vec sol, Vec work, PetscBool applytranspose)
1589674ae819SStefano Zampini {
1590674ae819SStefano Zampini   PetscErrorCode ierr;
1591674ae819SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)(pc->data);
1592674ae819SStefano Zampini 
1593674ae819SStefano Zampini   PetscFunctionBegin;
159420c7b377SStefano Zampini   if (applytranspose) {
1595674ae819SStefano Zampini     if (pcbddc->local_auxmat1) {
159620c7b377SStefano Zampini       ierr = MatMultTranspose(pcbddc->local_auxmat2,rhs,work);CHKERRQ(ierr);
159720c7b377SStefano Zampini       ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,work,rhs,rhs);CHKERRQ(ierr);
159820c7b377SStefano Zampini     }
159920c7b377SStefano Zampini     ierr = KSPSolveTranspose(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr);
160020c7b377SStefano Zampini   } else {
160120c7b377SStefano Zampini     ierr = KSPSolve(pcbddc->ksp_R,rhs,sol);CHKERRQ(ierr);
160220c7b377SStefano Zampini     if (pcbddc->local_auxmat1) {
160320c7b377SStefano Zampini       ierr = MatMult(pcbddc->local_auxmat1,sol,work);CHKERRQ(ierr);
160420c7b377SStefano Zampini       ierr = MatMultAdd(pcbddc->local_auxmat2,work,sol,sol);CHKERRQ(ierr);
160520c7b377SStefano Zampini     }
1606674ae819SStefano Zampini   }
1607674ae819SStefano Zampini   PetscFunctionReturn(0);
1608674ae819SStefano Zampini }
1609674ae819SStefano Zampini 
1610dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */
1611674ae819SStefano Zampini #undef __FUNCT__
1612674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner"
1613dc359a40SStefano Zampini PetscErrorCode  PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose)
1614674ae819SStefano Zampini {
1615674ae819SStefano Zampini   PetscErrorCode ierr;
1616674ae819SStefano Zampini   PC_BDDC*        pcbddc = (PC_BDDC*)(pc->data);
1617674ae819SStefano Zampini   PC_IS*            pcis = (PC_IS*)  (pc->data);
1618674ae819SStefano Zampini   const PetscScalar zero = 0.0;
1619674ae819SStefano Zampini 
1620674ae819SStefano Zampini   PetscFunctionBegin;
1621dc359a40SStefano Zampini   /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */
1622dc359a40SStefano Zampini   if (applytranspose) {
1623674ae819SStefano Zampini     ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr);
16248eeda7d8SStefano Zampini     if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); }
1625dc359a40SStefano Zampini   } else {
1626674ae819SStefano Zampini     ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr);
1627674ae819SStefano Zampini     if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); }
162815aaf578SStefano Zampini   }
162912edc857SStefano Zampini   /* start communications from local primal nodes to rhs of coarse solver */
163012edc857SStefano Zampini   ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr);
163112edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
163212edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
163312edc857SStefano Zampini 
16349f00e9b4SStefano Zampini   /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */
163512edc857SStefano Zampini   /* TODO remove null space when doing multilevel */
163612edc857SStefano Zampini   if (pcbddc->coarse_ksp) {
1637964fefecSStefano Zampini     Vec rhs,sol;
1638964fefecSStefano Zampini 
1639964fefecSStefano Zampini     ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr);
1640964fefecSStefano Zampini     ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr);
164112edc857SStefano Zampini     if (applytranspose) {
1642964fefecSStefano Zampini       ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr);
164312edc857SStefano Zampini     } else {
1644964fefecSStefano Zampini       ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr);
164512edc857SStefano Zampini     }
164612edc857SStefano Zampini   }
1647674ae819SStefano Zampini 
1648674ae819SStefano Zampini   /* Local solution on R nodes */
16499f00e9b4SStefano Zampini   if (pcis->n) {
1650674ae819SStefano Zampini     ierr = VecSet(pcbddc->vec1_R,zero);CHKERRQ(ierr);
1651674ae819SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
1652674ae819SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_B,pcis->vec1_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
16538eeda7d8SStefano Zampini     if (pcbddc->switch_static) {
1654674ae819SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
1655674ae819SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_D,pcis->vec1_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
1656674ae819SStefano Zampini     }
165720c7b377SStefano Zampini     ierr = PCBDDCSolveSubstructureCorrection(pc,pcbddc->vec1_R,pcbddc->vec2_R,pcbddc->vec1_C,applytranspose);CHKERRQ(ierr);
1658674ae819SStefano Zampini     ierr = VecSet(pcis->vec1_B,zero);CHKERRQ(ierr);
1659674ae819SStefano Zampini     ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1660674ae819SStefano Zampini     ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec2_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
16618eeda7d8SStefano Zampini     if (pcbddc->switch_static) {
1662674ae819SStefano Zampini       ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1663674ae819SStefano Zampini       ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec2_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1664674ae819SStefano Zampini     }
16659f00e9b4SStefano Zampini   }
1666674ae819SStefano Zampini 
16679f00e9b4SStefano Zampini   /* communications from coarse sol to local primal nodes */
16689f00e9b4SStefano Zampini   ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
166912edc857SStefano Zampini   ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
1670674ae819SStefano Zampini 
1671674ae819SStefano Zampini   /* Sum contributions from two levels */
1672dc359a40SStefano Zampini   if (applytranspose) {
1673dc359a40SStefano Zampini     ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr);
1674dc359a40SStefano Zampini     if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); }
1675dc359a40SStefano Zampini   } else {
1676674ae819SStefano Zampini     ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr);
16778eeda7d8SStefano Zampini     if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); }
1678dc359a40SStefano Zampini   }
1679674ae819SStefano Zampini   PetscFunctionReturn(0);
1680674ae819SStefano Zampini }
1681674ae819SStefano Zampini 
1682674ae819SStefano Zampini #undef __FUNCT__
1683674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin"
168412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode)
1685674ae819SStefano Zampini {
1686674ae819SStefano Zampini   PetscErrorCode ierr;
1687674ae819SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)(pc->data);
168858da7f69SStefano Zampini   PetscScalar    *array;
168912edc857SStefano Zampini   Vec            from,to;
1690674ae819SStefano Zampini 
1691674ae819SStefano Zampini   PetscFunctionBegin;
169212edc857SStefano Zampini   if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
169312edc857SStefano Zampini     from = pcbddc->coarse_vec;
169412edc857SStefano Zampini     to = pcbddc->vec1_P;
169512edc857SStefano Zampini     if (pcbddc->coarse_ksp) { /* get array from coarse processes */
169612edc857SStefano Zampini       Vec tvec;
169758da7f69SStefano Zampini 
169858da7f69SStefano Zampini       ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
169958da7f69SStefano Zampini       ierr = VecResetArray(tvec);CHKERRQ(ierr);
170012edc857SStefano Zampini       ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
170158da7f69SStefano Zampini       ierr = VecGetArray(tvec,&array);CHKERRQ(ierr);
170258da7f69SStefano Zampini       ierr = VecPlaceArray(from,array);CHKERRQ(ierr);
170358da7f69SStefano Zampini       ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr);
170412edc857SStefano Zampini     }
170512edc857SStefano Zampini   } else { /* from local to global -> put data in coarse right hand side */
170612edc857SStefano Zampini     from = pcbddc->vec1_P;
170712edc857SStefano Zampini     to = pcbddc->coarse_vec;
170812edc857SStefano Zampini   }
170912edc857SStefano Zampini   ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr);
1710674ae819SStefano Zampini   PetscFunctionReturn(0);
1711674ae819SStefano Zampini }
1712674ae819SStefano Zampini 
1713674ae819SStefano Zampini #undef __FUNCT__
1714674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd"
171512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode)
1716674ae819SStefano Zampini {
1717674ae819SStefano Zampini   PetscErrorCode ierr;
1718674ae819SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)(pc->data);
171958da7f69SStefano Zampini   PetscScalar    *array;
172012edc857SStefano Zampini   Vec            from,to;
1721674ae819SStefano Zampini 
1722674ae819SStefano Zampini   PetscFunctionBegin;
172312edc857SStefano Zampini   if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
172412edc857SStefano Zampini     from = pcbddc->coarse_vec;
172512edc857SStefano Zampini     to = pcbddc->vec1_P;
172612edc857SStefano Zampini   } else { /* from local to global -> put data in coarse right hand side */
172712edc857SStefano Zampini     from = pcbddc->vec1_P;
172812edc857SStefano Zampini     to = pcbddc->coarse_vec;
172912edc857SStefano Zampini   }
173012edc857SStefano Zampini   ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr);
173112edc857SStefano Zampini   if (smode == SCATTER_FORWARD) {
173212edc857SStefano Zampini     if (pcbddc->coarse_ksp) { /* get array from coarse processes */
173312edc857SStefano Zampini       Vec tvec;
173458da7f69SStefano Zampini 
173512edc857SStefano Zampini       ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr);
173658da7f69SStefano Zampini       ierr = VecGetArray(to,&array);CHKERRQ(ierr);
173758da7f69SStefano Zampini       ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr);
173858da7f69SStefano Zampini       ierr = VecRestoreArray(to,&array);CHKERRQ(ierr);
173958da7f69SStefano Zampini     }
174058da7f69SStefano Zampini   } else {
174158da7f69SStefano Zampini     if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */
174258da7f69SStefano Zampini      ierr = VecResetArray(from);CHKERRQ(ierr);
174312edc857SStefano Zampini     }
174412edc857SStefano Zampini   }
1745674ae819SStefano Zampini   PetscFunctionReturn(0);
1746674ae819SStefano Zampini }
1747674ae819SStefano Zampini 
1748984c4197SStefano Zampini /* uncomment for testing purposes */
1749984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */
1750674ae819SStefano Zampini #undef __FUNCT__
1751674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp"
1752674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc)
1753674ae819SStefano Zampini {
1754674ae819SStefano Zampini   PetscErrorCode    ierr;
1755674ae819SStefano Zampini   PC_IS*            pcis = (PC_IS*)(pc->data);
1756674ae819SStefano Zampini   PC_BDDC*          pcbddc = (PC_BDDC*)pc->data;
1757674ae819SStefano Zampini   Mat_IS*           matis = (Mat_IS*)pc->pmat->data;
1758984c4197SStefano Zampini   /* one and zero */
1759984c4197SStefano Zampini   PetscScalar       one=1.0,zero=0.0;
1760984c4197SStefano Zampini   /* space to store constraints and their local indices */
1761984c4197SStefano Zampini   PetscScalar       *temp_quadrature_constraint;
1762984c4197SStefano Zampini   PetscInt          *temp_indices,*temp_indices_to_constraint,*temp_indices_to_constraint_B;
1763984c4197SStefano Zampini   /* iterators */
1764984c4197SStefano Zampini   PetscInt          i,j,k,total_counts,temp_start_ptr;
1765984c4197SStefano Zampini   /* BLAS integers */
1766e310c8b4SStefano Zampini   PetscBLASInt      lwork,lierr;
1767e310c8b4SStefano Zampini   PetscBLASInt      Blas_N,Blas_M,Blas_K,Blas_one=1;
1768c4303822SStefano Zampini   PetscBLASInt      Blas_LDA,Blas_LDB,Blas_LDC;
1769727cdba6SStefano Zampini   /* reuse */
1770727cdba6SStefano Zampini   PetscInt          olocal_primal_size;
1771727cdba6SStefano Zampini   PetscInt          *oprimal_indices_local_idxs;
1772984c4197SStefano Zampini   /* change of basis */
1773984c4197SStefano Zampini   PetscInt          *aux_primal_numbering,*aux_primal_minloc,*global_indices;
1774a717540cSStefano Zampini   PetscBool         boolforchange,qr_needed;
1775a717540cSStefano Zampini   PetscBT           touched,change_basis,qr_needed_idx;
1776984c4197SStefano Zampini   /* auxiliary stuff */
177739e2fb2aSStefano Zampini   PetscInt          *nnz,*is_indices,*aux_primal_numbering_B;
1778a58a30b4SStefano Zampini   PetscInt          ncc,*gidxs=NULL,*permutation=NULL,*temp_indices_to_constraint_work=NULL;
1779a58a30b4SStefano Zampini   PetscScalar       *temp_quadrature_constraint_work=NULL;
1780984c4197SStefano Zampini   /* some quantities */
178145a1bb75SStefano Zampini   PetscInt          n_vertices,total_primal_vertices,valid_constraints;
1782a58a30b4SStefano Zampini   PetscInt          size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints;
1783984c4197SStefano Zampini 
1784674ae819SStefano Zampini   PetscFunctionBegin;
17858e61c736SStefano Zampini   /* Destroy Mat objects computed previously */
17868e61c736SStefano Zampini   ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
17878e61c736SStefano Zampini   ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
1788cf5a6209SStefano Zampini 
1789cf5a6209SStefano Zampini   /* print some info */
1790cf5a6209SStefano Zampini   if (pcbddc->dbg_flag) {
1791cf5a6209SStefano Zampini     IS       vertices;
1792cf5a6209SStefano Zampini     PetscInt nv,nedges,nfaces;
1793cf5a6209SStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr);
1794cf5a6209SStefano Zampini     ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr);
1795cf5a6209SStefano Zampini     ierr = ISDestroy(&vertices);CHKERRQ(ierr);
1796cf5a6209SStefano Zampini     ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
1797cf5a6209SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr);
1798cf5a6209SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr);
1799fd14bc51SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges    (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr);
1800fd14bc51SStefano Zampini     ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces    (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr);
1801cf5a6209SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
1802cf5a6209SStefano Zampini   }
1803cf5a6209SStefano Zampini 
1804cf5a6209SStefano Zampini   if (!pcbddc->adaptive_selection) {
1805cf5a6209SStefano Zampini     IS           ISForVertices,*ISForFaces,*ISForEdges,*used_IS;
1806cf5a6209SStefano Zampini     MatNullSpace nearnullsp;
1807cf5a6209SStefano Zampini     const Vec    *nearnullvecs;
1808cf5a6209SStefano Zampini     Vec          *localnearnullsp;
1809cf5a6209SStefano Zampini     PetscScalar  *array;
1810cf5a6209SStefano Zampini     PetscInt     n_ISForFaces,n_ISForEdges,nnsp_size;
1811cf5a6209SStefano Zampini     PetscBool    nnsp_has_cnst;
1812cf5a6209SStefano Zampini     /* LAPACK working arrays for SVD or POD */
1813cf5a6209SStefano Zampini     PetscBool    skip_lapack;
1814cf5a6209SStefano Zampini     PetscScalar  *work;
1815cf5a6209SStefano Zampini     PetscReal    *singular_vals;
1816cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX)
1817cf5a6209SStefano Zampini     PetscReal    *rwork;
1818cf5a6209SStefano Zampini #endif
1819cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
1820cf5a6209SStefano Zampini     PetscScalar  *temp_basis,*correlation_mat;
1821cf5a6209SStefano Zampini #else
1822964fefecSStefano Zampini     PetscBLASInt dummy_int=1;
1823964fefecSStefano Zampini     PetscScalar  dummy_scalar=1.;
1824cf5a6209SStefano Zampini #endif
1825cf5a6209SStefano Zampini 
1826674ae819SStefano Zampini     /* Get index sets for faces, edges and vertices from graph */
1827d06fc5fdSStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr);
1828d06fc5fdSStefano Zampini     /* free unneeded index sets */
1829d06fc5fdSStefano Zampini     if (!pcbddc->use_vertices) {
1830d06fc5fdSStefano Zampini       ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr);
1831674ae819SStefano Zampini     }
1832d06fc5fdSStefano Zampini     if (!pcbddc->use_edges) {
1833d06fc5fdSStefano Zampini       for (i=0;i<n_ISForEdges;i++) {
1834d06fc5fdSStefano Zampini         ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr);
1835d06fc5fdSStefano Zampini       }
1836d06fc5fdSStefano Zampini       ierr = PetscFree(ISForEdges);CHKERRQ(ierr);
1837d06fc5fdSStefano Zampini       n_ISForEdges = 0;
1838d06fc5fdSStefano Zampini     }
1839d06fc5fdSStefano Zampini     if (!pcbddc->use_faces) {
1840d06fc5fdSStefano Zampini       for (i=0;i<n_ISForFaces;i++) {
1841d06fc5fdSStefano Zampini         ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr);
1842d06fc5fdSStefano Zampini       }
1843d06fc5fdSStefano Zampini       ierr = PetscFree(ISForFaces);CHKERRQ(ierr);
1844d06fc5fdSStefano Zampini       n_ISForFaces = 0;
1845d06fc5fdSStefano Zampini     }
1846d06fc5fdSStefano Zampini     /* HACKS (the following two blocks of code) */
1847b9b85e73SStefano Zampini     if (!ISForVertices && pcbddc->NullSpace && !pcbddc->user_ChangeOfBasisMatrix) {
1848b8ffe317SStefano Zampini       pcbddc->use_change_of_basis = PETSC_TRUE;
1849d06fc5fdSStefano Zampini       if (!ISForEdges) {
1850d06fc5fdSStefano Zampini         pcbddc->use_change_on_faces = PETSC_TRUE;
1851d06fc5fdSStefano Zampini       }
1852b8ffe317SStefano Zampini     }
185398a51de6SStefano Zampini     if (pcbddc->NullSpace) {
185498a51de6SStefano Zampini       /* use_change_of_basis should be consistent among processors */
1855d06fc5fdSStefano Zampini       PetscBool tbool[2],gbool[2];
1856d06fc5fdSStefano Zampini       tbool[0] = pcbddc->use_change_of_basis;
1857d06fc5fdSStefano Zampini       tbool[1] = pcbddc->use_change_on_faces;
1858d06fc5fdSStefano Zampini       ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
1859d06fc5fdSStefano Zampini       pcbddc->use_change_of_basis = gbool[0];
1860d06fc5fdSStefano Zampini       pcbddc->use_change_on_faces = gbool[1];
186198a51de6SStefano Zampini     }
186208122e43SStefano Zampini 
1863674ae819SStefano Zampini     /* check if near null space is attached to global mat */
1864674ae819SStefano Zampini     ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr);
1865674ae819SStefano Zampini     if (nearnullsp) {
1866674ae819SStefano Zampini       ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr);
1867f4ddd8eeSStefano Zampini       /* remove any stored info */
1868f4ddd8eeSStefano Zampini       ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr);
1869f4ddd8eeSStefano Zampini       ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr);
1870f4ddd8eeSStefano Zampini       /* store information for BDDC solver reuse */
1871f4ddd8eeSStefano Zampini       ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr);
1872f4ddd8eeSStefano Zampini       pcbddc->onearnullspace = nearnullsp;
1873473ba861SJed Brown       ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr);
1874f4ddd8eeSStefano Zampini       for (i=0;i<nnsp_size;i++) {
1875f4ddd8eeSStefano Zampini         ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr);
1876f4ddd8eeSStefano Zampini       }
1877984c4197SStefano Zampini     } else { /* if near null space is not provided BDDC uses constants by default */
1878984c4197SStefano Zampini       nnsp_size = 0;
1879674ae819SStefano Zampini       nnsp_has_cnst = PETSC_TRUE;
1880674ae819SStefano Zampini     }
1881984c4197SStefano Zampini     /* get max number of constraints on a single cc */
1882984c4197SStefano Zampini     max_constraints = nnsp_size;
1883984c4197SStefano Zampini     if (nnsp_has_cnst) max_constraints++;
1884984c4197SStefano Zampini 
1885674ae819SStefano Zampini     /*
1886674ae819SStefano Zampini          Evaluate maximum storage size needed by the procedure
1887674ae819SStefano Zampini          - temp_indices will contain start index of each constraint stored as follows
1888911cabfeSStefano Zampini          - temp_indices_to_constraint  [temp_indices[i],...,temp_indices[i+1]-1] will contain the indices (in local numbering) on which the constraint acts
1889911cabfeSStefano Zampini          - temp_indices_to_constraint_B[temp_indices[i],...,temp_indices[i+1]-1] will contain the indices (in boundary numbering) on which the constraint acts
1890911cabfeSStefano Zampini          - temp_quadrature_constraint  [temp_indices[i],...,temp_indices[i+1]-1] will contain the scalars representing the constraint itself
1891674ae819SStefano Zampini                                                                                                                                                            */
1892674ae819SStefano Zampini     total_counts = n_ISForFaces+n_ISForEdges;
1893984c4197SStefano Zampini     total_counts *= max_constraints;
1894674ae819SStefano Zampini     n_vertices = 0;
1895674ae819SStefano Zampini     if (ISForVertices) {
1896674ae819SStefano Zampini       ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr);
1897674ae819SStefano Zampini     }
1898674ae819SStefano Zampini     total_counts += n_vertices;
1899854ce69bSBarry Smith     ierr = PetscMalloc1(total_counts+1,&temp_indices);CHKERRQ(ierr);
19004641a718SStefano Zampini     ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr);
1901674ae819SStefano Zampini     total_counts = 0;
1902674ae819SStefano Zampini     max_size_of_constraint = 0;
1903674ae819SStefano Zampini     for (i=0;i<n_ISForEdges+n_ISForFaces;i++) {
1904674ae819SStefano Zampini       if (i<n_ISForEdges) {
1905674ae819SStefano Zampini         used_IS = &ISForEdges[i];
1906674ae819SStefano Zampini       } else {
1907674ae819SStefano Zampini         used_IS = &ISForFaces[i-n_ISForEdges];
1908674ae819SStefano Zampini       }
1909674ae819SStefano Zampini       ierr = ISGetSize(*used_IS,&j);CHKERRQ(ierr);
1910674ae819SStefano Zampini       total_counts += j;
1911674ae819SStefano Zampini       max_size_of_constraint = PetscMax(j,max_size_of_constraint);
1912674ae819SStefano Zampini     }
1913984c4197SStefano Zampini     total_counts *= max_constraints;
1914674ae819SStefano Zampini     total_counts += n_vertices;
1915906d46d4SStefano Zampini     ierr = PetscMalloc3(total_counts,&temp_quadrature_constraint,total_counts,&temp_indices_to_constraint,total_counts,&temp_indices_to_constraint_B);CHKERRQ(ierr);
1916984c4197SStefano Zampini     /* get local part of global near null space vectors */
1917785e854fSJed Brown     ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr);
1918984c4197SStefano Zampini     for (k=0;k<nnsp_size;k++) {
1919984c4197SStefano Zampini       ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr);
1920984c4197SStefano Zampini       ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1921984c4197SStefano Zampini       ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
1922984c4197SStefano Zampini     }
1923674ae819SStefano Zampini 
1924242a89d7SStefano Zampini     /* whether or not to skip lapack calls */
1925242a89d7SStefano Zampini     skip_lapack = PETSC_TRUE;
1926a773dcb8SStefano Zampini     if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE;
1927242a89d7SStefano Zampini 
19288f1c130eSStefano Zampini     /* allocate some auxiliary stuff */
19298f1c130eSStefano Zampini     if (!skip_lapack || pcbddc->use_qr_single) {
19308f1c130eSStefano Zampini       ierr = PetscMalloc4(max_size_of_constraint,&gidxs,max_size_of_constraint,&permutation,max_size_of_constraint,&temp_indices_to_constraint_work,max_size_of_constraint,&temp_quadrature_constraint_work);CHKERRQ(ierr);
1931f9208b2eSJed Brown     } else {
1932f9208b2eSJed Brown       gidxs = NULL;
1933f9208b2eSJed Brown       permutation = NULL;
1934f9208b2eSJed Brown       temp_indices_to_constraint_work = NULL;
1935f9208b2eSJed Brown       temp_quadrature_constraint_work = NULL;
19368f1c130eSStefano Zampini     }
19378f1c130eSStefano Zampini 
1938984c4197SStefano Zampini     /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */
1939a773dcb8SStefano Zampini     if (!skip_lapack) {
1940674ae819SStefano Zampini       PetscScalar temp_work;
1941911cabfeSStefano Zampini 
1942674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
1943984c4197SStefano Zampini       /* Proper Orthogonal Decomposition (POD) using the snapshot method */
1944785e854fSJed Brown       ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr);
1945785e854fSJed Brown       ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr);
1946785e854fSJed Brown       ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr);
1947674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX)
1948785e854fSJed Brown       ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr);
1949674ae819SStefano Zampini #endif
1950674ae819SStefano Zampini       /* now we evaluate the optimal workspace using query with lwork=-1 */
1951c8244a33SStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr);
1952c8244a33SStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr);
1953674ae819SStefano Zampini       lwork = -1;
1954674ae819SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
1955674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
1956c8244a33SStefano Zampini       PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr));
1957674ae819SStefano Zampini #else
1958c8244a33SStefano Zampini       PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr));
1959674ae819SStefano Zampini #endif
1960674ae819SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
1961984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr);
1962674ae819SStefano Zampini #else /* on missing GESVD */
1963674ae819SStefano Zampini       /* SVD */
1964674ae819SStefano Zampini       PetscInt max_n,min_n;
1965674ae819SStefano Zampini       max_n = max_size_of_constraint;
1966984c4197SStefano Zampini       min_n = max_constraints;
1967984c4197SStefano Zampini       if (max_size_of_constraint < max_constraints) {
1968674ae819SStefano Zampini         min_n = max_size_of_constraint;
1969984c4197SStefano Zampini         max_n = max_constraints;
1970674ae819SStefano Zampini       }
1971785e854fSJed Brown       ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr);
1972674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX)
1973785e854fSJed Brown       ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr);
1974674ae819SStefano Zampini #endif
1975674ae819SStefano Zampini       /* now we evaluate the optimal workspace using query with lwork=-1 */
1976674ae819SStefano Zampini       lwork = -1;
1977e310c8b4SStefano Zampini       ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr);
1978e310c8b4SStefano Zampini       ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr);
1979b7d8b9f8SStefano Zampini       ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr);
1980674ae819SStefano Zampini       ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
1981674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
1982cf5a6209SStefano Zampini       PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr));
1983674ae819SStefano Zampini #else
1984cf5a6209SStefano Zampini       PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr));
1985674ae819SStefano Zampini #endif
1986674ae819SStefano Zampini       ierr = PetscFPTrapPop();CHKERRQ(ierr);
1987984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr);
1988984c4197SStefano Zampini #endif /* on missing GESVD */
1989674ae819SStefano Zampini       /* Allocate optimal workspace */
1990674ae819SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr);
1991854ce69bSBarry Smith       ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr);
1992674ae819SStefano Zampini     }
1993674ae819SStefano Zampini     /* Now we can loop on constraining sets */
1994674ae819SStefano Zampini     total_counts = 0;
1995674ae819SStefano Zampini     temp_indices[0] = 0;
1996674ae819SStefano Zampini     /* vertices */
1997674ae819SStefano Zampini     if (ISForVertices) {
1998674ae819SStefano Zampini       ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
1999674ae819SStefano Zampini       if (nnsp_has_cnst) { /* consider all vertices */
200039e2fb2aSStefano Zampini         ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr);
2001674ae819SStefano Zampini         for (i=0;i<n_vertices;i++) {
2002674ae819SStefano Zampini           temp_quadrature_constraint[temp_indices[total_counts]]=1.0;
2003674ae819SStefano Zampini           temp_indices[total_counts+1]=temp_indices[total_counts]+1;
2004674ae819SStefano Zampini           total_counts++;
2005674ae819SStefano Zampini         }
2006674ae819SStefano Zampini       } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */
2007984c4197SStefano Zampini         PetscBool used_vertex;
2008674ae819SStefano Zampini         for (i=0;i<n_vertices;i++) {
2009674ae819SStefano Zampini           used_vertex = PETSC_FALSE;
2010674ae819SStefano Zampini           k = 0;
2011674ae819SStefano Zampini           while (!used_vertex && k<nnsp_size) {
2012984c4197SStefano Zampini             ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
2013984c4197SStefano Zampini             if (PetscAbsScalar(array[is_indices[i]])>0.0) {
2014674ae819SStefano Zampini               temp_indices_to_constraint[temp_indices[total_counts]]=is_indices[i];
2015674ae819SStefano Zampini               temp_quadrature_constraint[temp_indices[total_counts]]=1.0;
2016674ae819SStefano Zampini               temp_indices[total_counts+1]=temp_indices[total_counts]+1;
2017674ae819SStefano Zampini               total_counts++;
2018674ae819SStefano Zampini               used_vertex = PETSC_TRUE;
2019674ae819SStefano Zampini             }
2020984c4197SStefano Zampini             ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
2021674ae819SStefano Zampini             k++;
2022674ae819SStefano Zampini           }
2023674ae819SStefano Zampini         }
2024674ae819SStefano Zampini       }
2025674ae819SStefano Zampini       ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr);
2026674ae819SStefano Zampini       n_vertices = total_counts;
2027674ae819SStefano Zampini     }
2028984c4197SStefano Zampini 
2029674ae819SStefano Zampini     /* edges and faces */
2030911cabfeSStefano Zampini     for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) {
2031911cabfeSStefano Zampini       if (ncc<n_ISForEdges) {
2032911cabfeSStefano Zampini         used_IS = &ISForEdges[ncc];
2033984c4197SStefano Zampini         boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */
2034674ae819SStefano Zampini       } else {
2035911cabfeSStefano Zampini         used_IS = &ISForFaces[ncc-n_ISForEdges];
2036984c4197SStefano Zampini         boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */
2037674ae819SStefano Zampini       }
2038674ae819SStefano Zampini       temp_constraints = 0;          /* zero the number of constraints I have on this conn comp */
2039674ae819SStefano Zampini       temp_start_ptr = total_counts; /* need to know the starting index of constraints stored */
2040674ae819SStefano Zampini       ierr = ISGetSize(*used_IS,&size_of_constraint);CHKERRQ(ierr);
2041674ae819SStefano Zampini       ierr = ISGetIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr);
2042984c4197SStefano Zampini       /* change of basis should not be performed on local periodic nodes */
2043984c4197SStefano Zampini       if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE;
2044674ae819SStefano Zampini       if (nnsp_has_cnst) {
20455b08dc53SStefano Zampini         PetscScalar quad_value;
2046674ae819SStefano Zampini         temp_constraints++;
2047a773dcb8SStefano Zampini         if (!pcbddc->use_nnsp_true) {
2048674ae819SStefano Zampini           quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint));
2049a773dcb8SStefano Zampini         } else {
2050a773dcb8SStefano Zampini           quad_value = 1.0;
2051a773dcb8SStefano Zampini         }
205239e2fb2aSStefano Zampini         ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
2053674ae819SStefano Zampini         for (j=0;j<size_of_constraint;j++) {
2054674ae819SStefano Zampini           temp_quadrature_constraint[temp_indices[total_counts]+j]=quad_value;
2055674ae819SStefano Zampini         }
205608122e43SStefano Zampini         /* sort by global ordering if using lapack subroutines (not needed!) */
20578f1c130eSStefano Zampini         if (!skip_lapack || pcbddc->use_qr_single) {
2058911cabfeSStefano Zampini           ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr);
2059911cabfeSStefano Zampini           for (j=0;j<size_of_constraint;j++) {
2060911cabfeSStefano Zampini             permutation[j]=j;
2061911cabfeSStefano Zampini           }
2062911cabfeSStefano Zampini           ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr);
2063911cabfeSStefano Zampini           for (j=0;j<size_of_constraint;j++) {
206408122e43SStefano Zampini             if (permutation[j]!=j) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP,"This should not happen");
206508122e43SStefano Zampini           }
206608122e43SStefano Zampini           for (j=0;j<size_of_constraint;j++) {
2067911cabfeSStefano Zampini             temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]];
2068911cabfeSStefano Zampini             temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]];
2069911cabfeSStefano Zampini           }
2070911cabfeSStefano Zampini           ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
2071911cabfeSStefano Zampini           ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr);
2072911cabfeSStefano Zampini         }
2073674ae819SStefano Zampini         temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint;  /* store new starting point */
2074674ae819SStefano Zampini         total_counts++;
2075674ae819SStefano Zampini       }
2076674ae819SStefano Zampini       for (k=0;k<nnsp_size;k++) {
2077984c4197SStefano Zampini         PetscReal real_value;
2078984c4197SStefano Zampini         ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
207939e2fb2aSStefano Zampini         ierr = PetscMemcpy(&temp_indices_to_constraint[temp_indices[total_counts]],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
2080674ae819SStefano Zampini         for (j=0;j<size_of_constraint;j++) {
2081984c4197SStefano Zampini           temp_quadrature_constraint[temp_indices[total_counts]+j]=array[is_indices[j]];
2082674ae819SStefano Zampini         }
2083984c4197SStefano Zampini         ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr);
2084984c4197SStefano Zampini         /* check if array is null on the connected component */
2085e310c8b4SStefano Zampini         ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
2086e310c8b4SStefano Zampini         PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_one));
20875b08dc53SStefano Zampini         if (real_value > 0.0) { /* keep indices and values */
2088911cabfeSStefano Zampini           /* sort by global ordering if using lapack subroutines */
20898f1c130eSStefano Zampini           if (!skip_lapack || pcbddc->use_qr_single) {
2090911cabfeSStefano Zampini             ierr = ISLocalToGlobalMappingApply(matis->mapping,size_of_constraint,temp_indices_to_constraint+temp_indices[total_counts],gidxs);CHKERRQ(ierr);
2091911cabfeSStefano Zampini             for (j=0;j<size_of_constraint;j++) {
2092911cabfeSStefano Zampini               permutation[j]=j;
2093911cabfeSStefano Zampini             }
2094911cabfeSStefano Zampini             ierr = PetscSortIntWithPermutation(size_of_constraint,gidxs,permutation);CHKERRQ(ierr);
2095911cabfeSStefano Zampini             for (j=0;j<size_of_constraint;j++) {
2096911cabfeSStefano Zampini               temp_indices_to_constraint_work[j] = temp_indices_to_constraint[temp_indices[total_counts]+permutation[j]];
2097911cabfeSStefano Zampini               temp_quadrature_constraint_work[j] = temp_quadrature_constraint[temp_indices[total_counts]+permutation[j]];
2098911cabfeSStefano Zampini             }
2099911cabfeSStefano Zampini             ierr = PetscMemcpy(temp_indices_to_constraint+temp_indices[total_counts],temp_indices_to_constraint_work,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr);
2100911cabfeSStefano Zampini             ierr = PetscMemcpy(temp_quadrature_constraint+temp_indices[total_counts],temp_quadrature_constraint_work,size_of_constraint*sizeof(PetscScalar));CHKERRQ(ierr);
2101911cabfeSStefano Zampini           }
2102674ae819SStefano Zampini           temp_constraints++;
2103674ae819SStefano Zampini           temp_indices[total_counts+1]=temp_indices[total_counts]+size_of_constraint;  /* store new starting point */
2104674ae819SStefano Zampini           total_counts++;
2105674ae819SStefano Zampini         }
2106674ae819SStefano Zampini       }
2107674ae819SStefano Zampini       ierr = ISRestoreIndices(*used_IS,(const PetscInt**)&is_indices);CHKERRQ(ierr);
210845a1bb75SStefano Zampini       valid_constraints = temp_constraints;
2109eb97c9d2SStefano Zampini       if (!pcbddc->use_nnsp_true && temp_constraints) {
2110a773dcb8SStefano Zampini         if (temp_constraints == 1) { /* just normalize the constraint */
2111a773dcb8SStefano Zampini           PetscScalar norm;
2112a773dcb8SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
2113a773dcb8SStefano Zampini           PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one));
2114a773dcb8SStefano Zampini           norm = 1.0/PetscSqrtReal(PetscRealPart(norm));
2115a773dcb8SStefano Zampini           PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,temp_quadrature_constraint+temp_indices[temp_start_ptr],&Blas_one));
2116a773dcb8SStefano Zampini         } else { /* perform SVD */
2117984c4197SStefano Zampini           PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */
2118674ae819SStefano Zampini 
2119674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
2120984c4197SStefano Zampini           /* SVD: Y = U*S*V^H                -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag
2121984c4197SStefano Zampini              POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2)
2122984c4197SStefano Zampini              -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined
2123984c4197SStefano Zampini                 the constraints basis will differ (by a complex factor with absolute value equal to 1)
2124984c4197SStefano Zampini                 from that computed using LAPACKgesvd
2125984c4197SStefano Zampini              -> This is due to a different computation of eigenvectors in LAPACKheev
2126984c4197SStefano Zampini              -> The quality of the POD-computed basis will be the same */
2127984c4197SStefano Zampini           ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr);
2128674ae819SStefano Zampini           /* Store upper triangular part of correlation matrix */
2129e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
2130984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
2131674ae819SStefano Zampini           for (j=0;j<temp_constraints;j++) {
2132674ae819SStefano Zampini             for (k=0;k<j+1;k++) {
2133cf5a6209SStefano Zampini               PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k]=BLASdot_(&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr+k]],&Blas_one,&temp_quadrature_constraint[temp_indices[temp_start_ptr+j]],&Blas_one));
2134674ae819SStefano Zampini             }
2135674ae819SStefano Zampini           }
2136e310c8b4SStefano Zampini           /* compute eigenvalues and eigenvectors of correlation matrix */
2137e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
2138e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr);
2139674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
2140c8244a33SStefano Zampini           PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr));
2141674ae819SStefano Zampini #else
2142c8244a33SStefano Zampini           PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr));
2143674ae819SStefano Zampini #endif
2144674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2145984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr);
2146984c4197SStefano Zampini           /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */
2147674ae819SStefano Zampini           j = 0;
2148984c4197SStefano Zampini           while (j < temp_constraints && singular_vals[j] < tol) j++;
2149674ae819SStefano Zampini           total_counts = total_counts-j;
215045a1bb75SStefano Zampini           valid_constraints = temp_constraints-j;
2151e310c8b4SStefano Zampini           /* scale and copy POD basis into used quadrature memory */
2152c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
2153c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
2154c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr);
2155c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2156c4303822SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr);
2157c4303822SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr);
2158674ae819SStefano Zampini           if (j<temp_constraints) {
2159984c4197SStefano Zampini             PetscInt ii;
2160984c4197SStefano Zampini             for (k=j;k<temp_constraints;k++) singular_vals[k]=1.0/PetscSqrtReal(singular_vals[k]);
2161674ae819SStefano Zampini             ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
2162c4303822SStefano Zampini             PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC));
2163674ae819SStefano Zampini             ierr = PetscFPTrapPop();CHKERRQ(ierr);
2164984c4197SStefano Zampini             for (k=0;k<temp_constraints-j;k++) {
2165674ae819SStefano Zampini               for (ii=0;ii<size_of_constraint;ii++) {
2166984c4197SStefano Zampini                 temp_quadrature_constraint[temp_indices[temp_start_ptr+k]+ii]=singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii];
2167674ae819SStefano Zampini               }
2168674ae819SStefano Zampini             }
2169674ae819SStefano Zampini           }
2170674ae819SStefano Zampini #else  /* on missing GESVD */
2171e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
2172e310c8b4SStefano Zampini           ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr);
2173b7d8b9f8SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2174674ae819SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
2175674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX)
2176cf5a6209SStefano Zampini           PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr));
2177674ae819SStefano Zampini #else
2178cf5a6209SStefano Zampini           PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&temp_quadrature_constraint[temp_indices[temp_start_ptr]],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr));
2179674ae819SStefano Zampini #endif
2180984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr);
2181674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2182984c4197SStefano Zampini           /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */
2183e310c8b4SStefano Zampini           k = temp_constraints;
2184e310c8b4SStefano Zampini           if (k > size_of_constraint) k = size_of_constraint;
2185674ae819SStefano Zampini           j = 0;
2186e310c8b4SStefano Zampini           while (j < k && singular_vals[k-j-1] < tol) j++;
218745a1bb75SStefano Zampini           valid_constraints = k-j;
2188911cabfeSStefano Zampini           total_counts = total_counts-temp_constraints+valid_constraints;
2189984c4197SStefano Zampini #endif /* on missing GESVD */
2190674ae819SStefano Zampini         }
2191a773dcb8SStefano Zampini       }
219245a1bb75SStefano Zampini       /* setting change_of_basis flag is safe now */
219345a1bb75SStefano Zampini       if (boolforchange) {
219445a1bb75SStefano Zampini         for (j=0;j<valid_constraints;j++) {
219545a1bb75SStefano Zampini           PetscBTSet(change_basis,total_counts-j-1);
219645a1bb75SStefano Zampini         }
219745a1bb75SStefano Zampini       }
2198674ae819SStefano Zampini     }
2199984c4197SStefano Zampini     /* free workspace */
22008f1c130eSStefano Zampini     if (!skip_lapack || pcbddc->use_qr_single) {
2201911cabfeSStefano Zampini       ierr = PetscFree4(gidxs,permutation,temp_indices_to_constraint_work,temp_quadrature_constraint_work);CHKERRQ(ierr);
22028f1c130eSStefano Zampini     }
22038f1c130eSStefano Zampini     if (!skip_lapack) {
2204984c4197SStefano Zampini       ierr = PetscFree(work);CHKERRQ(ierr);
2205984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX)
2206984c4197SStefano Zampini       ierr = PetscFree(rwork);CHKERRQ(ierr);
2207984c4197SStefano Zampini #endif
2208984c4197SStefano Zampini       ierr = PetscFree(singular_vals);CHKERRQ(ierr);
2209984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD)
2210984c4197SStefano Zampini       ierr = PetscFree(correlation_mat);CHKERRQ(ierr);
2211984c4197SStefano Zampini       ierr = PetscFree(temp_basis);CHKERRQ(ierr);
2212984c4197SStefano Zampini #endif
2213984c4197SStefano Zampini     }
2214984c4197SStefano Zampini     for (k=0;k<nnsp_size;k++) {
2215984c4197SStefano Zampini       ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr);
2216984c4197SStefano Zampini     }
2217984c4197SStefano Zampini     ierr = PetscFree(localnearnullsp);CHKERRQ(ierr);
2218cf5a6209SStefano Zampini     /* free index sets of faces, edges and vertices */
2219cf5a6209SStefano Zampini     for (i=0;i<n_ISForFaces;i++) {
2220cf5a6209SStefano Zampini       ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr);
2221cf5a6209SStefano Zampini     }
2222cf5a6209SStefano Zampini     if (n_ISForFaces) {
2223cf5a6209SStefano Zampini       ierr = PetscFree(ISForFaces);CHKERRQ(ierr);
2224cf5a6209SStefano Zampini     }
2225cf5a6209SStefano Zampini     for (i=0;i<n_ISForEdges;i++) {
2226cf5a6209SStefano Zampini       ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr);
2227cf5a6209SStefano Zampini     }
2228cf5a6209SStefano Zampini     if (n_ISForEdges) {
2229cf5a6209SStefano Zampini       ierr = PetscFree(ISForEdges);CHKERRQ(ierr);
2230cf5a6209SStefano Zampini     }
2231cf5a6209SStefano Zampini     ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr);
223208122e43SStefano Zampini   } else {
223308122e43SStefano Zampini     PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
223408122e43SStefano Zampini     PetscInt        cum = 0;
223508122e43SStefano Zampini 
223608122e43SStefano Zampini     total_counts = 0;
223708122e43SStefano Zampini     n_vertices = 0;
22388bec7fa6SStefano Zampini     if (sub_schurs->is_Ej_com && pcbddc->use_vertices) {
223908122e43SStefano Zampini       ierr = ISGetLocalSize(sub_schurs->is_Ej_com,&n_vertices);CHKERRQ(ierr);
224008122e43SStefano Zampini     }
224108122e43SStefano Zampini     max_constraints = 0;
224208122e43SStefano Zampini     for (i=0;i<sub_schurs->n_subs+n_vertices;i++) {
224308122e43SStefano Zampini       total_counts += pcbddc->adaptive_constraints_n[i];
224408122e43SStefano Zampini       max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]);
224508122e43SStefano Zampini     }
224608122e43SStefano Zampini     temp_indices = pcbddc->adaptive_constraints_ptrs;
224708122e43SStefano Zampini     temp_indices_to_constraint = pcbddc->adaptive_constraints_idxs;
224808122e43SStefano Zampini     temp_quadrature_constraint = pcbddc->adaptive_constraints_data;
224908122e43SStefano Zampini 
225008122e43SStefano Zampini #if 0
225108122e43SStefano Zampini     printf("Found %d totals\n",total_counts);
225208122e43SStefano Zampini     for (i=0;i<total_counts;i++) {
225308122e43SStefano Zampini       printf("const %d, start %d",i,temp_indices[i]);
225408122e43SStefano Zampini       printf(" end %d:\n",temp_indices[i+1]);
225508122e43SStefano Zampini       for (j=temp_indices[i];j<temp_indices[i+1];j++) {
225608122e43SStefano Zampini         printf("  idxs %d",temp_indices_to_constraint[j]);
225708122e43SStefano Zampini         printf("  data %1.2e\n",temp_quadrature_constraint[j]);
225808122e43SStefano Zampini       }
225908122e43SStefano Zampini     }
22601b968477SStefano Zampini     for (i=0;i<n_vertices;i++) {
22618bec7fa6SStefano Zampini       PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]);
22621b968477SStefano Zampini     }
22631b968477SStefano Zampini     for (i=0;i<sub_schurs->n_subs;i++) {
22648bec7fa6SStefano 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]);
22651b968477SStefano Zampini     }
226608122e43SStefano Zampini #endif
226708122e43SStefano Zampini 
22688bec7fa6SStefano Zampini     max_size_of_constraint = 0;
226908122e43SStefano Zampini     for (i=0;i<total_counts;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,temp_indices[i+1]-temp_indices[i]);
227008122e43SStefano Zampini     ierr = PetscMalloc1(temp_indices[total_counts],&temp_indices_to_constraint_B);CHKERRQ(ierr);
227108122e43SStefano Zampini     /* Change of basis */
227208122e43SStefano Zampini     ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr);
227308122e43SStefano Zampini     if (pcbddc->use_change_of_basis) {
2274bbb9e6c6SStefano Zampini       cum = n_vertices;
227508122e43SStefano Zampini       for (i=0;i<sub_schurs->n_subs;i++) {
227608122e43SStefano Zampini         if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) {
2277bbb9e6c6SStefano Zampini           for (j=0;j<pcbddc->adaptive_constraints_n[i+n_vertices];j++) {
227808122e43SStefano Zampini             ierr = PetscBTSet(change_basis,cum+j);CHKERRQ(ierr);
227908122e43SStefano Zampini           }
228008122e43SStefano Zampini         }
2281bbb9e6c6SStefano Zampini         cum += pcbddc->adaptive_constraints_n[i+n_vertices];
228208122e43SStefano Zampini       }
228308122e43SStefano Zampini     }
228408122e43SStefano Zampini   }
228508122e43SStefano Zampini 
228608122e43SStefano Zampini   /* map temp_indices_to_constraint in boundary numbering */
228708122e43SStefano Zampini   ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,temp_indices[total_counts],temp_indices_to_constraint,&i,temp_indices_to_constraint_B);CHKERRQ(ierr);
228808122e43SStefano Zampini   if (i != temp_indices[total_counts]) {
228908122e43SStefano Zampini     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",temp_indices[total_counts],i);
229008122e43SStefano Zampini   }
2291984c4197SStefano Zampini 
2292727cdba6SStefano Zampini   /* set quantities in pcbddc data structure and store previous primal size */
2293984c4197SStefano Zampini   /* n_vertices defines the number of subdomain corners in the primal space */
2294674ae819SStefano Zampini   /* n_constraints defines the number of averages (they can be point primal dofs if change of basis is requested) */
2295727cdba6SStefano Zampini   olocal_primal_size = pcbddc->local_primal_size;
2296984c4197SStefano Zampini   pcbddc->local_primal_size = total_counts;
2297674ae819SStefano Zampini   pcbddc->n_vertices = n_vertices;
2298984c4197SStefano Zampini   pcbddc->n_constraints = pcbddc->local_primal_size-pcbddc->n_vertices;
2299674ae819SStefano Zampini 
2300674ae819SStefano Zampini   /* Create constraint matrix */
2301674ae819SStefano Zampini   /* The constraint matrix is used to compute the l2g map of primal dofs */
2302674ae819SStefano Zampini   /* so we need to set it up properly either with or without change of basis */
2303674ae819SStefano Zampini   ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr);
230416f15bc4SStefano Zampini   ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr);
2305984c4197SStefano Zampini   ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr);
2306984c4197SStefano Zampini   /* array to compute a local numbering of constraints : vertices first then constraints */
2307785e854fSJed Brown   ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_numbering);CHKERRQ(ierr);
2308984c4197SStefano Zampini   /* array to select the proper local node (of minimum index with respect to global ordering) when changing the basis */
2309984c4197SStefano Zampini   /* note: it should not be needed since IS for faces and edges are already sorted by global ordering when analyzing the graph but... just in case */
2310785e854fSJed Brown   ierr = PetscMalloc1(pcbddc->local_primal_size,&aux_primal_minloc);CHKERRQ(ierr);
2311984c4197SStefano Zampini   /* auxiliary stuff for basis change */
2312785e854fSJed Brown   ierr = PetscMalloc1(max_size_of_constraint,&global_indices);CHKERRQ(ierr);
23134641a718SStefano Zampini   ierr = PetscBTCreate(pcis->n_B,&touched);CHKERRQ(ierr);
2314984c4197SStefano Zampini 
2315984c4197SStefano Zampini   /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */
2316984c4197SStefano Zampini   total_primal_vertices=0;
2317984c4197SStefano Zampini   for (i=0;i<pcbddc->local_primal_size;i++) {
2318674ae819SStefano Zampini     size_of_constraint=temp_indices[i+1]-temp_indices[i];
2319984c4197SStefano Zampini     if (size_of_constraint == 1) {
23204641a718SStefano Zampini       ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]]);CHKERRQ(ierr);
2321984c4197SStefano Zampini       aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]];
2322984c4197SStefano Zampini       aux_primal_minloc[total_primal_vertices]=0;
2323984c4197SStefano Zampini       total_primal_vertices++;
23244641a718SStefano Zampini     } else if (PetscBTLookup(change_basis,i)) { /* Same procedure used in PCBDDCGetPrimalConstraintsLocalIdx */
2325984c4197SStefano Zampini       PetscInt min_loc,min_index;
2326984c4197SStefano Zampini       ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],global_indices);CHKERRQ(ierr);
2327984c4197SStefano Zampini       /* find first untouched local node */
2328674ae819SStefano Zampini       k = 0;
23294641a718SStefano Zampini       while (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) k++;
2330984c4197SStefano Zampini       min_index = global_indices[k];
2331984c4197SStefano Zampini       min_loc = k;
2332984c4197SStefano Zampini       /* search the minimum among global nodes already untouched on the cc */
2333984c4197SStefano Zampini       for (k=1;k<size_of_constraint;k++) {
2334984c4197SStefano Zampini         /* there can be more than one constraint on a single connected component */
23354641a718SStefano Zampini         if (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k]) && min_index > global_indices[k]) {
2336984c4197SStefano Zampini           min_index = global_indices[k];
2337984c4197SStefano Zampini           min_loc = k;
2338674ae819SStefano Zampini         }
2339674ae819SStefano Zampini       }
23404641a718SStefano Zampini       ierr = PetscBTSet(touched,temp_indices_to_constraint_B[temp_indices[i]+min_loc]);CHKERRQ(ierr);
2341984c4197SStefano Zampini       aux_primal_numbering[total_primal_vertices]=temp_indices_to_constraint[temp_indices[i]+min_loc];
2342984c4197SStefano Zampini       aux_primal_minloc[total_primal_vertices]=min_loc;
2343984c4197SStefano Zampini       total_primal_vertices++;
2344984c4197SStefano Zampini     }
2345984c4197SStefano Zampini   }
2346a717540cSStefano Zampini   /* determine if a QR strategy is needed for change of basis */
2347a717540cSStefano Zampini   qr_needed = PETSC_FALSE;
2348a717540cSStefano Zampini   ierr = PetscBTCreate(pcbddc->local_primal_size,&qr_needed_idx);CHKERRQ(ierr);
2349a717540cSStefano Zampini   for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) {
2350a717540cSStefano Zampini     if (PetscBTLookup(change_basis,i)) {
2351ac632422SStefano Zampini       if (!pcbddc->use_qr_single && !pcbddc->faster_deluxe) {
2352a717540cSStefano Zampini         size_of_constraint = temp_indices[i+1]-temp_indices[i];
2353a717540cSStefano Zampini         j = 0;
2354a717540cSStefano Zampini         for (k=0;k<size_of_constraint;k++) {
2355a717540cSStefano Zampini           if (PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+k])) {
2356a717540cSStefano Zampini             j++;
2357a717540cSStefano Zampini           }
2358a717540cSStefano Zampini         }
2359a717540cSStefano Zampini         /* found more than one primal dof on the cc */
2360a717540cSStefano Zampini         if (j > 1) {
2361a717540cSStefano Zampini           PetscBTSet(qr_needed_idx,i);
2362a717540cSStefano Zampini           qr_needed = PETSC_TRUE;
2363a717540cSStefano Zampini         }
2364fa434743SStefano Zampini       } else {
2365fa434743SStefano Zampini         PetscBTSet(qr_needed_idx,i);
2366fa434743SStefano Zampini         qr_needed = PETSC_TRUE;
2367fa434743SStefano Zampini       }
2368a717540cSStefano Zampini     }
2369a717540cSStefano Zampini   }
2370984c4197SStefano Zampini   /* free workspace */
2371984c4197SStefano Zampini   ierr = PetscFree(global_indices);CHKERRQ(ierr);
237245a1bb75SStefano Zampini 
2373674ae819SStefano Zampini   /* permute indices in order to have a sorted set of vertices */
2374a717540cSStefano Zampini   ierr = PetscSortInt(total_primal_vertices,aux_primal_numbering);CHKERRQ(ierr);
2375984c4197SStefano Zampini 
2376984c4197SStefano Zampini   /* nonzero structure of constraint matrix */
2377785e854fSJed Brown   ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr);
2378984c4197SStefano Zampini   for (i=0;i<total_primal_vertices;i++) nnz[i]=1;
2379984c4197SStefano Zampini   j=total_primal_vertices;
2380984c4197SStefano Zampini   for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) {
23814641a718SStefano Zampini     if (!PetscBTLookup(change_basis,i)) {
2382674ae819SStefano Zampini       nnz[j]=temp_indices[i+1]-temp_indices[i];
2383674ae819SStefano Zampini       j++;
2384674ae819SStefano Zampini     }
2385674ae819SStefano Zampini   }
2386674ae819SStefano Zampini   ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr);
2387674ae819SStefano Zampini   ierr = PetscFree(nnz);CHKERRQ(ierr);
2388674ae819SStefano Zampini   /* set values in constraint matrix */
2389984c4197SStefano Zampini   for (i=0;i<total_primal_vertices;i++) {
2390984c4197SStefano Zampini     ierr = MatSetValue(pcbddc->ConstraintMatrix,i,aux_primal_numbering[i],1.0,INSERT_VALUES);CHKERRQ(ierr);
2391674ae819SStefano Zampini   }
2392984c4197SStefano Zampini   total_counts = total_primal_vertices;
2393984c4197SStefano Zampini   for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) {
23944641a718SStefano Zampini     if (!PetscBTLookup(change_basis,i)) {
2395674ae819SStefano Zampini       size_of_constraint=temp_indices[i+1]-temp_indices[i];
2396674ae819SStefano Zampini       ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&total_counts,size_of_constraint,&temp_indices_to_constraint[temp_indices[i]],&temp_quadrature_constraint[temp_indices[i]],INSERT_VALUES);CHKERRQ(ierr);
2397674ae819SStefano Zampini       total_counts++;
2398674ae819SStefano Zampini     }
2399674ae819SStefano Zampini   }
2400674ae819SStefano Zampini   /* assembling */
2401674ae819SStefano Zampini   ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2402674ae819SStefano Zampini   ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2403984c4197SStefano Zampini   /*
240445a1bb75SStefano Zampini   ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
2405984c4197SStefano Zampini   ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr);
2406984c4197SStefano Zampini   */
2407674ae819SStefano Zampini   /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */
2408674ae819SStefano Zampini   if (pcbddc->use_change_of_basis) {
2409026de310SStefano Zampini     /* dual and primal dofs on a single cc */
2410984c4197SStefano Zampini     PetscInt     dual_dofs,primal_dofs;
2411026de310SStefano Zampini     /* iterator on aux_primal_minloc (ordered as read from nearnullspace: vertices, edges and then constraints) */
2412026de310SStefano Zampini     PetscInt     primal_counter;
2413984c4197SStefano Zampini     /* working stuff for GEQRF */
241481d9aea3SBarry Smith     PetscScalar  *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t;
2415984c4197SStefano Zampini     PetscBLASInt lqr_work;
2416984c4197SStefano Zampini     /* working stuff for UNGQR */
2417984c4197SStefano Zampini     PetscScalar  *gqr_work,lgqr_work_t;
2418984c4197SStefano Zampini     PetscBLASInt lgqr_work;
2419984c4197SStefano Zampini     /* working stuff for TRTRS */
2420984c4197SStefano Zampini     PetscScalar  *trs_rhs;
24213f08241aSStefano Zampini     PetscBLASInt Blas_NRHS;
2422984c4197SStefano Zampini     /* pointers for values insertion into change of basis matrix */
2423984c4197SStefano Zampini     PetscInt     *start_rows,*start_cols;
2424984c4197SStefano Zampini     PetscScalar  *start_vals;
2425984c4197SStefano Zampini     /* working stuff for values insertion */
24264641a718SStefano Zampini     PetscBT      is_primal;
2427906d46d4SStefano Zampini     /* matrix sizes */
2428906d46d4SStefano Zampini     PetscInt     global_size,local_size;
2429906d46d4SStefano Zampini     /* temporary change of basis */
2430906d46d4SStefano Zampini     Mat          localChangeOfBasisMatrix;
2431cf5a6209SStefano Zampini     /* extra space for debugging */
2432cf5a6209SStefano Zampini     PetscScalar  *dbg_work;
2433984c4197SStefano Zampini 
2434906d46d4SStefano Zampini     /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */
2435906d46d4SStefano Zampini     ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr);
243616f15bc4SStefano Zampini     ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr);
2437bbb9e6c6SStefano Zampini     ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr);
2438906d46d4SStefano Zampini     /* nonzeros for local mat */
2439bbb9e6c6SStefano Zampini     ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr);
2440bbb9e6c6SStefano Zampini     for (i=0;i<pcis->n;i++) nnz[i]=1;
2441a717540cSStefano Zampini     for (i=pcbddc->n_vertices;i<pcbddc->local_primal_size;i++) {
2442a717540cSStefano Zampini       if (PetscBTLookup(change_basis,i)) {
2443a717540cSStefano Zampini         size_of_constraint = temp_indices[i+1]-temp_indices[i];
2444a717540cSStefano Zampini         if (PetscBTLookup(qr_needed_idx,i)) {
2445bbb9e6c6SStefano Zampini           for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint[temp_indices[i]+j]] = size_of_constraint;
2446a717540cSStefano Zampini         } else {
2447bbb9e6c6SStefano Zampini           for (j=0;j<size_of_constraint;j++) nnz[temp_indices_to_constraint[temp_indices[i]+j]] = 2;
2448a717540cSStefano Zampini           /* get local primal index on the cc */
2449a717540cSStefano Zampini           j = 0;
2450a717540cSStefano Zampini           while (!PetscBTLookup(touched,temp_indices_to_constraint_B[temp_indices[i]+j])) j++;
2451bbb9e6c6SStefano Zampini           nnz[temp_indices_to_constraint[temp_indices[i]+j]] = size_of_constraint;
2452a717540cSStefano Zampini         }
2453a717540cSStefano Zampini       }
2454a717540cSStefano Zampini     }
2455906d46d4SStefano Zampini     ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr);
2456bbb9e6c6SStefano Zampini     ierr = PetscFree(nnz);CHKERRQ(ierr);
2457a717540cSStefano Zampini     /* Set initial identity in the matrix */
2458bbb9e6c6SStefano Zampini     for (i=0;i<pcis->n;i++) {
2459906d46d4SStefano Zampini       ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr);
2460a717540cSStefano Zampini     }
2461a717540cSStefano Zampini 
2462a717540cSStefano Zampini     if (pcbddc->dbg_flag) {
2463a717540cSStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr);
2464a717540cSStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr);
2465a717540cSStefano Zampini     }
2466a717540cSStefano Zampini 
2467a717540cSStefano Zampini 
2468a717540cSStefano Zampini     /* Now we loop on the constraints which need a change of basis */
2469a717540cSStefano Zampini     /*
2470a717540cSStefano Zampini        Change of basis matrix is evaluated similarly to the FIRST APPROACH in
2471a717540cSStefano Zampini        Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1)
2472a717540cSStefano Zampini 
2473a6b551f4SStefano Zampini        Basic blocks of change of basis matrix T computed by
2474a717540cSStefano Zampini 
2475a6b551f4SStefano 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)
2476a6b551f4SStefano Zampini 
2477a6b551f4SStefano Zampini             | 1        0   ...        0         s_1/S |
2478a6b551f4SStefano Zampini             | 0        1   ...        0         s_2/S |
2479a717540cSStefano Zampini             |              ...                        |
2480a6b551f4SStefano Zampini             | 0        ...            1     s_{n-1}/S |
2481a6b551f4SStefano Zampini             | -s_1/s_n ...    -s_{n-1}/s_n      s_n/S |
2482a717540cSStefano Zampini 
2483a6b551f4SStefano Zampini             with S = \sum_{i=1}^n s_i^2
2484a6b551f4SStefano Zampini             NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering
2485a6b551f4SStefano Zampini                   in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering
2486a6b551f4SStefano Zampini 
2487a6b551f4SStefano Zampini           - QR decomposition of constraints otherwise
2488a717540cSStefano Zampini     */
2489a717540cSStefano Zampini     if (qr_needed) {
2490984c4197SStefano Zampini       /* space to store Q */
2491854ce69bSBarry Smith       ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr);
2492984c4197SStefano Zampini       /* first we issue queries for optimal work */
24933f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr);
24943f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr);
24953f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2496984c4197SStefano Zampini       lqr_work = -1;
24973f08241aSStefano Zampini       PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr));
2498984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr);
2499984c4197SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr);
2500785e854fSJed Brown       ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr);
2501984c4197SStefano Zampini       lgqr_work = -1;
25023f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr);
25033f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr);
25043f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr);
25053f08241aSStefano Zampini       ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
25063f08241aSStefano Zampini       if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */
25073f08241aSStefano Zampini       PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr));
2508984c4197SStefano Zampini       if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr);
2509984c4197SStefano Zampini       ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr);
2510785e854fSJed Brown       ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr);
2511984c4197SStefano Zampini       /* array to store scaling factors for reflectors */
2512785e854fSJed Brown       ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr);
2513984c4197SStefano Zampini       /* array to store rhs and solution of triangular solver */
2514785e854fSJed Brown       ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr);
2515a717540cSStefano Zampini       /* allocating workspace for check */
2516a717540cSStefano Zampini       if (pcbddc->dbg_flag) {
2517cf5a6209SStefano Zampini         ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr);
2518a717540cSStefano Zampini       }
2519a717540cSStefano Zampini     }
2520984c4197SStefano Zampini     /* array to store whether a node is primal or not */
25214641a718SStefano Zampini     ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr);
2522473ba861SJed Brown     ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr);
25235e8657edSStefano Zampini     ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,aux_primal_numbering,&i,aux_primal_numbering_B);CHKERRQ(ierr);
252439e2fb2aSStefano Zampini     if (i != total_primal_vertices) {
252539e2fb2aSStefano Zampini       SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i);
25264641a718SStefano Zampini     }
252739e2fb2aSStefano Zampini     for (i=0;i<total_primal_vertices;i++) {
252839e2fb2aSStefano Zampini       ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr);
252939e2fb2aSStefano Zampini     }
253039e2fb2aSStefano Zampini     ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr);
2531984c4197SStefano Zampini 
2532a717540cSStefano Zampini     /* loop on constraints and see whether or not they need a change of basis and compute it */
2533026de310SStefano Zampini     /* -> using implicit ordering contained in temp_indices data */
2534026de310SStefano Zampini     total_counts = pcbddc->n_vertices;
2535026de310SStefano Zampini     primal_counter = total_counts;
2536026de310SStefano Zampini     while (total_counts<pcbddc->local_primal_size) {
2537026de310SStefano Zampini       primal_dofs = 1;
25384641a718SStefano Zampini       if (PetscBTLookup(change_basis,total_counts)) {
2539026de310SStefano Zampini         /* get all constraints with same support: if more then one constraint is present on the cc then surely indices are stored contiguosly */
2540bbb9e6c6SStefano Zampini         while (total_counts+primal_dofs < pcbddc->local_primal_size && temp_indices_to_constraint[temp_indices[total_counts]] == temp_indices_to_constraint[temp_indices[total_counts+primal_dofs]]) {
2541026de310SStefano Zampini           primal_dofs++;
2542674ae819SStefano Zampini         }
2543984c4197SStefano Zampini         /* get constraint info */
2544026de310SStefano Zampini         size_of_constraint = temp_indices[total_counts+1]-temp_indices[total_counts];
2545984c4197SStefano Zampini         dual_dofs = size_of_constraint-primal_dofs;
2546984c4197SStefano Zampini 
2547984c4197SStefano Zampini         if (pcbddc->dbg_flag) {
2548a717540cSStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d to %d (incl) need a change of basis (size %d)\n",total_counts,total_counts+primal_dofs-1,size_of_constraint);CHKERRQ(ierr);
2549674ae819SStefano Zampini         }
2550984c4197SStefano Zampini 
2551fa434743SStefano Zampini         if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */
2552a717540cSStefano Zampini 
2553a717540cSStefano Zampini           /* copy quadrature constraints for change of basis check */
2554a717540cSStefano Zampini           if (pcbddc->dbg_flag) {
2555cf5a6209SStefano Zampini             ierr = PetscMemcpy(dbg_work,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
2556a717540cSStefano Zampini           }
2557984c4197SStefano Zampini           /* copy temporary constraints into larger work vector (in order to store all columns of Q) */
2558026de310SStefano Zampini           ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
2559984c4197SStefano Zampini 
2560984c4197SStefano Zampini           /* compute QR decomposition of constraints */
25613f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
25623f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
25633f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2564674ae819SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
25653f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr));
2566984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr);
2567674ae819SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2568984c4197SStefano Zampini 
2569984c4197SStefano Zampini           /* explictly compute R^-T */
2570984c4197SStefano Zampini           ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr);
2571984c4197SStefano Zampini           for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0;
25723f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
25733f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr);
25743f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
25753f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr);
2576984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
25773f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr));
2578984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr);
2579984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2580984c4197SStefano Zampini 
2581a717540cSStefano Zampini           /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */
25823f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
25833f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
25843f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr);
25853f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2586984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
25873f08241aSStefano Zampini           PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr));
2588984c4197SStefano Zampini           if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr);
2589984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2590984c4197SStefano Zampini 
2591984c4197SStefano Zampini           /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints
2592984c4197SStefano Zampini              i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below)
2593984c4197SStefano Zampini              where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */
25943f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr);
25953f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr);
25963f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr);
25973f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
25983f08241aSStefano Zampini           ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr);
25993f08241aSStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr);
2600984c4197SStefano Zampini           ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
2601c4303822SStefano Zampini           PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,&temp_quadrature_constraint[temp_indices[total_counts]],&Blas_LDC));
2602984c4197SStefano Zampini           ierr = PetscFPTrapPop();CHKERRQ(ierr);
2603026de310SStefano Zampini           ierr = PetscMemcpy(qr_basis,&temp_quadrature_constraint[temp_indices[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr);
2604984c4197SStefano Zampini 
2605984c4197SStefano Zampini           /* insert values in change of basis matrix respecting global ordering of new primal dofs */
2606bbb9e6c6SStefano Zampini           start_rows = &temp_indices_to_constraint[temp_indices[total_counts]];
2607984c4197SStefano Zampini           /* insert cols for primal dofs */
2608984c4197SStefano Zampini           for (j=0;j<primal_dofs;j++) {
2609984c4197SStefano Zampini             start_vals = &qr_basis[j*size_of_constraint];
2610bbb9e6c6SStefano Zampini             start_cols = &temp_indices_to_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter+j]];
2611906d46d4SStefano Zampini             ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr);
2612984c4197SStefano Zampini           }
2613984c4197SStefano Zampini           /* insert cols for dual dofs */
2614984c4197SStefano Zampini           for (j=0,k=0;j<dual_dofs;k++) {
26154641a718SStefano Zampini             if (!PetscBTLookup(is_primal,temp_indices_to_constraint_B[temp_indices[total_counts]+k])) {
2616984c4197SStefano Zampini               start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint];
2617bbb9e6c6SStefano Zampini               start_cols = &temp_indices_to_constraint[temp_indices[total_counts]+k];
2618906d46d4SStefano Zampini               ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr);
2619984c4197SStefano Zampini               j++;
2620674ae819SStefano Zampini             }
2621674ae819SStefano Zampini           }
2622984c4197SStefano Zampini 
2623984c4197SStefano Zampini           /* check change of basis */
2624984c4197SStefano Zampini           if (pcbddc->dbg_flag) {
2625984c4197SStefano Zampini             PetscInt   ii,jj;
2626984c4197SStefano Zampini             PetscBool valid_qr=PETSC_TRUE;
2627c4303822SStefano Zampini             ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr);
2628c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
2629c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr);
2630c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr);
2631c4303822SStefano Zampini             ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr);
2632c4303822SStefano Zampini             ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr);
2633984c4197SStefano Zampini             ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
2634cf5a6209SStefano 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));
2635984c4197SStefano Zampini             ierr = PetscFPTrapPop();CHKERRQ(ierr);
2636984c4197SStefano Zampini             for (jj=0;jj<size_of_constraint;jj++) {
2637984c4197SStefano Zampini               for (ii=0;ii<primal_dofs;ii++) {
2638cf5a6209SStefano Zampini                 if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE;
2639cf5a6209SStefano 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;
2640674ae819SStefano Zampini               }
2641674ae819SStefano Zampini             }
2642984c4197SStefano Zampini             if (!valid_qr) {
264322d5777bSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr);
2644984c4197SStefano Zampini               for (jj=0;jj<size_of_constraint;jj++) {
2645984c4197SStefano Zampini                 for (ii=0;ii<primal_dofs;ii++) {
2646cf5a6209SStefano Zampini                   if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) {
2647cf5a6209SStefano 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]));
2648674ae819SStefano Zampini                   }
2649cf5a6209SStefano Zampini                   if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) {
2650cf5a6209SStefano 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]));
2651984c4197SStefano Zampini                   }
2652984c4197SStefano Zampini                 }
2653984c4197SStefano Zampini               }
2654674ae819SStefano Zampini             } else {
265522d5777bSStefano Zampini               ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr);
2656674ae819SStefano Zampini             }
2657674ae819SStefano Zampini           }
2658a717540cSStefano Zampini         } else { /* simple transformation block */
2659a717540cSStefano Zampini           PetscInt    row,col;
2660a6b551f4SStefano Zampini           PetscScalar val,norm;
2661a6b551f4SStefano Zampini 
2662a6b551f4SStefano Zampini           ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr);
2663a6b551f4SStefano Zampini           PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one,temp_quadrature_constraint+temp_indices[total_counts],&Blas_one));
2664a717540cSStefano Zampini           for (j=0;j<size_of_constraint;j++) {
2665bbb9e6c6SStefano Zampini             PetscInt row_B = temp_indices_to_constraint_B[temp_indices[total_counts]+j];
2666bbb9e6c6SStefano Zampini             row = temp_indices_to_constraint[temp_indices[total_counts]+j];
2667bbb9e6c6SStefano Zampini             if (!PetscBTLookup(is_primal,row_B)) {
2668bbb9e6c6SStefano Zampini               col = temp_indices_to_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]];
2669906d46d4SStefano Zampini               ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr);
2670a6b551f4SStefano Zampini               ierr = MatSetValue(localChangeOfBasisMatrix,row,col,temp_quadrature_constraint[temp_indices[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr);
2671a717540cSStefano Zampini             } else {
2672a717540cSStefano Zampini               for (k=0;k<size_of_constraint;k++) {
2673bbb9e6c6SStefano Zampini                 col = temp_indices_to_constraint[temp_indices[total_counts]+k];
2674a717540cSStefano Zampini                 if (row != col) {
2675a717540cSStefano Zampini                   val = -temp_quadrature_constraint[temp_indices[total_counts]+k]/temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]];
2676a717540cSStefano Zampini                 } else {
2677a6b551f4SStefano Zampini                   val = temp_quadrature_constraint[temp_indices[total_counts]+aux_primal_minloc[primal_counter]]/norm;
2678a717540cSStefano Zampini                 }
2679906d46d4SStefano Zampini                 ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr);
2680a717540cSStefano Zampini               }
2681a717540cSStefano Zampini             }
2682a717540cSStefano Zampini           }
268398a51de6SStefano Zampini           if (pcbddc->dbg_flag) {
268422d5777bSStefano Zampini             ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr);
2685a717540cSStefano Zampini           }
2686674ae819SStefano Zampini         }
2687026de310SStefano Zampini         /* increment primal counter */
2688026de310SStefano Zampini         primal_counter += primal_dofs;
2689984c4197SStefano Zampini       } else {
2690984c4197SStefano Zampini         if (pcbddc->dbg_flag) {
2691026de310SStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,temp_indices[total_counts+1]-temp_indices[total_counts]);CHKERRQ(ierr);
2692674ae819SStefano Zampini         }
2693674ae819SStefano Zampini       }
2694026de310SStefano Zampini       /* increment constraint counter total_counts */
2695026de310SStefano Zampini       total_counts += primal_dofs;
2696674ae819SStefano Zampini     }
2697a717540cSStefano Zampini 
2698a717540cSStefano Zampini     /* free workspace */
2699a717540cSStefano Zampini     if (qr_needed) {
2700984c4197SStefano Zampini       if (pcbddc->dbg_flag) {
2701cf5a6209SStefano Zampini         ierr = PetscFree(dbg_work);CHKERRQ(ierr);
2702984c4197SStefano Zampini       }
2703984c4197SStefano Zampini       ierr = PetscFree(trs_rhs);CHKERRQ(ierr);
2704984c4197SStefano Zampini       ierr = PetscFree(qr_tau);CHKERRQ(ierr);
2705984c4197SStefano Zampini       ierr = PetscFree(qr_work);CHKERRQ(ierr);
2706984c4197SStefano Zampini       ierr = PetscFree(gqr_work);CHKERRQ(ierr);
2707984c4197SStefano Zampini       ierr = PetscFree(qr_basis);CHKERRQ(ierr);
2708674ae819SStefano Zampini     }
2709a717540cSStefano Zampini     ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr);
2710906d46d4SStefano Zampini     ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2711906d46d4SStefano Zampini     ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2712906d46d4SStefano Zampini 
2713906d46d4SStefano Zampini     /* assembling of global change of variable */
2714bbb9e6c6SStefano Zampini     {
2715bbb9e6c6SStefano Zampini       Mat      tmat;
271616f15bc4SStefano Zampini       PetscInt bs;
271716f15bc4SStefano Zampini 
2718906d46d4SStefano Zampini       ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr);
2719906d46d4SStefano Zampini       ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr);
2720bbb9e6c6SStefano Zampini       ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr);
2721bbb9e6c6SStefano Zampini       ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr);
2722bbb9e6c6SStefano Zampini       ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
2723bbb9e6c6SStefano Zampini       ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr);
272416f15bc4SStefano Zampini       ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr);
272516f15bc4SStefano Zampini       ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr);
2726906d46d4SStefano Zampini       ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr);
2727bbb9e6c6SStefano Zampini       ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr);
2728bbb9e6c6SStefano Zampini       ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
2729bbb9e6c6SStefano Zampini       ierr = MatDestroy(&tmat);CHKERRQ(ierr);
2730bbb9e6c6SStefano Zampini       ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
2731bbb9e6c6SStefano Zampini       ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr);
2732bbb9e6c6SStefano Zampini       ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2733bbb9e6c6SStefano Zampini       ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2734bbb9e6c6SStefano Zampini       ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr);
2735bbb9e6c6SStefano Zampini       ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr);
2736906d46d4SStefano Zampini     }
2737906d46d4SStefano Zampini     /* check */
2738906d46d4SStefano Zampini     if (pcbddc->dbg_flag) {
2739906d46d4SStefano Zampini       PetscReal error;
2740906d46d4SStefano Zampini       Vec       x,x_change;
2741906d46d4SStefano Zampini 
2742906d46d4SStefano Zampini       ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr);
2743906d46d4SStefano Zampini       ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr);
2744906d46d4SStefano Zampini       ierr = VecSetRandom(x,NULL);CHKERRQ(ierr);
2745906d46d4SStefano Zampini       ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr);
2746bbb9e6c6SStefano Zampini       ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2747bbb9e6c6SStefano Zampini       ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
2748bbb9e6c6SStefano Zampini       ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr);
2749bbb9e6c6SStefano Zampini       ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2750bbb9e6c6SStefano Zampini       ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
2751906d46d4SStefano Zampini       ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr);
2752906d46d4SStefano Zampini       ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr);
2753906d46d4SStefano Zampini       ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr);
2754906d46d4SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2755bbb9e6c6SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr);
2756906d46d4SStefano Zampini       ierr = VecDestroy(&x);CHKERRQ(ierr);
2757906d46d4SStefano Zampini       ierr = VecDestroy(&x_change);CHKERRQ(ierr);
2758906d46d4SStefano Zampini     }
2759b96c3477SStefano Zampini 
2760b96c3477SStefano Zampini     /* adapt sub_schurs computed (if any) */
2761b96c3477SStefano Zampini     if (pcbddc->use_deluxe_scaling) {
2762b96c3477SStefano Zampini       PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs;
2763b96c3477SStefano Zampini       if (sub_schurs->S_Ej_all) {
2764ac632422SStefano Zampini         Mat S_new,tmat;
2765bbb9e6c6SStefano Zampini         IS is_all_N;
2766bbb9e6c6SStefano Zampini 
2767bbb9e6c6SStefano Zampini         ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr);
2768d648f858SStefano Zampini         ierr = MatGetSubMatrixUnsorted(localChangeOfBasisMatrix,is_all_N,is_all_N,&tmat);CHKERRQ(ierr);
2769bbb9e6c6SStefano Zampini         ierr = ISDestroy(&is_all_N);CHKERRQ(ierr);
2770ac632422SStefano Zampini         ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr);
2771b96c3477SStefano Zampini         ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr);
2772ac632422SStefano Zampini         ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr);
2773ac632422SStefano Zampini         sub_schurs->S_Ej_all = S_new;
2774ac632422SStefano Zampini         ierr = MatDestroy(&S_new);CHKERRQ(ierr);
2775ac632422SStefano Zampini         if (sub_schurs->sum_S_Ej_all) {
2776ac632422SStefano Zampini           ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr);
2777b96c3477SStefano Zampini           ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr);
2778ac632422SStefano Zampini           ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr);
2779ac632422SStefano Zampini           sub_schurs->sum_S_Ej_all = S_new;
2780ac632422SStefano Zampini           ierr = MatDestroy(&S_new);CHKERRQ(ierr);
2781ac632422SStefano Zampini         }
2782b96c3477SStefano Zampini         ierr = MatDestroy(&tmat);CHKERRQ(ierr);
2783b96c3477SStefano Zampini       }
2784b96c3477SStefano Zampini     }
2785906d46d4SStefano Zampini     ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr);
2786906d46d4SStefano Zampini   } else if (pcbddc->user_ChangeOfBasisMatrix) {
2787b9b85e73SStefano Zampini     ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
2788b9b85e73SStefano Zampini     pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix;
2789b9b85e73SStefano Zampini   }
2790906d46d4SStefano Zampini 
2791906d46d4SStefano Zampini   /* set up change of basis context */
2792906d46d4SStefano Zampini   if (pcbddc->ChangeOfBasisMatrix) {
2793906d46d4SStefano Zampini     PCBDDCChange_ctx change_ctx;
2794906d46d4SStefano Zampini 
2795906d46d4SStefano Zampini     if (!pcbddc->new_global_mat) {
2796906d46d4SStefano Zampini       PetscInt global_size,local_size;
2797906d46d4SStefano Zampini 
2798906d46d4SStefano Zampini       ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr);
2799906d46d4SStefano Zampini       ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr);
2800906d46d4SStefano Zampini       ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr);
2801906d46d4SStefano Zampini       ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr);
2802906d46d4SStefano Zampini       ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr);
2803906d46d4SStefano Zampini       ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr);
2804906d46d4SStefano Zampini       ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr);
2805906d46d4SStefano Zampini       ierr = PetscNew(&change_ctx);CHKERRQ(ierr);
2806906d46d4SStefano Zampini       ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr);
2807906d46d4SStefano Zampini     } else {
2808906d46d4SStefano Zampini       ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr);
2809906d46d4SStefano Zampini       ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr);
2810906d46d4SStefano Zampini       ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr);
2811906d46d4SStefano Zampini     }
2812906d46d4SStefano Zampini     if (!pcbddc->user_ChangeOfBasisMatrix) {
2813906d46d4SStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr);
2814906d46d4SStefano Zampini       change_ctx->global_change = pcbddc->ChangeOfBasisMatrix;
2815906d46d4SStefano Zampini     } else {
2816906d46d4SStefano Zampini       ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr);
2817906d46d4SStefano Zampini       change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix;
2818906d46d4SStefano Zampini     }
2819906d46d4SStefano Zampini     ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr);
2820906d46d4SStefano Zampini     ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr);
2821906d46d4SStefano Zampini     ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2822906d46d4SStefano Zampini     ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
2823b9b85e73SStefano Zampini   }
2824a717540cSStefano Zampini 
2825727cdba6SStefano Zampini   /* get indices in local ordering for vertices and constraints */
2826727cdba6SStefano Zampini   if (olocal_primal_size == pcbddc->local_primal_size) { /* if this is true, I need to check if a new primal space has been introduced */
2827473ba861SJed Brown     ierr = PetscMalloc1(olocal_primal_size,&oprimal_indices_local_idxs);CHKERRQ(ierr);
2828727cdba6SStefano Zampini     ierr = PetscMemcpy(oprimal_indices_local_idxs,pcbddc->primal_indices_local_idxs,olocal_primal_size*sizeof(PetscInt));CHKERRQ(ierr);
2829727cdba6SStefano Zampini   }
2830727cdba6SStefano Zampini   ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr);
2831f347579bSStefano Zampini   ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
2832473ba861SJed Brown   ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr);
2833727cdba6SStefano Zampini   ierr = PCBDDCGetPrimalVerticesLocalIdx(pc,&i,&aux_primal_numbering);CHKERRQ(ierr);
2834727cdba6SStefano Zampini   ierr = PetscMemcpy(pcbddc->primal_indices_local_idxs,aux_primal_numbering,i*sizeof(PetscInt));CHKERRQ(ierr);
2835727cdba6SStefano Zampini   ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr);
2836727cdba6SStefano Zampini   ierr = PCBDDCGetPrimalConstraintsLocalIdx(pc,&j,&aux_primal_numbering);CHKERRQ(ierr);
2837727cdba6SStefano Zampini   ierr = PetscMemcpy(&pcbddc->primal_indices_local_idxs[i],aux_primal_numbering,j*sizeof(PetscInt));CHKERRQ(ierr);
2838727cdba6SStefano Zampini   ierr = PetscFree(aux_primal_numbering);CHKERRQ(ierr);
2839e9189074SStefano Zampini   /* set quantities in PCBDDC data struct */
2840e9189074SStefano Zampini   pcbddc->n_actual_vertices = i;
2841727cdba6SStefano Zampini   /* check if a new primal space has been introduced */
2842727cdba6SStefano Zampini   pcbddc->new_primal_space_local = PETSC_TRUE;
2843727cdba6SStefano Zampini   if (olocal_primal_size == pcbddc->local_primal_size) {
2844727cdba6SStefano Zampini     ierr = PetscMemcmp(pcbddc->primal_indices_local_idxs,oprimal_indices_local_idxs,olocal_primal_size,&pcbddc->new_primal_space_local);CHKERRQ(ierr);
2845c1c8e736SStefano Zampini     pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
2846727cdba6SStefano Zampini     ierr = PetscFree(oprimal_indices_local_idxs);CHKERRQ(ierr);
2847727cdba6SStefano Zampini   }
2848727cdba6SStefano Zampini   /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */
2849727cdba6SStefano Zampini   ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
2850727cdba6SStefano Zampini 
2851a717540cSStefano Zampini   /* flush dbg viewer */
2852b8ffe317SStefano Zampini   if (pcbddc->dbg_flag) {
2853b8ffe317SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
2854b8ffe317SStefano Zampini   }
2855a717540cSStefano Zampini 
2856e310c8b4SStefano Zampini   /* free workspace */
2857a717540cSStefano Zampini   ierr = PetscBTDestroy(&touched);CHKERRQ(ierr);
2858a717540cSStefano Zampini   ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr);
2859984c4197SStefano Zampini   ierr = PetscFree(aux_primal_minloc);CHKERRQ(ierr);
28604641a718SStefano Zampini   ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr);
286108122e43SStefano Zampini   if (!pcbddc->adaptive_selection) {
2862906d46d4SStefano Zampini     ierr = PetscFree(temp_indices);CHKERRQ(ierr);
2863906d46d4SStefano Zampini     ierr = PetscFree3(temp_quadrature_constraint,temp_indices_to_constraint,temp_indices_to_constraint_B);CHKERRQ(ierr);
286408122e43SStefano Zampini   } else {
286508122e43SStefano Zampini     ierr = PetscFree4(pcbddc->adaptive_constraints_n,
286608122e43SStefano Zampini                       pcbddc->adaptive_constraints_ptrs,
286708122e43SStefano Zampini                       pcbddc->adaptive_constraints_idxs,
286808122e43SStefano Zampini                       pcbddc->adaptive_constraints_data);CHKERRQ(ierr);
286908122e43SStefano Zampini     ierr = PetscFree(temp_indices_to_constraint_B);CHKERRQ(ierr);
287008122e43SStefano Zampini   }
2871674ae819SStefano Zampini   PetscFunctionReturn(0);
2872674ae819SStefano Zampini }
2873674ae819SStefano Zampini 
2874674ae819SStefano Zampini #undef __FUNCT__
2875674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface"
2876674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc)
2877674ae819SStefano Zampini {
2878674ae819SStefano Zampini   PC_BDDC     *pcbddc = (PC_BDDC*)pc->data;
2879674ae819SStefano Zampini   PC_IS       *pcis = (PC_IS*)pc->data;
2880674ae819SStefano Zampini   Mat_IS      *matis  = (Mat_IS*)pc->pmat->data;
288163602bcaSStefano Zampini   PetscInt    ierr,i,vertex_size;
2882674ae819SStefano Zampini   PetscViewer viewer=pcbddc->dbg_viewer;
2883674ae819SStefano Zampini 
2884674ae819SStefano Zampini   PetscFunctionBegin;
28858e61c736SStefano Zampini   /* Reset previously computed graph */
28868e61c736SStefano Zampini   ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr);
2887674ae819SStefano Zampini   /* Init local Graph struct */
2888674ae819SStefano Zampini   ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping);CHKERRQ(ierr);
2889674ae819SStefano Zampini 
2890575ad6abSStefano Zampini   /* Check validity of the csr graph passed in by the user */
2891575ad6abSStefano Zampini   if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) {
2892575ad6abSStefano Zampini     ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr);
2893575ad6abSStefano Zampini   }
28949577ea80SStefano Zampini 
2895674ae819SStefano Zampini   /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */
28964d379d7bSStefano Zampini   if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) {
28974d379d7bSStefano Zampini     PetscInt  *xadj,*adjncy;
28984d379d7bSStefano Zampini     PetscInt  nvtxs;
2899e496cd5dSStefano Zampini     PetscBool flg_row=PETSC_FALSE;
2900674ae819SStefano Zampini 
29014d379d7bSStefano Zampini     if (pcbddc->use_local_adj) {
29022fffb893SStefano Zampini 
29032fffb893SStefano Zampini       ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
29042fffb893SStefano Zampini       if (flg_row) {
29054d379d7bSStefano Zampini         ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr);
2906b96c3477SStefano Zampini         pcbddc->computed_rowadj = PETSC_TRUE;
29072fffb893SStefano Zampini       }
29082fffb893SStefano Zampini       ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
29092fffb893SStefano Zampini     } else if (pcbddc->current_level) { /* just compute subdomain's connected components for coarser levels */
29104d379d7bSStefano Zampini       IS                     is_dummy;
29114d379d7bSStefano Zampini       ISLocalToGlobalMapping l2gmap_dummy;
29124d379d7bSStefano Zampini       PetscInt               j,sum;
29134d379d7bSStefano Zampini       PetscInt               *cxadj,*cadjncy;
29144d379d7bSStefano Zampini       const PetscInt         *idxs;
29154d379d7bSStefano Zampini       PCBDDCGraph            graph;
29164d379d7bSStefano Zampini       PetscBT                is_on_boundary;
29174d379d7bSStefano Zampini 
29184d379d7bSStefano Zampini       ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr);
29194d379d7bSStefano Zampini       ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr);
29204d379d7bSStefano Zampini       ierr = ISDestroy(&is_dummy);CHKERRQ(ierr);
29214d379d7bSStefano Zampini       ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr);
29224d379d7bSStefano Zampini       ierr = PCBDDCGraphInit(graph,l2gmap_dummy);CHKERRQ(ierr);
29234d379d7bSStefano Zampini       ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr);
2924e496cd5dSStefano Zampini       ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
2925e496cd5dSStefano Zampini       if (flg_row) {
29264d379d7bSStefano Zampini         graph->xadj = xadj;
29274d379d7bSStefano Zampini         graph->adjncy = adjncy;
2928e496cd5dSStefano Zampini       }
29294d379d7bSStefano Zampini       ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr);
29304d379d7bSStefano Zampini       ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr);
2931e496cd5dSStefano Zampini       ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr);
29324d379d7bSStefano Zampini 
29334d379d7bSStefano Zampini       if (pcbddc->dbg_flag) {
29344d379d7bSStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains\n",PetscGlobalRank,graph->ncc);CHKERRQ(ierr);
29354d379d7bSStefano Zampini         for (i=0;i<graph->ncc;i++) {
29364d379d7bSStefano Zampini           ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr);
29374d379d7bSStefano Zampini         }
29384d379d7bSStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
29394d379d7bSStefano Zampini       }
29404d379d7bSStefano Zampini 
2941e496cd5dSStefano Zampini       ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr);
29424d379d7bSStefano Zampini       ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
29434d379d7bSStefano Zampini       for (i=0;i<pcis->n_B;i++) {
29444d379d7bSStefano Zampini         ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr);
29454d379d7bSStefano Zampini       }
29464d379d7bSStefano Zampini       ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr);
29474d379d7bSStefano Zampini 
2948e496cd5dSStefano Zampini       ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr);
29494d379d7bSStefano Zampini       sum = 0;
29504d379d7bSStefano Zampini       for (i=0;i<graph->ncc;i++) {
29514d379d7bSStefano Zampini         PetscInt sizecc = 0;
29524d379d7bSStefano Zampini         for (j=graph->cptr[i];j<graph->cptr[i+1];j++) {
29534d379d7bSStefano Zampini           if (PetscBTLookup(is_on_boundary,graph->queue[j])) {
29544d379d7bSStefano Zampini             sizecc++;
29554d379d7bSStefano Zampini           }
29564d379d7bSStefano Zampini         }
29574d379d7bSStefano Zampini         for (j=graph->cptr[i];j<graph->cptr[i+1];j++) {
29584d379d7bSStefano Zampini           if (PetscBTLookup(is_on_boundary,graph->queue[j])) {
29594d379d7bSStefano Zampini             cxadj[graph->queue[j]] = sizecc;
29604d379d7bSStefano Zampini           }
29614d379d7bSStefano Zampini         }
29624d379d7bSStefano Zampini         sum += sizecc*sizecc;
29634d379d7bSStefano Zampini       }
29644d379d7bSStefano Zampini       ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr);
29654d379d7bSStefano Zampini       sum = 0;
2966e496cd5dSStefano Zampini       for (i=0;i<pcis->n;i++) {
29674d379d7bSStefano Zampini         PetscInt temp = cxadj[i];
29684d379d7bSStefano Zampini         cxadj[i] = sum;
29694d379d7bSStefano Zampini         sum += temp;
29704d379d7bSStefano Zampini       }
2971e496cd5dSStefano Zampini       cxadj[pcis->n] = sum;
29724d379d7bSStefano Zampini       for (i=0;i<graph->ncc;i++) {
29734d379d7bSStefano Zampini         for (j=graph->cptr[i];j<graph->cptr[i+1];j++) {
29744d379d7bSStefano Zampini           if (PetscBTLookup(is_on_boundary,graph->queue[j])) {
29754d379d7bSStefano Zampini             PetscInt k,sizecc = 0;
29764d379d7bSStefano Zampini             for (k=graph->cptr[i];k<graph->cptr[i+1];k++) {
29774d379d7bSStefano Zampini               if (PetscBTLookup(is_on_boundary,graph->queue[k])) {
29784d379d7bSStefano Zampini                 cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k];
29794d379d7bSStefano Zampini                 sizecc++;
29804d379d7bSStefano Zampini               }
29814d379d7bSStefano Zampini             }
29824d379d7bSStefano Zampini           }
29834d379d7bSStefano Zampini         }
29844d379d7bSStefano Zampini       }
2985e496cd5dSStefano Zampini       if (pcis->n) {
2986e496cd5dSStefano Zampini         ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr);
29874d379d7bSStefano Zampini       } else {
29884d379d7bSStefano Zampini         ierr = PetscFree(cxadj);CHKERRQ(ierr);
29894d379d7bSStefano Zampini         ierr = PetscFree(cadjncy);CHKERRQ(ierr);
29904d379d7bSStefano Zampini       }
29914d379d7bSStefano Zampini       graph->xadj = 0;
29924d379d7bSStefano Zampini       graph->adjncy = 0;
29934d379d7bSStefano Zampini       ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr);
29944d379d7bSStefano Zampini       ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr);
29954d379d7bSStefano Zampini     }
2996674ae819SStefano Zampini   }
2997674ae819SStefano Zampini 
299863602bcaSStefano Zampini   /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */
2999674ae819SStefano Zampini   vertex_size = 1;
300063602bcaSStefano Zampini   if (pcbddc->user_provided_isfordofs) {
300163602bcaSStefano Zampini     if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */
300295ecbf38SStefano Zampini       ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr);
300363602bcaSStefano Zampini       for (i=0;i<pcbddc->n_ISForDofs;i++) {
3004a7dc3881SStefano Zampini         ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr);
300563602bcaSStefano Zampini         ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr);
3006674ae819SStefano Zampini       }
300763602bcaSStefano Zampini       pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs;
300863602bcaSStefano Zampini       pcbddc->n_ISForDofs = 0;
300963602bcaSStefano Zampini       ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr);
3010674ae819SStefano Zampini     }
301163602bcaSStefano Zampini     /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */
3012674ae819SStefano Zampini     ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr);
301363602bcaSStefano Zampini   } else {
301463602bcaSStefano Zampini     if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */
301563602bcaSStefano Zampini       ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr);
3016854ce69bSBarry Smith       ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr);
301763602bcaSStefano Zampini       for (i=0;i<pcbddc->n_ISForDofsLocal;i++) {
301863602bcaSStefano Zampini         ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr);
301963602bcaSStefano Zampini       }
302063602bcaSStefano Zampini     }
3021674ae819SStefano Zampini   }
3022674ae819SStefano Zampini 
3023674ae819SStefano Zampini   /* Setup of Graph */
3024785d1243SStefano Zampini   if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */
3025a7dc3881SStefano Zampini     ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr);
3026785d1243SStefano Zampini   }
3027785d1243SStefano Zampini   if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */
3028a7dc3881SStefano Zampini     ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr);
3029785d1243SStefano Zampini   }
3030302440fdSBarry Smith   ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr);
3031674ae819SStefano Zampini 
3032674ae819SStefano Zampini   /* Graph's connected components analysis */
3033674ae819SStefano Zampini   ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr);
3034674ae819SStefano Zampini 
3035674ae819SStefano Zampini   /* print some info to stdout */
3036674ae819SStefano Zampini   if (pcbddc->dbg_flag) {
3037302440fdSBarry Smith     ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr);
3038674ae819SStefano Zampini   }
3039fb180af4SStefano Zampini 
3040fb180af4SStefano Zampini   /* mark topography has done */
3041fb180af4SStefano Zampini   pcbddc->recompute_topography = PETSC_FALSE;
3042674ae819SStefano Zampini   PetscFunctionReturn(0);
3043674ae819SStefano Zampini }
3044674ae819SStefano Zampini 
3045674ae819SStefano Zampini #undef __FUNCT__
3046674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalVerticesLocalIdx"
3047f34684f1SStefano Zampini PetscErrorCode  PCBDDCGetPrimalVerticesLocalIdx(PC pc, PetscInt *n_vertices, PetscInt **vertices_idx)
3048674ae819SStefano Zampini {
3049674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)(pc->data);
3050674ae819SStefano Zampini   PetscInt       *vertices,*row_cmat_indices,n,i,size_of_constraint,local_primal_size;
3051674ae819SStefano Zampini   PetscErrorCode ierr;
3052674ae819SStefano Zampini 
3053674ae819SStefano Zampini   PetscFunctionBegin;
3054674ae819SStefano Zampini   n = 0;
3055674ae819SStefano Zampini   vertices = 0;
3056674ae819SStefano Zampini   if (pcbddc->ConstraintMatrix) {
3057674ae819SStefano Zampini     ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&i);CHKERRQ(ierr);
3058b120a5c6SStefano Zampini     for (i=0;i<local_primal_size;i++) {
3059b120a5c6SStefano Zampini       ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr);
3060b120a5c6SStefano Zampini       if (size_of_constraint == 1) n++;
3061b120a5c6SStefano Zampini       ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr);
3062b120a5c6SStefano Zampini     }
3063811e8ca2SStefano Zampini     if (vertices_idx) {
3064785e854fSJed Brown       ierr = PetscMalloc1(n,&vertices);CHKERRQ(ierr);
3065b120a5c6SStefano Zampini       n = 0;
3066674ae819SStefano Zampini       for (i=0;i<local_primal_size;i++) {
3067674ae819SStefano Zampini         ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr);
3068674ae819SStefano Zampini         if (size_of_constraint == 1) {
3069674ae819SStefano Zampini           vertices[n++]=row_cmat_indices[0];
3070674ae819SStefano Zampini         }
3071674ae819SStefano Zampini         ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr);
3072674ae819SStefano Zampini       }
3073674ae819SStefano Zampini     }
3074811e8ca2SStefano Zampini   }
3075674ae819SStefano Zampini   *n_vertices = n;
3076811e8ca2SStefano Zampini   if (vertices_idx) *vertices_idx = vertices;
3077674ae819SStefano Zampini   PetscFunctionReturn(0);
3078674ae819SStefano Zampini }
3079674ae819SStefano Zampini 
3080674ae819SStefano Zampini #undef __FUNCT__
3081674ae819SStefano Zampini #define __FUNCT__ "PCBDDCGetPrimalConstraintsLocalIdx"
3082f34684f1SStefano Zampini PetscErrorCode  PCBDDCGetPrimalConstraintsLocalIdx(PC pc, PetscInt *n_constraints, PetscInt **constraints_idx)
3083674ae819SStefano Zampini {
3084674ae819SStefano Zampini   PC_BDDC        *pcbddc = (PC_BDDC*)(pc->data);
3085674ae819SStefano Zampini   PetscInt       *constraints_index,*row_cmat_indices,*row_cmat_global_indices;
3086674ae819SStefano Zampini   PetscInt       n,i,j,size_of_constraint,local_primal_size,local_size,max_size_of_constraint,min_index,min_loc;
30874641a718SStefano Zampini   PetscBT        touched;
3088674ae819SStefano Zampini   PetscErrorCode ierr;
3089674ae819SStefano Zampini 
3090f34684f1SStefano Zampini     /* This function assumes that the number of local constraints per connected component
3091f34684f1SStefano Zampini        is not greater than the number of nodes defined for the connected component
3092f34684f1SStefano Zampini        (otherwise we will surely have linear dependence between constraints and thus a singular coarse problem) */
3093674ae819SStefano Zampini   PetscFunctionBegin;
3094674ae819SStefano Zampini   n = 0;
3095674ae819SStefano Zampini   constraints_index = 0;
3096674ae819SStefano Zampini   if (pcbddc->ConstraintMatrix) {
3097674ae819SStefano Zampini     ierr = MatGetSize(pcbddc->ConstraintMatrix,&local_primal_size,&local_size);CHKERRQ(ierr);
3098674ae819SStefano Zampini     max_size_of_constraint = 0;
3099674ae819SStefano Zampini     for (i=0;i<local_primal_size;i++) {
3100674ae819SStefano Zampini       ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr);
3101674ae819SStefano Zampini       if (size_of_constraint > 1) {
3102674ae819SStefano Zampini         n++;
3103674ae819SStefano Zampini       }
3104674ae819SStefano Zampini       max_size_of_constraint = PetscMax(size_of_constraint,max_size_of_constraint);
3105674ae819SStefano Zampini       ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,NULL,NULL);CHKERRQ(ierr);
3106674ae819SStefano Zampini     }
3107811e8ca2SStefano Zampini     if (constraints_idx) {
3108785e854fSJed Brown       ierr = PetscMalloc1(n,&constraints_index);CHKERRQ(ierr);
3109785e854fSJed Brown       ierr = PetscMalloc1(max_size_of_constraint,&row_cmat_global_indices);CHKERRQ(ierr);
31104641a718SStefano Zampini       ierr = PetscBTCreate(local_size,&touched);CHKERRQ(ierr);
3111674ae819SStefano Zampini       n = 0;
3112674ae819SStefano Zampini       for (i=0;i<local_primal_size;i++) {
3113674ae819SStefano Zampini         ierr = MatGetRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr);
3114674ae819SStefano Zampini         if (size_of_constraint > 1) {
3115674ae819SStefano Zampini           ierr = ISLocalToGlobalMappingApply(pcbddc->mat_graph->l2gmap,size_of_constraint,row_cmat_indices,row_cmat_global_indices);CHKERRQ(ierr);
311682d3d8afSStefano Zampini           /* find first untouched local node */
311782d3d8afSStefano Zampini           j = 0;
31184641a718SStefano Zampini           while (PetscBTLookup(touched,row_cmat_indices[j])) j++;
311982d3d8afSStefano Zampini           min_index = row_cmat_global_indices[j];
312082d3d8afSStefano Zampini           min_loc = j;
312182d3d8afSStefano Zampini           /* search the minimum among nodes not yet touched on the connected component
312282d3d8afSStefano Zampini              since there can be more than one constraint on a single cc */
3123674ae819SStefano Zampini           for (j=1;j<size_of_constraint;j++) {
31244641a718SStefano Zampini             if (!PetscBTLookup(touched,row_cmat_indices[j]) && min_index > row_cmat_global_indices[j]) {
3125674ae819SStefano Zampini               min_index = row_cmat_global_indices[j];
3126674ae819SStefano Zampini               min_loc = j;
3127674ae819SStefano Zampini             }
3128674ae819SStefano Zampini           }
31294641a718SStefano Zampini           ierr = PetscBTSet(touched,row_cmat_indices[min_loc]);CHKERRQ(ierr);
3130674ae819SStefano Zampini           constraints_index[n++] = row_cmat_indices[min_loc];
3131674ae819SStefano Zampini         }
3132674ae819SStefano Zampini         ierr = MatRestoreRow(pcbddc->ConstraintMatrix,i,&size_of_constraint,(const PetscInt**)&row_cmat_indices,NULL);CHKERRQ(ierr);
3133674ae819SStefano Zampini       }
31344641a718SStefano Zampini       ierr = PetscBTDestroy(&touched);CHKERRQ(ierr);
3135674ae819SStefano Zampini       ierr = PetscFree(row_cmat_global_indices);CHKERRQ(ierr);
3136811e8ca2SStefano Zampini     }
3137811e8ca2SStefano Zampini   }
3138674ae819SStefano Zampini   *n_constraints = n;
3139811e8ca2SStefano Zampini   if (constraints_idx) *constraints_idx = constraints_index;
3140674ae819SStefano Zampini   PetscFunctionReturn(0);
3141674ae819SStefano Zampini }
3142674ae819SStefano Zampini 
3143674ae819SStefano Zampini #undef __FUNCT__
3144674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering"
3145674ae819SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(MPI_Comm comm,ISLocalToGlobalMapping l2gmap, PetscInt n_local_dofs, PetscInt local_dofs[], PetscInt local_dofs_mult[], PetscInt* n_global_subset, PetscInt* global_numbering_subset[])
3146674ae819SStefano Zampini {
3147674ae819SStefano Zampini   Vec            local_vec,global_vec;
3148674ae819SStefano Zampini   IS             seqis,paris;
3149674ae819SStefano Zampini   VecScatter     scatter_ctx;
3150674ae819SStefano Zampini   PetscScalar    *array;
3151674ae819SStefano Zampini   PetscInt       *temp_global_dofs;
3152674ae819SStefano Zampini   PetscScalar    globalsum;
3153674ae819SStefano Zampini   PetscInt       i,j,s;
315417ccb9bdSStefano Zampini   PetscInt       nlocals,first_index,old_index,max_local,max_global;
315517ccb9bdSStefano Zampini   PetscMPIInt    rank_prec_comm,size_prec_comm;
315617ccb9bdSStefano Zampini   PetscInt       *dof_sizes,*dof_displs;
3157674ae819SStefano Zampini   PetscBool      first_found;
3158674ae819SStefano Zampini   PetscErrorCode ierr;
3159674ae819SStefano Zampini 
3160674ae819SStefano Zampini   PetscFunctionBegin;
3161674ae819SStefano Zampini   /* mpi buffers */
3162b9b85e73SStefano Zampini   ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr);
3163b9b85e73SStefano Zampini   ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr);
3164674ae819SStefano Zampini   j = ( !rank_prec_comm ? size_prec_comm : 0);
316517ccb9bdSStefano Zampini   ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr);
3166674ae819SStefano Zampini   /* get maximum size of subset */
3167785e854fSJed Brown   ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr);
3168674ae819SStefano Zampini   ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr);
3169674ae819SStefano Zampini   max_local = 0;
3170b9b85e73SStefano Zampini   for (i=0;i<n_local_dofs;i++) {
3171674ae819SStefano Zampini     if (max_local < temp_global_dofs[i] ) {
3172674ae819SStefano Zampini       max_local = temp_global_dofs[i];
3173674ae819SStefano Zampini     }
3174674ae819SStefano Zampini   }
3175b9b85e73SStefano Zampini   ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr);
3176674ae819SStefano Zampini   max_global++;
3177674ae819SStefano Zampini   max_local = 0;
3178b9b85e73SStefano Zampini   for (i=0;i<n_local_dofs;i++) {
3179674ae819SStefano Zampini     if (max_local < local_dofs[i] ) {
3180674ae819SStefano Zampini       max_local = local_dofs[i];
3181674ae819SStefano Zampini     }
3182674ae819SStefano Zampini   }
3183674ae819SStefano Zampini   max_local++;
3184674ae819SStefano Zampini   /* allocate workspace */
3185674ae819SStefano Zampini   ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr);
3186674ae819SStefano Zampini   ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr);
3187674ae819SStefano Zampini   ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr);
3188674ae819SStefano Zampini   ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr);
3189674ae819SStefano Zampini   ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr);
3190674ae819SStefano Zampini   ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr);
3191674ae819SStefano Zampini   /* create scatter */
3192674ae819SStefano Zampini   ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr);
3193674ae819SStefano Zampini   ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr);
3194674ae819SStefano Zampini   ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr);
3195674ae819SStefano Zampini   ierr = ISDestroy(&seqis);CHKERRQ(ierr);
3196674ae819SStefano Zampini   ierr = ISDestroy(&paris);CHKERRQ(ierr);
3197674ae819SStefano Zampini   /* init array */
3198674ae819SStefano Zampini   ierr = VecSet(global_vec,0.0);CHKERRQ(ierr);
3199674ae819SStefano Zampini   ierr = VecSet(local_vec,0.0);CHKERRQ(ierr);
3200674ae819SStefano Zampini   ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr);
3201674ae819SStefano Zampini   if (local_dofs_mult) {
3202674ae819SStefano Zampini     for (i=0;i<n_local_dofs;i++) {
3203674ae819SStefano Zampini       array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i];
3204674ae819SStefano Zampini     }
3205674ae819SStefano Zampini   } else {
3206674ae819SStefano Zampini     for (i=0;i<n_local_dofs;i++) {
3207674ae819SStefano Zampini       array[local_dofs[i]]=1.0;
3208674ae819SStefano Zampini     }
3209674ae819SStefano Zampini   }
3210674ae819SStefano Zampini   ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr);
3211674ae819SStefano Zampini   /* scatter into global vec and get total number of global dofs */
3212674ae819SStefano Zampini   ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
3213674ae819SStefano Zampini   ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
3214674ae819SStefano Zampini   ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr);
32155b08dc53SStefano Zampini   *n_global_subset = (PetscInt)PetscRealPart(globalsum);
3216674ae819SStefano Zampini   /* Fill global_vec with cumulative function for global numbering */
3217674ae819SStefano Zampini   ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr);
3218674ae819SStefano Zampini   ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr);
3219674ae819SStefano Zampini   nlocals = 0;
3220674ae819SStefano Zampini   first_index = -1;
3221674ae819SStefano Zampini   first_found = PETSC_FALSE;
3222674ae819SStefano Zampini   for (i=0;i<s;i++) {
3223b9b85e73SStefano Zampini     if (!first_found && PetscRealPart(array[i]) > 0.1) {
3224674ae819SStefano Zampini       first_found = PETSC_TRUE;
3225674ae819SStefano Zampini       first_index = i;
3226674ae819SStefano Zampini     }
32275b08dc53SStefano Zampini     nlocals += (PetscInt)PetscRealPart(array[i]);
3228674ae819SStefano Zampini   }
3229674ae819SStefano Zampini   ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr);
3230674ae819SStefano Zampini   if (!rank_prec_comm) {
3231674ae819SStefano Zampini     dof_displs[0]=0;
3232674ae819SStefano Zampini     for (i=1;i<size_prec_comm;i++) {
3233674ae819SStefano Zampini       dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1];
3234674ae819SStefano Zampini     }
3235674ae819SStefano Zampini   }
3236674ae819SStefano Zampini   ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr);
3237674ae819SStefano Zampini   if (first_found) {
3238674ae819SStefano Zampini     array[first_index] += (PetscScalar)nlocals;
3239674ae819SStefano Zampini     old_index = first_index;
3240674ae819SStefano Zampini     for (i=first_index+1;i<s;i++) {
3241b9b85e73SStefano Zampini       if (PetscRealPart(array[i]) > 0.1) {
3242674ae819SStefano Zampini         array[i] += array[old_index];
3243674ae819SStefano Zampini         old_index = i;
3244674ae819SStefano Zampini       }
3245674ae819SStefano Zampini     }
3246674ae819SStefano Zampini   }
3247674ae819SStefano Zampini   ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr);
3248674ae819SStefano Zampini   ierr = VecSet(local_vec,0.0);CHKERRQ(ierr);
3249674ae819SStefano Zampini   ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
3250674ae819SStefano Zampini   ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
3251674ae819SStefano Zampini   /* get global ordering of local dofs */
32522b095fd8SStefano Zampini   ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr);
3253674ae819SStefano Zampini   if (local_dofs_mult) {
3254674ae819SStefano Zampini     for (i=0;i<n_local_dofs;i++) {
32555b08dc53SStefano Zampini       temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i];
3256674ae819SStefano Zampini     }
3257674ae819SStefano Zampini   } else {
3258674ae819SStefano Zampini     for (i=0;i<n_local_dofs;i++) {
32595b08dc53SStefano Zampini       temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1;
3260674ae819SStefano Zampini     }
3261674ae819SStefano Zampini   }
32622b095fd8SStefano Zampini   ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr);
3263674ae819SStefano Zampini   /* free workspace */
3264674ae819SStefano Zampini   ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr);
3265674ae819SStefano Zampini   ierr = VecDestroy(&local_vec);CHKERRQ(ierr);
3266674ae819SStefano Zampini   ierr = VecDestroy(&global_vec);CHKERRQ(ierr);
326717ccb9bdSStefano Zampini   ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr);
3268674ae819SStefano Zampini   /* return pointer to global ordering of local dofs */
3269674ae819SStefano Zampini   *global_numbering_subset = temp_global_dofs;
3270674ae819SStefano Zampini   PetscFunctionReturn(0);
3271674ae819SStefano Zampini }
32729a7d3425SStefano Zampini 
32739a7d3425SStefano Zampini #undef __FUNCT__
32749a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs"
32759a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[])
32769a7d3425SStefano Zampini {
32779a7d3425SStefano Zampini   PetscInt       i,j;
32789a7d3425SStefano Zampini   PetscScalar    *alphas;
32799a7d3425SStefano Zampini   PetscErrorCode ierr;
32809a7d3425SStefano Zampini 
32819a7d3425SStefano Zampini   PetscFunctionBegin;
32829a7d3425SStefano Zampini   /* this implements stabilized Gram-Schmidt */
3283785e854fSJed Brown   ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr);
32849a7d3425SStefano Zampini   for (i=0;i<n;i++) {
32859a7d3425SStefano Zampini     ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr);
32869a7d3425SStefano Zampini     if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); }
32879a7d3425SStefano Zampini     for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); }
32889a7d3425SStefano Zampini   }
32899a7d3425SStefano Zampini   ierr = PetscFree(alphas);CHKERRQ(ierr);
32909a7d3425SStefano Zampini   PetscFunctionReturn(0);
32919a7d3425SStefano Zampini }
32929a7d3425SStefano Zampini 
3293e7931f94SStefano Zampini #undef __FUNCT__
329470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern"
329528143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends)
3296e7931f94SStefano Zampini {
3297e7931f94SStefano Zampini   Mat             subdomain_adj;
3298e7931f94SStefano Zampini   IS              new_ranks,ranks_send_to;
3299e7931f94SStefano Zampini   MatPartitioning partitioner;
3300e7931f94SStefano Zampini   Mat_IS          *matis;
3301e7931f94SStefano Zampini   PetscInt        n_neighs,*neighs,*n_shared,**shared;
3302e7931f94SStefano Zampini   PetscInt        prank;
3303e7931f94SStefano Zampini   PetscMPIInt     size,rank,color;
3304e7931f94SStefano Zampini   PetscInt        *xadj,*adjncy,*oldranks;
3305e7931f94SStefano Zampini   PetscInt        *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx;
33063837a79fSStefano Zampini   PetscInt        i,local_size,threshold=0;
3307e7931f94SStefano Zampini   PetscErrorCode  ierr;
33082b510759SStefano Zampini   PetscBool       use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE;
3309e7931f94SStefano Zampini   PetscSubcomm    subcomm;
3310a57a6d2fSStefano Zampini 
3311e7931f94SStefano Zampini   PetscFunctionBegin;
33122b510759SStefano Zampini   ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr);
33132b510759SStefano Zampini   ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr);
33142b510759SStefano Zampini   ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr);
3315e7931f94SStefano Zampini 
3316e7931f94SStefano Zampini   /* Get info on mapping */
3317e7931f94SStefano Zampini   matis = (Mat_IS*)(mat->data);
3318e7931f94SStefano Zampini   ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr);
3319e7931f94SStefano Zampini   ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr);
3320e7931f94SStefano Zampini 
3321e7931f94SStefano Zampini   /* build local CSR graph of subdomains' connectivity */
3322785e854fSJed Brown   ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr);
3323e7931f94SStefano Zampini   xadj[0] = 0;
3324e7931f94SStefano Zampini   xadj[1] = PetscMax(n_neighs-1,0);
3325785e854fSJed Brown   ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr);
3326785e854fSJed Brown   ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr);
3327e7931f94SStefano Zampini 
33282b510759SStefano Zampini   if (threshold) {
3329d023bfaeSStefano Zampini     PetscInt xadj_count = 0;
33302b510759SStefano Zampini     for (i=1;i<n_neighs;i++) {
3331d023bfaeSStefano Zampini       if (n_shared[i] > threshold) {
3332d023bfaeSStefano Zampini         adjncy[xadj_count] = neighs[i];
3333d023bfaeSStefano Zampini         adjncy_wgt[xadj_count] = n_shared[i];
3334d023bfaeSStefano Zampini         xadj_count++;
3335e7931f94SStefano Zampini       }
3336e7931f94SStefano Zampini     }
3337d023bfaeSStefano Zampini     xadj[1] = xadj_count;
3338c8587f34SStefano Zampini   } else {
3339e7931f94SStefano Zampini     if (xadj[1]) {
3340e7931f94SStefano Zampini       ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr);
3341e7931f94SStefano Zampini       ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr);
3342c8587f34SStefano Zampini     }
3343e7931f94SStefano Zampini   }
3344e7931f94SStefano Zampini   ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr);
3345e7931f94SStefano Zampini   if (use_square) {
3346e7931f94SStefano Zampini     for (i=0;i<xadj[1];i++) {
3347e7931f94SStefano Zampini       adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i];
3348e7931f94SStefano Zampini     }
3349e7931f94SStefano Zampini   }
3350e7931f94SStefano Zampini   ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr);
3351e7931f94SStefano Zampini 
33523837a79fSStefano Zampini   ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr);
3353e7931f94SStefano Zampini 
3354e7931f94SStefano Zampini   /*
3355e7931f94SStefano Zampini     Restrict work on active processes only.
3356e7931f94SStefano Zampini   */
3357e7931f94SStefano Zampini   ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr);
3358e7931f94SStefano Zampini   ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */
3359e7931f94SStefano Zampini   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr);
33602b510759SStefano Zampini   ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr);
3361d3531aaaSJed Brown   ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr);
3362e7931f94SStefano Zampini   if (color) {
3363e7931f94SStefano Zampini     ierr = PetscFree(xadj);CHKERRQ(ierr);
3364e7931f94SStefano Zampini     ierr = PetscFree(adjncy);CHKERRQ(ierr);
3365e7931f94SStefano Zampini     ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr);
3366c8587f34SStefano Zampini   } else {
336728143c3dSStefano Zampini     PetscInt coarsening_ratio;
3368306c2d5bSBarry Smith     ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr);
3369785e854fSJed Brown     ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr);
3370e7931f94SStefano Zampini     prank = rank;
3371306c2d5bSBarry Smith     ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr);
33728002ef2cSStefano Zampini     /*
3373e7931f94SStefano Zampini     for (i=0;i<size;i++) {
3374e7931f94SStefano Zampini       PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]);
3375c8587f34SStefano Zampini     }
33768002ef2cSStefano Zampini     */
3377e7931f94SStefano Zampini     for (i=0;i<xadj[1];i++) {
3378e7931f94SStefano Zampini       ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr);
3379c8587f34SStefano Zampini     }
3380e7931f94SStefano Zampini     ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr);
3381306c2d5bSBarry Smith     ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr);
338222b6e8a2SStefano Zampini     /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */
3383e7931f94SStefano Zampini 
3384e7931f94SStefano Zampini     /* Partition */
3385306c2d5bSBarry Smith     ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr);
3386e7931f94SStefano Zampini     ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr);
3387e7931f94SStefano Zampini     if (use_vwgt) {
33883837a79fSStefano Zampini       ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr);
3389e7931f94SStefano Zampini       v_wgt[0] = local_size;
3390e7931f94SStefano Zampini       ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr);
3391c8587f34SStefano Zampini     }
339228143c3dSStefano Zampini     n_subdomains = PetscMin((PetscInt)size,n_subdomains);
339328143c3dSStefano Zampini     coarsening_ratio = size/n_subdomains;
339428143c3dSStefano Zampini     ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr);
3395e7931f94SStefano Zampini     ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr);
3396e7931f94SStefano Zampini     ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr);
339722b6e8a2SStefano Zampini     /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */
3398e7931f94SStefano Zampini 
3399e7931f94SStefano Zampini     ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr);
340028143c3dSStefano Zampini     if (contiguous) {
340128143c3dSStefano Zampini       ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */
340228143c3dSStefano Zampini     } else {
340384ba6060SStefano Zampini       ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */
340428143c3dSStefano Zampini     }
3405e7931f94SStefano Zampini     ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr);
3406e7931f94SStefano Zampini     /* clean up */
3407e7931f94SStefano Zampini     ierr = PetscFree(oldranks);CHKERRQ(ierr);
3408e7931f94SStefano Zampini     ierr = ISDestroy(&new_ranks);CHKERRQ(ierr);
3409e7931f94SStefano Zampini     ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr);
3410e7931f94SStefano Zampini     ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr);
3411e7931f94SStefano Zampini   }
3412e7931f94SStefano Zampini   ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr);
3413e7931f94SStefano Zampini 
3414e7931f94SStefano Zampini   /* assemble parallel IS for sends */
3415e7931f94SStefano Zampini   i = 1;
3416e7931f94SStefano Zampini   if (color) i=0;
3417e7931f94SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr);
3418e7931f94SStefano Zampini 
3419e7931f94SStefano Zampini   /* get back IS */
3420e7931f94SStefano Zampini   *is_sends = ranks_send_to;
3421e7931f94SStefano Zampini   PetscFunctionReturn(0);
3422e7931f94SStefano Zampini }
3423e7931f94SStefano Zampini 
3424e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate;
3425e7931f94SStefano Zampini 
3426e7931f94SStefano Zampini #undef __FUNCT__
3427e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble"
342828143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[])
3429e7931f94SStefano Zampini {
343070cf5478SStefano Zampini   Mat                    local_mat;
3431e7931f94SStefano Zampini   Mat_IS                 *matis;
3432e7931f94SStefano Zampini   IS                     is_sends_internal;
34339d30be91SStefano Zampini   PetscInt               rows,cols,new_local_rows;
343428143c3dSStefano Zampini   PetscInt               i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals;
34359d30be91SStefano Zampini   PetscBool              ismatis,isdense,newisdense,destroy_mat;
3436e7931f94SStefano Zampini   ISLocalToGlobalMapping l2gmap;
3437e7931f94SStefano Zampini   PetscInt*              l2gmap_indices;
3438e7931f94SStefano Zampini   const PetscInt*        is_indices;
3439e7931f94SStefano Zampini   MatType                new_local_type;
3440e7931f94SStefano Zampini   /* buffers */
3441e7931f94SStefano Zampini   PetscInt               *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs;
344228143c3dSStefano Zampini   PetscInt               *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is;
34439d30be91SStefano Zampini   PetscInt               *recv_buffer_idxs_local;
3444e7931f94SStefano Zampini   PetscScalar            *ptr_vals,*send_buffer_vals,*recv_buffer_vals;
3445e7931f94SStefano Zampini   /* MPI */
344628143c3dSStefano Zampini   MPI_Comm               comm,comm_n;
344728143c3dSStefano Zampini   PetscSubcomm           subcomm;
3448e7931f94SStefano Zampini   PetscMPIInt            n_sends,n_recvs,commsize;
344928143c3dSStefano Zampini   PetscMPIInt            *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is;
345028143c3dSStefano Zampini   PetscMPIInt            *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals;
345128143c3dSStefano Zampini   PetscMPIInt            len,tag_idxs,tag_idxs_is,tag_vals,source_dest;
345228143c3dSStefano Zampini   MPI_Request            *send_req_idxs,*send_req_idxs_is,*send_req_vals;
345328143c3dSStefano Zampini   MPI_Request            *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals;
3454e7931f94SStefano Zampini   PetscErrorCode         ierr;
3455e7931f94SStefano Zampini 
3456e7931f94SStefano Zampini   PetscFunctionBegin;
345728143c3dSStefano Zampini   /* TODO: add missing checks */
345828143c3dSStefano Zampini   PetscValidLogicalCollectiveInt(mat,n_subdomains,3);
345928143c3dSStefano Zampini   PetscValidLogicalCollectiveBool(mat,restrict_comm,4);
346028143c3dSStefano Zampini   PetscValidLogicalCollectiveEnum(mat,reuse,5);
346128143c3dSStefano Zampini   PetscValidLogicalCollectiveInt(mat,nis,7);
3462e7931f94SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr);
346328143c3dSStefano Zampini   if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__);
3464e7931f94SStefano Zampini   ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr);
3465e7931f94SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr);
3466e7931f94SStefano Zampini   if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE");
3467e7931f94SStefano Zampini   ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr);
3468e7931f94SStefano Zampini   if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square");
346928143c3dSStefano Zampini   if (reuse == MAT_REUSE_MATRIX && *mat_n) {
347070cf5478SStefano Zampini     PetscInt mrows,mcols,mnrows,mncols;
347170cf5478SStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr);
347228143c3dSStefano Zampini     if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS");
347370cf5478SStefano Zampini     ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr);
347470cf5478SStefano Zampini     ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr);
347570cf5478SStefano Zampini     if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows);
347670cf5478SStefano Zampini     if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols);
347770cf5478SStefano Zampini   }
3478e7931f94SStefano Zampini   ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr);
3479e7931f94SStefano Zampini   PetscValidLogicalCollectiveInt(mat,bs,0);
3480e7931f94SStefano Zampini   /* prepare IS for sending if not provided */
3481e7931f94SStefano Zampini   if (!is_sends) {
348228143c3dSStefano Zampini     PetscBool pcontig = PETSC_TRUE;
348328143c3dSStefano Zampini     if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains");
348428143c3dSStefano Zampini     ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr);
3485c8587f34SStefano Zampini   } else {
3486e7931f94SStefano Zampini     ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr);
3487e7931f94SStefano Zampini     is_sends_internal = is_sends;
3488c8587f34SStefano Zampini   }
3489e7931f94SStefano Zampini 
3490e7931f94SStefano Zampini   /* get pointer of MATIS data */
3491e7931f94SStefano Zampini   matis = (Mat_IS*)mat->data;
3492e7931f94SStefano Zampini 
3493e7931f94SStefano Zampini   /* get comm */
3494a316fed8SStefano Zampini   ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
3495e7931f94SStefano Zampini 
3496e7931f94SStefano Zampini   /* compute number of sends */
3497e7931f94SStefano Zampini   ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr);
3498e7931f94SStefano Zampini   ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr);
3499e7931f94SStefano Zampini 
3500e7931f94SStefano Zampini   /* compute number of receives */
3501e7931f94SStefano Zampini   ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr);
3502785e854fSJed Brown   ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr);
3503e7931f94SStefano Zampini   ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr);
3504e7931f94SStefano Zampini   ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr);
3505e7931f94SStefano Zampini   for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1;
3506e7931f94SStefano Zampini   ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr);
3507e7931f94SStefano Zampini   ierr = PetscFree(iflags);CHKERRQ(ierr);
3508e7931f94SStefano Zampini 
350928143c3dSStefano Zampini   /* restrict comm if requested */
351028143c3dSStefano Zampini   subcomm = 0;
351128143c3dSStefano Zampini   destroy_mat = PETSC_FALSE;
351228143c3dSStefano Zampini   if (restrict_comm) {
3513779c1cceSStefano Zampini     PetscMPIInt color,subcommsize;
3514779c1cceSStefano Zampini 
351528143c3dSStefano Zampini     color = 0;
3516779c1cceSStefano Zampini     if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm */
351728143c3dSStefano Zampini     ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr);
351828143c3dSStefano Zampini     subcommsize = commsize - subcommsize;
351928143c3dSStefano Zampini     /* check if reuse has been requested */
352028143c3dSStefano Zampini     if (reuse == MAT_REUSE_MATRIX) {
352128143c3dSStefano Zampini       if (*mat_n) {
352228143c3dSStefano Zampini         PetscMPIInt subcommsize2;
352328143c3dSStefano Zampini         ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr);
352428143c3dSStefano Zampini         if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2);
352528143c3dSStefano Zampini         comm_n = PetscObjectComm((PetscObject)*mat_n);
352628143c3dSStefano Zampini       } else {
352728143c3dSStefano Zampini         comm_n = PETSC_COMM_SELF;
352828143c3dSStefano Zampini       }
352928143c3dSStefano Zampini     } else { /* MAT_INITIAL_MATRIX */
3530779c1cceSStefano Zampini       PetscMPIInt rank;
3531779c1cceSStefano Zampini 
3532779c1cceSStefano Zampini       ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
353328143c3dSStefano Zampini       ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr);
353428143c3dSStefano Zampini       ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr);
353528143c3dSStefano Zampini       ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr);
3536306c2d5bSBarry Smith       comm_n = PetscSubcommChild(subcomm);
353728143c3dSStefano Zampini     }
353828143c3dSStefano Zampini     /* flag to destroy *mat_n if not significative */
353928143c3dSStefano Zampini     if (color) destroy_mat = PETSC_TRUE;
354028143c3dSStefano Zampini   } else {
354128143c3dSStefano Zampini     comm_n = comm;
354228143c3dSStefano Zampini   }
354328143c3dSStefano Zampini 
3544e7931f94SStefano Zampini   /* prepare send/receive buffers */
3545785e854fSJed Brown   ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr);
3546e7931f94SStefano Zampini   ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr);
3547785e854fSJed Brown   ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr);
3548e7931f94SStefano Zampini   ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr);
354928143c3dSStefano Zampini   if (nis) {
3550854ce69bSBarry Smith     ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr);
355128143c3dSStefano Zampini   }
3552e7931f94SStefano Zampini 
355328143c3dSStefano Zampini   /* Get data from local matrices */
3554e7931f94SStefano Zampini   if (!isdense) {
3555a26c9d0eSStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented");
3556e7931f94SStefano Zampini     /* TODO: See below some guidelines on how to prepare the local buffers */
3557e7931f94SStefano Zampini     /*
3558e7931f94SStefano Zampini        send_buffer_vals should contain the raw values of the local matrix
3559e7931f94SStefano Zampini        send_buffer_idxs should contain:
3560e7931f94SStefano Zampini        - MatType_PRIVATE type
3561e7931f94SStefano Zampini        - PetscInt        size_of_l2gmap
3562e7931f94SStefano Zampini        - PetscInt        global_row_indices[size_of_l2gmap]
3563e7931f94SStefano Zampini        - PetscInt        all_other_info_which_is_needed_to_compute_preallocation_and_set_values
3564e7931f94SStefano Zampini     */
3565e7931f94SStefano Zampini   } else {
3566e7931f94SStefano Zampini     ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr);
3567e7931f94SStefano Zampini     ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr);
3568854ce69bSBarry Smith     ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr);
3569e7931f94SStefano Zampini     send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE;
3570e7931f94SStefano Zampini     send_buffer_idxs[1] = i;
3571e7931f94SStefano Zampini     ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr);
3572e7931f94SStefano Zampini     ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr);
3573e7931f94SStefano Zampini     ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr);
3574e7931f94SStefano Zampini     ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr);
3575e7931f94SStefano Zampini     for (i=0;i<n_sends;i++) {
3576e7931f94SStefano Zampini       ilengths_vals[is_indices[i]] = len*len;
3577e7931f94SStefano Zampini       ilengths_idxs[is_indices[i]] = len+2;
3578c8587f34SStefano Zampini     }
3579c8587f34SStefano Zampini   }
3580e7931f94SStefano Zampini   ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr);
358128143c3dSStefano Zampini   /* additional is (if any) */
358228143c3dSStefano Zampini   if (nis) {
358328143c3dSStefano Zampini     PetscMPIInt psum;
358428143c3dSStefano Zampini     PetscInt j;
358528143c3dSStefano Zampini     for (j=0,psum=0;j<nis;j++) {
358628143c3dSStefano Zampini       PetscInt plen;
358728143c3dSStefano Zampini       ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr);
358828143c3dSStefano Zampini       ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr);
358928143c3dSStefano Zampini       psum += len+1; /* indices + lenght */
359028143c3dSStefano Zampini     }
3591854ce69bSBarry Smith     ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr);
359228143c3dSStefano Zampini     for (j=0,psum=0;j<nis;j++) {
359328143c3dSStefano Zampini       PetscInt plen;
359428143c3dSStefano Zampini       const PetscInt *is_array_idxs;
359528143c3dSStefano Zampini       ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr);
359628143c3dSStefano Zampini       send_buffer_idxs_is[psum] = plen;
359728143c3dSStefano Zampini       ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr);
359828143c3dSStefano Zampini       ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr);
359928143c3dSStefano Zampini       ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr);
360028143c3dSStefano Zampini       psum += plen+1; /* indices + lenght */
360128143c3dSStefano Zampini     }
360228143c3dSStefano Zampini     for (i=0;i<n_sends;i++) {
360328143c3dSStefano Zampini       ilengths_idxs_is[is_indices[i]] = psum;
360428143c3dSStefano Zampini     }
360528143c3dSStefano Zampini     ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr);
360628143c3dSStefano Zampini   }
360728143c3dSStefano Zampini 
3608e7931f94SStefano Zampini   buf_size_idxs = 0;
3609e7931f94SStefano Zampini   buf_size_vals = 0;
361028143c3dSStefano Zampini   buf_size_idxs_is = 0;
3611e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
3612e7931f94SStefano Zampini     buf_size_idxs += (PetscInt)olengths_idxs[i];
3613e7931f94SStefano Zampini     buf_size_vals += (PetscInt)olengths_vals[i];
361428143c3dSStefano Zampini     if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i];
3615e7931f94SStefano Zampini   }
3616785e854fSJed Brown   ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr);
3617785e854fSJed Brown   ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr);
361895ecbf38SStefano Zampini   ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr);
3619e7931f94SStefano Zampini 
3620e7931f94SStefano Zampini   /* get new tags for clean communications */
3621e7931f94SStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr);
3622e7931f94SStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr);
362328143c3dSStefano Zampini   ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr);
3624e7931f94SStefano Zampini 
3625e7931f94SStefano Zampini   /* allocate for requests */
3626785e854fSJed Brown   ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr);
3627785e854fSJed Brown   ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr);
362895ecbf38SStefano Zampini   ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr);
3629785e854fSJed Brown   ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr);
3630785e854fSJed Brown   ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr);
363195ecbf38SStefano Zampini   ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr);
3632e7931f94SStefano Zampini 
3633e7931f94SStefano Zampini   /* communications */
3634e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
3635e7931f94SStefano Zampini   ptr_vals = recv_buffer_vals;
363628143c3dSStefano Zampini   ptr_idxs_is = recv_buffer_idxs_is;
3637e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
3638e7931f94SStefano Zampini     source_dest = onodes[i];
3639e7931f94SStefano Zampini     ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr);
3640e7931f94SStefano Zampini     ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr);
3641e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
3642e7931f94SStefano Zampini     ptr_vals += olengths_vals[i];
364328143c3dSStefano Zampini     if (nis) {
364428143c3dSStefano 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);
364528143c3dSStefano Zampini       ptr_idxs_is += olengths_idxs_is[i];
364628143c3dSStefano Zampini     }
3647e7931f94SStefano Zampini   }
3648e7931f94SStefano Zampini   for (i=0;i<n_sends;i++) {
3649e7931f94SStefano Zampini     ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr);
3650e7931f94SStefano Zampini     ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr);
3651e7931f94SStefano Zampini     ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr);
365228143c3dSStefano Zampini     if (nis) {
365328143c3dSStefano 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);
365428143c3dSStefano Zampini     }
3655e7931f94SStefano Zampini   }
3656e7931f94SStefano Zampini   ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr);
3657e7931f94SStefano Zampini   ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr);
3658e7931f94SStefano Zampini 
3659e7931f94SStefano Zampini   /* assemble new l2g map */
3660e7931f94SStefano Zampini   ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
3661e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
36629d30be91SStefano Zampini   new_local_rows = 0;
3663e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
36649d30be91SStefano Zampini     new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
3665e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
3666e7931f94SStefano Zampini   }
36679d30be91SStefano Zampini   ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr);
3668e7931f94SStefano Zampini   ptr_idxs = recv_buffer_idxs;
36699d30be91SStefano Zampini   new_local_rows = 0;
3670e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
36719d30be91SStefano Zampini     ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr);
36729d30be91SStefano Zampini     new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
3673e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
3674e7931f94SStefano Zampini   }
36759d30be91SStefano Zampini   ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr);
36769d30be91SStefano Zampini   ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr);
3677e7931f94SStefano Zampini   ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr);
3678e7931f94SStefano Zampini 
3679e7931f94SStefano Zampini   /* infer new local matrix type from received local matrices type */
3680e7931f94SStefano Zampini   /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */
3681e7931f94SStefano 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) */
3682e7931f94SStefano Zampini   if (n_recvs) {
368328143c3dSStefano Zampini     MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0];
3684e7931f94SStefano Zampini     ptr_idxs = recv_buffer_idxs;
3685e7931f94SStefano Zampini     for (i=0;i<n_recvs;i++) {
3686e7931f94SStefano Zampini       if ((PetscInt)new_local_type_private != *ptr_idxs) {
3687e7931f94SStefano Zampini         new_local_type_private = MATAIJ_PRIVATE;
3688e7931f94SStefano Zampini         break;
3689e7931f94SStefano Zampini       }
3690e7931f94SStefano Zampini       ptr_idxs += olengths_idxs[i];
3691e7931f94SStefano Zampini     }
3692e7931f94SStefano Zampini     switch (new_local_type_private) {
369328143c3dSStefano Zampini       case MATDENSE_PRIVATE:
369428143c3dSStefano Zampini         if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */
3695e7931f94SStefano Zampini           new_local_type = MATSEQAIJ;
3696e7931f94SStefano Zampini           bs = 1;
369728143c3dSStefano Zampini         } else { /* if I receive only 1 dense matrix */
369828143c3dSStefano Zampini           new_local_type = MATSEQDENSE;
369928143c3dSStefano Zampini           bs = 1;
370028143c3dSStefano Zampini         }
3701e7931f94SStefano Zampini         break;
3702e7931f94SStefano Zampini       case MATAIJ_PRIVATE:
3703e7931f94SStefano Zampini         new_local_type = MATSEQAIJ;
3704e7931f94SStefano Zampini         bs = 1;
3705e7931f94SStefano Zampini         break;
3706e7931f94SStefano Zampini       case MATBAIJ_PRIVATE:
3707e7931f94SStefano Zampini         new_local_type = MATSEQBAIJ;
3708e7931f94SStefano Zampini         break;
3709e7931f94SStefano Zampini       case MATSBAIJ_PRIVATE:
3710e7931f94SStefano Zampini         new_local_type = MATSEQSBAIJ;
3711e7931f94SStefano Zampini         break;
3712e7931f94SStefano Zampini       default:
37139d30be91SStefano Zampini         SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__);
3714e7931f94SStefano Zampini         break;
3715e7931f94SStefano Zampini     }
371628143c3dSStefano Zampini   } else { /* by default, new_local_type is seqdense */
371728143c3dSStefano Zampini     new_local_type = MATSEQDENSE;
371828143c3dSStefano Zampini     bs = 1;
3719e7931f94SStefano Zampini   }
3720e7931f94SStefano Zampini 
372170cf5478SStefano Zampini   /* create MATIS object if needed */
372270cf5478SStefano Zampini   if (reuse == MAT_INITIAL_MATRIX) {
3723e7931f94SStefano Zampini     ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr);
372428143c3dSStefano Zampini     ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr);
372570cf5478SStefano Zampini   } else {
372670cf5478SStefano Zampini     /* it also destroys the local matrices */
372770cf5478SStefano Zampini     ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr);
372870cf5478SStefano Zampini   }
372970cf5478SStefano Zampini   ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr);
3730e7931f94SStefano Zampini   ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr);
37319d30be91SStefano Zampini 
37329d30be91SStefano Zampini   ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
37339d30be91SStefano Zampini 
37349d30be91SStefano Zampini   /* Global to local map of received indices */
37359d30be91SStefano Zampini   ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */
37369d30be91SStefano Zampini   ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr);
37379d30be91SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr);
37389d30be91SStefano Zampini 
37399d30be91SStefano Zampini   /* restore attributes -> type of incoming data and its size */
37409d30be91SStefano Zampini   buf_size_idxs = 0;
37419d30be91SStefano Zampini   for (i=0;i<n_recvs;i++) {
37429d30be91SStefano Zampini     recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs];
37439d30be91SStefano Zampini     recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1];
37449d30be91SStefano Zampini     buf_size_idxs += (PetscInt)olengths_idxs[i];
37459d30be91SStefano Zampini   }
37469d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr);
37479d30be91SStefano Zampini 
37489d30be91SStefano Zampini   /* set preallocation */
37499d30be91SStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr);
37509d30be91SStefano Zampini   if (!newisdense) {
37519d30be91SStefano Zampini     PetscInt *new_local_nnz=0;
37529d30be91SStefano Zampini 
37539d30be91SStefano Zampini     ptr_vals = recv_buffer_vals;
37549d30be91SStefano Zampini     ptr_idxs = recv_buffer_idxs_local;
37559d30be91SStefano Zampini     if (n_recvs) {
37569d30be91SStefano Zampini       ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr);
37579d30be91SStefano Zampini     }
37589d30be91SStefano Zampini     for (i=0;i<n_recvs;i++) {
37599d30be91SStefano Zampini       PetscInt j;
37609d30be91SStefano Zampini       if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */
37619d30be91SStefano Zampini         for (j=0;j<*(ptr_idxs+1);j++) {
37629d30be91SStefano Zampini           new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1);
37639d30be91SStefano Zampini         }
37649d30be91SStefano Zampini       } else {
37659d30be91SStefano Zampini         /* TODO */
37669d30be91SStefano Zampini       }
37679d30be91SStefano Zampini       ptr_idxs += olengths_idxs[i];
37689d30be91SStefano Zampini     }
37699d30be91SStefano Zampini     if (new_local_nnz) {
37709d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows);
37719d30be91SStefano Zampini       ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr);
37729d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs;
37739d30be91SStefano Zampini       ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr);
37749d30be91SStefano Zampini       for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0);
37759d30be91SStefano Zampini       ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr);
37769d30be91SStefano Zampini     } else {
37779d30be91SStefano Zampini       ierr = MatSetUp(local_mat);CHKERRQ(ierr);
37789d30be91SStefano Zampini     }
37799d30be91SStefano Zampini     ierr = PetscFree(new_local_nnz);CHKERRQ(ierr);
37809d30be91SStefano Zampini   } else {
37819d30be91SStefano Zampini     ierr = MatSetUp(local_mat);CHKERRQ(ierr);
37829d30be91SStefano Zampini   }
3783e7931f94SStefano Zampini 
3784e7931f94SStefano Zampini   /* set values */
3785e7931f94SStefano Zampini   ptr_vals = recv_buffer_vals;
37869d30be91SStefano Zampini   ptr_idxs = recv_buffer_idxs_local;
3787e7931f94SStefano Zampini   for (i=0;i<n_recvs;i++) {
3788e7931f94SStefano Zampini     if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */
3789e7931f94SStefano Zampini       ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr);
37909d30be91SStefano Zampini       ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr);
3791e7931f94SStefano Zampini       ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr);
3792e7931f94SStefano Zampini       ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr);
3793e7931f94SStefano Zampini       ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr);
379428143c3dSStefano Zampini     } else {
379528143c3dSStefano Zampini       /* TODO */
3796e7931f94SStefano Zampini     }
3797e7931f94SStefano Zampini     ptr_idxs += olengths_idxs[i];
3798e7931f94SStefano Zampini     ptr_vals += olengths_vals[i];
3799e7931f94SStefano Zampini   }
3800e7931f94SStefano Zampini   ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
3801e7931f94SStefano Zampini   ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
380270cf5478SStefano Zampini   ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
380370cf5478SStefano Zampini   ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
38049d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr);
38059d30be91SStefano Zampini   ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr);
3806e7931f94SStefano Zampini 
3807dfd14d43SStefano Zampini #if 0
380828143c3dSStefano Zampini   if (!restrict_comm) { /* check */
3809e7931f94SStefano Zampini     Vec       lvec,rvec;
3810e7931f94SStefano Zampini     PetscReal infty_error;
3811e7931f94SStefano Zampini 
38122a7a6963SBarry Smith     ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr);
3813e7931f94SStefano Zampini     ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr);
3814e7931f94SStefano Zampini     ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr);
3815e7931f94SStefano Zampini     ierr = VecScale(lvec,-1.0);CHKERRQ(ierr);
381670cf5478SStefano Zampini     ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr);
3817e7931f94SStefano Zampini     ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr);
3818e7931f94SStefano Zampini     ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error);
3819e7931f94SStefano Zampini     ierr = VecDestroy(&rvec);CHKERRQ(ierr);
3820e7931f94SStefano Zampini     ierr = VecDestroy(&lvec);CHKERRQ(ierr);
3821e7931f94SStefano Zampini   }
382228143c3dSStefano Zampini #endif
3823e7931f94SStefano Zampini 
382428143c3dSStefano Zampini   /* assemble new additional is (if any) */
382528143c3dSStefano Zampini   if (nis) {
382628143c3dSStefano Zampini     PetscInt **temp_idxs,*count_is,j,psum;
382728143c3dSStefano Zampini 
382828143c3dSStefano Zampini     ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
3829854ce69bSBarry Smith     ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr);
383028143c3dSStefano Zampini     ptr_idxs = recv_buffer_idxs_is;
383128143c3dSStefano Zampini     psum = 0;
383228143c3dSStefano Zampini     for (i=0;i<n_recvs;i++) {
383328143c3dSStefano Zampini       for (j=0;j<nis;j++) {
383428143c3dSStefano Zampini         PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
383528143c3dSStefano Zampini         count_is[j] += plen; /* increment counting of buffer for j-th IS */
383628143c3dSStefano Zampini         psum += plen;
383728143c3dSStefano Zampini         ptr_idxs += plen+1; /* shift pointer to received data */
383828143c3dSStefano Zampini       }
383928143c3dSStefano Zampini     }
3840854ce69bSBarry Smith     ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr);
3841854ce69bSBarry Smith     ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr);
384228143c3dSStefano Zampini     for (i=1;i<nis;i++) {
384328143c3dSStefano Zampini       temp_idxs[i] = temp_idxs[i-1]+count_is[i-1];
384428143c3dSStefano Zampini     }
384528143c3dSStefano Zampini     ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr);
384628143c3dSStefano Zampini     ptr_idxs = recv_buffer_idxs_is;
384728143c3dSStefano Zampini     for (i=0;i<n_recvs;i++) {
384828143c3dSStefano Zampini       for (j=0;j<nis;j++) {
384928143c3dSStefano Zampini         PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
385028143c3dSStefano Zampini         ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr);
385128143c3dSStefano Zampini         count_is[j] += plen; /* increment starting point of buffer for j-th IS */
385228143c3dSStefano Zampini         ptr_idxs += plen+1; /* shift pointer to received data */
385328143c3dSStefano Zampini       }
385428143c3dSStefano Zampini     }
385528143c3dSStefano Zampini     for (i=0;i<nis;i++) {
385628143c3dSStefano Zampini       ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
385728143c3dSStefano Zampini       ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr);
385828143c3dSStefano Zampini       ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr);
385928143c3dSStefano Zampini     }
386028143c3dSStefano Zampini     ierr = PetscFree(count_is);CHKERRQ(ierr);
386128143c3dSStefano Zampini     ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr);
386228143c3dSStefano Zampini     ierr = PetscFree(temp_idxs);CHKERRQ(ierr);
386328143c3dSStefano Zampini   }
3864e7931f94SStefano Zampini   /* free workspace */
386528143c3dSStefano Zampini   ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr);
3866e7931f94SStefano Zampini   ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
3867e7931f94SStefano Zampini   ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr);
3868e7931f94SStefano Zampini   ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
3869e7931f94SStefano Zampini   if (isdense) {
3870e7931f94SStefano Zampini     ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr);
3871e7931f94SStefano Zampini     ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr);
3872e7931f94SStefano Zampini   } else {
3873e7931f94SStefano Zampini     /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */
3874e7931f94SStefano Zampini   }
387528143c3dSStefano Zampini   if (nis) {
387628143c3dSStefano Zampini     ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
387728143c3dSStefano Zampini     ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr);
387828143c3dSStefano Zampini   }
3879e7931f94SStefano Zampini   ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr);
3880e7931f94SStefano Zampini   ierr = PetscFree(recv_req_vals);CHKERRQ(ierr);
388128143c3dSStefano Zampini   ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr);
3882e7931f94SStefano Zampini   ierr = PetscFree(send_req_idxs);CHKERRQ(ierr);
3883e7931f94SStefano Zampini   ierr = PetscFree(send_req_vals);CHKERRQ(ierr);
388428143c3dSStefano Zampini   ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr);
3885e7931f94SStefano Zampini   ierr = PetscFree(ilengths_vals);CHKERRQ(ierr);
3886e7931f94SStefano Zampini   ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr);
3887e7931f94SStefano Zampini   ierr = PetscFree(olengths_vals);CHKERRQ(ierr);
3888e7931f94SStefano Zampini   ierr = PetscFree(olengths_idxs);CHKERRQ(ierr);
3889e7931f94SStefano Zampini   ierr = PetscFree(onodes);CHKERRQ(ierr);
389028143c3dSStefano Zampini   if (nis) {
389128143c3dSStefano Zampini     ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr);
389228143c3dSStefano Zampini     ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr);
389328143c3dSStefano Zampini     ierr = PetscFree(onodes_is);CHKERRQ(ierr);
389428143c3dSStefano Zampini   }
389528143c3dSStefano Zampini   ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr);
389628143c3dSStefano Zampini   if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */
389728143c3dSStefano Zampini     ierr = MatDestroy(mat_n);CHKERRQ(ierr);
389828143c3dSStefano Zampini     for (i=0;i<nis;i++) {
389928143c3dSStefano Zampini       ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
390028143c3dSStefano Zampini     }
390128143c3dSStefano Zampini   }
3902e7931f94SStefano Zampini   PetscFunctionReturn(0);
3903e7931f94SStefano Zampini }
3904a57a6d2fSStefano Zampini 
390512edc857SStefano Zampini /* temporary hack into ksp private data structure */
3906af0996ceSBarry Smith #include <petsc/private/kspimpl.h>
390712edc857SStefano Zampini 
3908c8587f34SStefano Zampini #undef __FUNCT__
3909c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver"
3910c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals)
3911c8587f34SStefano Zampini {
3912c8587f34SStefano Zampini   PC_BDDC                *pcbddc = (PC_BDDC*)pc->data;
3913c8587f34SStefano Zampini   PC_IS                  *pcis = (PC_IS*)pc->data;
391420a2ab83SStefano Zampini   Mat                    coarse_mat,coarse_mat_is,coarse_submat_dense;
39159881197aSStefano Zampini   MatNullSpace           CoarseNullSpace=NULL;
391620a2ab83SStefano Zampini   ISLocalToGlobalMapping coarse_islg;
39176e683305SStefano Zampini   IS                     coarse_is,*isarray;
39186e683305SStefano Zampini   PetscInt               i,im_active=-1,active_procs=-1;
39196e683305SStefano Zampini   PetscInt               nis,nisdofs,nisneu;
3920f9eb5b7dSStefano Zampini   PC                     pc_temp;
3921c8587f34SStefano Zampini   PCType                 coarse_pc_type;
3922c8587f34SStefano Zampini   KSPType                coarse_ksp_type;
3923f9eb5b7dSStefano Zampini   PetscBool              multilevel_requested,multilevel_allowed;
39244f3a063dSStefano Zampini   PetscBool              isredundant,isbddc,isnn,coarse_reuse;
39256e683305SStefano Zampini   Mat                    t_coarse_mat_is;
39266e683305SStefano Zampini   PetscInt               void_procs,ncoarse_ml,ncoarse_ds,ncoarse;
39276e683305SStefano Zampini   PetscMPIInt            all_procs;
392874e2c79eSStefano Zampini   PetscBool              csin_ml,csin_ds,csin,csin_type_simple,redist;
392968457ee5SStefano Zampini   PetscBool              compute_vecs = PETSC_FALSE;
393022bc73bbSStefano Zampini   PetscScalar            *array;
39319881197aSStefano Zampini   PetscErrorCode         ierr;
3932fdc09c96SStefano Zampini 
3933c8587f34SStefano Zampini   PetscFunctionBegin;
3934c8587f34SStefano Zampini   /* Assign global numbering to coarse dofs */
393568457ee5SStefano 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 */
3936fa7f1dd8SStefano Zampini     PetscInt ocoarse_size;
39375a75c04eSSatish Balay     compute_vecs = PETSC_TRUE;
3938fa7f1dd8SStefano Zampini     ocoarse_size = pcbddc->coarse_size;
3939f4ddd8eeSStefano Zampini     ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr);
3940f4ddd8eeSStefano Zampini     ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr);
3941f4ddd8eeSStefano Zampini     /* see if we can avoid some work */
3942fa7f1dd8SStefano Zampini     if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */
3943fa7f1dd8SStefano Zampini       if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */
3944727cdba6SStefano Zampini         ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr);
3945fa7f1dd8SStefano Zampini         coarse_reuse = PETSC_FALSE;
3946fa7f1dd8SStefano Zampini       } else { /* we can safely reuse already computed coarse matrix */
3947fa7f1dd8SStefano Zampini         coarse_reuse = PETSC_TRUE;
3948f4ddd8eeSStefano Zampini       }
3949fa7f1dd8SStefano Zampini     } else { /* there's no coarse ksp, so we need to create the coarse matrix too */
3950fa7f1dd8SStefano Zampini       coarse_reuse = PETSC_FALSE;
3951f4ddd8eeSStefano Zampini     }
395270cf5478SStefano Zampini     /* reset any subassembling information */
395370cf5478SStefano Zampini     ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr);
39546e683305SStefano Zampini     ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
39556e683305SStefano Zampini   } else { /* primal space is unchanged, so we can reuse coarse matrix */
3956fa7f1dd8SStefano Zampini     coarse_reuse = PETSC_TRUE;
3957f4ddd8eeSStefano Zampini   }
3958c8587f34SStefano Zampini 
39596e683305SStefano Zampini   /* count "active" (i.e. with positive local size) and "void" processes */
39602b510759SStefano Zampini   im_active = !!(pcis->n);
39612b510759SStefano Zampini   ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
39626e683305SStefano Zampini   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr);
39636e683305SStefano Zampini   void_procs = all_procs-active_procs;
39646e683305SStefano Zampini   csin_type_simple = PETSC_TRUE;
396574e2c79eSStefano Zampini   redist = PETSC_FALSE;
396622bc73bbSStefano Zampini   if (pcbddc->current_level && void_procs) {
39676e683305SStefano Zampini     csin_ml = PETSC_TRUE;
39686e683305SStefano Zampini     ncoarse_ml = void_procs;
3969779c1cceSStefano Zampini     /* it has no sense to redistribute on a set of processors larger than the number of active processes */
3970779c1cceSStefano Zampini     if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) {
397118a45a71SStefano Zampini       csin_ds = PETSC_TRUE;
397218a45a71SStefano Zampini       ncoarse_ds = pcbddc->redistribute_coarse;
397318a45a71SStefano Zampini       redist = PETSC_TRUE;
397418a45a71SStefano Zampini     } else {
39756e683305SStefano Zampini       csin_ds = PETSC_TRUE;
3976779c1cceSStefano Zampini       ncoarse_ds = active_procs;
3977779c1cceSStefano Zampini       redist = PETSC_TRUE;
397818a45a71SStefano Zampini     }
39796e683305SStefano Zampini   } else {
39806e683305SStefano Zampini     csin_ml = PETSC_FALSE;
39816e683305SStefano Zampini     ncoarse_ml = all_procs;
39826e683305SStefano Zampini     if (void_procs) {
39836e683305SStefano Zampini       csin_ds = PETSC_TRUE;
39846e683305SStefano Zampini       ncoarse_ds = void_procs;
39856e683305SStefano Zampini       csin_type_simple = PETSC_FALSE;
39866e683305SStefano Zampini     } else {
3987779c1cceSStefano Zampini       if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) {
398874e2c79eSStefano Zampini         csin_ds = PETSC_TRUE;
398974e2c79eSStefano Zampini         ncoarse_ds = pcbddc->redistribute_coarse;
399074e2c79eSStefano Zampini         redist = PETSC_TRUE;
399174e2c79eSStefano Zampini       } else {
39926e683305SStefano Zampini         csin_ds = PETSC_FALSE;
39936e683305SStefano Zampini         ncoarse_ds = all_procs;
39946e683305SStefano Zampini       }
39956e683305SStefano Zampini     }
399674e2c79eSStefano Zampini   }
39976e683305SStefano Zampini 
39986e683305SStefano Zampini   /*
39996e683305SStefano Zampini     test if we can go multilevel: three conditions must be satisfied:
40006e683305SStefano Zampini     - we have not exceeded the number of levels requested
40016e683305SStefano Zampini     - we can actually subassemble the active processes
40026e683305SStefano Zampini     - we can find a suitable number of MPI processes where we can place the subassembled problem
40036e683305SStefano Zampini   */
40046e683305SStefano Zampini   multilevel_allowed = PETSC_FALSE;
40056e683305SStefano Zampini   multilevel_requested = PETSC_FALSE;
40066e683305SStefano Zampini   if (pcbddc->current_level < pcbddc->max_levels) {
40076e683305SStefano Zampini     multilevel_requested = PETSC_TRUE;
40086e683305SStefano Zampini     if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) {
4009f9eb5b7dSStefano Zampini       multilevel_allowed = PETSC_FALSE;
40102b510759SStefano Zampini     } else {
4011f9eb5b7dSStefano Zampini       multilevel_allowed = PETSC_TRUE;
4012c8587f34SStefano Zampini     }
4013c8587f34SStefano Zampini   }
40146e683305SStefano Zampini   /* determine number of process partecipating to coarse solver */
40156e683305SStefano Zampini   if (multilevel_allowed) {
40166e683305SStefano Zampini     ncoarse = ncoarse_ml;
40176e683305SStefano Zampini     csin = csin_ml;
401858da7f69SStefano Zampini     redist = PETSC_FALSE;
40196e683305SStefano Zampini   } else {
40206e683305SStefano Zampini     ncoarse = ncoarse_ds;
40216e683305SStefano Zampini     csin = csin_ds;
40226e683305SStefano Zampini   }
4023e7931f94SStefano Zampini 
4024abbbba34SStefano Zampini   /* creates temporary l2gmap and IS for coarse indexes */
4025abbbba34SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr);
4026abbbba34SStefano Zampini   ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr);
4027abbbba34SStefano Zampini 
4028abbbba34SStefano Zampini   /* creates temporary MATIS object for coarse matrix */
402922bc73bbSStefano Zampini   ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr);
403022bc73bbSStefano Zampini   ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr);
403122bc73bbSStefano Zampini   ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr);
403222bc73bbSStefano Zampini   ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr);
4033b9b85e73SStefano Zampini #if 0
4034b9b85e73SStefano Zampini   {
4035b9b85e73SStefano Zampini     PetscViewer viewer;
4036b9b85e73SStefano Zampini     char filename[256];
4037b9b85e73SStefano Zampini     sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank);
4038b9b85e73SStefano Zampini     ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr);
4039b9b85e73SStefano Zampini     ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
4040b9b85e73SStefano Zampini     ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr);
4041b9b85e73SStefano Zampini     ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr);
4042b9b85e73SStefano Zampini   }
4043b9b85e73SStefano Zampini #endif
40446e683305SStefano Zampini   ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,&t_coarse_mat_is);CHKERRQ(ierr);
40456e683305SStefano Zampini   ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr);
40466e683305SStefano Zampini   ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
40476e683305SStefano Zampini   ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
4048abbbba34SStefano Zampini   ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr);
4049abbbba34SStefano Zampini 
40506e683305SStefano Zampini   /* compute dofs splitting and neumann boundaries for coarse dofs */
40516e683305SStefano Zampini   if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */
40526e683305SStefano Zampini     PetscInt               *tidxs,*tidxs2,nout,tsize,i;
40536e683305SStefano Zampini     const PetscInt         *idxs;
40546e683305SStefano Zampini     ISLocalToGlobalMapping tmap;
40556e683305SStefano Zampini 
40566e683305SStefano Zampini     /* create map between primal indices (in local representative ordering) and local primal numbering */
40570be93d54SStefano Zampini     ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr);
40586e683305SStefano Zampini     /* allocate space for temporary storage */
4059854ce69bSBarry Smith     ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr);
4060854ce69bSBarry Smith     ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr);
40616e683305SStefano Zampini     /* allocate for IS array */
40626e683305SStefano Zampini     nisdofs = pcbddc->n_ISForDofsLocal;
40636e683305SStefano Zampini     nisneu = !!pcbddc->NeumannBoundariesLocal;
40646e683305SStefano Zampini     nis = nisdofs + nisneu;
4065854ce69bSBarry Smith     ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr);
40666e683305SStefano Zampini     /* dofs splitting */
40676e683305SStefano Zampini     for (i=0;i<nisdofs;i++) {
40686e683305SStefano Zampini       /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */
40696e683305SStefano Zampini       ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr);
40706e683305SStefano Zampini       ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr);
40716e683305SStefano Zampini       ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr);
40726e683305SStefano Zampini       ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr);
40736e683305SStefano Zampini       ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr);
40746e683305SStefano Zampini       ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr);
40756e683305SStefano Zampini       /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */
40766e683305SStefano Zampini     }
40776e683305SStefano Zampini     /* neumann boundaries */
40786e683305SStefano Zampini     if (pcbddc->NeumannBoundariesLocal) {
40796e683305SStefano Zampini       /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */
40806e683305SStefano Zampini       ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr);
40816e683305SStefano Zampini       ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr);
40826e683305SStefano Zampini       ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr);
40836e683305SStefano Zampini       ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr);
40846e683305SStefano Zampini       ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr);
40856e683305SStefano Zampini       ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr);
40866e683305SStefano Zampini       /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */
40876e683305SStefano Zampini     }
40886e683305SStefano Zampini     /* free memory */
40896e683305SStefano Zampini     ierr = PetscFree(tidxs);CHKERRQ(ierr);
40906e683305SStefano Zampini     ierr = PetscFree(tidxs2);CHKERRQ(ierr);
40916e683305SStefano Zampini     ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr);
40926e683305SStefano Zampini   } else {
40936e683305SStefano Zampini     nis = 0;
40946e683305SStefano Zampini     nisdofs = 0;
40956e683305SStefano Zampini     nisneu = 0;
40966e683305SStefano Zampini     isarray = NULL;
40976e683305SStefano Zampini   }
40986e683305SStefano Zampini   /* destroy no longer needed map */
40996e683305SStefano Zampini   ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr);
41006e683305SStefano Zampini 
41016e683305SStefano Zampini   /* restrict on coarse candidates (if needed) */
41026e683305SStefano Zampini   coarse_mat_is = NULL;
41036e683305SStefano Zampini   if (csin) {
41046e683305SStefano Zampini     if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */
410574e2c79eSStefano Zampini       if (redist) {
410674e2c79eSStefano Zampini         PetscMPIInt rank;
4107779c1cceSStefano Zampini         PetscInt    spc,n_spc_p1,dest[1],destsize;
410874e2c79eSStefano Zampini 
410974e2c79eSStefano Zampini         ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr);
411058da7f69SStefano Zampini         spc = active_procs/ncoarse;
411158da7f69SStefano Zampini         n_spc_p1 = active_procs%ncoarse;
4112779c1cceSStefano Zampini         if (im_active) {
4113779c1cceSStefano Zampini           destsize = 1;
411474e2c79eSStefano Zampini           if (rank > n_spc_p1*(spc+1)-1) {
411574e2c79eSStefano Zampini             dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc;
411674e2c79eSStefano Zampini           } else {
411774e2c79eSStefano Zampini             dest[0] = rank/(spc+1);
411874e2c79eSStefano Zampini           }
411974e2c79eSStefano Zampini         } else {
4120779c1cceSStefano Zampini           destsize = 0;
41216e683305SStefano Zampini         }
4122779c1cceSStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
4123779c1cceSStefano Zampini       } else if (csin_type_simple) {
41246e683305SStefano Zampini         PetscMPIInt rank;
41256e683305SStefano Zampini         PetscInt    issize,isidx;
4126779c1cceSStefano Zampini 
41276e683305SStefano Zampini         ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr);
41286e683305SStefano Zampini         if (im_active) {
41296e683305SStefano Zampini           issize = 1;
41306e683305SStefano Zampini           isidx = (PetscInt)rank;
41316e683305SStefano Zampini         } else {
41326e683305SStefano Zampini           issize = 0;
41336e683305SStefano Zampini           isidx = -1;
41346e683305SStefano Zampini         }
41356e683305SStefano Zampini         ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
4136779c1cceSStefano Zampini       } else { /* get a suitable subassembling pattern from MATIS code */
41376e683305SStefano Zampini         ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr);
41386e683305SStefano Zampini       }
4139779c1cceSStefano Zampini 
4140779c1cceSStefano Zampini       /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */
4141779c1cceSStefano Zampini       if (!redist || ncoarse <= void_procs) {
4142779c1cceSStefano Zampini         PetscInt ncoarse_cand,tissize,*nisindices;
4143779c1cceSStefano Zampini         PetscInt *coarse_candidates;
4144779c1cceSStefano Zampini         const PetscInt* tisindices;
4145779c1cceSStefano Zampini 
4146779c1cceSStefano Zampini         /* get coarse candidates' ranks in pc communicator */
4147779c1cceSStefano Zampini         ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr);
4148779c1cceSStefano Zampini         ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
4149779c1cceSStefano Zampini         for (i=0,ncoarse_cand=0;i<all_procs;i++) {
4150779c1cceSStefano Zampini           if (!coarse_candidates[i]) {
4151779c1cceSStefano Zampini             coarse_candidates[ncoarse_cand++]=i;
4152779c1cceSStefano Zampini           }
4153779c1cceSStefano Zampini         }
4154779c1cceSStefano Zampini         if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse);
4155779c1cceSStefano Zampini 
4156779c1cceSStefano Zampini 
41576e683305SStefano Zampini         if (pcbddc->dbg_flag) {
41586e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
41596e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr);
41606e683305SStefano Zampini           ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr);
41616e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr);
4162779c1cceSStefano Zampini           for (i=0;i<ncoarse_cand;i++) {
41636e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr);
41646e683305SStefano Zampini           }
41656e683305SStefano Zampini           ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr);
41666e683305SStefano Zampini           ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
41676e683305SStefano Zampini         }
41686e683305SStefano Zampini         /* shift the pattern on coarse candidates */
41696e683305SStefano Zampini         ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr);
41706e683305SStefano Zampini         ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr);
4171854ce69bSBarry Smith         ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr);
41726e683305SStefano Zampini         for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]];
41736e683305SStefano Zampini         ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr);
41746e683305SStefano Zampini         ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr);
41756e683305SStefano Zampini         ierr = PetscFree(coarse_candidates);CHKERRQ(ierr);
41766e683305SStefano Zampini       }
41776e683305SStefano Zampini       if (pcbddc->dbg_flag) {
41786e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
41796e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr);
41806e683305SStefano Zampini         ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr);
41816e683305SStefano Zampini         ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
41826e683305SStefano Zampini       }
4183779c1cceSStefano Zampini     }
41846e683305SStefano Zampini     /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */
41856e683305SStefano Zampini     ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr);
41866e683305SStefano Zampini   } else {
41876e683305SStefano Zampini     if (pcbddc->dbg_flag) {
41886e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
41896e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr);
41906e683305SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
41916e683305SStefano Zampini     }
41926e683305SStefano Zampini     ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr);
41936e683305SStefano Zampini     coarse_mat_is = t_coarse_mat_is;
41946e683305SStefano Zampini   }
41956e683305SStefano Zampini 
41966e683305SStefano Zampini   /* create local to global scatters for coarse problem */
419768457ee5SStefano Zampini   if (compute_vecs) {
41986e683305SStefano Zampini     PetscInt lrows;
41996e683305SStefano Zampini     ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr);
42006e683305SStefano Zampini     if (coarse_mat_is) {
42016e683305SStefano Zampini       ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr);
42026e683305SStefano Zampini     } else {
42036e683305SStefano Zampini       lrows = 0;
42046e683305SStefano Zampini     }
42056e683305SStefano Zampini     ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr);
42066e683305SStefano Zampini     ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr);
42076e683305SStefano Zampini     ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr);
42086e683305SStefano Zampini     ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
42096e683305SStefano Zampini     ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr);
42106e683305SStefano Zampini   }
42116e683305SStefano Zampini   ierr = ISDestroy(&coarse_is);CHKERRQ(ierr);
42126e683305SStefano Zampini   ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr);
4213c8587f34SStefano Zampini 
4214f9eb5b7dSStefano Zampini   /* set defaults for coarse KSP and PC */
4215f9eb5b7dSStefano Zampini   if (multilevel_allowed) {
4216f9eb5b7dSStefano Zampini     coarse_ksp_type = KSPRICHARDSON;
4217f9eb5b7dSStefano Zampini     coarse_pc_type = PCBDDC;
4218f9eb5b7dSStefano Zampini   } else {
4219f9eb5b7dSStefano Zampini     coarse_ksp_type = KSPPREONLY;
4220f9eb5b7dSStefano Zampini     coarse_pc_type = PCREDUNDANT;
4221c8587f34SStefano Zampini   }
4222c8587f34SStefano Zampini 
42236e683305SStefano Zampini   /* print some info if requested */
42246e683305SStefano Zampini   if (pcbddc->dbg_flag) {
42256e683305SStefano Zampini     if (!multilevel_allowed) {
42266e683305SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
42276e683305SStefano Zampini       if (multilevel_requested) {
42286e683305SStefano 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);
42296e683305SStefano Zampini       } else if (pcbddc->max_levels) {
42306e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr);
42316e683305SStefano Zampini       }
42326e683305SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
42336e683305SStefano Zampini     }
42346e683305SStefano Zampini   }
42356e683305SStefano Zampini 
4236f9eb5b7dSStefano Zampini   /* create the coarse KSP object only once with defaults */
42376e683305SStefano Zampini   if (coarse_mat_is) {
42386e683305SStefano Zampini     MatReuse coarse_mat_reuse;
42396a1308c2SStefano Zampini     PetscViewer dbg_viewer = NULL;
42406e683305SStefano Zampini     if (pcbddc->dbg_flag) {
42416e683305SStefano Zampini       dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is));
42426e683305SStefano Zampini       ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr);
42436e683305SStefano Zampini     }
4244f9eb5b7dSStefano Zampini     if (!pcbddc->coarse_ksp) {
4245312be037SStefano Zampini       char prefix[256],str_level[16];
4246e604994aSStefano Zampini       size_t len;
42476e683305SStefano Zampini       ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr);
4248c8587f34SStefano Zampini       ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr);
4249f9eb5b7dSStefano Zampini       ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr);
42505f76c7aeSStefano Zampini       ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr);
4251c8587f34SStefano Zampini       ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr);
42526e683305SStefano Zampini       ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr);
4253c8587f34SStefano Zampini       ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr);
4254c8587f34SStefano Zampini       ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr);
42553e3c6dadSStefano Zampini       ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr);
4256e604994aSStefano Zampini       /* prefix */
4257e604994aSStefano Zampini       ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr);
4258e604994aSStefano Zampini       ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr);
4259e604994aSStefano Zampini       if (!pcbddc->current_level) {
4260e604994aSStefano Zampini         ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr);
4261e604994aSStefano Zampini         ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr);
4262c8587f34SStefano Zampini       } else {
4263e604994aSStefano Zampini         ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr);
4264312be037SStefano Zampini         if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
4265312be037SStefano Zampini         if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
426634d6797cSStefano Zampini         ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr);
4267312be037SStefano Zampini         sprintf(str_level,"l%d_",(int)(pcbddc->current_level));
4268e604994aSStefano Zampini         ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr);
4269e604994aSStefano Zampini       }
4270e604994aSStefano Zampini       ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr);
42713e3c6dadSStefano Zampini       /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
42723e3c6dadSStefano Zampini       ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr);
42733e3c6dadSStefano Zampini       ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr);
42743e3c6dadSStefano Zampini       ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr);
4275f9eb5b7dSStefano Zampini       /* allow user customization */
4276f9eb5b7dSStefano Zampini       ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr);
42773e3c6dadSStefano Zampini     }
42783e3c6dadSStefano Zampini     /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
42793e3c6dadSStefano Zampini     if (nisdofs) {
42803e3c6dadSStefano Zampini       ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr);
42813e3c6dadSStefano Zampini       for (i=0;i<nisdofs;i++) {
42823e3c6dadSStefano Zampini         ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr);
42833e3c6dadSStefano Zampini       }
42843e3c6dadSStefano Zampini     }
42853e3c6dadSStefano Zampini     if (nisneu) {
42863e3c6dadSStefano Zampini       ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr);
42873e3c6dadSStefano Zampini       ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr);
4288312be037SStefano Zampini     }
4289f9eb5b7dSStefano Zampini 
4290f9eb5b7dSStefano Zampini     /* get some info after set from options */
4291f9eb5b7dSStefano Zampini     ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr);
4292f9eb5b7dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr);
4293f9eb5b7dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr);
42944f3a063dSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr);
42956e683305SStefano Zampini     if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */
4296f9eb5b7dSStefano Zampini       ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr);
4297f9eb5b7dSStefano Zampini       isbddc = PETSC_FALSE;
4298f9eb5b7dSStefano Zampini     }
42994f3a063dSStefano Zampini     if (isredundant) {
43004f3a063dSStefano Zampini       KSP inner_ksp;
43014f3a063dSStefano Zampini       PC inner_pc;
43024f3a063dSStefano Zampini       ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr);
43034f3a063dSStefano Zampini       ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr);
43044f3a063dSStefano Zampini       ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr);
43054f3a063dSStefano Zampini     }
4306f9eb5b7dSStefano Zampini 
4307f9eb5b7dSStefano Zampini     /* assemble coarse matrix */
4308fa7f1dd8SStefano Zampini     if (coarse_reuse) {
430981d9aea3SBarry Smith       ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr);
4310fa7f1dd8SStefano Zampini       ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr);
43116e683305SStefano Zampini       coarse_mat_reuse = MAT_REUSE_MATRIX;
4312fa7f1dd8SStefano Zampini     } else {
43136e683305SStefano Zampini       coarse_mat_reuse = MAT_INITIAL_MATRIX;
4314fa7f1dd8SStefano Zampini     }
4315c8587f34SStefano Zampini     if (isbddc || isnn) {
431622bc73bbSStefano Zampini       if (pcbddc->coarsening_ratio > 1) {
431770cf5478SStefano Zampini         if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */
431828143c3dSStefano Zampini           ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr);
431922b6e8a2SStefano Zampini           if (pcbddc->dbg_flag) {
43206e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
43216e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr);
43226e683305SStefano Zampini             ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr);
43236e683305SStefano Zampini             ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr);
432422b6e8a2SStefano Zampini           }
432570cf5478SStefano Zampini         }
43266e683305SStefano Zampini         ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr);
432770cf5478SStefano Zampini       } else {
432822bc73bbSStefano Zampini         ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr);
432922bc73bbSStefano Zampini         coarse_mat = coarse_mat_is;
433022bc73bbSStefano Zampini       }
433122bc73bbSStefano Zampini     } else {
43322e1e5fa4SStefano Zampini       ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr);
4333c8587f34SStefano Zampini     }
4334c8587f34SStefano Zampini     ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr);
4335c8587f34SStefano Zampini 
4336c8587f34SStefano Zampini     /* propagate symmetry info to coarse matrix */
4337b9d89cd5SStefano Zampini     ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr);
43385a16e3a0SStefano Zampini     ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr);
4339c8587f34SStefano Zampini 
43406e683305SStefano Zampini     /* set operators */
43415f76c7aeSStefano Zampini     ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr);
43426e683305SStefano Zampini     if (pcbddc->dbg_flag) {
43436e683305SStefano Zampini       ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr);
43446e683305SStefano Zampini     }
43456e683305SStefano Zampini   } else { /* processes non partecipating to coarse solver (if any) */
43466e683305SStefano Zampini     coarse_mat = 0;
43476e683305SStefano Zampini   }
43486e683305SStefano Zampini   ierr = PetscFree(isarray);CHKERRQ(ierr);
4349b9b85e73SStefano Zampini #if 0
4350b9b85e73SStefano Zampini   {
4351b9b85e73SStefano Zampini     PetscViewer viewer;
4352b9b85e73SStefano Zampini     char filename[256];
4353b9b85e73SStefano Zampini     sprintf(filename,"coarse_mat.m");
4354b9b85e73SStefano Zampini     ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr);
4355b9b85e73SStefano Zampini     ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr);
4356b9b85e73SStefano Zampini     ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr);
4357b9b85e73SStefano Zampini     ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr);
4358b9b85e73SStefano Zampini   }
4359b9b85e73SStefano Zampini #endif
4360c8587f34SStefano Zampini 
4361c8587f34SStefano Zampini   /* Compute coarse null space (special handling by BDDC only) */
4362c8587f34SStefano Zampini   if (pcbddc->NullSpace) {
4363c8587f34SStefano Zampini     ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr);
436498a51de6SStefano Zampini   }
436598a51de6SStefano Zampini 
436698a51de6SStefano Zampini   if (pcbddc->coarse_ksp) {
436798a51de6SStefano Zampini     Vec crhs,csol;
436898a51de6SStefano Zampini     PetscBool ispreonly;
436998a51de6SStefano Zampini     if (CoarseNullSpace) {
4370c8587f34SStefano Zampini       if (isbddc) {
4371c8587f34SStefano Zampini         ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr);
4372c8587f34SStefano Zampini       } else {
4373c8587f34SStefano Zampini         ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr);
4374c8587f34SStefano Zampini       }
4375c8587f34SStefano Zampini     }
4376f9eb5b7dSStefano Zampini     /* setup coarse ksp */
4377f9eb5b7dSStefano Zampini     ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr);
4378f347579bSStefano Zampini     ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr);
4379f347579bSStefano Zampini     ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr);
43806e683305SStefano Zampini     /* hack */
4381f347579bSStefano Zampini     if (!csol) {
43822a7a6963SBarry Smith       ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr);
4383f9eb5b7dSStefano Zampini     }
4384f347579bSStefano Zampini     if (!crhs) {
43852a7a6963SBarry Smith       ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr);
4386f347579bSStefano Zampini     }
4387cbcc2c2aSStefano Zampini     /* Check coarse problem if in debug mode or if solving with an iterative method */
4388cbcc2c2aSStefano Zampini     ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr);
43896e683305SStefano Zampini     if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) {
4390c8587f34SStefano Zampini       KSP       check_ksp;
43912b510759SStefano Zampini       KSPType   check_ksp_type;
4392c8587f34SStefano Zampini       PC        check_pc;
43936e683305SStefano Zampini       Vec       check_vec,coarse_vec;
43946a1308c2SStefano Zampini       PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0;
43952b510759SStefano Zampini       PetscInt  its;
43966e683305SStefano Zampini       PetscBool compute_eigs;
43976e683305SStefano Zampini       PetscReal *eigs_r,*eigs_c;
43986e683305SStefano Zampini       PetscInt  neigs;
43998e185a42SStefano Zampini       const char *prefix;
4400c8587f34SStefano Zampini 
44012b510759SStefano Zampini       /* Create ksp object suitable for estimation of extreme eigenvalues */
44026e683305SStefano Zampini       ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr);
440323ee1639SBarry Smith       ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr);
4404f4ddd8eeSStefano Zampini       ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr);
44052b510759SStefano Zampini       if (ispreonly) {
44062b510759SStefano Zampini         check_ksp_type = KSPPREONLY;
44076e683305SStefano Zampini         compute_eigs = PETSC_FALSE;
44082b510759SStefano Zampini       } else {
4409cbcc2c2aSStefano Zampini         check_ksp_type = KSPGMRES;
44106e683305SStefano Zampini         compute_eigs = PETSC_TRUE;
4411c8587f34SStefano Zampini       }
4412c8587f34SStefano Zampini       ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr);
44136e683305SStefano Zampini       ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr);
44146e683305SStefano Zampini       ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr);
44156e683305SStefano Zampini       ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr);
4416a7dc3881SStefano Zampini       ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr);
4417a7dc3881SStefano Zampini       ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr);
4418a7dc3881SStefano Zampini       ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr);
4419a7dc3881SStefano Zampini       ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr);
4420c8587f34SStefano Zampini       ierr = KSPSetUp(check_ksp);CHKERRQ(ierr);
4421c8587f34SStefano Zampini       ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr);
4422c8587f34SStefano Zampini       ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr);
4423c8587f34SStefano Zampini       /* create random vec */
44246e683305SStefano Zampini       ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr);
44256e683305SStefano Zampini       ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr);
4426c8587f34SStefano Zampini       ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr);
4427c8587f34SStefano Zampini       if (CoarseNullSpace) {
4428c8587f34SStefano Zampini         ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr);
4429c8587f34SStefano Zampini       }
44306e683305SStefano Zampini       ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr);
4431c8587f34SStefano Zampini       /* solve coarse problem */
44326e683305SStefano Zampini       ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr);
4433c8587f34SStefano Zampini       if (CoarseNullSpace) {
44346e683305SStefano Zampini         ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr);
4435c8587f34SStefano Zampini       }
4436cbcc2c2aSStefano Zampini       /* set eigenvalue estimation if preonly has not been requested */
44376e683305SStefano Zampini       if (compute_eigs) {
4438854ce69bSBarry Smith         ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr);
4439854ce69bSBarry Smith         ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr);
44406e683305SStefano Zampini         ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr);
44416e683305SStefano Zampini         lambda_max = eigs_r[neigs-1];
44426e683305SStefano Zampini         lambda_min = eigs_r[0];
44436e683305SStefano Zampini         if (pcbddc->use_coarse_estimates) {
44446e683305SStefano Zampini           if (lambda_max>lambda_min) {
4445cbcc2c2aSStefano Zampini             ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr);
4446cbcc2c2aSStefano Zampini             ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr);
4447cbcc2c2aSStefano Zampini           }
4448c8587f34SStefano Zampini         }
4449c8587f34SStefano Zampini       }
4450cbcc2c2aSStefano Zampini 
4451c8587f34SStefano Zampini       /* check coarse problem residual error */
44526e683305SStefano Zampini       if (pcbddc->dbg_flag) {
44536e683305SStefano Zampini         PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp));
44546e683305SStefano Zampini         ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr);
44556e683305SStefano Zampini         ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr);
4456c8587f34SStefano Zampini         ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr);
44576e683305SStefano Zampini         ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr);
44586e683305SStefano Zampini         ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr);
4459c8587f34SStefano Zampini         ierr = VecDestroy(&check_vec);CHKERRQ(ierr);
4460779c1cceSStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr);
44616e683305SStefano Zampini         ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr);
44626e683305SStefano Zampini         ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr);
44636e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error   : %1.6e\n",infty_error);CHKERRQ(ierr);
44646e683305SStefano Zampini         ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr);
44656e683305SStefano Zampini         if (compute_eigs) {
44666e683305SStefano Zampini           PetscReal lambda_max_s,lambda_min_s;
4467deec49d1SStefano Zampini           ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr);
4468c8587f34SStefano Zampini           ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr);
44696e683305SStefano Zampini           ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr);
44706e683305SStefano 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);
44716e683305SStefano Zampini           for (i=0;i<neigs;i++) {
44726e683305SStefano Zampini             ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr);
4473c8587f34SStefano Zampini           }
44746e683305SStefano Zampini         }
44756e683305SStefano Zampini         ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr);
44766e683305SStefano Zampini         ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr);
44776e683305SStefano Zampini       }
4478c8587f34SStefano Zampini       ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr);
44796e683305SStefano Zampini       if (compute_eigs) {
44806e683305SStefano Zampini         ierr = PetscFree(eigs_r);CHKERRQ(ierr);
44816e683305SStefano Zampini         ierr = PetscFree(eigs_c);CHKERRQ(ierr);
4482c8587f34SStefano Zampini       }
44836e683305SStefano Zampini     }
44846e683305SStefano Zampini   }
4485cbcc2c2aSStefano Zampini   /* print additional info */
4486cbcc2c2aSStefano Zampini   if (pcbddc->dbg_flag) {
44876e683305SStefano Zampini     /* waits until all processes reaches this point */
44886e683305SStefano Zampini     ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr);
4489cbcc2c2aSStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr);
4490cbcc2c2aSStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4491cbcc2c2aSStefano Zampini   }
4492cbcc2c2aSStefano Zampini 
44932b510759SStefano Zampini   /* free memory */
4494c8587f34SStefano Zampini   ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr);
4495fdc635d7SStefano Zampini   ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr);
4496c8587f34SStefano Zampini   PetscFunctionReturn(0);
4497c8587f34SStefano Zampini }
4498674ae819SStefano Zampini 
4499f34684f1SStefano Zampini #undef __FUNCT__
4500f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering"
4501f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n)
4502f34684f1SStefano Zampini {
4503f34684f1SStefano Zampini   PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
4504f34684f1SStefano Zampini   PC_IS*         pcis = (PC_IS*)pc->data;
4505f34684f1SStefano Zampini   Mat_IS*        matis = (Mat_IS*)pc->pmat->data;
450673be2a3aSStefano Zampini   PetscInt       i,coarse_size=0;
450773be2a3aSStefano Zampini   PetscInt       *local_primal_indices=NULL;
4508f34684f1SStefano Zampini   PetscErrorCode ierr;
4509f34684f1SStefano Zampini 
4510f34684f1SStefano Zampini   PetscFunctionBegin;
4511f34684f1SStefano Zampini   /* Compute global number of coarse dofs */
451289c96988SStefano Zampini   if (!pcbddc->primal_indices_local_idxs && pcbddc->local_primal_size) {
451389c96988SStefano Zampini     SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Local primal indices have not been created");
4514727cdba6SStefano Zampini   }
4515727cdba6SStefano Zampini   ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,NULL,&coarse_size,&local_primal_indices);CHKERRQ(ierr);
4516f34684f1SStefano Zampini 
4517f34684f1SStefano Zampini   /* check numbering */
4518f34684f1SStefano Zampini   if (pcbddc->dbg_flag) {
4519f34684f1SStefano Zampini     PetscScalar coarsesum,*array;
4520b9b85e73SStefano Zampini     PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE;
4521f34684f1SStefano Zampini 
4522f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4523f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr);
4524f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr);
45250fccc4e9SStefano Zampini     ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
4526f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr);
4527f34684f1SStefano Zampini     for (i=0;i<pcbddc->local_primal_size;i++) {
4528727cdba6SStefano Zampini       ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr);
4529f34684f1SStefano Zampini     }
4530f34684f1SStefano Zampini     ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr);
4531f34684f1SStefano Zampini     ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr);
4532f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
4533f34684f1SStefano Zampini     ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4534f34684f1SStefano Zampini     ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4535f34684f1SStefano Zampini     ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4536f34684f1SStefano Zampini     ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4537f34684f1SStefano Zampini     ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr);
4538f34684f1SStefano Zampini     for (i=0;i<pcis->n;i++) {
4539f34684f1SStefano Zampini       if (array[i] == 1.0) {
4540b9b85e73SStefano Zampini         set_error = PETSC_TRUE;
4541f34684f1SStefano Zampini         ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr);
4542f34684f1SStefano Zampini       }
4543f34684f1SStefano Zampini     }
4544b9b85e73SStefano Zampini     ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr);
4545f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4546f34684f1SStefano Zampini     for (i=0;i<pcis->n;i++) {
4547f34684f1SStefano Zampini       if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]);
4548f34684f1SStefano Zampini     }
4549f34684f1SStefano Zampini     ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr);
4550f34684f1SStefano Zampini     ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr);
4551f34684f1SStefano Zampini     ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4552f34684f1SStefano Zampini     ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
4553f34684f1SStefano Zampini     ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr);
4554f34684f1SStefano Zampini     ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr);
4555b9b85e73SStefano Zampini     if (pcbddc->dbg_flag > 1 || set_error_reduced) {
4556ca8b9ea9SStefano Zampini       PetscInt *gidxs;
4557ca8b9ea9SStefano Zampini 
4558ca8b9ea9SStefano Zampini       ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr);
4559ca8b9ea9SStefano Zampini       ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr);
4560f34684f1SStefano Zampini       ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr);
4561f34684f1SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4562f34684f1SStefano Zampini       ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr);
4563f34684f1SStefano Zampini       for (i=0;i<pcbddc->local_primal_size;i++) {
45644bc2dc4bSStefano 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);
4565f34684f1SStefano Zampini       }
4566f34684f1SStefano Zampini       ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4567ca8b9ea9SStefano Zampini       ierr = PetscFree(gidxs);CHKERRQ(ierr);
4568f34684f1SStefano Zampini     }
4569f34684f1SStefano Zampini     ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
4570302440fdSBarry Smith     if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed");
4571f34684f1SStefano Zampini   }
45728bec7fa6SStefano Zampini   /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */
4573f34684f1SStefano Zampini   /* get back data */
4574f34684f1SStefano Zampini   *coarse_size_n = coarse_size;
4575f34684f1SStefano Zampini   *local_primal_indices_n = local_primal_indices;
4576674ae819SStefano Zampini   PetscFunctionReturn(0);
4577674ae819SStefano Zampini }
4578674ae819SStefano Zampini 
4579e456f2a8SStefano Zampini #undef __FUNCT__
4580e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal"
4581a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis)
4582e456f2a8SStefano Zampini {
4583e456f2a8SStefano Zampini   IS             localis_t;
4584a7dc3881SStefano Zampini   PetscInt       i,lsize,*idxs,n;
4585e456f2a8SStefano Zampini   PetscScalar    *vals;
4586e456f2a8SStefano Zampini   PetscErrorCode ierr;
4587e456f2a8SStefano Zampini 
4588e456f2a8SStefano Zampini   PetscFunctionBegin;
4589a7dc3881SStefano Zampini   /* get indices in local ordering exploiting local to global map */
4590e456f2a8SStefano Zampini   ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr);
4591854ce69bSBarry Smith   ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr);
4592e456f2a8SStefano Zampini   for (i=0;i<lsize;i++) vals[i] = 1.0;
4593e456f2a8SStefano Zampini   ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr);
4594a7dc3881SStefano Zampini   ierr = VecSet(gwork,0.0);CHKERRQ(ierr);
4595a7dc3881SStefano Zampini   ierr = VecSet(lwork,0.0);CHKERRQ(ierr);
45961035eff8SStefano Zampini   if (idxs) { /* multilevel guard */
4597a7dc3881SStefano Zampini     ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr);
45981035eff8SStefano Zampini   }
4599a7dc3881SStefano Zampini   ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr);
4600e456f2a8SStefano Zampini   ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr);
4601e456f2a8SStefano Zampini   ierr = PetscFree(vals);CHKERRQ(ierr);
4602a7dc3881SStefano Zampini   ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr);
4603a7dc3881SStefano Zampini   /* now compute set in local ordering */
4604a7dc3881SStefano Zampini   ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4605a7dc3881SStefano Zampini   ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
4606a7dc3881SStefano Zampini   ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr);
4607a7dc3881SStefano Zampini   ierr = VecGetSize(lwork,&n);CHKERRQ(ierr);
4608a7dc3881SStefano Zampini   for (i=0,lsize=0;i<n;i++) {
4609ff92baa0SMatthew G. Knepley     if (PetscRealPart(vals[i]) > 0.5) {
4610e456f2a8SStefano Zampini       lsize++;
4611e456f2a8SStefano Zampini     }
4612e456f2a8SStefano Zampini   }
4613854ce69bSBarry Smith   ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr);
4614a7dc3881SStefano Zampini   for (i=0,lsize=0;i<n;i++) {
4615ff92baa0SMatthew G. Knepley     if (PetscRealPart(vals[i]) > 0.5) {
4616e456f2a8SStefano Zampini       idxs[lsize++] = i;
4617e456f2a8SStefano Zampini     }
4618e456f2a8SStefano Zampini   }
4619a7dc3881SStefano Zampini   ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr);
4620a7dc3881SStefano Zampini   ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr);
4621e456f2a8SStefano Zampini   *localis = localis_t;
4622e456f2a8SStefano Zampini   PetscFunctionReturn(0);
4623e456f2a8SStefano Zampini }
4624906d46d4SStefano Zampini 
4625906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */
4626906d46d4SStefano Zampini #undef __FUNCT__
4627906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private"
4628906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y)
4629906d46d4SStefano Zampini {
4630906d46d4SStefano Zampini   PCBDDCChange_ctx change_ctx;
4631906d46d4SStefano Zampini   PetscErrorCode   ierr;
4632906d46d4SStefano Zampini 
4633906d46d4SStefano Zampini   PetscFunctionBegin;
4634906d46d4SStefano Zampini   ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr);
4635906d46d4SStefano Zampini   ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr);
4636906d46d4SStefano Zampini   ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr);
4637906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr);
4638906d46d4SStefano Zampini   PetscFunctionReturn(0);
4639906d46d4SStefano Zampini }
4640906d46d4SStefano Zampini 
4641906d46d4SStefano Zampini #undef __FUNCT__
4642906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private"
4643906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y)
4644906d46d4SStefano Zampini {
4645906d46d4SStefano Zampini   PCBDDCChange_ctx change_ctx;
4646906d46d4SStefano Zampini   PetscErrorCode   ierr;
4647906d46d4SStefano Zampini 
4648906d46d4SStefano Zampini   PetscFunctionBegin;
4649906d46d4SStefano Zampini   ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr);
4650906d46d4SStefano Zampini   ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr);
4651906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr);
4652906d46d4SStefano Zampini   ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr);
4653906d46d4SStefano Zampini   PetscFunctionReturn(0);
4654906d46d4SStefano Zampini }
4655b96c3477SStefano Zampini 
4656b96c3477SStefano Zampini #undef __FUNCT__
4657b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs"
465808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc)
4659b96c3477SStefano Zampini {
4660b96c3477SStefano Zampini   PC_BDDC             *pcbddc=(PC_BDDC*)pc->data;
4661b96c3477SStefano Zampini   PCBDDCSubSchurs     sub_schurs=pcbddc->sub_schurs;
4662b96c3477SStefano Zampini   PetscInt            *used_xadj,*used_adjncy;
4663b96c3477SStefano Zampini   PetscBool           free_used_adj;
4664b96c3477SStefano Zampini   PetscErrorCode      ierr;
4665b96c3477SStefano Zampini 
4666b96c3477SStefano Zampini   PetscFunctionBegin;
4667b96c3477SStefano Zampini   /* decide the adjacency to be used for determining internal problems for local schur on subsets */
4668b96c3477SStefano Zampini   free_used_adj = PETSC_FALSE;
466908122e43SStefano Zampini   if (pcbddc->sub_schurs_layers == -1) {
4670b96c3477SStefano Zampini     used_xadj = NULL;
4671b96c3477SStefano Zampini     used_adjncy = NULL;
4672b96c3477SStefano Zampini   } else {
467308122e43SStefano Zampini     if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) {
467408122e43SStefano Zampini       used_xadj = pcbddc->mat_graph->xadj;
467508122e43SStefano Zampini       used_adjncy = pcbddc->mat_graph->adjncy;
467608122e43SStefano Zampini     } else if (pcbddc->computed_rowadj) {
4677b96c3477SStefano Zampini       used_xadj = pcbddc->mat_graph->xadj;
4678b96c3477SStefano Zampini       used_adjncy = pcbddc->mat_graph->adjncy;
4679b96c3477SStefano Zampini     } else {
46802fffb893SStefano Zampini       PetscBool      flg_row=PETSC_FALSE;
4681b96c3477SStefano Zampini       const PetscInt *xadj,*adjncy;
4682b96c3477SStefano Zampini       PetscInt       nvtxs;
4683b96c3477SStefano Zampini 
46842fffb893SStefano Zampini       ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
46852fffb893SStefano Zampini       if (flg_row) {
4686b96c3477SStefano Zampini         ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr);
4687b96c3477SStefano Zampini         ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr);
4688b96c3477SStefano Zampini         ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr);
4689b96c3477SStefano Zampini         free_used_adj = PETSC_TRUE;
46902fffb893SStefano Zampini       } else {
46912fffb893SStefano Zampini         pcbddc->sub_schurs_layers = -1;
46922fffb893SStefano Zampini         used_xadj = NULL;
46932fffb893SStefano Zampini         used_adjncy = NULL;
46942fffb893SStefano Zampini       }
46952fffb893SStefano Zampini       ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr);
4696b96c3477SStefano Zampini     }
4697b96c3477SStefano Zampini   }
4698d5574798SStefano Zampini 
4699d5574798SStefano Zampini   /* set Schur complement solver for interior variables (used only when MUMPS is not present) */
4700d5574798SStefano Zampini   if (!sub_schurs->use_mumps) {
4701d5574798SStefano Zampini     ierr = MatSchurComplementSetKSP(sub_schurs->S,pcbddc->ksp_D);CHKERRQ(ierr);
4702d5574798SStefano Zampini   }
4703d5574798SStefano Zampini 
4704d5574798SStefano Zampini   /* setup sub_schurs data */
4705c3dc5de7SStefano Zampini   ierr = PCBDDCSubSchursSetUp(sub_schurs,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr);
4706b96c3477SStefano Zampini 
4707b96c3477SStefano Zampini   /* free adjacency */
4708b96c3477SStefano Zampini   if (free_used_adj) {
4709b96c3477SStefano Zampini     ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr);
4710b96c3477SStefano Zampini   }
4711b96c3477SStefano Zampini   PetscFunctionReturn(0);
4712b96c3477SStefano Zampini }
4713b96c3477SStefano Zampini 
4714b96c3477SStefano Zampini #undef __FUNCT__
4715b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs"
471608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc)
4717b96c3477SStefano Zampini {
4718b96c3477SStefano Zampini   PC_IS               *pcis=(PC_IS*)pc->data;
4719b96c3477SStefano Zampini   PC_BDDC             *pcbddc=(PC_BDDC*)pc->data;
4720b96c3477SStefano Zampini   PCBDDCSubSchurs     sub_schurs=pcbddc->sub_schurs;
4721b96c3477SStefano Zampini   PCBDDCGraph         graph;
4722b96c3477SStefano Zampini   Mat                 S_j;
4723b96c3477SStefano Zampini   PetscErrorCode      ierr;
4724b96c3477SStefano Zampini 
4725b96c3477SStefano Zampini   PetscFunctionBegin;
4726b96c3477SStefano Zampini   /* attach interface graph for determining subsets */
472708122e43SStefano Zampini   if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */
4728b96c3477SStefano Zampini     IS verticesIS;
4729b96c3477SStefano Zampini 
4730b96c3477SStefano Zampini     ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr);
4731b96c3477SStefano Zampini     ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr);
4732b96c3477SStefano Zampini     ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap);CHKERRQ(ierr);
4733b96c3477SStefano Zampini     ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr);
4734b96c3477SStefano Zampini     ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr);
4735b96c3477SStefano Zampini     ierr = ISDestroy(&verticesIS);CHKERRQ(ierr);
4736b96c3477SStefano Zampini /*
4737b96c3477SStefano Zampini     if (pcbddc->dbg_flag) {
4738b96c3477SStefano Zampini       ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr);
4739b96c3477SStefano Zampini     }
4740b96c3477SStefano Zampini */
4741b96c3477SStefano Zampini   } else {
4742b96c3477SStefano Zampini     graph = pcbddc->mat_graph;
4743b96c3477SStefano Zampini   }
4744b96c3477SStefano Zampini 
4745b96c3477SStefano Zampini   /* Create Schur complement matrix */
4746b96c3477SStefano Zampini   ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr);
4747b96c3477SStefano Zampini 
4748b96c3477SStefano Zampini   /* sub_schurs init */
47495a95e1ceSStefano Zampini   ierr = PCBDDCSubSchursInit(sub_schurs,pcbddc->local_mat,S_j,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr);
4750b96c3477SStefano Zampini   ierr = MatDestroy(&S_j);CHKERRQ(ierr);
4751b96c3477SStefano Zampini   /* free graph struct */
475208122e43SStefano Zampini   if (pcbddc->sub_schurs_rebuild) {
4753b96c3477SStefano Zampini     ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr);
4754b96c3477SStefano Zampini   }
4755b96c3477SStefano Zampini   PetscFunctionReturn(0);
4756b96c3477SStefano Zampini }
4757